feat(governance): add durable Owner authorization projection

active-runtime/app/aios_all_modified_files_packager.py

@@ -0,0 +1,470 @@
+#!/usr/bin/env python3
+"""Canonical AIOS all-modified-files evidence packager.
+
+The bundle contract is intentionally compatible with the historical
+aios_package_all_modified_files_v1.sh output while keeping secret scanning in
+version-controlled, tested code.
+"""
+
+from __future__ import annotations
+
+from datetime import datetime, timezone
+import argparse
+import hashlib
+import json
+import os
+from pathlib import Path
+import re
+import subprocess
+import tempfile
+import zipfile
+from typing import Iterable
+
+OID_RE = re.compile(r"^[0-9a-f]{40}$")
+TASK_RE = re.compile(r"^[A-Za-z0-9][A-Za-z0-9._:-]{2,179}$")
+
+# A PEM private-key finding requires the complete PEM header. Source code,
+# tests, docs, or error text containing only "BEGIN PRIVATE KEY" are not
+# credentials and must not be rejected.
+PEM_PRIVATE_KEY_RE = re.compile(
+    rb"-----BEGIN (?:RSA |EC |DSA |OPENSSH )?PRIVATE KEY-----",
+    re.IGNORECASE,
+)
+BEARER_RE = re.compile(
+    rb"authorization\s*:\s*bearer\s+[A-Za-z0-9._~+/=-]{20,}",
+    re.IGNORECASE,
+)
+KEY_VALUE_SECRET_RE = re.compile(
+    rb"(?:password|passwd|secret|api[_-]?key)\s*[:=]\s*[^\s,;]{12,}",
+    re.IGNORECASE,
+)
+
+DOCUMENT_SUFFIXES = {".md", ".rst", ".adoc", ".txt"}
+DOCUMENT_PREFIXES = (
+    "README",
+    "CHANGELOG",
+    "RUNBOOK",
+    "ADR",
+    "ARCHITECTURE",
+    "DEPLOYMENT",
+    "TROUBLESHOOTING",
+)
+
+
+class PackagerError(RuntimeError):
+    pass
+
+
+def sha256(data: bytes) -> str:
+    return hashlib.sha256(data).hexdigest()
+
+
+def secret_finding(data: bytes) -> str | None:
+    if PEM_PRIVATE_KEY_RE.search(data):
+        return "PEM_PRIVATE_KEY"
+    if BEARER_RE.search(data):
+        return "AUTHORIZATION_BEARER"
+    if KEY_VALUE_SECRET_RE.search(data):
+        return "KEY_VALUE_SECRET"
+    return None
+
+
+def scan_secret(data: bytes, label: str) -> None:
+    finding = secret_finding(data)
+    if finding is not None:
+        raise PackagerError(
+            f"MODIFIED_FILES_SECRET_SCAN_MATCH:{finding}:{label}"
+        )
+
+
+def git_bytes(git_dir: Path, oid: str, path: str) -> bytes:
+    try:
+        return subprocess.check_output(
+            [
+                "/usr/bin/git",
+                f"--git-dir={git_dir}",
+                "show",
+                f"{oid}:{path}",
+            ],
+            timeout=30,
+        )
+    except subprocess.CalledProcessError as exc:
+        raise PackagerError(
+            "MODIFIED_FILES_GIT_SHOW_FAILED:" + path
+        ) from exc
+
+
+def git_text(git_dir: Path, *args: str) -> str:
+    try:
+        return subprocess.check_output(
+            ["/usr/bin/git", f"--git-dir={git_dir}", *args],
+            text=True,
+            timeout=30,
+        )
+    except subprocess.CalledProcessError as exc:
+        raise PackagerError(
+            "MODIFIED_FILES_GIT_COMMAND_FAILED:" + " ".join(args)
+        ) from exc
+
+
+def validate_lineage(git_dir: Path, base: str, candidate: str) -> None:
+    if OID_RE.fullmatch(base) is None or OID_RE.fullmatch(candidate) is None:
+        raise PackagerError("MODIFIED_FILES_OID_INVALID")
+    for oid in (base, candidate):
+        try:
+            subprocess.run(
+                [
+                    "/usr/bin/git",
+                    f"--git-dir={git_dir}",
+                    "cat-file",
+                    "-e",
+                    f"{oid}^{{commit}}",
+                ],
+                check=True,
+                timeout=30,
+            )
+        except subprocess.CalledProcessError as exc:
+            raise PackagerError("MODIFIED_FILES_COMMIT_MISSING:" + oid) from exc
+    if subprocess.run(
+        [
+            "/usr/bin/git",
+            f"--git-dir={git_dir}",
+            "merge-base",
+            "--is-ancestor",
+            base,
+            candidate,
+        ],
+        check=False,
+        timeout=30,
+    ).returncode != 0:
+        raise PackagerError("MODIFIED_FILES_LINEAGE_INVALID")
+
+
+def safe_destination(root: Path, relative: str) -> Path:
+    path = Path(relative)
+    if path.is_absolute() or ".." in path.parts:
+        raise PackagerError("MODIFIED_FILE_PATH_INVALID:" + relative)
+    destination = root / path
+    destination.parent.mkdir(parents=True, exist_ok=True)
+    return destination
+
+
+def is_document_path(path: str) -> bool:
+    value = Path(path)
+    if path.startswith("docs/"):
+        return True
+    if value.suffix.lower() in DOCUMENT_SUFFIXES:
+        return True
+    upper = value.name.upper()
+    return any(upper.startswith(prefix) for prefix in DOCUMENT_PREFIXES)
+
+
+def diff_entries(
+    git_dir: Path,
+    base: str,
+    candidate: str,
+    before_root: Path,
+    after_root: Path,
+) -> list[dict[str, object]]:
+    raw = git_text(
+        git_dir,
+        "diff",
+        "--name-status",
+        "-M",
+        "-C",
+        base,
+        candidate,
+    )
+    entries: list[dict[str, object]] = []
+    for line in raw.splitlines():
+        if not line.strip():
+            continue
+        parts = line.split("\t")
+        status = parts[0]
+        code = status[0]
+        old_path: str | None = None
+        new_path: str | None = None
+        if code in {"R", "C"}:
+            if len(parts) != 3:
+                raise PackagerError(
+                    "MODIFIED_FILES_DIFF_PARSE_INVALID:" + line
+                )
+            old_path, new_path = parts[1], parts[2]
+        else:
+            if len(parts) != 2:
+                raise PackagerError(
+                    "MODIFIED_FILES_DIFF_PARSE_INVALID:" + line
+                )
+            new_path = parts[1] if code != "D" else None
+            old_path = parts[1] if code != "A" else None
+
+        entry: dict[str, object] = {"status": status}
+        if old_path is not None:
+            entry["previous_path"] = old_path
+            before = git_bytes(git_dir, base, old_path)
+            scan_secret(before, "before:" + old_path)
+            safe_destination(before_root, old_path).write_bytes(before)
+            entry["before_sha256"] = sha256(before)
+            entry["before_bytes"] = len(before)
+        if new_path is not None:
+            entry["path"] = new_path
+            after = git_bytes(git_dir, candidate, new_path)
+            scan_secret(after, "after:" + new_path)
+            safe_destination(after_root, new_path).write_bytes(after)
+            entry["after_sha256"] = sha256(after)
+            entry["after_bytes"] = len(after)
+        entry["deleted"] = code == "D"
+        entry["added"] = code == "A"
+        entry["renamed"] = code == "R"
+        entry["copied"] = code == "C"
+        entries.append(entry)
+    return entries
+
+
+def execution_entries(
+    extra_root: Path | None,
+    execution_root: Path,
+) -> list[dict[str, object]]:
+    if extra_root is None or not extra_root.exists():
+        return []
+    if extra_root.is_symlink() or not extra_root.is_dir():
+        raise PackagerError("EXTRA_FILES_DIR_INVALID")
+
+    entries: list[dict[str, object]] = []
+    for path in sorted(extra_root.rglob("*")):
+        if path.is_symlink():
+            raise PackagerError(
+                "EXTRA_FILE_SYMLINK_REJECTED:" + str(path)
+            )
+        if not path.is_file():
+            continue
+        relative = path.relative_to(extra_root).as_posix()
+        data = path.read_bytes()
+        scan_secret(data, "execution:" + relative)
+        safe_destination(execution_root, relative).write_bytes(data)
+        entries.append(
+            {
+                "path": relative,
+                "sha256": sha256(data),
+                "bytes": len(data),
+            }
+        )
+    return entries
+
+
+def write_checksums(bundle: Path) -> None:
+    checks: list[tuple[str, str]] = []
+    for path in sorted(bundle.rglob("*")):
+        if path.is_file() and path.name != "checksums.sha256":
+            checks.append(
+                (
+                    sha256(path.read_bytes()),
+                    path.relative_to(bundle).as_posix(),
+                )
+            )
+    (bundle / "checksums.sha256").write_text(
+        "".join(f"{digest}  {name}\n" for digest, name in checks),
+        encoding="utf-8",
+    )
+
+
+def verify_checksums(bundle: Path) -> None:
+    checksum_path = bundle / "checksums.sha256"
+    for line in checksum_path.read_text(encoding="utf-8").splitlines():
+        digest, relative = line.split("  ", 1)
+        path = bundle / relative
+        if not path.is_file() or sha256(path.read_bytes()) != digest:
+            raise PackagerError(
+                "MODIFIED_FILES_CHECKSUM_CLOSURE_INVALID:" + relative
+            )
+
+
+def build_bundle(
+    *,
+    git_dir: Path,
+    base: str,
+    candidate: str,
+    task_id: str,
+    extra_dir: Path | None = None,
+    output_root: Path | None = None,
+    now: datetime | None = None,
+) -> dict[str, object]:
+    if TASK_RE.fullmatch(task_id) is None:
+        raise PackagerError("TASK_ID_INVALID")
+    validate_lineage(git_dir, base, candidate)
+
+    timestamp = (now or datetime.now(timezone.utc)).strftime(
+        "%Y%m%dT%H%M%SZ"
+    )
+    safe_task = re.sub(r"[^A-Za-z0-9._-]+", "-", task_id).strip("-")
+    if output_root is None:
+        output_root = Path(
+            tempfile.mkdtemp(
+                prefix=f"aios-modified-files-bundle-{safe_task}-{timestamp}-",
+                dir="/var/tmp",
+            )
+        )
+    else:
+        output_root.mkdir(parents=True, exist_ok=True)
+
+    bundle = output_root / "bundle"
+    before_root = bundle / "repo" / "before"
+    after_root = bundle / "repo" / "after"
+    execution_root = bundle / "execution"
+    before_root.mkdir(parents=True)
+    after_root.mkdir(parents=True)
+    execution_root.mkdir(parents=True)
+
+    changes = diff_entries(
+        git_dir,
+        base,
+        candidate,
+        before_root,
+        after_root,
+    )
+    execution_files = execution_entries(extra_dir, execution_root)
+
+    document_changes = [
+        entry
+        for entry in changes
+        if is_document_path(
+            str(entry.get("path") or entry.get("previous_path") or "")
+        )
+    ]
+    execution_documents = [
+        entry
+        for entry in execution_files
+        if is_document_path(str(entry["path"]))
+    ]
+
+    documents_index = {
+        "schema": "aios.modified-documents.index.v1",
+        "task_id": task_id,
+        "base_oid": base,
+        "candidate_oid": candidate,
+        "repository_documents": document_changes,
+        "execution_documents": execution_documents,
+        "repository_document_count": len(document_changes),
+        "execution_document_count": len(execution_documents),
+    }
+    (bundle / "documents-index.json").write_text(
+        json.dumps(documents_index, indent=2, sort_keys=True) + "\n",
+        encoding="utf-8",
+    )
+
+    manifest = {
+        "schema": "aios.all-modified-files.bundle.v1",
+        "task_id": task_id,
+        "base_oid": base,
+        "candidate_oid": candidate,
+        "generated_at": (
+            now or datetime.now(timezone.utc)
+        ).isoformat(),
+        "changed_file_count": len(changes),
+        "execution_file_count": len(execution_files),
+        "repository_document_count": len(document_changes),
+        "execution_document_count": len(execution_documents),
+        "changes": changes,
+        "execution_files": execution_files,
+    }
+    (bundle / "all-modified-files-manifest.json").write_text(
+        json.dumps(manifest, indent=2, sort_keys=True) + "\n",
+        encoding="utf-8",
+    )
+    write_checksums(bundle)
+    verify_checksums(bundle)
+
+    zip_name = (
+        f"AIOS_{task_id}_ALL_MODIFIED_FILES_{candidate[:12]}.zip"
+        .replace("/", "-")
+    )
+    zip_path = output_root / zip_name
+    with zipfile.ZipFile(
+        zip_path,
+        "w",
+        compression=zipfile.ZIP_DEFLATED,
+    ) as archive:
+        for path in sorted(bundle.rglob("*")):
+            if path.is_symlink():
+                raise PackagerError(
+                    "BUNDLE_SYMLINK_REJECTED:" + str(path)
+                )
+            if path.is_file():
+                archive.write(
+                    path,
+                    arcname=path.relative_to(bundle).as_posix(),
+                )
+
+    zip_sha = sha256(zip_path.read_bytes())
+    summary = {
+        "schema": "aios.all-modified-files.bundle-summary.v1",
+        "task_id": task_id,
+        "base_oid": base,
+        "candidate_oid": candidate,
+        "bundle_path": str(zip_path),
+        "bundle_sha256": zip_sha,
+        "changed_file_count": len(changes),
+        "execution_file_count": len(execution_files),
+        "repository_document_count": len(document_changes),
+        "execution_document_count": len(execution_documents),
+    }
+    (output_root / "ALL_MODIFIED_FILES_BUNDLE.json").write_text(
+        json.dumps(summary, indent=2, sort_keys=True) + "\n",
+        encoding="utf-8",
+    )
+    return summary
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser()
+    parser.add_argument("base")
+    parser.add_argument("candidate")
+    parser.add_argument("task_id")
+    parser.add_argument("extra_dir", nargs="?")
+    parser.add_argument(
+        "--authority-git-dir",
+        default=os.environ.get(
+            "AIOS_AUTHORITY_GIT_DIR",
+            "/data/aios-gitops/repo.git",
+        ),
+    )
+    parser.add_argument("--output-root")
+    args = parser.parse_args()
+
+    try:
+        summary = build_bundle(
+            git_dir=Path(args.authority_git_dir),
+            base=args.base,
+            candidate=args.candidate,
+            task_id=args.task_id,
+            extra_dir=Path(args.extra_dir) if args.extra_dir else None,
+            output_root=(
+                Path(args.output_root) if args.output_root else None
+            ),
+        )
+    except PackagerError as exc:
+        print("FATAL:" + str(exc))
+        return 20
+
+    print("ALL_MODIFIED_FILES_BUNDLE=PASS")
+    print("ALL_MODIFIED_FILES_BUNDLE_PATH=" + str(summary["bundle_path"]))
+    print("ALL_MODIFIED_FILES_BUNDLE_SHA256=" + str(summary["bundle_sha256"]))
+    print("CHANGED_FILE_COUNT=" + str(summary["changed_file_count"]))
+    print("EXECUTION_FILE_COUNT=" + str(summary["execution_file_count"]))
+    print(
+        "REPOSITORY_DOCUMENT_COUNT="
+        + str(summary["repository_document_count"])
+    )
+    print(
+        "EXECUTION_DOCUMENT_COUNT="
+        + str(summary["execution_document_count"])
+    )
+    print(
+        "ALL_MODIFIED_FILES_EVIDENCE="
+        + str(Path(str(summary["bundle_path"])).parent)
+    )
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

active-runtime/app/aios_durable_owner_authorization.py

@@ -0,0 +1,770 @@
+#!/usr/bin/env python3
+"""Durable Owner Authorization broker for unattended AIOS preparation work.
+
+This module deliberately separates durable Owner intent authorization from
+cryptographic release approval. A durable grant may authorize preparation
+operations such as governed-task projection and release-candidate creation,
+but it can never authorize deployment, tag publication, baseline finalization
+or runtime switching.
+"""
+
+from __future__ import annotations
+
+from dataclasses import dataclass
+from datetime import datetime, timezone
+import argparse
+import hashlib
+import hmac
+from http.server import BaseHTTPRequestHandler
+import json
+import os
+from pathlib import Path
+import re
+import sqlite3
+import socketserver
+import stat
+import tempfile
+import threading
+import urllib.error
+import urllib.request
+from typing import Any, Mapping
+
+POLICY_SCHEMA = "aios.owner-durable-authorization.v1"
+RECEIPT_SCHEMA = "aios.owner-durable-authorization-receipt.v1"
+DECISION_SCHEMA = "aios.owner-durable-authorization-decision.v1"
+
+ALLOWED_PERSISTENT_CAPABILITIES = frozenset({
+    "governed-task:project",
+    "release:candidate",
+})
+SAFE_PROJECTED_TASK_CAPABILITIES = frozenset({
+    "git.read",
+    "git.test",
+    "governed.write.branch",
+    "governed.write.file",
+})
+FORBIDDEN_PERSISTENT_CAPABILITIES = frozenset({
+    "submitOwnerApproval",
+    "startDeployment",
+    "tag_publish",
+    "baseline_registry_finalize",
+    "runtime_switch",
+    "main_write",
+    "production_deploy",
+    "force_push",
+    "credential_admin",
+})
+ACCOUNT_RE = re.compile(r"^[A-Za-z0-9][A-Za-z0-9._:@/-]{3,255}$")
+GRANT_RE = re.compile(r"^[A-Za-z0-9][A-Za-z0-9._:@/-]{7,255}$")
+OID_RE = re.compile(r"^[0-9a-f]{40}$")
+VERSION_RE = re.compile(r"^V[0-9]+R[0-9]+C[0-9]+$")
+TASK_RE = re.compile(r"^[A-Za-z0-9][A-Za-z0-9._:-]{2,179}$")
+BRANCH_RE = re.compile(r"^agent/[A-Za-z0-9][A-Za-z0-9._/-]{2,179}$")
+JOB_RE = re.compile(r"^job-[0-9a-f]{32}$")
+MAX_BODY = 512 * 1024
+
+
+class AuthorizationError(RuntimeError):
+    def __init__(self, code: str, status: int = 400) -> None:
+        super().__init__(code)
+        self.code = code
+        self.status = status
+
+
+def canonical(value: object) -> bytes:
+    return json.dumps(
+        value,
+        ensure_ascii=False,
+        sort_keys=True,
+        separators=(",", ":"),
+    ).encode("utf-8")
+
+
+def sha256(value: bytes) -> str:
+    return hashlib.sha256(value).hexdigest()
+
+
+def parse_time(value: object, *, field: str) -> datetime:
+    if not isinstance(value, str) or not value.strip():
+        raise AuthorizationError("POLICY_" + field.upper() + "_INVALID")
+    text = value.strip()
+    if text.endswith("Z"):
+        text = text[:-1] + "+00:00"
+    try:
+        result = datetime.fromisoformat(text)
+    except ValueError as exc:
+        raise AuthorizationError("POLICY_" + field.upper() + "_INVALID") from exc
+    if result.tzinfo is None:
+        raise AuthorizationError("POLICY_" + field.upper() + "_INVALID")
+    return result.astimezone(timezone.utc)
+
+
+def validate_root_file(path: Path, *, max_size: int = 256 * 1024) -> Path:
+    if path.is_symlink() or not path.is_file():
+        raise AuthorizationError("POLICY_FILE_INVALID", 503)
+    info = path.stat(follow_symlinks=False)
+    if info.st_uid != 0 or info.st_gid != 0:
+        raise AuthorizationError("POLICY_FILE_OWNERSHIP_INVALID", 503)
+    if stat.S_IMODE(info.st_mode) != 0o600:
+        raise AuthorizationError("POLICY_FILE_MODE_INVALID", 503)
+    if info.st_size < 16 or info.st_size > max_size:
+        raise AuthorizationError("POLICY_FILE_SIZE_INVALID", 503)
+    return path
+
+
+@dataclass(frozen=True)
+class Constraints:
+    repository: str
+    account: str
+    allowed_target_prefix: str
+    release_version_prefix: str
+    max_task_risk: str
+
+
+@dataclass(frozen=True)
+class DurablePolicy:
+    grant_id: str
+    owner_identity: str
+    account: str
+    capabilities: frozenset[str]
+    issued_at: datetime
+    not_before: datetime
+    persistent_until_revoked: bool
+    constraints: Constraints
+    policy_sha256: str
+
+    @classmethod
+    def load(
+        cls,
+        policy_path: Path,
+        revocation_path: Path | None = None,
+        *,
+        now: datetime | None = None,
+    ) -> "DurablePolicy":
+        validate_root_file(policy_path)
+        try:
+            value = json.loads(policy_path.read_text(encoding="utf-8"))
+        except Exception as exc:
+            raise AuthorizationError("POLICY_JSON_INVALID", 503) from exc
+        if not isinstance(value, Mapping):
+            raise AuthorizationError("POLICY_JSON_INVALID", 503)
+        allowed_fields = {
+            "schema",
+            "grant_id",
+            "owner_identity",
+            "account",
+            "capabilities",
+            "issued_at",
+            "not_before",
+            "persistent_until_revoked",
+            "constraints",
+        }
+        if set(value) != allowed_fields or value.get("schema") != POLICY_SCHEMA:
+            raise AuthorizationError("POLICY_FIELDS_INVALID", 503)
+
+        grant_id = str(value.get("grant_id") or "")
+        owner_identity = str(value.get("owner_identity") or "")
+        account = str(value.get("account") or "")
+        if GRANT_RE.fullmatch(grant_id) is None:
+            raise AuthorizationError("POLICY_GRANT_ID_INVALID", 503)
+        if ACCOUNT_RE.fullmatch(owner_identity) is None:
+            raise AuthorizationError("POLICY_OWNER_IDENTITY_INVALID", 503)
+        if ACCOUNT_RE.fullmatch(account) is None:
+            raise AuthorizationError("POLICY_ACCOUNT_INVALID", 503)
+
+        raw_caps = value.get("capabilities")
+        if not isinstance(raw_caps, list) or not raw_caps:
+            raise AuthorizationError("POLICY_CAPABILITIES_INVALID", 503)
+        capabilities = frozenset(str(item) for item in raw_caps)
+        if len(capabilities) != len(raw_caps):
+            raise AuthorizationError("POLICY_CAPABILITIES_DUPLICATE", 503)
+        if not capabilities.issubset(ALLOWED_PERSISTENT_CAPABILITIES):
+            raise AuthorizationError("POLICY_CAPABILITY_NOT_PERSISTABLE", 403)
+        if capabilities & FORBIDDEN_PERSISTENT_CAPABILITIES:
+            raise AuthorizationError("POLICY_FORBIDDEN_CAPABILITY", 403)
+
+        issued_at = parse_time(value.get("issued_at"), field="issued_at")
+        not_before = parse_time(value.get("not_before"), field="not_before")
+        current = now or datetime.now(timezone.utc)
+        if issued_at > not_before or current < not_before:
+            raise AuthorizationError("POLICY_NOT_ACTIVE", 403)
+        if value.get("persistent_until_revoked") is not True:
+            raise AuthorizationError("POLICY_PERSISTENCE_INVALID", 503)
+
+        raw_constraints = value.get("constraints")
+        if not isinstance(raw_constraints, Mapping) or set(raw_constraints) != {
+            "repository",
+            "allowed_target_prefix",
+            "release_version_prefix",
+            "max_task_risk",
+        }:
+            raise AuthorizationError("POLICY_CONSTRAINTS_INVALID", 503)
+        if raw_constraints.get("repository") != "aios/aios":
+            raise AuthorizationError("POLICY_REPOSITORY_INVALID", 503)
+        prefix = str(raw_constraints.get("allowed_target_prefix") or "")
+        if prefix != "agent/":
+            raise AuthorizationError("POLICY_TARGET_PREFIX_INVALID", 503)
+        version_prefix = str(raw_constraints.get("release_version_prefix") or "")
+        if version_prefix != "V2R2C":
+            raise AuthorizationError("POLICY_RELEASE_VERSION_PREFIX_INVALID", 503)
+        max_risk = str(raw_constraints.get("max_task_risk") or "")
+        if max_risk not in {"low", "engineering"}:
+            raise AuthorizationError("POLICY_MAX_RISK_INVALID", 503)
+
+        if revocation_path is not None and revocation_path.exists():
+            validate_root_file(revocation_path, max_size=64 * 1024)
+            try:
+                revoked = json.loads(revocation_path.read_text(encoding="utf-8"))
+            except Exception as exc:
+                raise AuthorizationError("REVOCATION_JSON_INVALID", 503) from exc
+            if not isinstance(revoked, Mapping):
+                raise AuthorizationError("REVOCATION_JSON_INVALID", 503)
+            grants = revoked.get("revoked_grant_ids")
+            if not isinstance(grants, list) or not all(isinstance(x, str) for x in grants):
+                raise AuthorizationError("REVOCATION_FIELDS_INVALID", 503)
+            if grant_id in grants:
+                raise AuthorizationError("OWNER_GRANT_REVOKED", 403)
+
+        return cls(
+            grant_id=grant_id,
+            owner_identity=owner_identity,
+            account=account,
+            capabilities=capabilities,
+            issued_at=issued_at,
+            not_before=not_before,
+            persistent_until_revoked=True,
+            constraints=Constraints(
+                repository="aios/aios",
+                account=account,
+                allowed_target_prefix=prefix,
+                release_version_prefix=version_prefix,
+                max_task_risk=max_risk,
+            ),
+            policy_sha256=sha256(policy_path.read_bytes()),
+        )
+
+    def authorize(self, capability: str, request: Mapping[str, Any]) -> dict[str, Any]:
+        if capability not in self.capabilities:
+            raise AuthorizationError("CAPABILITY_NOT_AUTHORIZED", 403)
+        if capability in FORBIDDEN_PERSISTENT_CAPABILITIES:
+            raise AuthorizationError("CAPABILITY_REQUIRES_SEPARATE_OWNER_APPROVAL", 403)
+
+        if capability == "governed-task:project":
+            required = {
+                "task_id",
+                "repository",
+                "base_oid",
+                "expected_branch_oid",
+                "target_branch",
+                "generation",
+                "risk",
+                "description",
+                "idempotency_key",
+                "priority",
+                "max_attempts",
+                "require_online_task_projection",
+                "online_task",
+                "execution",
+            }
+            optional = {"branch_mode"}
+            if set(request) - (required | optional) or not required.issubset(request):
+                raise AuthorizationError("GOVERNED_TASK_FIELDS_INVALID")
+            if request["repository"] != self.constraints.repository:
+                raise AuthorizationError("REPOSITORY_NOT_AUTHORIZED", 403)
+            if not TASK_RE.fullmatch(str(request["task_id"])):
+                raise AuthorizationError("TASK_ID_INVALID")
+            base = str(request["base_oid"])
+            expected = str(request["expected_branch_oid"])
+            if not OID_RE.fullmatch(base) or expected != base:
+                raise AuthorizationError("BASE_BINDING_INVALID")
+            branch = str(request["target_branch"])
+            if BRANCH_RE.fullmatch(branch) is None or not branch.startswith(
+                self.constraints.allowed_target_prefix
+            ):
+                raise AuthorizationError("TARGET_BRANCH_NOT_AUTHORIZED", 403)
+            if int(request["generation"]) != 1:
+                raise AuthorizationError("GENERATION_NOT_AUTHORIZED", 403)
+            risk = str(request["risk"])
+            allowed_risks = {"low"} if self.constraints.max_task_risk == "low" else {
+                "low",
+                "engineering",
+            }
+            if risk not in allowed_risks:
+                raise AuthorizationError("TASK_RISK_NOT_AUTHORIZED", 403)
+            if request.get("branch_mode", "CREATE_NEW") != "CREATE_NEW":
+                raise AuthorizationError("BRANCH_MODE_NOT_AUTHORIZED", 403)
+            priority = request["priority"]
+            if (
+                not isinstance(priority, int)
+                or isinstance(priority, bool)
+                or not -100 <= priority <= 100
+            ):
+                raise AuthorizationError("TASK_PRIORITY_INVALID")
+            max_attempts = request["max_attempts"]
+            if (
+                not isinstance(max_attempts, int)
+                or isinstance(max_attempts, bool)
+                or not 1 <= max_attempts <= 3
+            ):
+                raise AuthorizationError("TASK_MAX_ATTEMPTS_INVALID")
+            if request["require_online_task_projection"] is not True:
+                raise AuthorizationError("ONLINE_TASK_PROJECTION_REQUIRED", 403)
+
+            online = request["online_task"]
+            execution = request["execution"]
+            if not isinstance(online, Mapping) or not isinstance(execution, Mapping):
+                raise AuthorizationError("ONLINE_TASK_ENVELOPE_INVALID")
+            online_allowed = {
+                "goal",
+                "instructions",
+                "success_criteria",
+                "constraints",
+                "required_capabilities",
+                "priority",
+                "max_attempts",
+            }
+            if set(online) != online_allowed:
+                raise AuthorizationError("ONLINE_TASK_FIELDS_INVALID")
+            if (
+                not isinstance(online["goal"], str)
+                or not online["goal"].strip()
+                or len(online["goal"]) > 12_000
+                or not isinstance(online["instructions"], str)
+                or len(online["instructions"]) > 24_000
+                or not isinstance(online["success_criteria"], list)
+                or len(online["success_criteria"]) > 32
+                or not isinstance(online["constraints"], Mapping)
+            ):
+                raise AuthorizationError("ONLINE_TASK_CONTENT_INVALID")
+            projected_caps = online["required_capabilities"]
+            if (
+                not isinstance(projected_caps, list)
+                or not projected_caps
+                or len(projected_caps) != len(set(projected_caps))
+                or not frozenset(projected_caps).issubset(
+                    SAFE_PROJECTED_TASK_CAPABILITIES
+                )
+            ):
+                raise AuthorizationError("ONLINE_TASK_CAPABILITIES_NOT_AUTHORIZED", 403)
+            if online["priority"] != priority or online["max_attempts"] != max_attempts:
+                raise AuthorizationError("ONLINE_TASK_LIMITS_MISMATCH")
+
+            execution_allowed = {
+                "action",
+                "adapter",
+                "capability_scope",
+                "source_signals",
+            }
+            if set(execution) != execution_allowed:
+                raise AuthorizationError("EXECUTION_FIELDS_INVALID")
+            if execution.get("action") != "source_capture" or execution.get(
+                "adapter"
+            ) != "source_capture":
+                raise AuthorizationError("INITIAL_EXECUTION_ACTION_NOT_AUTHORIZED", 403)
+            scope = execution.get("capability_scope")
+            if (
+                not isinstance(scope, list)
+                or sorted(scope) != sorted(projected_caps)
+            ):
+                raise AuthorizationError("EXECUTION_CAPABILITY_SCOPE_MISMATCH")
+            signals = execution.get("source_signals")
+            if (
+                not isinstance(signals, list)
+                or len(signals) > 32
+                or not all(isinstance(item, str) and len(item) <= 1_000 for item in signals)
+            ):
+                raise AuthorizationError("SOURCE_SIGNALS_INVALID")
+
+        elif capability == "release:candidate":
+            required = {"candidate_id", "source_commit", "version"}
+            if set(request) != required:
+                raise AuthorizationError("RELEASE_CANDIDATE_FIELDS_INVALID")
+            if not GRANT_RE.fullmatch(str(request["candidate_id"])):
+                raise AuthorizationError("RELEASE_CANDIDATE_ID_INVALID")
+            if OID_RE.fullmatch(str(request["source_commit"])) is None:
+                raise AuthorizationError("RELEASE_SOURCE_COMMIT_INVALID")
+            version = str(request["version"])
+            if VERSION_RE.fullmatch(version) is None or not version.startswith(
+                self.constraints.release_version_prefix
+            ):
+                raise AuthorizationError("RELEASE_VERSION_NOT_AUTHORIZED", 403)
+        else:
+            raise AuthorizationError("CAPABILITY_NOT_SUPPORTED", 403)
+
+        return {
+            "schema": DECISION_SCHEMA,
+            "decision": "ALLOW",
+            "grant_id": self.grant_id,
+            "owner_identity": self.owner_identity,
+            "account": self.account,
+            "capability": capability,
+            "policy_sha256": self.policy_sha256,
+            "request_sha256": sha256(canonical(request)),
+            "persistent_until_revoked": True,
+        }
+
+
+class ReceiptStore:
+    def __init__(self, path: Path) -> None:
+        self.path = path
+        self.path.parent.mkdir(parents=True, exist_ok=True)
+        with sqlite3.connect(self.path) as con:
+            con.execute(
+                """
+                CREATE TABLE IF NOT EXISTS receipts(
+                  idempotency_key TEXT PRIMARY KEY,
+                  capability TEXT NOT NULL,
+                  request_sha256 TEXT NOT NULL,
+                  response_sha256 TEXT NOT NULL,
+                  response_json TEXT NOT NULL,
+                  created_at TEXT NOT NULL
+                )
+                """
+            )
+            con.commit()
+        self._lock = threading.RLock()
+
+    def get(self, key: str) -> dict[str, Any] | None:
+        with sqlite3.connect(self.path) as con:
+            row = con.execute(
+                "SELECT request_sha256,response_json FROM receipts WHERE idempotency_key=?",
+                (key,),
+            ).fetchone()
+        if row is None:
+            return None
+        return {"request_sha256": row[0], "response": json.loads(row[1])}
+
+    def put(
+        self,
+        key: str,
+        capability: str,
+        request_sha256: str,
+        response: Mapping[str, Any],
+    ) -> None:
+        encoded = json.dumps(response, sort_keys=True, separators=(",", ":"))
+        with self._lock, sqlite3.connect(self.path) as con:
+            con.execute(
+                """
+                INSERT INTO receipts(
+                  idempotency_key,capability,request_sha256,response_sha256,
+                  response_json,created_at
+                ) VALUES(?,?,?,?,?,?)
+                """,
+                (
+                    key,
+                    capability,
+                    request_sha256,
+                    sha256(encoded.encode()),
+                    encoded,
+                    datetime.now(timezone.utc).isoformat(),
+                ),
+            )
+            con.commit()
+
+
+@dataclass(frozen=True)
+class Upstream:
+    bridge_url: str
+    bridge_token: str
+
+    def post_bridge(
+        self,
+        path: str,
+        body: Mapping[str, Any],
+    ) -> tuple[int, dict[str, Any]]:
+        req = urllib.request.Request(
+            self.bridge_url.rstrip("/") + path,
+            method="POST",
+            data=canonical(body),
+            headers={
+                "Authorization": "Bearer " + self.bridge_token,
+                "Content-Type": "application/json",
+                "Accept": "application/json",
+            },
+        )
+        try:
+            with urllib.request.urlopen(req, timeout=30) as response:
+                raw, status = response.read(), response.status
+        except urllib.error.HTTPError as exc:
+            raw, status = exc.read(), exc.code
+        except Exception as exc:
+            raise AuthorizationError("UPSTREAM_TRANSPORT_FAILED", 503) from exc
+        try:
+            value = json.loads(raw.decode())
+        except Exception as exc:
+            raise AuthorizationError("UPSTREAM_RESPONSE_INVALID", 502) from exc
+        if not isinstance(value, Mapping):
+            raise AuthorizationError("UPSTREAM_RESPONSE_INVALID", 502)
+        return status, dict(value)
+
+    def project_governed_task(
+        self,
+        body: Mapping[str, Any],
+    ) -> tuple[int, dict[str, Any]]:
+        return self.post_bridge("/v2/governed-task", body)
+
+
+class Broker:
+    def __init__(
+        self,
+        *,
+        policy_path: Path,
+        revocation_path: Path,
+        store: ReceiptStore,
+        upstream: Upstream,
+        expected_account: str,
+    ) -> None:
+        self.policy_path = policy_path
+        self.revocation_path = revocation_path
+        self.store = store
+        self.upstream = upstream
+        self.expected_account = expected_account
+
+    def policy(self) -> DurablePolicy:
+        policy = DurablePolicy.load(self.policy_path, self.revocation_path)
+        if policy.account != self.expected_account:
+            raise AuthorizationError("POLICY_ACCOUNT_MISMATCH", 403)
+        return policy
+
+    def execute(
+        self,
+        capability: str,
+        request: Mapping[str, Any],
+        *,
+        idempotency_key: str,
+    ) -> dict[str, Any]:
+        if not GRANT_RE.fullmatch(idempotency_key):
+            raise AuthorizationError("IDEMPOTENCY_KEY_INVALID")
+        policy = self.policy()
+        decision = policy.authorize(capability, request)
+        request_sha = decision["request_sha256"]
+
+        prior = self.store.get(idempotency_key)
+        if prior is not None:
+            if prior["request_sha256"] != request_sha:
+                raise AuthorizationError("IDEMPOTENCY_CONFLICT", 409)
+            return prior["response"]
+
+        if capability == "governed-task:project":
+            status, value = self.upstream.project_governed_task(request)
+            if status not in {200, 201, 202}:
+                raise AuthorizationError(
+                    "GOVERNED_TASK_PROJECTION_FAILED",
+                    409 if status == 409 else 503,
+                )
+            if value.get("task_id") != request["task_id"]:
+                raise AuthorizationError("GOVERNED_TASK_RESPONSE_INVALID", 502)
+            operation_state = "EXECUTED"
+            upstream_status: int | None = status
+            upstream_response: Mapping[str, Any] | None = value
+        elif capability == "release:candidate":
+            # The trusted release builder lives outside the repository control
+            # plane. This broker persists/attests Owner intent but does not
+            # invent or proxy a release-builder endpoint. The external
+            # connector must project the account's release:candidate scope and
+            # may consume this grant/decision as authorization evidence.
+            operation_state = "AUTHORIZED_PENDING_EXTERNAL_CAPABILITY_PROJECTION"
+            upstream_status = None
+            upstream_response = None
+        else:
+            raise AuthorizationError("CAPABILITY_NOT_SUPPORTED", 403)
+
+        response = {
+            "schema": RECEIPT_SCHEMA,
+            "grant_id": policy.grant_id,
+            "owner_identity": policy.owner_identity,
+            "account": policy.account,
+            "capability": capability,
+            "policy_sha256": policy.policy_sha256,
+            "request_sha256": request_sha,
+            "idempotency_key": idempotency_key,
+            "operation_state": operation_state,
+            "upstream_status": upstream_status,
+            "upstream_response": upstream_response,
+            "external_capability_projection_required": (
+                capability == "release:candidate"
+            ),
+            "separate_owner_release_approval_still_required": True,
+        }
+        self.store.put(idempotency_key, capability, request_sha, response)
+        return response
+
+
+def load_secret(path: Path, *, code: str) -> str:
+    if path.is_symlink() or not path.is_file():
+        raise RuntimeError(code + "_UNAVAILABLE")
+    value = path.read_text(encoding="utf-8").strip()
+    if len(value) < 32 or any(ch.isspace() for ch in value):
+        raise RuntimeError(code + "_INVALID")
+    return value
+
+
+class Handler(BaseHTTPRequestHandler):
+    broker: Broker | None = None
+
+    def log_message(self, *_args: Any) -> None:
+        return
+
+    def send_json(self, value: Mapping[str, Any], status: int) -> None:
+        raw = canonical(value)
+        self.send_response(status)
+        self.send_header("Content-Type", "application/json")
+        self.send_header("Cache-Control", "no-store")
+        self.send_header("Content-Length", str(len(raw)))
+        self.end_headers()
+        self.wfile.write(raw)
+
+    def caller_account(self) -> str:
+        return self.headers.get("X-AIOS-Account", "").strip()
+
+    def account_authorized(self) -> bool:
+        if self.broker is None:
+            return False
+        return self.caller_account() == self.broker.expected_account
+
+    def body(self) -> dict[str, Any]:
+        try:
+            length = int(self.headers.get("Content-Length", "-1"))
+        except ValueError as exc:
+            raise AuthorizationError("CONTENT_LENGTH_INVALID") from exc
+        if length < 1 or length > MAX_BODY:
+            raise AuthorizationError("REQUEST_SIZE_INVALID", 413)
+        try:
+            value = json.loads(self.rfile.read(length).decode())
+        except Exception as exc:
+            raise AuthorizationError("REQUEST_JSON_INVALID") from exc
+        if not isinstance(value, Mapping):
+            raise AuthorizationError("REQUEST_OBJECT_REQUIRED")
+        return dict(value)
+
+    def do_GET(self) -> None:
+        if self.path == "/health":
+            self.send_json(
+                {
+                    "status": "ok",
+                    "service": "aios-owner-durable-authorization-broker",
+                    "transport": "unix",
+                },
+                200,
+            )
+            return
+        if not self.account_authorized():
+            self.send_json({"error": "ACCOUNT_NOT_AUTHORIZED"}, 403)
+            return
+        if self.path != "/v1/capabilities" or self.broker is None:
+            self.send_json({"error": "NOT_FOUND"}, 404)
+            return
+        try:
+            policy = self.broker.policy()
+            self.send_json(
+                {
+                    "schema": POLICY_SCHEMA,
+                    "grant_id": policy.grant_id,
+                    "account": policy.account,
+                    "capabilities": sorted(policy.capabilities),
+                    "persistent_until_revoked": True,
+                    "separate_owner_release_approval_still_required": True,
+                },
+                200,
+            )
+        except AuthorizationError as exc:
+            self.send_json({"error": exc.code}, exc.status)
+
+    def do_POST(self) -> None:
+        if not self.account_authorized():
+            self.send_json({"error": "ACCOUNT_NOT_AUTHORIZED"}, 403)
+            return
+        routes = {
+            "/v1/project-governed-task": "governed-task:project",
+            "/v1/authorize-release-candidate": "release:candidate",
+        }
+        capability = routes.get(self.path)
+        if capability is None or self.broker is None:
+            self.send_json({"error": "NOT_FOUND"}, 404)
+            return
+        try:
+            body = self.body()
+            if set(body) != {"idempotency_key", "request"}:
+                raise AuthorizationError("BROKER_FIELDS_INVALID")
+            if not isinstance(body["request"], Mapping):
+                raise AuthorizationError("BROKER_REQUEST_INVALID")
+            value = self.broker.execute(
+                capability,
+                dict(body["request"]),
+                idempotency_key=str(body["idempotency_key"]),
+            )
+            self.send_json(value, 200)
+        except AuthorizationError as exc:
+            self.send_json({"error": exc.code}, exc.status)
+
+
+class UnixHTTPServer(socketserver.ThreadingUnixStreamServer):
+    daemon_threads = True
+
+def serve() -> None:
+    credentials = Path(os.environ.get("CREDENTIALS_DIRECTORY", "/run/credentials"))
+    policy_path = Path(
+        os.environ.get(
+            "AIOS_OWNER_DURABLE_AUTHORIZATION_POLICY",
+            "/etc/aios-gitops/owner-durable-authorization.json",
+        )
+    )
+    revocation_path = Path(
+        os.environ.get(
+            "AIOS_OWNER_DURABLE_AUTHORIZATION_REVOCATIONS",
+            "/etc/aios-gitops/owner-durable-authorization-revocations.json",
+        )
+    )
+    bridge_token = load_secret(
+        Path(
+            os.environ.get(
+                "AIOS_AGENT_TASK_CONTROL_TOKEN_FILE",
+                str(credentials / "agent_task_control_token"),
+            )
+        ),
+        code="BRIDGE_CONTROL_CREDENTIAL",
+    )
+    account = os.environ.get(
+        "AIOS_OWNER_AUTH_BROKER_ACCOUNT",
+        "chatgpt-governed-implementer-singlehost",
+    ).strip()
+    state_root = Path(
+        os.environ.get(
+            "AIOS_OWNER_AUTH_BROKER_STATE_ROOT",
+            "/var/lib/aios-owner-authorization-broker",
+        )
+    )
+    Handler.broker = Broker(
+        policy_path=policy_path,
+        revocation_path=revocation_path,
+        store=ReceiptStore(state_root / "receipts.sqlite3"),
+        upstream=Upstream(
+            bridge_url="http://127.0.0.1:18892",
+            bridge_token=bridge_token,
+        ),
+        expected_account=account,
+    )
+    socket_path = Path(
+        os.environ.get(
+            "AIOS_OWNER_AUTH_BROKER_SOCKET",
+            "/run/aios-owner-authorization-broker/broker.sock",
+        )
+    )
+    if not socket_path.is_absolute() or socket_path.is_symlink():
+        raise RuntimeError("BROKER_SOCKET_PATH_INVALID")
+    socket_path.parent.mkdir(parents=True, exist_ok=True)
+    if socket_path.exists():
+        if not stat.S_ISSOCK(socket_path.stat().st_mode):
+            raise RuntimeError("BROKER_SOCKET_CONFLICT")
+        socket_path.unlink()
+    server = UnixHTTPServer(str(socket_path), Handler)
+    os.chmod(socket_path, 0o660)
+    try:
+        server.serve_forever()
+    finally:
+        server.server_close()
+        socket_path.unlink(missing_ok=True)
+
+
+if __name__ == "__main__":
+    serve()

active-runtime/app/aios_external_agent_runtime.py

@@ -16,6 +16,11 @@ import subprocess
 import tempfile
 import time
 from typing import Any, Mapping, Sequence
+from aios_model_governance import (
+    ModelGovernanceError,
+    ModelGovernanceStore,
+    RouteRequest,
+)
 
 REQUEST_SCHEMA = "aios.external-agent-runtime.request.v1"
 RESULT_SCHEMA = "aios.external-agent-runtime.result.v1"
@@ -62,6 +67,15 @@ MAX_PROFILE_BYTES = 64 * 1024
 MAX_CREDENTIAL_BYTES = 64 * 1024
 DEFAULT_PROFILE_ROOT = Path("/etc/aios/external-agent-runtime.d")
 DEFAULT_EXCHANGE_ROOT = Path("/var/lib/aios-external-agent-executor")
+DEFAULT_MODEL_GOVERNANCE_DB = Path(
+    os.environ.get(
+        "AIOS_MODEL_GOVERNANCE_DB",
+        "/var/lib/aios-model-governance/state.sqlite3",
+    )
+)
+MODEL_GOVERNANCE_ENABLED = os.environ.get(
+    "AIOS_MODEL_GOVERNANCE_ENABLED", "0"
+).strip().lower() in {"1", "true", "yes", "on"}
 ALLOWED_EXECUTABLE_ROOTS = (
     Path("/usr/bin"),
     Path("/usr/local/bin"),
@@ -208,6 +222,172 @@ def _profile_credential_environment(
     return tuple(sorted(result))
 
