Compare commits

...
Author SHA1 Message Date
Codex 6b0db21a85 review(codex): 固化 V2R1 最新发布审查结论 2026-07-21 12:34:00 +08:00
Codex be4726178a task(agent): 转入 provenance 后正式 Gate 复审 2026-07-21 11:23:44 +08:00
AIOS-Owner 0834b272f0 task(agent): 审查四项清障并保留 provenance 门禁 2026-07-20 18:18:36 +08:00
Codex Supervisor e28245f411 chore(evidence): refine AIOS V2 signing-material blocker 2026-07-20 17:05:10 +08:00
Codex Supervisor 8a9d033e9c chore(agent): record final V2R1 first-round blocked review 2026-07-20 17:03:08 +08:00
Codex Supervisor e567699283 chore(agent): claim V2R1 Codex supervision and block preproduction 2026-07-20 16:51:25 +08:00
Codex c935aac3b8 task(codex): 初始化无人值守监督状态 2026-07-20 12:08:06 +08:00
Codex 75f0148878 task(codex): 写入生产监督无人值守任务载荷 2026-07-20 12:07:58 +08:00
Codex 91afbfaf76 task(codex): launch architecture supervision and blocker resolution 2026-07-20 11:49:13 +08:00
Codex f23b694ad3 tasks(aios-v2): add Codex production supervisor task 2026-07-20 11:46:43 +08:00
Codex f97dea55ac tasks(aios-v2): add OpenClaw production deployment task 2026-07-20 11:46:39 +08:00
Codex efd08d7771 docs(aios-v2): add ADR-003 GitOps-only flow 2026-07-20 11:46:31 +08:00
Codex ff6cb2e217 docs(aios-v2): add ADR-005 documentation sync gate 2026-07-20 11:45:16 +08:00
Codex d774620fea docs(aios-v2): add ADR-002 runtime decoupling decision 2026-07-20 11:45:12 +08:00
Codex b26ff010fc docs(aios-v2): add ADR-001 control plane decision 2026-07-20 11:45:09 +08:00
Codex 4b49f2ba4b docs(aios-v2): add ADR index 2026-07-20 11:45:04 +08:00
Codex b4b8fe8828 docs(aios-v2): add traceability matrix 2026-07-20 11:43:46 +08:00
Codex 4e8ab051a0 docs(aios-v2): add module contracts 2026-07-20 11:41:39 +08:00
Codex f61e35b218 docs(aios-v2): add migration and delivery strategy 2026-07-20 11:34:33 +08:00
Codex b2f90f24a3 docs(aios-v2): add target architecture 2026-07-20 11:32:16 +08:00
Codex 6721eb4752 docs(aios-v2): add user requirements baseline 2026-07-20 11:29:07 +08:00
Codex a9e4fd1e40 docs(aios-v2): establish architecture source of truth 2026-07-20 11:24:12 +08:00
Codex 621750976a deploy(candidate): add guarded release rollback 2026-07-20 11:22:04 +08:00
Codex 5359e88806 feat(candidate): expose sanitized GitOps status API 2026-07-20 11:19:37 +08:00
Codex 76636931ba feat(candidate): add sanitized GitOps status provider 2026-07-20 11:17:18 +08:00
Codex 180c1fc66e deploy(candidate): add end-to-end healthcheck 2026-07-20 11:14:53 +08:00
Codex c0ff052dab deploy(candidate): add Docker Compose runtime 2026-07-20 11:12:42 +08:00
Codex 94cffb6ba3 deploy(candidate): add hardened container image 2026-07-20 11:10:36 +08:00
Codex 9ad5628e69 deploy(candidate): add environment configuration template 2026-07-20 11:08:37 +08:00
Codex 6d397a7894 deploy(candidate): add hardened systemd service 2026-07-20 11:07:06 +08:00
Codex 165fea6ba9 fix(candidate): prevent Hermes completion heartbeat revision race 2026-07-20 11:05:45 +08:00
Codex d079e2087c test(candidate): add end-to-end runtime and persistence tests 2026-07-20 11:01:07 +08:00
Codex e3b36ebdcc feat(candidate): add production application entrypoint 2026-07-20 10:57:42 +08:00
Codex 5e791b11c0 feat(candidate): add tenant-safe web console 2026-07-20 10:54:43 +08:00
Codex 50be65b76e feat(candidate): add durable SQLite control-plane persistence 2026-07-20 10:50:31 +08:00
Codex 7bb86e30bf feat(candidate): export orchestration and execution tracking 2026-07-20 10:42:49 +08:00
Codex ea56b45b7e feat(candidate): add persistent agent execution tracking 2026-07-20 10:40:22 +08:00
Codex f53682da7a feat(candidate): add unified AIOS execution orchestrator 2026-07-20 10:38:06 +08:00
Codex 71884ac088 feat(candidate): add Hermes worker loop and lease heartbeat 2026-07-20 10:35:55 +08:00
Codex 969c278cb0 feat(candidate): implement Hermes queue and lease adapter 2026-07-20 10:34:25 +08:00
Codex 6fe2a1e036 feat(candidate): expose Hermes runtime package 2026-07-20 10:31:07 +08:00
Codex 9a60e0c68e feat(candidate): add Hermes runtime contracts 2026-07-20 10:27:47 +08:00
Codex 46f553ca11 feat(candidate): expose Hermes runtime boundary package 2026-07-20 10:17:02 +08:00
Codex 4f0a0c2184 feat(candidate): add OpenClaw audit callback sink 2026-07-20 10:07:54 +08:00
Codex c107b193d9 feat(candidate): add OpenClaw runtime health policy 2026-07-20 10:06:36 +08:00
Codex 2db7ced7a4 feat(candidate): add OpenClaw runtime transports 2026-07-20 10:01:59 +08:00
Codex 6361d9a334 feat(candidate): implement OpenClaw execution adapter 2026-07-20 09:59:53 +08:00
Codex 2cf54f04ef feat(candidate): expose OpenClaw adapter package 2026-07-20 09:56:06 +08:00
Codex f840ce5c94 feat(integration): add OpenClaw command contract 2026-07-20 09:49:15 +08:00
Codex 35e7b55490 integration(aios-v2): create CTO production candidate branch 2026-07-20 09:47:01 +08:00
Codex 1f8d1f23b3 feat(aios-v2): add tenant-aware HTTP control plane gateway 2026-07-20 02:24:19 +08:00
Codex e93011791e feat(aios-v2): add tenant-isolated control plane domain 2026-07-20 02:19:16 +08:00
Codex 4460564a48 feat(aios-v2): initialize control plane package 2026-07-20 01:11:11 +08:00
Codex 4d6658136e docs(task): add AIOS control plane gateway task contract 2026-07-20 01:06:11 +08:00
Codex 5767e665c2 task(codex): implement AIOS V2 control plane gateway v1 2026-07-20 01:06:04 +08:00
51 changed files with 6759 additions and 0 deletions
+57
View File
@@ -0,0 +1,57 @@
{
"schema": "aios.v2.agent-blockers.v1",
"task_id": "AIOS-V2R1-PRODUCTION-SUPERVISOR-20260720",
"updated_at": "2026-07-21T04:30:00Z",
"blockers": [
{
"blocker_id": "BLK-OPENCLAW-PLAINTEXT-SECRETS",
"severity": "high",
"status": "resolved",
"root_cause": "OpenClaw 三处 managed plaintext。",
"fix": "迁移为单一 file Secret Provider 并轮换 Gateway Token。",
"verification": "最终 audit 四类计数均为 0,OneAPI 实际调用通过。",
"commit": "bf01a13f3aaf42a386232ed2f5409feb0f79fcf6",
"resolved_at": "2026-07-20T10:18:36Z"
},
{
"blocker_id": "BLK-AIOS-V2-SIGNING-MATERIAL-MISSING",
"severity": "high",
"status": "resolved",
"root_cause": "Ed25519 Key ID 与 HMAC 实现不一致且材料缺失。",
"fix": "真实 Ed25519 合同及受保护材料。",
"verification": "0600、sign/verify、篡改拒绝通过。",
"commit": "8a36f7832bc105e4b9d32ffbe018abed2ff1b1ab",
"resolved_at": "2026-07-20T10:18:36Z"
},
{
"blocker_id": "BLK-CANDIDATE-NOT-MERGED",
"severity": "high",
"status": "resolved",
"root_cause": "OpenClaw 任务未包含 Candidate。",
"fix": "--no-ff 合并最新 Candidate。",
"verification": "Candidate 为 OpenClaw HEAD 祖先且 OID 同步。",
"commit": "f55dfd975e9d716b690c7693742e7cf2beaa55e7",
"resolved_at": "2026-07-20T10:18:36Z"
},
{
"blocker_id": "BLK-OPENCLAW-RUN-PROVENANCE-MISSING",
"severity": "high",
"status": "resolved",
"root_cause": "旧插件只有进程内 runId/sender Map,缺少 Owner-signed Schema、持久 Store 和限定 Git Writer。",
"fix": "部署 aios-tenant-boundary rc2,增加 Ed25519 issuer/store、Run/Sender/OID/期限绑定、一次性 permit 与七路径专用 Writer。",
"verification": "Provenance openclaw-v2r1-20260721T031523Z 完成 heartbeat 写入、Commit、普通 Push 和审计;README.md 越权写入被拒绝;Node 测试 9/9。",
"commit": "25c25abd3ed19091fe078d4d7a271c57f23e87fb",
"resolved_at": "2026-07-21T03:23:44Z"
},
{
"blocker_id": "BLK-CONTROLLER-LEGACY-GITEA-CHECK",
"severity": "high",
"status": "resolved",
"root_cause": "硬编码旧仓库。",
"fix": "配置化 V2R1 OID 健康检查并使用只读 Credential。",
"verification": "12 项测试、detached clean、fully_synced。",
"commit": "6188da26453505dacde9d6d298191437164c8d27",
"resolved_at": "2026-07-20T10:18:36Z"
}
]
}
+13
View File
@@ -0,0 +1,13 @@
{
"schema": "aios.v2.agent-claim.v1",
"task_id": "AIOS-V2R1-PRODUCTION-SUPERVISOR-20260720",
"executor": "codex",
"executor_instance": "windows-chatgpt-owner-supervisor",
"task_branch": "task/codex/AIOS-V2R1-PRODUCTION-SUPERVISOR-20260720",
"candidate_branch": "integration/aios-v2r1-cto-production-candidate-20260720",
"candidate_commit": "ba7669c3b49f98cf3500af6ea966ba0db472ff3b",
"claimed_at": "2026-07-20T08:51:25Z",
"owner_observation": "五项 Blocker、Gate A/B、离线验收、OpenClaw 在线合同、双服务重启与回滚均通过;Hermes 完整在线 Worker 生命周期仍缺服务端点证据,Codex 维持 STOP_PREPRODUCTION。",
"owner_observation_at": "2026-07-21T04:30:00Z",
"reviewed_openclaw_commit_before_evidence": "b158041b8633e658c34ce182f3c8aa97b4bf5f9a"
}
+18
View File
@@ -0,0 +1,18 @@
{
"schema": "aios.agent.credential-status.v1",
"task_id": "task/codex/AIOS-V2R1-PRODUCTION-SUPERVISOR-20260720",
"last_updated": "2026-07-21T04:30:00Z",
"plaintext_findings_before": 3,
"plaintext_findings_after": 0,
"gateway_token_rotated": true,
"secret_values_recorded": false,
"signing_material_configured": true,
"signing_key_id": "openclaw-v2r1-candidate-ed25519",
"runtime_sign_verify_test": "passed",
"controller_credential_scope": "read:repository",
"run_provenance_status": "resolved",
"provenance_id": "openclaw-v2r1-20260721T031523Z",
"openclaw_runtime_plugin_version": "2.1.0-rc.3",
"runtime_public_key_only_hot_path": true,
"secret_values_recorded_in_evidence": false
}
+10
View File
@@ -0,0 +1,10 @@
{
"schema": "aios.v2.agent-heartbeat.v1",
"task_id": "AIOS-V2R1-PRODUCTION-SUPERVISOR-20260720",
"executor": "codex",
"heartbeat_at": "2026-07-21T04:30:00Z",
"state": "reviewing",
"current_step": "FINAL_REVIEW_HERMES_ONLINE_GAP",
"reviewed_commit": "b158041b8633e658c34ce182f3c8aa97b4bf5f9a",
"openclaw_evidence_commit": "0be3ba09e42b4f6db56461e3ca3a3fcc0a487f31"
}
+73
View File
@@ -0,0 +1,73 @@
{
"schema": "aios.v2.agent-progress.v1",
"task_id": "AIOS-V2R1-PRODUCTION-SUPERVISOR-20260720",
"updated_at": "2026-07-21T04:30:00Z",
"state": "reviewing",
"candidate_oid": "634f6b164f5b20372ab9ec8144bbf76df97de335",
"openclaw_task_oid_at_review": "0be3ba09e42b4f6db56461e3ca3a3fcc0a487f31",
"codex_task_oid_before_evidence": "be4726178aa8fa59321aafaabcff8a83653cc1b8",
"gitea_sync": {
"candidate_match": true,
"openclaw_task_match": true,
"codex_task_match_before_this_evidence": true
},
"completed_checks": [
"worktree",
"candidate_merge",
"secrets_audit_zero",
"ed25519_sign_verify",
"controller_fully_synced",
"owner_signed_provenance_writer",
"unauthorized_path_denied",
"provenance_audit_chain",
"gate_a",
"gate_b_10_of_10",
"offline_acceptance_8_0_1",
"openclaw_online_contract",
"aios_openclaw_restart_recovery",
"controlled_rollback_and_forward"
],
"review_conclusion": "五项 Blocker、文档/代码 Gate、候选测试、离线验收、OpenClaw 在线合同、重启与回滚均通过;Hermes 完整在线 Worker 生命周期未形成服务端点证据,预发布仍停止。",
"check_time_window": {
"started_at": "2026-07-20T08:24:30Z",
"ended_at": "2026-07-20T09:03:08Z"
},
"operator": "windows-chatgpt-owner-supervisor",
"reviewer": "codex-supervisor + Spark architect/explorer/implementer/validator/docs_reporter",
"approver": "AIOS Owner gate recorded by Windows ChatGPT Owner Supervisor",
"evidence_artifacts": [
"task/openclaw/AIOS-V2R1-PRODUCTION-DEPLOY-20260720@1aa71f233d2c2a6ed7290d0a3ddcb04dcd84febe",
"task/codex/AIOS-V2R1-PRODUCTION-SUPERVISOR-20260720:.agent/",
"/data/aios-gitops/runtime/backups/gitea-db-before-openclaw-token-rotation-20260720T084251Z.sqlite3",
"/home/aiman/.ssh/known_hosts.aios-v2r1-backup-20260720T083647Z",
"/var/log/aios-openclaw-v2r1-evidence-repair-20260720.json"
],
"independent_reviews": {
"architect": "REJECT;只允许修复+复验,禁止生产放行。",
"explorer": "OpenClaw evidence OID 已重新闭环,但 Candidate 未合并、Secret 与预发布仍未确认。",
"implementer": "IMPLEMENTATION_SAFE_WITH_BLOCKERS;已实施动作无需立即回滚。",
"validator": "BLOCKED;Secret、环境、预发布门禁未通过。",
"docs_reporter": "已补齐时间窗、证据路径、operator/reviewer/approver、blocker 与 final_decision。"
},
"final_decision": "STOP_PREPRODUCTION",
"aios_v2_env_status": {
"exists": true,
"owner": "root",
"group": "gitops",
"mode": "0640",
"four_fields_nonempty": true,
"credential_ref_target_exists": true,
"signing_material_target_exists": true
},
"open_checks": [
"hermes_online_claim_start_heartbeat_complete_fail_cancel",
"hermes_online_lease_expiry_worker_crash_recovery",
"hermes_online_cell_profile_and_invalid_result_rejection"
],
"gate_a": "passed",
"gate_b": "passed_10_of_10",
"offline_acceptance": "passed_8_failed_0_warning_1",
"online_openclaw": "passed",
"restart_recovery": "passed",
"rollback_drill": "passed"
}
+27
View File
@@ -0,0 +1,27 @@
{
"schema": "aios.v2.agent-result.v1",
"task_id": "AIOS-V2R1-PRODUCTION-SUPERVISOR-20260720",
"final_status": "PARTIAL",
"candidate_commit": "ba7669c3b49f98cf3500af6ea966ba0db472ff3b",
"reviewed_commit": "b158041b8633e658c34ce182f3c8aa97b4bf5f9a",
"gitea_sync": {
"candidate_match": true,
"openclaw_task_match": true,
"codex_task_match_before_this_evidence": true
},
"credential_rotation_status": "secretref_resolved",
"test_status": "Gate A 通过;Gate B 10/10;离线 8/0/1;OpenClaw 在线、重启恢复、回滚通过;Hermes 完整在线 Worker 生命周期未覆盖",
"deployment_status": "STOP_PREPRODUCTION",
"rollback_status": "Previous 2a2b702 与 Current b158041 双向演练通过且业务状态保留",
"documentation_sync_status": "updated",
"open_blockers": [],
"release_recommendation": "STOP_PREPRODUCTION",
"completed_at": "2026-07-21T04:30:00Z",
"openclaw_evidence_commit": "0be3ba09e42b4f6db56461e3ca3a3fcc0a487f31",
"independent_validation": "PASS_FOR_IMPLEMENTED_OPENCLAW_BOUNDARY",
"operator": "windows-chatgpt-owner-supervisor",
"reviewer": "codex-supervisor + Spark roles",
"approver": "AIOS Owner gate recorded",
"remaining_risk": "Hermes 完整在线 Worker 生命周期缺少可调用服务端点与实时证据",
"formal_online_acceptance": "partial"
}
+18
View File
@@ -0,0 +1,18 @@
{
"schema": "aios.v2.agent-status.v1",
"task_id": "AIOS-V2R1-PRODUCTION-SUPERVISOR-20260720",
"executor": "codex",
"state": "reviewing",
"claimed_at": "2026-07-20T08:51:25Z",
"running_at": "2026-07-20T08:51:25Z",
"updated_at": "2026-07-21T04:30:00Z",
"last_heartbeat": "2026-07-21T04:30:00Z",
"current_step": "FINAL_REVIEW_HERMES_ONLINE_GAP",
"candidate_commit": "ba7669c3b49f98cf3500af6ea966ba0db472ff3b",
"reviewed_commit": "b158041b8633e658c34ce182f3c8aa97b4bf5f9a",