 
+def _profile_model_governance(
+    value: object,
+) -> tuple[bool, tuple[str, ...], tuple[str, ...], int, int]:
+    if value is None:
+        return False, (), (), 0, 0
+    if not isinstance(value, Mapping):
+        raise ExternalAgentRuntimeError("EXTERNAL_AGENT_MODEL_GOVERNANCE_PROFILE_INVALID")
+    allowed = {
+        "required",
+        "provider_ids",
+        "model_capabilities",
+        "max_output_tokens",
+        "input_reservation_floor_tokens",
+    }
+    if set(value) - allowed:
+        raise ExternalAgentRuntimeError("EXTERNAL_AGENT_MODEL_GOVERNANCE_PROFILE_FIELDS_INVALID")
+    required = bool(value.get("required", False))
+    providers = value.get("provider_ids") or []
+    capabilities = value.get("model_capabilities") or []
+    if (
+        not isinstance(providers, Sequence)
+        or isinstance(providers, (str, bytes))
+        or not all(isinstance(item, str) and item.strip() for item in providers)
+    ):
+        raise ExternalAgentRuntimeError("EXTERNAL_AGENT_MODEL_PROVIDER_IDS_INVALID")
+    if (
+        not isinstance(capabilities, Sequence)
+        or isinstance(capabilities, (str, bytes))
+        or not all(isinstance(item, str) and item.strip() for item in capabilities)
+    ):
+        raise ExternalAgentRuntimeError("EXTERNAL_AGENT_MODEL_CAPABILITIES_INVALID")
+    max_output_tokens = int(value.get("max_output_tokens") or 0)
+    input_reservation_floor_tokens = int(
+        value.get("input_reservation_floor_tokens") or 0
+    )
+    if max_output_tokens < 0 or max_output_tokens > 262144:
+        raise ExternalAgentRuntimeError("EXTERNAL_AGENT_MODEL_OUTPUT_LIMIT_INVALID")
+    if (
+        input_reservation_floor_tokens < 0
+        or input_reservation_floor_tokens > 262144
+    ):
+        raise ExternalAgentRuntimeError(
+            "EXTERNAL_AGENT_MODEL_INPUT_RESERVATION_FLOOR_INVALID"
+        )
+    if required and not providers:
+        raise ExternalAgentRuntimeError("EXTERNAL_AGENT_MODEL_PROVIDER_IDS_REQUIRED")
+    if required and max_output_tokens < 1:
+        raise ExternalAgentRuntimeError("EXTERNAL_AGENT_MODEL_OUTPUT_LIMIT_REQUIRED")
+    return (
+        required,
+        tuple(sorted({item.strip() for item in providers})),
+        tuple(sorted({item.strip() for item in capabilities})),
+        max_output_tokens,
+        input_reservation_floor_tokens,
+    )
+
+
+def _estimate_tokens(text: str) -> int:
+    # Conservative byte-based estimate; runtime-reported usage supersedes this at settlement.
+    data = text.encode("utf-8", "replace")
+    return max(1, (len(data) + 3) // 4)
+
+
+def _extract_runtime_usage(
+    output_json: object,
+) -> tuple[int | None, int | None, int | None]:
+    if not isinstance(output_json, Mapping):
+        return None, None, None
+
+    candidates: list[Mapping[str, Any]] = []
+    direct = output_json.get("usage")
+    if isinstance(direct, Mapping):
+        candidates.append(direct)
+
+    meta = output_json.get("meta")
+    if isinstance(meta, Mapping):
+        agent_meta = meta.get("agentMeta")
+        if isinstance(agent_meta, Mapping):
+            nested = agent_meta.get("usage")
+            if isinstance(nested, Mapping):
+                candidates.append(nested)
+            last_call = agent_meta.get("lastCallUsage")
+            if isinstance(last_call, Mapping):
+                candidates.append(last_call)
+
+    def number(mapping: Mapping[str, Any], *names: str) -> int | None:
+        for name in names:
+            value = mapping.get(name)
+            if isinstance(value, int) and not isinstance(value, bool) and value >= 0:
+                return value
+        return None
+
+    input_tokens = None
+    output_tokens = None
+    cost_micros = None
+    for usage in candidates:
+        if input_tokens is None:
+            input_tokens = number(usage, "input_tokens", "prompt_tokens", "input")
+        if output_tokens is None:
+            output_tokens = number(
+                usage, "output_tokens", "completion_tokens", "output"
+            )
+        if cost_micros is None:
+            cost_micros = number(usage, "cost_micros", "total_cost_micros")
+
+    if cost_micros is None and isinstance(meta, Mapping):
+        agent_meta = meta.get("agentMeta")
+        if isinstance(agent_meta, Mapping):
+            cost_usd = agent_meta.get("costUsd")
+            if (
+                isinstance(cost_usd, (int, float))
+                and not isinstance(cost_usd, bool)
+                and cost_usd >= 0
+            ):
+                cost_micros = int(round(float(cost_usd) * 1_000_000))
+
+    return input_tokens, output_tokens, cost_micros
+
+
+def _scope_from_payload(payload: Mapping[str, Any]) -> tuple[str, str | None, str | None, str]:
+    scope = payload.get("scope")
+    if not isinstance(scope, Mapping):
+        raise ExternalAgentRuntimeError("EXTERNAL_AGENT_MODEL_SCOPE_MISSING")
+    scope_type = str(scope.get("scope_type") or "legacy")
+    tenant_id = scope.get("tenant_id")
+    project_id = scope.get("project_id")
+    trace_id = str(payload.get("trace_id") or "").strip()
+    if not trace_id:
+        raise ExternalAgentRuntimeError("EXTERNAL_AGENT_MODEL_TRACE_ID_MISSING")
+    return (
+        scope_type,
+        str(tenant_id) if tenant_id not in (None, "") else None,
+        str(project_id) if project_id not in (None, "") else None,
+        trace_id,
+    )
+
+
+def _model_replacements(decision: object | None) -> dict[str, str]:
+    if decision is None:
+        return {}
+    return {
+        "{provider_id}": str(getattr(decision, "provider_id")),
+        "{model_id}": str(getattr(decision, "model_id")),
+        "{model_version}": str(getattr(decision, "model_version")),
+    }
+
+
+def _bind_model_values(
+    values: Sequence[tuple[str, str]],
+    decision: object | None,
+) -> tuple[tuple[str, str], ...]:
+    replacements = _model_replacements(decision)
+    result: list[tuple[str, str]] = []
+    for key, value in values:
+        bound = value
+        for marker, replacement in replacements.items():
+            bound = bound.replace(marker, replacement)
+        result.append((key, bound))
+    return tuple(result)
+
+
+def _runtime_model_binding_present(profile: "RuntimeProfile") -> bool:
+    markers = ("{model_id}", "{model_version}", "{provider_id}")
+    return any(marker in item for item in (*profile.argv, *(value for _, value in profile.environment)) for marker in markers)
+
+
 @dataclass(frozen=True)
 class RuntimeProfile:
     runtime: str
@@ -218,6 +398,11 @@ class RuntimeProfile:
     max_output_bytes: int
     environment: tuple[tuple[str, str], ...]
     credential_env: tuple[tuple[str, str, str, str], ...]
+    model_governance_required: bool
+    model_provider_ids: tuple[str, ...]
+    model_capabilities: tuple[str, ...]
+    model_max_output_tokens: int
+    model_input_reservation_floor_tokens: int
 
     @classmethod
     def load(
@@ -275,7 +460,14 @@ class RuntimeProfile:
             raise ExternalAgentRuntimeError("EXTERNAL_AGENT_OUTPUT_LIMIT_INVALID")
         environment = _profile_environment(value.get("environment"))
         credential_env = _profile_credential_environment(value.get("credential_env"))
-        return cls(
+        (
+            model_governance_required,
+            model_provider_ids,
+            model_capabilities,
+            model_max_output_tokens,
+            model_input_reservation_floor_tokens,
+        ) = _profile_model_governance(value.get("model_governance"))
+        profile = cls(
             runtime,
             tuple(normalized),
             input_mode,
@@ -284,7 +476,15 @@ class RuntimeProfile:
             max_output,
             environment,
             credential_env,
+            model_governance_required,
+            model_provider_ids,
+            model_capabilities,
+            model_max_output_tokens,
+            model_input_reservation_floor_tokens,
         )
+        if profile.model_governance_required and not _runtime_model_binding_present(profile):
+            raise ExternalAgentRuntimeError("EXTERNAL_AGENT_MODEL_BINDING_REQUIRED")
+        return profile
 
 
 def validate_request(value: Mapping[str, Any]) -> dict[str, Any]:
@@ -295,9 +495,16 @@ def validate_request(value: Mapping[str, Any]) -> dict[str, Any]:
         raise ExternalAgentRuntimeError("EXTERNAL_AGENT_REQUEST_ID_INVALID")
     runtime = _safe_runtime(value.get("runtime"))
     task_id = str(value.get("task_id") or "")
+    attempt = value.get("attempt")
     execution_digest = str(value.get("execution_digest") or "")
     if not task_id or not re.fullmatch(r"[A-Za-z0-9][A-Za-z0-9._:-]{2,179}", task_id):
         raise ExternalAgentRuntimeError("EXTERNAL_AGENT_TASK_ID_INVALID")
+    if (
+        not isinstance(attempt, int)
+        or isinstance(attempt, bool)
+        or not 1 <= attempt <= 1000
+    ):
+        raise ExternalAgentRuntimeError("EXTERNAL_AGENT_ATTEMPT_INVALID")
     if re.fullmatch(r"[0-9a-f]{64}", execution_digest) is None:
         raise ExternalAgentRuntimeError("EXTERNAL_AGENT_EXECUTION_DIGEST_INVALID")
     payload = value.get("payload")
@@ -317,6 +524,7 @@ def validate_request(value: Mapping[str, Any]) -> dict[str, Any]:
         "request_id": request_id,
         "runtime": runtime,
         "task_id": task_id,
+        "attempt": attempt,
         "execution_digest": execution_digest,
         "payload": dict(payload),
     }
@@ -351,11 +559,13 @@ def _argv_with_files(
     input_path: Path,
     output_path: Path,
     prompt: str,
+    model_decision: object | None = None,
 ) -> list[str]:
     replacements = {
         "{input_json}": str(input_path),
         "{output_json}": str(output_path),
         "{prompt}": prompt,
+        **_model_replacements(model_decision),
     }
     result = []
     for item in profile.argv:
@@ -471,7 +681,12 @@ def execute_request(
         request = validate_request(json.loads(raw_request.decode("utf-8")))
     except json.JSONDecodeError as exc:
         raise ExternalAgentRuntimeError("EXTERNAL_AGENT_REQUEST_JSON_INVALID") from exc
+
     profile = RuntimeProfile.load(request["runtime"], profile_root)
+    governance_store: ModelGovernanceStore | None = None
+    route_request: RouteRequest | None = None
+    decision = None
+    governance_attempts: list[dict[str, Any]] = []
 
     with tempfile.TemporaryDirectory(prefix=f"aios-{request['runtime']}-") as temporary:
         temp_root = Path(temporary)
@@ -480,78 +695,203 @@ def execute_request(
         output_path = temp_root / "output.json"
         input_path.write_bytes(canonical(request["payload"]))
         prompt = render_prompt(request["payload"])
-        argv = _argv_with_files(profile, input_path, output_path, prompt)
 
-        stdin_data = None
-        if profile.input_mode == "stdin_json":
-            stdin_data = canonical(request["payload"]).decode("utf-8")
-        elif profile.input_mode == "stdin_text":
-            stdin_data = prompt
-
-        child_env = {
-            "PATH": "/usr/bin:/usr/local/bin:/bin",
-            "LC_ALL": "C.UTF-8",
-            "LANG": "C.UTF-8",
-            "HOME": str(temp_root / "home"),
-            "NO_COLOR": "1",
-        }
-        child_env.update(dict(profile.environment))
-        child_env.update(_credential_environment(profile))
-
-        started = time.monotonic()
-        try:
-            completed = subprocess.run(
-                argv,
-                input=stdin_data,
-                capture_output=True,
-                text=True,
-                check=False,
-                shell=False,
-                timeout=profile.timeout_seconds,
-                cwd=str(temp_root),
-                env=child_env,
+        if profile.model_governance_required:
+            if not MODEL_GOVERNANCE_ENABLED:
+                raise ExternalAgentRuntimeError("EXTERNAL_AGENT_MODEL_GOVERNANCE_DISABLED")
+            scope_type, tenant_id, project_id, trace_id = _scope_from_payload(request["payload"])
+            governance_store = ModelGovernanceStore(DEFAULT_MODEL_GOVERNANCE_DB)
+            prompt_estimated_input_tokens = _estimate_tokens(prompt)
+            reserved_estimated_input_tokens = max(
+                prompt_estimated_input_tokens,
+                profile.model_input_reservation_floor_tokens,
             )
-        except subprocess.TimeoutExpired as exc:
-            raise ExternalAgentRuntimeError("EXTERNAL_AGENT_TIMEOUT") from exc
-
-        if completed.returncode != 0:
-            raise ExternalAgentRuntimeError(
-                f"EXTERNAL_AGENT_PROCESS_FAILED:{completed.returncode}:"
-                + _bounded_text(completed.stderr, 2048)
+            route_request = RouteRequest(
+                request_id=request["request_id"] + ":model",
+                task_id=request["task_id"],
+                trace_id=trace_id,
+                scope_type=scope_type,
+                tenant_id=tenant_id,
+                project_id=project_id,
+                required_capabilities=tuple(
+                    sorted(
+                        {
+                            f"agent.runtime.{request['runtime']}",
+                            *profile.model_capabilities,
+                        }
+                    )
+                ),
+                estimated_input_tokens=reserved_estimated_input_tokens,
+                max_output_tokens=profile.model_max_output_tokens,
+            )
+            try:
+                decision = governance_store.route(route_request)
+            except ModelGovernanceError as exc:
+                raise ExternalAgentRuntimeError(exc.code) from exc
+            if profile.model_provider_ids and decision.provider_id not in profile.model_provider_ids:
+                raise ExternalAgentRuntimeError("EXTERNAL_AGENT_MODEL_PROVIDER_PROFILE_DENY")
+            governance_attempts.append(
+                {
+                    "request_id": decision.request_id,
+                    "provider_id": decision.provider_id,
+                    "model_id": decision.model_id,
+                    "model_version": decision.model_version,
+                    "fallback_index": decision.fallback_index,
+                }
             )
 
-        output_text, output_json = _parse_output(profile, completed.stdout, output_path)
-        result: dict[str, Any] = {
-            "schema": RESULT_SCHEMA,
-            "request_id": request["request_id"],
-            "request_sha256": sha256(canonical(request)),
-            "runtime": request["runtime"],
-            "task_id": request["task_id"],
-            "execution_digest": request["execution_digest"],
-            "exit_code": completed.returncode,
-            "duration_ms": int((time.monotonic() - started) * 1000),
-            "output_text": output_text,
-            "output_sha256": sha256(output_text.encode("utf-8", "replace")),
-            "stderr_sha256": sha256(completed.stderr.encode("utf-8", "replace")),
-            "credential_material_exposed": False,
-            "control_plane_access_granted": False,
-        }
-        if output_json is not None:
-            result["output_json"] = output_json
-        result["result_sha256"] = sha256(canonical(result))
-        result_path.parent.mkdir(parents=True, exist_ok=True)
-        fd, name = tempfile.mkstemp(prefix=f".{result_path.name}.", dir=result_path.parent)
         try:
-            os.fchmod(fd, 0o640)
-            with os.fdopen(fd, "w", encoding="utf-8") as handle:
-                json.dump(result, handle, ensure_ascii=False, sort_keys=True, separators=(",", ":"))
-                handle.write("\n")
-                handle.flush()
-                os.fsync(handle.fileno())
-            os.replace(name, result_path)
+            while True:
+                argv = _argv_with_files(
+                    profile,
+                    input_path,
+                    output_path,
+                    prompt,
+                    decision,
+                )
+                stdin_data = None
+                if profile.input_mode == "stdin_json":
+                    stdin_data = canonical(request["payload"]).decode("utf-8")
+                elif profile.input_mode == "stdin_text":
+                    stdin_data = prompt
+
+                child_env = {
+                    "PATH": "/usr/bin:/usr/local/bin:/bin",
+                    "LC_ALL": "C.UTF-8",
+                    "LANG": "C.UTF-8",
+                    "HOME": str(temp_root / "home"),
+                    "NO_COLOR": "1",
+                }
+                child_env.update(dict(_bind_model_values(profile.environment, decision)))
+                child_env.update(_credential_environment(profile))
+
+                started = time.monotonic()
+                failure_class = None
+                try:
+                    completed = subprocess.run(
+                        argv,
+                        input=stdin_data,
+                        capture_output=True,
+                        text=True,
+                        check=False,
+                        shell=False,
+                        timeout=profile.timeout_seconds,
+                        cwd=str(temp_root),
+                        env=child_env,
+                    )
+                    if completed.returncode != 0:
+                        failure_class = f"process_exit_{completed.returncode}"
+                except subprocess.TimeoutExpired:
+                    completed = None
+                    failure_class = "timeout"
+
+                if failure_class is None and completed is not None:
+                    break
+
+                if governance_store is None or decision is None or route_request is None:
+                    if failure_class == "timeout":
+                        raise ExternalAgentRuntimeError("EXTERNAL_AGENT_TIMEOUT")
+                    raise ExternalAgentRuntimeError(
+                        f"EXTERNAL_AGENT_PROCESS_FAILED:{completed.returncode}:"
+                        + _bounded_text(completed.stderr, 2048)
+                    )
+
+                try:
+                    governance_store.record_failure(decision.request_id, failure_class)
+                    decision = governance_store.fallback(route_request, decision.request_id)
+                except ModelGovernanceError as exc:
+                    raise ExternalAgentRuntimeError(exc.code) from exc
+                if profile.model_provider_ids and decision.provider_id not in profile.model_provider_ids:
+                    raise ExternalAgentRuntimeError("EXTERNAL_AGENT_MODEL_PROVIDER_PROFILE_DENY")
+                governance_attempts.append(
+                    {
+                        "request_id": decision.request_id,
+                        "provider_id": decision.provider_id,
+                        "model_id": decision.model_id,
+                        "model_version": decision.model_version,
+                        "fallback_index": decision.fallback_index,
+                        "previous_failure_class": failure_class,
+                    }
+                )
+
+            output_text, output_json = _parse_output(profile, completed.stdout, output_path)
+            model_governance: dict[str, Any] | None = None
+            if governance_store is not None and decision is not None:
+                input_tokens, output_tokens, actual_cost = _extract_runtime_usage(output_json)
+                try:
+                    settlement = governance_store.record_success(
+                        decision.request_id,
+                        input_tokens=input_tokens,
+                        output_tokens=output_tokens,
+                        actual_cost_micros=actual_cost,
+                        usage_source="runtime",
+                    )
+                except ModelGovernanceError as exc:
+                    raise ExternalAgentRuntimeError(exc.code) from exc
+                model_governance = {
+                    "decision_id": decision.decision_id,
+                    "request_id": decision.request_id,
+                    "provider_id": decision.provider_id,
+                    "model_id": decision.model_id,
+                    "model_version": decision.model_version,
+                    "policy_revision": decision.policy_revision,
+                    "model_revision": decision.model_revision,
+                    "fallback_index": decision.fallback_index,
+                    "prompt_estimated_input_tokens": prompt_estimated_input_tokens,
+                    "input_reservation_floor_tokens": (
+                        profile.model_input_reservation_floor_tokens
+                    ),
+                    "reserved_estimated_input_tokens": (
+                        reserved_estimated_input_tokens
+                    ),
+                    "settlement": settlement,
+                    "attempts": governance_attempts,
+                    "audit_chain_valid": governance_store.verify_audit_chain(),
+                }
+
+            result: dict[str, Any] = {
+                "schema": RESULT_SCHEMA,
+                "request_id": request["request_id"],
+                "request_sha256": sha256(canonical(request)),
+                "runtime": request["runtime"],
+                "task_id": request["task_id"],
+                "attempt": request["attempt"],
+                "execution_digest": request["execution_digest"],
+                "exit_code": completed.returncode,
+                "duration_ms": int((time.monotonic() - started) * 1000),
+                "output_text": output_text,
+                "output_sha256": sha256(output_text.encode("utf-8", "replace")),
+                "stderr_sha256": sha256(completed.stderr.encode("utf-8", "replace")),
+                "credential_material_exposed": False,
+                "control_plane_access_granted": False,
+            }
+            if output_json is not None:
+                result["output_json"] = output_json
+            if model_governance is not None:
+                result["model_governance"] = model_governance
+            result["result_sha256"] = sha256(canonical(result))
+            result_path.parent.mkdir(parents=True, exist_ok=True)
+            fd, name = tempfile.mkstemp(prefix=f".{result_path.name}.", dir=result_path.parent)
+            try:
+                os.fchmod(fd, 0o640)
+                with os.fdopen(fd, "w", encoding="utf-8") as handle:
+                    json.dump(
+                        result,
+                        handle,
+                        ensure_ascii=False,
+                        sort_keys=True,
+                        separators=(",", ":"),
+                    )
+                    handle.write("\n")
+                    handle.flush()
+                    os.fsync(handle.fileno())
+                os.replace(name, result_path)
+            finally:
+                if os.path.exists(name):
+                    os.unlink(name)
         finally:
-            if os.path.exists(name):
-                os.unlink(name)
+            if governance_store is not None:
+                governance_store.close()
 
 
 class ExternalAgentExchange:
@@ -566,6 +906,7 @@ class ExternalAgentExchange:
         request_id: str,
         runtime: str,
         task_id: str,
+        attempt: int,
         execution_digest: str,
         payload: Mapping[str, Any],
         timeout_seconds: int = 1800,
@@ -576,6 +917,7 @@ class ExternalAgentExchange:
                 "request_id": request_id,
                 "runtime": runtime,
                 "task_id": task_id,
+                "attempt": attempt,
                 "execution_digest": execution_digest,
                 "payload": dict(payload),
             }
@@ -651,6 +993,7 @@ class ExternalAgentExchange:
             or value.get("request_sha256") != request_sha
             or value.get("runtime") != request["runtime"]
             or value.get("task_id") != request["task_id"]
+            or value.get("attempt") != request["attempt"]
             or value.get("execution_digest") != request["execution_digest"]
         ):
             raise ExternalAgentRuntimeError("EXTERNAL_AGENT_RESULT_BINDING_INVALID")

active-runtime/app/aios_gpt_https_edge.py

@@ -0,0 +1,456 @@
+#!/usr/bin/env python3
+"""Disabled-by-default TLS edge for the read-only GPT Online Task ingress.
+
+The internal AIOS ingress remains loopback-only. This service is a separate edge
+that may bind non-loopback only when the owner explicitly approves it. It
+forwards only the tiny read-only ingress API and never proxies the canonical
+Bridge or Task API.
+"""
+
+from __future__ import annotations
+
+from collections import defaultdict, deque
+from dataclasses import dataclass
+import hashlib
+import hmac
+from http.server import BaseHTTPRequestHandler, ThreadingHTTPServer
+import json
+import os
+from pathlib import Path
+import re
+import secrets
+import ssl
+import sys
+import threading
+import time
+import urllib.error
+import urllib.request
+from typing import Any, Mapping
+from urllib.parse import urlsplit
+
+UPSTREAM = "http://127.0.0.1:18945"
+MAX_BODY = 64 * 1024
+MAX_RESPONSE = 256 * 1024
+MAX_CONCURRENT_REQUESTS = 32
+REQUEST_TIMEOUT_SECONDS = 20
+RATE_WINDOW_SECONDS = 60
+RATE_LIMIT_REQUESTS = 60
+RATE_BURST_REQUESTS = 15
+RATE_MAX_CLIENTS = 4096
+_AUDIT_SALT = secrets.token_bytes(32)
+PUBLIC_APPROVAL_ENV = "AIOS_GPT_EDGE_PUBLIC_BIND_APPROVED"
+PUBLIC_DOMAIN_ENV = "AIOS_GPT_EDGE_PUBLIC_DOMAIN"
+TASK_PATH_RE = re.compile(r"^/v1/read-only-tasks(?:/[A-Za-z0-9][A-Za-z0-9._:-]{7,127})?$")
+
+
+def _client_pseudonym(value: str) -> str:
+    return hashlib.sha256(_AUDIT_SALT + value.encode("utf-8", "replace")).hexdigest()[:16]
+
+
+def audit(event: str, **fields: object) -> None:
+    safe = {
+        "event": event,
+        "ts_unix_ms": int(time.time() * 1000),
+    }
+    for key, value in fields.items():
+        if key in {"authorization", "body", "token", "cookie", "client_ip"}:
+            continue
+        safe[key] = value
+    print(
+        json.dumps(safe, sort_keys=True, separators=(",", ":")),
+        file=sys.stderr,
+        flush=True,
+    )
+
+
+class EdgeError(RuntimeError):
+    def __init__(self, code: str, status: int = 400) -> None:
+        super().__init__(code)
+        self.code = code
+        self.status = status
+
+
+def canonical(value: object) -> bytes:
+    return json.dumps(
+        value,
+        ensure_ascii=False,
+        sort_keys=True,
+        separators=(",", ":"),
+    ).encode("utf-8")
+
+
+def public_bind_allowed(host: str, approval: str) -> str:
+    value = host.strip()
+    if value in {"127.0.0.1", "::1", "localhost"}:
+        return value
+    if approval != "1":
+        raise RuntimeError("GPT_EDGE_PUBLIC_BIND_OWNER_APPROVAL_REQUIRED")
+    if value not in {"0.0.0.0", "::"}:
+        raise RuntimeError("GPT_EDGE_BIND_ADDRESS_INVALID")
+    return value
+
+
+def validated_public_domain(value: str) -> str:
+    domain = value.strip().lower()
+    if not domain:
+        return ""
+    if len(domain) > 253 or re.fullmatch(
+        r"(?=.{1,253}\Z)(?:[a-z0-9](?:[a-z0-9-]{0,61}[a-z0-9])?\.)+[a-z0-9](?:[a-z0-9-]{0,61}[a-z0-9])?",
+        domain,
+    ) is None:
+        raise RuntimeError("GPT_EDGE_PUBLIC_DOMAIN_INVALID")
+    return domain
+
+
+def validated_edge_identity(host: str, approval: str, domain: str) -> tuple[str, str]:
+    bind_host = public_bind_allowed(host, approval)
+    public_domain = validated_public_domain(domain)
+    if bind_host not in {"127.0.0.1", "::1", "localhost"} and not public_domain:
+        raise RuntimeError("GPT_EDGE_PUBLIC_DOMAIN_REQUIRED")
+    return bind_host, public_domain
+
+
+def validate_tls_file(path: Path, *, key: bool) -> Path:
+    if path.is_symlink() or not path.is_file():
+        raise RuntimeError("GPT_EDGE_TLS_FILE_INVALID")
+    size = path.stat().st_size
+    if size < 16 or size > 256 * 1024:
+        raise RuntimeError("GPT_EDGE_TLS_FILE_INVALID")
+    if key and (path.stat().st_mode & 0o077):
+        raise RuntimeError("GPT_EDGE_TLS_KEY_MODE_INVALID")
+    return path
+
+
+def build_tls_context(cert_path: Path, key_path: Path) -> ssl.SSLContext:
+    cert = validate_tls_file(cert_path, key=False)
+    key = validate_tls_file(key_path, key=True)
+    context = ssl.SSLContext(ssl.PROTOCOL_TLS_SERVER)
+    context.minimum_version = ssl.TLSVersion.TLSv1_2
+    context.options |= ssl.OP_NO_COMPRESSION
+    context.load_cert_chain(certfile=str(cert), keyfile=str(key))
+    return context
+
+
+def allowed_route(method: str, raw_path: str) -> str:
+    parsed = urlsplit(raw_path)
+    if parsed.query or parsed.fragment:
+        raise EdgeError("QUERY_OR_FRAGMENT_DENY", 400)
+    path = parsed.path
+    if method == "GET" and path in {"/health", "/openapi.json"}:
+        return path
+    if method == "POST" and path == "/v1/read-only-tasks":
+        return path
+    if method == "GET" and TASK_PATH_RE.fullmatch(path):
+        return path
+    raise EdgeError("ROUTE_DENY", 404)
+
+
+def route_label(method: str, path: str) -> str:
+    if method == "GET" and path == "/health":
+        return "health"
+    if method == "GET" and path == "/openapi.json":
+        return "openapi"
+    if method == "POST" and path == "/v1/read-only-tasks":
+        return "task_create"
+    if method == "GET" and path.startswith("/v1/read-only-tasks/"):
+        return "task_status"
+    return "unknown"
+
+
+def forwarded_headers(headers: Mapping[str, str], *, body: bool) -> dict[str, str]:
+    result = {"Accept": "application/json"}
+    authorization = headers.get("Authorization")
+    if not isinstance(authorization, str) or not authorization.startswith("Bearer "):
+        raise EdgeError("AUTHORIZATION_REQUIRED", 401)
+    token = authorization[7:]
+    if len(token) < 32 or len(token) > 4096 or any(ch.isspace() for ch in token):
+        raise EdgeError("AUTHORIZATION_INVALID", 401)
+    result["Authorization"] = authorization
+    if body:
+        content_type = headers.get("Content-Type", "")
+        if content_type.split(";", 1)[0].strip().lower() != "application/json":
+            raise EdgeError("CONTENT_TYPE_INVALID", 415)
+        result["Content-Type"] = "application/json"
+    return result
+
+
+class RateLimiter:
+    def __init__(
+        self,
+        *,
+        window_seconds: int = RATE_WINDOW_SECONDS,
+        limit: int = RATE_LIMIT_REQUESTS,
+        burst: int = RATE_BURST_REQUESTS,
+        max_clients: int = RATE_MAX_CLIENTS,
+        now_fn=time.monotonic,
+    ) -> None:
+        self.window_seconds = window_seconds
+        self.limit = limit
+        self.burst = burst
+        self.max_clients = max_clients
+        self.now_fn = now_fn
+        self._lock = threading.Lock()
+        self._events: dict[str, deque[float]] = defaultdict(deque)
+
+    def _prune_stale(self, cutoff: float) -> None:
+        stale = [
+            key
+            for key, events in self._events.items()
+            if not events or events[-1] <= cutoff
+        ]
+        for key in stale:
+            self._events.pop(key, None)
+
+    def allowed(self, key: str) -> bool:
+        now = self.now_fn()
+        cutoff = now - self.window_seconds
+        with self._lock:
+            if key not in self._events and len(self._events) >= self.max_clients:
+                self._prune_stale(cutoff)
+                if len(self._events) >= self.max_clients:
+                    return False
+            events = self._events[key]
+            while events and events[0] <= cutoff:
+                events.popleft()
+            if len(events) >= self.limit + self.burst:
+                return False
+            events.append(now)
+            return True
+
+
+@dataclass(frozen=True)
+class Proxy:
+    upstream: str = UPSTREAM
+    timeout: int = REQUEST_TIMEOUT_SECONDS
+
+    def request(
+        self,
+        method: str,
+        path: str,
+        headers: Mapping[str, str],
+        body: bytes | None,
+    ) -> tuple[int, bytes, str]:
+        if self.upstream != UPSTREAM:
+            raise EdgeError("UPSTREAM_DRIFT_DENY", 503)
+        req = urllib.request.Request(
+            self.upstream + path,
+            method=method,
+            data=body,
+            headers=dict(headers),
+        )
+        try:
+            with urllib.request.urlopen(req, timeout=self.timeout) as response:
+                raw = response.read(MAX_RESPONSE + 1)
+                status = response.status
+                content_type = response.headers.get("Content-Type", "application/json")
+        except urllib.error.HTTPError as exc:
+            raw = exc.read(MAX_RESPONSE + 1)
+            status = exc.code
+            content_type = exc.headers.get("Content-Type", "application/json")
+        except Exception as exc:
+            raise EdgeError("UPSTREAM_UNAVAILABLE", 502) from exc
+        if len(raw) > MAX_RESPONSE:
+            raise EdgeError("UPSTREAM_RESPONSE_TOO_LARGE", 502)
+        if content_type.split(";", 1)[0].strip().lower() != "application/json":
+            raise EdgeError("UPSTREAM_CONTENT_TYPE_INVALID", 502)
+        return status, raw, "application/json"
+
+
+class EdgeServer(ThreadingHTTPServer):
+    daemon_threads = True
+    request_queue_size = 64
+    allow_reuse_address = True
+
+
+class Handler(BaseHTTPRequestHandler):
+    proxy = Proxy()
+    limiter = RateLimiter()
+    concurrency = threading.BoundedSemaphore(MAX_CONCURRENT_REQUESTS)
+    expected_domain = ""
+
+    server_version = "AIOSGPTIngressEdge/1.0"
+    sys_version = ""
+
+    def log_message(self, *_args: Any) -> None:
+        return
+
+    def respond_json(self, value: Mapping[str, Any], status: int) -> None:
+        raw = canonical(value)
+        self.send_response(status)
+        self.send_header("Content-Type", "application/json")
+        self.send_header("Cache-Control", "no-store")
+        self.send_header("X-Content-Type-Options", "nosniff")
+        self.send_header("Referrer-Policy", "no-referrer")
+        self.send_header("Content-Security-Policy", "default-src 'none'")
+        self.send_header("Strict-Transport-Security", "max-age=31536000")
+        self.send_header("Content-Length", str(len(raw)))
+        self.end_headers()
+        self.wfile.write(raw)
+
+    def send_proxy_response(self, status: int, raw: bytes, content_type: str) -> None:
+        self.send_response(status)
+        self.send_header("Content-Type", content_type)
+        self.send_header("Cache-Control", "no-store")
+        self.send_header("X-Content-Type-Options", "nosniff")
+        self.send_header("Referrer-Policy", "no-referrer")
+        self.send_header("Content-Security-Policy", "default-src 'none'")
+        self.send_header("Strict-Transport-Security", "max-age=31536000")
+        self.send_header("Content-Length", str(len(raw)))
+        self.end_headers()
+        self.wfile.write(raw)
+
+    def client_key(self) -> str:
+        return str(self.client_address[0])
+
+    def validate_host(self) -> None:
+        if not self.expected_domain:
+            return
+        host = self.headers.get("Host", "").split(":", 1)[0].lower()
+        if not hmac.compare_digest(host, self.expected_domain):
+            raise EdgeError("HOST_HEADER_DENY", 421)
+
+    def body(self) -> bytes:
+        try:
+            length = int(self.headers.get("Content-Length", "-1"))
+        except ValueError as exc:
+            raise EdgeError("CONTENT_LENGTH_INVALID") from exc
+        if length < 1 or length > MAX_BODY:
+            raise EdgeError("REQUEST_SIZE_INVALID", 413)
+        raw = self.rfile.read(length)
+        if len(raw) != length:
+            raise EdgeError("REQUEST_BODY_TRUNCATED", 400)
+        # Validate JSON at the edge so malformed bodies do not consume upstream work.
+        try:
+            value = json.loads(raw.decode("utf-8"))
+        except Exception as exc:
+            raise EdgeError("REQUEST_JSON_INVALID", 400) from exc
+        if not isinstance(value, Mapping):
+            raise EdgeError("REQUEST_JSON_OBJECT_REQUIRED", 400)
+        return canonical(value)
+
+    def handle_proxy(self, method: str) -> None:
+        client_id = _client_pseudonym(self.client_key())
+        started = time.monotonic()
+        acquired = self.concurrency.acquire(blocking=False)
+        if not acquired:
+            audit("EDGE_BUSY", client_id=client_id, method=method)
+            self.respond_json({"error": "EDGE_BUSY"}, 503)
+            return
+        label = "unknown"
+        try:
+            self.validate_host()
+            path = allowed_route(method, self.path)
+            label = route_label(method, path)
+            if path == "/health" and method == "GET":
+                self.respond_json(
+                    {
+                        "status": "ok",
+                        "service": "aios-gpt-https-edge",
+                        "upstream": "loopback-readonly-ingress",
+                    },
+                    200,
+                )
+                audit(
+                    "EDGE_REQUEST",
+                    client_id=client_id,
+                    method=method,
+                    route=label,
+                    status=200,
+                    duration_ms=int((time.monotonic() - started) * 1000),
+                )
+                return
+            if not self.limiter.allowed(self.client_key()):
+                audit(
+                    "EDGE_RATE_LIMIT",
+                    client_id=client_id,
+                    method=method,
+                    route=label,
+                )
+                self.respond_json({"error": "RATE_LIMITED"}, 429)
+                return
+            has_body = method == "POST"
+            headers = forwarded_headers(self.headers, body=has_body)
+            body = self.body() if has_body else None
+            status, raw, content_type = self.proxy.request(
+                method,
+                path,
+                headers,
+                body,
+            )
+            self.send_proxy_response(status, raw, content_type)
+            audit(
+                "EDGE_REQUEST",
+                client_id=client_id,
+                method=method,
+                route=label,
+                status=status,
+                duration_ms=int((time.monotonic() - started) * 1000),
+            )
+        except EdgeError as exc:
+            audit(
+                "EDGE_REJECT",
+                client_id=client_id,
+                method=method,
+                route=label,
+                status=exc.status,
+                error=exc.code,
+            )
+            self.respond_json({"error": exc.code}, exc.status)
+        finally:
+            self.concurrency.release()
+
+    def do_GET(self) -> None:
+        self.handle_proxy("GET")
+
+    def do_POST(self) -> None:
+        self.handle_proxy("POST")
+
+    def do_PUT(self) -> None:
+        self.respond_json({"error": "METHOD_DENY"}, 405)
+
+    def do_PATCH(self) -> None:
+        self.respond_json({"error": "METHOD_DENY"}, 405)
+
+    def do_DELETE(self) -> None:
+        self.respond_json({"error": "METHOD_DENY"}, 405)
+
+    def do_CONNECT(self) -> None:
+        self.respond_json({"error": "METHOD_DENY"}, 405)
+
+    def do_OPTIONS(self) -> None:
+        self.respond_json({"error": "METHOD_DENY"}, 405)
+
+
+def serve() -> None:
+    host, public_domain = validated_edge_identity(
+        os.environ.get("AIOS_GPT_EDGE_HOST", "127.0.0.1"),
+        os.environ.get(PUBLIC_APPROVAL_ENV, "0"),
+        os.environ.get(PUBLIC_DOMAIN_ENV, ""),
+    )
+    port = int(os.environ.get("AIOS_GPT_EDGE_PORT", "18443"))
+    if not 1024 <= port <= 65535:
+        raise RuntimeError("GPT_EDGE_PORT_INVALID")
+    Handler.expected_domain = public_domain
+
+    credentials = Path(os.environ.get("CREDENTIALS_DIRECTORY", "/run/credentials"))
+    cert_path = Path(
+        os.environ.get(
+            "AIOS_GPT_EDGE_TLS_CERT_FILE",
+            str(credentials / "tls_cert"),
+        )
+    )
+    key_path = Path(
+        os.environ.get(
+            "AIOS_GPT_EDGE_TLS_KEY_FILE",
+            str(credentials / "tls_key"),
+        )
+    )
+    context = build_tls_context(cert_path, key_path)
+
+    server = EdgeServer((host, port), Handler)
+    server.socket = context.wrap_socket(server.socket, server_side=True)
+    server.serve_forever()
+
+
+if __name__ == "__main__":
+    serve()

active-runtime/app/aios_gpt_https_edge_activation.py

@@ -0,0 +1,561 @@
+#!/usr/bin/env python3
+"""Owner-approved activation/rollback controller for the AIOS GPT HTTPS edge.
+
+This controller never creates approval, TLS, or ingress credentials. It consumes
+owner-managed inputs, applies one exact public-bind overlay, verifies the live
+edge, and provides an idempotent rollback path.
+"""
+
+from __future__ import annotations
+
+from dataclasses import dataclass
+from datetime import datetime, timezone
+import argparse
+import hashlib
+import json
+import os
+from pathlib import Path
+import re
+import socket
+import ssl
+import stat
+import subprocess
+import tempfile
+import time
+import urllib.request
+from typing import Any, Mapping
+
+APPROVAL_SCHEMA = "aios.gpt-https-edge.public-approval.v1"
+STATE_SCHEMA = "aios.gpt-https-edge.activation-state.v1"
+
+OID_RE = re.compile(r"^[0-9a-f]{40}$")
+SHA_RE = re.compile(r"^[0-9a-f]{64}$")
+APPROVAL_ID_RE = re.compile(r"^[A-Za-z0-9][A-Za-z0-9._:-]{7,127}$")
+DOMAIN_RE = re.compile(
+    r"(?=.{1,253}\Z)"
+    r"(?:[a-z0-9](?:[a-z0-9-]{0,61}[a-z0-9])?\.)+"
+    r"[a-z0-9](?:[a-z0-9-]{0,61}[a-z0-9])?"
+)
+
+EDGE_UNIT = "aios-gpt-https-edge.service"
+INGRESS_UNIT = "aios-gpt-online-task-ingress.service"
+PUBLIC_HOST = "0.0.0.0"
+PUBLIC_PORT = 18443
+INGRESS_HEALTH = "http://127.0.0.1:18945/health"
+DROPIN_NAME = "zzzzzzzzzz-owner-approved-public.conf"
+
+
+class ActivationError(RuntimeError):
+    pass
+
+
+def canonical(value: object) -> bytes:
+    return json.dumps(
+        value,
+        ensure_ascii=False,
+        sort_keys=True,
+        separators=(",", ":"),
+    ).encode("utf-8")
+
+
+def sha256_bytes(data: bytes) -> str:
+    return hashlib.sha256(data).hexdigest()
+
+
+def sha256_file(path: Path) -> str:
+    return sha256_bytes(path.read_bytes())
+
+
+def parse_time(value: object, *, field: str) -> datetime:
+    if not isinstance(value, str) or not value.strip():
+        raise ActivationError("APPROVAL_" + field.upper() + "_INVALID")
+    text = value.strip()
+    if text.endswith("Z"):
+        text = text[:-1] + "+00:00"
+    try:
+        parsed = datetime.fromisoformat(text)
+    except ValueError as exc:
+        raise ActivationError("APPROVAL_" + field.upper() + "_INVALID") from exc
+    if parsed.tzinfo is None:
+        raise ActivationError("APPROVAL_" + field.upper() + "_INVALID")
+    return parsed.astimezone(timezone.utc)
+
+
+def validate_domain(value: object) -> str:
+    if not isinstance(value, str):
+        raise ActivationError("APPROVAL_DOMAIN_INVALID")
+    domain = value.strip().lower()
+    if DOMAIN_RE.fullmatch(domain) is None:
+        raise ActivationError("APPROVAL_DOMAIN_INVALID")
+    return domain
+
+
+def validate_root_file(
+    path: Path,
+    *,
+    exact_mode: int | None = None,
+    max_size: int = 256 * 1024,
+) -> Path:
+    if path.is_symlink() or not path.is_file():
+        raise ActivationError("OWNER_FILE_INVALID")
+    info = path.stat()
+    if info.st_uid != 0 or info.st_gid != 0:
+        raise ActivationError("OWNER_FILE_OWNERSHIP_INVALID")
+    mode = stat.S_IMODE(info.st_mode)
+    if exact_mode is not None and mode != exact_mode:
+        raise ActivationError("OWNER_FILE_MODE_INVALID")
+    if info.st_size < 16 or info.st_size > max_size:
+        raise ActivationError("OWNER_FILE_SIZE_INVALID")
+    return path
+
+
+def certificate_dns_names(cert_path: Path) -> tuple[str, ...]:
+    try:
+        decoded = ssl._ssl._test_decode_cert(str(cert_path))
+    except Exception as exc:
+        raise ActivationError("TLS_CERT_DECODE_INVALID") from exc
+    names = [
+        value
+        for kind, value in decoded.get("subjectAltName", ())
+        if kind == "DNS"
+    ]
+    if not names:
+        for rdn in decoded.get("subject", ()):
+            for key, value in rdn:
+                if key == "commonName":
+                    names.append(value)
+    return tuple(names)
+
+
+def cert_matches_domain(cert_path: Path, domain: str) -> bool:
+    names = certificate_dns_names(cert_path)
+    try:
+        return bool(names) and any(ssl._dnsname_match(name, domain) for name in names)
+    except ssl.CertificateError:
+        return False
+
+
+def validate_tls_pair(
+    cert_path: Path,
+    key_path: Path,
+    domain: str,
+    *,
+    now_fn=time.time,
+) -> dict[str, Any]:
+    validate_root_file(cert_path)
+    validate_root_file(key_path, exact_mode=0o600)
+    context = ssl.SSLContext(ssl.PROTOCOL_TLS_SERVER)
+    context.minimum_version = ssl.TLSVersion.TLSv1_2
+    context.options |= ssl.OP_NO_COMPRESSION
+    try:
+        context.load_cert_chain(str(cert_path), str(key_path))
+    except Exception as exc:
+        raise ActivationError("TLS_CERT_KEY_PAIR_INVALID") from exc
+    if not cert_matches_domain(cert_path, domain):
+        raise ActivationError("TLS_CERT_DOMAIN_MISMATCH")
+    decoded = ssl._ssl._test_decode_cert(str(cert_path))
+    not_after = decoded.get("notAfter")
+    if not not_after:
+        raise ActivationError("TLS_CERT_EXPIRY_MISSING")
+    remaining = ssl.cert_time_to_seconds(not_after) - now_fn()
+    if remaining < 7 * 86400:
+        raise ActivationError("TLS_CERT_EXPIRY_TOO_CLOSE")
+    return {
+        "cert_sha256": sha256_file(cert_path),
+        "remaining_days": int(remaining // 86400),
+    }
+
+
+@dataclass(frozen=True)
+class Approval:
+    approval_id: str
+    candidate_oid: str
+    domain: str
+    bind_host: str
+    port: int
+    tls_cert_sha256: str
+    owner: str
+    issued_at: datetime
+    not_before: datetime
+    expires_at: datetime
+    raw_sha256: str
+
+    @classmethod
+    def load(
+        cls,
+        path: Path,
+        *,
+        candidate_oid: str,
+        cert_path: Path,
+        now: datetime | None = None,
+    ) -> "Approval":
+        validate_root_file(path, exact_mode=0o600, max_size=64 * 1024)
+        try:
+            value = json.loads(path.read_text(encoding="utf-8"))
+        except Exception as exc:
+            raise ActivationError("APPROVAL_JSON_INVALID") from exc
+        if not isinstance(value, Mapping):
+            raise ActivationError("APPROVAL_JSON_INVALID")
+        allowed = {
+            "schema",
+            "approval_id",
+            "candidate_oid",
+            "domain",
+            "public_bind",
+            "bind_host",
+            "port",
+            "tls_cert_sha256",
+            "issued_at",
+            "not_before",
+            "expires_at",
+            "owner",
+        }
+        if set(value) != allowed or value.get("schema") != APPROVAL_SCHEMA:
+            raise ActivationError("APPROVAL_FIELDS_INVALID")
+
+        approval_id = value.get("approval_id")
+        if not isinstance(approval_id, str) or APPROVAL_ID_RE.fullmatch(approval_id) is None:
+            raise ActivationError("APPROVAL_ID_INVALID")
+        if not OID_RE.fullmatch(str(value.get("candidate_oid") or "")):
+            raise ActivationError("APPROVAL_CANDIDATE_INVALID")
+        if value["candidate_oid"] != candidate_oid:
+            raise ActivationError("APPROVAL_CANDIDATE_MISMATCH")
+        if value.get("public_bind") is not True:
+            raise ActivationError("APPROVAL_PUBLIC_BIND_REQUIRED")
+        if value.get("bind_host") != PUBLIC_HOST or value.get("port") != PUBLIC_PORT:
+            raise ActivationError("APPROVAL_BINDING_INVALID")
+
+        cert_sha = str(value.get("tls_cert_sha256") or "")
+        if SHA_RE.fullmatch(cert_sha) is None:
+            raise ActivationError("APPROVAL_CERT_SHA_INVALID")
+        if not cert_path.is_file() or sha256_file(cert_path) != cert_sha:
+            raise ActivationError("APPROVAL_CERT_SHA_MISMATCH")
+
+        domain = validate_domain(value.get("domain"))
+        owner = value.get("owner")
+        if not isinstance(owner, str) or not owner.strip() or len(owner) > 255:
+            raise ActivationError("APPROVAL_OWNER_INVALID")
+
+        issued_at = parse_time(value.get("issued_at"), field="issued_at")
+        not_before = parse_time(value.get("not_before"), field="not_before")
+        expires_at = parse_time(value.get("expires_at"), field="expires_at")
+        current = now or datetime.now(timezone.utc)
+        if not (issued_at <= not_before <= expires_at):
+            raise ActivationError("APPROVAL_TIME_ORDER_INVALID")
+        if current < not_before:
+            raise ActivationError("APPROVAL_NOT_YET_VALID")
+        if current >= expires_at:
+            raise ActivationError("APPROVAL_EXPIRED")
+        if expires_at - issued_at > __import__("datetime").timedelta(days=7):
+            raise ActivationError("APPROVAL_VALIDITY_TOO_LONG")
+
+        return cls(
+            approval_id=approval_id,
+            candidate_oid=candidate_oid,
+            domain=domain,
+            bind_host=PUBLIC_HOST,
+            port=PUBLIC_PORT,
+            tls_cert_sha256=cert_sha,
+            owner=owner.strip(),
+            issued_at=issued_at,
+            not_before=not_before,
+            expires_at=expires_at,
+            raw_sha256=sha256_file(path),
+        )
+
+
+def validate_release(release_root: Path) -> dict[str, Path]:
+    if release_root.is_symlink() or not release_root.is_dir():
+        raise ActivationError("IMMUTABLE_RELEASE_INVALID")
+    edge = release_root / "active-runtime/app/aios_gpt_https_edge.py"
+    controller = release_root / "active-runtime/app/aios_gpt_https_edge_activation.py"
+    for path in (edge, controller):
+        if path.is_symlink() or not path.is_file():
+            raise ActivationError("IMMUTABLE_RELEASE_INVALID")
+    return {"edge": edge, "controller": controller}
+
+
+def render_dropin(approval: Approval, release_root: Path) -> str:
+    app = release_root / "active-runtime/app/aios_gpt_https_edge.py"
+    return (
+        "[Service]\n"
+        f"Environment=AIOS_GPT_EDGE_HOST={approval.bind_host}\n"
+        f"Environment=AIOS_GPT_EDGE_PORT={approval.port}\n"
+        "Environment=AIOS_GPT_EDGE_PUBLIC_BIND_APPROVED=1\n"
+        f"Environment=AIOS_GPT_EDGE_PUBLIC_DOMAIN={approval.domain}\n"
+        "ExecStart=\n"
+        f"ExecStart=/usr/bin/python3 -B {app}\n"
+        f"ReadOnlyPaths={release_root}\n"
+    )
+
+
+def atomic_write(path: Path, data: str, mode: int = 0o600) -> None:
+    path.parent.mkdir(parents=True, exist_ok=True)
+    fd, name = tempfile.mkstemp(prefix="." + path.name + ".", dir=path.parent)
+    try:
+        os.fchmod(fd, mode)
+        with os.fdopen(fd, "w", encoding="utf-8") as handle:
+            handle.write(data)
+            handle.flush()
+            os.fsync(handle.fileno())
+        os.replace(name, path)
+    finally:
+        if os.path.exists(name):
+            os.unlink(name)
+
+
+class Systemd:
+    def run(self, *args: str, check: bool = True) -> subprocess.CompletedProcess[str]:
+        return subprocess.run(
+            ["/usr/bin/systemctl", *args],
+            check=check,
+            text=True,
+            capture_output=True,
+            timeout=30,
+        )
+
+    def active(self, unit: str) -> bool:
+        return self.run("is-active", "--quiet", unit, check=False).returncode == 0
+
+    def enabled(self, unit: str) -> bool:
+        return self.run("is-enabled", "--quiet", unit, check=False).returncode == 0
+
+    def show(self, unit: str, prop: str) -> str:
+        return self.run("show", unit, f"--property={prop}", "--value").stdout.strip()
+
+
+def require_ingress_ready(systemd: Systemd) -> None:
+    if not systemd.active(INGRESS_UNIT):
+        raise ActivationError("INGRESS_SERVICE_NOT_ACTIVE")
+    try:
+        with urllib.request.urlopen(INGRESS_HEALTH, timeout=5) as response:
+            if response.status != 200:
+                raise ActivationError("INGRESS_HEALTH_FAILED")
+            value = json.loads(response.read().decode())
+    except ActivationError:
+        raise
+    except Exception as exc:
+        raise ActivationError("INGRESS_HEALTH_FAILED") from exc
+    if value.get("status") != "ok":
+        raise ActivationError("INGRESS_HEALTH_FAILED")
+
+
+def tls_probe(domain: str, cert_path: Path, *, port: int = PUBLIC_PORT) -> None:
+    context = ssl.create_default_context(cafile=str(cert_path))
+    context.check_hostname = True
+    context.verify_mode = ssl.CERT_REQUIRED
+    try:
+        with socket.create_connection(("127.0.0.1", port), timeout=5) as sock:
+            with context.wrap_socket(sock, server_hostname=domain) as tls_sock:
+                if tls_sock.version() not in {"TLSv1.2", "TLSv1.3"}:
+                    raise ActivationError("EDGE_TLS_VERSION_INVALID")
+    except ActivationError:
+        raise
+    except Exception as exc:
+        raise ActivationError("EDGE_TLS_PROBE_FAILED") from exc
+
+
+def receipt_for(
+    approval: Approval,
+    release_root: Path,
+    *,
+    state: str,
+) -> dict[str, Any]:
+    return {
+        "schema": STATE_SCHEMA,
+        "state": state,
+        "approval_id": approval.approval_id,
+        "approval_sha256": approval.raw_sha256,
+        "candidate_oid": approval.candidate_oid,
+        "domain": approval.domain,
+        "bind_host": approval.bind_host,
+        "port": approval.port,
+        "tls_cert_sha256": approval.tls_cert_sha256,
+        "release_root": str(release_root),
+        "updated_at": datetime.now(timezone.utc).isoformat(),
+    }
+
+
+def activate(
+    *,
+    approval: Approval,
+    release_root: Path,
+    cert_path: Path,
+    key_path: Path,
+    dropin_path: Path,
+    state_path: Path,
+    systemd: Systemd,
+) -> dict[str, Any]:
+    if os.geteuid() != 0:
+        raise ActivationError("ROOT_REQUIRED")
+    release = validate_release(release_root)
+    app = release["edge"]
+    require_ingress_ready(systemd)
+    validate_tls_pair(cert_path, key_path, approval.domain)
+
+    if state_path.is_file():
+        try:
+            existing = json.loads(state_path.read_text(encoding="utf-8"))
+        except Exception:
+            existing = {}
+        if (
+            existing.get("schema") == STATE_SCHEMA
+            and existing.get("state") == "ACTIVE"
+            and existing.get("approval_sha256") == approval.raw_sha256
+            and existing.get("candidate_oid") == approval.candidate_oid
+            and systemd.active(EDGE_UNIT)
+        ):
+            return existing
+
+    dropin = render_dropin(approval, release_root)
+    modified = False
+    try:
+        atomic_write(dropin_path, dropin, mode=0o644)
+        modified = True
+        systemd.run("daemon-reload")
+        systemd.run("enable", EDGE_UNIT)
+        systemd.run("restart", EDGE_UNIT)
+        if not systemd.active(EDGE_UNIT):
+            raise ActivationError("EDGE_SERVICE_NOT_ACTIVE")
+        exec_start = systemd.show(EDGE_UNIT, "ExecStart")
+        environment = systemd.show(EDGE_UNIT, "Environment")
+        if str(app) not in exec_start:
+            raise ActivationError("EDGE_EFFECTIVE_EXECSTART_INVALID")
+        for expected in (
+            f"AIOS_GPT_EDGE_HOST={approval.bind_host}",
+            f"AIOS_GPT_EDGE_PORT={approval.port}",
+            "AIOS_GPT_EDGE_PUBLIC_BIND_APPROVED=1",
+            f"AIOS_GPT_EDGE_PUBLIC_DOMAIN={approval.domain}",
+        ):
+            if expected not in environment:
+                raise ActivationError("EDGE_EFFECTIVE_ENVIRONMENT_INVALID")
+        tls_probe(approval.domain, cert_path, port=approval.port)
+        receipt = receipt_for(approval, release_root, state="ACTIVE")
+        atomic_write(state_path, json.dumps(receipt, indent=2, sort_keys=True) + "\n")
+        return receipt
+    except Exception:
+        if modified:
+            rollback(
+                dropin_path=dropin_path,
+                state_path=state_path,
+                systemd=systemd,
+                reason="activation_failed",
+            )
+        raise
+
+
+def rollback(
+    *,
+    dropin_path: Path,
+    state_path: Path,
+    systemd: Systemd,
+    reason: str,
+) -> dict[str, Any]:
+    if os.geteuid() != 0:
+        raise ActivationError("ROOT_REQUIRED")
+    systemd.run("stop", EDGE_UNIT, check=False)
+    systemd.run("disable", EDGE_UNIT, check=False)
+    dropin_path.unlink(missing_ok=True)
+    systemd.run("daemon-reload")
+    if systemd.active(EDGE_UNIT):
+        raise ActivationError("EDGE_ROLLBACK_SERVICE_STILL_ACTIVE")
+    previous: dict[str, Any] = {}
+    if state_path.is_file():
+        try:
+            previous = json.loads(state_path.read_text(encoding="utf-8"))
+        except Exception:
+            previous = {}
+    receipt = {
+        "schema": STATE_SCHEMA,
+        "state": "ROLLED_BACK",
+        "reason": reason[:255],
+        "candidate_oid": previous.get("candidate_oid"),
+        "approval_id": previous.get("approval_id"),
+        "domain": previous.get("domain"),
+        "updated_at": datetime.now(timezone.utc).isoformat(),
+    }
+    atomic_write(state_path, json.dumps(receipt, indent=2, sort_keys=True) + "\n")
+    return receipt
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser()
+    parser.add_argument("command", choices=("check", "activate", "rollback"))
+    parser.add_argument("--candidate")
+    parser.add_argument("--approval", default="/etc/aios-gitops/gpt-https-edge.approved")
+    parser.add_argument("--tls-cert", default="/etc/aios-gitops/gpt-https-edge-tls.crt")
+    parser.add_argument("--tls-key", default="/etc/aios-gitops/gpt-https-edge-tls.key")
+    parser.add_argument("--release-root")
+    parser.add_argument(
+        "--dropin",
+        default=f"/etc/systemd/system/{EDGE_UNIT}.d/{DROPIN_NAME}",
+    )
+    parser.add_argument(
+        "--state",
+        default="/var/lib/aios-gpt-https-edge-activation/state.json",
+    )
+    parser.add_argument("--reason", default="owner_requested")
+    args = parser.parse_args()
+
+    systemd = Systemd()
+    if args.command == "rollback":
+        value = rollback(
+            dropin_path=Path(args.dropin),
+            state_path=Path(args.state),
+            systemd=systemd,
+            reason=args.reason,
+        )
+        print(json.dumps(value, indent=2, sort_keys=True))
+        return 0
+
+    candidate = str(args.candidate or "")
+    if OID_RE.fullmatch(candidate) is None:
+        raise SystemExit("CANDIDATE_REQUIRED")
+    release_root = (
+        Path(args.release_root)
+        if args.release_root
+        else Path(f"/opt/aios/releases/{candidate}-gpt-https-edge-public")
+    )
+    cert_path = Path(args.tls_cert)
+    key_path = Path(args.tls_key)
+    approval = Approval.load(
+        Path(args.approval),
+        candidate_oid=candidate,
+        cert_path=cert_path,
+    )
+    tls = validate_tls_pair(cert_path, key_path, approval.domain)
+    release = validate_release(release_root)
+    require_ingress_ready(systemd)
+
+    if args.command == "check":
+        value = {
+            "status": "READY",
+            "candidate_oid": candidate,
+            "approval_id": approval.approval_id,
+            "approval_sha256": approval.raw_sha256,
+            "domain": approval.domain,
+            "bind_host": approval.bind_host,
+            "port": approval.port,
+            "edge_sha256": sha256_file(release["edge"]),
+            "controller_sha256": sha256_file(release["controller"]),
+            **tls,
+            "public_network_mutation": False,
+        }
+        print(json.dumps(value, indent=2, sort_keys=True))
+        return 0
+
+    value = activate(
+        approval=approval,
+        release_root=release_root,
+        cert_path=cert_path,
+        key_path=key_path,
+        dropin_path=Path(args.dropin),
+        state_path=Path(args.state),
+        systemd=systemd,
+    )
+    print(json.dumps(value, indent=2, sort_keys=True))
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

active-runtime/app/aios_gpt_online_task_ingress.py

@@ -0,0 +1,816 @@
+#!/usr/bin/env python3
+"""Credential-isolated host-side ingress for externally scheduled read-only Online Tasks.
+
+This service is intentionally not a public AIOS control plane. It accepts a tiny,
+authenticated external intent and translates it to the canonical loopback Bridge.
+Callers cannot select repository coordinates, capabilities, execution action,
+worker, branch policy, credentials, verification decisions, or write permissions.
+"""
+
+from __future__ import annotations
+
+from dataclasses import dataclass
+from datetime import datetime, timezone
+import hashlib
+import hmac
+from http.server import BaseHTTPRequestHandler, ThreadingHTTPServer
+import json
+import os
+from pathlib import Path
+import re
+import sqlite3
+import threading
+import urllib.error
+import urllib.request
+from typing import Any, Mapping
+
+SCHEMA = "aios.gpt-online-task-ingress.request.v1"
+RESPONSE_SCHEMA = "aios.gpt-online-task-ingress.response.v1"
+OPENAPI_VERSION = "3.1.0"
+
+REQUEST_ID_RE = re.compile(r"^[A-Za-z0-9][A-Za-z0-9._:-]{7,127}$")
+TASK_ID_RE = re.compile(r"^[A-Za-z0-9][A-Za-z0-9._:-]{2,179}$")
+JOB_ID_RE = re.compile(r"^job-[0-9a-f]{32}$")
+OID_RE = re.compile(r"^[0-9a-f]{40}$")
+MAX_BODY = 64 * 1024
+MAX_GOAL = 12_000
+MAX_INSTRUCTIONS = 24_000
+MAX_CRITERIA = 16
+MAX_CRITERION_TEXT = 1_000
+
+READONLY_CAPABILITIES = ("git.read", "git.test")
+REPOSITORY = "aios/aios"
+
+
+class IngressError(RuntimeError):
+    def __init__(self, code: str, status: int = 400) -> None:
+        super().__init__(code)
+        self.code = code
+        self.status = status
+
+
+def canonical(value: object) -> bytes:
+    return json.dumps(
+        value,
+        ensure_ascii=False,
+        sort_keys=True,
+        separators=(",", ":"),
+    ).encode("utf-8")
+
+
+def digest(value: object) -> str:
+    return hashlib.sha256(canonical(value)).hexdigest()
+
+
+def now() -> str:
+    return datetime.now(timezone.utc).isoformat()
+
+
+def load_secret(path: Path, *, code: str) -> str:
+    if path.is_symlink() or not path.is_file():
+        raise RuntimeError(code + "_UNAVAILABLE")
+    value = path.read_text(encoding="utf-8").strip()
+    if len(value) < 32 or "\n" in value or "\r" in value:
+        raise RuntimeError(code + "_INVALID")
+    return value
+
+
+def normalized_request(value: Mapping[str, Any]) -> dict[str, Any]:
+    allowed = {
+        "schema",
+        "request_id",
+        "goal",
+        "instructions",
+        "success_criteria",
+        "constraints",
+    }
+    if set(value) - allowed:
+        raise IngressError("UNKNOWN_FIELDS")
+    if value.get("schema") != SCHEMA:
+        raise IngressError("SCHEMA_INVALID")
+
+    request_id = value.get("request_id")
+    if not isinstance(request_id, str) or REQUEST_ID_RE.fullmatch(request_id) is None:
+        raise IngressError("REQUEST_ID_INVALID")
+
+    goal = value.get("goal")
+    if not isinstance(goal, str) or not goal.strip() or len(goal) > MAX_GOAL:
+        raise IngressError("GOAL_INVALID")
+    instructions = value.get("instructions", "")
+    if not isinstance(instructions, str) or len(instructions) > MAX_INSTRUCTIONS:
+        raise IngressError("INSTRUCTIONS_INVALID")
+
+    criteria = value.get("success_criteria") or []
+    if not isinstance(criteria, list) or len(criteria) > MAX_CRITERIA:
+        raise IngressError("SUCCESS_CRITERIA_INVALID")
+    normalized_criteria = []
+    for item in criteria:
+        if not isinstance(item, Mapping) or set(item) != {"id", "description"}:
+            raise IngressError("SUCCESS_CRITERIA_INVALID")
+        criterion_id = item.get("id")
+        description = item.get("description")
+        if (
+            not isinstance(criterion_id, str)
+            or not criterion_id.strip()
+            or len(criterion_id) > 120
+            or not isinstance(description, str)
+            or not description.strip()
+            or len(description) > MAX_CRITERION_TEXT
+        ):
+            raise IngressError("SUCCESS_CRITERIA_INVALID")
+        normalized_criteria.append(
+            {"id": criterion_id.strip(), "description": description.strip()}
+        )
+
+    constraints = value.get("constraints") or {}
+    if not isinstance(constraints, Mapping):
+        raise IngressError("CONSTRAINTS_INVALID")
+    allowed_constraints = {
+        "max_files",
+        "focus_paths",
+        "notes",
+    }
+    if set(constraints) - allowed_constraints:
+        raise IngressError("CONSTRAINTS_FIELDS_INVALID")
+    normalized_constraints: dict[str, Any] = {}
+    if "max_files" in constraints:
+        max_files = constraints["max_files"]
+        if (
+            not isinstance(max_files, int)
+            or isinstance(max_files, bool)
+            or not 1 <= max_files <= 500
+        ):
+            raise IngressError("MAX_FILES_INVALID")
+        normalized_constraints["max_files"] = max_files
+    if "focus_paths" in constraints:
+        paths = constraints["focus_paths"]
+        if (
+            not isinstance(paths, list)
+            or len(paths) > 32
+            or not all(
+                isinstance(path, str)
+                and path
+                and len(path) <= 512
+                and not path.startswith("/")
+                and ".." not in Path(path).parts
+                for path in paths
+            )
+        ):
+            raise IngressError("FOCUS_PATHS_INVALID")
+        normalized_constraints["focus_paths"] = paths
+    if "notes" in constraints:
+        notes = constraints["notes"]
+        if not isinstance(notes, str) or len(notes) > 4_000:
+            raise IngressError("CONSTRAINT_NOTES_INVALID")
+        normalized_constraints["notes"] = notes
+
+    return {
+        "schema": SCHEMA,
+        "request_id": request_id,
+        "goal": goal.strip(),
+        "instructions": instructions,
+        "success_criteria": normalized_criteria,
+        "constraints": normalized_constraints,
+    }
+
+
+def validated_bind_host(value: str) -> str:
+    host = value.strip()
+    if host not in {"127.0.0.1", "::1", "localhost"}:
+        raise RuntimeError("GPT_INGRESS_NON_LOOPBACK_BIND_DENY")
+    return host
+
+
+def task_id_for(request_id: str) -> str:
+    suffix = hashlib.sha256(request_id.encode("utf-8")).hexdigest()[:24]
+    return "gpt-readonly-" + suffix
+
+
+def branch_for(task_id: str) -> str:
+    return "agent/external-gpt-readonly/" + task_id[-24:]
+
+
+@dataclass(frozen=True)
+class Upstream:
+    bridge_base: str
+    task_base: str
+    control_token: str
+    timeout: int = 20
+
+    def _request(
+        self,
+        base: str,
+        method: str,
+        path: str,
+        body: Mapping[str, Any] | None = None,
+    ) -> tuple[int, dict[str, Any]]:
+        raw_body = None if body is None else canonical(body)
+        headers = {
+            "Authorization": "Bearer " + self.control_token,
+            "Accept": "application/json",
+        }
+        if raw_body is not None:
+            headers["Content-Type"] = "application/json"
+        req = urllib.request.Request(
+            base.rstrip("/") + path,
+            method=method,
+            data=raw_body,
+            headers=headers,
+        )
+        try:
+            with urllib.request.urlopen(req, timeout=self.timeout) as response:
+                raw = response.read()
+                status = response.status
+        except urllib.error.HTTPError as exc:
+            raw = exc.read()
+            status = exc.code
+        except Exception as exc:
+            raise IngressError("UPSTREAM_TRANSPORT_FAILED", 503) from exc
+        try:
+            value = json.loads(raw.decode("utf-8"))
+        except Exception as exc:
+            raise IngressError("UPSTREAM_RESPONSE_INVALID", 502) from exc
+        if not isinstance(value, Mapping):
+            raise IngressError("UPSTREAM_RESPONSE_INVALID", 502)
+        return status, dict(value)
+
+    def control_snapshot(self) -> dict[str, Any]:
+        status, value = self._request(
+            self.bridge_base,
+            "GET",
+            "/v2/control-plane-snapshot",
+        )
+        if status != 200:
+            raise IngressError("CONTROL_SNAPSHOT_UNAVAILABLE", 503)
+        return value
+
+    def create_governed_task(
+        self,
+        body: Mapping[str, Any],
+    ) -> tuple[int, dict[str, Any]]:
+        return self._request(
+            self.bridge_base,
+            "POST",
+            "/v2/governed-task",
+            body,
+        )
+
+    def execution_envelope(
+        self,
+        body: Mapping[str, Any],
+    ) -> tuple[int, dict[str, Any]]:
+        return self._request(
+            self.bridge_base,
+            "POST",
+            "/v2/execution-envelope",
+            body,
+        )
+
+    def get_task(self, task_id: str) -> tuple[int, dict[str, Any]]:
+        return self._request(
+            self.task_base,
+            "GET",
+            "/v1/tasks/" + task_id,
+        )
+
+
+class Store:
+    def __init__(self, path: Path) -> None:
+        self.path = Path(path)
+        self._lock = threading.RLock()
+        self.path.parent.mkdir(parents=True, exist_ok=True)
+        with self._connect() as con:
+            con.execute(
+                """
+                CREATE TABLE IF NOT EXISTS external_task_requests(
+                  request_id TEXT PRIMARY KEY,
+                  request_sha256 TEXT NOT NULL,
+                  task_id TEXT NOT NULL UNIQUE,
+                  job_id TEXT,
+                  base_oid TEXT,
+                  state TEXT NOT NULL,
+                  created_at TEXT NOT NULL,
+                  updated_at TEXT NOT NULL,
+                  response_json TEXT
+                )
+                """
+            )
+            con.commit()
+
+    def _connect(self) -> sqlite3.Connection:
+        con = sqlite3.connect(self.path, timeout=30)
+        con.row_factory = sqlite3.Row
+        con.execute("PRAGMA busy_timeout=30000")
+        return con
+
+    def get(self, request_id: str) -> sqlite3.Row | None:
+        with self._connect() as con:
+            return con.execute(
+                "SELECT * FROM external_task_requests WHERE request_id=?",
+                (request_id,),
+            ).fetchone()
+
+    def reserve(self, request: Mapping[str, Any], task_id: str) -> sqlite3.Row:
+        request_sha = digest(request)
+        timestamp = now()
+        with self._lock, self._connect() as con:
+            con.execute("BEGIN IMMEDIATE")
+            row = con.execute(
+                "SELECT * FROM external_task_requests WHERE request_id=?",
+                (request["request_id"],),
+            ).fetchone()
+            if row is not None:
+                if row["request_sha256"] != request_sha:
+                    raise IngressError("IDEMPOTENCY_CONFLICT", 409)
+                con.commit()
+                return row
+            con.execute(
+                """
+                INSERT INTO external_task_requests(
+                  request_id,request_sha256,task_id,state,created_at,updated_at
+                ) VALUES(?,?,?,?,?,?)
+                """,
+                (
+                    request["request_id"],
+                    request_sha,
+                    task_id,
+                    "PENDING",
+                    timestamp,
+                    timestamp,
+                ),
+            )
+            con.commit()
+        row = self.get(request["request_id"])
+        assert row is not None
+        return row
+
+    def bound(
+        self,
+        request_id: str,
+        *,
+        job_id: str,
+        base_oid: str,
+        response: Mapping[str, Any],
+    ) -> None:
+        with self._lock, self._connect() as con:
+            con.execute(
+                """
+                UPDATE external_task_requests
+                   SET job_id=?,base_oid=?,state='BOUND',response_json=?,updated_at=?
+                 WHERE request_id=?
+                """,
+                (
+                    job_id,
+                    base_oid,
+                    json.dumps(response, sort_keys=True, separators=(",", ":")),
+                    now(),
+                    request_id,
+                ),
+            )
+            con.commit()
+
+
+class IngressService:
+    def __init__(self, upstream: Upstream, store: Store) -> None:
+        self.upstream = upstream
+        self.store = store
+
+    @staticmethod
+    def _main_oid(snapshot: Mapping[str, Any]) -> str:
+        candidates = [
+            snapshot.get("authoritative_head_oid"),
+            snapshot.get("head"),
+            (snapshot.get("aggregate") or {}).get("authoritative_head_oid")
+            if isinstance(snapshot.get("aggregate"), Mapping)
+            else None,
+        ]
+        for candidate in candidates:
+            if isinstance(candidate, str) and OID_RE.fullmatch(candidate):
+                return candidate
+        raise IngressError("AUTHORITATIVE_HEAD_UNAVAILABLE", 503)
+
+    @staticmethod
+    def _public_response(
+        request_id: str,
+        task_id: str,
+        job_id: str,
+        state: str,
+    ) -> dict[str, Any]:
+        return {
+            "schema": RESPONSE_SCHEMA,
+            "request_id": request_id,
+            "task_id": task_id,
+            "job_id": job_id,
+            "state": state,
+            "execution_mode": "source_capture",
+            "capabilities": list(READONLY_CAPABILITIES),
+            "write_access": False,
+            "official_verification_required": True,
+        }
+
+    def _recover_existing(
+        self,
+        request: Mapping[str, Any],
+        row: sqlite3.Row,
+    ) -> dict[str, Any] | None:
+        task_id = str(row["task_id"])
+        status, task = self.upstream.get_task(task_id)
+        if status == 404:
+            return None
+        if status != 200:
+            raise IngressError("TASK_RECOVERY_FAILED", 503)
+        metadata = task.get("metadata")
+        execution = (
+            metadata.get("execution")
+            if isinstance(metadata, Mapping)
+            and isinstance(metadata.get("execution"), Mapping)
+            else {}
+        )
+        job_id = str(execution.get("job_id") or row["job_id"] or "")
+        base_oid = str(execution.get("base_oid") or row["base_oid"] or "")
+        if (
+            task.get("task_id") != task_id
+            or not JOB_ID_RE.fullmatch(job_id)
+            or not OID_RE.fullmatch(base_oid)
+            or execution.get("action") != "source_capture"
+            or sorted(execution.get("capability_scope") or [])
+            != sorted(READONLY_CAPABILITIES)
+            or (execution.get("governance") or {}).get("writer_lease") is not False
+        ):
+            raise IngressError("TASK_RECOVERY_BINDING_INVALID", 409)
+        response = self._public_response(
+            request["request_id"],
+            task_id,
+            job_id,
+            str(task.get("status") or "BOUND"),
+        )
+        self.store.bound(
+            request["request_id"],
+            job_id=job_id,
+            base_oid=base_oid,
+            response=response,
+        )
+        return response
+
+    def create(self, raw: Mapping[str, Any]) -> dict[str, Any]:
+        request = normalized_request(raw)
+        task_id = task_id_for(request["request_id"])
+        row = self.store.reserve(request, task_id)
+        if row["state"] == "BOUND" and row["response_json"]:
+            return json.loads(row["response_json"])
+
+        recovered = self._recover_existing(request, row)
+        if recovered is not None:
+            return recovered
+
+        snapshot = self.upstream.control_snapshot()
+        base_oid = self._main_oid(snapshot)
+        target_branch = branch_for(task_id)
+        binding_key = "gpt-readonly-binding-" + digest(
+            {"request_id": request["request_id"], "base_oid": base_oid}
+        )[:24]
+        constraints = {
+            **request["constraints"],
+            "repository_mutation": "forbidden",
+            "control_plane_side_effects": "forbidden",
+            "credential_access": "forbidden",
+        }
+        body = {
+            "task_id": task_id,
+            "repository": REPOSITORY,
+            "base_oid": base_oid,
+            "expected_branch_oid": base_oid,
+            "target_branch": target_branch,
+            "branch_mode": "CREATE_NEW",
+            "generation": 1,
+            "risk": "low",
+            "description": "Externally requested read-only governed Online Task.",
+            "idempotency_key": binding_key,
+            "priority": 0,
+            "max_attempts": 3,
+            "require_online_task_projection": True,
+            "online_task": {
+                "goal": request["goal"],
+                "instructions": request["instructions"],
+                "success_criteria": request["success_criteria"],
+                "constraints": constraints,
+                "required_capabilities": list(READONLY_CAPABILITIES),
+                "priority": 0,
+                "max_attempts": 3,
+            },
+            "execution": {
+                "action": "source_capture",
+                "adapter": "source_capture",
+                "capability_scope": list(READONLY_CAPABILITIES),
+                "source_signals": [
+                    "external read-only source analysis request",
+                    "no repository mutation no credential access no control-plane side effects",
+                ],
+            },
+        }
+        status, value = self.upstream.create_governed_task(body)
+        if status not in {200, 201, 202}:
+            if status == 409:
+                recovered = self._recover_existing(request, row)
+                if recovered is not None:
+                    return recovered
+            raise IngressError(
+                "GOVERNED_TASK_CREATE_FAILED",
+                409 if status == 409 else 503,
+            )
+        job_id = str(value.get("job_id") or "")
+        if value.get("task_id") != task_id or JOB_ID_RE.fullmatch(job_id) is None:
+            raise IngressError("GOVERNED_TASK_RESPONSE_INVALID", 502)
+
+        envelope_status, envelope = self.upstream.execution_envelope(
+            {
+                "task_id": task_id,
+                "job_id": job_id,
+                "required_capabilities": list(READONLY_CAPABILITIES),
+                "execution": body["execution"],
+            }
+        )
+        if envelope_status != 200:
+            raise IngressError("EXECUTION_PROJECTION_FAILED", 503)
+        if (
+            envelope.get("task_id") != task_id
+            or envelope.get("job_id") != job_id
+            or envelope.get("base_oid") != base_oid
+            or envelope.get("action") != "source_capture"
+            or sorted(envelope.get("capability_scope") or [])
+            != sorted(READONLY_CAPABILITIES)
+            or (envelope.get("governance") or {}).get("writer_lease") is not False
+        ):
+            raise IngressError("EXECUTION_PROJECTION_INVALID", 502)
+
+        response = self._public_response(
+            request["request_id"],
+            task_id,
+            job_id,
+            "BOUND",
+        )
+        self.store.bound(
+            request["request_id"],
+            job_id=job_id,
+            base_oid=base_oid,
+            response=response,
+        )
+        return response
+
+    def status(self, request_id: str) -> dict[str, Any]:
+        if REQUEST_ID_RE.fullmatch(request_id) is None:
+            raise IngressError("REQUEST_ID_INVALID")
+        row = self.store.get(request_id)
+        if row is None:
+            raise IngressError("REQUEST_NOT_FOUND", 404)
+        task_id = str(row["task_id"])
+        status, task = self.upstream.get_task(task_id)
+        if status == 200 and not row["job_id"]:
+            metadata = task.get("metadata")
+            execution = (
+                metadata.get("execution")
+                if isinstance(metadata, Mapping)
+                and isinstance(metadata.get("execution"), Mapping)
+                else {}
+            )
+            job_id = str(execution.get("job_id") or "")
+            base_oid = str(execution.get("base_oid") or "")
+            if (
+                JOB_ID_RE.fullmatch(job_id)
+                and OID_RE.fullmatch(base_oid)
+                and execution.get("action") == "source_capture"
+                and sorted(execution.get("capability_scope") or [])
+                == sorted(READONLY_CAPABILITIES)
+                and (execution.get("governance") or {}).get("writer_lease") is False
+            ):
+                recovered = self._public_response(
+                    request_id,
+                    task_id,
+                    job_id,
+                    str(task.get("status") or "BOUND"),
+                )
+                self.store.bound(
+                    request_id,
+                    job_id=job_id,
+                    base_oid=base_oid,
+                    response=recovered,
+                )
+                row = self.store.get(request_id) or row
+        if status == 404:
+            return {
+                "schema": RESPONSE_SCHEMA,
+                "request_id": request_id,
+                "task_id": task_id,
+                "state": "PENDING",
+                "official_verification_required": True,
+            }
+        if status != 200:
+            raise IngressError("TASK_STATUS_UNAVAILABLE", 503)
+
+        result = task.get("result") if isinstance(task.get("result"), Mapping) else {}
+        verification = (
+            task.get("verification")
+            if isinstance(task.get("verification"), Mapping)
+            else {}
+        )
+        safe = {
+            "schema": RESPONSE_SCHEMA,
+            "request_id": request_id,
+            "task_id": task_id,
+            "job_id": str(row["job_id"] or ""),
+            "state": str(task.get("status") or "UNKNOWN"),
+            "attempt": int(task.get("attempt") or 0),
+            "claimed": bool(task.get("claimed_by")),
+            "result_status": result.get("status"),
+            "result_classification": result.get("classification"),
+            "official_verification": verification.get("decision"),
+            "official_classification": verification.get("classification"),
+            "write_access": False,
+        }
+        if result.get("execution_mode") == "source_capture":
+            snapshot = result.get("source_snapshot")
+            if isinstance(snapshot, Mapping):
+                safe["source_snapshot_sha256"] = snapshot.get("payload_sha256")
+                safe["source_snapshot_file_count"] = snapshot.get("file_count")
+        return safe
+
+
+def openapi_document() -> dict[str, Any]:
+    return {
+        "openapi": OPENAPI_VERSION,
+        "info": {
+            "title": "AIOS Read-Only Online Task Ingress",
+            "version": "1.0.0",
+        },
+        "paths": {
+            "/v1/read-only-tasks": {
+                "post": {
+                    "operationId": "createReadOnlyOnlineTask",
+                    "security": [{"bearerAuth": []}],
+                    "requestBody": {"required": True},
+                    "responses": {"200": {"description": "Task accepted"}},
+                }
+            },
+            "/v1/read-only-tasks/{request_id}": {
+                "get": {
+                    "operationId": "getReadOnlyOnlineTaskStatus",
+                    "security": [{"bearerAuth": []}],
+                    "responses": {"200": {"description": "Sanitized task status"}},
+                }
+            },
+        },
+        "components": {
+            "securitySchemes": {
+                "bearerAuth": {
+                    "type": "http",
+                    "scheme": "bearer",
+                }
+            }
+        },
+    }
+
+
+class Handler(BaseHTTPRequestHandler):
+    service: IngressService | None = None
+    ingress_token: str = ""
+
+    def log_message(self, *_args: Any) -> None:
+        return
+
+    def respond(self, value: Mapping[str, Any], status: int) -> None:
+        raw = canonical(value)
+        self.send_response(status)
+        self.send_header("Content-Type", "application/json")
+        self.send_header("Cache-Control", "no-store")
+        self.send_header("Content-Length", str(len(raw)))
+        self.end_headers()
+        self.wfile.write(raw)
+
+    def authorized(self) -> bool:
+        header = self.headers.get("Authorization", "")
+        return (
+            header.startswith("Bearer ")
+            and bool(self.ingress_token)
+            and hmac.compare_digest(header[7:], self.ingress_token)
+        )
+
+    def body(self) -> dict[str, Any]:
+        try:
+            length = int(self.headers.get("Content-Length", "-1"))
+        except ValueError as exc:
+            raise IngressError("CONTENT_LENGTH_INVALID") from exc
+        if length < 1 or length > MAX_BODY:
+            raise IngressError("REQUEST_SIZE_INVALID", 413)
+        try:
+            value = json.loads(self.rfile.read(length).decode("utf-8"))
+        except Exception as exc:
+            raise IngressError("REQUEST_JSON_INVALID") from exc
+        if not isinstance(value, Mapping):
+            raise IngressError("REQUEST_OBJECT_REQUIRED")
+        return dict(value)
+
+    def do_GET(self) -> None:
+        if self.path == "/health":
+            self.respond(
+                {
+                    "status": "ok",
+                    "service": "aios-gpt-online-task-ingress",
+                    "mode": "read-only",
+                },
+                200,
+            )
+            return
+        if self.path == "/openapi.json":
+            self.respond(openapi_document(), 200)
+            return
+        if not self.authorized():
+            self.respond({"error": "UNAUTHORIZED"}, 401)
+            return
+        match = re.fullmatch(r"/v1/read-only-tasks/([^/?#]+)", self.path)
+        if match is None:
+            self.respond({"error": "NOT_FOUND"}, 404)
+            return
+        try:
+            if self.service is None:
+                raise IngressError("RUNTIME_UNAVAILABLE", 503)
+            self.respond(self.service.status(match.group(1)), 200)
+        except IngressError as exc:
+            self.respond({"error": exc.code}, exc.status)
+
+    def do_POST(self) -> None:
+        if self.path != "/v1/read-only-tasks":
+            self.respond({"error": "NOT_FOUND"}, 404)
+            return
+        if not self.authorized():
+            self.respond({"error": "UNAUTHORIZED"}, 401)
+            return
+        try:
+            if self.service is None:
+                raise IngressError("RUNTIME_UNAVAILABLE", 503)
+            value = self.service.create(self.body())
+            self.respond(value, 200)
+        except IngressError as exc:
+            self.respond({"error": exc.code}, exc.status)
+
+
+def serve() -> None:
+    ingress_token_path = Path(
+        os.environ.get(
+            "AIOS_GPT_INGRESS_TOKEN_FILE",
+            str(
+                Path(os.environ.get("CREDENTIALS_DIRECTORY", "/run/credentials"))
+                / "gpt_ingress_token"
+            ),
+        )
+    )
+    control_token_path = Path(
+        os.environ.get(
+            "AIOS_AGENT_TASK_CONTROL_TOKEN_FILE",
+            str(
+                Path(os.environ.get("CREDENTIALS_DIRECTORY", "/run/credentials"))
+                / "agent_task_control_token"
+            ),
+        )
+    )
+    ingress_token = load_secret(
+        ingress_token_path,
+        code="GPT_INGRESS_CREDENTIAL",
+    )
+    control_token = load_secret(
+        control_token_path,
+        code="CONTROL_PLANE_CREDENTIAL",
+    )
+
+    state_root = Path(
+        os.environ.get(
+            "AIOS_GPT_INGRESS_STATE_ROOT",
+            "/var/lib/aios-gpt-online-task-ingress",
+        )
+    )
+    host = validated_bind_host(
+        os.environ.get("AIOS_GPT_INGRESS_HOST", "127.0.0.1")
+    )
+    port = int(os.environ.get("AIOS_GPT_INGRESS_PORT", "18945"))
+    if not 1024 <= port <= 65535:
+        raise RuntimeError("GPT_INGRESS_PORT_INVALID")
+
+    Handler.service = IngressService(
+        Upstream(
+            bridge_base=os.environ.get(
+                "AIOS_GOVERNED_WRITE_API_URL",
+                "http://127.0.0.1:18892",
+            ),
+            task_base=os.environ.get(
+                "AIOS_AGENT_TASK_API_URL",
+                "http://127.0.0.1:18940",
+            ),
+            control_token=control_token,
+        ),
+        Store(state_root / "state.sqlite3"),
+    )
+    Handler.ingress_token = ingress_token
+    ThreadingHTTPServer((host, port), Handler).serve_forever()
+
+
+if __name__ == "__main__":
+    serve()