"release_recommendation": "STOP_PREPRODUCTION",
"openclaw_evidence_commit": "0be3ba09e42b4f6db56461e3ca3a3fcc0a487f31",
"status": "reviewing"
,"formal_online_acceptance": "partial"
,"remaining_gap": "Hermes full online worker lifecycle"
}
+31
View File
@@ -0,0 +1,31 @@
{
"schema": "aios.v2.agent-task.v1",
"task_id": "AIOS-V2R1-PRODUCTION-SUPERVISOR-20260720",
"display_name": "AIOS V2R1 Codex 架构监督、阻塞处理与发布审查任务",
"language": "zh-CN",
"executor": "codex",
"role": "architecture-supervisor",
"state": "queued",
"candidate_branch": "integration/aios-v2r1-cto-production-candidate-20260720",
"candidate_commit": "634f6b164f5b20372ab9ec8144bbf76df97de335",
"canonical_task": "tasks/codex-production-supervisor-task.yaml",
"supervises": "task/openclaw/AIOS-V2R1-PRODUCTION-DEPLOY-20260720",
"claim_required": true,
"heartbeat_interval_seconds": 300,
"evidence_directory": ".agent/evidence",
"instructions": [
"领取前读取候选分支完整中文文档、ADR、追踪矩阵和候选报告",
"领取后写入 .agent/claim.json 并把状态改为 claimed",
"持续审阅 OpenClaw 的 claim、heartbeat、progress、blocker、commit 和 result",
"分析失败测试、架构偏移、多租户、安全和文档同步阻塞",
"需要修改时创建 Patch 分支、增加复现测试并同步文档",
"所有阻塞和审查结论必须提交到系统统一审视",
"禁止缺少证据时批准发布或削弱安全门禁"
],
"acceptance_commands": [
"python3 tools/aios_v2_docs_code_check.py --base-ref integration/aios-v2r1-cto-production-candidate-20260720",
"python3 -m unittest discover -s tests/aios_v2_candidate -p test_*.py -v",
"bash deploy/aios-v2/acceptance/run-all.sh"
],
"created_at": "2026-07-20T00:00:00Z"
}
@@ -0,0 +1,25 @@
"""AIOS-managed Hermes runtime boundary."""
from .adapter import (
HermesAdapter,
HermesConflict,
HermesNotFound,
HermesRuntimeError,
InMemoryHermesStore,
)
from .contract import HermesHealth, HermesJob, HermesJobState, HermesResult
from .worker import HermesWorker, WorkerRunResult
__all__ = [
"HermesAdapter",
"HermesConflict",
"HermesHealth",
"HermesJob",
"HermesJobState",
"HermesNotFound",
"HermesResult",
"HermesRuntimeError",
"HermesWorker",
"InMemoryHermesStore",
"WorkerRunResult",
]
+437
View File
@@ -0,0 +1,437 @@
"""Tenant-partitioned Hermes queue and worker lease boundary."""
from __future__ import annotations
import hashlib
import json
import threading
from dataclasses import replace
from datetime import datetime, timedelta, timezone
from typing import Iterable, Mapping
from .contract import (
ALLOWED_TRANSITIONS,
TERMINAL_STATES,
HermesHealth,
HermesJob,
HermesJobState,
HermesResult,
utc_now,
)
class HermesRuntimeError(RuntimeError):
"""Base Hermes runtime-boundary failure."""
class HermesNotFound(HermesRuntimeError):
"""A tenant-scoped job was not found."""
class HermesConflict(HermesRuntimeError):
"""A state, lease, idempotency, or revision conflict occurred."""
class InMemoryHermesStore:
"""Thread-safe reference store; production backends implement this contract."""
def __init__(self) -> None:
self.lock = threading.RLock()
self.jobs: dict[str, HermesJob] = {}
self.results: dict[str, HermesResult] = {}
self.idempotency: dict[tuple[str, str], tuple[str, str]] = {}
self.worker_heartbeats: dict[str, tuple[str, str | None]] = {}
class HermesAdapter:
"""Queue and reconcile Hermes jobs without owning AIOS policy."""
def __init__(
self,
store: InMemoryHermesStore | None = None,
*,
runtime_id: str = "hermes-runtime",
runtime_version: str = "v1",
default_lease_seconds: int = 60,
) -> None:
if default_lease_seconds < 5:
raise ValueError("default_lease_seconds must be >= 5")
self.store = store or InMemoryHermesStore()
self.runtime_id = runtime_id
self.runtime_version = runtime_version
self.default_lease_seconds = int(default_lease_seconds)
@staticmethod
def _digest(job: HermesJob) -> str:
raw = json.dumps(
{
"tenant_id": job.tenant_id,
"cell_id": job.cell_id,
"profile_id": job.profile_id,
"task_id": job.task_id,
"trace_id": job.trace_id,
"action": job.action,
"payload": dict(job.payload),
"priority": job.priority,
},
ensure_ascii=False,
sort_keys=True,
separators=(",", ":"),
default=str,
).encode("utf-8")
return hashlib.sha256(raw).hexdigest()
@staticmethod
def _parse(value: str | None) -> datetime | None:
if value is None:
return None
normalized = value[:-1] + "+00:00" if value.endswith("Z") else value
parsed = datetime.fromisoformat(normalized)
if parsed.tzinfo is None:
raise ValueError("timestamp must include timezone")
return parsed.astimezone(timezone.utc)
@staticmethod
def _format(value: datetime) -> str:
return value.astimezone(timezone.utc).isoformat().replace("+00:00", "Z")
@staticmethod
def _assert_tenant(job: HermesJob, tenant_id: str) -> None:
if job.tenant_id != tenant_id:
raise HermesNotFound("job not found")
def enqueue(self, job: HermesJob) -> HermesJob:
if job.state != HermesJobState.QUEUED or job.revision != 1:
raise HermesConflict("new job must be QUEUED at revision 1")
digest = self._digest(job)
idem_ref = (job.tenant_id, job.idempotency_key)
with self.store.lock:
existing = self.store.jobs.get(job.job_id)
if existing is not None:
if self._digest(existing) != digest:
raise HermesConflict("job_id reused with different content")
return existing
prior = self.store.idempotency.get(idem_ref)
if prior is not None:
prior_id, prior_digest = prior
if prior_digest != digest:
raise HermesConflict("idempotency key reused with different content")
return self.store.jobs[prior_id]
self.store.jobs[job.job_id] = job
self.store.idempotency[idem_ref] = (job.job_id, digest)
return job
def get(self, job_id: str, *, tenant_id: str) -> HermesJob:
with self.store.lock:
job = self.store.jobs.get(job_id)
if job is None:
raise HermesNotFound("job not found")
self._assert_tenant(job, tenant_id)
return job
def list_jobs(
self,
*,
tenant_id: str,
cell_id: str | None = None,
profile_id: str | None = None,
states: Iterable[HermesJobState | str] | None = None,
) -> list[HermesJob]:
state_filter = {HermesJobState(item) for item in states} if states else None
with self.store.lock:
self._requeue_expired_locked()
jobs = [
job
for job in self.store.jobs.values()
if job.tenant_id == tenant_id
and (cell_id is None or job.cell_id == cell_id)
and (profile_id is None or job.profile_id == profile_id)
and (state_filter is None or job.state in state_filter)
]
return sorted(jobs, key=lambda item: (item.priority, item.created_at, item.job_id))
def claim(
self,
*,
tenant_id: str,
cell_id: str,
profile_id: str,
worker_id: str,
lease_seconds: int | None = None,
) -> HermesJob | None:
worker_id = str(worker_id or "").strip()
if not worker_id:
raise ValueError("worker_id is required")
duration = int(lease_seconds or self.default_lease_seconds)
if duration < 5:
raise ValueError("lease_seconds must be >= 5")
with self.store.lock:
self._requeue_expired_locked()
candidates = [
job
for job in self.store.jobs.values()
if job.tenant_id == tenant_id
and job.cell_id == cell_id
and job.profile_id == profile_id
and job.state == HermesJobState.QUEUED
]
self.store.worker_heartbeats[worker_id] = (utc_now(), None)
if not candidates:
return None
current = sorted(
candidates,
key=lambda item: (item.priority, item.created_at, item.job_id),
)[0]
claimed = self._transition_locked(
current,
HermesJobState.CLAIMED,
assigned_worker_id=worker_id,
lease_expires_at=self._format(
datetime.now(timezone.utc) + timedelta(seconds=duration)
),
)
self.store.worker_heartbeats[worker_id] = (utc_now(), claimed.job_id)
return claimed
def start(
self,
job_id: str,
*,
tenant_id: str,
worker_id: str,
expected_revision: int | None = None,
) -> HermesJob:
with self.store.lock:
current = self._owned_job_locked(job_id, tenant_id, worker_id)
self._assert_revision(current, expected_revision)
return self._transition_locked(current, HermesJobState.RUNNING)
def heartbeat(
self,
job_id: str,
*,
tenant_id: str,
worker_id: str,
lease_seconds: int | None = None,
) -> HermesJob:
duration = int(lease_seconds or self.default_lease_seconds)
with self.store.lock:
current = self._owned_job_locked(job_id, tenant_id, worker_id)
if current.state not in {HermesJobState.CLAIMED, HermesJobState.RUNNING}:
raise HermesConflict("job does not accept heartbeat")
updated = replace(
current,
revision=current.revision + 1,
lease_expires_at=self._format(
datetime.now(timezone.utc) + timedelta(seconds=duration)
),
updated_at=utc_now(),
)
self.store.jobs[job_id] = updated
self.store.worker_heartbeats[worker_id] = (utc_now(), job_id)
return updated
def complete(
self,
job_id: str,
*,
tenant_id: str,
worker_id: str,
output: Mapping[str, object],
expected_revision: int | None = None,
) -> HermesResult:
return self._finish(
job_id,
tenant_id=tenant_id,
worker_id=worker_id,
state=HermesJobState.COMPLETED,
output=output,
expected_revision=expected_revision,
)
def fail(
self,
job_id: str,
*,
tenant_id: str,
worker_id: str,
error: Mapping[str, object],
expected_revision: int | None = None,
) -> HermesResult:
return self._finish(
job_id,
tenant_id=tenant_id,
worker_id=worker_id,
state=HermesJobState.FAILED,
error=error,
expected_revision=expected_revision,
)
def cancel(
self,
job_id: str,
*,
tenant_id: str,
expected_revision: int | None = None,
) -> HermesResult:
with self.store.lock:
current = self.store.jobs.get(job_id)
if current is None:
raise HermesNotFound("job not found")
self._assert_tenant(current, tenant_id)
self._assert_revision(current, expected_revision)
if current.state in TERMINAL_STATES:
result = self.store.results.get(job_id)
if result is None:
raise HermesConflict("terminal result missing")
return result
updated = self._transition_locked(current, HermesJobState.CANCELLED)
result = HermesResult(
job_id=updated.job_id,
tenant_id=updated.tenant_id,
task_id=updated.task_id,
trace_id=updated.trace_id,
state=HermesJobState.CANCELLED,
worker_id=updated.assigned_worker_id,
error={"code": "cancelled", "message": "cancelled by AIOS"},
finished_at=utc_now(),
)
self.store.results[job_id] = result
return result
def result(self, job_id: str, *, tenant_id: str) -> HermesResult:
with self.store.lock:
job = self.store.jobs.get(job_id)
if job is None:
raise HermesNotFound("job not found")
self._assert_tenant(job, tenant_id)
result = self.store.results.get(job_id)
if result is None:
raise HermesConflict("job has no terminal result")
return result
def health(self) -> HermesHealth:
with self.store.lock:
self._requeue_expired_locked()
active_jobs = sum(
1 for job in self.store.jobs.values()
if job.state in {HermesJobState.CLAIMED, HermesJobState.RUNNING}
)
queued_jobs = sum(
1 for job in self.store.jobs.values()
if job.state == HermesJobState.QUEUED
)
partitions = {
(job.tenant_id, job.cell_id, job.profile_id)
for job in self.store.jobs.values()
}
active_workers = sum(
1 for heartbeat, _ in self.store.worker_heartbeats.values()
if self._parse(heartbeat)
and self._parse(heartbeat) >= datetime.now(timezone.utc) - timedelta(seconds=120)
)
return HermesHealth(
runtime_id=self.runtime_id,
status="healthy",
version=self.runtime_version,
heartbeat_at=utc_now(),
active_workers=active_workers,
active_jobs=active_jobs,
queued_jobs=queued_jobs,
tenant_partitions=len(partitions),
profile_isolation_enforced=True,
details={"store": type(self.store).__name__},
)
def _finish(
self,
job_id: str,
*,
tenant_id: str,
worker_id: str,
state: HermesJobState,
output: Mapping[str, object] | None = None,
error: Mapping[str, object] | None = None,
expected_revision: int | None,
) -> HermesResult:
with self.store.lock:
current = self._owned_job_locked(job_id, tenant_id, worker_id)
self._assert_revision(current, expected_revision)
updated = self._transition_locked(current, state)
result = HermesResult(
job_id=updated.job_id,
tenant_id=updated.tenant_id,
task_id=updated.task_id,
trace_id=updated.trace_id,
state=state,
worker_id=worker_id,
output=output,
error=error,
started_at=current.updated_at,
finished_at=utc_now(),
)
self.store.results[job_id] = result
self.store.worker_heartbeats[worker_id] = (utc_now(), None)
return result
def _owned_job_locked(self, job_id: str, tenant_id: str, worker_id: str) -> HermesJob:
current = self.store.jobs.get(job_id)
if current is None:
raise HermesNotFound("job not found")
self._assert_tenant(current, tenant_id)
if current.assigned_worker_id != worker_id:
raise HermesConflict("worker does not own job lease")
return current
@staticmethod
def _assert_revision(job: HermesJob, expected: int | None) -> None:
if expected is not None and job.revision != expected:
raise HermesConflict("job revision changed")
def _transition_locked(
self,
current: HermesJob,
target: HermesJobState,
*,
assigned_worker_id: str | None = None,
lease_expires_at: str | None = None,
) -> HermesJob:
if target not in ALLOWED_TRANSITIONS[current.state]:
raise HermesConflict(
f"invalid transition {current.state.value} -> {target.value}"
)
updated = replace(
current,
state=target,
revision=current.revision + 1,
assigned_worker_id=(
assigned_worker_id
if assigned_worker_id is not None
else current.assigned_worker_id
),
lease_expires_at=(
lease_expires_at
if lease_expires_at is not None
else (None if target in TERMINAL_STATES else current.lease_expires_at)
),
updated_at=utc_now(),
)
self.store.jobs[current.job_id] = updated
return updated
def _requeue_expired_locked(self) -> None:
now = datetime.now(timezone.utc)
for current in list(self.store.jobs.values()):
expiry = self._parse(current.lease_expires_at)
if (
current.state in {HermesJobState.CLAIMED, HermesJobState.RUNNING}
and expiry is not None
and expiry <= now
):
self.store.jobs[current.job_id] = replace(
current,
state=HermesJobState.QUEUED,
revision=current.revision + 1,
assigned_worker_id=None,
lease_expires_at=None,
updated_at=utc_now(),
)
+153
View File
@@ -0,0 +1,153 @@
"""Tenant-scoped Hermes runtime contracts."""
from __future__ import annotations
import json
import uuid
from dataclasses import dataclass, field
from datetime import datetime, timezone
from enum import Enum
from typing import Any, Mapping
def utc_now() -> str:
return datetime.now(timezone.utc).isoformat().replace("+00:00", "Z")
def _required(value: str, label: str) -> str:
normalized = str(value or "").strip()
if not normalized:
raise ValueError(f"{label} is required")
return normalized
def _safe_mapping(value: Mapping[str, Any] | None) -> dict[str, Any]:
return json.loads(json.dumps(dict(value or {}), ensure_ascii=False, sort_keys=True, default=str))
class HermesJobState(str, Enum):
QUEUED = "QUEUED"
CLAIMED = "CLAIMED"
RUNNING = "RUNNING"
COMPLETED = "COMPLETED"
FAILED = "FAILED"
CANCELLED = "CANCELLED"
ALLOWED_TRANSITIONS: Mapping[HermesJobState, frozenset[HermesJobState]] = {
HermesJobState.QUEUED: frozenset({HermesJobState.CLAIMED, HermesJobState.CANCELLED}),
HermesJobState.CLAIMED: frozenset({HermesJobState.RUNNING, HermesJobState.CANCELLED}),
HermesJobState.RUNNING: frozenset(
{HermesJobState.COMPLETED, HermesJobState.FAILED, HermesJobState.CANCELLED}
),
HermesJobState.COMPLETED: frozenset(),
HermesJobState.FAILED: frozenset(),
HermesJobState.CANCELLED: frozenset(),
}
TERMINAL_STATES = frozenset(
{HermesJobState.COMPLETED, HermesJobState.FAILED, HermesJobState.CANCELLED}
)
@dataclass(frozen=True)
class HermesJob:
job_id: str
tenant_id: str
cell_id: str
profile_id: str
task_id: str
trace_id: str
action: str
payload: Mapping[str, Any] = field(default_factory=dict)
idempotency_key: str = ""
priority: int = 100
state: HermesJobState = HermesJobState.QUEUED
revision: int = 1
assigned_worker_id: str | None = None
lease_expires_at: str | None = None
created_at: str = field(default_factory=utc_now)
updated_at: str = field(default_factory=utc_now)
def __post_init__(self) -> None:
for name in (
"job_id", "tenant_id", "cell_id", "profile_id", "task_id",
"trace_id", "action", "idempotency_key", "created_at", "updated_at",
):
object.__setattr__(self, name, _required(getattr(self, name), name))
object.__setattr__(self, "payload", _safe_mapping(self.payload))
object.__setattr__(self, "state", HermesJobState(self.state))
if not 0 <= int(self.priority) <= 1000:
raise ValueError("priority must be between 0 and 1000")
if self.revision < 1:
raise ValueError("revision must be >= 1")
@classmethod
def create(
cls,
*,
tenant_id: str,
cell_id: str,
profile_id: str,
task_id: str,
trace_id: str,
action: str,
payload: Mapping[str, Any] | None,
idempotency_key: str,
priority: int = 100,
) -> "HermesJob":
return cls(
job_id=f"hermes_{uuid.uuid4().hex}",
tenant_id=tenant_id,
cell_id=cell_id,
profile_id=profile_id,
task_id=task_id,
trace_id=trace_id,
action=action,
payload=_safe_mapping(payload),
idempotency_key=idempotency_key,
priority=priority,
)
@dataclass(frozen=True)
class HermesResult:
job_id: str
tenant_id: str
task_id: str
trace_id: str
state: HermesJobState
worker_id: str | None = None
output: Mapping[str, Any] | None = None
error: Mapping[str, Any] | None = None
started_at: str | None = None
finished_at: str | None = None
def __post_init__(self) -> None:
for name in ("job_id", "tenant_id", "task_id", "trace_id"):
object.__setattr__(self, name, _required(getattr(self, name), name))
object.__setattr__(self, "state", HermesJobState(self.state))
if self.output is not None:
object.__setattr__(self, "output", _safe_mapping(self.output))
if self.error is not None:
object.__setattr__(self, "error", _safe_mapping(self.error))
@dataclass(frozen=True)
class HermesHealth:
runtime_id: str
status: str
version: str
heartbeat_at: str
active_workers: int = 0
active_jobs: int = 0
queued_jobs: int = 0
tenant_partitions: int = 0
profile_isolation_enforced: bool = True
details: Mapping[str, Any] = field(default_factory=dict)
def __post_init__(self) -> None:
for name in ("runtime_id", "status", "version", "heartbeat_at"):
object.__setattr__(self, name, _required(getattr(self, name), name))
for name in ("active_workers", "active_jobs", "queued_jobs", "tenant_partitions"):
if int(getattr(self, name)) < 0:
raise ValueError(f"{name} must be >= 0")
object.__setattr__(self, "details", _safe_mapping(self.details))
+155
View File
@@ -0,0 +1,155 @@
"""Hermes worker loop with lease renewal and bounded recovery."""
from __future__ import annotations
import threading
from dataclasses import dataclass
from typing import Any, Callable, Mapping
from .adapter import HermesAdapter, HermesConflict
from .contract import HermesJob, HermesJobState
JobHandler = Callable[[HermesJob], Mapping[str, Any]]
@dataclass(frozen=True)
class WorkerRunResult:
claimed: bool
job_id: str | None = None
state: str | None = None
error: str | None = None
class HermesWorker:
"""Poll one tenant/cell/profile partition and execute jobs safely."""
def __init__(
self,
adapter: HermesAdapter,
handler: JobHandler,
*,
worker_id: str,
tenant_id: str,
cell_id: str,
profile_id: str,
poll_seconds: float = 1.0,
heartbeat_seconds: float = 10.0,
) -> None:
if poll_seconds <= 0 or heartbeat_seconds <= 0:
raise ValueError("poll_seconds and heartbeat_seconds must be positive")
for name, value in {
"worker_id": worker_id,
"tenant_id": tenant_id,
"cell_id": cell_id,
"profile_id": profile_id,
}.items():
if not str(value or "").strip():
raise ValueError(f"{name} is required")
self.adapter = adapter
self.handler = handler
self.worker_id = worker_id
self.tenant_id = tenant_id
self.cell_id = cell_id
self.profile_id = profile_id
self.poll_seconds = float(poll_seconds)
self.heartbeat_seconds = float(heartbeat_seconds)
self._stop = threading.Event()
def stop(self) -> None:
self._stop.set()
def run_once(self) -> WorkerRunResult:
job = self.adapter.claim(
tenant_id=self.tenant_id,
cell_id=self.cell_id,
profile_id=self.profile_id,
worker_id=self.worker_id,
)
if job is None:
return WorkerRunResult(claimed=False)
running = self.adapter.start(
job.job_id,
tenant_id=self.tenant_id,
worker_id=self.worker_id,
expected_revision=job.revision,
)
heartbeat_stop = threading.Event()
heartbeat_thread = threading.Thread(
target=self._heartbeat_loop,
args=(running.job_id, heartbeat_stop),
name=f"hermes-heartbeat-{self.worker_id}",
daemon=True,
)
heartbeat_thread.start()
output: Mapping[str, Any] | None = None
execution_error: Exception | None = None
try:
output = dict(self.handler(running))
except Exception as exc: # runtime boundary: normalize below
execution_error = exc
finally:
# Stop the lease heartbeat before reading the final revision. This
# prevents a revision increment racing with complete/fail.
heartbeat_stop.set()
heartbeat_thread.join(timeout=max(self.heartbeat_seconds * 2, 1.0))
try:
current = self.adapter.get(running.job_id, tenant_id=self.tenant_id)
if execution_error is None:
result = self.adapter.complete(
running.job_id,
tenant_id=self.tenant_id,
worker_id=self.worker_id,
output=dict(output or {}),
expected_revision=current.revision,
)
return WorkerRunResult(
claimed=True,
job_id=result.job_id,
state=result.state.value,
)
result = self.adapter.fail(
running.job_id,
tenant_id=self.tenant_id,
worker_id=self.worker_id,
error={
"code": "worker_error",
"message": str(execution_error),
},
expected_revision=current.revision,
)
return WorkerRunResult(
claimed=True,
job_id=result.job_id,
state=result.state.value,
error=str(execution_error),
)
except HermesConflict as exc:
return WorkerRunResult(
claimed=True,
job_id=running.job_id,
state=HermesJobState.FAILED.value,
error=f"reconciliation_conflict: {exc}",
)
def run_forever(self) -> None:
while not self._stop.is_set():
result = self.run_once()
if not result.claimed:
self._stop.wait(self.poll_seconds)
def _heartbeat_loop(self, job_id: str, stop_event: threading.Event) -> None:
while not stop_event.wait(self.heartbeat_seconds):
try:
self.adapter.heartbeat(
job_id,
tenant_id=self.tenant_id,
worker_id=self.worker_id,
)
except Exception:
stop_event.set()
return
@@ -0,0 +1,37 @@
"""AIOS-managed OpenClaw execution adapter."""
from .adapter import (
AdapterError,
CommandConflict,
CommandNotFound,
CommandRejected,
InMemoryCallbackSink,
OpenClawAdapter,
)
from .contract import (
CommandState,
OpenClawCommand,
OpenClawResult,
RuntimeHealth,
sign_command,
verify_command_signature,
)
from .transport import HTTPRuntimeTransport, InProcessRuntimeTransport, RuntimeTransport
__all__ = [
"AdapterError",
"CommandConflict",
"CommandNotFound",
"CommandRejected",
"CommandState",
"HTTPRuntimeTransport",
"InMemoryCallbackSink",
"InProcessRuntimeTransport",
"OpenClawAdapter",
"OpenClawCommand",
"OpenClawResult",
"RuntimeHealth",
"RuntimeTransport",
"sign_command",
"verify_command_signature",
]
+228
View File
@@ -0,0 +1,228 @@
"""Signed, tenant-isolated execution adapter for OpenClaw runtimes."""
from __future__ import annotations
import threading
from dataclasses import replace
from typing import Callable, Mapping, Protocol
from .contract import (
CommandState,
OpenClawCommand,
OpenClawResult,
RuntimeHealth,
utc_now,
verify_command_signature,
)
class AdapterError(RuntimeError):
"""Base adapter failure."""
class CommandRejected(AdapterError):
"""A command failed signature or contract validation."""
class CommandNotFound(AdapterError):
"""A tenant-scoped command is not visible."""
class CommandConflict(AdapterError):
"""A command or idempotency key conflicts with an existing execution."""
class RuntimeTransport(Protocol):
def execute(self, command: OpenClawCommand, timeout_seconds: float) -> OpenClawResult: ...
def cancel(self, command_id: str) -> bool: ...
def health(self) -> RuntimeHealth: ...
class CallbackSink(Protocol):
def on_started(self, result: OpenClawResult) -> None: ...
def on_completed(self, result: OpenClawResult) -> None: ...
def on_failed(self, result: OpenClawResult) -> None: ...
class InMemoryCallbackSink:
def __init__(self) -> None:
self._lock = threading.RLock()
self.started: list[OpenClawResult] = []
self.completed: list[OpenClawResult] = []
self.failed: list[OpenClawResult] = []
def on_started(self, result: OpenClawResult) -> None:
with self._lock:
self.started.append(result)
def on_completed(self, result: OpenClawResult) -> None:
with self._lock:
self.completed.append(result)
def on_failed(self, result: OpenClawResult) -> None:
with self._lock:
self.failed.append(result)
SigningMaterialResolver = Callable[[str, str], bytes | str]
class OpenClawAdapter:
"""Verify, de-duplicate, execute, cancel and reconcile commands."""