active-runtime/app/aios_model_governance_admin.py

@@ -0,0 +1,232 @@
+#!/usr/bin/env python3
+"""Root-owned Model Governance configuration reconciler.
+
+The config contains provider/model/policy metadata only. Raw credentials are rejected by
+aios_model_governance and never stored in the governance database.
+"""
+from __future__ import annotations
+
+import argparse
+import hashlib
+import json
+from pathlib import Path
+from typing import Any, Mapping
+
+from aios_model_governance import (
+    ModelGovernanceError,
+    ModelGovernanceStore,
+    ModelRecord,
+    ProviderRecord,
+    ScopePolicy,
+)
+
+SCHEMA = "aios.model-governance.bootstrap.v1"
+MAX_CONFIG_BYTES = 512 * 1024
+
+
+def canonical(value: object) -> bytes:
+    return json.dumps(
+        value, ensure_ascii=False, sort_keys=True, separators=(",", ":")
+    ).encode("utf-8")
+
+
+def fail(code: str) -> None:
+    raise SystemExit(code)
+
+
+def _load(path: Path) -> dict[str, Any]:
+    if path.is_symlink() or not path.is_file() or path.stat().st_size > MAX_CONFIG_BYTES:
+        fail("MODEL_GOVERNANCE_CONFIG_INVALID")
+    try:
+        value = json.loads(path.read_text(encoding="utf-8"))
+    except json.JSONDecodeError:
+        fail("MODEL_GOVERNANCE_CONFIG_INVALID")
+    if not isinstance(value, Mapping) or value.get("schema") != SCHEMA:
+        fail("MODEL_GOVERNANCE_CONFIG_SCHEMA_INVALID")
+    return dict(value)
+
+
+def _provider_desired(value: Mapping[str, Any]) -> ProviderRecord:
+    return ProviderRecord(
+        provider_id=str(value.get("provider_id") or ""),
+        credential_ref=value.get("credential_ref"),
+        enabled=bool(value.get("enabled", True)),
+        failure_domain=str(value.get("failure_domain") or "default"),
+        max_failures=int(value.get("max_failures") or 3),
+        cooldown_seconds=int(value.get("cooldown_seconds") or 60),
+    )
+
+
+def _model_desired(value: Mapping[str, Any]) -> ModelRecord:
+    return ModelRecord(
+        model_id=str(value.get("model_id") or ""),
+        provider_id=str(value.get("provider_id") or ""),
+        version=str(value.get("version") or ""),
+        capabilities=tuple(value.get("capabilities") or ()),
+        input_cost_micros_per_1k=int(value.get("input_cost_micros_per_1k") or 0),
+        output_cost_micros_per_1k=int(value.get("output_cost_micros_per_1k") or 0),
+        max_input_tokens=int(value.get("max_input_tokens") or 0),
+        max_output_tokens=int(value.get("max_output_tokens") or 0),
+        evaluation_state=str(value.get("evaluation_state") or "APPROVED"),
+        evaluation_sha256=str(value.get("evaluation_sha256") or ""),
+        enabled=bool(value.get("enabled", True)),
+        priority=int(value.get("priority") or 100),
+    )
+
+
+def _policy_desired(value: Mapping[str, Any]) -> ScopePolicy:
+    return ScopePolicy(
+        scope_type=str(value.get("scope_type") or ""),
+        tenant_id=value.get("tenant_id"),
+        project_id=value.get("project_id"),
+        allowed_models=tuple(value.get("allowed_models") or ()),
+        denied_providers=tuple(value.get("denied_providers") or ()),
+        required_capabilities=tuple(value.get("required_capabilities") or ()),
+        pinned_model_id=value.get("pinned_model_id"),
+        pinned_model_version=value.get("pinned_model_version"),
+        max_input_tokens_per_request=int(value.get("max_input_tokens_per_request") or 0),
+        max_output_tokens_per_request=int(value.get("max_output_tokens_per_request") or 0),
+        daily_token_budget=int(value.get("daily_token_budget") or 0),
+        daily_cost_budget_micros=int(value.get("daily_cost_budget_micros") or 0),
+        fallback_limit=int(value.get("fallback_limit") or 0),
+    )
+
+
+def _row_dict(row: object) -> dict[str, Any] | None:
+    return dict(row) if row is not None else None
+
+
+def _provider_matches(row: Mapping[str, Any], desired: ProviderRecord) -> bool:
+    return (
+        row["credential_ref"] == desired.credential_ref
+        and bool(row["enabled"]) == desired.enabled
+        and row["failure_domain"] == desired.failure_domain
+        and int(row["max_failures"]) == desired.max_failures
+        and int(row["cooldown_seconds"]) == desired.cooldown_seconds
+    )
+
+
+def _model_matches(row: Mapping[str, Any], desired: ModelRecord) -> bool:
+    return (
+        row["provider_id"] == desired.provider_id
+        and row["version"] == desired.version
+        and tuple(json.loads(row["capabilities_json"])) == tuple(sorted(set(desired.capabilities)))
+        and int(row["input_cost_micros_per_1k"]) == desired.input_cost_micros_per_1k
+        and int(row["output_cost_micros_per_1k"]) == desired.output_cost_micros_per_1k
+        and int(row["max_input_tokens"]) == desired.max_input_tokens
+        and int(row["max_output_tokens"]) == desired.max_output_tokens
+        and row["evaluation_state"] == desired.evaluation_state
+        and row["evaluation_sha256"] == desired.evaluation_sha256
+        and bool(row["enabled"]) == desired.enabled
+        and int(row["priority"]) == desired.priority
+    )
+
+
+def _policy_matches(row: Mapping[str, Any], desired: ScopePolicy) -> bool:
+    return (
+        row["scope_type"] == desired.scope_type
+        and row["tenant_id"] == desired.tenant_id
+        and row["project_id"] == desired.project_id
+        and tuple(json.loads(row["allowed_models_json"])) == tuple(sorted(set(desired.allowed_models)))
+        and tuple(json.loads(row["denied_providers_json"])) == tuple(sorted(set(desired.denied_providers)))
+        and tuple(json.loads(row["required_capabilities_json"])) == tuple(sorted(set(desired.required_capabilities)))
+        and row["pinned_model_id"] == desired.pinned_model_id
+        and row["pinned_model_version"] == desired.pinned_model_version
+        and int(row["max_input_tokens_per_request"]) == desired.max_input_tokens_per_request
+        and int(row["max_output_tokens_per_request"]) == desired.max_output_tokens_per_request
+        and int(row["daily_token_budget"]) == desired.daily_token_budget
+        and int(row["daily_cost_budget_micros"]) == desired.daily_cost_budget_micros
+        and int(row["fallback_limit"]) == desired.fallback_limit
+    )
+
+
+def apply_config(db_path: Path, config_path: Path) -> dict[str, Any]:
+    config = _load(config_path)
+    config_sha = hashlib.sha256(canonical(config)).hexdigest()
+    providers = config.get("providers") or []
+    models = config.get("models") or []
+    policies = config.get("policies") or []
+    if not all(isinstance(value, list) for value in (providers, models, policies)):
+        fail("MODEL_GOVERNANCE_CONFIG_COLLECTION_INVALID")
+
+    store = ModelGovernanceStore(db_path)
+    changes = {"providers": 0, "models": 0, "policies": 0, "unchanged": 0}
+    try:
+        for item in providers:
+            if not isinstance(item, Mapping):
+                fail("MODEL_GOVERNANCE_PROVIDER_CONFIG_INVALID")
+            desired = _provider_desired(item)
+            row = store.conn.execute(
+                "SELECT * FROM model_governance_providers WHERE provider_id=?",
+                (desired.provider_id,),
+            ).fetchone()
+            if row is not None and _provider_matches(row, desired):
+                changes["unchanged"] += 1
+                continue
+            expected = int(row["revision"]) if row is not None else 0
+            store.register_provider(desired, expected_revision=expected)
+            changes["providers"] += 1
+
+        for item in models:
+            if not isinstance(item, Mapping):
+                fail("MODEL_GOVERNANCE_MODEL_CONFIG_INVALID")
+            desired = _model_desired(item)
+            row = store.conn.execute(
+                "SELECT * FROM model_governance_models WHERE model_id=?",
+                (desired.model_id,),
+            ).fetchone()
+            if row is not None and _model_matches(row, desired):
+                changes["unchanged"] += 1
+                continue
+            expected = int(row["revision"]) if row is not None else 0
+            store.register_model(desired, expected_revision=expected)
+            changes["models"] += 1
+
+        for item in policies:
+            if not isinstance(item, Mapping):
+                fail("MODEL_GOVERNANCE_POLICY_CONFIG_INVALID")
+            desired = _policy_desired(item)
+            row = store.conn.execute(
+                "SELECT * FROM model_governance_scope_policies WHERE scope_key=?",
+                (desired.scope_key,),
+            ).fetchone()
+            if row is not None and _policy_matches(row, desired):
+                changes["unchanged"] += 1
+                continue
+            expected = int(row["revision"]) if row is not None else 0
+            store.set_scope_policy(desired, expected_revision=expected)
+            changes["policies"] += 1
+
+        if not store.verify_audit_chain():
+            fail("MODEL_GOVERNANCE_AUDIT_CHAIN_INVALID")
+        return {
+            "schema": "aios.model-governance.bootstrap-result.v1",
+            "database": str(db_path),
+            "config_sha256": config_sha,
+            "changes": changes,
+            "audit_chain_valid": True,
+        }
+    finally:
+        store.close()
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--db", required=True)
+    parser.add_argument("--config", required=True)
+    parser.add_argument("--output")
+    args = parser.parse_args()
+    try:
+        result = apply_config(Path(args.db), Path(args.config))
+    except ModelGovernanceError as exc:
+        print(json.dumps({"status": "FAILED", "error": exc.code}, sort_keys=True))
+        return 20
+    data = json.dumps(result, indent=2, sort_keys=True) + "\n"
+    if args.output:
+        Path(args.output).write_text(data, encoding="utf-8")
+    print(data, end="")
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

active-runtime/app/aios_online_agent_checkpoint.py

@@ -17,6 +17,7 @@ PHASE_ORDER = {
     "claimed": 10,
     "workspace_prepared": 20,
     "verification_passed": 30,
+    "agent_runtime_done": 40,
     "source_capture_done": 40,
     "commit_done": 40,
     "reconciliation_required": 45,

active-runtime/app/aios_online_agent_execution.py

@@ -179,7 +179,10 @@ def _agent_runtime_input(value: object) -> dict[str, Any]:
     if not isinstance(value, Mapping):
         _error("AGENT_RUNTIME_INPUT_REQUIRED")
     raw = dict(value)
-    allowed = {"goal", "instructions", "success_criteria", "constraints"}
+    allowed = {
+        "goal", "instructions", "success_criteria", "constraints",
+        "scope", "trace_id",
+    }
     if set(raw) - allowed:
         _error("AGENT_RUNTIME_INPUT_FIELDS_INVALID")
     goal = _string(raw.get("goal"), "AGENT_RUNTIME_GOAL_REQUIRED")
@@ -208,11 +211,34 @@ def _agent_runtime_input(value: object) -> dict[str, Any]:
         lowered = str(key).lower()
         if any(marker in lowered for marker in sensitive_markers):
             _error("AGENT_RUNTIME_SECRET_FIELD_DENY")
+
+    scope = raw.get("scope", {})
+    if not isinstance(scope, Mapping):
+        _error("AGENT_RUNTIME_SCOPE_INVALID")
+    scope = {
+        "scope_type": str(scope.get("scope_type") or "legacy"),
+        "tenant_id": scope.get("tenant_id"),
+        "project_id": scope.get("project_id"),
+    }
+    if scope["scope_type"] not in {"legacy", "system", "tenant"}:
+        _error("AGENT_RUNTIME_SCOPE_INVALID")
+    if scope["scope_type"] == "tenant":
+        if not isinstance(scope["tenant_id"], str) or not scope["tenant_id"].strip():
+            _error("AGENT_RUNTIME_SCOPE_INVALID")
+    elif scope["tenant_id"] not in (None, "") or scope["project_id"] not in (None, ""):
+        _error("AGENT_RUNTIME_SCOPE_INVALID")
+
+    trace_id = str(raw.get("trace_id") or "").strip()
+    if not trace_id or len(trace_id) > 255:
+        _error("AGENT_RUNTIME_TRACE_ID_INVALID")
+
     normalized = {
         "goal": goal[:12000],
         "instructions": instructions[:32000],
         "success_criteria": normalized_criteria,
         "constraints": dict(constraints),
+        "scope": scope,
+        "trace_id": trace_id,
     }
     if len(_canonical(normalized).encode("utf-8")) > MAX_AGENT_RUNTIME_INPUT_BYTES:
         _error("AGENT_RUNTIME_INPUT_TOO_LARGE")
@@ -425,6 +451,12 @@ def project_online_task(
         if adapter not in AGENT_RUNTIME_ADAPTERS:
             _error("AGENT_RUNTIME_ADAPTER_INVALID")
         runtime_input = execution_value.get("runtime_input")
+        tenant_id = request.get("tenant_id")
+        project_id = request.get("project_id")
+        declared_scope = str(
+            request.get("scope_type")
+            or ("tenant" if isinstance(tenant_id, str) and tenant_id.strip() else "legacy")
+        )
         if runtime_input is None:
             runtime_input = {
                 "goal": request.get("goal"),
@@ -432,6 +464,20 @@ def project_online_task(
                 "success_criteria": request.get("success_criteria") or [],
                 "constraints": request.get("constraints") or {},
             }
+        elif not isinstance(runtime_input, Mapping):
+            _error("AGENT_RUNTIME_INPUT_REQUIRED")
+        runtime_input = dict(runtime_input)
+        # Scope/trace are authoritative Task identity, never caller-selected runtime hints.
+        runtime_input["scope"] = {
+            "scope_type": declared_scope,
+            "tenant_id": tenant_id,
+            "project_id": project_id,
+        }
+        runtime_input["trace_id"] = str(
+            request.get("trace_id")
+            or request.get("run_id")
+            or binding.task_id
+        )
         execution_value["runtime_input"] = runtime_input
     execution = project_execution_envelope(
         binding,

active-runtime/app/aios_online_agent_worker.py

@@ -236,6 +236,64 @@ def atomic_json(path: Path, value: Mapping[str, Any], mode: int = 0o640) -> None
             os.unlink(temporary)
 
 
+
+WORKER_LIVENESS_PATH = Path(
+    os.environ.get(
+        "AIOS_WORKER_LIVENESS_PATH",
+        str(STATE_ROOT / "worker-liveness.json"),
+    )
+)
+_LIVENESS_LOCK = threading.Lock()
+_LIVENESS_SEQUENCE = 0
+ENGINEERING_WRITE_CAPABILITIES = frozenset({
+    "governed.write.branch",
+    "governed.write.file",
+})
+
+
+def write_worker_liveness(
+    state: str,
+    *,
+    visible_count: int | None = None,
+    task_id: str | None = None,
+    stage: str | None = None,
+    error: str | None = None,
+) -> None:
+    """Write idle/busy liveness without touching Task/Job state."""
+    global _LIVENESS_SEQUENCE
+    with _LIVENESS_LOCK:
+        _LIVENESS_SEQUENCE += 1
+        epoch = time.time()
+        payload = {
+            "schema": "aios.online-agent.worker-liveness.v1",
+            "agent_id": AGENT_ID,
+            "agent_type": AGENT_TYPE,
+            "agent_version": AGENT_VERSION,
+            "pid": os.getpid(),
+            "state": state,
+            "sequence": _LIVENESS_SEQUENCE,
+            "last_seen_epoch": epoch,
+            "last_seen_at": datetime.fromtimestamp(epoch, timezone.utc).isoformat(),
+            "poll_seconds": POLL_SECONDS,
+            "heartbeat_seconds": HEARTBEAT_SECONDS,
+            "api_base": API_BASE,
+            "bridge_base": BRIDGE_BASE,
+            "visible_count": visible_count,
+            "task_id": task_id,
+            "stage": stage,
+            "error": error,
+        }
+        atomic_json(WORKER_LIVENESS_PATH, payload)
+
+
+def source_capture_requires_deterministic_plan(envelope: "ExecutionEnvelope") -> bool:
+    """Treat write-capable source_capture as a binding placeholder, never executable work."""
+    return (
+        envelope.action == "source_capture"
+        and bool(ENGINEERING_WRITE_CAPABILITIES.intersection(envelope.capability_scope))
+    )
+
+
 def safe_relative_path(value: Any, *, field: str = "path") -> str:
     if not isinstance(value, str) or not value or len(value) > 240:
         raise WorkerError(f"{field.upper()}_INVALID", phase="validation")
@@ -525,11 +583,44 @@ class ExecutionEnvelope:
         )
 
 
+def external_agent_request_id(
+    envelope: ExecutionEnvelope,
+    runtime_name: str,
+    attempt: int,
+) -> str:
+    if not isinstance(attempt, int) or isinstance(attempt, bool) or attempt < 1:
+        raise WorkerError("TASK_ATTEMPT_INVALID", phase="external_agent")
+    if runtime_name not in {"hermes", "openclaw"}:
+        raise WorkerError(
+            "EXTERNAL_AGENT_RUNTIME_UNSUPPORTED",
+            phase="external_agent",
+            details={"runtime": runtime_name},
+        )
+    return (
+        f"external-{runtime_name}-"
+        + digest_json(
+            {
+                "task_id": envelope.task_id,
+                "job_id": envelope.job_id,
+                "execution_digest": envelope.digest,
+                "runtime": runtime_name,
+                "attempt": attempt,
+            }
+        )[:32]
+    )
+
+
 class RuntimeAdapter(ABC):
     name: str
 
     @abstractmethod
-    def execute(self, envelope: ExecutionEnvelope, worktree: Path) -> dict[str, Any]:
+    def execute(
+        self,
+        envelope: ExecutionEnvelope,
+        worktree: Path,
+        *,
+        attempt: int,
+    ) -> dict[str, Any]:
         raise NotImplementedError
 
 
@@ -550,7 +641,13 @@ class DeterministicExecutor(RuntimeAdapter):
             if os.path.exists(temporary):
                 os.unlink(temporary)
 
-    def execute(self, envelope: ExecutionEnvelope, worktree: Path) -> dict[str, Any]:
+    def execute(
+        self,
+        envelope: ExecutionEnvelope,
+        worktree: Path,
+        *,
+        attempt: int,
+    ) -> dict[str, Any]:
         changed: set[str] = set()
         for change in envelope.change_set:
             relative = str(change["path"])
@@ -605,7 +702,13 @@ class SourceCaptureExecutor(RuntimeAdapter):
         "governed", "trace", "tool", "credential", "scope", "message", "handoff",
     }
 
-    def execute(self, envelope: ExecutionEnvelope, worktree: Path) -> dict[str, Any]:
+    def execute(
+        self,
+        envelope: ExecutionEnvelope,
+        worktree: Path,
+        *,
+        attempt: int,
+    ) -> dict[str, Any]:
         terms = set(self.defaults)
         for signal in envelope.source_signals:
             terms.update(token.lower() for token in re.findall(r"[A-Za-z][A-Za-z0-9_.-]{2,63}", signal))
@@ -674,7 +777,13 @@ class CompactInterfaceMapExecutor(RuntimeAdapter):
 
     name = "compact_interface_map"
 
-    def execute(self, envelope: ExecutionEnvelope, worktree: Path) -> dict[str, Any]:
+    def execute(
+        self,
+        envelope: ExecutionEnvelope,
+        worktree: Path,
+        *,
+        attempt: int,
+    ) -> dict[str, Any]:
         if envelope.action != "source_capture":
             raise WorkerError("COMPACT_INTERFACE_MAP_ACTION_INVALID", phase="interface_map")
         try:
@@ -704,7 +813,13 @@ class CompactInterfaceMapExecutor(RuntimeAdapter):
 class CodexExecutor(RuntimeAdapter):
     name = "codex"
 
-    def execute(self, envelope: ExecutionEnvelope, worktree: Path) -> dict[str, Any]:
+    def execute(
+        self,
+        envelope: ExecutionEnvelope,
+        worktree: Path,
+        *,
+        attempt: int,
+    ) -> dict[str, Any]:
         raise WorkerError(
             "CODEX_ADAPTER_REQUIRES_EXPLICIT_POLICY",
             phase="adapter",
@@ -715,7 +830,13 @@ class CodexExecutor(RuntimeAdapter):
 class ExternalAgentExecutor(RuntimeAdapter):
     runtime_name: str
 
-    def execute(self, envelope: ExecutionEnvelope, worktree: Path) -> dict[str, Any]:
+    def execute(
+        self,
+        envelope: ExecutionEnvelope,
+        worktree: Path,
+        *,
+        attempt: int,
+    ) -> dict[str, Any]:
         if envelope.action != "agent_runtime":
             raise WorkerError(
                 "EXTERNAL_AGENT_ACTION_INVALID",
@@ -729,16 +850,10 @@ class ExternalAgentExecutor(RuntimeAdapter):
             )
         if not envelope.runtime_input:
             raise WorkerError("EXTERNAL_AGENT_RUNTIME_INPUT_MISSING", phase="adapter")
-        request_id = (
-            f"external-{self.runtime_name}-"
-            + digest_json(
-                {
-                    "task_id": envelope.task_id,
-                    "job_id": envelope.job_id,
-                    "execution_digest": envelope.digest,
-                    "runtime": self.runtime_name,
-                }
-            )[:32]
+        request_id = external_agent_request_id(
+            envelope,
+            self.runtime_name,
+            attempt,
         )
         exchange = ExternalAgentExchange(EXTERNAL_AGENT_EXECUTOR_ROOT)
         try:
@@ -746,6 +861,7 @@ class ExternalAgentExecutor(RuntimeAdapter):
                 request_id=request_id,
                 runtime=self.runtime_name,
                 task_id=envelope.task_id,
+                attempt=attempt,
                 execution_digest=envelope.digest,
                 payload=envelope.runtime_input,
                 timeout_seconds=EXTERNAL_AGENT_EXECUTOR_TIMEOUT_SECONDS,
@@ -761,6 +877,7 @@ class ExternalAgentExecutor(RuntimeAdapter):
             task_id=envelope.task_id,
             adapter=self.runtime_name,
             request_id=request_id,
+            attempt=attempt,
             output_sha256=external.get("output_sha256"),
             duration_ms=external.get("duration_ms"),
             credential_material_exposed=False,
@@ -890,6 +1007,11 @@ class LeaseHeartbeat:
             self.failed = str(response.get("error") or status)
             raise WorkerError("TASK_HEARTBEAT_FAILED", phase="heartbeat", details={"api_status": status, "error": self.failed})
         audit("TASK_HEARTBEAT", task_id=self.task_id, stage=self.stage, api_status=status)
+        write_worker_liveness(
+            "BUSY",
+            task_id=self.task_id,
+            stage=self.stage,
+        )
 
     def start(self) -> None:
         self.thread.start()
@@ -909,6 +1031,11 @@ class LeaseHeartbeat:
                     audit("TASK_HEARTBEAT_FAILED", task_id=self.task_id, stage=stage, api_status=status, error=self.failed)
                 else:
                     audit("TASK_HEARTBEAT", task_id=self.task_id, stage=stage, api_status=status)
+                    write_worker_liveness(
+                        "BUSY",
+                        task_id=self.task_id,
+                        stage=stage,
+                    )
             except Exception as exc:  # pragma: no cover - defensive runtime guard
                 self.failed = type(exc).__name__
                 audit("TASK_HEARTBEAT_FAILED", task_id=self.task_id, stage=stage, error=self.failed)
@@ -1286,7 +1413,13 @@ class IsolatedDeterministicExecutor(RuntimeAdapter):
             if os.path.exists(temporary):
                 os.unlink(temporary)
 
-    def execute(self, envelope: ExecutionEnvelope, worktree: Path) -> dict[str, Any]:
+    def execute(
+        self,
+        envelope: ExecutionEnvelope,
+        worktree: Path,
+        *,
+        attempt: int,
+    ) -> dict[str, Any]:
         if envelope.action != "deterministic":
             raise WorkerError("ISOLATED_EXECUTOR_ACTION_INVALID", phase="adapter")
         root = ISOLATED_EXECUTOR_ROOT
@@ -1772,12 +1905,25 @@ class WorkerSupervisor:
             if not isinstance(execution, Mapping):
                 raise WorkerError("EXECUTION_ENVELOPE_SCHEMA_VALIDATION", phase="validation")
             persisted = ExecutionEnvelope.from_task(task, worker_capabilities=WORKER_CAPABILITIES)
-            # Deterministic tasks must re-resolve through the plan registry.  A
-            # read-only source-capture task has no plan to replace its payload: replay
-            # the already validated durable envelope so source signals remain part of
-            # the canonical digest and the worker does not invent a different capture.
-            projection_input = execution if persisted.action == "source_capture" else None
+            # Read-only source_capture/agent_runtime tasks replay their persisted
+            # payload. A write-capable source_capture is only a Bridge-created
+            # engineering placeholder: never pass that stale payload back to the
+            # resolver, and never claim it until a deterministic plan exists.
+            pending_plan = source_capture_requires_deterministic_plan(persisted)
+            projection_input = (
+                execution
+                if persisted.action in {"source_capture", "agent_runtime"} and not pending_plan
+                else None
+            )
             status, authoritative = self.bridge.project_execution(task, projection_input)
+            if pending_plan and (
+                status == 404
+                or authoritative.get("error") == "TASK_EXECUTION_PROJECTION_MISSING"
+            ):
+                raise WorkerError(
+                    "TASK_SKIPPED_PENDING_DETERMINISTIC_PLAN",
+                    phase="projection",
+                )
             if not 200 <= status < 300:
                 raise WorkerError("TASK_PERSISTED_BINDING_UNVERIFIED", phase="projection", details={"api_status": status, "error": authoritative.get("error")})
             canonical = ExecutionEnvelope.from_task(task, authoritative, WORKER_CAPABILITIES)
@@ -1806,6 +1952,11 @@ class WorkerSupervisor:
                     execution_digest=canonical.digest,
                 )
                 return canonical, "plan_registry_verified"
+            if pending_plan:
+                raise WorkerError(
+                    "TASK_SKIPPED_PENDING_DETERMINISTIC_PLAN",
+                    phase="projection",
+                )
             return persisted, "persisted_verified"
         status, projected = self.bridge.project_execution(task)
         if status == 404 or projected.get("error") == "TASK_EXECUTION_PROJECTION_MISSING":
@@ -2107,7 +2258,11 @@ class WorkerSupervisor:
                     save_task_state(task_id, state)
                     save_checkpoint(task_id, state, envelope, attempt)
                     adapter = adapter_for(envelope.adapter)
-                    execution = adapter.execute(envelope, worktree)
+                    execution = adapter.execute(
+                        envelope,
+                        worktree,
+                        attempt=attempt,
+                    )
                     state.update(
                         {
                             "phase": "verification_passed",
@@ -2290,8 +2445,13 @@ class WorkerSupervisor:
     def once(self) -> int:
         try:
             tasks = self.agent.poll()
+            write_worker_liveness("POLL_OK", visible_count=len(tasks))
             audit("TASK_POLL", visible_count=len(tasks))
         except WorkerError as exc:
+            write_worker_liveness(
+                "POLL_FAILED",
+                error=exc.code,
+            )
             audit("TASK_POLL_FAILED", error=exc.code, phase=exc.phase, details=exc.details)
             return 10
         dynamic_tasks: list[tuple[dict[str, Any], ExecutionEnvelope]] = []
@@ -2300,7 +2460,11 @@ class WorkerSupervisor:
                 envelope, source = self.resolve_envelope(task)
             except WorkerError as exc:
                 event = "TASK_SKIPPED_ENVELOPE_INVALID"
-                if exc.code in {"TASK_SKIPPED_PROJECTION_MISSING", "TASK_SKIPPED_POLICY_DENY"}:
+                if exc.code in {
+                    "TASK_SKIPPED_PROJECTION_MISSING",
+                    "TASK_SKIPPED_POLICY_DENY",
+                    "TASK_SKIPPED_PENDING_DETERMINISTIC_PLAN",
+                }:
                     event = exc.code
                 elif exc.code == "CAPABILITY_MISMATCH_DENY":
                     event = "TASK_SKIPPED_CAPABILITY_MISMATCH"
@@ -2334,6 +2498,7 @@ class WorkerSupervisor:
         ensure_roots()
         recover_stale_workspaces()
         checkpoint_store()
+        write_worker_liveness("STARTING")
         audit("WORKER_STARTED", task_agnostic=True, api_base=API_BASE, bridge_base=BRIDGE_BASE, adapter_names=["isolated-deterministic", "source_capture", "codex", "hermes"], runtime_only_credentials=True, isolated_executor=True, durable_checkpoint=True)
         while True:
             self.once()

active-runtime/app/aios_worker_availability_guard.py

@@ -0,0 +1,462 @@
+"""Self-healing availability guard for persistent Online Agent workers.
+
+The guard never claims Tasks and never writes Task/Job/Git databases. It only:
+- ensures required worker/executor systemd units are enabled + active;
+- checks worker-owned liveness JSON written by the real poll/heartbeat loop;
+- restarts an active worker only when its liveness is stale;
+- honors an explicit maintenance inhibit file;
+- writes a bounded status JSON under /run for observability.
+
+Idle liveness is the successful poll loop itself. Busy liveness is refreshed by the
+Task lease-heartbeat thread, so long-running Tasks are not mistaken for hung workers.
+"""
+from __future__ import annotations
+
+from dataclasses import asdict, dataclass
+import argparse
+import json
+import os
+from pathlib import Path
+import shlex
+import subprocess
+import tempfile
+import time
+from typing import Any, Callable, Mapping, Sequence
+
+SCHEMA = "aios.worker-availability.status.v1"
+DEFAULT_WORKER_UNITS = (
+    "aios-online-agent-worker-20260815.service",
+    "aios-hermes-online-agent-worker.service",
+    "aios-openclaw-online-agent-worker.service",
+)
+DEFAULT_DEPENDENCY_UNITS = (
+    "aios-external-agent-executor.service",
+)
+DEFAULT_RUN_ROOT = Path("/run/aios-worker-availability")
+DEFAULT_STATUS = DEFAULT_RUN_ROOT / "status.json"
+DEFAULT_RECOVERY = DEFAULT_RUN_ROOT / "recovery.json"
+DEFAULT_INHIBIT = DEFAULT_RUN_ROOT / "maintenance.lock"
+DEFAULT_MAX_STALE_SECONDS = 90
+DEFAULT_RESTART_COOLDOWN_SECONDS = 60
+DEFAULT_STARTUP_WAIT_SECONDS = 10
+ACCEPTED_ENABLED_STATES = frozenset({"enabled", "enabled-runtime", "static", "indirect", "generated"})
+
+
+class AvailabilityGuardError(RuntimeError):
+    pass
+
+
+@dataclass(frozen=True)
+class UnitSnapshot:
+    unit: str
+    role: str
+    load_state: str
+    active_state: str
+    unit_file_state: str
+    enabled_ok: bool
+    liveness_path: str | None = None
+    liveness_state: str | None = None
+    last_seen_epoch: float | None = None
+    liveness_age_seconds: float | None = None
+    fresh: bool | None = None
+    action: str = "NONE"
+    detail: str | None = None
+
+
+class Systemd:
+    def __init__(
+        self,
+        runner: Callable[..., subprocess.CompletedProcess[str]] | None = None,
+    ) -> None:
+        self.runner = runner or subprocess.run
+
+    def _run(self, args: Sequence[str], *, check: bool = False) -> subprocess.CompletedProcess[str]:
+        result = self.runner(
+            ["/usr/bin/systemctl", *args],
+            text=True,
+            capture_output=True,
+            check=False,
+            timeout=20,
+        )
+        if check and result.returncode != 0:
+            raise AvailabilityGuardError(
+                "SYSTEMD_COMMAND_FAILED:"
+                + ":".join(args)
+                + ":"
+                + str(result.returncode)
+            )
+        return result
+
+    def property(self, unit: str, name: str) -> str:
+        result = self._run(["show", unit, f"--property={name}", "--value"])
+        if result.returncode != 0:
+            return ""
+        return result.stdout.strip()
+
+    def environment(self, unit: str) -> dict[str, str]:
+        text = self.property(unit, "Environment")
+        result: dict[str, str] = {}
+        for item in shlex.split(text):
+            if "=" not in item:
+                continue
+            key, value = item.split("=", 1)
+            result[key] = value
+        return result
+
+    def enable(self, unit: str) -> None:
+        self._run(["enable", unit], check=True)
+
+    def start(self, unit: str) -> None:
+        self._run(["start", unit], check=True)
+
+    def restart(self, unit: str) -> None:
+        self._run(["restart", unit], check=True)
+
+
+def _atomic_json(path: Path, value: Mapping[str, Any]) -> None:
+    path.parent.mkdir(parents=True, exist_ok=True, mode=0o755)
+    if path.parent.is_symlink() or path.is_symlink():
+        raise AvailabilityGuardError("STATUS_PATH_SYMLINK_REJECTED")
+    fd, temporary = tempfile.mkstemp(prefix=f".{path.name}.", dir=path.parent)
+    try:
+        os.fchmod(fd, 0o644)
+        with os.fdopen(fd, "w", encoding="utf-8") as handle:
+            json.dump(value, handle, ensure_ascii=False, sort_keys=True, indent=2)
+            handle.write("\n")
+            handle.flush()
+            os.fsync(handle.fileno())
+        os.replace(temporary, path)
+        os.chmod(path, 0o644)
+    finally:
+        if os.path.exists(temporary):
+            os.unlink(temporary)
+
+
+def _load_json(path: Path) -> dict[str, Any] | None:
+    if path.is_symlink() or not path.is_file():
+        return None
+    try:
+        value = json.loads(path.read_text(encoding="utf-8"))
+    except (OSError, json.JSONDecodeError):
+        return None
+    return dict(value) if isinstance(value, Mapping) else None
+
+
+def _state_root(environment: Mapping[str, str], unit: str) -> Path:
+    explicit = environment.get("AIOS_WORKER_STATE_ROOT")
+    if explicit:
+        return Path(explicit)
+    fallbacks = {
+        "aios-online-agent-worker-20260815.service": Path(
+            "/var/lib/aios-online-agent-worker-20260815"
+        ),
+        "aios-hermes-online-agent-worker.service": Path(
+            "/var/lib/aios-online-agent-hermes"
+        ),
+        "aios-openclaw-online-agent-worker.service": Path(
+            "/var/lib/aios-online-agent-openclaw"
+        ),
+    }
+    if unit not in fallbacks:
+        raise AvailabilityGuardError("WORKER_STATE_ROOT_UNAVAILABLE:" + unit)
+    return fallbacks[unit]
+
+
+def _liveness_path(environment: Mapping[str, str], unit: str) -> Path:
+    explicit = environment.get("AIOS_WORKER_LIVENESS_PATH")
+    if explicit:
+        return Path(explicit)
+    return _state_root(environment, unit) / "worker-liveness.json"
+
+
+class AvailabilityGuard:
+    def __init__(
+        self,
+        *,
+        systemd: Systemd | None = None,
+        worker_units: Sequence[str] = DEFAULT_WORKER_UNITS,
+        dependency_units: Sequence[str] = DEFAULT_DEPENDENCY_UNITS,
+        status_path: Path = DEFAULT_STATUS,
+        recovery_path: Path = DEFAULT_RECOVERY,
+        inhibit_path: Path = DEFAULT_INHIBIT,
+        max_stale_seconds: int = DEFAULT_MAX_STALE_SECONDS,
+        restart_cooldown_seconds: int = DEFAULT_RESTART_COOLDOWN_SECONDS,
+        startup_wait_seconds: int = DEFAULT_STARTUP_WAIT_SECONDS,
+        now_fn: Callable[[], float] = time.time,
+        sleep_fn: Callable[[float], None] = time.sleep,
+    ) -> None:
+        self.systemd = systemd or Systemd()
+        self.worker_units = tuple(worker_units)
+        self.dependency_units = tuple(dependency_units)
+        self.status_path = Path(status_path)
+        self.recovery_path = Path(recovery_path)
+        self.inhibit_path = Path(inhibit_path)
+        self.max_stale_seconds = int(max_stale_seconds)
+        self.restart_cooldown_seconds = int(restart_cooldown_seconds)
+        self.startup_wait_seconds = int(startup_wait_seconds)
+        self.now_fn = now_fn
+        self.sleep_fn = sleep_fn
+        if self.max_stale_seconds < 15:
+            raise AvailabilityGuardError("MAX_STALE_SECONDS_TOO_SMALL")
+        if self.restart_cooldown_seconds < 15:
+            raise AvailabilityGuardError("RESTART_COOLDOWN_TOO_SMALL")
+        if self.startup_wait_seconds < 0 or self.startup_wait_seconds > 60:
+            raise AvailabilityGuardError("STARTUP_WAIT_INVALID")
+
+    def _recovery_state(self) -> dict[str, float]:
+        value = _load_json(self.recovery_path) or {}
+        raw = value.get("last_restart_epoch")
+        if not isinstance(raw, Mapping):
+            return {}
+        result: dict[str, float] = {}
+        for unit, epoch in raw.items():
+            try:
+                result[str(unit)] = float(epoch)
+            except (TypeError, ValueError):
+                continue
+        return result
+
+    def _save_recovery_state(self, state: Mapping[str, float]) -> None:
+        _atomic_json(
+            self.recovery_path,
+            {
+                "schema": "aios.worker-availability.recovery.v1",
+                "last_restart_epoch": dict(state),
+            },
+        )
+
+    def _basic_snapshot(self, unit: str, role: str) -> tuple[str, str, str, bool]:
+        load = self.systemd.property(unit, "LoadState")
+        active = self.systemd.property(unit, "ActiveState")
+        enabled = self.systemd.property(unit, "UnitFileState")
+        return load, active, enabled, enabled in ACCEPTED_ENABLED_STATES
+
+    def _ensure_enabled_active(self, unit: str, role: str) -> tuple[str, bool]:
+        load, active, enabled, enabled_ok = self._basic_snapshot(unit, role)
+        if load != "loaded":
+            raise AvailabilityGuardError("REQUIRED_UNIT_NOT_LOADED:" + unit)
+        action = "NONE"
+        if not enabled_ok:
+            self.systemd.enable(unit)
+            action = "ENABLE"
+        if active != "active":
+            self.systemd.start(unit)
+            action = "START" if action == "NONE" else "ENABLE_START"
+        return action, active == "active"
+
+    def _wait_for_liveness(self, path: Path) -> dict[str, Any] | None:
+        deadline = self.now_fn() + self.startup_wait_seconds
+        while True:
+            value = _load_json(path)
+            if value is not None:
+                try:
+                    epoch = float(value.get("last_seen_epoch"))
+                except (TypeError, ValueError):
+                    epoch = 0.0
+                if epoch > 0 and self.now_fn() - epoch <= self.max_stale_seconds:
+                    return value
+            if self.now_fn() >= deadline:
+                return value
+            self.sleep_fn(0.5)
+
+    def run(self, *, heal: bool) -> dict[str, Any]:
+        now = self.now_fn()
+        if self.inhibit_path.exists():
+            result = {
+                "schema": SCHEMA,
+                "status": "MAINTENANCE_INHIBITED",
+                "generated_at_epoch": now,
+                "maintenance_inhibit": str(self.inhibit_path),
+                "workers": [],
+                "dependencies": [],
+                "all_fresh": False,
+                "actions": [],
+            }
+            _atomic_json(self.status_path, result)
+            return result
+
+        actions: list[dict[str, str]] = []
+        dependencies: list[UnitSnapshot] = []
+        for unit in self.dependency_units:
+            load, active, enabled, enabled_ok = self._basic_snapshot(unit, "dependency")
+            action = "NONE"
+            if heal:
+                if load != "loaded":
+                    raise AvailabilityGuardError("REQUIRED_UNIT_NOT_LOADED:" + unit)
+                if not enabled_ok:
+                    self.systemd.enable(unit)
+                    action = "ENABLE"
+                if active != "active":
+                    self.systemd.start(unit)
+                    action = "START" if action == "NONE" else "ENABLE_START"
+                if action != "NONE":
+                    actions.append({"unit": unit, "action": action})
+                load, active, enabled, enabled_ok = self._basic_snapshot(unit, "dependency")
+            dependencies.append(
+                UnitSnapshot(
+                    unit=unit,
+                    role="dependency",
+                    load_state=load,
+                    active_state=active,
+                    unit_file_state=enabled,
+                    enabled_ok=enabled_ok,
+                    fresh=(active == "active"),
+                    action=action,
+                )
+            )
+
+        recovery = self._recovery_state()
+        workers: list[UnitSnapshot] = []
+        for unit in self.worker_units:
+            load, active, enabled, enabled_ok = self._basic_snapshot(unit, "worker")
+            environment = self.systemd.environment(unit) if load == "loaded" else {}
+            path = _liveness_path(environment, unit) if load == "loaded" else None
+            action = "NONE"
+
+            if heal:
+                if load != "loaded":
+                    raise AvailabilityGuardError("REQUIRED_UNIT_NOT_LOADED:" + unit)
+                if not enabled_ok:
+                    self.systemd.enable(unit)
+                    action = "ENABLE"
+                if active != "active":
+                    self.systemd.start(unit)
+                    action = "START" if action == "NONE" else "ENABLE_START"
+                    actions.append({"unit": unit, "action": action})
+                    if path is not None:
+                        self._wait_for_liveness(path)
+                load, active, enabled, enabled_ok = self._basic_snapshot(unit, "worker")
+
+            live = _load_json(path) if path is not None else None
+            last_seen: float | None = None
+            state: str | None = None
+            age: float | None = None
+            fresh = False
+            detail: str | None = None
+            if live is not None:
+                state = str(live.get("state") or "")
+                try:
+                    last_seen = float(live.get("last_seen_epoch"))
+                except (TypeError, ValueError):
+                    last_seen = None
+                if last_seen is not None:
+                    age = max(0.0, now - last_seen)
+                    fresh = age <= self.max_stale_seconds
+                if live.get("agent_id") in (None, ""):
+                    fresh = False
+                    detail = "LIVENESS_AGENT_ID_MISSING"
+            else:
+                detail = "LIVENESS_MISSING"
+
+            if heal and active == "active" and not fresh:
+                previous = float(recovery.get(unit, 0.0))
+                if now - previous >= self.restart_cooldown_seconds:
+                    self.systemd.restart(unit)
+                    recovery[unit] = now
+                    action = "RESTART_STALE"
+                    actions.append({"unit": unit, "action": action})
+                    if path is not None:
+                        live = self._wait_for_liveness(path)
+                    load, active, enabled, enabled_ok = self._basic_snapshot(unit, "worker")
+                    if live is not None:
+                        state = str(live.get("state") or "")
+                        try:
+                            last_seen = float(live.get("last_seen_epoch"))
+                        except (TypeError, ValueError):
+                            last_seen = None
+                        if last_seen is not None:
+                            age = max(0.0, self.now_fn() - last_seen)
+                            fresh = age <= self.max_stale_seconds
+                            detail = None if fresh else "LIVENESS_STALE_AFTER_RESTART"
+                else:
+                    detail = "RESTART_COOLDOWN_ACTIVE"
+
+            workers.append(
+                UnitSnapshot(
+                    unit=unit,
+                    role="worker",
+                    load_state=load,
+                    active_state=active,
+                    unit_file_state=enabled,
+                    enabled_ok=enabled_ok,
+                    liveness_path=str(path) if path is not None else None,
+                    liveness_state=state,
+                    last_seen_epoch=last_seen,
+                    liveness_age_seconds=age,
+                    fresh=bool(active == "active" and enabled_ok and fresh),
+                    action=action,
+                    detail=detail,
+                )
+            )
+
+        if heal:
+            self._save_recovery_state(recovery)
+
+        all_fresh = (
+            all(item.active_state == "active" and item.enabled_ok and item.fresh for item in workers)
+            and all(item.active_state == "active" for item in dependencies)
+        )
+        result = {
+            "schema": SCHEMA,
+            "status": "HEALTHY" if all_fresh else "DEGRADED",
+            "generated_at_epoch": self.now_fn(),
+            "max_stale_seconds": self.max_stale_seconds,
+            "maintenance_inhibit": str(self.inhibit_path),
+            "workers": [asdict(item) for item in workers],
+            "dependencies": [asdict(item) for item in dependencies],
+            "all_fresh": all_fresh,
+            "actions": actions,
+        }
+        _atomic_json(self.status_path, result)
+        return result
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser()
+    mode = parser.add_mutually_exclusive_group(required=True)
+    mode.add_argument("--heal", action="store_true")
+    mode.add_argument("--check", action="store_true")
+    parser.add_argument("--status-file", default=str(DEFAULT_STATUS))
+    parser.add_argument("--recovery-file", default=str(DEFAULT_RECOVERY))
+    parser.add_argument("--inhibit-file", default=str(DEFAULT_INHIBIT))
+    parser.add_argument(
+        "--max-stale-seconds",
+        type=int,
+        default=int(
+            os.environ.get(
+                "AIOS_WORKER_AVAILABILITY_MAX_STALE_SECONDS",
+                str(DEFAULT_MAX_STALE_SECONDS),
+            )
+        ),
+    )
+    parser.add_argument(
+        "--restart-cooldown-seconds",
+        type=int,
+        default=int(
+            os.environ.get(
+                "AIOS_WORKER_AVAILABILITY_RESTART_COOLDOWN_SECONDS",
+                str(DEFAULT_RESTART_COOLDOWN_SECONDS),
+            )
+        ),
+    )
+    args = parser.parse_args()
+    guard = AvailabilityGuard(
+        status_path=Path(args.status_file),
+        recovery_path=Path(args.recovery_file),
+        inhibit_path=Path(args.inhibit_file),
+        max_stale_seconds=args.max_stale_seconds,
+        restart_cooldown_seconds=args.restart_cooldown_seconds,
+    )
+    try:
+        result = guard.run(heal=args.heal)
+    except AvailabilityGuardError as exc:
+        print(json.dumps({"status": "FAILED", "error": str(exc)}, sort_keys=True))
+        return 20
+    print(json.dumps(result, ensure_ascii=False, sort_keys=True))
+    if args.check and not result.get("all_fresh"):
+        return 10
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

config/external-agent-runtime.d/hermes.profile.json

@@ -6,7 +6,7 @@
     "--provider",
     "oneapi",
     "-m",
-    "deepseek-v4-flash",
+    "{model_id}",
     "--no-restore-cwd"
   ],
   "credential_env": {
@@ -21,6 +21,16 @@
   },
   "input_mode": "stdin_text",
   "max_output_bytes": 131072,
+  "model_governance": {
+    "max_output_tokens": 512,
+    "model_capabilities": [
+      "text"
+    ],
+    "provider_ids": [
+      "oneapi"
+    ],
+    "required": true
+  },
   "output_mode": "stdout_text",
   "runtime": "hermes",
   "schema": "aios.external-agent-runtime.profile.v1",

config/external-agent-runtime.d/openclaw.profile.json

@@ -12,7 +12,7 @@
     "--message",
     "{prompt}",
     "--model",
-    "openai/deepseek-v4-flash",
+    "{model_id}",
     "--timeout",
     "180"
   ],
@@ -30,6 +30,17 @@
   },
   "input_mode": "stdin_text",
   "max_output_bytes": 131072,
+  "model_governance": {
+    "input_reservation_floor_tokens": 16384,
+    "max_output_tokens": 256,
+    "model_capabilities": [
+      "text"
+    ],
+    "provider_ids": [
+      "openai"
+    ],
+    "required": true
+  },
   "output_mode": "stdout_json",
   "runtime": "openclaw",
   "schema": "aios.external-agent-runtime.profile.v1",

config/model-governance/model-governance.staging.json

@@ -0,0 +1,95 @@
+{
+  "models": [
+    {
+      "capabilities": [
+        "agent.runtime.hermes",
+        "text"
+      ],
+      "enabled": true,
+      "evaluation_sha256": "43ae0db1640daeb6d33e2787769732844f366d36070065d3a2239905bd9a823b",
+      "evaluation_state": "APPROVED",
+      "input_cost_micros_per_1k": 0,
+      "max_input_tokens": 32768,
+      "max_output_tokens": 512,
+      "model_id": "deepseek-v4-flash",
+      "output_cost_micros_per_1k": 0,
+      "priority": 10,
+      "provider_id": "oneapi",
+      "version": "profile-sha:54615e533bfdb6a2e8b870ae74bb8cd7fc45177495bcf6e8c45a50837cefb85a"
+    },
+    {
+      "capabilities": [
+        "agent.runtime.openclaw",
+        "text"
+      ],
+      "enabled": true,
+      "evaluation_sha256": "7b1c89634eb9c0e7cb5a3d75323d904ebbd91b3ae5494ce938bd8316a4cf5b12",
+      "evaluation_state": "APPROVED",
+      "input_cost_micros_per_1k": 0,
+      "max_input_tokens": 32768,
+      "max_output_tokens": 256,
+      "model_id": "openai/deepseek-v4-flash",
+      "output_cost_micros_per_1k": 0,
+      "priority": 10,
+      "provider_id": "openai",
+      "version": "profile-sha:2a31c480226aa78d44ed9692bddcedc99c3aa03c23c0dd5f031a5cbbb8bbd7ba"
+    }
+  ],
+  "policies": [
+    {
+      "allowed_models": [
+        "deepseek-v4-flash",
+        "openai/deepseek-v4-flash"
+      ],
+      "daily_cost_budget_micros": 0,
+      "daily_token_budget": 250000,
+      "denied_providers": [],
+      "fallback_limit": 1,
+      "max_input_tokens_per_request": 16384,
+      "max_output_tokens_per_request": 512,
+      "project_id": null,
+      "required_capabilities": [
+        "text"
+      ],
+      "scope_type": "legacy",
+      "tenant_id": null
+    },
+    {
+      "allowed_models": [
+        "deepseek-v4-flash",
+        "openai/deepseek-v4-flash"
+      ],
+      "daily_cost_budget_micros": 0,
+      "daily_token_budget": 250000,
+      "denied_providers": [],
+      "fallback_limit": 1,
+      "max_input_tokens_per_request": 16384,
+      "max_output_tokens_per_request": 512,
+      "project_id": "staging",
+      "required_capabilities": [
+        "text"
+      ],
+      "scope_type": "tenant",
+      "tenant_id": "model-governance-live"
+    }
+  ],
+  "providers": [
+    {
+      "cooldown_seconds": 60,
+      "credential_ref": "runtime-profile:hermes",
+      "enabled": true,
+      "failure_domain": "oneapi",
+      "max_failures": 2,
+      "provider_id": "oneapi"
+    },
+    {
+      "cooldown_seconds": 60,
+      "credential_ref": "runtime-profile:openclaw",
+      "enabled": true,
+      "failure_domain": "oneapi",
+      "max_failures": 2,
+      "provider_id": "openai"
+    }
+  ],
+  "schema": "aios.model-governance.bootstrap.v1"
+}

deploy/systemd/aios-external-agent-executor.service

@@ -33,7 +33,10 @@ InaccessiblePaths=/data/agent/secrets
 ReadOnlyPaths=/etc/aios/external-agent-runtime.d
 ReadOnlyPaths=/opt/aios-hermes-agent
 ReadOnlyPaths=/opt/aios/openclaw-engineering
+Environment=AIOS_MODEL_GOVERNANCE_DB=/var/lib/aios-model-governance/state.sqlite3
+Environment=AIOS_MODEL_GOVERNANCE_ENABLED=1
 ReadWritePaths=/var/lib/aios-external-agent-executor
+ReadWritePaths=/var/lib/aios-model-governance
 UMask=0027
 
 [Install]

deploy/systemd/aios-gpt-https-edge.service

@@ -0,0 +1,41 @@
+[Unit]
+Description=AIOS disabled-by-default HTTPS edge for GPT read-only Online Task ingress
+After=network-online.target aios-gpt-online-task-ingress.service
+Wants=network-online.target aios-gpt-online-task-ingress.service
+ConditionPathExists=/etc/aios-gitops/gpt-https-edge.approved
+
+[Service]
+Type=simple
+DynamicUser=yes
+LoadCredential=tls_cert:/etc/aios-gitops/gpt-https-edge-tls.crt
+LoadCredential=tls_key:/etc/aios-gitops/gpt-https-edge-tls.key
+Environment=AIOS_GPT_EDGE_TLS_CERT_FILE=%d/tls_cert
+Environment=AIOS_GPT_EDGE_TLS_KEY_FILE=%d/tls_key
+Environment=AIOS_GPT_EDGE_HOST=127.0.0.1
+Environment=AIOS_GPT_EDGE_PORT=18443
+Environment=AIOS_GPT_EDGE_PUBLIC_BIND_APPROVED=0
+Environment=AIOS_GPT_EDGE_PUBLIC_DOMAIN=
+ExecStart=/usr/bin/python3 -B /opt/aios/current/active-runtime/app/aios_gpt_https_edge.py
+Restart=on-failure
+RestartSec=2
+NoNewPrivileges=yes
+PrivateTmp=yes
+PrivateDevices=yes
+ProtectSystem=strict
+ProtectHome=yes
+ProtectKernelTunables=yes
+ProtectKernelModules=yes
+ProtectControlGroups=yes
+ProtectKernelLogs=yes
+ProtectClock=yes
+LockPersonality=yes
+RestrictSUIDSGID=yes
+RestrictRealtime=yes
+RestrictNamespaces=yes
+MemoryDenyWriteExecute=yes
+RestrictAddressFamilies=AF_INET AF_INET6
+InaccessiblePaths=/data/aios-gitops /etc/aios-gitops /home /var/lib/aios-agent-task
+UMask=0077
+
+[Install]
+WantedBy=multi-user.target

deploy/systemd/aios-gpt-online-task-ingress.service

@@ -0,0 +1,45 @@
+[Unit]
+Description=AIOS external GPT read-only Online Task ingress
+After=network.target aios-chatgpt-bridge-v2-staging.service aios-agent-task-api.service
+Wants=aios-chatgpt-bridge-v2-staging.service aios-agent-task-api.service
+
+[Service]
+Type=simple
+DynamicUser=yes
+StateDirectory=aios-gpt-online-task-ingress
+StateDirectoryMode=0750
+LoadCredential=gpt_ingress_token:/etc/aios-gitops/gpt-online-task-ingress.token
+LoadCredential=agent_task_control_token:/etc/aios-gitops/agent-task-control.token
+Environment=AIOS_GPT_INGRESS_TOKEN_FILE=%d/gpt_ingress_token
+Environment=AIOS_AGENT_TASK_CONTROL_TOKEN_FILE=%d/agent_task_control_token
+Environment=AIOS_GPT_INGRESS_STATE_ROOT=%S/aios-gpt-online-task-ingress
+Environment=AIOS_GPT_INGRESS_HOST=127.0.0.1
+Environment=AIOS_GPT_INGRESS_PORT=18945
+Environment=AIOS_GOVERNED_WRITE_API_URL=http://127.0.0.1:18892
+Environment=AIOS_AGENT_TASK_API_URL=http://127.0.0.1:18940
+ExecStart=/usr/bin/python3 -B /opt/aios/current/active-runtime/app/aios_gpt_online_task_ingress.py
+Restart=on-failure
+RestartSec=2
+NoNewPrivileges=yes
+PrivateTmp=yes
+PrivateDevices=yes
+ProtectSystem=strict
+ProtectHome=yes
+ProtectKernelTunables=yes
+ProtectKernelModules=yes
+ProtectControlGroups=yes
+ProtectKernelLogs=yes
+ProtectClock=yes
+LockPersonality=yes
+RestrictSUIDSGID=yes
+RestrictRealtime=yes
+RestrictNamespaces=yes
+MemoryDenyWriteExecute=yes
+RestrictAddressFamilies=AF_UNIX AF_INET
+IPAddressDeny=any
+IPAddressAllow=127.0.0.0/8
+InaccessiblePaths=/data/aios-gitops /etc/aios-gitops /home
+UMask=0077
+
+[Install]
+WantedBy=multi-user.target

deploy/systemd/aios-owner-durable-authorization-broker.service

@@ -0,0 +1,47 @@
+[Unit]
+Description=AIOS durable Owner authorization broker
+After=network.target aios-chatgpt-bridge-v2-staging.service
+Wants=aios-chatgpt-bridge-v2-staging.service
+
+[Service]
+Type=simple
+User=root
+Group=aios-executor-exchange
+RuntimeDirectory=aios-owner-authorization-broker
+RuntimeDirectoryMode=0750
+StateDirectory=aios-owner-authorization-broker
+StateDirectoryMode=0750
+LoadCredential=agent_task_control_token:/etc/aios-gitops/agent-task-control.token
+Environment=AIOS_AGENT_TASK_CONTROL_TOKEN_FILE=%d/agent_task_control_token
+Environment=AIOS_OWNER_DURABLE_AUTHORIZATION_POLICY=/etc/aios-gitops/owner-durable-authorization.json
+Environment=AIOS_OWNER_DURABLE_AUTHORIZATION_REVOCATIONS=/etc/aios-gitops/owner-durable-authorization-revocations.json
+Environment=AIOS_OWNER_AUTH_BROKER_ACCOUNT=chatgpt-governed-implementer-singlehost
+Environment=AIOS_OWNER_AUTH_BROKER_SOCKET=%t/aios-owner-authorization-broker/broker.sock
+ExecStart=/usr/bin/python3 -B /opt/aios/current/active-runtime/app/aios_durable_owner_authorization.py
+Restart=on-failure
+RestartSec=2
+NoNewPrivileges=yes
+PrivateTmp=yes
+PrivateDevices=yes
+ProtectSystem=strict
+ProtectHome=yes
+ProtectKernelTunables=yes
+ProtectKernelModules=yes
+ProtectControlGroups=yes
+ProtectKernelLogs=yes
+ProtectClock=yes
+LockPersonality=yes
+RestrictSUIDSGID=yes
+RestrictRealtime=yes
+RestrictNamespaces=yes
+MemoryDenyWriteExecute=yes
+RestrictAddressFamilies=AF_UNIX AF_INET
+IPAddressDeny=any
+IPAddressAllow=127.0.0.0/8
+ReadOnlyPaths=/etc/aios-gitops/owner-durable-authorization.json
+ReadOnlyPaths=/etc/aios-gitops/owner-durable-authorization-revocations.json
+InaccessiblePaths=/data/aios-gitops /home
+UMask=0007
+
+[Install]
+WantedBy=multi-user.target

deploy/systemd/aios-worker-availability-guard.service

@@ -0,0 +1,27 @@
+[Unit]
+Description=AIOS persistent Online Agent worker availability guard
+After=local-fs.target
+
+[Service]
+Type=oneshot
+User=root
+Group=root
+Environment=PYTHONUNBUFFERED=1
+Environment=AIOS_WORKER_AVAILABILITY_MAX_STALE_SECONDS=90
+Environment=AIOS_WORKER_AVAILABILITY_RESTART_COOLDOWN_SECONDS=60
+ExecStart=/usr/bin/python3 -B /opt/aios/current/active-runtime/app/aios_worker_availability_guard.py --heal
+NoNewPrivileges=true
+PrivateTmp=true
+ProtectSystem=strict
+ProtectHome=true
+ProtectKernelTunables=true
+ProtectKernelModules=true
+ProtectControlGroups=true
+LockPersonality=true
+RestrictSUIDSGID=true
+RestrictRealtime=true
+RestrictNamespaces=true
+MemoryDenyWriteExecute=true
+RestrictAddressFamilies=AF_UNIX
+ReadWritePaths=/run/aios-worker-availability
+UMask=0022

deploy/systemd/aios-worker-availability-guard.timer

@@ -0,0 +1,13 @@
+[Unit]
+Description=Periodically ensure AIOS Online Agent workers remain available
+
+[Timer]
+OnBootSec=15s
+OnUnitActiveSec=30s
+AccuracySec=5s
+RandomizedDelaySec=3s
+Persistent=true
+Unit=aios-worker-availability-guard.service
+
+[Install]
+WantedBy=timers.target

deploy/tmpfiles.d/aios-model-governance.conf

@@ -0,0 +1 @@
+d /var/lib/aios-model-governance 0750 aios-engineering-executor aios-executor-exchange -

deploy/tmpfiles.d/aios-worker-availability.conf

@@ -0,0 +1 @@
+d /run/aios-worker-availability 0755 root root -

docs/superpowers/specs/2026-08-18-external-agent-attempt-request-identity.md

@@ -0,0 +1,123 @@
+# External Agent Attempt-Scoped Request Identity
+
+## Problem
+
+External Agent execution previously derived its durable request ID from:
+
+```text
+task_id
+job_id
+execution_digest
+runtime
+```
+
+The Online Task attempt was missing.
+
+That is incorrect for retryable external side effects. A failed Model Governance route reservation is
+durable and terminal for that request ID. Reusing the same external request ID on a later official
+Task attempt therefore reuses the old route decision/reservation and can fail settlement with:
+
+```text
+MODEL_RESERVATION_STATE_INVALID
+```
+
+Archiving an exchange failed marker does not fix the root cause because the Model Governance
+reservation remains intentionally durable.
+
+## Contract
+
+External request identity is now:
+
+```text
+hash(
+  task_id,
+  job_id,
+  execution_digest,
+  runtime,
+  attempt
+)
+```
+
+The attempt is a **transient execution coordinate**. It is not added to the signed governed execution
+envelope and does not alter the Task/Job binding or execution digest.
+
+The worker gets the authoritative attempt from the successful claim response and passes it through:
+
+```text
+process_claim
+→ RuntimeAdapter.execute(..., attempt=N)
+→ ExternalAgentExecutor
+→ ExternalAgentExchange
+→ external request JSON
+→ external result JSON
+```
+
+The external result must bind the same attempt.
+
+## Idempotency
+
+This gives the intended two-level behavior:
+
+```text
+same Task/Job/execution/runtime + same attempt
+→ same external request ID
+→ same external result/reservation can be replayed safely
+
+same Task/Job/execution/runtime + next official attempt
+→ different external request ID
+→ different Model Governance request ID
+→ new reservation
+```
+
+Old FAILED reservations are retained for audit. No Model Governance DB cleanup is required or
+permitted.
+
+## Model Governance
+
+Model Governance already derives its route request ID from:
+
+```text
+external_request_id + ":model"
+```
+
+Therefore making the external request ID attempt-scoped automatically makes route decisions and
+reservations attempt-scoped without creating another state authority.
+
+## Recovery
+
+For the current Hermes live Task:
+
+```text
+attempt 6
+→ old request/reservation remains FAILED
+
+official NEEDS_REWORK
+→ bounded retry extension 6 → 7
+
+attempt 7
+→ new external request ID
+→ new route request ID
+→ new reservation
+→ provider execution
+→ settlement
+```
+
+No replacement Task/Job, direct DB rewrite, failed-reservation deletion or marker deletion is
+required.
+
+## Worker Availability
+
+The repair must preserve the already verified Worker Availability runtime. Runtime switches use the
+maintenance inhibit and must finish with the guard healthy and all three workers fresh.
+
+## Verification
+
+Required regression coverage:
+
+- request ID stable within one attempt;
+- request ID different across attempts;
+- request schema requires a valid attempt;
+- external result binds attempt;
+- old failed Model Governance reservation remains FAILED;
+- next attempt gets a new reservation and can settle;
+- existing external-agent and Model Governance runtime tests still pass.

docs/superpowers/specs/2026-08-18-model-governance-policy-profile-floor-compatibility.md

@@ -0,0 +1,109 @@
+# Model Governance Policy / Profile Floor Compatibility
+
+## Problem
+
+The root-owned OpenClaw runtime profile reserves a minimum of 16,384 input tokens so Model
+Governance can cover OpenClaw-injected workspace, skills and system context before provider
+execution.
+
+The staging scope policies still limited input to 16,000 tokens per request. Route selection
+therefore failed before reservation with:
+
+```text
+MODEL_INPUT_TOKEN_LIMIT_DENY
+```
+
+This was a configuration invariant violation:
+
+```text
+scope max_input_tokens_per_request
+<
+root-owned profile input_reservation_floor_tokens
+```
+
+The registered OpenClaw model itself supports 32,768 input tokens, so the model capability was not
+the blocker.
+
+## Invariant
+
+For every Model Governance scope that allows a model/runtime whose root-owned profile declares an
+input reservation floor:
+
+```text
+scope.max_input_tokens_per_request
+>=
+profile.model_governance.input_reservation_floor_tokens
+```
+
+The profile floor remains root-owned. Task payloads cannot alter it.
+
+## Staging policy
+
+The two staging policies that allow `openai/deepseek-v4-flash` are updated from:
+
+```text
+16000
+```
+
+to:
+
+```text
+16384
+```
+
+This is the minimum compatible value. It intentionally remains below the model's 32,768-token
+technical input ceiling.
+
+The daily token budget is unchanged.
+
+## Governance path
+
+The live SQLite policy must be changed only through the repository bootstrap and
+`aios_model_governance_admin.py`.
+
+The admin reconciler:
+
+- reads the governed repository config;
+- uses current policy revision as CAS;
+- calls `ModelGovernanceStore.set_scope_policy`;
+- appends normal Model Governance audit evidence;
+- verifies the audit chain.
+
+Direct SQL updates remain forbidden.
+
+## Exhausted OpenClaw recovery
+
+The fixed OpenClaw live Task has exhausted its normal three attempts.
+
+After the governed policy candidate is officially verified and the live policy is reconciled:
+
+```text
+attempt 3 VERIFYING/FAILED result
+→ official NEEDS_REWORK
+→ Task becomes FAILED 3/3
+→ official bounded retry extension 3 → 4
+→ same Task / same Job returns QUEUED
+→ persistent OpenClaw worker auto-claims attempt 4
+```
+
+AR1 attempt-scoped external request identity ensures attempt 4 receives a new external request and a
+new Model Governance reservation.
+
+No historical reservation is deleted or rewritten.
+
+## Acceptance
+
+Attempt 4 must prove:
+
+```text
+scope max_input_tokens_per_request=16384
+profile input_reservation_floor_tokens=16384
+reserved_estimated_input_tokens>=16384
+actual input <= reserved input
+settlement=SETTLED
+usage_source=runtime
+audit_chain_valid=true
+```
+
+If attempt 4 exposes a new failure, stop. No additional retry extension is authorized by this
+repair.

docs/superpowers/specs/2026-08-18-model-governance-runtime.md

@@ -0,0 +1,211 @@
+# AIOS Model Governance Runtime
+
+## Status and purpose
+
+This component turns the previously `PARTIAL` Model Governance interface map into an enforceable
+runtime policy for the already verified Hermes/OpenClaw Online Agent path.
+
+It is intentionally a **single SQLite library**, not a new network service. The existing external
+Agent executor calls the library immediately before launching a model-backed runtime and settles the
+same reservation immediately after the result.
+
+The authoritative execution chain remains:
+
+```text
+Online Task scope / trace
+→ dedicated worker claim + heartbeat
+→ external Agent executor
+→ ModelGovernanceStore.route()
+→ root-owned runtime profile binds provider/model/version
+→ Hermes / OpenClaw process
+→ ModelGovernanceStore.record_success() / record_failure()
+→ durable Online Agent checkpoint/result
+```
+
+P0 scope/trace/credential, P1 checkpoint/retry, P2 fleet scheduling and the existing external Agent
+credential boundary remain authoritative.
+
+## Ten Model Governance contracts
+
+### MODEL_PROVIDER_REGISTRY
+
+`model_governance_providers` stores provider identity, failure domain, circuit policy and an opaque
+credential reference. Raw credentials are rejected.
+
+Allowed reference prefixes:
+
+- `systemd-credential:`
+- `runtime-profile:`
+- `credential-broker:`
+
+### MODEL_SELECTION_POLICY
+
+`model_governance_scope_policies` constrains eligible models/providers/capabilities. Routing is
+deterministic: explicit preferred models, then configured priority, then stable provider/model IDs.
+
+### MODEL_VERSION_PINNING
+
+A registry row always includes a model version. A scope can additionally pin a model ID + exact
+version. Every route decision persists both model ID and version plus model/policy revisions.
+
+### MODEL_BUDGET_ACCOUNTING
+
+A route first reserves estimated input/output tokens and estimated cost inside `BEGIN IMMEDIATE`.
+Settled usage and outstanding reservations both count against the daily scope budget, preventing two
+concurrent requests from oversubscribing the same budget.
+
+Runtime-reported usage is preferred. When telemetry is absent, the reserved upper bound is charged
+rather than silently under-counting.
+
+OpenClaw's verified result shape is supported directly:
+
+```text
+meta.agentMeta.usage.input
+meta.agentMeta.usage.output
+meta.agentMeta.costUsd
+```
+
+Hermes currently has no structured usage in the verified output, so it uses conservative
+reservation settlement.
+
+### MODEL_FALLBACK_POLICY
+
+Provider failures are durable. Repeated failures open a circuit for the configured cooldown. A
+fallback request excludes the failed provider and increments `fallback_index`; the scope policy
+bounds the number of fallback attempts.
+
+### MODEL_SCOPE_ISOLATION
+
+The signed `agent_runtime` execution envelope now carries authoritative Task scope + trace identity.
+
+The projection layer always overwrites `runtime_input.scope` and `runtime_input.trace_id` from the
+Online Task, even if an execution request attempts to supply different values.
+
+Tenant/project lookup may fall back only to the same tenant's policy. It never falls back to a
+system or legacy policy. `system` and `legacy` require their own explicit policies, avoiding fake
+tenant identities.
+
+### MODEL_CREDENTIAL_BOUNDARY
+
+Model governance stores credential references only. The root-owned external-runtime profile still
+owns the actual provider credential projection. Task payloads cannot select executable, argv,
+environment, provider credential, or credential path.
+
+The external child process remains unable to reach the AIOS control plane.
+
+### MODEL_TRACE_AUDIT
+
+Every route, failure, provider/model/policy update and settlement appends to a SHA-256 hash-chained
+audit table with task ID, trace ID and scope key. `verify_audit_chain()` detects tampering.
+
+### MODEL_EVALUATION_GATE
+
+A model cannot enter the registry unless its state is `APPROVED` and it has a 64-hex evaluation
+evidence digest.
+
+The staging bootstrap uses the already verified Hermes/OpenClaw E2E external-result SHA256 values as
+evaluation evidence, rather than a configuration-file hash.
+
+### MODEL_ROUTING_RESILIENCE
+
+Route request IDs are idempotent and bound to a canonical request hash. Conflicting replay is
+rejected. Provider health, reservations, usage settlement and policy/provider/model updates use
+transaction/CAS semantics.
+
+## Runtime profile contract
+
+A root-owned Hermes/OpenClaw profile may declare:
+
+```json
+{
+  "model_governance": {
+    "required": true,
+    "provider_ids": ["provider-a"],
+    "model_capabilities": ["text"],
+    "max_output_tokens": 512
+  }
+}
+```
+
+When `required=true`:
+
+- `provider_ids` must be non-empty;
+- `max_output_tokens` must be positive;
+- argv/environment must contain at least one governance-controlled marker:
+  `{provider_id}`, `{model_id}`, or `{model_version}`;
+- `AIOS_MODEL_GOVERNANCE_ENABLED` must be enabled;
+- a matching scope policy/model/provider must exist;
+- budget reservation must succeed before the child process is launched.
+
+The task cannot provide these markers or their values.
+
+## SQLite authority
+
+Default database:
+
+```text
+/var/lib/aios-model-governance/state.sqlite3
+```
+
+The executor process already owns the only runtime mutation path. The directory is created by
+tmpfiles and is the only new writable path added to that executor.
+
+No direct Task/Job/Git database writes are introduced.
+
+## Bootstrap/reconciliation
+
+`aios_model_governance_admin.py` applies a root-owned JSON bootstrap idempotently. It compares desired
+state to current rows and uses the current revision as the CAS input only when a change is actually
+required.
+
+The repository staging bootstrap is not a Production policy. It exists only to reproduce and
+govern the already verified non-production Hermes/OpenClaw runtimes.
+
+## Cost caveat
+
+The current verified OneAPI-backed runtime evidence reports zero authoritative provider cost for
+OpenClaw and no structured cost for Hermes. The staging bootstrap therefore does **not** invent
+pricing and sets cost rates/budget to zero/unbounded.
+
+Token budgets remain enforced. Real monetary budget enforcement becomes effective as soon as
+authoritative per-model pricing is entered into the root-owned registry.
+
+## OpenClaw JIT/WASM compatibility
+
+The previous live acceptance proved that OpenClaw Node/WASM requires
+`MemoryDenyWriteExecute=no`. Model Governance must preserve that already-evidenced runtime exception;
+it must not further weaken `NoNewPrivileges`, `ProtectSystem=strict`, `ProtectHome`, loopback deny,
+credential isolation or repository isolation.
+
+## Failure/recovery invariants
+
+- Budget denial occurs before provider execution.
+- Missing scope policy fails closed.
+- Missing/disabled governance fails closed when a profile requires it.
+- Actual usage cannot exceed the reservation.
+- Provider failure releases that reservation into `FAILED` state before fallback.
+- Settlement is idempotent.
+- Route replay is idempotent only for the exact canonical request.
+- Audit corruption is detectable.
+- Online Agent checkpoint/result retry remains unchanged and remains the outer recovery authority.
+
+## Verification
+
+Required unit/integration coverage includes:
+
+- registry/evaluation/CAS;
+- scope isolation and no system→tenant fallback;
+- model/version pinning;
+- atomic budget reservation;
+- settlement/replay;
+- missing telemetry conservative settlement;
+- provider circuit/fallback;
+- trace/audit tamper detection;
+- authoritative scope/trace projection;
+- task runtime-input scope override attack;
+- root-owned provider/model binding;
+- OpenClaw nested usage extraction;
+- real Hermes/OpenClaw staging live tasks with Model Governance evidence.
+
+The governed implementation is not complete without the unified all-modified-files bundle for the
+candidate delta and the cumulative lineage bundle.

docs/superpowers/specs/2026-08-18-openclaw-input-reservation-floor.md

@@ -0,0 +1,80 @@
+# OpenClaw Input Reservation Floor
+
+## Problem
+
+Model Governance originally estimated input reservation from the user-visible prompt only.
+
+The verified OpenClaw runtime injects additional root-owned/system-controlled context including
+workspace, skills and system instructions. In live acceptance the prompt estimate reserved only
+153 input tokens while OpenClaw reported approximately 7,770 actual input tokens.
+
+The settlement correctly failed closed with:
+
+```text
+MODEL_USAGE_EXCEEDS_RESERVATION
+```
+
+The provider call itself had succeeded. The defect was therefore pre-execution reservation
+underestimation, not provider connectivity.
+
+## Contract
+
+A root-owned external Agent profile may now declare:
+
+```json
+{
+  "model_governance": {
+    "input_reservation_floor_tokens": 16384
+  }
+}
+```
+
+The effective input reservation is:
+
+```text
+max(
+  byte-based prompt estimate,
+  profile input_reservation_floor_tokens
+)
+```
+
+Only the root-owned runtime profile can choose this floor. Online Task payloads cannot lower or
+override it.
+
+The OpenClaw staging profile uses 16,384 tokens. This is greater than twice the observed 7,770-token
+live input high-water mark while remaining below the registered 32,768-token model input limit.
+
+Hermes has no floor unless explicitly configured and therefore preserves its existing behavior.
+
+## Fail-closed behavior
+
+The floor is not an unlimited post-execution top-up.
+
+If actual runtime usage exceeds the reserved amount, settlement still fails with
+`MODEL_USAGE_EXCEEDS_RESERVATION`.
+
+This keeps budget enforcement pre-execution and avoids silently authorizing usage after it has
+already happened.
+
+## Evidence
+
+External results expose:
+
+```text
+prompt_estimated_input_tokens
+input_reservation_floor_tokens
+reserved_estimated_input_tokens
+```
+
+alongside the normal settlement and audit evidence.
+
+## Retry
+
+The current OpenClaw attempt 1 failure remains preserved. After this code repair is governed and
+activated, official `NEEDS_REWORK` returns the same fixed Task to the normal retry path. Because AR1
+already made external request identity attempt-scoped, attempt 2 receives a new external request ID
+and a new reservation.
+
+No retry-extension is required because the Task is currently attempt 1 of 3.
+
+No Model Governance reservation is deleted or rewritten.

docs/superpowers/specs/2026-08-18-worker-availability-runtime.md

@@ -0,0 +1,212 @@
+# AIOS Worker Availability Runtime
+
+## Priority
+
+This runtime exists to make the already verified Online Agent execution path continuously usable
+without repeated human intervention.
+
+The system already proved that generic, Hermes and OpenClaw workers can claim and complete real
+Online Tasks. The remaining availability defect was operational: evidence-freeze workflows left the
+worker services stopped, while registry `ACTIVE` records remained stale.
+
+The availability contract therefore separates:
+
+```text
+AUTO_CLAIM_MECHANISM=VERIFIED
+CURRENT_WORKER_AVAILABILITY=<fresh runtime state>
+```
+
+A historical successful claim is not current liveness.
+
+## Source of truth for idle liveness
+
+An idle worker may have no Task lease heartbeat. Therefore Task heartbeat alone cannot prove that
+the worker is continuously polling.
+
+`aios_online_agent_worker.py` now writes:
+
+```text
+<AIOS_WORKER_STATE_ROOT>/worker-liveness.json
+```
+
+on:
+
+- process start: `STARTING`;
+- every successful poll: `POLL_OK`;
+- poll failure: `POLL_FAILED`;
+- every successful Task lease heartbeat while executing: `BUSY`.
+
+The file contains:
+
+- agent ID/type/version;
+- process PID;
+- monotonically increasing local sequence;
+- `last_seen_epoch`;
+- poll/heartbeat intervals;
+- current Task/stage while busy.
+
+The liveness file is worker-owned observability state only. It does not write or mutate Task, Job,
+Git, WriterLease or registry databases.
+
+## Self-healing guard
+
+`aios-worker-availability-guard.timer` invokes a root-owned oneshot guard every 30 seconds.
+
+The guard manages:
+
+```text
+aios-online-agent-worker-20260815.service
+aios-hermes-online-agent-worker.service
+aios-openclaw-online-agent-worker.service
+```
+
+and the required external runtime dependency:
+
+```text
+aios-external-agent-executor.service
+```
+
+For each worker:
+
+1. missing systemd unit → fail closed;
+2. disabled → `systemctl enable`;
+3. inactive → `systemctl start`;
+4. active but liveness older than 90 seconds → restart;
+5. repeated stale restart within 60 seconds → do not restart again; report degraded.
+
+Long-running Tasks are safe because the Task lease-heartbeat thread refreshes `BUSY` liveness.
+
+The guard never claims a Task and never touches Task/Job/Git databases.
+
+## Maintenance inhibit
+
+Planned maintenance creates:
+
+```text
+/run/aios-worker-availability/maintenance.lock
+```
+
+While this file exists the guard performs no start/restart actions and reports:
+
+```text
+MAINTENANCE_INHIBITED
+```
+
+Maintenance tooling must remove the file when done. This prevents the self-healer from fighting
+intentional service changes.
+
+## Status
+
+The guard writes:
+
+```text
+/run/aios-worker-availability/status.json
+```
+
+with current worker/dependency states, liveness ages and recovery actions.
+
+A worker is currently available only when:
+
+```text
+systemd load_state=loaded
+systemd active_state=active
+unit enabled/static
+worker-liveness.json present
+last_seen_age <= 90s
+```
+
+Registry `ACTIVE` without fresh poll/busy liveness is not sufficient.
+
+## Boot/restart behavior
+
+The timer is:
+
+```text
+OnBootSec=15s
+OnUnitActiveSec=30s
+Persistent=true
+```
+
+Therefore reboot, manual stop left behind by another workflow, process crash, and hung poll loop are
+all repaired without requiring a human to notice and intervene.
+
+Normal worker units retain their own `Restart=always`; the availability guard handles the cases that
+`Restart=always` cannot, especially a deliberate `systemctl stop` left after an evidence workflow.
+
+## Acceptance
+
+The live acceptance requires all of the following:
+
+```text
+guard timer active/enabled
+external executor active/enabled
+generic worker active/enabled + fresh liveness
+Hermes worker active/enabled + fresh liveness
+OpenClaw worker active/enabled + fresh liveness
+```
+
+Then a new read-only Online Task is projected through canonical Bridge `18892`.
+
+Success requires:
+
+```text
+TASK_CREATED
+→ generic worker auto claim
+→ TASK_HEARTBEAT
+→ RESULT_SUBMITTED
+→ official TASK_VERIFIED(SUCCEEDED)
+```
+
+After the Task, the services must remain active and liveness must continue advancing. The acceptance
+also stops a worker intentionally once and proves the guard starts it again with a new fresh
+liveness sequence.
+
+Only after these gates may current status be upgraded from:
+
+```text
+OVERALL=PARTIAL
+```
+
+to:
+
+```text
+PERSISTENT_WORKER_LIVENESS=VERIFIED
+NEW_TASK_AUTO_CONSUMPTION=VERIFIED
+WORKER_AVAILABILITY_RUNTIME=VERIFIED_CANDIDATE
+```
+
+Protected main and Production remain out of scope.
+
+
+## Forced source-capture projection compatibility
+
+The deployed canonical Bridge currently projects a default `source_capture` envelope even when a
+governed binding request asks for binding-only behavior. This must not be allowed to make always-on
+workers consume engineering Tasks before their deterministic plan is registered.
+
+The worker therefore distinguishes a real read-only source capture from an engineering placeholder:
+
+```text
+source_capture
++ capability_scope contains governed.write.branch or governed.write.file
+= PENDING_DETERMINISTIC_PLAN
+```
+
+For such a placeholder the worker:
+
+1. re-resolves through the canonical plan registry **without replaying the stale source-capture
+   envelope**;
+2. if no deterministic plan exists, emits `TASK_SKIPPED_PENDING_DETERMINISTIC_PLAN` and does not
+   claim;
+3. if the registered canonical envelope is deterministic and the governed binding still matches,
+   uses that deterministic envelope and claims normally.
+
+A real read-only source-capture Task with only `git.read`/`git.test` still replays its exact persisted
+payload and remains claimable.
+
+This makes persistent workers safe with the currently deployed Bridge while preserving the same
+Task/Job binding. No Task DB rewrite, cancellation, replacement Task, fake NEEDS_REWORK, or
+pre-claim metadata mutation is required.
+
+A later Bridge cleanup may implement true binding-only projection, but Worker Availability does not
+depend on that cleanup.

docs/superpowers/specs/2026-08-19-all-modified-files-packager-secret-scan.md

@@ -0,0 +1,96 @@
+# Canonical all-modified-files packager and secret scanning
+
+## Problem
+
+The historical all-modified-files evidence packager lived only as a copied handoff script.
+
+Its private-key scanner matched the fragment:
+
+```text
+BEGIN PRIVATE KEY
+```
+
+instead of a complete PEM header. A regression test containing that literal therefore caused a
+false secret finding even though the candidate contained no credential.
+
+A task had to use a temporary evidence-only compatibility packager to complete its bundle. That is
+acceptable as an emergency workaround but not as a durable unattended path.
+
+## Repository source of truth
+
+The canonical implementation is now:
+
+```text
+active-runtime/app/aios_all_modified_files_packager.py
+```
+
+It preserves the historical output/bundle schemas:
+
+```text
+aios.all-modified-files.bundle.v1
+aios.all-modified-files.bundle-summary.v1
+aios.modified-documents.index.v1
+```
+
+and the host-visible status lines:
+
+```text
+ALL_MODIFIED_FILES_BUNDLE=PASS
+ALL_MODIFIED_FILES_BUNDLE_PATH=...
+ALL_MODIFIED_FILES_BUNDLE_SHA256=...
+```
+
+Future governed tasks should use this repository version instead of copying an unversioned scanner.
+
+## Secret scanning invariant
+
+Private-key detection requires a complete PEM header:
+
+```text
+-----BEGIN [RSA|EC|DSA|OPENSSH ]PRIVATE KEY-----
+```
+
+The following source/test/document text is not secret material and must remain packageable:
+
+```text
+BEGIN PRIVATE KEY
+PRIVATE KEY
+PEM private-key header
+```
+
+Bearer and key/value credential detection remains fail-closed:
+
+```text
+Authorization: Bearer <20+ token chars>
+password=<12+ non-space chars>
+passwd=<12+ non-space chars>
+secret=<12+ non-space chars>
+api_key=<12+ non-space chars>
+api-key=<12+ non-space chars>
+```
+
+Tests construct the real PEM header dynamically so the scanner's own test source does not contain a
+literal credential-looking header.
+
+## Bundle integrity
+
+The packager continues to enforce:
+
+- exact base→candidate ancestry;
+- before/after file capture;
+- rename/copy/delete handling;
+- path traversal denial;
+- execution evidence symlink denial;
+- secret scan over repository before/after and execution evidence;
+- document index;
+- SHA-256 closure;
+- ZIP without symlinks.
+
+No secret finding is downgraded to a warning.
+
+## Migration
+
+The old handoff script may remain for historical reproducibility, but new handoffs should call the
+repository packager from the verified candidate lineage.
+
+A temporary compatibility scanner is no longer needed once this candidate is in the lineage.

docs/superpowers/specs/2026-08-19-durable-owner-authorization.md

@@ -0,0 +1,173 @@
+# Durable Owner Authorization and capability projection
+
+## Problem
+
+AIOS already supports unattended worker execution after a Task is executable, but Owner intent was
+not durable at the capability layer. The same Owner repeatedly had to authorize or repair:
+
+- `governed-task:project`;
+- `release:candidate`.
+
+This caused valid one-time Owner intent to be lost between conversations and forced repeated human
+interaction.
+
+## Design
+
+Durable authorization is represented by a root-owned mode `0600` policy outside Git. The policy is
+persistent until explicitly revoked.
+
+The first persistable capabilities are deliberately narrow:
+
+```text
+governed-task:project
+release:candidate
+```
+
+They cover preparation work only.
+
+The following capabilities can never be inherited from this durable grant:
+
+```text
+submitOwnerApproval
+startDeployment
+tag_publish
+baseline_registry_finalize
+runtime_switch
+main_write
+production_deploy
+force_push
+credential_admin
+```
+
+Those remain separate high-risk release/Owner-signature gates.
+
+## Policy
+
+Example schema:
+
+```json
+{
+  "schema": "aios.owner-durable-authorization.v1",
+  "grant_id": "owner-permanent-preparation-20260819",
+  "owner_identity": "owner@example",
+  "account": "chatgpt-governed-implementer-singlehost",
+  "capabilities": [
+    "governed-task:project",
+    "release:candidate"
+  ],
+  "issued_at": "2026-08-19T05:00:00Z",
+  "not_before": "2026-08-19T05:00:00Z",
+  "persistent_until_revoked": true,
+  "constraints": {
+    "repository": "aios/aios",
+    "allowed_target_prefix": "agent/",
+    "release_version_prefix": "V2R2C",
+    "max_task_risk": "engineering"
+  }
+}
+```
+
+The repository never stores the live Owner policy or release credential.
+
+## Revocation
+
+A root-owned mode `0600` revocation file lists `revoked_grant_ids`.
+
+Revocation is checked on every operation, not only at service startup. A running broker therefore
+fails closed immediately after revocation without needing a new conversation.
+
+## Broker boundary
+
+The broker exposes only an AF_UNIX socket:
+
+```text
+/run/aios-owner-authorization-broker/broker.sock
+```
+
+The socket is mode `0660` and owned by the service's `aios-executor-exchange` group. This avoids introducing another long-lived bearer secret solely for local broker access.
+
+It exposes only:
+
+```text
+GET  /v1/capabilities
+POST /v1/project-governed-task
+POST /v1/authorize-release-candidate
+```
+
+`project-governed-task` calls only the canonical Bridge.
+
+`authorize-release-candidate` persists an Owner authorization decision only. It deliberately does
+**not** invent a local trusted-builder endpoint. The external release connector remains responsible
+for projecting the account's real `release:candidate` scope. Until that external projection exists,
+the decision state is `AUTHORIZED_PENDING_EXTERNAL_CAPABILITY_PROJECTION`.
+
+The broker holds only the existing host control credential needed for canonical governed-task
+projection. It never holds an Owner signing key or a release/deployment credential. The caller also supplies the expected account identifier over the Unix transport; the OS socket permission is the local access boundary.
+
+
+## Canonical Online Task projection
+
+A durable `governed-task:project` grant may create a complete canonical Online Task rather than a
+bare Task. It requires:
+
+```text
+require_online_task_projection=true
+generation=1
+branch_mode=CREATE_NEW
+max_attempts<=3
+risk<=engineering
+initial action=source_capture
+```
+
+Projected task capabilities are limited to:
+
+```text
+git.read
+git.test
+governed.write.branch
+governed.write.file
+```
+
+A write-capable task still begins as `source_capture`. Existing placeholder-deferral prevents the
+generic worker from claiming it until a deterministic plan is registered. Repository mutation
+therefore still requires the normal deterministic plan, WriterLease and governed commit path.
+
+This fixes the previous split-brain behavior where a bare Task could exist without the canonical
+governed execution projection.
+
+## Idempotency and audit
+
+Every broker operation requires an idempotency key. A receipt SQLite database stores only:
+
+- idempotency key;
+- capability;
+- request SHA-256;
+- response SHA-256;
+- sanitized upstream response;
+- timestamp.
+
+An exact retry returns the same receipt. Reusing an idempotency key with different content fails.
+
+## Unattended invariant
+
+Once the Owner policy and narrow service credentials are provisioned:
+
+```text
+Owner durable grant
+→ broker decision
+→ canonical governed task projection / release candidate preparation
+→ worker / trusted builder
+```
+
+No repeated Owner interaction is required for those preparation capabilities.
+
+Production execution remains:
+
+```text
+release candidate
+→ immutable approval envelope
+→ Owner signature
+→ deployment
+```
+
+The durable grant never substitutes for the Owner signature.

docs/superpowers/specs/2026-08-19-gpt-https-edge-activation.md

@@ -0,0 +1,114 @@
+# Owner-approved GPT HTTPS edge activation and rollback
+
+## Scope
+
+This controller is the final host-side control mechanism before the disabled-by-default GPT HTTPS
+edge may bind publicly.
+
+Repository implementation and public activation remain separate states:
+
+```text
+IMPLEMENTED
+≠
+OWNER_APPROVED
+≠
+ACTIVATED
+≠
+RUNTIME_VERIFIED
+```
+
+The controller never creates owner approval, TLS assets, DNS, firewall rules or ingress bearer
+credentials.
+
+## Approval binding
+
+The owner approval file is root-owned mode `0600` and has schema:
+
+```json
+{
+  "schema": "aios.gpt-https-edge.public-approval.v1",
+  "approval_id": "owner-public-edge-...",
+  "candidate_oid": "<exact governed edge candidate>",
+  "domain": "tasks.example.com",
+  "public_bind": true,
+  "bind_host": "0.0.0.0",
+  "port": 18443,
+  "tls_cert_sha256": "<exact certificate sha256>",
+  "issued_at": "2026-08-19T00:00:00Z",
+  "not_before": "2026-08-19T00:00:00Z",
+  "expires_at": "2026-08-20T00:00:00Z",
+  "owner": "<owner identity>"
+}
+```
+
+Approval lifetime is at most seven days. Candidate/domain/bind/port/certificate are immutable within
+one approval.
+
+Changing any of those requires a new owner approval.
+
+## Check
+
+```text
+aios_gpt_https_edge_activation.py check --candidate <GE>
+```
+
+Checks:
+
+- approval ownership/mode/schema/time window;
+- exact candidate binding;
+- exact certificate hash;
+- certificate/private-key pair;
+- certificate hostname;
+- certificate remaining lifetime;
+- loopback GPT ingress service health.
+
+It performs no public network mutation.
+
+## Activate
+
+`activate` additionally:
+
+- writes one exact systemd public-bind drop-in atomically;
+- enables/restarts only the HTTPS edge unit;
+- verifies effective ExecStart points at the immutable candidate release;
+- verifies effective public-bind/domain environment;
+- performs a TLS handshake to `127.0.0.1:18443` using the approved public hostname;
+- writes a durable activation receipt.
+
+Any failure after drop-in mutation automatically invokes rollback.
+
+Exact replay of an already-active matching approval returns the durable receipt without restarting
+or creating duplicate state.
+
+## Rollback
+
+`rollback`:
+
+- stops and disables only the HTTPS edge;
+- removes only the controller-owned public-bind drop-in;
+- daemon-reloads systemd;
+- verifies edge inactive;
+- writes a durable rollback receipt.
+
+It does not delete:
+
+- owner approval;
+- TLS certificate/private key;
+- ingress bearer token;
+- internal loopback ingress;
+- Task/Job state;
+- Git branches.
+
+Rollback is therefore replay-safe and does not require reconstructing secrets.
+
+## Not covered
+
+This controller intentionally does not own:
+
+- DNS records;
+- host firewall/NAT/security-group changes;
+- certificate issuance/renewal;
+- OpenAI Custom GPT domain verification;
+- public Internet acceptance from outside the host.
+
+Those remain separate owner/environment gates.