def __init__(
self,
transport: RuntimeTransport,
signing_material_resolver: SigningMaterialResolver,
*,
callback_sink: CallbackSink | None = None,
default_timeout_seconds: float = 300.0,
) -> None:
if default_timeout_seconds <= 0:
raise ValueError("default_timeout_seconds must be positive")
self.transport = transport
self.signing_material_resolver = signing_material_resolver
self.callback_sink = callback_sink or InMemoryCallbackSink()
self.default_timeout_seconds = float(default_timeout_seconds)
self._lock = threading.RLock()
self._commands: dict[str, OpenClawCommand] = {}
self._results: dict[str, OpenClawResult] = {}
self._idempotency: dict[tuple[str, str], tuple[str, str]] = {}
def _verify(self, command: OpenClawCommand) -> None:
try:
key_material = self.signing_material_resolver(command.tenant_id, command.key_id)
except Exception as exc:
raise CommandRejected("signing material unavailable") from exc
if not verify_command_signature(command, key_material):
raise CommandRejected("invalid or expired command signature")
@staticmethod
def _validate_result(command: OpenClawCommand, result: OpenClawResult) -> None:
if (
result.command_id != command.command_id
or result.tenant_id != command.tenant_id
or result.task_id != command.task_id
or result.trace_id != command.trace_id
):
raise AdapterError("runtime returned mismatched command identity")
if result.state not in {
CommandState.COMPLETED,
CommandState.FAILED,
CommandState.CANCELLED,
CommandState.TIMED_OUT,
}:
raise AdapterError("runtime returned a non-terminal result")
def execute(
self,
command: OpenClawCommand,
*,
timeout_seconds: float | None = None,
) -> OpenClawResult:
timeout = self.default_timeout_seconds if timeout_seconds is None else float(timeout_seconds)
if timeout <= 0:
raise ValueError("timeout_seconds must be positive")
self._verify(command)
digest = command.signature
idem_ref = (command.tenant_id, command.idempotency_key)
with self._lock:
prior_command = self._commands.get(command.command_id)
if prior_command is not None:
if prior_command.signature != command.signature:
raise CommandConflict("command_id reused with different content")
return self._results[command.command_id]
prior_idem = self._idempotency.get(idem_ref)
if prior_idem is not None:
prior_id, prior_digest = prior_idem
if prior_digest != digest:
raise CommandConflict("idempotency key reused with different content")
return self._results[prior_id]
running = OpenClawResult(
command_id=command.command_id,
tenant_id=command.tenant_id,
task_id=command.task_id,
trace_id=command.trace_id,
state=CommandState.RUNNING,
started_at=utc_now(),
)
self._commands[command.command_id] = command
self._results[command.command_id] = running
self._idempotency[idem_ref] = (command.command_id, digest)
self.callback_sink.on_started(running)
try:
result = self.transport.execute(command, timeout)
self._validate_result(command, result)
except TimeoutError as exc:
result = replace(
running,
state=CommandState.TIMED_OUT,
error={"code": "runtime_timeout", "message": str(exc) or "runtime timeout"},
finished_at=utc_now(),
)
except Exception as exc:
result = replace(
running,
state=CommandState.FAILED,
error={"code": "runtime_error", "message": str(exc)},
finished_at=utc_now(),
)
if result.started_at is None:
result = replace(result, started_at=running.started_at)
if result.finished_at is None:
result = replace(result, finished_at=utc_now())
with self._lock:
self._results[command.command_id] = result
if result.state == CommandState.COMPLETED:
self.callback_sink.on_completed(result)
else:
self.callback_sink.on_failed(result)
return result
def status(self, command_id: str, *, tenant_id: str) -> OpenClawResult:
with self._lock:
command = self._commands.get(command_id)
result = self._results.get(command_id)
if command is None or result is None or command.tenant_id != tenant_id:
raise CommandNotFound("command not found")
return result
def cancel(self, command_id: str, *, tenant_id: str) -> OpenClawResult:
with self._lock:
command = self._commands.get(command_id)
current = self._results.get(command_id)
if command is None or current is None or command.tenant_id != tenant_id:
raise CommandNotFound("command not found")
if current.state in {
CommandState.COMPLETED,
CommandState.FAILED,
CommandState.CANCELLED,
CommandState.TIMED_OUT,
CommandState.REJECTED,
}:
return current
if not self.transport.cancel(command_id):
raise AdapterError("runtime rejected cancellation")
cancelled = replace(
current,
state=CommandState.CANCELLED,
error={"code": "cancelled", "message": "cancelled by AIOS"},
finished_at=utc_now(),
)
with self._lock:
self._results[command_id] = cancelled
self.callback_sink.on_failed(cancelled)
return cancelled
def health(self) -> RuntimeHealth:
return self.transport.health()
def snapshot(self, *, tenant_id: str) -> Mapping[str, OpenClawResult]:
with self._lock:
return {
command_id: result
for command_id, result in self._results.items()
if self._commands[command_id].tenant_id == tenant_id
}
@@ -0,0 +1,97 @@
"""Audit-oriented lifecycle callbacks for OpenClaw execution."""
from __future__ import annotations
import json
import threading
from dataclasses import dataclass, field
from typing import Any, Mapping, Protocol
from .contract import OpenClawResult, utc_now
class EventPublisher(Protocol):
def publish(self, event: Mapping[str, Any]) -> None: ...
@dataclass(frozen=True)
class ExecutionEvent:
event_type: str
tenant_id: str
task_id: str
trace_id: str
command_id: str
state: str
created_at: str = field(default_factory=utc_now)
details: Mapping[str, Any] = field(default_factory=dict)
def as_dict(self) -> dict[str, Any]:
return {
"event_type": self.event_type,
"tenant_id": self.tenant_id,
"task_id": self.task_id,
"trace_id": self.trace_id,
"command_id": self.command_id,
"state": self.state,
"created_at": self.created_at,
"details": json.loads(
json.dumps(dict(self.details), ensure_ascii=False, default=str)
),
}
class InMemoryEventPublisher:
"""Thread-safe publisher used by tests and first integration slice."""
def __init__(self) -> None:
self._lock = threading.RLock()
self.events: list[dict[str, Any]] = []
def publish(self, event: Mapping[str, Any]) -> None:
with self._lock:
self.events.append(
json.loads(json.dumps(dict(event), ensure_ascii=False, default=str))
)
class AuditCallbackSink:
"""Convert runtime callbacks into tenant-scoped audit events."""
def __init__(self, publisher: EventPublisher) -> None:
self.publisher = publisher
@staticmethod
def _event(
event_type: str,
result: OpenClawResult,
) -> ExecutionEvent:
details: dict[str, Any] = {
"runtime_id": result.runtime_id,
"attempts": result.attempts,
}
if result.error:
details["error"] = dict(result.error)
if result.output:
details["output_present"] = True
return ExecutionEvent(
event_type=event_type,
tenant_id=result.tenant_id,
task_id=result.task_id,
trace_id=result.trace_id,
command_id=result.command_id,
state=result.state.value,
details=details,
)
def on_started(self, result: OpenClawResult) -> None:
self.publisher.publish(self._event("OPENCLAW_STARTED", result).as_dict())
def on_completed(self, result: OpenClawResult) -> None:
self.publisher.publish(self._event("OPENCLAW_COMPLETED", result).as_dict())
def on_failed(self, result: OpenClawResult) -> None:
event_type = {
"CANCELLED": "OPENCLAW_CANCELLED",
"TIMED_OUT": "OPENCLAW_TIMED_OUT",
"REJECTED": "OPENCLAW_REJECTED",
}.get(result.state.value, "OPENCLAW_FAILED")
self.publisher.publish(self._event(event_type, result).as_dict())
@@ -0,0 +1,213 @@
"""Versioned AIOS to OpenClaw execution contracts."""
from __future__ import annotations
import hashlib
import hmac
import json
import uuid
from dataclasses import dataclass, field, replace
from datetime import datetime, timezone
from enum import Enum
from typing import Any, Mapping
CONTRACT_VERSION = "aios.openclaw.command.v1"
DEFAULT_CLOCK_SKEW_SECONDS = 300
def utc_now() -> str:
return datetime.now(timezone.utc).isoformat().replace("+00:00", "Z")
def _parse_timestamp(value: str) -> datetime:
normalized = str(value or "").strip()
if not normalized:
raise ValueError("timestamp is required")
if normalized.endswith("Z"):
normalized = normalized[:-1] + "+00:00"
parsed = datetime.fromisoformat(normalized)
if parsed.tzinfo is None:
raise ValueError("timestamp must include a timezone")
return parsed.astimezone(timezone.utc)
def _required(value: str, label: str) -> str:
normalized = str(value or "").strip()
if not normalized:
raise ValueError(f"{label} is required")
return normalized
def _safe_mapping(value: Mapping[str, Any] | None) -> dict[str, Any]:
return json.loads(json.dumps(dict(value or {}), ensure_ascii=False, sort_keys=True, default=str))
class CommandState(str, Enum):
ACCEPTED = "ACCEPTED"
RUNNING = "RUNNING"
COMPLETED = "COMPLETED"
FAILED = "FAILED"
CANCELLED = "CANCELLED"
TIMED_OUT = "TIMED_OUT"
REJECTED = "REJECTED"
@dataclass(frozen=True)
class OpenClawCommand:
command_id: str
tenant_id: str
task_id: str
trace_id: str
action: str
payload: Mapping[str, Any] = field(default_factory=dict)
idempotency_key: str = ""
created_at: str = field(default_factory=utc_now)
expires_at: str | None = None
contract_version: str = CONTRACT_VERSION
key_id: str = "default"
signature: str = ""
def __post_init__(self) -> None:
for field_name in (
"command_id", "tenant_id", "task_id", "trace_id", "action",
"idempotency_key", "created_at", "contract_version", "key_id",
):
object.__setattr__(self, field_name, _required(getattr(self, field_name), field_name))
if self.contract_version != CONTRACT_VERSION:
raise ValueError(f"unsupported contract_version: {self.contract_version}")
object.__setattr__(self, "payload", _safe_mapping(self.payload))
_parse_timestamp(self.created_at)
if self.expires_at is not None:
object.__setattr__(self, "expires_at", _required(self.expires_at, "expires_at"))
_parse_timestamp(self.expires_at)
@classmethod
def create(
cls,
*,
tenant_id: str,
task_id: str,
trace_id: str,
action: str,
payload: Mapping[str, Any] | None,
idempotency_key: str,
expires_at: str | None = None,
key_id: str = "default",
command_id: str | None = None,
) -> "OpenClawCommand":
return cls(
command_id=command_id or f"cmd_{uuid.uuid4().hex}",
tenant_id=tenant_id,
task_id=task_id,
trace_id=trace_id,
action=action,
payload=_safe_mapping(payload),
idempotency_key=idempotency_key,
expires_at=expires_at,
key_id=key_id,
)
def unsigned_dict(self) -> dict[str, Any]:
return {
"command_id": self.command_id,
"tenant_id": self.tenant_id,
"task_id": self.task_id,
"trace_id": self.trace_id,
"action": self.action,
"payload": _safe_mapping(self.payload),
"idempotency_key": self.idempotency_key,
"created_at": self.created_at,
"expires_at": self.expires_at,
"contract_version": self.contract_version,
"key_id": self.key_id,
}
@dataclass(frozen=True)
class OpenClawResult:
command_id: str
tenant_id: str
task_id: str
trace_id: str
state: CommandState
output: Mapping[str, Any] | None = None
error: Mapping[str, Any] | None = None
started_at: str | None = None
finished_at: str | None = None
runtime_id: str | None = None
attempts: int = 1
def __post_init__(self) -> None:
for field_name in ("command_id", "tenant_id", "task_id", "trace_id"):
object.__setattr__(self, field_name, _required(getattr(self, field_name), field_name))
object.__setattr__(self, "state", CommandState(self.state))
if self.output is not None:
object.__setattr__(self, "output", _safe_mapping(self.output))
if self.error is not None:
object.__setattr__(self, "error", _safe_mapping(self.error))
if self.started_at is not None:
_parse_timestamp(self.started_at)
if self.finished_at is not None:
_parse_timestamp(self.finished_at)
if self.attempts < 1:
raise ValueError("attempts must be >= 1")
@dataclass(frozen=True)
class RuntimeHealth:
runtime: str
runtime_id: str
status: str
version: str
heartbeat_at: str
active_tasks: int = 0
workspace_isolated: bool = False
credential_ref_configured: bool = False
details: Mapping[str, Any] = field(default_factory=dict)
def __post_init__(self) -> None:
for field_name in ("runtime", "runtime_id", "status", "version", "heartbeat_at"):
object.__setattr__(self, field_name, _required(getattr(self, field_name), field_name))
_parse_timestamp(self.heartbeat_at)
if self.active_tasks < 0:
raise ValueError("active_tasks must be >= 0")
object.__setattr__(self, "details", _safe_mapping(self.details))
def canonical_command_bytes(command: OpenClawCommand) -> bytes:
return json.dumps(
command.unsigned_dict(), ensure_ascii=False, sort_keys=True, separators=(",", ":")
).encode("utf-8")
def sign_command(command: OpenClawCommand, key_material: bytes | str) -> OpenClawCommand:
key = key_material.encode("utf-8") if isinstance(key_material, str) else bytes(key_material)
if not key:
raise ValueError("signing material is required")
signature = hmac.new(key, canonical_command_bytes(command), hashlib.sha256).hexdigest()
return replace(command, signature=signature)
def verify_command_signature(
command: OpenClawCommand,
key_material: bytes | str,
*,
now: datetime | None = None,
clock_skew_seconds: int = DEFAULT_CLOCK_SKEW_SECONDS,
) -> bool:
if not command.signature:
return False
key = key_material.encode("utf-8") if isinstance(key_material, str) else bytes(key_material)
if not key:
return False
expected = hmac.new(key, canonical_command_bytes(command), hashlib.sha256).hexdigest()
if not hmac.compare_digest(expected, command.signature):
return False
current = (now or datetime.now(timezone.utc)).astimezone(timezone.utc)
created = _parse_timestamp(command.created_at)
if created.timestamp() - current.timestamp() > clock_skew_seconds:
return False
if command.expires_at is not None:
expires = _parse_timestamp(command.expires_at)
if current.timestamp() - expires.timestamp() > clock_skew_seconds:
return False
return True
+103
View File
@@ -0,0 +1,103 @@
"""Health policy and self-healing decisions for OpenClaw runtimes."""
from __future__ import annotations
from dataclasses import dataclass, field
from datetime import datetime, timezone
from typing import Mapping
from .contract import RuntimeHealth
def _parse(value: str) -> datetime:
normalized = value[:-1] + "+00:00" if value.endswith("Z") else value
parsed = datetime.fromisoformat(normalized)
if parsed.tzinfo is None:
raise ValueError("heartbeat timestamp must include timezone")
return parsed.astimezone(timezone.utc)
@dataclass(frozen=True)
class HealthPolicy:
max_heartbeat_age_seconds: int = 90
require_workspace_isolation: bool = True
require_credential_ref: bool = True
accepted_statuses: frozenset[str] = field(
default_factory=lambda: frozenset({"healthy", "ready"})
)
def __post_init__(self) -> None:
if self.max_heartbeat_age_seconds < 5:
raise ValueError("max_heartbeat_age_seconds must be >= 5")
if not self.accepted_statuses:
raise ValueError("accepted_statuses must not be empty")
@dataclass(frozen=True)
class HealthAssessment:
healthy: bool
reasons: tuple[str, ...]
recoverable: bool
recommended_action: str
details: Mapping[str, object] = field(default_factory=dict)
class RuntimeHealthMonitor:
"""Evaluate runtime health and recommend bounded recovery actions."""
def __init__(self, policy: HealthPolicy | None = None) -> None:
self.policy = policy or HealthPolicy()
def assess(
self,
health: RuntimeHealth,
*,
now: datetime | None = None,
) -> HealthAssessment:
current = (now or datetime.now(timezone.utc)).astimezone(timezone.utc)
heartbeat = _parse(health.heartbeat_at)
age = max(0.0, current.timestamp() - heartbeat.timestamp())
reasons: list[str] = []
if health.status.lower() not in self.policy.accepted_statuses:
reasons.append("runtime_status_unhealthy")
if age > self.policy.max_heartbeat_age_seconds:
reasons.append("heartbeat_stale")
if self.policy.require_workspace_isolation and not health.workspace_isolated:
reasons.append("workspace_not_isolated")
if self.policy.require_credential_ref and not health.credential_ref_configured:
reasons.append("credential_reference_missing")
if not reasons:
return HealthAssessment(
healthy=True,
reasons=(),
recoverable=True,
recommended_action="none",
details={"heartbeat_age_seconds": age},
)
non_recoverable = {
"workspace_not_isolated",
"credential_reference_missing",
}.intersection(reasons)
if non_recoverable:
action = "quarantine_runtime"
recoverable = False
elif "heartbeat_stale" in reasons:
action = "restart_runtime"
recoverable = True
else:
action = "drain_and_restart"
recoverable = True
return HealthAssessment(
healthy=False,
reasons=tuple(reasons),
recoverable=recoverable,
recommended_action=action,
details={
"heartbeat_age_seconds": age,
"runtime_id": health.runtime_id,
"active_tasks": health.active_tasks,
},
)
@@ -0,0 +1,214 @@
"""Runtime transport implementations for the OpenClaw adapter."""
from __future__ import annotations
import json
import threading
import urllib.error
import urllib.parse
import urllib.request
from typing import Any, Callable, Mapping, Protocol
from .contract import (
CommandState,
OpenClawCommand,
OpenClawResult,
RuntimeHealth,
utc_now,
)
class RuntimeTransport(Protocol):
def execute(self, command: OpenClawCommand, timeout_seconds: float) -> OpenClawResult: ...
def cancel(self, command_id: str) -> bool: ...
def health(self) -> RuntimeHealth: ...
class TransportError(RuntimeError):
"""Runtime transport or protocol failure."""
RuntimeHandler = Callable[[OpenClawCommand], Mapping[str, Any] | OpenClawResult]
class InProcessRuntimeTransport:
"""Deterministic runtime used by tests and controlled local integrations."""
def __init__(
self,
handler: RuntimeHandler,
*,
runtime_id: str = "openclaw-in-process",
version: str = "v1",
workspace_isolated: bool = True,
credential_ref_configured: bool = True,
) -> None:
self.handler = handler
self.runtime_id = runtime_id
self.version = version
self.workspace_isolated = workspace_isolated
self.credential_ref_configured = credential_ref_configured
self._lock = threading.RLock()
self._running: set[str] = set()
self._cancelled: set[str] = set()
def execute(self, command: OpenClawCommand, timeout_seconds: float) -> OpenClawResult:
if timeout_seconds <= 0:
raise ValueError("timeout_seconds must be positive")
with self._lock:
if command.command_id in self._cancelled:
return OpenClawResult(
command_id=command.command_id,
tenant_id=command.tenant_id,
task_id=command.task_id,
trace_id=command.trace_id,
state=CommandState.CANCELLED,
error={"code": "cancelled", "message": "cancelled before execution"},
finished_at=utc_now(),
runtime_id=self.runtime_id,
)
self._running.add(command.command_id)
try:
raw = self.handler(command)
if isinstance(raw, OpenClawResult):
return raw
if not isinstance(raw, Mapping):
raise TransportError("runtime handler must return a mapping or OpenClawResult")
return OpenClawResult(
command_id=command.command_id,
tenant_id=command.tenant_id,
task_id=command.task_id,
trace_id=command.trace_id,
state=CommandState(str(raw.get("state", "COMPLETED"))),
output=raw.get("output"),
error=raw.get("error"),
started_at=str(raw.get("started_at") or utc_now()),
finished_at=str(raw.get("finished_at") or utc_now()),
runtime_id=str(raw.get("runtime_id") or self.runtime_id),
attempts=int(raw.get("attempts", 1)),
)
finally:
with self._lock:
self._running.discard(command.command_id)
def cancel(self, command_id: str) -> bool:
with self._lock:
self._cancelled.add(command_id)
return command_id in self._running
def health(self) -> RuntimeHealth:
with self._lock:
active = len(self._running)
return RuntimeHealth(
runtime="openclaw",
runtime_id=self.runtime_id,
status="healthy",
version=self.version,
heartbeat_at=utc_now(),
active_tasks=active,
workspace_isolated=self.workspace_isolated,
credential_ref_configured=self.credential_ref_configured,
details={"transport": "in_process"},
)
class HTTPRuntimeTransport:
"""Private HTTP transport for an OpenClaw runtime service."""
def __init__(
self,
base_url: str,
*,
runtime_id: str = "openclaw-http",
credential_ref: str | None = None,
max_response_bytes: int = 2_097_152,
) -> None:
normalized = str(base_url).rstrip("/")
if not normalized.startswith(("http://", "https://")):
raise ValueError("base_url must use http or https")
self.base_url = normalized
self.runtime_id = runtime_id
self.credential_ref = credential_ref
self.max_response_bytes = int(max_response_bytes)
def _request(
self,
method: str,
path: str,
*,
payload: Mapping[str, Any] | None,
timeout_seconds: float,
) -> Mapping[str, Any]:
body = None
headers = {"Accept": "application/json", "User-Agent": "AIOS-OpenClaw/1"}
if payload is not None:
body = json.dumps(dict(payload), separators=(",", ":")).encode("utf-8")
headers["Content-Type"] = "application/json"
if self.credential_ref:
headers["X-AIOS-Credential-Ref"] = self.credential_ref
request = urllib.request.Request(
f"{self.base_url}{path}",
data=body,
method=method,
headers=headers,
)
try:
with urllib.request.urlopen(request, timeout=timeout_seconds) as response:
raw = response.read(self.max_response_bytes + 1)
except urllib.error.HTTPError as exc:
raise TransportError(f"runtime returned HTTP {exc.code}") from exc
except urllib.error.URLError as exc:
raise TransportError(f"runtime unavailable: {exc.reason}") from exc
if len(raw) > self.max_response_bytes:
raise TransportError("runtime response exceeded size limit")
decoded = json.loads(raw.decode("utf-8"))
if not isinstance(decoded, Mapping):
raise TransportError("runtime response must be a JSON object")
return decoded
@staticmethod
def _command_payload(command: OpenClawCommand) -> dict[str, Any]:
return {**command.unsigned_dict(), "signature": command.signature}
def execute(self, command: OpenClawCommand, timeout_seconds: float) -> OpenClawResult:
value = self._request(
"POST",
"/v1/execute",
payload=self._command_payload(command),
timeout_seconds=timeout_seconds,
)
return OpenClawResult(
command_id=str(value["command_id"]),
tenant_id=str(value["tenant_id"]),
task_id=str(value["task_id"]),
trace_id=str(value["trace_id"]),
state=CommandState(str(value["state"])),
output=value.get("output"),
error=value.get("error"),
started_at=value.get("started_at"),
finished_at=value.get("finished_at"),
runtime_id=str(value.get("runtime_id") or self.runtime_id),
attempts=int(value.get("attempts", 1)),
)
def cancel(self, command_id: str) -> bool:
value = self._request(
"POST",
f"/v1/commands/{urllib.parse.quote(command_id, safe='')}/cancel",
payload={},
timeout_seconds=10.0,
)
return bool(value.get("accepted"))
def health(self) -> RuntimeHealth:
value = self._request("GET", "/v1/health", payload=None, timeout_seconds=5.0)
return RuntimeHealth(
runtime=str(value.get("runtime") or "openclaw"),
runtime_id=str(value.get("runtime_id") or self.runtime_id),
status=str(value.get("status") or "unknown"),
version=str(value.get("version") or "unknown"),
heartbeat_at=str(value.get("heartbeat_at") or utc_now()),
active_tasks=int(value.get("active_tasks", 0)),
workspace_isolated=bool(value.get("workspace_isolated", False)),
credential_ref_configured=bool(value.get("credential_ref_configured", False)),
details=value.get("details") if isinstance(value.get("details"), Mapping) else {},
)
+46
View File
@@ -0,0 +1,46 @@
"""AIOS V2 control plane and execution orchestration."""
from .control_plane import (
AgentRecord,
AuditEvent,
ControlPlaneError,
ControlPlaneService,
IdempotencyConflict,
InMemoryControlPlaneStore,
InvalidTransition,
NotFound,
PolicyDenied,
RequestContext,
RevisionConflict,
TaskRecord,
TaskState,
TenantIsolationError,
TenantRecord,
Unauthorized,
)
from .execution_tracking import ExecutionEvidence, ExecutionTracker
from .orchestrator import DispatchRecord, ExecutionOrchestrator, OrchestrationError
__all__ = [
"AgentRecord",
"AuditEvent",
"ControlPlaneError",
"ControlPlaneService",
"DispatchRecord",
"ExecutionEvidence",
"ExecutionOrchestrator",
"ExecutionTracker",
"IdempotencyConflict",
"InMemoryControlPlaneStore",
"InvalidTransition",
"NotFound",
"OrchestrationError",
"PolicyDenied",
"RequestContext",
"RevisionConflict",
"TaskRecord",
"TaskState",
"TenantIsolationError",
"TenantRecord",
"Unauthorized",
]
+357
View File
@@ -0,0 +1,357 @@
"""Production-candidate AIOS V2 application entrypoint.
This process owns the public control-plane API and read-only web console. It
loads durable control-plane state, dispatches approved work through runtime
adapters, exposes tenant-scoped execution evidence and sanitized GitOps status,
and never exposes runtime credentials to the browser.
"""
from __future__ import annotations
import dataclasses
import json
import os
import stat
import urllib.parse
from enum import Enum
from http import HTTPStatus
from http.server import BaseHTTPRequestHandler, ThreadingHTTPServer
from pathlib import Path
from typing import Any, Mapping
from aios_hermes import HermesAdapter, HermesJobState, HermesResult
from aios_openclaw import HTTPRuntimeTransport, OpenClawAdapter
from aios_openclaw.callback import AuditCallbackSink, EventPublisher
from .console import CONSOLE_HTML
from .control_plane import AuditEvent, RequestContext
from .execution_tracking import ExecutionTracker
from .gateway import GatewayApplication, error_response
from .gitops_status import collect_gitops_status
from .orchestrator import ExecutionOrchestrator
from .persistence import DurableControlPlaneService, SQLiteSnapshotRepository
def _json_value(value: Any) -> Any:
if dataclasses.is_dataclass(value):
return {
item.name: _json_value(getattr(value, item.name))
for item in dataclasses.fields(value)
}
if isinstance(value, Enum):
return value.value
if isinstance(value, Mapping):
return {str(key): _json_value(item) for key, item in value.items()}
if isinstance(value, (list, tuple, set, frozenset)):
return [_json_value(item) for item in value]
return value
class FileSigningMaterialResolver:
"""Resolve runtime signing material from a root-owned file reference."""
def __init__(self, path: str | Path) -> None:
self.path = Path(path).resolve()
def __call__(self, tenant_id: str, key_id: str) -> bytes:
del tenant_id, key_id
metadata = self.path.stat()
if not stat.S_ISREG(metadata.st_mode):
raise RuntimeError("signing material reference is not a regular file")
if metadata.st_mode & 0o077:
raise RuntimeError("signing material file permissions must be 0600 or stricter")
value = self.path.read_bytes().strip()
if len(value) < 32:
raise RuntimeError("signing material is too short")
return value
class ControlPlaneEventPublisher(EventPublisher):
"""Append OpenClaw runtime events to the tenant audit stream."""
def __init__(self, service: DurableControlPlaneService) -> None:
self.service = service
def publish(self, event: Mapping[str, Any]) -> None:
tenant_id = str(event.get("tenant_id") or "").strip()
trace_id = str(event.get("trace_id") or "").strip()
command_id = str(event.get("command_id") or "").strip()
event_type = str(event.get("event_type") or "OPENCLAW_EVENT").strip()
if not tenant_id or not trace_id or not command_id:
raise ValueError("runtime audit event is missing tenant identity")
state_value = str(event.get("state") or "unknown").lower()
outcome = "allowed" if state_value == "completed" else state_value
audit = AuditEvent(
event_id=f"runtime_{command_id}_{len(self.service.store.audit_events) + 1}",
tenant_id=tenant_id,
principal_id="openclaw-runtime",
trace_id=trace_id,
action=event_type,
resource_type="openclaw_command",
resource_id=command_id,
outcome=outcome,
policy_version="runtime-v1",
details=dict(event.get("details") or {}),
)
with self.service.store.lock:
self.service.store.audit_events.append(audit)
self.service.repository.save(self.service._snapshot())
class CandidateApplication:
"""Compose durable control plane, runtimes, evidence and console routes."""