docs/superpowers/specs/2026-08-19-gpt-https-edge.md

@@ -0,0 +1,120 @@
+# Disabled-by-default HTTPS edge for external GPT Online Task access
+
+## Purpose
+
+The internal ingress (`127.0.0.1:18945`) remains loopback-only. This edge is the only component
+allowed to become Internet-facing, and only after a separate owner approval.
+
+It forwards only:
+
+```text
+POST /v1/read-only-tasks
+GET  /v1/read-only-tasks/{request_id}
+GET  /openapi.json
+GET  /health
+```
+
+It never proxies:
+
+```text
+18892 canonical Bridge
+18940 Task API
+WriterLease
+commit
+verification
+release
+tag
+baseline
+arbitrary URLs
+```
+
+## Default state
+
+The systemd unit is disabled by policy unless the owner creates:
+
+```text
+/etc/aios-gitops/gpt-https-edge.approved
+```
+
+The base unit still binds only to:
+
+```text
+127.0.0.1:18443
+AIOS_GPT_EDGE_PUBLIC_BIND_APPROVED=0
+```
+
+Non-loopback bind is rejected by code unless an owner-controlled drop-in sets:
+
+```text
+AIOS_GPT_EDGE_PUBLIC_BIND_APPROVED=1
+AIOS_GPT_EDGE_HOST=0.0.0.0
+AIOS_GPT_EDGE_PUBLIC_DOMAIN=<approved hostname>
+```
+
+This repository candidate does not create that drop-in or approval marker.
+
+## TLS
+
+TLS certificate/key are owner-managed systemd credentials:
+
+```text
+/etc/aios-gitops/gpt-https-edge-tls.crt
+/etc/aios-gitops/gpt-https-edge-tls.key
+```
+
+The service requires TLS 1.2 or newer, disables TLS compression, and never logs or returns key
+material.
+
+Certificate/key issuance and renewal are external owner responsibilities until a separately
+governed ACME workflow is implemented.
+
+## Request protections
+
+The edge provides:
+
+- 64 KiB request cap;
+- 256 KiB upstream response cap;
+- JSON validation before proxying;
+- fixed upstream `http://127.0.0.1:18945`;
+- Authorization-only header forwarding;
+- no Cookie/X-Forwarded-* trust;
+- route/method allowlist;
+- per-source-IP in-memory rate limiting;
+- bounded concurrent request handling;
+- optional exact Host-header enforcement;
+- HSTS / no-store / nosniff / no-referrer response headers.
+
+The rate limiter is intentionally simple and local. Multi-node/global abuse protection is outside
+the current one-host architecture.
+
+## Authentication
+
+The HTTPS edge does not create a new authentication scheme. It forwards the external Bearer
+credential to the loopback GPT ingress, which verifies `gpt_ingress_token`.
+
+The edge never receives the AIOS control-plane token.
+
+## Deployment boundary
+
+Public activation is a high-risk network operation and requires explicit owner approval. Before
+activation verify:
+
+```text
+host-side GPT ingress E2E=VERIFIED
+owner ingress token provisioned
+TLS certificate/key provisioned
+public hostname owned and DNS configured
+firewall/network exposure reviewed
+explicit public-bind approval recorded
+rollback plan available
+```
+
+Only after those gates may an owner-authorized runtime drop-in bind this service publicly.
+
+Until then:
+
+```text
+GPT_HTTPS_EDGE_IMPLEMENTATION=candidate
+PUBLIC_NETWORK_EXPOSURE=NONE
+PUBLIC_CUSTOM_GPT_DIRECT_AI0S_ACCESS=NOT_VERIFIED
+```

docs/superpowers/specs/2026-08-19-gpt-online-task-ingress.md

@@ -0,0 +1,138 @@
+# External GPT → Online Task ingress
+
+## Scope
+
+This component closes the current architectural gap between an external Custom GPT and AIOS
+without exposing the internal Task API (`18940`) or canonical Bridge (`18892`) directly.
+
+The first supported public-facing capability is intentionally read-only:
+
+```text
+authenticated external intent
+→ host-side ingress
+→ canonical /v2/governed-task
+→ canonical /v2/execution-envelope
+→ persistent generic worker
+→ source_capture result
+→ official control-plane verification
+```
+
+The ingress itself never claims, heartbeats, verifies, commits, acquires WriterLease, or writes
+protected refs.
+
+## Security boundary
+
+The service binds only to loopback (`127.0.0.1:18945`). A future owner-approved TLS reverse proxy
+or tunnel may expose this service externally. Direct non-loopback binding is denied by code.
+
+Two independent credentials are projected by systemd:
+
+- `gpt_ingress_token`: authenticates the external caller to this ingress.
+- `agent_task_control_token`: lets the ingress call the loopback AIOS control plane.
+
+Neither credential is stored in Git, SQLite, task payloads, results, logs, or API responses.
+
+The ingress has no repository filesystem access.
+
+## Caller-controlled fields
+
+Allowed:
+
+```text
+request_id
+goal
+instructions
+success_criteria
+constraints.max_files
+constraints.focus_paths
+constraints.notes
+```
+
+Not caller-controlled:
+
+```text
+repository
+base_oid
+expected_branch_oid
+target_branch
+branch_mode
+generation
+risk
+priority
+max_attempts
+capabilities
+action/adapter
+WriterLease
+worker identity
+credentials
+verification decision
+main/production/tag/baseline permissions
+```
+
+All external tasks are hard-bound to:
+
+```text
+repository=aios/aios
+capabilities=[git.read, git.test]
+action=source_capture
+writer_lease=false
+repository_mutation=forbidden
+control_plane_side_effects=forbidden
+credential_access=forbidden
+```
+
+## Idempotency
+
+`request_id` is the durable external idempotency boundary.
+
+The ingress stores only:
+
+```text
+request hash
+derived AIOS task id
+canonical job id
+base oid
+sanitized receipt
+state
+```
+
+An exact retry returns the same receipt. Reusing a `request_id` with different content returns
+`IDEMPOTENCY_CONFLICT`.
+
+A crash after canonical task creation can recover the same binding from the Task API rather than
+creating a replacement Task.
+
+## Status
+
+`GET /v1/read-only-tasks/{request_id}` returns only sanitized state:
+
+```text
+task/job ids
+status
+attempt
+whether claimed
+result classification
+official verification decision/classification
+source snapshot hash/count when available
+```
+
+It does not proxy raw Task metadata, execution envelopes, result text, credentials, or source files.
+
+## Public exposure
+
+This candidate does **not** make AIOS publicly reachable. Public exposure requires a separately
+approved HTTPS ingress/reverse proxy with:
+
+- TLS;
+- external authentication lifecycle;
+- rate limiting / abuse controls;
+- request size limits;
+- observability;
+- explicit owner approval for the network exposure.
+
+Until that step is verified:
+
+```text
+HOST_SIDE_EXTERNAL_GPT_ADAPTER=implemented/staging-candidate
+PUBLIC_CUSTOM_GPT_DIRECT_AI0S_ACCESS=not verified
+```

tests/test_all_modified_files_packager.py

@@ -0,0 +1,252 @@
+from __future__ import annotations
+
+from datetime import datetime, timezone
+import json
+from pathlib import Path
+import subprocess
+import sys
+import zipfile
+
+import pytest
+
+ROOT=Path(__file__).resolve().parents[1]
+APP=ROOT/"active-runtime/app"
+sys.path.insert(0,str(APP))
+
+import aios_all_modified_files_packager as packager
+
+
+def init_repo(tmp_path: Path) -> tuple[Path, Path, str]:
+    bare=tmp_path/"repo.git"
+    work=tmp_path/"work"
+    subprocess.run(["git","init","--bare",str(bare)],check=True,capture_output=True)
+    subprocess.run(["git","clone",str(bare),str(work)],check=True,capture_output=True)
+    subprocess.run(["git","-C",str(work),"config","user.email","test@example.invalid"],check=True)
+    subprocess.run(["git","-C",str(work),"config","user.name","AIOS Test"],check=True)
+    (work/"README.md").write_text("base\n",encoding="utf-8")
+    subprocess.run(["git","-C",str(work),"add","README.md"],check=True)
+    subprocess.run(["git","-C",str(work),"commit","-m","base"],check=True,capture_output=True)
+    subprocess.run(["git","-C",str(work),"push","origin","HEAD:refs/heads/main"],check=True,capture_output=True)
+    base=subprocess.check_output(["git","-C",str(work),"rev-parse","HEAD"],text=True).strip()
+    return bare,work,base
+
+
+def commit(work: Path, message: str) -> str:
+    subprocess.run(["git","-C",str(work),"add","-A"],check=True)
+    subprocess.run(["git","-C",str(work),"commit","-m",message],check=True,capture_output=True)
+    candidate=subprocess.check_output(["git","-C",str(work),"rev-parse","HEAD"],text=True).strip()
+    subprocess.run(["git","-C",str(work),"push","origin","HEAD:refs/heads/main"],check=True,capture_output=True)
+    return candidate
+
+
+def real_pem_header() -> bytes:
+    return b"-----" + b"BEGIN " + b"PRIVATE KEY" + b"-----"
+
+
+def bearer_value() -> bytes:
+    return b"Authorization" + b": Bearer " + (b"a"*40)
+
+
+def key_value_secret() -> bytes:
+    return b"api_key" + b"=" + (b"x"*24)
+
+
+def test_plain_private_key_words_are_not_secret_findings():
+    for value in (
+        b"BEGIN PRIVATE KEY",
+        b"PRIVATE KEY",
+        b"PEM private-key header",
+        b"assert 'BEGIN PRIVATE KEY' in message",
+    ):
+        assert packager.secret_finding(value) is None
+
+
+def test_complete_pem_private_key_header_is_rejected():
+    assert packager.secret_finding(real_pem_header())=="PEM_PRIVATE_KEY"
+    assert packager.secret_finding(
+        b"-----BEGIN OPENSSH " + b"PRIVATE KEY-----"
+    )=="PEM_PRIVATE_KEY"
+
+
+def test_bearer_and_key_value_secret_detection_remains_fail_closed():
+    assert packager.secret_finding(bearer_value())=="AUTHORIZATION_BEARER"
+    assert packager.secret_finding(key_value_secret())=="KEY_VALUE_SECRET"
+
+
+def test_short_or_descriptive_secret_words_do_not_false_positive():
+    assert packager.secret_finding(b"secret=short") is None
+    assert packager.secret_finding(b"password field is validated") is None
+    assert packager.secret_finding(b"api_key example placeholder") is None
+
+
+def test_end_to_end_bundle_accepts_literal_begin_private_key_text(tmp_path):
+    bare,work,base=init_repo(tmp_path)
+    (work/"tests").mkdir()
+    (work/"tests"/"test_text.py").write_text(
+        'ERROR = "BEGIN PRIVATE KEY is not a complete PEM header"\n',
+        encoding="utf-8",
+    )
+    (work/"docs").mkdir()
+    (work/"docs"/"note.md").write_text("Scanner regression.\n",encoding="utf-8")
+    candidate=commit(work,"candidate")
+
+    extra=tmp_path/"execution"
+    extra.mkdir()
+    (extra/"evidence.json").write_text('{"status":"PASS"}\n',encoding="utf-8")
+    out=tmp_path/"out"
+
+    summary=packager.build_bundle(
+        git_dir=bare,
+        base=base,
+        candidate=candidate,
+        task_id="r42-packager-test",
+        extra_dir=extra,
+        output_root=out,
+        now=datetime(2026,8,19,tzinfo=timezone.utc),
+    )
+
+    assert summary["changed_file_count"]==2
+    assert summary["execution_file_count"]==1
+    assert summary["repository_document_count"]==1
+    assert summary["execution_document_count"]==0
+    bundle_path=Path(summary["bundle_path"])
+    assert bundle_path.is_file()
+    assert packager.sha256(bundle_path.read_bytes())==summary["bundle_sha256"]
+
+    with zipfile.ZipFile(bundle_path) as archive:
+        names=set(archive.namelist())
+        assert "repo/after/tests/test_text.py" in names
+        assert "repo/after/docs/note.md" in names
+        assert "execution/evidence.json" in names
+        assert "documents-index.json" in names
+        assert "all-modified-files-manifest.json" in names
+        assert "checksums.sha256" in names
+        manifest=json.loads(archive.read("all-modified-files-manifest.json"))
+        assert manifest["schema"]=="aios.all-modified-files.bundle.v1"
+        assert manifest["candidate_oid"]==candidate
+
+    packager.verify_checksums(out/"bundle")
+
+
+def test_end_to_end_bundle_rejects_real_pem_header(tmp_path):
+    bare,work,base=init_repo(tmp_path)
+    (work/"credential.txt").write_bytes(
+        real_pem_header()+b"\nnot-a-real-key-body\n"
+    )
+    candidate=commit(work,"credential-like candidate")
+
+    with pytest.raises(
+        packager.PackagerError,
+        match="PEM_PRIVATE_KEY:after:credential.txt",
+    ):
+        packager.build_bundle(
+            git_dir=bare,
+            base=base,
+            candidate=candidate,
+            task_id="r42-packager-pem-reject",
+            output_root=tmp_path/"out",
+        )
+
+
+def test_execution_evidence_bearer_is_rejected(tmp_path):
+    bare,work,base=init_repo(tmp_path)
+    (work/"normal.txt").write_text("normal\n",encoding="utf-8")
+    candidate=commit(work,"normal")
+    extra=tmp_path/"execution"
+    extra.mkdir()
+    (extra/"unsafe.txt").write_bytes(bearer_value()+b"\n")
+
+    with pytest.raises(
+        packager.PackagerError,
+        match="AUTHORIZATION_BEARER:execution:unsafe.txt",
+    ):
+        packager.build_bundle(
+            git_dir=bare,
+            base=base,
+            candidate=candidate,
+            task_id="r42-packager-bearer-reject",
+            extra_dir=extra,
+            output_root=tmp_path/"out",
+        )
+
+
+def test_execution_symlink_is_rejected(tmp_path):
+    bare,work,base=init_repo(tmp_path)
+    (work/"normal.txt").write_text("normal\n",encoding="utf-8")
+    candidate=commit(work,"normal")
+    extra=tmp_path/"execution"
+    extra.mkdir()
+    target=extra/"target.txt"
+    target.write_text("safe\n",encoding="utf-8")
+    (extra/"link.txt").symlink_to(target)
+
+    with pytest.raises(
+        packager.PackagerError,
+        match="EXTRA_FILE_SYMLINK_REJECTED",
+    ):
+        packager.build_bundle(
+            git_dir=bare,
+            base=base,
+            candidate=candidate,
+            task_id="r42-packager-symlink-reject",
+            extra_dir=extra,
+            output_root=tmp_path/"out",
+        )
+
+
+def test_non_ancestor_lineage_is_rejected(tmp_path):
+    bare,work,base=init_repo(tmp_path)
+    (work/"one.txt").write_text("one\n",encoding="utf-8")
+    first=commit(work,"first")
+
+    subprocess.run(["git","-C",str(work),"checkout","--detach",base],check=True,capture_output=True)
+    (work/"other.txt").write_text("other\n",encoding="utf-8")
+    subprocess.run(["git","-C",str(work),"add","other.txt"],check=True)
+    subprocess.run(["git","-C",str(work),"commit","-m","other"],check=True,capture_output=True)
+    other=subprocess.check_output(["git","-C",str(work),"rev-parse","HEAD"],text=True).strip()
+    subprocess.run(
+        ["git","-C",str(work),"push","origin","HEAD:refs/heads/other"],
+        check=True,
+        capture_output=True,
+    )
+
+    with pytest.raises(packager.PackagerError,match="LINEAGE_INVALID"):
+        packager.validate_lineage(bare,first,other)
+
+
+def test_bundle_checksum_closure_detects_tamper(tmp_path):
+    bare,work,base=init_repo(tmp_path)
+    (work/"normal.txt").write_text("normal\n",encoding="utf-8")
+    candidate=commit(work,"normal")
+    out=tmp_path/"out"
+    packager.build_bundle(
+        git_dir=bare,
+        base=base,
+        candidate=candidate,
+        task_id="r42-packager-checksum",
+        output_root=out,
+    )
+    (out/"bundle"/"repo"/"after"/"normal.txt").write_text(
+        "tampered\n",
+        encoding="utf-8",
+    )
+    with pytest.raises(
+        packager.PackagerError,
+        match="CHECKSUM_CLOSURE_INVALID",
+    ):
+        packager.verify_checksums(out/"bundle")
+
+
+def test_source_does_not_contain_complete_pem_header_literal():
+    source=(ROOT/"active-runtime/app/aios_all_modified_files_packager.py").read_bytes()
+    tests=(ROOT/"tests/test_all_modified_files_packager.py").read_bytes()
+    complete=real_pem_header()
+    assert complete not in source
+    assert complete not in tests
+
+
+def test_document_path_contract():
+    assert packager.is_document_path("docs/spec.md")
+    assert packager.is_document_path("README")
+    assert packager.is_document_path("CHANGELOG.md")
+    assert not packager.is_document_path("active-runtime/app/main.py")

tests/test_durable_owner_authorization.py

@@ -0,0 +1,442 @@
+from __future__ import annotations
+
+from datetime import datetime, timezone
+import json
+from pathlib import Path
+import sqlite3
+import sys
+
+import pytest
+
+ROOT=Path(__file__).resolve().parents[1]
+APP=ROOT/"active-runtime/app"
+sys.path.insert(0,str(APP))
+
+import aios_durable_owner_authorization as auth
+
+
+NOW=datetime(2026,8,19,6,0,0,tzinfo=timezone.utc)
+ACCOUNT="chatgpt-governed-implementer-singlehost"
+
+
+def write_root_file(path: Path, value: object) -> Path:
+    path.parent.mkdir(parents=True,exist_ok=True)
+    path.write_text(json.dumps(value,sort_keys=True)+"\n",encoding="utf-8")
+    path.chmod(0o600)
+    return path
+
+
+def policy_value(**overrides):
+    value={
+        "schema":auth.POLICY_SCHEMA,
+        "grant_id":"owner-permanent-preparation-20260819",
+        "owner_identity":"owner-singlehost",
+        "account":ACCOUNT,
+        "capabilities":[
+            "governed-task:project",
+            "release:candidate",
+        ],
+        "issued_at":"2026-08-19T05:00:00Z",
+        "not_before":"2026-08-19T05:00:00Z",
+        "persistent_until_revoked":True,
+        "constraints":{
+            "repository":"aios/aios",
+            "allowed_target_prefix":"agent/",
+            "release_version_prefix":"V2R2C",
+            "max_task_risk":"engineering",
+        },
+    }
+    value.update(overrides)
+    return value
+
+
+def load_policy(tmp_path,monkeypatch,**overrides):
+    path=write_root_file(tmp_path/"policy.json",policy_value(**overrides))
+    revocations=write_root_file(
+        tmp_path/"revocations.json",
+        {"revoked_grant_ids":[]},
+    )
+    monkeypatch.setattr(auth,"validate_root_file",lambda path,**kwargs:path)
+    return auth.DurablePolicy.load(
+        path,
+        revocations,
+        now=NOW,
+    ),path,revocations
+
+
+def governed_request(**overrides):
+    value={
+        "task_id":"r42-owner-auth-projection-test-20260819",
+        "repository":"aios/aios",
+        "base_oid":"8"*40,
+        "expected_branch_oid":"8"*40,
+        "target_branch":"agent/r42-owner-auth-projection-test-20260819",
+        "generation":1,
+        "risk":"engineering",
+        "description":"project existing governed task",
+        "idempotency_key":"binding-owner-auth-projection-v1",
+        "branch_mode":"CREATE_NEW",
+        "priority":100,
+        "max_attempts":3,
+        "require_online_task_projection":True,
+        "online_task":{
+            "goal":"Analyze source read-only.",
+            "instructions":"Use source capture first.",
+            "success_criteria":[
+                {"id":"source","description":"Return source snapshot."}
+            ],
+            "constraints":{"repository_mutation":"forbidden"},
+            "required_capabilities":["git.read","git.test"],
+            "priority":100,
+            "max_attempts":3,
+        },
+        "execution":{
+            "action":"source_capture",
+            "adapter":"source_capture",
+            "capability_scope":["git.read","git.test"],
+            "source_signals":["durable authorization projection"],
+        },
+    }
+    value.update(overrides)
+    return value
+
+
+def release_request(**overrides):
+    value={
+        "candidate_id":"r42-pf1-release-candidate-20260819",
+        "source_commit":"8"*40,
+        "version":"V2R2C2",
+    }
+    value.update(overrides)
+    return value
+
+
+class FakeUpstream:
+    def __init__(self):
+        self.project_calls=[]
+
+    def project_governed_task(self,body):
+        self.project_calls.append(dict(body))
+        return 200,{
+            "task_id":body["task_id"],
+            "job_id":"job-"+"a"*32,
+            "state":"BOUND",
+        }
+
+
+def broker(tmp_path,monkeypatch):
+    policy,path,revocations=load_policy(tmp_path,monkeypatch)
+    store=auth.ReceiptStore(tmp_path/"receipts.sqlite3")
+    upstream=FakeUpstream()
+    instance=auth.Broker(
+        policy_path=path,
+        revocation_path=revocations,
+        store=store,
+        upstream=upstream,
+        expected_account=ACCOUNT,
+    )
+    monkeypatch.setattr(
+        auth.DurablePolicy,
+        "load",
+        classmethod(lambda cls,*args,**kwargs:policy),
+    )
+    return instance,upstream,store,policy
+
+
+def test_allowed_persistent_capabilities_are_only_preparation_scopes():
+    assert auth.ALLOWED_PERSISTENT_CAPABILITIES=={
+        "governed-task:project",
+        "release:candidate",
+    }
+    assert auth.ALLOWED_PERSISTENT_CAPABILITIES.isdisjoint(
+        auth.FORBIDDEN_PERSISTENT_CAPABILITIES
+    )
+
+
+@pytest.mark.parametrize(
+    "capability",
+    sorted(auth.FORBIDDEN_PERSISTENT_CAPABILITIES),
+)
+def test_policy_rejects_high_risk_persistent_capability(
+    tmp_path,monkeypatch,capability
+):
+    bad=policy_value(
+        capabilities=["governed-task:project",capability]
+    )
+    path=write_root_file(tmp_path/"policy.json",bad)
+    revocations=write_root_file(
+        tmp_path/"revocations.json",
+        {"revoked_grant_ids":[]},
+    )
+    monkeypatch.setattr(auth,"validate_root_file",lambda path,**kwargs:path)
+    with pytest.raises(
+        auth.AuthorizationError,
+        match="POLICY_CAPABILITY_NOT_PERSISTABLE",
+    ):
+        auth.DurablePolicy.load(path,revocations,now=NOW)
+
+
+def test_policy_is_persistent_until_revoked(tmp_path,monkeypatch):
+    policy,_,_=load_policy(tmp_path,monkeypatch)
+    assert policy.persistent_until_revoked is True
+    assert policy.account==ACCOUNT
+    assert policy.capabilities=={
+        "governed-task:project",
+        "release:candidate",
+    }
+
+
+def test_revocation_fails_closed(tmp_path,monkeypatch):
+    path=write_root_file(tmp_path/"policy.json",policy_value())
+    revocations=write_root_file(
+        tmp_path/"revocations.json",
+        {"revoked_grant_ids":["owner-permanent-preparation-20260819"]},
+    )
+    monkeypatch.setattr(auth,"validate_root_file",lambda path,**kwargs:path)
+    with pytest.raises(auth.AuthorizationError,match="OWNER_GRANT_REVOKED"):
+        auth.DurablePolicy.load(path,revocations,now=NOW)
+
+
+def test_future_policy_is_not_active(tmp_path,monkeypatch):
+    with pytest.raises(auth.AuthorizationError,match="POLICY_NOT_ACTIVE"):
+        load_policy(
+            tmp_path,
+            monkeypatch,
+            issued_at="2026-08-20T05:00:00Z",
+            not_before="2026-08-20T05:00:00Z",
+        )
+
+
+def test_governed_projection_request_is_tightly_constrained(
+    tmp_path,monkeypatch
+):
+    policy,_,_=load_policy(tmp_path,monkeypatch)
+    decision=policy.authorize(
+        "governed-task:project",
+        governed_request(),
+    )
+    assert decision["decision"]=="ALLOW"
+    assert decision["persistent_until_revoked"] is True
+
+    bad_caps=governed_request()
+    bad_caps["online_task"]={
+        **bad_caps["online_task"],
+        "required_capabilities":["release:candidate"],
+    }
+    bad_action=governed_request()
+    bad_action["execution"]={
+        **bad_action["execution"],
+        "action":"deterministic",
+        "adapter":"deterministic",
+    }
+    bad_scope=governed_request()
+    bad_scope["execution"]={
+        **bad_scope["execution"],
+        "capability_scope":["git.read"],
+    }
+    for bad in (
+        governed_request(repository="other/repo"),
+        governed_request(target_branch="main"),
+        governed_request(generation=2),
+        governed_request(risk="high"),
+        governed_request(expected_branch_oid="9"*40),
+        governed_request(branch_mode="CONTINUE_EXISTING"),
+        governed_request(priority=101),
+        governed_request(max_attempts=4),
+        governed_request(require_online_task_projection=False),
+        bad_caps,
+        bad_action,
+        bad_scope,
+    ):
+        with pytest.raises(auth.AuthorizationError):
+            policy.authorize("governed-task:project",bad)
+
+
+
+def test_write_capable_projection_is_allowed_only_as_source_capture_placeholder(
+    tmp_path,monkeypatch
+):
+    policy,_,_=load_policy(tmp_path,monkeypatch)
+    request=governed_request()
+    request["online_task"]={
+        **request["online_task"],
+        "required_capabilities":[
+            "git.read",
+            "git.test",
+            "governed.write.branch",
+            "governed.write.file",
+        ],
+    }
+    request["execution"]={
+        **request["execution"],
+        "capability_scope":[
+            "git.read",
+            "git.test",
+            "governed.write.branch",
+            "governed.write.file",
+        ],
+    }
+    decision=policy.authorize("governed-task:project",request)
+    assert decision["decision"]=="ALLOW"
+    assert request["execution"]["action"]=="source_capture"
+
+def test_release_candidate_request_is_tightly_constrained(
+    tmp_path,monkeypatch
+):
+    policy,_,_=load_policy(tmp_path,monkeypatch)
+    decision=policy.authorize("release:candidate",release_request())
+    assert decision["decision"]=="ALLOW"
+
+    with pytest.raises(
+        auth.AuthorizationError,
+        match="RELEASE_VERSION_NOT_AUTHORIZED",
+    ):
+        policy.authorize(
+            "release:candidate",
+            release_request(version="V3R1C1"),
+        )
+    with pytest.raises(
+        auth.AuthorizationError,
+        match="RELEASE_CANDIDATE_FIELDS_INVALID",
+    ):
+        policy.authorize(
+            "release:candidate",
+            {**release_request(),"execute":True},
+        )
+
+
+def test_governed_task_projection_executes_canonical_upstream(
+    tmp_path,monkeypatch
+):
+    instance,upstream,store,_=broker(tmp_path,monkeypatch)
+    request=governed_request()
+    response=instance.execute(
+        "governed-task:project",
+        request,
+        idempotency_key="owner-auth-projection-operation-0001",
+    )
+    assert len(upstream.project_calls)==1
+    assert upstream.project_calls[0]==request
+    assert response["operation_state"]=="EXECUTED"
+    assert response["external_capability_projection_required"] is False
+    assert response["separate_owner_release_approval_still_required"] is True
+
+
+def test_release_candidate_authorization_does_not_invent_builder_upstream(
+    tmp_path,monkeypatch
+):
+    instance,upstream,store,_=broker(tmp_path,monkeypatch)
+    response=instance.execute(
+        "release:candidate",
+        release_request(),
+        idempotency_key="owner-auth-release-candidate-operation-0001",
+    )
+    assert upstream.project_calls==[]
+    assert response["operation_state"]==(
+        "AUTHORIZED_PENDING_EXTERNAL_CAPABILITY_PROJECTION"
+    )
+    assert response["upstream_status"] is None
+    assert response["upstream_response"] is None
+    assert response["external_capability_projection_required"] is True
+    assert response["separate_owner_release_approval_still_required"] is True
+
+
+def test_exact_retry_returns_same_receipt_without_second_projection(
+    tmp_path,monkeypatch
+):
+    instance,upstream,_,_=broker(tmp_path,monkeypatch)
+    request=governed_request()
+    first=instance.execute(
+        "governed-task:project",
+        request,
+        idempotency_key="owner-auth-projection-operation-0002",
+    )
+    second=instance.execute(
+        "governed-task:project",
+        request,
+        idempotency_key="owner-auth-projection-operation-0002",
+    )
+    assert second==first
+    assert len(upstream.project_calls)==1
+
+
+def test_idempotency_conflict_fails_closed(tmp_path,monkeypatch):
+    instance,_,_,_=broker(tmp_path,monkeypatch)
+    instance.execute(
+        "release:candidate",
+        release_request(),
+        idempotency_key="owner-auth-release-operation-0002",
+    )
+    with pytest.raises(auth.AuthorizationError,match="IDEMPOTENCY_CONFLICT"):
+        instance.execute(
+            "release:candidate",
+            release_request(version="V2R2C3"),
+            idempotency_key="owner-auth-release-operation-0002",
+        )
+
+
+def test_receipt_store_contains_no_owner_policy_or_credentials(
+    tmp_path,monkeypatch
+):
+    instance,_,store,_=broker(tmp_path,monkeypatch)
+    instance.execute(
+        "release:candidate",
+        release_request(),
+        idempotency_key="owner-auth-release-operation-0003",
+    )
+    con=sqlite3.connect(store.path)
+    columns=[row[1] for row in con.execute("pragma table_info(receipts)")]
+    joined=" ".join(columns).lower()
+    assert "token" not in joined
+    assert "credential" not in joined
+    assert "policy_json" not in columns
+
+
+def test_systemd_broker_uses_unix_socket_and_only_host_control_credential():
+    unit=(
+        ROOT/"deploy/systemd/aios-owner-durable-authorization-broker.service"
+    ).read_text()
+    assert "LoadCredential=agent_task_control_token:" in unit
+    assert "release_candidate_token" not in unit
+    assert "18947" not in unit
+    assert "AIOS_RELEASE_CANDIDATE_API_URL" not in unit
+    assert "AIOS_OWNER_AUTH_BROKER_SOCKET=%t/aios-owner-authorization-broker/broker.sock" in unit
+    assert "Group=aios-executor-exchange" in unit
+    assert "RestrictAddressFamilies=AF_UNIX AF_INET" in unit
+    assert "IPAddressDeny=any" in unit
+    assert "IPAddressAllow=127.0.0.0/8" in unit
+    assert "InaccessiblePaths=/data/aios-gitops /home" in unit
+
+
+def test_source_contains_no_release_builder_endpoint_or_high_risk_execution():
+    source=(
+        ROOT/"active-runtime/app/aios_durable_owner_authorization.py"
+    ).read_text()
+    assert "/createReleaseCandidate" not in source
+    assert "18947" not in source
+    assert "ThreadingHTTPServer" not in source
+    assert "ThreadingUnixStreamServer" in source
+    assert "startDeployment" in source
+    assert "tag_publish" in source
+    assert "baseline_registry_finalize" in source
+    assert "runtime_switch" in source
+    assert "AUTHORIZED_PENDING_EXTERNAL_CAPABILITY_PROJECTION" in source
+    assert "SAFE_PROJECTED_TASK_CAPABILITIES" in source
+    assert "INITIAL_EXECUTION_ACTION_NOT_AUTHORIZED" in source
+
+
+
+def test_docs_avoid_new_broker_bearer_secret():
+    doc=(
+        ROOT/"docs/superpowers/specs/2026-08-19-durable-owner-authorization.md"
+    ).read_text()
+    assert "AF_UNIX" in doc
+    assert "avoids introducing another long-lived bearer secret" in doc
+
+def test_docs_preserve_separate_owner_release_approval():
+    doc=(
+        ROOT/"docs/superpowers/specs/2026-08-19-durable-owner-authorization.md"
+    ).read_text()
+    assert "persistent until explicitly revoked" in doc.lower()
+    assert "Owner signature" in doc
+    assert "AUTHORIZED_PENDING_EXTERNAL_CAPABILITY_PROJECTION" in doc

tests/test_external_agent_attempt_identity.py

@@ -0,0 +1,232 @@
+from __future__ import annotations
+
+import json
+from pathlib import Path
+import sys
+
+import pytest
+
+ROOT=Path(__file__).resolve().parents[1]
+APP=ROOT/"active-runtime/app"
+sys.path.insert(0,str(APP))
+
+import aios_online_agent_worker as worker
+from aios_external_agent_runtime import (
+    ExternalAgentExchange,
+    ExternalAgentRuntimeError,
+    validate_request,
+)
+
+
+def envelope():
+    value=object.__new__(worker.ExecutionEnvelope)
+    object.__setattr__(value,"task_id","task-attempt-identity")
+    object.__setattr__(value,"job_id","job-attempt-identity")
+    object.__setattr__(value,"digest","d"*64)
+    return value
+
+
+def test_external_request_id_is_stable_within_attempt_and_changes_across_attempts():
+    env=envelope()
+    a1=worker.external_agent_request_id(env,"hermes",6)
+    a1_replay=worker.external_agent_request_id(env,"hermes",6)
+    a2=worker.external_agent_request_id(env,"hermes",7)
+    assert a1==a1_replay
+    assert a1!=a2
+    assert a1.startswith("external-hermes-")
+    assert a2.startswith("external-hermes-")
+
+
+def test_external_request_id_rejects_invalid_attempt():
+    env=envelope()
+    with pytest.raises(worker.WorkerError,match="TASK_ATTEMPT_INVALID"):
+        worker.external_agent_request_id(env,"hermes",0)
+
+
+def test_external_request_schema_binds_attempt():
+    value=validate_request({
+        "schema":"aios.external-agent-runtime.request.v1",
+        "request_id":"external-hermes-0123456789abcdef",
+        "runtime":"hermes",
+        "task_id":"task-attempt-identity",
+        "attempt":7,
+        "execution_digest":"d"*64,
+        "payload":{
+            "goal":"Return ok.",
+            "instructions":"",
+            "success_criteria":[],
+            "constraints":{},
+        },
+    })
+    assert value["attempt"]==7
+
+
+def test_external_request_schema_rejects_missing_attempt():
+    with pytest.raises(ExternalAgentRuntimeError,match="EXTERNAL_AGENT_ATTEMPT_INVALID"):
+        validate_request({
+            "schema":"aios.external-agent-runtime.request.v1",
+            "request_id":"external-hermes-0123456789abcdef",
+            "runtime":"hermes",
+            "task_id":"task-attempt-identity",
+            "execution_digest":"d"*64,
+            "payload":{
+                "goal":"Return ok.",
+                "instructions":"",
+                "success_criteria":[],
+                "constraints":{},
+            },
+        })
+
+
+def _write_bound_result(path, request):
+    import hashlib
+    from aios_external_agent_runtime import canonical, sha256
+    material={
+        "schema":"aios.external-agent-runtime.result.v1",
+        "request_id":request["request_id"],
+        "request_sha256":sha256(canonical(request)),
+        "runtime":request["runtime"],
+        "task_id":request["task_id"],
+        "attempt":request["attempt"],
+        "execution_digest":request["execution_digest"],
+        "exit_code":0,
+        "duration_ms":1,
+        "output_text":"ok",
+        "output_sha256":hashlib.sha256(b"ok").hexdigest(),
+        "stderr_sha256":hashlib.sha256(b"").hexdigest(),
+        "credential_material_exposed":False,
+        "control_plane_access_granted":False,
+    }
+    material["result_sha256"]=sha256(canonical(material))
+    path.write_text(json.dumps(material,sort_keys=True),encoding="utf-8")
+
+
+def test_exchange_result_binding_rejects_wrong_attempt(tmp_path):
+    exchange=ExternalAgentExchange(tmp_path)
+    request={
+        "schema":"aios.external-agent-runtime.request.v1",
+        "request_id":"external-hermes-0123456789abcdef",
+        "runtime":"hermes",
+        "task_id":"task-attempt-identity",
+        "attempt":7,
+        "execution_digest":"d"*64,
+        "payload":{
+            "goal":"Return ok.",
+            "instructions":"",
+            "success_criteria":[],
+            "constraints":{},
+        },
+    }
+    request=validate_request(request)
+    result_path=tmp_path/"result.json"
+    _write_bound_result(result_path,request)
+    value=json.loads(result_path.read_text())
+    value["attempt"]=6
+    # Recompute digest so the binding check, not the digest check, is what fails.
+    from aios_external_agent_runtime import canonical,sha256
+    material=dict(value); material.pop("result_sha256")
+    value["result_sha256"]=sha256(canonical(material))
+    result_path.write_text(json.dumps(value),encoding="utf-8")
+    with pytest.raises(ExternalAgentRuntimeError,match="EXTERNAL_AGENT_RESULT_BINDING_INVALID"):
+        exchange._validated_result(
+            result_path,
+            request,
+            sha256(canonical(request)),
+        )
+
+
+def test_model_governance_request_id_is_attempt_scoped():
+    env=envelope()
+    attempt6=worker.external_agent_request_id(env,"hermes",6)+":model"
+    attempt7=worker.external_agent_request_id(env,"hermes",7)+":model"
+    assert attempt6!=attempt7
+
+
+def test_failed_previous_attempt_does_not_poison_next_model_governance_reservation(tmp_path):
+    from aios_model_governance import (
+        ModelGovernanceStore,
+        ModelRecord,
+        ProviderRecord,
+        RouteRequest,
+        ScopePolicy,
+    )
+
+    store=ModelGovernanceStore(tmp_path/"model-governance.sqlite3",now_fn=lambda:1_800_000_000.0)
+    store.register_provider(
+        ProviderRecord(
+            "oneapi",
+            "runtime-profile:hermes",
+            max_failures=5,
+            cooldown_seconds=60,
+        )
+    )
+    store.register_model(
+        ModelRecord(
+            "deepseek-v4-flash",
+            "oneapi",
+            "v1",
+            ("agent.runtime.hermes","text"),
+            input_cost_micros_per_1k=1,
+            output_cost_micros_per_1k=1,
+            max_input_tokens=10000,
+            max_output_tokens=1000,
+            evaluation_sha256="a"*64,
+        )
+    )
+    store.set_scope_policy(
+        ScopePolicy(
+            "tenant",
+            "model-governance-live",
+            "staging",
+            allowed_models=("deepseek-v4-flash",),
+            required_capabilities=("text",),
+            daily_token_budget=100000,
+            daily_cost_budget_micros=100000,
+            fallback_limit=1,
+        )
+    )
+
+    def route(request_id):
+        return store.route(
+            RouteRequest(
+                request_id=request_id,
+                task_id="r42-model-governance-hermes-live-20260818",
+                trace_id="trace-hermes-model-governance-live",
+                scope_type="tenant",
+                tenant_id="model-governance-live",
+                project_id="staging",
+                required_capabilities=("agent.runtime.hermes",),
+                estimated_input_tokens=100,
+                max_output_tokens=50,
+            )
+        )
+
+    attempt6=route("external-hermes-attempt6:model")
+    store.record_failure(attempt6.request_id,"provider_process_failed")
+    row6=store.conn.execute(
+        "select state from model_governance_reservations where request_id=?",
+        (attempt6.request_id,),
+    ).fetchone()
+    assert row6["state"]=="FAILED"
+
+    attempt7=route("external-hermes-attempt7:model")
+    assert attempt7.request_id!=attempt6.request_id
+    row7=store.conn.execute(
+        "select state from model_governance_reservations where request_id=?",
+        (attempt7.request_id,),
+    ).fetchone()
+    assert row7["state"]=="RESERVED"
+
+    settled=store.record_success(
+        attempt7.request_id,
+        input_tokens=90,
+        output_tokens=40,
+        usage_source="runtime",
+    )
+    assert settled["state"]=="SETTLED"
+    assert store.conn.execute(
+        "select state from model_governance_reservations where request_id=?",
+        (attempt6.request_id,),
+    ).fetchone()["state"]=="FAILED"
+    assert store.verify_audit_chain() is True
+    store.close()

tests/test_external_agent_online_runtime.py

@@ -110,6 +110,7 @@ def make_text_request(tmp_path, runtime, payload=None):
         "request_id":f"external-{runtime}-0123456789abcdef",
         "runtime":runtime,
         "task_id":f"task-{runtime}-runtime",
+        "attempt":1,
         "execution_digest":"d"*64,
         "payload":payload or {
             "goal":"Return ok.",
@@ -193,6 +194,7 @@ def test_external_executor_runs_root_owned_profile_without_shell(tmp_path,monkey
         "request_id":f"external-{runtime}-0123456789abcdef",
         "runtime":runtime,
         "task_id":f"task-{runtime}-runtime",
+        "attempt":1,
         "execution_digest":"d"*64,
         "payload":{
             "goal":"Return ok.",
@@ -326,6 +328,7 @@ def test_worker_external_adapter_builds_readonly_result(monkeypatch,tmp_path,run
                 "request_sha256":"e"*64,
                 "runtime":runtime,
                 "task_id":kwargs["task_id"],
+                "attempt":kwargs["attempt"],
                 "execution_digest":kwargs["execution_digest"],
                 "exit_code":0,
                 "duration_ms":10,
@@ -339,7 +342,7 @@ def test_worker_external_adapter_builds_readonly_result(monkeypatch,tmp_path,run
     monkeypatch.setattr(worker,"ExternalAgentExchange",FakeExchange)
     monkeypatch.setattr(worker,"audit",lambda *a,**k:None)
     executor=worker.HermesExecutor() if runtime=="hermes" else worker.OpenClawExecutor()
-    execution=executor.execute(envelope,tmp_path)
+    execution=executor.execute(envelope,tmp_path,attempt=1)
     result=worker.build_agent_runtime_result(envelope.task_id,envelope,execution)
     assert result["status"]=="SUCCEEDED"
     assert result["classification"]=="EXTERNAL_AGENT_RUNTIME_SUCCEEDED"

tests/test_gpt_https_edge.py

@@ -0,0 +1,205 @@
+from __future__ import annotations
+
+from email.message import Message
+import json
+from pathlib import Path
+import stat
+import sys
+
+import pytest
+
+ROOT=Path(__file__).resolve().parents[1]
+APP=ROOT/"active-runtime/app"
+sys.path.insert(0,str(APP))
+
+import aios_gpt_https_edge as edge
+
+
+class Headers(dict):
+    def get(self,key,default=None):
+        return super().get(key,default)
+
+
+def test_default_and_public_bind_require_explicit_owner_approval():
+    assert edge.public_bind_allowed("127.0.0.1","0")=="127.0.0.1"
+    assert edge.public_bind_allowed("::1","0")=="::1"
+    with pytest.raises(RuntimeError,match="OWNER_APPROVAL_REQUIRED"):
+        edge.public_bind_allowed("0.0.0.0","0")
+    with pytest.raises(RuntimeError,match="OWNER_APPROVAL_REQUIRED"):
+        edge.public_bind_allowed("::","")
+    assert edge.public_bind_allowed("0.0.0.0","1")=="0.0.0.0"
+
+
+def test_arbitrary_public_bind_address_is_denied_even_with_approval():
+    with pytest.raises(RuntimeError,match="BIND_ADDRESS_INVALID"):
+        edge.public_bind_allowed("192.0.2.10","1")
+
+
+def test_public_domain_validation():
+    assert edge.validated_public_domain("tasks.example.com")=="tasks.example.com"
+    assert edge.validated_public_domain("")==""
+    with pytest.raises(RuntimeError,match="PUBLIC_DOMAIN_INVALID"):
+        edge.validated_public_domain("https://example.com")
+    with pytest.raises(RuntimeError,match="PUBLIC_DOMAIN_INVALID"):
+        edge.validated_public_domain("localhost")
+
+
+def test_route_allowlist_accepts_only_readonly_api():
+    assert edge.allowed_route("POST","/v1/read-only-tasks")=="/v1/read-only-tasks"
+    assert edge.allowed_route("GET","/v1/read-only-tasks/external-request-0001")==(
+        "/v1/read-only-tasks/external-request-0001"
+    )
+    assert edge.allowed_route("GET","/health")=="/health"
+    assert edge.allowed_route("GET","/openapi.json")=="/openapi.json"
+    for method,path in [
+        ("POST","/v2/governed-task"),
+        ("GET","/v1/tasks/x"),
+        ("POST","/v1/read-only-tasks/x"),
+        ("DELETE","/v1/read-only-tasks/x"),
+    ]:
+        with pytest.raises(edge.EdgeError,match="ROUTE_DENY"):
+            edge.allowed_route(method,path)
+
+
+def test_query_and_fragment_are_denied():
+    with pytest.raises(edge.EdgeError,match="QUERY_OR_FRAGMENT_DENY"):
+        edge.allowed_route("GET","/health?x=1")
+    with pytest.raises(edge.EdgeError,match="QUERY_OR_FRAGMENT_DENY"):
+        edge.allowed_route("GET","/openapi.json#x")
+
+
+def test_forwarded_headers_are_minimal_and_require_bearer():
+    incoming=Headers({
+        "Authorization":"Bearer "+"a"*40,
+        "Content-Type":"application/json; charset=utf-8",
+        "Cookie":"secret=1",
+        "X-Forwarded-For":"203.0.113.10",
+        "X-Api-Key":"do-not-forward",
+    })
+    forwarded=edge.forwarded_headers(incoming,body=True)
+    assert forwarded=={
+        "Accept":"application/json",
+        "Authorization":"Bearer "+"a"*40,
+        "Content-Type":"application/json",
+    }
+    with pytest.raises(edge.EdgeError,match="AUTHORIZATION_REQUIRED"):
+        edge.forwarded_headers(Headers({}),body=False)
+
+
+def test_invalid_content_type_is_denied():
+    with pytest.raises(edge.EdgeError,match="CONTENT_TYPE_INVALID"):
+        edge.forwarded_headers(
+            Headers({
+                "Authorization":"Bearer "+"a"*40,
+                "Content-Type":"text/plain",
+            }),
+            body=True,
+        )
+
+
+def test_rate_limiter_caps_window_and_allows_expiry():
+    now=[100.0]
+    limiter=edge.RateLimiter(window_seconds=60,limit=2,burst=1,now_fn=lambda:now[0])
+    assert limiter.allowed("client")
+    assert limiter.allowed("client")
+    assert limiter.allowed("client")
+    assert not limiter.allowed("client")
+    now[0]=161.0
+    assert limiter.allowed("client")
+
+
+
+def test_rate_limiter_client_table_is_bounded():
+    now=[100.0]
+    limiter=edge.RateLimiter(
+        window_seconds=60,
+        limit=10,
+        burst=0,
+        max_clients=2,
+        now_fn=lambda:now[0],
+    )
+    assert limiter.allowed("a")
+    assert limiter.allowed("b")
+    assert not limiter.allowed("c")
+    now[0]=161.0
+    assert limiter.allowed("c")
+    assert len(limiter._events) <= 2
+
+
+def test_audit_source_does_not_log_sensitive_request_material():
+    source=(ROOT/"active-runtime/app/aios_gpt_https_edge.py").read_text()
+    assert '"authorization", "body", "token", "cookie", "client_ip"' in source
+    assert "_client_pseudonym" in source
+
+def test_proxy_upstream_is_fixed():
+    proxy=edge.Proxy()
+    assert proxy.upstream=="http://127.0.0.1:18945"
+    bad=edge.Proxy(upstream="http://127.0.0.1:18892")
+    with pytest.raises(edge.EdgeError,match="UPSTREAM_DRIFT_DENY"):
+        bad.request("GET","/health",{"Authorization":"Bearer "+"a"*40},None)
+
+
+
+def test_public_bind_requires_approved_domain():
+    assert edge.validated_edge_identity("127.0.0.1","0","")==("127.0.0.1","")
+    with pytest.raises(RuntimeError,match="PUBLIC_DOMAIN_REQUIRED"):
+        edge.validated_edge_identity("0.0.0.0","1","")
+    assert edge.validated_edge_identity(
+        "0.0.0.0","1","tasks.example.com"
+    )==("0.0.0.0","tasks.example.com")
+
+def test_tls_key_requires_no_group_or_other_permissions(tmp_path):
+    key=tmp_path/"key.pem"
+    key.write_text("x"*64,encoding="utf-8")
+    key.chmod(0o600)
+    assert edge.validate_tls_file(key,key=True)==key
+    key.chmod(0o644)
+    with pytest.raises(RuntimeError,match="TLS_KEY_MODE_INVALID"):
+        edge.validate_tls_file(key,key=True)
+
+
+def test_tls_symlink_is_denied(tmp_path):
+    real=tmp_path/"cert.pem"
+    real.write_text("x"*64,encoding="utf-8")
+    link=tmp_path/"link.pem"
+    link.symlink_to(real)
+    with pytest.raises(RuntimeError,match="TLS_FILE_INVALID"):
+        edge.validate_tls_file(link,key=False)
+
+
+def test_systemd_unit_is_disabled_by_owner_approval_condition_and_loopback_default():
+    unit=(ROOT/"deploy/systemd/aios-gpt-https-edge.service").read_text()
+    assert "ConditionPathExists=/etc/aios-gitops/gpt-https-edge.approved" in unit
+    assert "AIOS_GPT_EDGE_HOST=127.0.0.1" in unit
+    assert "AIOS_GPT_EDGE_PUBLIC_BIND_APPROVED=0" in unit
+    assert "AIOS_GPT_EDGE_PORT=18443" in unit
+    assert "LoadCredential=tls_cert:" in unit
+    assert "LoadCredential=tls_key:" in unit
+    assert "InaccessiblePaths=/data/aios-gitops" in unit
+    assert "InaccessiblePaths=" in unit and "/var/lib/aios-agent-task" in unit
+    assert "NoNewPrivileges=yes" in unit
+    assert "ProtectSystem=strict" in unit
+
+
+def test_systemd_unit_does_not_reference_internal_control_credentials():
+    unit=(ROOT/"deploy/systemd/aios-gpt-https-edge.service").read_text()
+    assert "agent_task_control" not in unit
+    assert "gitea" not in unit.lower()
+    assert "18892" not in unit
+    assert "18940" not in unit
+
+
+def test_edge_source_does_not_proxy_internal_bridge_or_task_api():
+    source=(ROOT/"active-runtime/app/aios_gpt_https_edge.py").read_text()
+    assert 'UPSTREAM = "http://127.0.0.1:18945"' in source
+    assert "18892" not in source
+    assert "18940" not in source
+    assert "/v2/governed-task" not in source
+    assert "/v1/tasks/" not in source
+
+
+def test_docs_keep_public_activation_separately_approved():
+    doc=(ROOT/"docs/superpowers/specs/2026-08-19-gpt-https-edge.md").read_text()
+    assert "explicit owner approval" in doc.lower()
+    assert "PUBLIC_NETWORK_EXPOSURE=NONE" in doc
+    assert "PUBLIC_CUSTOM_GPT_DIRECT_AI0S_ACCESS=NOT_VERIFIED" in doc