def __init__(
self,
gateway: GatewayApplication,
orchestrator: ExecutionOrchestrator,
tracker: ExecutionTracker,
hermes: HermesAdapter,
openclaw: OpenClawAdapter | None,
) -> None:
self.gateway = gateway
self.orchestrator = orchestrator
self.tracker = tracker
self.hermes = hermes
self.openclaw = openclaw
def context(self, headers: Mapping[str, str]) -> RequestContext:
return self.gateway.context_from_headers(headers)
def handle_json(
self,
method: str,
raw_path: str,
headers: Mapping[str, str],
body: Any,
) -> tuple[int, Mapping[str, Any]]:
parsed = urllib.parse.urlparse(raw_path)
path = parsed.path.rstrip("/") or "/"
method = method.upper()
if method == "GET" and path == "/api/v2/gitops":
context = self.context(headers)
return HTTPStatus.OK, {
"tenant_id": context.tenant_id,
"gitops": collect_gitops_status().as_dict(),
}
if method == "GET" and path == "/api/v2/runtime":
context = self.context(headers)
runtimes: list[dict[str, Any]] = []
if self.openclaw is None:
runtimes.append(
{
"name": "openclaw",
"status": "disabled",
"details": {"reason": "runtime_not_configured"},
}
)
else:
try:
health = self.openclaw.health()
runtimes.append(
{
"name": "openclaw",
"status": health.status,
"details": _json_value(health),
}
)
except Exception as exc:
runtimes.append(
{
"name": "openclaw",
"status": "unavailable",
"details": {"message": str(exc)},
}
)
hermes_health = self.hermes.health()
runtimes.append(
{
"name": "hermes",
"status": hermes_health.status,
"details": _json_value(hermes_health),
}
)
return HTTPStatus.OK, {
"tenant_id": context.tenant_id,
"runtimes": runtimes,
}
if method == "GET" and path == "/api/v2/executions":
context = self.context(headers)
return HTTPStatus.OK, {
"executions": _json_value(
self.tracker.list(tenant_id=context.tenant_id)
)
}
dispatch_match = __import__("re").fullmatch(
r"/api/v2/tasks/([^/]+)/dispatch",
path,
)
if method == "POST" and dispatch_match:
context = self.context(headers)
task_id = urllib.parse.unquote(dispatch_match.group(1))
payload = body if isinstance(body, Mapping) else {}
timeout = payload.get("timeout_seconds")
record = self.orchestrator.dispatch(
context,
task_id=task_id,
timeout_seconds=float(timeout) if timeout is not None else None,
)
return HTTPStatus.ACCEPTED, {"dispatch": _json_value(record)}
if method == "POST" and path == "/api/v2/hermes/results/reconcile":
context = self.context(headers)
if not isinstance(body, Mapping):
raise ValueError("JSON body must be an object")
result = HermesResult(
job_id=str(body["job_id"]),
tenant_id=str(body["tenant_id"]),
task_id=str(body["task_id"]),
trace_id=str(body["trace_id"]),
state=HermesJobState(str(body["state"])),
worker_id=(
str(body["worker_id"])
if body.get("worker_id") is not None
else None
),
output=body.get("output"),
error=body.get("error"),
started_at=body.get("started_at"),
finished_at=body.get("finished_at"),
)
task = self.orchestrator.reconcile_hermes(context, result)
return HTTPStatus.OK, {"task": _json_value(task)}
return self.gateway.handle(method, raw_path, headers, body)
class CandidateRequestHandler(BaseHTTPRequestHandler):
server_version = "AIOSV2Candidate/1.0"
max_body_bytes = 1_048_576
@property
def application(self) -> CandidateApplication:
return self.server.application # type: ignore[attr-defined]
def _read_body(self) -> Any:
length = int(self.headers.get("Content-Length", "0") or "0")
if length < 0 or length > self.max_body_bytes:
raise ValueError("request body is too large")
if length == 0:
return {}
content_type = self.headers.get("Content-Type", "").split(";", 1)[0].strip()
if content_type != "application/json":
raise ValueError("Content-Type must be application/json")
return json.loads(self.rfile.read(length).decode("utf-8"))
def _security_headers(self) -> None:
self.send_header("Cache-Control", "no-store")
self.send_header("X-Content-Type-Options", "nosniff")
self.send_header("X-Frame-Options", "DENY")
self.send_header("Referrer-Policy", "no-referrer")
self.send_header(
"Content-Security-Policy",
"default-src 'self'; style-src 'unsafe-inline'; script-src 'unsafe-inline'; connect-src 'self'",
)
def _send_json(self, status_code: int, payload: Mapping[str, Any]) -> None:
encoded = json.dumps(
_json_value(payload),
ensure_ascii=False,
separators=(",", ":"),
).encode("utf-8")
self.send_response(int(status_code))
self.send_header("Content-Type", "application/json; charset=utf-8")
self.send_header("Content-Length", str(len(encoded)))
self._security_headers()
self.end_headers()
self.wfile.write(encoded)
def _send_console(self) -> None:
encoded = CONSOLE_HTML.encode("utf-8")
self.send_response(HTTPStatus.OK)
self.send_header("Content-Type", "text/html; charset=utf-8")
self.send_header("Content-Length", str(len(encoded)))
self._security_headers()
self.end_headers()
self.wfile.write(encoded)
def _dispatch(self, method: str) -> None:
path = urllib.parse.urlparse(self.path).path.rstrip("/") or "/"
if method == "GET" and path in {"/", "/console"}:
self._send_console()
return
try:
body = self._read_body() if method == "POST" else {}
status_code, payload = self.application.handle_json(
method,
self.path,
{key: value for key, value in self.headers.items()},
body,
)
except BaseException as exc:
status_code, payload = error_response(exc)
self._send_json(status_code, payload)
def do_GET(self) -> None: # noqa: N802
self._dispatch("GET")
def do_POST(self) -> None: # noqa: N802
self._dispatch("POST")
class CandidateHTTPServer(ThreadingHTTPServer):
daemon_threads = True
allow_reuse_address = True
def __init__(self, address: tuple[str, int], application: CandidateApplication) -> None:
self.application = application
super().__init__(address, CandidateRequestHandler)
def build_application() -> CandidateApplication:
state_dir = Path(os.getenv("AIOS_V2_STATE_DIR", "/var/lib/aios-v2")).resolve()
state_dir.mkdir(parents=True, exist_ok=True)
repository = SQLiteSnapshotRepository(state_dir / "control-plane.sqlite3")
service = DurableControlPlaneService(repository)
tracker = ExecutionTracker(state_dir / "executions")
hermes = HermesAdapter()
openclaw_url = os.getenv("AIOS_OPENCLAW_URL", "").strip()
material_file = os.getenv("AIOS_OPENCLAW_SIGNING_MATERIAL_FILE", "").strip()
openclaw = None
resolver = None
if openclaw_url and material_file:
resolver = FileSigningMaterialResolver(material_file)
transport = HTTPRuntimeTransport(
openclaw_url,
credential_ref=os.getenv("AIOS_OPENCLAW_CREDENTIAL_REF") or None,
)
openclaw = OpenClawAdapter(
transport,
resolver,
callback_sink=AuditCallbackSink(ControlPlaneEventPublisher(service)),
)
gateway = GatewayApplication(service)
orchestrator = ExecutionOrchestrator(
service,
openclaw=openclaw,
hermes=hermes,
openclaw_signing_material_resolver=resolver,
)
return CandidateApplication(gateway, orchestrator, tracker, hermes, openclaw)
def main() -> None:
host = os.getenv("AIOS_V2_HOST", "127.0.0.1")
port = int(os.getenv("AIOS_V2_PORT", "18920"))
server = CandidateHTTPServer((host, port), build_application())
print(f"AIOS V2 candidate listening on {host}:{port}")
server.serve_forever()
if __name__ == "__main__":
main()
+69
View File
@@ -0,0 +1,69 @@
"""Dependency-free AIOS V2 web console assets.
Authentication and tenant identity are injected by the trusted reverse proxy.
The browser never selects a tenant and never calls OpenClaw or Hermes directly.
"""
CONSOLE_HTML = r'''<!doctype html>
<html lang="en">
<head>
<meta charset="utf-8">
<meta name="viewport" content="width=device-width,initial-scale=1">
<meta name="referrer" content="no-referrer">
<title>AIOS V2 Console</title>
<style>
:root{font-family:Inter,ui-sans-serif,system-ui,sans-serif;color:#152033;background:#f4f7fb}
*{box-sizing:border-box} body{margin:0} header{background:#111827;color:white;padding:18px 24px;display:flex;justify-content:space-between;align-items:center}
header h1{font-size:20px;margin:0} header span{font-size:12px;color:#cbd5e1}
main{padding:24px;max-width:1440px;margin:auto}.notice{background:#fff7ed;border:1px solid #fed7aa;padding:12px 14px;border-radius:10px;margin-bottom:18px;font-size:13px}
.cards{display:grid;grid-template-columns:repeat(auto-fit,minmax(190px,1fr));gap:14px;margin-bottom:20px}.card,.panel{background:white;border:1px solid #e5e7eb;border-radius:12px;box-shadow:0 1px 2px rgba(0,0,0,.04)}
.card{padding:16px}.card b{display:block;font-size:28px;margin-top:8px}.card small{color:#64748b}.grid{display:grid;grid-template-columns:repeat(auto-fit,minmax(420px,1fr));gap:16px}.panel{padding:16px;overflow:auto}.panel h2{font-size:16px;margin:0 0 12px}
table{width:100%;border-collapse:collapse;font-size:12px}th,td{text-align:left;border-bottom:1px solid #edf0f4;padding:8px;vertical-align:top}th{color:#64748b;font-weight:600}.state{font-weight:700}.ok{color:#047857}.bad{color:#b91c1c}.muted{color:#64748b}
button{border:0;border-radius:8px;padding:9px 13px;background:#2563eb;color:white;cursor:pointer}button:disabled{opacity:.5}.toolbar{display:flex;gap:8px;align-items:center;margin-bottom:16px}.error{color:#b91c1c;white-space:pre-wrap}
</style>
</head>
<body>
<header><h1>AIOS V2 Control Plane</h1><span id="updated">Not loaded</span></header>
<main>
<div class="notice">Read-only candidate console. Identity and tenant context must be injected by the trusted reverse proxy. The console never calls OpenClaw or Hermes directly.</div>
<div class="toolbar"><button id="refresh">Refresh</button><span id="error" class="error"></span></div>
<section class="cards">
<div class="card"><small>Agents</small><b id="agentCount">—</b></div>
<div class="card"><small>Tasks</small><b id="taskCount">—</b></div>
<div class="card"><small>Audit events</small><b id="auditCount">—</b></div>
<div class="card"><small>Executions</small><b id="executionCount">—</b></div>
</section>
<section class="grid">
<div class="panel"><h2>Runtime health</h2><div id="runtime"></div></div>
<div class="panel"><h2>Tasks</h2><div id="tasks"></div></div>
<div class="panel"><h2>Agents</h2><div id="agents"></div></div>
<div class="panel"><h2>Execution evidence</h2><div id="executions"></div></div>
<div class="panel"><h2>Audit</h2><div id="audit"></div></div>
</section>
</main>
<script>
const esc=v=>String(v??'').replace(/[&<>"']/g,c=>({'&':'&amp;','<':'&lt;','>':'&gt;','"':'&quot;',"'":'&#39;'}[c]));
const table=(cols,rows)=>`<table><thead><tr>${cols.map(c=>`<th>${esc(c[0])}</th>`).join('')}</tr></thead><tbody>${rows.map(r=>`<tr>${cols.map(c=>`<td>${esc(c[1](r))}</td>`).join('')}</tr>`).join('')}</tbody></table>`;
async function api(path){const r=await fetch(path,{credentials:'same-origin',headers:{'Accept':'application/json'}});const data=await r.json().catch(()=>({}));if(!r.ok)throw new Error(`${r.status} ${data.message||data.error||path}`);return data}
async function refresh(){
const btn=document.getElementById('refresh');btn.disabled=true;document.getElementById('error').textContent='';
try{
const [agents,tasks,audit,runtime,executions]=await Promise.all([
api('/api/v2/agents'),api('/api/v2/tasks'),api('/api/v2/audit'),api('/api/v2/runtime'),api('/api/v2/executions')
]);
document.getElementById('agentCount').textContent=agents.agents.length;
document.getElementById('taskCount').textContent=tasks.tasks.length;
document.getElementById('auditCount').textContent=audit.events.length;
document.getElementById('executionCount').textContent=executions.executions.length;
document.getElementById('agents').innerHTML=table([['ID',x=>x.agent_id],['Runtime',x=>x.runtime],['Enabled',x=>x.enabled]],agents.agents);
document.getElementById('tasks').innerHTML=table([['Task',x=>x.task_id],['Action',x=>x.action],['State',x=>x.state],['Rev',x=>x.revision]],tasks.tasks);
document.getElementById('audit').innerHTML=table([['Time',x=>x.created_at],['Action',x=>x.action],['Outcome',x=>x.outcome],['Resource',x=>x.resource_id]],audit.events.slice(-100).reverse());
document.getElementById('executions').innerHTML=table([['Executor',x=>x.executor],['Task',x=>x.task_id],['Status',x=>x.status],['Progress',x=>`${x.progress_completed}/${x.progress_total}`],['Commit',x=>x.last_commit||'']],executions.executions);
document.getElementById('runtime').innerHTML=table([['Runtime',x=>x.name],['Status',x=>x.status],['Details',x=>JSON.stringify(x.details)]],runtime.runtimes);
document.getElementById('updated').textContent=`Updated ${new Date().toLocaleString()}`;
}catch(e){document.getElementById('error').textContent=e.message||String(e)}finally{btn.disabled=false}
}
document.getElementById('refresh').addEventListener('click',refresh);refresh();setInterval(refresh,30000);
</script>
</body>
</html>'''
+627
View File
@@ -0,0 +1,627 @@
"""Tenant-isolated AIOS V2 control-plane domain service.
This module deliberately contains no OpenClaw or Hermes implementation details.
It owns authoritative tenant, agent, task-lifecycle, idempotency and audit state.
Runtime adapters consume approved tasks through a separate boundary.
"""
from __future__ import annotations
import copy
import hashlib
import json
import threading
import uuid
from dataclasses import dataclass, field
from datetime import datetime, timezone
from enum import Enum
from typing import Any, Callable, Iterable, Mapping
def _utc_now() -> str:
return datetime.now(timezone.utc).isoformat().replace("+00:00", "Z")
def _new_id(prefix: str) -> str:
return f"{prefix}_{uuid.uuid4().hex}"
def _canonical_digest(value: Mapping[str, Any]) -> str:
encoded = json.dumps(
value,
ensure_ascii=False,
sort_keys=True,
separators=(",", ":"),
default=str,
).encode("utf-8")
return hashlib.sha256(encoded).hexdigest()
class ControlPlaneError(RuntimeError):
"""Base error returned by the control-plane domain."""
class Unauthorized(ControlPlaneError):
"""The caller is not authenticated or lacks the required role."""
class TenantIsolationError(ControlPlaneError):
"""A request attempted to cross a tenant boundary."""
class NotFound(ControlPlaneError):
"""A tenant-scoped resource does not exist or is not visible."""
class InvalidTransition(ControlPlaneError):
"""A task state transition violates the lifecycle contract."""
class IdempotencyConflict(ControlPlaneError):
"""An idempotency key was reused for a different command."""
class RevisionConflict(ControlPlaneError):
"""Optimistic task revision does not match current state."""
class PolicyDenied(ControlPlaneError):
"""The policy gate denied a control-plane operation."""
@dataclass(frozen=True)
class RequestContext:
"""Authenticated request identity propagated across all AIOS layers."""
tenant_id: str
principal_id: str
trace_id: str
roles: frozenset[str] = field(default_factory=frozenset)
session_id: str | None = None
policy_version: str = "v1"
def __post_init__(self) -> None:
for name in ("tenant_id", "principal_id", "trace_id"):
if not str(getattr(self, name, "")).strip():
raise ValueError(f"{name} is required")
@classmethod
def create(
cls,
*,
tenant_id: str,
principal_id: str,
roles: Iterable[str] = (),
trace_id: str | None = None,
session_id: str | None = None,
policy_version: str = "v1",
) -> "RequestContext":
return cls(
tenant_id=tenant_id.strip(),
principal_id=principal_id.strip(),
trace_id=trace_id or _new_id("trace"),
roles=frozenset(str(role).strip() for role in roles if str(role).strip()),
session_id=session_id,
policy_version=policy_version,
)
@dataclass(frozen=True)
class TenantRecord:
tenant_id: str
display_name: str
status: str = "active"
metadata: Mapping[str, Any] = field(default_factory=dict)
created_at: str = field(default_factory=_utc_now)
updated_at: str = field(default_factory=_utc_now)
@dataclass(frozen=True)
class AgentRecord:
agent_id: str
tenant_id: str
display_name: str
runtime: str
capabilities: tuple[str, ...] = field(default_factory=tuple)
enabled: bool = True
metadata: Mapping[str, Any] = field(default_factory=dict)
created_at: str = field(default_factory=_utc_now)
updated_at: str = field(default_factory=_utc_now)
class TaskState(str, Enum):
CREATED = "CREATED"
APPROVED = "APPROVED"
RUNNING = "RUNNING"
COMPLETED = "COMPLETED"
FAILED = "FAILED"
CANCELLED = "CANCELLED"
TERMINAL_TASK_STATES = frozenset(
{TaskState.COMPLETED, TaskState.FAILED, TaskState.CANCELLED}
)
ALLOWED_TASK_TRANSITIONS: Mapping[TaskState, frozenset[TaskState]] = {
TaskState.CREATED: frozenset({TaskState.APPROVED, TaskState.CANCELLED}),
TaskState.APPROVED: frozenset({TaskState.RUNNING, TaskState.CANCELLED}),
TaskState.RUNNING: frozenset(
{TaskState.COMPLETED, TaskState.FAILED, TaskState.CANCELLED}
),
TaskState.COMPLETED: frozenset(),
TaskState.FAILED: frozenset(),
TaskState.CANCELLED: frozenset(),
}
@dataclass(frozen=True)
class TaskRecord:
task_id: str
tenant_id: str
agent_id: str
action: str
payload: Mapping[str, Any]
state: TaskState = TaskState.CREATED
revision: int = 1
idempotency_key: str | None = None
request_digest: str = ""
result: Mapping[str, Any] | None = None
error: Mapping[str, Any] | None = None
created_by: str = ""
approved_by: str | None = None
created_at: str = field(default_factory=_utc_now)
updated_at: str = field(default_factory=_utc_now)
@dataclass(frozen=True)
class AuditEvent:
event_id: str
tenant_id: str
principal_id: str
trace_id: str
action: str
resource_type: str
resource_id: str
outcome: str
policy_version: str
details: Mapping[str, Any] = field(default_factory=dict)
created_at: str = field(default_factory=_utc_now)
@dataclass(frozen=True)
class PolicyDecision:
allowed: bool
reason: str
policy_version: str
PolicyEvaluator = Callable[[RequestContext, str, Mapping[str, Any]], PolicyDecision]
class InMemoryControlPlaneStore:
"""Thread-safe reference store used by tests and the first gateway slice.
Production backends must preserve the same tenant-scoped interface and add
database row-level security, migrations, backup and HA evidence.
"""
def __init__(self) -> None:
self.lock = threading.RLock()
self.tenants: dict[str, TenantRecord] = {}
self.agents: dict[str, AgentRecord] = {}
self.tasks: dict[str, TaskRecord] = {}
self.idempotency: dict[tuple[str, str], tuple[str, str]] = {}
self.audit_events: list[AuditEvent] = []
class ControlPlaneService:
"""Authoritative tenant, agent and task lifecycle service."""
def __init__(
self,
store: InMemoryControlPlaneStore | None = None,
*,
policy_evaluator: PolicyEvaluator | None = None,
) -> None:
self.store = store or InMemoryControlPlaneStore()
self.policy_evaluator = policy_evaluator or self._default_policy
@staticmethod
def _default_policy(
context: RequestContext,
action: str,
resource: Mapping[str, Any],
) -> PolicyDecision:
del resource
role_requirements = {
"tenant.register": {"platform_admin"},
"tenant.read": {"tenant_admin", "operator", "auditor", "platform_admin"},
"agent.register": {"tenant_admin", "operator", "platform_admin"},
"agent.read": {"tenant_admin", "operator", "auditor", "platform_admin"},
"task.create": {"tenant_admin", "operator", "platform_admin"},
"task.read": {"tenant_admin", "operator", "auditor", "platform_admin"},
"task.approve": {"tenant_admin", "approver", "platform_admin"},
"task.start": {"runtime", "operator", "platform_admin"},
"task.complete": {"runtime", "operator", "platform_admin"},
"task.fail": {"runtime", "operator", "platform_admin"},
"task.cancel": {"tenant_admin", "operator", "approver", "platform_admin"},
"audit.read": {"tenant_admin", "auditor", "platform_admin"},
}
required = role_requirements.get(action, {"platform_admin"})
allowed = bool(context.roles.intersection(required))
return PolicyDecision(
allowed=allowed,
reason="role_allowed" if allowed else "required_role_missing",
policy_version=context.policy_version,
)
def _require_context(self, context: RequestContext) -> None:
if not context.principal_id or not context.tenant_id or not context.trace_id:
raise Unauthorized("authenticated tenant context is required")
def _authorize(
self,
context: RequestContext,
action: str,
resource: Mapping[str, Any],
) -> PolicyDecision:
self._require_context(context)
decision = self.policy_evaluator(context, action, resource)
if not decision.allowed:
self._audit(
context,
action=action,
resource_type=str(resource.get("resource_type", "policy")),
resource_id=str(resource.get("resource_id", "unknown")),
outcome="denied",
details={"reason": decision.reason},
)
raise PolicyDenied(f"{action} denied: {decision.reason}")
return decision
@staticmethod
def _assert_same_tenant(context: RequestContext, tenant_id: str) -> None:
if context.tenant_id != tenant_id and "platform_admin" not in context.roles:
raise TenantIsolationError("cross-tenant access denied")
def _audit(
self,
context: RequestContext,
*,
action: str,
resource_type: str,
resource_id: str,
outcome: str,
details: Mapping[str, Any] | None = None,
) -> AuditEvent:
event = AuditEvent(
event_id=_new_id("audit"),
tenant_id=context.tenant_id,
principal_id=context.principal_id,
trace_id=context.trace_id,
action=action,
resource_type=resource_type,
resource_id=resource_id,
outcome=outcome,
policy_version=context.policy_version,
details=copy.deepcopy(dict(details or {})),
)
self.store.audit_events.append(event)
return event
def register_tenant(
self,
context: RequestContext,
*,
tenant_id: str,
display_name: str,
metadata: Mapping[str, Any] | None = None,
) -> TenantRecord:
tenant_id = tenant_id.strip()
if not tenant_id or not display_name.strip():
raise ValueError("tenant_id and display_name are required")
self._authorize(
context,
"tenant.register",
{"resource_type": "tenant", "resource_id": tenant_id},
)
self._assert_same_tenant(context, tenant_id)
with self.store.lock:
if tenant_id in self.store.tenants:
raise ControlPlaneError(f"tenant already exists: {tenant_id}")
tenant = TenantRecord(
tenant_id=tenant_id,
display_name=display_name.strip(),
metadata=copy.deepcopy(dict(metadata or {})),
)
self.store.tenants[tenant_id] = tenant
self._audit(
context,
action="tenant.register",
resource_type="tenant",
resource_id=tenant_id,
outcome="allowed",
)
return tenant
def get_tenant(self, context: RequestContext, tenant_id: str | None = None) -> TenantRecord:
tenant_id = tenant_id or context.tenant_id
self._authorize(
context,
"tenant.read",
{"resource_type": "tenant", "resource_id": tenant_id},
)
self._assert_same_tenant(context, tenant_id)
with self.store.lock:
tenant = self.store.tenants.get(tenant_id)
if tenant is None:
raise NotFound("tenant not found")
return tenant
def register_agent(
self,
context: RequestContext,
*,
agent_id: str,
display_name: str,
runtime: str,
capabilities: Iterable[str] = (),
metadata: Mapping[str, Any] | None = None,
) -> AgentRecord:
self._authorize(
context,
"agent.register",
{"resource_type": "agent", "resource_id": agent_id},
)
tenant = self.get_tenant(context)
if tenant.status != "active":
raise ControlPlaneError("tenant is not active")
normalized_capabilities = tuple(
sorted({str(item).strip() for item in capabilities if str(item).strip()})
)
if not agent_id.strip() or not display_name.strip() or not runtime.strip():
raise ValueError("agent_id, display_name and runtime are required")
with self.store.lock:
existing = self.store.agents.get(agent_id)
if existing is not None and existing.tenant_id != context.tenant_id:
raise TenantIsolationError("agent id belongs to another tenant")
if existing is not None:
raise ControlPlaneError(f"agent already exists: {agent_id}")
agent = AgentRecord(
agent_id=agent_id.strip(),
tenant_id=context.tenant_id,
display_name=display_name.strip(),
runtime=runtime.strip(),
capabilities=normalized_capabilities,
metadata=copy.deepcopy(dict(metadata or {})),
)
self.store.agents[agent.agent_id] = agent
self._audit(
context,
action="agent.register",
resource_type="agent",
resource_id=agent.agent_id,
outcome="allowed",
details={"runtime": agent.runtime},
)
return agent
def get_agent(self, context: RequestContext, agent_id: str) -> AgentRecord:
self._authorize(
context,
"agent.read",
{"resource_type": "agent", "resource_id": agent_id},
)
with self.store.lock:
agent = self.store.agents.get(agent_id)
if agent is None:
raise NotFound("agent not found")
self._assert_same_tenant(context, agent.tenant_id)
return agent
def list_agents(self, context: RequestContext) -> list[AgentRecord]:
self._authorize(
context,
"agent.read",
{"resource_type": "agent_collection", "resource_id": context.tenant_id},
)
with self.store.lock:
return [
agent
for agent in self.store.agents.values()
if agent.tenant_id == context.tenant_id
]
def create_task(
self,
context: RequestContext,
*,
agent_id: str,
action: str,
payload: Mapping[str, Any] | None = None,
idempotency_key: str | None = None,
) -> TaskRecord:
self._authorize(
context,
"task.create",
{"resource_type": "task", "resource_id": idempotency_key or "new"},
)
agent = self.get_agent(context, agent_id)
if not agent.enabled:
raise ControlPlaneError("agent is disabled")
if action not in agent.capabilities and "*" not in agent.capabilities:
raise PolicyDenied(f"agent does not advertise capability: {action}")
normalized_payload = copy.deepcopy(dict(payload or {}))
request_digest = _canonical_digest(
{"agent_id": agent_id, "action": action, "payload": normalized_payload}
)
normalized_key = idempotency_key.strip() if idempotency_key else None
with self.store.lock:
if normalized_key:
idempotency_ref = (context.tenant_id, normalized_key)
existing = self.store.idempotency.get(idempotency_ref)
if existing is not None:
existing_task_id, existing_digest = existing
if existing_digest != request_digest:
raise IdempotencyConflict(
"idempotency key was reused with a different request"
)
return self.store.tasks[existing_task_id]
task = TaskRecord(
task_id=_new_id("task"),
tenant_id=context.tenant_id,
agent_id=agent_id,
action=action,
payload=normalized_payload,
idempotency_key=normalized_key,
request_digest=request_digest,
created_by=context.principal_id,
)
self.store.tasks[task.task_id] = task
if normalized_key:
self.store.idempotency[(context.tenant_id, normalized_key)] = (
task.task_id,
request_digest,
)
self._audit(
context,
action="task.create",
resource_type="task",
resource_id=task.task_id,
outcome="allowed",
details={"agent_id": agent_id, "action": action},
)
return task
def get_task(self, context: RequestContext, task_id: str) -> TaskRecord:
self._authorize(
context,
"task.read",
{"resource_type": "task", "resource_id": task_id},
)
with self.store.lock:
task = self.store.tasks.get(task_id)
if task is None:
raise NotFound("task not found")
self._assert_same_tenant(context, task.tenant_id)
return task
def list_tasks(
self,
context: RequestContext,
*,
states: Iterable[TaskState | str] | None = None,
) -> list[TaskRecord]:
self._authorize(
context,
"task.read",
{"resource_type": "task_collection", "resource_id": context.tenant_id},
)
state_filter = (
{TaskState(item) for item in states}
if states is not None
else None
)
with self.store.lock:
return [
task
for task in self.store.tasks.values()
if task.tenant_id == context.tenant_id
and (state_filter is None or task.state in state_filter)
]
def transition_task(
self,
context: RequestContext,
*,
task_id: str,
target_state: TaskState | str,
expected_revision: int | None = None,
result: Mapping[str, Any] | None = None,
error: Mapping[str, Any] | None = None,
) -> TaskRecord:
target = TaskState(target_state)
action_name = {
TaskState.APPROVED: "task.approve",
TaskState.RUNNING: "task.start",
TaskState.COMPLETED: "task.complete",
TaskState.FAILED: "task.fail",
TaskState.CANCELLED: "task.cancel",
}.get(target)
if action_name is None:
raise InvalidTransition(f"cannot transition to {target.value}")
self._authorize(
context,
action_name,
{"resource_type": "task", "resource_id": task_id},
)
with self.store.lock:
current = self.store.tasks.get(task_id)
if current is None:
raise NotFound("task not found")
self._assert_same_tenant(context, current.tenant_id)
if expected_revision is not None and current.revision != expected_revision:
raise RevisionConflict(
f"expected revision {expected_revision}, current {current.revision}"
)
if target not in ALLOWED_TASK_TRANSITIONS[current.state]:
raise InvalidTransition(
f"invalid transition {current.state.value} -> {target.value}"
)
if target == TaskState.COMPLETED and result is None:
raise ValueError("completed task requires result")
if target == TaskState.FAILED and error is None:
raise ValueError("failed task requires error")
updated = TaskRecord(
task_id=current.task_id,
tenant_id=current.tenant_id,
agent_id=current.agent_id,
action=current.action,
payload=current.payload,
state=target,
revision=current.revision + 1,
idempotency_key=current.idempotency_key,
request_digest=current.request_digest,
result=copy.deepcopy(dict(result)) if result is not None else current.result,
error=copy.deepcopy(dict(error)) if error is not None else current.error,
created_by=current.created_by,
approved_by=(
context.principal_id
if target == TaskState.APPROVED
else current.approved_by
),
created_at=current.created_at,
updated_at=_utc_now(),
)
self.store.tasks[task_id] = updated
self._audit(
context,
action=action_name,
resource_type="task",
resource_id=task_id,
outcome="allowed",
details={
"from": current.state.value,
"to": target.value,
"revision": updated.revision,
},
)
return updated
def list_audit_events(
self,
context: RequestContext,
*,
resource_id: str | None = None,
) -> list[AuditEvent]:
self._authorize(
context,
"audit.read",
{"resource_type": "audit", "resource_id": resource_id or context.tenant_id},
)
with self.store.lock:
return [
event
for event in self.store.audit_events
if event.tenant_id == context.tenant_id
and (resource_id is None or event.resource_id == resource_id)
]
@@ -0,0 +1,195 @@
"""Git-backed agent execution evidence for AIOS V2.
Evidence is tenant-scoped, append-friendly, and contains no runtime credentials.
It provides a deterministic source for claim, heartbeat, progress, result and
verification state across Codex, OpenClaw and Hermes executors.
"""
from __future__ import annotations
import json
import os
import threading
from dataclasses import asdict, dataclass, field, replace
from datetime import datetime, timezone
from pathlib import Path
from typing import Any, Mapping
def utc_now() -> str:
return datetime.now(timezone.utc).isoformat().replace("+00:00", "Z")
def _required(value: str, label: str) -> str:
normalized = str(value or "").strip()
if not normalized:
raise ValueError(f"{label} is required")
return normalized
def _safe(value: Mapping[str, Any] | None) -> dict[str, Any]:
return json.loads(json.dumps(dict(value or {}), ensure_ascii=False, default=str))
@dataclass(frozen=True)
class ExecutionEvidence:
execution_id: str
tenant_id: str
task_id: str
agent_id: str
executor: str
branch: str
status: str = "claimed"
progress_completed: int = 0
progress_total: int = 1
current_step: str = "claim"
last_commit: str | None = None
test_status: str = "not_run"
test_summary: Mapping[str, Any] = field(default_factory=dict)
result_summary: Mapping[str, Any] = field(default_factory=dict)
claimed_at: str = field(default_factory=utc_now)
heartbeat_at: str = field(default_factory=utc_now)
completed_at: str | None = None
def __post_init__(self) -> None:
for name in (
"execution_id",
"tenant_id",
"task_id",
"agent_id",
"executor",
"branch",
"status",
"current_step",
"claimed_at",
"heartbeat_at",
):
object.__setattr__(self, name, _required(getattr(self, name), name))
if self.progress_completed < 0 or self.progress_total < 1:
raise ValueError("invalid progress values")
if self.progress_completed > self.progress_total:
raise ValueError("progress_completed cannot exceed progress_total")
object.__setattr__(self, "test_summary", _safe(self.test_summary))
object.__setattr__(self, "result_summary", _safe(self.result_summary))
class ExecutionTracker:
"""Persist one JSON evidence record per tenant and execution."""