tests/test_gpt_https_edge_activation.py

@@ -0,0 +1,362 @@
+from __future__ import annotations
+
+from datetime import datetime, timedelta, timezone
+import hashlib
+import json
+from pathlib import Path
+import sys
+import subprocess
+
+import pytest
+
+ROOT=Path(__file__).resolve().parents[1]
+APP=ROOT/"active-runtime/app"
+sys.path.insert(0,str(APP))
+
+import aios_gpt_https_edge_activation as activation
+
+
+CANDIDATE="69a51d50c7f6cd9045346d41e338acebe4b6f1e8"
+NOW=datetime(2026,8,19,2,0,0,tzinfo=timezone.utc)
+
+
+def approval_value(cert_sha: str, **overrides):
+    value={
+        "schema":"aios.gpt-https-edge.public-approval.v1",
+        "approval_id":"owner-edge-approval-0001",
+        "candidate_oid":CANDIDATE,
+        "domain":"tasks.example.com",
+        "public_bind":True,
+        "bind_host":"0.0.0.0",
+        "port":18443,
+        "tls_cert_sha256":cert_sha,
+        "issued_at":"2026-08-19T01:00:00Z",
+        "not_before":"2026-08-19T01:00:00Z",
+        "expires_at":"2026-08-20T01:00:00Z",
+        "owner":"owner@example",
+    }
+    value.update(overrides)
+    return value
+
+
+def write_owner_file(path: Path, data: bytes, mode=0o600):
+    path.write_bytes(data)
+    path.chmod(mode)
+    return path
+
+
+def load_approval(tmp_path, monkeypatch, **overrides):
+    tmp_path.mkdir(parents=True,exist_ok=True)
+    cert=write_owner_file(tmp_path/"cert.pem",b"certificate-material-"*4,0o600)
+    value=approval_value(hashlib.sha256(cert.read_bytes()).hexdigest(),**overrides)
+    approval=write_owner_file(
+        tmp_path/"approval.json",
+        (json.dumps(value,sort_keys=True)+"\n").encode(),
+        0o600,
+    )
+    monkeypatch.setattr(
+        activation,
+        "validate_root_file",
+        lambda path, **kwargs: path,
+    )
+    return activation.Approval.load(
+        approval,
+        candidate_oid=CANDIDATE,
+        cert_path=cert,
+        now=NOW,
+    ),approval,cert
+
+
+class FakeSystemd:
+    def __init__(self):
+        self.active_units={activation.INGRESS_UNIT}
+        self.enabled_units=set()
+        self.calls=[]
+        self.exec_start=""
+        self.environment=""
+        self.fail_restart=False
+
+    def run(self,*args,check=True):
+        self.calls.append(args)
+        class Result:
+            returncode=0
+            stdout=""
+        if args[:2]==("restart",activation.EDGE_UNIT):
+            if self.fail_restart:
+                raise subprocess.CalledProcessError(1,args)
+            self.active_units.add(activation.EDGE_UNIT)
+        elif args[:2]==("stop",activation.EDGE_UNIT):
+            self.active_units.discard(activation.EDGE_UNIT)
+        elif args[:2]==("enable",activation.EDGE_UNIT):
+            self.enabled_units.add(activation.EDGE_UNIT)
+        elif args[:2]==("disable",activation.EDGE_UNIT):
+            self.enabled_units.discard(activation.EDGE_UNIT)
+        return Result()
+
+    def active(self,unit):
+        return unit in self.active_units
+
+    def enabled(self,unit):
+        return unit in self.enabled_units
+
+    def show(self,unit,prop):
+        if prop=="ExecStart":
+            return self.exec_start
+        if prop=="Environment":
+            return self.environment
+        return ""
+
+
+def test_approval_binds_candidate_domain_bind_port_and_cert(tmp_path,monkeypatch):
+    approval,_,cert=load_approval(tmp_path,monkeypatch)
+    assert approval.candidate_oid==CANDIDATE
+    assert approval.domain=="tasks.example.com"
+    assert approval.bind_host=="0.0.0.0"
+    assert approval.port==18443
+    assert approval.tls_cert_sha256==hashlib.sha256(cert.read_bytes()).hexdigest()
+    assert approval.expires_at>approval.not_before
+
+
+def test_approval_rejects_candidate_mismatch(tmp_path,monkeypatch):
+    cert=write_owner_file(tmp_path/"cert.pem",b"certificate-material-"*4)
+    value=approval_value(hashlib.sha256(cert.read_bytes()).hexdigest())
+    approval=write_owner_file(tmp_path/"approval.json",json.dumps(value).encode())
+    monkeypatch.setattr(activation,"validate_root_file",lambda path,**kwargs:path)
+    with pytest.raises(activation.ActivationError,match="CANDIDATE_MISMATCH"):
+        activation.Approval.load(
+            approval,
+            candidate_oid="0"*40,
+            cert_path=cert,
+            now=NOW,
+        )
+
+
+def test_approval_rejects_cert_hash_mismatch(tmp_path,monkeypatch):
+    cert=write_owner_file(tmp_path/"cert.pem",b"certificate-material-"*4)
+    value=approval_value("a"*64)
+    approval=write_owner_file(tmp_path/"approval.json",json.dumps(value).encode())
+    monkeypatch.setattr(activation,"validate_root_file",lambda path,**kwargs:path)
+    with pytest.raises(activation.ActivationError,match="CERT_SHA_MISMATCH"):
+        activation.Approval.load(
+            approval,
+            candidate_oid=CANDIDATE,
+            cert_path=cert,
+            now=NOW,
+        )
+
+
+def test_approval_rejects_expired_not_yet_valid_and_long_lived(tmp_path,monkeypatch):
+    with pytest.raises(activation.ActivationError,match="APPROVAL_EXPIRED"):
+        load_approval(
+            tmp_path/"expired",
+            monkeypatch,
+            issued_at="2026-08-17T01:00:00Z",
+            not_before="2026-08-17T01:00:00Z",
+            expires_at="2026-08-18T01:00:00Z",
+        )
+    # Use distinct temp roots because helper writes fixed names.
+
+    with pytest.raises(activation.ActivationError,match="NOT_YET_VALID"):
+        load_approval(
+            tmp_path/"future",
+            monkeypatch,
+            issued_at="2026-08-19T03:00:00Z",
+            not_before="2026-08-19T03:00:00Z",
+            expires_at="2026-08-20T03:00:00Z",
+        )
+    with pytest.raises(activation.ActivationError,match="VALIDITY_TOO_LONG"):
+        load_approval(
+            tmp_path/"long",
+            monkeypatch,
+            issued_at="2026-08-18T01:00:00Z",
+            not_before="2026-08-18T01:00:00Z",
+            expires_at="2026-08-26T01:00:00Z",
+        )
+
+
+
+def test_validate_release_requires_edge_and_controller_and_denies_symlink(tmp_path):
+    release=tmp_path/"release"
+    app=release/"active-runtime/app"
+    app.mkdir(parents=True)
+    edge_file=app/"aios_gpt_https_edge.py"
+    controller=app/"aios_gpt_https_edge_activation.py"
+    edge_file.write_text("edge\n")
+    controller.write_text("controller\n")
+    value=activation.validate_release(release)
+    assert value["edge"]==edge_file
+    assert value["controller"]==controller
+    controller.unlink()
+    with pytest.raises(activation.ActivationError,match="IMMUTABLE_RELEASE_INVALID"):
+        activation.validate_release(release)
+
+def test_render_dropin_is_exact_and_contains_no_secret(tmp_path,monkeypatch):
+    approval,_,_=load_approval(tmp_path,monkeypatch)
+    release=Path("/opt/aios/releases")/(CANDIDATE+"-gpt-https-edge-public")
+    text=activation.render_dropin(approval,release)
+    assert "AIOS_GPT_EDGE_HOST=0.0.0.0" in text
+    assert "AIOS_GPT_EDGE_PORT=18443" in text
+    assert "AIOS_GPT_EDGE_PUBLIC_BIND_APPROVED=1" in text
+    assert "AIOS_GPT_EDGE_PUBLIC_DOMAIN=tasks.example.com" in text
+    assert CANDIDATE in text
+    for forbidden in ("Bearer ","agent_task_control","tls_key","gpt_ingress_token"):
+        assert forbidden not in text
+
+
+def test_receipt_contains_binding_but_no_owner_secret_material(tmp_path,monkeypatch):
+    approval,_,_=load_approval(tmp_path,monkeypatch)
+    receipt=activation.receipt_for(
+        approval,
+        Path("/opt/aios/releases/test"),
+        state="ACTIVE",
+    )
+    assert receipt["approval_id"]==approval.approval_id
+    assert receipt["candidate_oid"]==CANDIDATE
+    assert receipt["domain"]=="tasks.example.com"
+    serialized=json.dumps(receipt)
+    assert "Bearer " not in serialized
+    assert "private" not in serialized.lower()
+
+
+def test_activate_is_idempotent_for_same_active_approval(tmp_path,monkeypatch):
+    approval,_,cert=load_approval(tmp_path,monkeypatch)
+    release=tmp_path/"release"
+    app=release/"active-runtime/app/aios_gpt_https_edge.py"
+    controller=release/"active-runtime/app/aios_gpt_https_edge_activation.py"
+    app.parent.mkdir(parents=True)
+    app.write_text("print('x')\n")
+    controller.write_text("print('controller')\n")
+    key=write_owner_file(tmp_path/"key.pem",b"private-key-material-"*4,0o600)
+    dropin=tmp_path/"edge.d"/activation.DROPIN_NAME
+    state=tmp_path/"state.json"
+    systemd=FakeSystemd()
+
+    monkeypatch.setattr(activation.os,"geteuid",lambda:0)
+    monkeypatch.setattr(activation,"require_ingress_ready",lambda _:None)
+    monkeypatch.setattr(activation,"validate_tls_pair",lambda *a,**k:{"cert_sha256":approval.tls_cert_sha256})
+    monkeypatch.setattr(activation,"tls_probe",lambda *a,**k:None)
+
+    systemd.exec_start=str(app)
+    systemd.environment=(
+        "AIOS_GPT_EDGE_HOST=0.0.0.0 "
+        "AIOS_GPT_EDGE_PORT=18443 "
+        "AIOS_GPT_EDGE_PUBLIC_BIND_APPROVED=1 "
+        "AIOS_GPT_EDGE_PUBLIC_DOMAIN=tasks.example.com"
+    )
+    first=activation.activate(
+        approval=approval,
+        release_root=release,
+        cert_path=cert,
+        key_path=key,
+        dropin_path=dropin,
+        state_path=state,
+        systemd=systemd,
+    )
+    calls_after_first=len(systemd.calls)
+    second=activation.activate(
+        approval=approval,
+        release_root=release,
+        cert_path=cert,
+        key_path=key,
+        dropin_path=dropin,
+        state_path=state,
+        systemd=systemd,
+    )
+    assert first["state"]=="ACTIVE"
+    assert second==json.loads(state.read_text())
+    assert len(systemd.calls)==calls_after_first
+
+
+def test_activate_failure_after_mutation_rolls_back(tmp_path,monkeypatch):
+    approval,_,cert=load_approval(tmp_path,monkeypatch)
+    release=tmp_path/"release"
+    app=release/"active-runtime/app/aios_gpt_https_edge.py"
+    controller=release/"active-runtime/app/aios_gpt_https_edge_activation.py"
+    app.parent.mkdir(parents=True)
+    app.write_text("print('x')\n")
+    controller.write_text("print('controller')\n")
+    key=write_owner_file(tmp_path/"key.pem",b"private-key-material-"*4,0o600)
+    dropin=tmp_path/"edge.d"/activation.DROPIN_NAME
+    state=tmp_path/"state.json"
+    systemd=FakeSystemd()
+    systemd.fail_restart=True
+
+    monkeypatch.setattr(activation.os,"geteuid",lambda:0)
+    monkeypatch.setattr(activation,"require_ingress_ready",lambda _:None)
+    monkeypatch.setattr(activation,"validate_tls_pair",lambda *a,**k:{"cert_sha256":approval.tls_cert_sha256})
+    monkeypatch.setattr(activation,"tls_probe",lambda *a,**k:None)
+
+    with pytest.raises(subprocess.CalledProcessError):
+        activation.activate(
+            approval=approval,
+            release_root=release,
+            cert_path=cert,
+            key_path=key,
+            dropin_path=dropin,
+            state_path=state,
+            systemd=systemd,
+        )
+    assert not dropin.exists()
+    assert not systemd.active(activation.EDGE_UNIT)
+    assert json.loads(state.read_text())["state"]=="ROLLED_BACK"
+
+
+def test_rollback_is_replay_safe_and_leaves_owner_assets_alone(tmp_path,monkeypatch):
+    approval_file=tmp_path/"approval.json"
+    cert=tmp_path/"cert.pem"
+    key=tmp_path/"key.pem"
+    for path in (approval_file,cert,key):
+        path.write_text("owner-asset"*4)
+    dropin=tmp_path/"edge.d"/activation.DROPIN_NAME
+    dropin.parent.mkdir(parents=True)
+    dropin.write_text("[Service]\n")
+    state=tmp_path/"state.json"
+    state.write_text(json.dumps({
+        "schema":activation.STATE_SCHEMA,
+        "state":"ACTIVE",
+        "candidate_oid":CANDIDATE,
+        "approval_id":"owner-edge-approval-0001",
+        "domain":"tasks.example.com",
+    }))
+    systemd=FakeSystemd()
+    systemd.active_units.add(activation.EDGE_UNIT)
+    systemd.enabled_units.add(activation.EDGE_UNIT)
+    monkeypatch.setattr(activation.os,"geteuid",lambda:0)
+
+    first=activation.rollback(
+        dropin_path=dropin,
+        state_path=state,
+        systemd=systemd,
+        reason="owner_requested",
+    )
+    second=activation.rollback(
+        dropin_path=dropin,
+        state_path=state,
+        systemd=systemd,
+        reason="owner_requested",
+    )
+    assert first["state"]=="ROLLED_BACK"
+    assert second["state"]=="ROLLED_BACK"
+    assert not dropin.exists()
+    assert not systemd.active(activation.EDGE_UNIT)
+    assert not systemd.enabled(activation.EDGE_UNIT)
+    for path in (approval_file,cert,key):
+        assert path.exists()
+
+
+def test_activation_source_never_creates_owner_approval_or_tls_secret():
+    source=(ROOT/"active-runtime/app/aios_gpt_https_edge_activation.py").read_text()
+    assert "secrets.token" not in source
+    assert "openssl" not in source.lower()
+    assert "BEGIN PRIVATE KEY" not in source
+    assert "gpt-https-edge.approved" in source
+    assert "gpt-https-edge-tls.key" in source
+
+
+def test_docs_separate_implementation_activation_and_runtime_verification():
+    doc=(ROOT/"docs/superpowers/specs/2026-08-19-gpt-https-edge-activation.md").read_text()
+    assert "IMPLEMENTED" in doc
+    assert "OWNER_APPROVED" in doc
+    assert "ACTIVATED" in doc
+    assert "RUNTIME_VERIFIED" in doc
+    assert "does not own" in doc.lower()

tests/test_gpt_online_task_ingress.py

@@ -0,0 +1,342 @@
+from __future__ import annotations
+
+import json
+from pathlib import Path
+import sqlite3
+import sys
+
+import pytest
+
+ROOT=Path(__file__).resolve().parents[1]
+APP=ROOT/"active-runtime/app"
+sys.path.insert(0,str(APP))
+
+from aios_gpt_online_task_ingress import (
+    IngressError,
+    IngressService,
+    Store,
+    READONLY_CAPABILITIES,
+    branch_for,
+    normalized_request,
+    openapi_document,
+    task_id_for,
+    validated_bind_host,
+)
+
+
+MAIN="3244f0f17f8845592ac94d585f5ebd65693c9f29"
+JOB="job-0123456789abcdef0123456789abcdef"
+
+
+def request(request_id="external-request-0001", **overrides):
+    value={
+        "schema":"aios.gpt-online-task-ingress.request.v1",
+        "request_id":request_id,
+        "goal":"Inspect the repository for durable worker state.",
+        "instructions":"Read only. Return a concise source analysis.",
+        "success_criteria":[
+            {"id":"analysis","description":"Return a source-backed analysis."}
+        ],
+        "constraints":{
+            "max_files":50,
+            "focus_paths":["active-runtime/app","tests"],
+            "notes":"Do not mutate repository state.",
+        },
+    }
+    value.update(overrides)
+    return value
+
+
+class FakeUpstream:
+    def __init__(self):
+        self.created=[]
+        self.projected=[]
+        self.tasks={}
+        self.snapshot={"authoritative_head_oid":MAIN}
+
+    def control_snapshot(self):
+        return dict(self.snapshot)
+
+    def create_governed_task(self, body):
+        self.created.append(dict(body))
+        task_id=body["task_id"]
+        job_id=JOB
+        self.tasks[task_id]={
+            "task_id":task_id,
+            "status":"PROJECTED",
+            "attempt":0,
+            "metadata":{
+                "execution":{
+                    "task_id":task_id,
+                    "job_id":job_id,
+                    "base_oid":body["base_oid"],
+                    "action":"source_capture",
+                    "capability_scope":list(READONLY_CAPABILITIES),
+                    "governance":{"writer_lease":False},
+                }
+            },
+        }
+        return 201,{"task_id":task_id,"job_id":job_id,"state":"BOUND"}
+
+    def execution_envelope(self, body):
+        self.projected.append(dict(body))
+        task_id=body["task_id"]
+        return 200,{
+            "schema":"aios.online-agent.execution.v2",
+            "task_id":task_id,
+            "job_id":body["job_id"],
+            "repository":"aios/aios",
+            "base_oid":MAIN,
+            "action":"source_capture",
+            "capability_scope":list(READONLY_CAPABILITIES),
+            "governance":{"writer_lease":False},
+        }
+
+    def get_task(self, task_id):
+        value=self.tasks.get(task_id)
+        return (404,{}) if value is None else (200,dict(value))
+
+
+def service(tmp_path):
+    upstream=FakeUpstream()
+    store=Store(tmp_path/"state.sqlite3")
+    return IngressService(upstream,store),upstream,store
+
+
+def test_request_schema_is_narrow_and_rejects_capability_or_execution_control():
+    value=normalized_request(request())
+    assert value["goal"].startswith("Inspect")
+    for field in ("capability_scope","required_capabilities","action","adapter","repository","base_oid"):
+        bad=request()
+        bad[field]="evil"
+        with pytest.raises(IngressError,match="UNKNOWN_FIELDS"):
+            normalized_request(bad)
+
+
+def test_constraints_are_bounded_and_path_traversal_is_denied():
+    bad=request(constraints={"focus_paths":["../etc/passwd"]})
+    with pytest.raises(IngressError,match="FOCUS_PATHS_INVALID"):
+        normalized_request(bad)
+    bad=request(constraints={"max_files":1000})
+    with pytest.raises(IngressError,match="MAX_FILES_INVALID"):
+        normalized_request(bad)
+
+
+def test_non_loopback_bind_is_denied():
+    assert validated_bind_host("127.0.0.1")=="127.0.0.1"
+    with pytest.raises(RuntimeError,match="GPT_INGRESS_NON_LOOPBACK_BIND_DENY"):
+        validated_bind_host("0.0.0.0")
+    with pytest.raises(RuntimeError,match="GPT_INGRESS_NON_LOOPBACK_BIND_DENY"):
+        validated_bind_host("192.0.2.10")
+
+
+def test_task_and_branch_ids_are_deterministic_and_bounded():
+    a=task_id_for("external-request-0001")
+    b=task_id_for("external-request-0001")
+    c=task_id_for("external-request-0002")
+    assert a==b
+    assert a!=c
+    assert a.startswith("gpt-readonly-")
+    branch=branch_for(a)
+    assert branch.startswith("agent/external-gpt-readonly/")
+    assert len(a)<180
+
+
+def test_create_translates_to_canonical_readonly_governed_task(tmp_path):
+    svc,upstream,store=service(tmp_path)
+    response=svc.create(request())
+    assert response["state"]=="BOUND"
+    assert response["write_access"] is False
+    assert response["official_verification_required"] is True
+    assert response["capabilities"]==list(READONLY_CAPABILITIES)
+
+    assert len(upstream.created)==1
+    body=upstream.created[0]
+    assert body["repository"]=="aios/aios"
+    assert body["base_oid"]==MAIN
+    assert body["expected_branch_oid"]==MAIN
+    assert body["branch_mode"]=="CREATE_NEW"
+    assert body["risk"]=="low"
+    assert body["priority"]==0
+    assert body["max_attempts"]==3
+    assert body["require_online_task_projection"] is True
+    assert body["online_task"]["required_capabilities"]==list(READONLY_CAPABILITIES)
+    assert body["online_task"]["constraints"]["repository_mutation"]=="forbidden"
+    assert body["online_task"]["constraints"]["credential_access"]=="forbidden"
+    assert body["execution"]["action"]=="source_capture"
+    assert body["execution"]["adapter"]=="source_capture"
+    assert body["execution"]["capability_scope"]==list(READONLY_CAPABILITIES)
+    assert "writer_lease" not in body["execution"]
+
+    assert len(upstream.projected)==1
+    projection=upstream.projected[0]
+    assert projection["task_id"]==response["task_id"]
+    assert projection["job_id"]==JOB
+    assert projection["required_capabilities"]==list(READONLY_CAPABILITIES)
+
+    row=store.get("external-request-0001")
+    assert row is not None
+    assert row["state"]=="BOUND"
+    assert row["job_id"]==JOB
+    assert row["base_oid"]==MAIN
+
+
+def test_exact_retry_is_idempotent_without_duplicate_bridge_call(tmp_path):
+    svc,upstream,_=service(tmp_path)
+    first=svc.create(request())
+    second=svc.create(request())
+    assert second==first
+    assert len(upstream.created)==1
+    assert len(upstream.projected)==1
+
+
+def test_same_request_id_with_different_content_is_conflict(tmp_path):
+    svc,upstream,_=service(tmp_path)
+    svc.create(request())
+    with pytest.raises(IngressError,match="IDEMPOTENCY_CONFLICT"):
+        svc.create(request(goal="Different goal."))
+    assert len(upstream.created)==1
+
+
+def test_pending_request_recovers_existing_canonical_task_after_crash(tmp_path):
+    svc,upstream,store=service(tmp_path)
+    raw=normalized_request(request())
+    task_id=task_id_for(raw["request_id"])
+    store.reserve(raw,task_id)
+
+    upstream.tasks[task_id]={
+        "task_id":task_id,
+        "status":"PROJECTED",
+        "attempt":0,
+        "metadata":{
+            "execution":{
+                "task_id":task_id,
+                "job_id":JOB,
+                "base_oid":MAIN,
+                "action":"source_capture",
+                "capability_scope":list(READONLY_CAPABILITIES),
+                "governance":{"writer_lease":False},
+            }
+        },
+    }
+    result=svc.create(request())
+    assert result["job_id"]==JOB
+    assert result["task_id"]==task_id
+    assert len(upstream.created)==0
+    row=store.get(raw["request_id"])
+    assert row["state"]=="BOUND"
+
+
+def test_status_is_sanitized_and_never_proxies_raw_execution_or_result_text(tmp_path):
+    svc,upstream,store=service(tmp_path)
+    created=svc.create(request())
+    task_id=created["task_id"]
+    upstream.tasks[task_id]={
+        "task_id":task_id,
+        "status":"SUCCEEDED",
+        "attempt":1,
+        "claimed_by":"online-engineering-worker-cn-network-mkt-generic-20260815",
+        "metadata":{
+            "execution":{
+                "job_id":JOB,
+                "base_oid":MAIN,
+                "action":"source_capture",
+                "capability_scope":list(READONLY_CAPABILITIES),
+                "governance":{"writer_lease":False},
+                "secret_marker":"do-not-return",
+            }
+        },
+        "result":{
+            "status":"SUCCEEDED",
+            "classification":"NEEDS_REWORK_SOURCE_CAPTURE",
+            "execution_mode":"source_capture",
+            "output_text":"raw source text that must not be proxied",
+            "source_snapshot":{
+                "payload_sha256":"a"*64,
+                "file_count":42,
+                "files":["secret/path"],
+            },
+        },
+        "verification":{
+            "decision":"SUCCEEDED",
+            "classification":"GENERIC_ONLINE_TASK_AUTO_CONSUMPTION_VERIFIED",
+            "private":"do-not-return",
+        },
+    }
+    status=svc.status("external-request-0001")
+    assert status["state"]=="SUCCEEDED"
+    assert status["attempt"]==1
+    assert status["claimed"] is True
+    assert status["official_verification"]=="SUCCEEDED"
+    assert status["source_snapshot_sha256"]=="a"*64
+    assert status["source_snapshot_file_count"]==42
+    serialized=json.dumps(status)
+    assert "raw source text" not in serialized
+    assert "secret_marker" not in serialized
+    assert "secret/path" not in serialized
+    assert "do-not-return" not in serialized
+
+
+def test_pending_status_can_recover_job_binding(tmp_path):
+    svc,upstream,store=service(tmp_path)
+    raw=normalized_request(request())
+    task_id=task_id_for(raw["request_id"])
+    store.reserve(raw,task_id)
+    upstream.tasks[task_id]={
+        "task_id":task_id,
+        "status":"PROJECTED",
+        "attempt":0,
+        "metadata":{
+            "execution":{
+                "job_id":JOB,
+                "base_oid":MAIN,
+                "action":"source_capture",
+                "capability_scope":list(READONLY_CAPABILITIES),
+                "governance":{"writer_lease":False},
+            }
+        },
+    }
+    value=svc.status(raw["request_id"])
+    assert value["job_id"]==JOB
+    assert store.get(raw["request_id"])["state"]=="BOUND"
+
+
+def test_openapi_exposes_only_create_readonly_and_status_operations():
+    doc=openapi_document()
+    assert doc["openapi"]=="3.1.0"
+    paths=doc["paths"]
+    assert set(paths)=={
+        "/v1/read-only-tasks",
+        "/v1/read-only-tasks/{request_id}",
+    }
+    rendered=json.dumps(doc)
+    for forbidden in ("writer-lease","commit","verify","production","tag","baseline"):
+        assert forbidden not in rendered.lower()
+
+
+def test_systemd_unit_is_loopback_credential_isolated_and_has_no_repository_access():
+    unit=(ROOT/"deploy/systemd/aios-gpt-online-task-ingress.service").read_text()
+    assert "DynamicUser=yes" in unit
+    assert "LoadCredential=gpt_ingress_token:" in unit
+    assert "LoadCredential=agent_task_control_token:" in unit
+    assert "AIOS_GPT_INGRESS_HOST=127.0.0.1" in unit
+    assert "AIOS_GOVERNED_WRITE_API_URL=http://127.0.0.1:18892" in unit
+    assert "AIOS_AGENT_TASK_API_URL=http://127.0.0.1:18940" in unit
+    assert "IPAddressDeny=any" in unit
+    assert "IPAddressAllow=127.0.0.0/8" in unit
+    assert "InaccessiblePaths=/data/aios-gitops /etc/aios-gitops /home" in unit
+    assert "ProtectSystem=strict" in unit
+    assert "NoNewPrivileges=yes" in unit
+    assert "MemoryDenyWriteExecute=yes" in unit
+    assert "0.0.0.0" not in unit
+
+
+def test_ingress_sqlite_contains_no_request_body_or_credentials(tmp_path):
+    svc,upstream,store=service(tmp_path)
+    svc.create(request(instructions="sensitive-looking but non-secret user text"))
+    con=sqlite3.connect(store.path)
+    columns=[row[1] for row in con.execute("pragma table_info(external_task_requests)")]
+    assert "request_json" not in columns
+    assert "goal" not in columns
+    assert "instructions" not in columns
+    assert "token" not in " ".join(columns).lower()

tests/test_model_governance.py

@@ -0,0 +1,325 @@
+from __future__ import annotations
+
+from pathlib import Path
+import sys
+
+import pytest
+
+ROOT=Path(__file__).resolve().parents[1]
+sys.path.insert(0,str(ROOT/"active-runtime/app"))
+
+from aios_model_governance import (
+    APPROVED,
+    ModelGovernanceError,
+    ModelGovernanceStore,
+    ModelRecord,
+    ProviderRecord,
+    RouteRequest,
+    ScopePolicy,
+)
+
+
+def setup_store(tmp_path, now=1_800_000_000.0):
+    store=ModelGovernanceStore(tmp_path/"model-governance.sqlite3",now_fn=lambda:now)
+    store.register_provider(
+        ProviderRecord(
+            "provider-a",
+            "runtime-profile:provider-a",
+            max_failures=1,
+            cooldown_seconds=120,
+        )
+    )
+    store.register_provider(
+        ProviderRecord(
+            "provider-b",
+            "runtime-profile:provider-b",
+            max_failures=2,
+            cooldown_seconds=60,
+        )
+    )
+    store.register_model(
+        ModelRecord(
+            "model-a",
+            "provider-a",
+            "2026-08-18",
+            ("agent.runtime.hermes","text"),
+            input_cost_micros_per_1k=100,
+            output_cost_micros_per_1k=200,
+            max_input_tokens=10000,
+            max_output_tokens=4000,
+            evaluation_state=APPROVED,
+            evaluation_sha256="a"*64,
+            priority=10,
+        )
+    )
+    store.register_model(
+        ModelRecord(
+            "model-b",
+            "provider-b",
+            "2026-08-17",
+            ("agent.runtime.hermes","text"),
+            input_cost_micros_per_1k=120,
+            output_cost_micros_per_1k=250,
+            max_input_tokens=10000,
+            max_output_tokens=4000,
+            evaluation_state=APPROVED,
+            evaluation_sha256="b"*64,
+            priority=20,
+        )
+    )
+    store.set_scope_policy(
+        ScopePolicy(
+            "tenant",
+            "tenant-a",
+            "project-a",
+            allowed_models=("model-a","model-b"),
+            required_capabilities=("text",),
+            max_input_tokens_per_request=5000,
+            max_output_tokens_per_request=2000,
+            daily_token_budget=12000,
+            daily_cost_budget_micros=4000,
+            fallback_limit=2,
+        )
+    )
+    return store
+
+
+def req(request_id="request-001", *, exclude=()):
+    return RouteRequest(
+        request_id=request_id,
+        task_id="task-a",
+        trace_id="trace-a",
+        scope_type="tenant",
+        tenant_id="tenant-a",
+        project_id="project-a",
+        required_capabilities=("agent.runtime.hermes",),
+        estimated_input_tokens=1000,
+        max_output_tokens=500,
+        preferred_models=(),
+        exclude_providers=exclude,
+    )
+
+
+def test_registry_rejects_raw_secret(tmp_path):
+    store=ModelGovernanceStore(tmp_path/"db.sqlite3")
+    with pytest.raises(ModelGovernanceError,match="MODEL_CREDENTIAL"):
+        store.register_provider(
+            ProviderRecord("provider-a","Bearer raw-secret-value")
+        )
+
+
+def test_model_registration_requires_evaluation_gate(tmp_path):
+    store=ModelGovernanceStore(tmp_path/"db.sqlite3")
+    store.register_provider(ProviderRecord("provider-a"))
+    with pytest.raises(ModelGovernanceError,match="MODEL_EVALUATION_GATE_DENY"):
+        store.register_model(
+            ModelRecord(
+                "model-a","provider-a","v1",("text",),
+                evaluation_state="PENDING",evaluation_sha256="a"*64
+            )
+        )
+
+
+def test_scope_policy_requires_version_with_pin(tmp_path):
+    store=ModelGovernanceStore(tmp_path/"db.sqlite3")
+    with pytest.raises(ModelGovernanceError,match="MODEL_POLICY_PIN_VERSION_REQUIRED"):
+        store.set_scope_policy(
+            ScopePolicy("tenant","t","p",pinned_model_id="model-a")
+        )
+
+
+def test_route_is_deterministic_version_pinned_and_idempotent(tmp_path):
+    store=setup_store(tmp_path)
+    first=store.route(req())
+    second=store.route(req())
+    assert first==second
+    assert first.model_id=="model-a"
+    assert first.model_version=="2026-08-18"
+    assert first.provider_credential_ref=="runtime-profile:provider-a"
+    assert first.policy_revision==1
+    assert first.model_revision==1
+    snapshot=store.usage_snapshot("tenant","tenant-a","project-a")
+    assert snapshot["reserved_input_tokens"]==1000
+    assert snapshot["reserved_output_tokens"]==500
+    assert snapshot["reserved_cost_micros"]>0
+
+
+def test_route_idempotency_payload_conflict_fails(tmp_path):
+    store=setup_store(tmp_path)
+    store.route(req())
+    changed=RouteRequest(
+        **{**req().__dict__,"estimated_input_tokens":2000}
+    )
+    with pytest.raises(ModelGovernanceError,match="MODEL_ROUTE_IDEMPOTENCY_CONFLICT"):
+        store.route(changed)
+
+
+def test_budget_reservation_is_atomic(tmp_path):
+    store=setup_store(tmp_path)
+    store.set_scope_policy(
+        ScopePolicy(
+            "tenant","tenant-a","project-a",
+            allowed_models=("model-a",),
+            required_capabilities=("text",),
+            daily_token_budget=1600,
+            daily_cost_budget_micros=10000,
+        ),
+        expected_revision=1,
+    )
+    store.route(req("request-1"))
+    with pytest.raises(ModelGovernanceError,match="MODEL_BUDGET_DENY"):
+        store.route(req("request-2"))
+
+
+def test_usage_settlement_is_idempotent_and_releases_reservation(tmp_path):
+    store=setup_store(tmp_path)
+    decision=store.route(req())
+    first=store.record_success(
+        decision.request_id,
+        input_tokens=900,
+        output_tokens=400,
+        usage_source="runtime",
+    )
+    second=store.record_success(
+        decision.request_id,
+        input_tokens=900,
+        output_tokens=400,
+        usage_source="runtime",
+    )
+    assert first["idempotent"] is False
+    assert second["idempotent"] is True
+    snapshot=store.usage_snapshot("tenant","tenant-a","project-a")
+    assert snapshot["settled_input_tokens"]==900
+    assert snapshot["settled_output_tokens"]==400
+    assert snapshot["reserved_input_tokens"]==0
+    assert snapshot["reserved_output_tokens"]==0
+
+
+def test_missing_runtime_usage_falls_back_to_reserved_upper_bound(tmp_path):
+    store=setup_store(tmp_path)
+    decision=store.route(req())
+    result=store.record_success(
+        decision.request_id,
+        input_tokens=None,
+        output_tokens=None,
+    )
+    assert result["usage_source"]=="reserved_upper_bound"
+    assert result["input_tokens"]==1000
+    assert result["output_tokens"]==500
+
+
+def test_usage_cannot_exceed_reserved_budget(tmp_path):
+    store=setup_store(tmp_path)
+    decision=store.route(req())
+    with pytest.raises(ModelGovernanceError,match="MODEL_USAGE_EXCEEDS_RESERVATION"):
+        store.record_success(
+            decision.request_id,
+            input_tokens=1001,
+            output_tokens=500,
+        )
+
+
+def test_provider_failure_opens_circuit_and_fallback_uses_other_provider(tmp_path):
+    store=setup_store(tmp_path)
+    original=req()
+    first=store.route(original)
+    assert first.provider_id=="provider-a"
+    failure=store.record_failure(first.request_id,"provider_timeout")
+    assert failure["open_until"]>0
+    fallback=store.fallback(original,first.request_id)
+    assert fallback.provider_id=="provider-b"
+    assert fallback.fallback_index==1
+
+
+def test_circuit_excludes_provider_for_new_request(tmp_path):
+    store=setup_store(tmp_path)
+    first=store.route(req("request-first"))
+    store.record_failure(first.request_id,"provider_timeout")
+    next_decision=store.route(req("request-next"))
+    assert next_decision.provider_id=="provider-b"
+
+
+def test_model_version_pin_is_enforced(tmp_path):
+    store=setup_store(tmp_path)
+    store.set_scope_policy(
+        ScopePolicy(
+            "tenant","tenant-a","project-a",
+            allowed_models=("model-a","model-b"),
+            pinned_model_id="model-b",
+            pinned_model_version="2026-08-17",
+            daily_token_budget=12000,
+            daily_cost_budget_micros=4000,
+        ),
+        expected_revision=1,
+    )
+    decision=store.route(req())
+    assert decision.model_id=="model-b"
+    assert decision.model_version=="2026-08-17"
+
+
+def test_scope_isolation_denies_missing_cross_scope_policy(tmp_path):
+    store=setup_store(tmp_path)
+    cross=RouteRequest(
+        request_id="request-cross",
+        task_id="task-a",
+        trace_id="trace-a",
+        scope_type="tenant",
+        tenant_id="tenant-b",
+        project_id="project-a",
+        required_capabilities=("agent.runtime.hermes",),
+        estimated_input_tokens=100,
+        max_output_tokens=50,
+    )
+    with pytest.raises(ModelGovernanceError,match="MODEL_SCOPE_POLICY_MISSING"):
+        store.route(cross)
+
+
+def test_audit_chain_detects_tampering(tmp_path):
+    store=setup_store(tmp_path)
+    decision=store.route(req())
+    store.record_success(decision.request_id,input_tokens=100,output_tokens=50)
+    assert store.verify_audit_chain() is True
+    store.conn.execute(
+        "UPDATE model_governance_audit SET payload_json='{}' WHERE sequence=1"
+    )
+    assert store.verify_audit_chain() is False
+
+
+def test_registry_and_policy_use_cas(tmp_path):
+    store=ModelGovernanceStore(tmp_path/"db.sqlite3")
+    store.register_provider(ProviderRecord("p"))
+    with pytest.raises(ModelGovernanceError,match="MODEL_PROVIDER_CAS_MISMATCH"):
+        store.register_provider(ProviderRecord("p"),expected_revision=7)
+    store.set_scope_policy(ScopePolicy("tenant","t","p"))
+    with pytest.raises(ModelGovernanceError,match="MODEL_POLICY_CAS_MISMATCH"):
+        store.set_scope_policy(
+            ScopePolicy("tenant","t","p"),
+            expected_revision=7,
+        )
+
+
+def test_system_policy_never_authorizes_missing_tenant_policy(tmp_path):
+    store=ModelGovernanceStore(tmp_path/"db.sqlite3")
+    store.register_provider(ProviderRecord("provider-a"))
+    store.register_model(
+        ModelRecord(
+            "model-a","provider-a","v1",("text",),
+            evaluation_sha256="a"*64,
+        )
+    )
+    store.set_scope_policy(
+        ScopePolicy("system",None,None,allowed_models=("model-a",))
+    )
+    request=RouteRequest(
+        request_id="request-tenant-no-policy",
+        task_id="task-a",
+        trace_id="trace-a",
+        scope_type="tenant",
+        tenant_id="tenant-a",
+        project_id="project-a",
+        required_capabilities=("text",),
+        estimated_input_tokens=10,
+        max_output_tokens=10,
+    )
+    with pytest.raises(ModelGovernanceError,match="MODEL_SCOPE_POLICY_MISSING"):
+        store.route(request)

tests/test_model_governance_admin.py

@@ -0,0 +1,82 @@
+from __future__ import annotations
+
+import json
+from pathlib import Path
+import sys
+
+ROOT=Path(__file__).resolve().parents[1]
+sys.path.insert(0,str(ROOT/"active-runtime/app"))
+
+from aios_model_governance_admin import apply_config
+from aios_model_governance import ModelGovernanceStore
+
+
+def config():
+    return {
+        "schema":"aios.model-governance.bootstrap.v1",
+        "providers":[{
+            "provider_id":"provider-a",
+            "credential_ref":"runtime-profile:provider-a",
+            "max_failures":2,
+            "cooldown_seconds":60,
+        }],
+        "models":[{
+            "model_id":"model-a",
+            "provider_id":"provider-a",
+            "version":"v1",
+            "capabilities":["text"],
+            "input_cost_micros_per_1k":1,
+            "output_cost_micros_per_1k":2,
+            "max_input_tokens":1000,
+            "max_output_tokens":100,
+            "evaluation_state":"APPROVED",
+            "evaluation_sha256":"a"*64,
+            "priority":10,
+        }],
+        "policies":[{
+            "scope_type":"legacy",
+            "tenant_id":None,
+            "project_id":None,
+            "allowed_models":["model-a"],
+            "required_capabilities":["text"],
+            "daily_token_budget":10000,
+            "daily_cost_budget_micros":1000,
+            "fallback_limit":0,
+        }],
+    }
+
+
+def test_admin_apply_is_idempotent(tmp_path):
+    cfg=tmp_path/"config.json"
+    cfg.write_text(json.dumps(config()),encoding="utf-8")
+    db=tmp_path/"db.sqlite3"
+    first=apply_config(db,cfg)
+    second=apply_config(db,cfg)
+    assert first["changes"]["providers"]==1
+    assert first["changes"]["models"]==1
+    assert first["changes"]["policies"]==1
+    assert second["changes"]["providers"]==0
+    assert second["changes"]["models"]==0
+    assert second["changes"]["policies"]==0
+    assert second["changes"]["unchanged"]==3
+    store=ModelGovernanceStore(db)
+    assert store.verify_audit_chain() is True
+    store.close()
+
+
+def test_admin_config_change_uses_cas_revision(tmp_path):
+    value=config()
+    cfg=tmp_path/"config.json"
+    cfg.write_text(json.dumps(value),encoding="utf-8")
+    db=tmp_path/"db.sqlite3"
+    apply_config(db,cfg)
+    value["providers"][0]["cooldown_seconds"]=120
+    cfg.write_text(json.dumps(value),encoding="utf-8")
+    result=apply_config(db,cfg)
+    assert result["changes"]["providers"]==1
+    store=ModelGovernanceStore(db)
+    row=store.conn.execute(
+        "select revision,cooldown_seconds from model_governance_providers where provider_id='provider-a'"
+    ).fetchone()
+    assert tuple(row)==(2,120)
+    store.close()

tests/test_model_governance_external_runtime.py

@@ -0,0 +1,554 @@
+from __future__ import annotations
+
+import json
+import os
+from pathlib import Path
+import sys
+
+import pytest
+
+ROOT=Path(__file__).resolve().parents[1]
+APP=ROOT/"active-runtime/app"
+sys.path.insert(0,str(APP))
+
+from aios_model_governance import (
+    ModelGovernanceStore,
+    ModelRecord,
+    ProviderRecord,
+    ScopePolicy,
+)
+from aios_online_agent_execution import GovernedExecutionBinding, project_online_task
+import aios_external_agent_runtime as runtime
+
+
+BASE="a"*40
+
+
+def binding(task_id):
+    return GovernedExecutionBinding.from_mapping({
+        "task_id":task_id,
+        "job_id":"job-"+task_id,
+        "repository":"aios/aios",
+        "base_oid":BASE,
+        "expected_branch_oid":BASE,
+        "target_branch":"agent/"+task_id,
+        "branch_mode":"CREATE_NEW",
+        "generation":1,
+        "idempotency_key":task_id+"-binding-v1",
+    })
+
+
+def projected(task_id="task-model-governed", *, scope_type="tenant", tenant_id="tenant-a", project_id="project-a"):
+    request={
+        "task_id":task_id,
+        "goal":"Return the requested sentinel.",
+        "instructions":"No repository or control-plane side effects.",
+        "success_criteria":[{"id":"done","description":"Return a result."}],
+        "constraints":{"repository_mutation":"forbidden"},
+        "scope_type":scope_type,
+        "tenant_id":tenant_id,
+        "project_id":project_id,
+        "trace_id":"trace-model-a",
+        "required_capabilities":["agent.runtime.hermes"],
+    }
+    return project_online_task(
+        binding(task_id),
+        request,
+        {
+            "action":"agent_runtime",
+            "adapter":"hermes",
+            "capability_scope":["agent.runtime.hermes"],
+        },
+    )
+
+
+def helper(tmp_path):
+    path=tmp_path/"model_runtime.py"
+    path.write_text(
+        "import json,sys\n"
+        "model=sys.argv[1]\n"
+        "provider=sys.argv[2]\n"
+        "payload=json.load(sys.stdin)\n"
+        "if model=='model-a' and payload.get('constraints',{}).get('force_fail_a'):\n"
+        "    raise SystemExit(7)\n"
+        "print(json.dumps({'answer':'ok','model':model,'provider':provider,"
+        "'usage':{'input_tokens':20,'output_tokens':10,'cost_micros':2}},sort_keys=True))\n",
+        encoding="utf-8",
+    )
+    return path
+
+
+def write_profile(tmp_path, helper_path, *, providers=("provider-a","provider-b")):
+    root=tmp_path/"profiles";root.mkdir(exist_ok=True)
+    value={
+        "schema":"aios.external-agent-runtime.profile.v1",
+        "runtime":"hermes",
+        "argv":["/usr/bin/python3",str(helper_path),"{model_id}","{provider_id}"],
+        "input_mode":"stdin_json",
+        "output_mode":"stdout_json",
+        "timeout_seconds":30,
+        "max_output_bytes":65536,
+        "environment":{},
+        "credential_env":{},
+        "model_governance":{
+            "required":True,
+            "provider_ids":list(providers),
+            "model_capabilities":["text"],
+            "max_output_tokens":100,
+        },
+    }
+    (root/"hermes.json").write_text(json.dumps(value),encoding="utf-8")
+    return root
+
+
+def seed(db_path, *, legacy=False, tiny_budget=False):
+    store=ModelGovernanceStore(db_path,now_fn=lambda:1_800_000_000.0)
+    for provider in ("provider-a","provider-b"):
+        store.register_provider(
+            ProviderRecord(
+                provider,
+                f"runtime-profile:{provider}",
+                max_failures=1 if provider=="provider-a" else 2,
+                cooldown_seconds=120,
+            )
+        )
+    store.register_model(
+        ModelRecord(
+            "model-a","provider-a","v1",
+            ("agent.runtime.hermes","text"),
+            input_cost_micros_per_1k=10,
+            output_cost_micros_per_1k=20,
+            max_input_tokens=5000,max_output_tokens=1000,
+            evaluation_sha256="a"*64,priority=1,
+        )
+    )
+    store.register_model(
+        ModelRecord(
+            "model-b","provider-b","v2",
+            ("agent.runtime.hermes","text"),
+            input_cost_micros_per_1k=12,
+            output_cost_micros_per_1k=22,
+            max_input_tokens=5000,max_output_tokens=1000,
+            evaluation_sha256="b"*64,priority=2,
+        )
+    )
+    if legacy:
+        policy=ScopePolicy(
+            "legacy",None,None,
+            allowed_models=("model-a","model-b"),
+            required_capabilities=("text",),
+            daily_token_budget=10000,
+            daily_cost_budget_micros=10000,
+            fallback_limit=2,
+        )
+    else:
+        policy=ScopePolicy(
+            "tenant","tenant-a","project-a",
+            allowed_models=("model-a","model-b"),
+            required_capabilities=("text",),
+            daily_token_budget=1 if tiny_budget else 10000,
+            daily_cost_budget_micros=10000,
+            fallback_limit=2,
+        )
+    store.set_scope_policy(policy)
+    store.close()
+
+
+def make_external_request(projected_task, *, force_fail=False, attempt=1):
+    execution=projected_task["metadata"]["execution"]
+    payload=dict(execution["runtime_input"])
+    if force_fail:
+        payload["constraints"]=dict(payload["constraints"])
+        payload["constraints"]["force_fail_a"]=True
+    return {
+        "schema":"aios.external-agent-runtime.request.v1",
+        "request_id":"external-model-governance-0123456789",
+        "runtime":"hermes",
+        "task_id":projected_task["task_id"],
+        "attempt":attempt,
+        "execution_digest":execution["execution_digest"],
+        "payload":payload,
+    }
+
+
+@pytest.fixture(autouse=True)
+def clean_env(monkeypatch):
+    for key in list(os.environ):
+        upper=key.upper()
+        if any(marker in upper for marker in ("TOKEN","SECRET","PASSWORD","CREDENTIAL","AUTHORIZATION","PRIVATE_KEY","API_KEY")):
+            monkeypatch.delenv(key,raising=False)
+    monkeypatch.setenv("CREDENTIALS_DIRECTORY","/definitely-not-present-model-governance")
+
+
+def test_execution_projection_binds_trusted_scope_and_trace():
+    task=projected()
+    runtime_input=task["metadata"]["execution"]["runtime_input"]
+    assert runtime_input["scope"]=={
+        "scope_type":"tenant",
+        "tenant_id":"tenant-a",
+        "project_id":"project-a",
+    }
+    assert runtime_input["trace_id"]=="trace-model-a"
+
+
+def test_legacy_scope_remains_explicit_not_fake_tenant():
+    task=projected(scope_type="legacy",tenant_id=None,project_id=None)
+    scope=task["metadata"]["execution"]["runtime_input"]["scope"]
+    assert scope=={"scope_type":"legacy","tenant_id":None,"project_id":None}
+
+
+def test_profile_requires_model_binding_when_governance_required(tmp_path):
+    path=tmp_path/"helper.py";path.write_text("print('ok')\n")
+    root=tmp_path/"profiles";root.mkdir()
+    (root/"hermes.json").write_text(json.dumps({
+        "schema":"aios.external-agent-runtime.profile.v1",
+        "runtime":"hermes",
+        "argv":["/usr/bin/python3",str(path)],
+        "input_mode":"stdin_text",
+        "output_mode":"stdout_text",
+        "model_governance":{"required":True,"provider_ids":["provider-a"],"max_output_tokens":100},
+    }))
+    with pytest.raises(runtime.ExternalAgentRuntimeError,match="EXTERNAL_AGENT_MODEL_BINDING_REQUIRED"):
+        runtime.RuntimeProfile.load("hermes",root)
+
+
+def test_governed_runtime_selects_model_and_settles_usage(tmp_path,monkeypatch):
+    helper_path=helper(tmp_path)
+    profile_root=write_profile(tmp_path,helper_path)
+    db=tmp_path/"governance.sqlite3";seed(db)
+    monkeypatch.setattr(runtime,"MODEL_GOVERNANCE_ENABLED",True)
+    monkeypatch.setattr(runtime,"DEFAULT_MODEL_GOVERNANCE_DB",db)
+
+    request_path=tmp_path/"request.json"
+    result_path=tmp_path/"result.json"
+    request_path.write_text(json.dumps(make_external_request(projected())),encoding="utf-8")
+    runtime.execute_request(request_path,result_path,profile_root=profile_root)
+    result=json.loads(result_path.read_text())
+    mg=result["model_governance"]
+    assert mg["model_id"]=="model-a"
+    assert mg["model_version"]=="v1"
+    assert mg["provider_id"]=="provider-a"
+    assert mg["settlement"]["usage_source"]=="runtime"
+    assert mg["settlement"]["input_tokens"]==20
+    assert mg["settlement"]["output_tokens"]==10
+    assert mg["audit_chain_valid"] is True
+    assert result["output_json"]["model"]=="model-a"
+
+
+def test_budget_denies_before_runtime_process(tmp_path,monkeypatch):
+    helper_path=helper(tmp_path)
+    profile_root=write_profile(tmp_path,helper_path)
+    db=tmp_path/"governance.sqlite3";seed(db,tiny_budget=True)
+    monkeypatch.setattr(runtime,"MODEL_GOVERNANCE_ENABLED",True)
+    monkeypatch.setattr(runtime,"DEFAULT_MODEL_GOVERNANCE_DB",db)
+
+    req=tmp_path/"request.json";res=tmp_path/"result.json"
+    req.write_text(json.dumps(make_external_request(projected())),encoding="utf-8")
+    with pytest.raises(runtime.ExternalAgentRuntimeError,match="MODEL_BUDGET_DENY"):
+        runtime.execute_request(req,res,profile_root=profile_root)
+    assert not res.exists()
+
+
+def test_provider_failure_routes_to_fallback_provider(tmp_path,monkeypatch):
+    helper_path=helper(tmp_path)
+    profile_root=write_profile(tmp_path,helper_path)
+    db=tmp_path/"governance.sqlite3";seed(db)
+    monkeypatch.setattr(runtime,"MODEL_GOVERNANCE_ENABLED",True)
+    monkeypatch.setattr(runtime,"DEFAULT_MODEL_GOVERNANCE_DB",db)
+
+    req=tmp_path/"request.json";res=tmp_path/"result.json"
+    req.write_text(json.dumps(make_external_request(projected(),force_fail=True)),encoding="utf-8")
+    runtime.execute_request(req,res,profile_root=profile_root)
+    result=json.loads(res.read_text())
+    mg=result["model_governance"]
+    assert len(mg["attempts"])==2
+    assert mg["attempts"][0]["provider_id"]=="provider-a"
+    assert mg["attempts"][1]["provider_id"]=="provider-b"
+    assert mg["provider_id"]=="provider-b"
+    assert result["output_json"]["provider"]=="provider-b"
+
+
+def test_profile_provider_allowlist_fails_closed(tmp_path,monkeypatch):
+    helper_path=helper(tmp_path)
+    profile_root=write_profile(tmp_path,helper_path,providers=("provider-b",))
+    db=tmp_path/"governance.sqlite3";seed(db)
+    monkeypatch.setattr(runtime,"MODEL_GOVERNANCE_ENABLED",True)
+    monkeypatch.setattr(runtime,"DEFAULT_MODEL_GOVERNANCE_DB",db)
+
+    req=tmp_path/"request.json";res=tmp_path/"result.json"
+    req.write_text(json.dumps(make_external_request(projected())),encoding="utf-8")
+    with pytest.raises(runtime.ExternalAgentRuntimeError,match="EXTERNAL_AGENT_MODEL_PROVIDER_PROFILE_DENY"):
+        runtime.execute_request(req,res,profile_root=profile_root)
+
+
+def test_governance_required_but_runtime_disabled_fails_closed(tmp_path,monkeypatch):
+    helper_path=helper(tmp_path)
+    profile_root=write_profile(tmp_path,helper_path)
+    monkeypatch.setattr(runtime,"MODEL_GOVERNANCE_ENABLED",False)
+    req=tmp_path/"request.json";res=tmp_path/"result.json"
+    req.write_text(json.dumps(make_external_request(projected())),encoding="utf-8")
+    with pytest.raises(runtime.ExternalAgentRuntimeError,match="EXTERNAL_AGENT_MODEL_GOVERNANCE_DISABLED"):
+        runtime.execute_request(req,res,profile_root=profile_root)
+
+
+def test_execution_request_cannot_override_authoritative_task_scope():
+    task_id="task-scope-override"
+    request={
+        "task_id":task_id,
+        "goal":"Return result.",
+        "instructions":"",
+        "success_criteria":[],
+        "constraints":{},
+        "scope_type":"tenant",
+        "tenant_id":"tenant-a",
+        "project_id":"project-a",
+        "trace_id":"trace-authoritative",
+        "required_capabilities":["agent.runtime.hermes"],
+    }
+    task=project_online_task(
+        binding(task_id),
+        request,
+        {
+            "action":"agent_runtime",
+            "adapter":"hermes",
+            "capability_scope":["agent.runtime.hermes"],
+            "runtime_input":{
+                "goal":"Return result.",
+                "instructions":"",
+                "success_criteria":[],
+                "constraints":{},
+                "scope":{
+                    "scope_type":"tenant",
+                    "tenant_id":"tenant-b",
+                    "project_id":"project-b",
+                },
+                "trace_id":"trace-forged",
+            },
+        },
+    )
+    runtime_input=task["metadata"]["execution"]["runtime_input"]
+    assert runtime_input["scope"]["tenant_id"]=="tenant-a"
+    assert runtime_input["scope"]["project_id"]=="project-a"
+    assert runtime_input["trace_id"]=="trace-authoritative"
+
+
+def test_openclaw_nested_usage_is_extracted_exactly():
+    value={
+        "meta":{
+            "agentMeta":{
+                "usage":{"input":7770,"output":13,"total":7783},
+                "lastCallUsage":{"input":7770,"output":13,"total":7783},
+                "costUsd":0,
+            }
+        }
+    }
+    assert runtime._extract_runtime_usage(value)==(7770,13,0)
+
+
+def test_governance_required_profile_needs_provider_allowlist_and_output_limit(tmp_path):
+    path=tmp_path/"helper.py";path.write_text("print('ok')\\n")
+    root=tmp_path/"profiles";root.mkdir()
+    (root/"hermes.json").write_text(json.dumps({
+        "schema":"aios.external-agent-runtime.profile.v1",
+        "runtime":"hermes",
+        "argv":["/usr/bin/python3",str(path),"{model_id}"],
+        "input_mode":"stdin_text",
+        "output_mode":"stdout_text",
+        "model_governance":{"required":True},
+    }))
+    with pytest.raises(
+        runtime.ExternalAgentRuntimeError,
+        match="EXTERNAL_AGENT_MODEL_PROVIDER_IDS_REQUIRED",
+    ):
+        runtime.RuntimeProfile.load("hermes",root)
+
+
+def _openclaw_context_helper(tmp_path, *, input_tokens=7770, output_tokens=13):
+    path=tmp_path/"openclaw_context_helper.py"
+    path.write_text(
+        "import json,sys\n"
+        "model=sys.argv[1]\n"
+        "provider=sys.argv[2]\n"
+        "json.load(sys.stdin)\n"
+        f"print(json.dumps({{'answer':'ok','model':model,'provider':provider,"
+        f"'meta':{{'agentMeta':{{'usage':{{'input':{input_tokens},'output':{output_tokens},"
+        f"'total':{input_tokens+output_tokens}}},'costUsd':0}}}}}},sort_keys=True))\n",
+        encoding="utf-8",
+    )
+    return path
+
+
+def _openclaw_floor_profile(tmp_path, helper_path, *, floor=16384):
+    root=tmp_path/"openclaw-profiles"
+    root.mkdir(exist_ok=True)
+    value={
+        "schema":"aios.external-agent-runtime.profile.v1",
+        "runtime":"openclaw",
+        "argv":["/usr/bin/python3",str(helper_path),"{model_id}","{provider_id}"],
+        "input_mode":"stdin_json",
+        "output_mode":"stdout_json",
+        "timeout_seconds":30,
+        "max_output_bytes":65536,
+        "environment":{},
+        "credential_env":{},
+        "model_governance":{
+            "required":True,
+            "provider_ids":["openai"],
+            "model_capabilities":["text"],
+            "max_output_tokens":256,
+            "input_reservation_floor_tokens":floor,
+        },
+    }
+    (root/"openclaw.json").write_text(json.dumps(value),encoding="utf-8")
+    return root
+
+
+def _seed_openclaw_floor_db(db_path):
+    store=ModelGovernanceStore(db_path,now_fn=lambda:1_800_000_000.0)
+    store.register_provider(
+        ProviderRecord(
+            "openai",
+            "runtime-profile:openclaw",
+            max_failures=2,
+            cooldown_seconds=60,
+        )
+    )
+    store.register_model(
+        ModelRecord(
+            "openai/deepseek-v4-flash",
+            "openai",
+            "v1",
+            ("agent.runtime.openclaw","text"),
+            input_cost_micros_per_1k=0,
+            output_cost_micros_per_1k=0,
+            max_input_tokens=32768,
+            max_output_tokens=256,
+            evaluation_sha256="c"*64,
+            priority=10,
+        )
+    )
+    store.set_scope_policy(
+        ScopePolicy(
+            "tenant",
+            "model-governance-live",
+            "staging",
+            allowed_models=("openai/deepseek-v4-flash",),
+            required_capabilities=("text",),
+            max_input_tokens_per_request=32768,
+            max_output_tokens_per_request=256,
+            daily_token_budget=250000,
+            daily_cost_budget_micros=0,
+            fallback_limit=1,
+        )
+    )
+    store.close()
+
+
+def _openclaw_floor_request(*, attempt=2):
+    return {
+        "schema":"aios.external-agent-runtime.request.v1",
+        "request_id":f"external-openclaw-floor-attempt-{attempt:02d}",
+        "runtime":"openclaw",
+        "task_id":"r42-model-governance-openclaw-live-20260818",
+        "attempt":attempt,
+        "execution_digest":"d"*64,
+        "payload":{
+            "goal":"Return AIOS_MODEL_GOVERNANCE_OPENCLAW_OK.",
+            "instructions":"No side effects.",
+            "success_criteria":[],
+            "constraints":{},
+            "scope":{
+                "scope_type":"tenant",
+                "tenant_id":"model-governance-live",
+                "project_id":"staging",
+            },
+            "trace_id":"trace-openclaw-model-governance-live",
+        },
+    }
+
+
+def test_openclaw_profile_reservation_floor_covers_injected_context(
+    tmp_path,monkeypatch
+):
+    helper_path=_openclaw_context_helper(tmp_path,input_tokens=7770,output_tokens=13)
+    profile_root=_openclaw_floor_profile(tmp_path,helper_path,floor=16384)
+    db=tmp_path/"openclaw-governance.sqlite3"
+    _seed_openclaw_floor_db(db)
+    monkeypatch.setattr(runtime,"MODEL_GOVERNANCE_ENABLED",True)
+    monkeypatch.setattr(runtime,"DEFAULT_MODEL_GOVERNANCE_DB",db)
+
+    request_path=tmp_path/"request.json"
+    result_path=tmp_path/"result.json"
+    request_path.write_text(json.dumps(_openclaw_floor_request()),encoding="utf-8")
+    runtime.execute_request(request_path,result_path,profile_root=profile_root)
+
+    result=json.loads(result_path.read_text())
+    mg=result["model_governance"]
+    assert mg["input_reservation_floor_tokens"]==16384
+    assert mg["prompt_estimated_input_tokens"] < 16384
+    assert mg["reserved_estimated_input_tokens"]==16384
+    assert mg["settlement"]["state"]=="SETTLED"
+    assert mg["settlement"]["usage_source"]=="runtime"
+    assert mg["settlement"]["input_tokens"]==7770
+    assert mg["settlement"]["output_tokens"]==13
+
+    store=ModelGovernanceStore(db)
+    row=store.conn.execute(
+        """
+        select reserved_input_tokens,state,actual_input_tokens,actual_output_tokens
+          from model_governance_reservations
+         where request_id=?
+        """,
+        (result["request_id"]+":model",),
+    ).fetchone()
+    assert int(row["reserved_input_tokens"])==16384
+    assert row["state"]=="SETTLED"
+    assert int(row["actual_input_tokens"])==7770
+    assert int(row["actual_output_tokens"])==13
+    store.close()
+
+
+def test_openclaw_reservation_floor_still_fails_closed_above_floor(
+    tmp_path,monkeypatch
+):
+    helper_path=_openclaw_context_helper(tmp_path,input_tokens=17000,output_tokens=13)
+    profile_root=_openclaw_floor_profile(tmp_path,helper_path,floor=16384)
+    db=tmp_path/"openclaw-governance.sqlite3"
+    _seed_openclaw_floor_db(db)
+    monkeypatch.setattr(runtime,"MODEL_GOVERNANCE_ENABLED",True)
+    monkeypatch.setattr(runtime,"DEFAULT_MODEL_GOVERNANCE_DB",db)
+
+    request_path=tmp_path/"request.json"
+    result_path=tmp_path/"result.json"
+    request_path.write_text(json.dumps(_openclaw_floor_request()),encoding="utf-8")
+    with pytest.raises(
+        runtime.ExternalAgentRuntimeError,
+        match="MODEL_USAGE_EXCEEDS_RESERVATION",
+    ):
+        runtime.execute_request(request_path,result_path,profile_root=profile_root)
+    assert not result_path.exists()
+
+
+def test_model_governance_profile_rejects_invalid_input_reservation_floor(tmp_path):
+    path=tmp_path/"helper.py"
+    path.write_text("print('ok')\n",encoding="utf-8")
+    root=tmp_path/"profiles"
+    root.mkdir()
+    (root/"openclaw.json").write_text(json.dumps({
+        "schema":"aios.external-agent-runtime.profile.v1",
+        "runtime":"openclaw",
+        "argv":["/usr/bin/python3",str(path),"{model_id}"],
+        "input_mode":"stdin_text",
+        "output_mode":"stdout_text",
+        "model_governance":{
+            "required":True,
+            "provider_ids":["openai"],
+            "max_output_tokens":256,
+            "input_reservation_floor_tokens":262145,
+        },
+    }))
+    with pytest.raises(
+        runtime.ExternalAgentRuntimeError,
+        match="EXTERNAL_AGENT_MODEL_INPUT_RESERVATION_FLOOR_INVALID",
+    ):
+        runtime.RuntimeProfile.load("openclaw",root)

tests/test_model_governance_runtime_config.py

@@ -0,0 +1,99 @@
+from __future__ import annotations
+
+import json
+from pathlib import Path
+
+ROOT=Path(__file__).resolve().parents[1]
+
+
+def test_executor_has_single_model_governance_sqlite_write_path():
+    unit=(ROOT/"deploy/systemd/aios-external-agent-executor.service").read_text()
+    assert "AIOS_MODEL_GOVERNANCE_DB=/var/lib/aios-model-governance/state.sqlite3" in unit
+    assert "AIOS_MODEL_GOVERNANCE_ENABLED=1" in unit
+    assert "ReadWritePaths=/var/lib/aios-model-governance" in unit
+    assert "ReadWritePaths=/var/lib/aios-external-agent-executor" in unit
+    assert "InaccessiblePaths=/etc/aios-gitops" in unit
+    assert "InaccessiblePaths=/data/aios-gitops" in unit
+
+
+def test_tmpfiles_owner_matches_external_executor():
+    line=(ROOT/"deploy/tmpfiles.d/aios-model-governance.conf").read_text().strip()
+    assert line == (
+        "d /var/lib/aios-model-governance 0750 "
+        "aios-engineering-executor aios-executor-exchange -"
+    )
+
+
+def test_staging_models_use_real_e2e_evaluation_evidence():
+    value=json.loads(
+        (ROOT/"config/model-governance/model-governance.staging.json").read_text()
+    )
+    assert value["schema"]=="aios.model-governance.bootstrap.v1"
+    models={item["model_id"]:item for item in value["models"]}
+    assert models["deepseek-v4-flash"]["evaluation_sha256"] == (
+        "43ae0db1640daeb6d33e2787769732844f366d36070065d3a2239905bd9a823b"
+    )
+    assert models["openai/deepseek-v4-flash"]["evaluation_sha256"] == (
+        "7b1c89634eb9c0e7cb5a3d75323d904ebbd91b3ae5494ce938bd8316a4cf5b12"
+    )
+
+
+def test_runtime_profiles_bind_model_from_governance_not_task():
+    for name in ("hermes","openclaw"):
+        value=json.loads(
+            (ROOT/"config/external-agent-runtime.d"/f"{name}.profile.json").read_text()
+        )
+        governance=value["model_governance"]
+        assert governance["required"] is True
+        assert governance["provider_ids"]
+        assert governance["max_output_tokens"] > 0
+        serialized=json.dumps(value["argv"])+json.dumps(value.get("environment",{}))
+        assert "{model_id}" in serialized or "{provider_id}" in serialized
+
+
+def test_staging_bootstrap_has_explicit_tenant_live_policy():
+    value=json.loads(
+        (ROOT/"config/model-governance/model-governance.staging.json").read_text()
+    )
+    keys={
+        (
+            item["scope_type"],
+            item.get("tenant_id"),
+            item.get("project_id"),
+        )
+        for item in value["policies"]
+    }
+    assert ("legacy",None,None) in keys
+    assert ("tenant","model-governance-live","staging") in keys
+
+
+def test_staging_policy_input_limit_covers_root_owned_openclaw_reservation_floor():
+    config=json.loads(
+        (ROOT/"config/model-governance/model-governance.staging.json").read_text()
+    )
+    profile=json.loads(
+        (ROOT/"config/external-agent-runtime.d/openclaw.profile.json").read_text()
+    )
+    floor=int(profile["model_governance"]["input_reservation_floor_tokens"])
+    assert floor==16384
+    openclaw_model="openai/deepseek-v4-flash"
+    relevant=[
+        policy for policy in config["policies"]
+        if openclaw_model in policy["allowed_models"]
+    ]
+    assert relevant
+    for policy in relevant:
+        assert int(policy["max_input_tokens_per_request"]) >= floor
+
+
+def test_staging_openclaw_scope_limit_is_least_privilege_floor_not_model_ceiling():
+    config=json.loads(
+        (ROOT/"config/model-governance/model-governance.staging.json").read_text()
+    )
+    models={item["model_id"]:item for item in config["models"]}
+    model=models["openai/deepseek-v4-flash"]
+    assert int(model["max_input_tokens"])==32768
+    for policy in config["policies"]:
+        if "openai/deepseek-v4-flash" in policy["allowed_models"]:
+            assert int(policy["max_input_tokens_per_request"])==16384
+            assert int(policy["max_input_tokens_per_request"]) < int(model["max_input_tokens"])

tests/test_online_agent_agent_runtime_replay.py

@@ -0,0 +1,85 @@
+from __future__ import annotations
+
+from pathlib import Path
+import sys
+
+ROOT = Path(__file__).resolve().parents[1]
+sys.path.insert(0, str(ROOT / "active-runtime/app"))
+
+from aios_online_agent_execution import (  # noqa: E402
+    GovernedExecutionBinding,
+    project_execution_envelope,
+)
+import aios_online_agent_worker as worker  # noqa: F401,E402
+from aios_online_agent_worker import WorkerSupervisor  # noqa: E402
+
+
+BASE = "a" * 40
+
+
+def test_agent_runtime_replays_the_durable_envelope_to_bridge(monkeypatch) -> None:
+    monkeypatch.setattr(
+        worker,
+        "WORKER_CAPABILITIES",
+        tuple(sorted((*worker.WORKER_CAPABILITIES, "agent.runtime.hermes"))),
+    )
+    binding = GovernedExecutionBinding.from_mapping(
+        {
+            "task_id": "r42-agent-runtime-replay-test-20260818",
+            "job_id": "job-agent-runtime-replay-test",
+            "repository": "aios/aios",
+            "base_oid": BASE,
+            "expected_branch_oid": BASE,
+            "target_branch": "agent/agent-runtime-replay-test",
+            "branch_mode": "CREATE_NEW",
+            "generation": 1,
+            "idempotency_key": "agent-runtime-replay-test-v1",
+        }
+    )
+    execution = project_execution_envelope(
+        binding,
+        {
+            "action": "agent_runtime",
+            "adapter": "hermes",
+            "capability_scope": ["agent.runtime.hermes"],
+            "change_set": [],
+            "runtime_input": {
+                "goal": "replay the exact durable runtime envelope",
+                "instructions": "return the bounded sentinel",
+                "success_criteria": [
+                    {
+                        "id": "sentinel",
+                        "description": "return the bounded sentinel",
+                    }
+                ],
+                "constraints": {
+                    "provider_boundary": "provider-credential-only",
+                    "change_set": [],
+                    "writer_lease": False,
+                },
+            },
+        },
+    )
+    task = {
+        **binding.public(),
+        "required_capabilities": ["agent.runtime.hermes"],
+        "metadata": {"execution": execution},
+    }
+
+    class Bridge:
+        def __init__(self) -> None:
+            self.calls: list[object] = []
+
+        def project_execution(self, _task: object, projection_input: object = None):
+            self.calls.append(projection_input)
+            return 200, dict(projection_input or {})
+
+    supervisor = object.__new__(WorkerSupervisor)
+    supervisor.bridge = Bridge()
+    resolved, source = supervisor.resolve_envelope(task)
+
+    assert source == "persisted_verified"
+    assert resolved.action == "agent_runtime"
+    assert resolved.adapter == "hermes"
+    assert resolved.digest == execution["execution_digest"]
+    assert supervisor.bridge.calls == [execution]

tests/test_online_agent_checkpoint_phases.py

@@ -0,0 +1,56 @@
+from __future__ import annotations
+
+from pathlib import Path
+
+from aios_online_agent_checkpoint import CheckpointBinding, CheckpointStore
+
+
+def _binding() -> CheckpointBinding:
+    return CheckpointBinding(
+        task_id="checkpoint-agent-runtime-test-20260818",
+        attempt=1,
+        execution_digest="a" * 64,
+        job_id="job-checkpoint-agent-runtime-test-20260818",
+        repository="aios/aios",
+        base_oid="b" * 40,
+        expected_branch_oid="b" * 40,
+        target_branch="agent/checkpoint-agent-runtime-test-20260818",
+        generation=1,
+    )
+
+
+def test_agent_runtime_checkpoint_phase_is_durable_before_result(tmp_path: Path) -> None:
+    store = CheckpointStore(tmp_path / "checkpoints.sqlite3")
+    binding = _binding()
+    phases = [
+        "claimed",
+        "workspace_prepared",
+        "verification_passed",
+        "agent_runtime_done",
+        "result_pending",
+        "result_submitted",
+    ]
+
+    rows = [
+        store.append(binding, phase, {"phase": phase, "task_id": binding.task_id})
+        for phase in phases
+    ]
+
+    assert [row["phase"] for row in rows] == phases
+    assert [int(row["phase_order"]) for row in rows] == [10, 20, 30, 40, 50, 60]
+    latest = store.latest_for_attempt(binding.task_id, binding.attempt)
+    assert latest is not None
+    assert latest["phase"] == "result_submitted"
+    assert latest["payload"] == {"phase": "result_submitted", "task_id": binding.task_id}
+
+
+def test_agent_runtime_checkpoint_append_is_idempotent(tmp_path: Path) -> None:
+    store = CheckpointStore(tmp_path / "checkpoints.sqlite3")
+    binding = _binding()
+    payload = {"phase": "agent_runtime_done", "result": {"status": "SUCCEEDED"}}
+
+    first = store.append(binding, "agent_runtime_done", payload)
+    replay = store.append(binding, "agent_runtime_done", payload)
+
+    assert replay["sequence"] == first["sequence"]
+    assert replay["payload_sha256"] == first["payload_sha256"]

tests/test_online_agent_placeholder_plan_deferral.py

@@ -0,0 +1,209 @@
+from __future__ import annotations
+
+from pathlib import Path
+import sys
+
+import pytest
+
+ROOT=Path(__file__).resolve().parents[1]
+sys.path.insert(0,str(ROOT/"active-runtime/app"))
+
+import aios_online_agent_worker as worker
+from aios_online_agent_execution import (
+    GovernedExecutionBinding,
+    project_execution_envelope,
+)
+
+
+
+
+@pytest.fixture(autouse=True)
+def _state_paths(monkeypatch, tmp_path):
+    root=tmp_path/"worker-state"
+    monkeypatch.setattr(worker,"STATE_ROOT",root)
+    monkeypatch.setattr(worker,"WORKER_LIVENESS_PATH",root/"worker-liveness.json")
+
+BASE="a"*40
+TASK_ID="task-worker-availability-placeholder"
+JOB_ID="job-worker-availability-placeholder"
+BRANCH="agent/task-worker-availability-placeholder"
+IDEMPOTENCY="task-worker-availability-placeholder-binding-v1"
+
+
+def binding():
+    return GovernedExecutionBinding.from_mapping({
+        "task_id":TASK_ID,
+        "job_id":JOB_ID,
+        "repository":"aios/aios",
+        "base_oid":BASE,
+        "expected_branch_oid":BASE,
+        "target_branch":BRANCH,
+        "branch_mode":"CREATE_NEW",
+        "generation":1,
+        "idempotency_key":IDEMPOTENCY,
+    })
+
+
+def placeholder():
+    return project_execution_envelope(
+        binding(),
+        {
+            "action":"source_capture",
+            "adapter":"source_capture",
+            "capability_scope":[
+                "git.read",
+                "git.test",
+                "governed.write.branch",
+                "governed.write.file",
+            ],
+        },
+    )
+
+
+def readonly_source_capture():
+    return project_execution_envelope(
+        binding(),
+        {
+            "action":"source_capture",
+            "adapter":"source_capture",
+            "capability_scope":["git.read","git.test"],
+            "source_signals":["worker availability"],
+        },
+    )
+
+
+def deterministic():
+    return project_execution_envelope(
+        binding(),
+        {
+            "action":"deterministic",
+            "adapter":"deterministic",
+            "capability_scope":[
+                "git.read",
+                "git.test",
+                "governed.write.branch",
+                "governed.write.file",
+            ],
+            "change_set":[{
+                "path":"active-runtime/app/example.py",
+                "operation":"write_file",
+                "content":"VALUE=1\\n",
+            }],
+            "verification":[{
+                "label":"unit",
+                "argv":["/usr/bin/python3","-m","pytest","-q"],
+                "cwd":".",
+                "timeout_seconds":120,
+            }],
+            "governance":{
+                "commit_idempotency_key":"task-worker-availability-placeholder-commit-v1",
+                "allowed_files":["active-runtime/app/example.py"],
+            },
+        },
+    )
+
+
+def task(execution):
+    return {
+        "task_id":TASK_ID,
+        "required_capabilities":list(execution["capability_scope"]),
+        "metadata":{
+            "execution":execution,
+            "execution_projection":{
+                "schema":"aios.online-agent.task-projection.v1",
+                "task_id":TASK_ID,
+                "job_id":JOB_ID,
+                "execution_digest":execution["execution_digest"],
+            },
+        },
+    }
+
+
+class FakeBridge:
+    def __init__(self,status,execution):
+        self.status=status
+        self.execution=execution
+        self.calls=[]
+
+    def project_execution(self,task_value,execution=None):
+        self.calls.append(execution)
+        return self.status,self.execution
+
+
+class FakeAgent:
+    def __init__(self,tasks):
+        self.tasks=tasks
+        self.claim_calls=[]
+
+    def poll(self):
+        return list(self.tasks)
+
+    def claim(self,task_id):
+        self.claim_calls.append(task_id)
+        return 500,{"error":"SHOULD_NOT_CLAIM"}
+
+
+def supervisor(bridge, agent=None):
+    value=object.__new__(worker.WorkerSupervisor)
+    value.bridge=bridge
+    value.agent=agent
+    return value
+
+
+def test_write_capable_source_capture_is_classified_as_placeholder():
+    envelope=worker.ExecutionEnvelope.from_task(
+        task(placeholder()),
+        worker_capabilities=tuple(placeholder()["capability_scope"]),
+    )
+    assert worker.source_capture_requires_deterministic_plan(envelope) is True
+
+
+def test_readonly_source_capture_is_not_placeholder():
+    envelope=worker.ExecutionEnvelope.from_task(
+        task(readonly_source_capture()),
+        worker_capabilities=tuple(readonly_source_capture()["capability_scope"]),
+    )
+    assert worker.source_capture_requires_deterministic_plan(envelope) is False
+
+
+def test_placeholder_without_registered_plan_is_skipped_before_claim():
+    persisted=placeholder()
+    bridge=FakeBridge(200,persisted)
+    agent=FakeAgent([task(persisted)])
+    sup=supervisor(bridge,agent)
+    assert sup.once()==0
+    assert bridge.calls==[None]
+    assert agent.claim_calls==[]
+
+
+def test_placeholder_with_missing_projection_is_skipped_before_claim():
+    persisted=placeholder()
+    bridge=FakeBridge(404,{"error":"TASK_EXECUTION_PROJECTION_MISSING"})
+    agent=FakeAgent([task(persisted)])
+    sup=supervisor(bridge,agent)
+    assert sup.once()==0
+    assert bridge.calls==[None]
+    assert agent.claim_calls==[]
+
+
+def test_registered_deterministic_plan_replaces_placeholder():
+    persisted=placeholder()
+    resolved=deterministic()
+    bridge=FakeBridge(200,resolved)
+    sup=supervisor(bridge)
+    envelope,source=sup.resolve_envelope(task(persisted))
+    assert source=="plan_registry_verified"
+    assert envelope.action=="deterministic"
+    assert envelope.digest==resolved["execution_digest"]
+    assert bridge.calls==[None]
+
+
+def test_real_readonly_source_capture_replays_exact_payload():
+    persisted=readonly_source_capture()
+    bridge=FakeBridge(200,persisted)
+    sup=supervisor(bridge)
+    envelope,source=sup.resolve_envelope(task(persisted))
+    assert source=="persisted_verified"
+    assert envelope.action=="source_capture"
+    assert envelope.digest==persisted["execution_digest"]
+    assert bridge.calls==[persisted]

tests/test_openclaw_input_reservation_profile.py

@@ -0,0 +1,16 @@
+from __future__ import annotations
+
+import json
+from pathlib import Path
+
+ROOT=Path(__file__).resolve().parents[1]
+
+
+def test_openclaw_profile_uses_conservative_input_reservation_floor():
+    value=json.loads(
+        (ROOT/"config/external-agent-runtime.d/openclaw.profile.json").read_text()
+    )
+    mg=value["model_governance"]
+    assert mg["required"] is True
+    assert mg["input_reservation_floor_tokens"]==16384
+    assert mg["max_output_tokens"]==256

tests/test_worker_availability_guard.py

@@ -0,0 +1,199 @@
+from __future__ import annotations
+
+import json
+from pathlib import Path
+import sys
+
+ROOT=Path(__file__).resolve().parents[1]
+sys.path.insert(0,str(ROOT/"active-runtime/app"))
+
+from aios_worker_availability_guard import AvailabilityGuard
+
+
+class FakeSystemd:
+    def __init__(self, roots, now=1000.0):
+        self.now=now
+        self.actions=[]
+        self.units={}
+        for unit,root in roots.items():
+            self.units[unit]={
+                "LoadState":"loaded",
+                "ActiveState":"active",
+                "UnitFileState":"enabled",
+                "Environment":f"AIOS_WORKER_STATE_ROOT={root}",
+            }
+        self.units["aios-external-agent-executor.service"]={
+            "LoadState":"loaded",
+            "ActiveState":"active",
+            "UnitFileState":"enabled",
+            "Environment":"",
+        }
+
+    def property(self, unit, name):
+        return self.units.get(unit,{}).get(name,"")
+
+    def environment(self, unit):
+        raw=self.property(unit,"Environment")
+        result={}
+        for item in raw.split():
+            if "=" in item:
+                k,v=item.split("=",1)
+                result[k]=v
+        return result
+
+    def enable(self, unit):
+        self.actions.append((unit,"enable"))
+        self.units[unit]["UnitFileState"]="enabled"
+
+    def start(self, unit):
+        self.actions.append((unit,"start"))
+        self.units[unit]["ActiveState"]="active"
+        env=self.environment(unit)
+        if "AIOS_WORKER_STATE_ROOT" not in env:
+            return
+        root=Path(env["AIOS_WORKER_STATE_ROOT"])
+        root.mkdir(parents=True,exist_ok=True)
+        (root/"worker-liveness.json").write_text(json.dumps({
+            "schema":"aios.online-agent.worker-liveness.v1",
+            "agent_id":"agent-"+unit,
+            "state":"POLL_OK",
+            "last_seen_epoch":self.now,
+        }),encoding="utf-8")
+
+    def restart(self, unit):
+        self.actions.append((unit,"restart"))
+        self.units[unit]["ActiveState"]="active"
+        env=self.environment(unit)
+        root=Path(env["AIOS_WORKER_STATE_ROOT"])
+        root.mkdir(parents=True,exist_ok=True)
+        (root/"worker-liveness.json").write_text(json.dumps({
+            "schema":"aios.online-agent.worker-liveness.v1",
+            "agent_id":"agent-"+unit,
+            "state":"POLL_OK",
+            "last_seen_epoch":self.now,
+        }),encoding="utf-8")
+
+
+def make_guard(tmp_path, fake):
+    worker_units=tuple(
+        unit for unit in fake.units
+        if unit!="aios-external-agent-executor.service"
+    )
+    return AvailabilityGuard(
+        systemd=fake,
+        worker_units=worker_units,
+        dependency_units=("aios-external-agent-executor.service",),
+        status_path=tmp_path/"run/status.json",
+        recovery_path=tmp_path/"run/recovery.json",
+        inhibit_path=tmp_path/"run/maintenance.lock",
+        max_stale_seconds=90,
+        restart_cooldown_seconds=60,
+        startup_wait_seconds=0,
+        now_fn=lambda:fake.now,
+        sleep_fn=lambda _seconds:None,
+    )
+
+
+def write_liveness(root, now, *, state="POLL_OK"):
+    root.mkdir(parents=True,exist_ok=True)
+    (root/"worker-liveness.json").write_text(json.dumps({
+        "schema":"aios.online-agent.worker-liveness.v1",
+        "agent_id":"agent",
+        "state":state,
+        "last_seen_epoch":now,
+    }),encoding="utf-8")
+
+
+def test_healthy_workers_need_no_action(tmp_path):
+    roots={
+        "generic.service":tmp_path/"generic",
+        "hermes.service":tmp_path/"hermes",
+        "openclaw.service":tmp_path/"openclaw",
+    }
+    fake=FakeSystemd(roots)
+    for root in roots.values():
+        write_liveness(root,fake.now)
+    result=make_guard(tmp_path,fake).run(heal=True)
+    assert result["status"]=="HEALTHY"
+    assert result["all_fresh"] is True
+    assert fake.actions==[]
+
+
+def test_inactive_worker_is_started_and_becomes_fresh(tmp_path):
+    roots={"generic.service":tmp_path/"generic"}
+    fake=FakeSystemd(roots)
+    fake.units["generic.service"]["ActiveState"]="inactive"
+    result=make_guard(tmp_path,fake).run(heal=True)
+    assert ("generic.service","start") in fake.actions
+    assert result["all_fresh"] is True
+
+
+def test_disabled_worker_is_enabled_and_started(tmp_path):
+    roots={"generic.service":tmp_path/"generic"}
+    fake=FakeSystemd(roots)
+    fake.units["generic.service"]["UnitFileState"]="disabled"
+    fake.units["generic.service"]["ActiveState"]="inactive"
+    result=make_guard(tmp_path,fake).run(heal=True)
+    assert ("generic.service","enable") in fake.actions
+    assert ("generic.service","start") in fake.actions
+    assert result["all_fresh"] is True
+
+
+def test_active_but_stale_worker_is_restarted(tmp_path):
+    roots={"generic.service":tmp_path/"generic"}
+    fake=FakeSystemd(roots)
+    write_liveness(roots["generic.service"],fake.now-500)
+    result=make_guard(tmp_path,fake).run(heal=True)
+    assert ("generic.service","restart") in fake.actions
+    assert result["all_fresh"] is True
+    assert result["workers"][0]["action"]=="RESTART_STALE"
+
+
+def test_restart_cooldown_prevents_restart_storm(tmp_path):
+    roots={"generic.service":tmp_path/"generic"}
+    fake=FakeSystemd(roots)
+    write_liveness(roots["generic.service"],fake.now-500)
+    guard=make_guard(tmp_path,fake)
+    guard.recovery_path.parent.mkdir(parents=True,exist_ok=True)
+    guard.recovery_path.write_text(json.dumps({
+        "schema":"aios.worker-availability.recovery.v1",
+        "last_restart_epoch":{"generic.service":fake.now-10},
+    }),encoding="utf-8")
+    result=guard.run(heal=True)
+    assert ("generic.service","restart") not in fake.actions
+    assert result["status"]=="DEGRADED"
+    assert result["workers"][0]["detail"]=="RESTART_COOLDOWN_ACTIVE"
+
+
+def test_maintenance_inhibit_disables_healing(tmp_path):
+    roots={"generic.service":tmp_path/"generic"}
+    fake=FakeSystemd(roots)
+    fake.units["generic.service"]["ActiveState"]="inactive"
+    guard=make_guard(tmp_path,fake)
+    guard.inhibit_path.parent.mkdir(parents=True,exist_ok=True)
+    guard.inhibit_path.write_text("planned maintenance\n")
+    result=guard.run(heal=True)
+    assert result["status"]=="MAINTENANCE_INHIBITED"
+    assert fake.actions==[]
+
+
+def test_dependency_is_enabled_and_started(tmp_path):
+    roots={"generic.service":tmp_path/"generic"}
+    fake=FakeSystemd(roots)
+    write_liveness(roots["generic.service"],fake.now)
+    fake.units["aios-external-agent-executor.service"]["UnitFileState"]="disabled"
+    fake.units["aios-external-agent-executor.service"]["ActiveState"]="inactive"
+    result=make_guard(tmp_path,fake).run(heal=True)
+    assert ("aios-external-agent-executor.service","enable") in fake.actions
+    assert ("aios-external-agent-executor.service","start") in fake.actions
+    assert result["all_fresh"] is True
+
+
+def test_check_reports_stale_without_mutation(tmp_path):
+    roots={"generic.service":tmp_path/"generic"}
+    fake=FakeSystemd(roots)
+    write_liveness(roots["generic.service"],fake.now-500)
+    result=make_guard(tmp_path,fake).run(heal=False)
+    assert result["status"]=="DEGRADED"
+    assert result["all_fresh"] is False
+    assert fake.actions==[]

tests/test_worker_availability_runtime_config.py

@@ -0,0 +1,56 @@
+from __future__ import annotations
+
+from pathlib import Path
+
+ROOT=Path(__file__).resolve().parents[1]
+
+
+def test_worker_source_has_idle_and_busy_liveness():
+    source=(ROOT/"active-runtime/app/aios_online_agent_worker.py").read_text()
+    assert "aios.online-agent.worker-liveness.v1" in source
+    assert 'write_worker_liveness("POLL_OK"' in source
+    assert '"POLL_FAILED"' in source
+    assert 'write_worker_liveness("STARTING")' in source
+    assert '"BUSY"' in source
+    assert "last_seen_epoch" in source
+
+
+def test_guard_timer_is_persistent_and_frequent():
+    timer=(ROOT/"deploy/systemd/aios-worker-availability-guard.timer").read_text()
+    assert "OnBootSec=15s" in timer
+    assert "OnUnitActiveSec=30s" in timer
+    assert "Persistent=true" in timer
+    assert "WantedBy=timers.target" in timer
+
+
+def test_guard_service_has_no_network_or_credentials():
+    unit=(ROOT/"deploy/systemd/aios-worker-availability-guard.service").read_text()
+    assert "User=root" in unit
+    assert "LoadCredential" not in unit
+    assert "RestrictAddressFamilies=AF_UNIX" in unit
+    assert "NoNewPrivileges=true" in unit
+    assert "ProtectSystem=strict" in unit
+    assert "ReadWritePaths=/run/aios-worker-availability" in unit
+    assert "AIOS_WORKER_AVAILABILITY_MAX_STALE_SECONDS=90" in unit
+
+
+def test_guard_manages_all_three_workers_and_external_executor():
+    source=(ROOT/"active-runtime/app/aios_worker_availability_guard.py").read_text()
+    for unit in (
+        "aios-online-agent-worker-20260815.service",
+        "aios-hermes-online-agent-worker.service",
+        "aios-openclaw-online-agent-worker.service",
+        "aios-external-agent-executor.service",
+    ):
+        assert unit in source
+    assert "maintenance.lock" in source
+    assert "RESTART_STALE" in source
+
+
+def test_worker_defers_write_capable_source_capture_until_plan_exists():
+    source=(ROOT/"active-runtime/app/aios_online_agent_worker.py").read_text()
+    assert "ENGINEERING_WRITE_CAPABILITIES" in source
+    assert "source_capture_requires_deterministic_plan" in source
+    assert "TASK_SKIPPED_PENDING_DETERMINISTIC_PLAN" in source
+    assert '"governed.write.branch"' in source
+    assert '"governed.write.file"' in source