def __init__(self, root: str | Path) -> None:
self.root = Path(root).resolve()
self.root.mkdir(parents=True, exist_ok=True)
self._lock = threading.RLock()
@staticmethod
def _component(value: str) -> str:
normalized = _required(value, "path component")
if normalized in {".", ".."} or "/" in normalized or "\\" in normalized:
raise ValueError("invalid evidence path component")
return normalized
def _path(self, tenant_id: str, execution_id: str) -> Path:
tenant = self._component(tenant_id)
execution = self._component(execution_id)
return self.root / tenant / f"{execution}.json"
def claim(self, evidence: ExecutionEvidence) -> ExecutionEvidence:
path = self._path(evidence.tenant_id, evidence.execution_id)
with self._lock:
if path.exists():
current = self._read(path)
if current.task_id != evidence.task_id or current.agent_id != evidence.agent_id:
raise RuntimeError("execution_id already belongs to another task")
return current
self._write(path, evidence)
return evidence
def heartbeat(
self,
*,
tenant_id: str,
execution_id: str,
status: str | None = None,
current_step: str | None = None,
progress_completed: int | None = None,
progress_total: int | None = None,
last_commit: str | None = None,
) -> ExecutionEvidence:
path = self._path(tenant_id, execution_id)
with self._lock:
current = self._read(path)
updated = replace(
current,
status=status or current.status,
current_step=current_step or current.current_step,
progress_completed=(
progress_completed
if progress_completed is not None
else current.progress_completed
),
progress_total=(
progress_total if progress_total is not None else current.progress_total
),
last_commit=last_commit or current.last_commit,
heartbeat_at=utc_now(),
)
self._write(path, updated)
return updated
def complete(
self,
*,
tenant_id: str,
execution_id: str,
status: str,
test_status: str,
test_summary: Mapping[str, Any] | None = None,
result_summary: Mapping[str, Any] | None = None,
last_commit: str | None = None,
) -> ExecutionEvidence:
path = self._path(tenant_id, execution_id)
with self._lock:
current = self._read(path)
updated = replace(
current,
status=_required(status, "status"),
current_step="complete",
progress_completed=current.progress_total,
last_commit=last_commit or current.last_commit,
test_status=_required(test_status, "test_status"),
test_summary=_safe(test_summary),
result_summary=_safe(result_summary),
heartbeat_at=utc_now(),
completed_at=utc_now(),
)
self._write(path, updated)
return updated
def get(self, *, tenant_id: str, execution_id: str) -> ExecutionEvidence:
return self._read(self._path(tenant_id, execution_id))
def list(self, *, tenant_id: str) -> list[ExecutionEvidence]:
tenant_dir = self.root / self._component(tenant_id)
if not tenant_dir.exists():
return []
records = [self._read(path) for path in tenant_dir.glob("*.json")]
return sorted(records, key=lambda item: (item.heartbeat_at, item.execution_id), reverse=True)
@staticmethod
def _read(path: Path) -> ExecutionEvidence:
try:
value = json.loads(path.read_text(encoding="utf-8"))
except FileNotFoundError as exc:
raise KeyError("execution evidence not found") from exc
if not isinstance(value, Mapping):
raise RuntimeError("invalid execution evidence file")
return ExecutionEvidence(**dict(value))
@staticmethod
def _write(path: Path, evidence: ExecutionEvidence) -> None:
path.parent.mkdir(parents=True, exist_ok=True)
temporary = path.with_suffix(f".tmp-{os.getpid()}")
temporary.write_text(
json.dumps(asdict(evidence), ensure_ascii=False, sort_keys=True, indent=2) + "\n",
encoding="utf-8",
)
os.replace(temporary, path)
+346
View File
@@ -0,0 +1,346 @@
"""HTTP/JSON gateway for the AIOS V2 control plane.
The gateway is intentionally runtime-neutral: browsers and external clients
call this API, while OpenClaw and Hermes remain behind authenticated adapters.
Identity headers are accepted only from a trusted reverse proxy in v1; direct
public exposure is prohibited until signed identity middleware is enabled.
"""
from __future__ import annotations
import dataclasses
import json
import os
import re
import urllib.parse
from enum import Enum
from http import HTTPStatus
from http.server import BaseHTTPRequestHandler, ThreadingHTTPServer
from typing import Any, Mapping
from .control_plane import (
ControlPlaneError,
ControlPlaneService,
IdempotencyConflict,
InvalidTransition,
NotFound,
PolicyDenied,
RequestContext,
RevisionConflict,
TaskState,
TenantIsolationError,
Unauthorized,
)
TENANT_HEADER = "X-AIOS-Tenant-ID"
PRINCIPAL_HEADER = "X-AIOS-Principal-ID"
ROLES_HEADER = "X-AIOS-Roles"
TRACE_HEADER = "X-AIOS-Trace-ID"
SESSION_HEADER = "X-AIOS-Session-ID"
POLICY_HEADER = "X-AIOS-Policy-Version"
AUTHENTICATED_HEADER = "X-AIOS-Authenticated"
_RESOURCE_ID_RE = re.compile(r"^[A-Za-z0-9][A-Za-z0-9._:-]{0,191}$")
def _json_value(value: Any) -> Any:
if dataclasses.is_dataclass(value):
return {
field.name: _json_value(getattr(value, field.name))
for field in dataclasses.fields(value)
}
if isinstance(value, Enum):
return value.value
if isinstance(value, Mapping):
return {str(key): _json_value(item) for key, item in value.items()}
if isinstance(value, (list, tuple, set, frozenset)):
return [_json_value(item) for item in value]
return value
def _require_mapping(value: Any, *, label: str = "JSON body") -> Mapping[str, Any]:
if not isinstance(value, Mapping):
raise ValueError(f"{label} must be an object")
return value
def _require_resource_id(value: Any, *, label: str) -> str:
normalized = str(value or "").strip()
if not _RESOURCE_ID_RE.fullmatch(normalized):
raise ValueError(f"invalid {label}")
return normalized
class GatewayApplication:
"""Pure request router that can be tested without a listening socket."""
def __init__(self, service: ControlPlaneService | None = None) -> None:
self.service = service or ControlPlaneService()
@staticmethod
def context_from_headers(headers: Mapping[str, str]) -> RequestContext:
authenticated = str(headers.get(AUTHENTICATED_HEADER, "")).strip().lower()
if authenticated not in {"1", "true", "yes"}:
raise Unauthorized("trusted authentication assertion is required")
roles = [
role.strip()
for role in str(headers.get(ROLES_HEADER, "")).split(",")
if role.strip()
]
return RequestContext.create(
tenant_id=str(headers.get(TENANT_HEADER, "")),
principal_id=str(headers.get(PRINCIPAL_HEADER, "")),
trace_id=str(headers.get(TRACE_HEADER, "")).strip() or None,
session_id=str(headers.get(SESSION_HEADER, "")).strip() or None,
policy_version=str(headers.get(POLICY_HEADER, "v1")).strip() or "v1",
roles=roles,
)
def handle(
self,
method: str,
raw_path: str,
headers: Mapping[str, str],
body: Any = None,
) -> tuple[int, Mapping[str, Any]]:
parsed = urllib.parse.urlparse(raw_path)
path = parsed.path.rstrip("/") or "/"
query = urllib.parse.parse_qs(parsed.query, keep_blank_values=False)
method = method.upper()
if method == "GET" and path in {"/healthz", "/api/v2/health"}:
return HTTPStatus.OK, {
"service": "aios-v2-control-plane",
"status": "ok",
"store": type(self.service.store).__name__,
}
context = self.context_from_headers(headers)
payload = _require_mapping(body or {})
if method == "POST" and path == "/api/v2/tenants":
tenant = self.service.register_tenant(
context,
tenant_id=_require_resource_id(payload.get("tenant_id"), label="tenant_id"),
display_name=str(payload.get("display_name", "")),
metadata=_require_mapping(payload.get("metadata", {}), label="metadata"),
)
return HTTPStatus.CREATED, {"tenant": _json_value(tenant)}
if method == "GET" and path == "/api/v2/tenants/current":
tenant = self.service.get_tenant(context)
return HTTPStatus.OK, {"tenant": _json_value(tenant)}
if method == "POST" and path == "/api/v2/agents":
agent = self.service.register_agent(
context,
agent_id=_require_resource_id(payload.get("agent_id"), label="agent_id"),
display_name=str(payload.get("display_name", "")),
runtime=str(payload.get("runtime", "")),
capabilities=payload.get("capabilities", ()),
metadata=_require_mapping(payload.get("metadata", {}), label="metadata"),
)
return HTTPStatus.CREATED, {"agent": _json_value(agent)}
if method == "GET" and path == "/api/v2/agents":
return HTTPStatus.OK, {
"agents": _json_value(self.service.list_agents(context))
}
agent_match = re.fullmatch(r"/api/v2/agents/([^/]+)", path)
if method == "GET" and agent_match:
agent_id = _require_resource_id(
urllib.parse.unquote(agent_match.group(1)), label="agent_id"
)
return HTTPStatus.OK, {
"agent": _json_value(self.service.get_agent(context, agent_id))
}
if method == "POST" and path == "/api/v2/tasks":
task = self.service.create_task(
context,
agent_id=_require_resource_id(payload.get("agent_id"), label="agent_id"),
action=str(payload.get("action", "")).strip(),
payload=_require_mapping(payload.get("payload", {}), label="payload"),
idempotency_key=(
str(payload["idempotency_key"]).strip()
if payload.get("idempotency_key") is not None
else None
),
)
return HTTPStatus.CREATED, {"task": _json_value(task)}
if method == "GET" and path == "/api/v2/tasks":
requested_states = query.get("state")
tasks = self.service.list_tasks(context, states=requested_states)
return HTTPStatus.OK, {"tasks": _json_value(tasks)}
task_match = re.fullmatch(r"/api/v2/tasks/([^/]+)", path)
if method == "GET" and task_match:
task_id = _require_resource_id(
urllib.parse.unquote(task_match.group(1)), label="task_id"
)
return HTTPStatus.OK, {
"task": _json_value(self.service.get_task(context, task_id))
}
transition_match = re.fullmatch(
r"/api/v2/tasks/([^/]+)/(approve|start|complete|fail|cancel)", path
)
if method == "POST" and transition_match:
task_id = _require_resource_id(
urllib.parse.unquote(transition_match.group(1)), label="task_id"
)
transition = transition_match.group(2)
target = {
"approve": TaskState.APPROVED,
"start": TaskState.RUNNING,
"complete": TaskState.COMPLETED,
"fail": TaskState.FAILED,
"cancel": TaskState.CANCELLED,
}[transition]
expected_revision = payload.get("expected_revision")
if expected_revision is not None:
expected_revision = int(expected_revision)
task = self.service.transition_task(
context,
task_id=task_id,
target_state=target,
expected_revision=expected_revision,
result=(
_require_mapping(payload.get("result"), label="result")
if payload.get("result") is not None
else None
),
error=(
_require_mapping(payload.get("error"), label="error")
if payload.get("error") is not None
else None
),
)
return HTTPStatus.OK, {"task": _json_value(task)}
if method == "GET" and path == "/api/v2/audit":
resource_id = query.get("resource_id", [None])[0]
events = self.service.list_audit_events(
context,
resource_id=(
_require_resource_id(resource_id, label="resource_id")
if resource_id
else None
),
)
return HTTPStatus.OK, {"events": _json_value(events)}
return HTTPStatus.NOT_FOUND, {
"error": "not_found",
"message": "route not found",
}
_ERROR_STATUS: tuple[tuple[type[BaseException], HTTPStatus], ...] = (
(Unauthorized, HTTPStatus.UNAUTHORIZED),
(PolicyDenied, HTTPStatus.FORBIDDEN),
(TenantIsolationError, HTTPStatus.NOT_FOUND),
(NotFound, HTTPStatus.NOT_FOUND),
(IdempotencyConflict, HTTPStatus.CONFLICT),
(RevisionConflict, HTTPStatus.CONFLICT),
(InvalidTransition, HTTPStatus.CONFLICT),
(ValueError, HTTPStatus.BAD_REQUEST),
(ControlPlaneError, HTTPStatus.BAD_REQUEST),
)
def error_response(exc: BaseException) -> tuple[int, Mapping[str, Any]]:
for error_type, status in _ERROR_STATUS:
if isinstance(exc, error_type):
return status, {
"error": error_type.__name__,
"message": str(exc),
}
return HTTPStatus.INTERNAL_SERVER_ERROR, {
"error": "InternalServerError",
"message": "internal server error",
}
class GatewayRequestHandler(BaseHTTPRequestHandler):
server_version = "AIOSV2Gateway/1.0"
max_body_bytes = 1_048_576
@property
def application(self) -> GatewayApplication:
return self.server.application # type: ignore[attr-defined]
def _read_body(self) -> Any:
length = int(self.headers.get("Content-Length", "0") or "0")
if length < 0 or length > self.max_body_bytes:
raise ValueError("request body is too large")
if length == 0:
return {}
content_type = self.headers.get("Content-Type", "").split(";", 1)[0].strip()
if content_type != "application/json":
raise ValueError("Content-Type must be application/json")
raw = self.rfile.read(length)
return json.loads(raw.decode("utf-8"))
def _send_json(self, status: int, payload: Mapping[str, Any]) -> None:
encoded = json.dumps(
_json_value(payload), ensure_ascii=False, separators=(",", ":")
).encode("utf-8")
self.send_response(int(status))
self.send_header("Content-Type", "application/json; charset=utf-8")
self.send_header("Content-Length", str(len(encoded)))
self.send_header("Cache-Control", "no-store")
self.send_header("X-Content-Type-Options", "nosniff")
self.end_headers()
self.wfile.write(encoded)
def _dispatch(self, method: str) -> None:
try:
body = self._read_body() if method in {"POST", "PUT", "PATCH"} else {}
status, payload = self.application.handle(
method,
self.path,
{key: value for key, value in self.headers.items()},
body,
)
except BaseException as exc: # handler boundary; sanitized below
status, payload = error_response(exc)
self._send_json(status, payload)
def do_GET(self) -> None: # noqa: N802
self._dispatch("GET")
def do_POST(self) -> None: # noqa: N802
self._dispatch("POST")
def log_message(self, format: str, *args: Any) -> None:
# Production deployment routes structured access logs through the
# reverse proxy; do not log identity headers or request bodies here.
super().log_message(format, *args)
class GatewayHTTPServer(ThreadingHTTPServer):
daemon_threads = True
allow_reuse_address = True
def __init__(
self,
server_address: tuple[str, int],
application: GatewayApplication,
) -> None:
self.application = application
super().__init__(server_address, GatewayRequestHandler)
def main() -> None:
host = os.getenv("AIOS_V2_GATEWAY_HOST", "127.0.0.1")
port = int(os.getenv("AIOS_V2_GATEWAY_PORT", "18920"))
server = GatewayHTTPServer((host, port), GatewayApplication())
print(f"AIOS V2 control-plane gateway listening on {host}:{port}")
server.serve_forever()
if __name__ == "__main__":
main()
+123
View File
@@ -0,0 +1,123 @@
"""Sanitized GitOps runtime status for the AIOS Web control plane."""
from __future__ import annotations
import os
import subprocess
from dataclasses import dataclass, asdict
from pathlib import Path
from typing import Any, Mapping
@dataclass(frozen=True)
class GitOpsStatus:
available: bool
service_status: str
workspace_mode: str
workspace_branch: str | None
workspace_ref: str | None
workspace_dirty: bool | None
candidate_branch: str
release_channel: str
release_version: str
sync_status: str
details: Mapping[str, Any]
def as_dict(self) -> dict[str, Any]:
return asdict(self)
def _run_git(worktree: Path, *args: str) -> tuple[int, str, str]:
process = subprocess.run(
["git", "-C", str(worktree), *args],
text=True,
capture_output=True,
timeout=5,
)
return process.returncode, process.stdout.strip(), process.stderr.strip()
def collect_gitops_status() -> GitOpsStatus:
"""Return only non-sensitive Git/workspace metadata.
Remote URLs, environment values, credentials, filesystem contents and
command stderr are deliberately excluded from the Web response.
"""
candidate_branch = os.getenv(
"AIOS_CANDIDATE_BRANCH",
"integration/aios-v2r1-cto-production-candidate-20260720",
)
release_channel = os.getenv("AIOS_RELEASE_CHANNEL", "candidate")
release_version = os.getenv("AIOS_RELEASE_VERSION", "V2R1C0-candidate")
sync_status = os.getenv("AIOS_GITOPS_SYNC_STATUS", "unknown")
worktree = Path(
os.getenv(
"AIOS_GITOPS_CONTROLLER_WORKTREE",
"/data/aios-gitops/worktrees/controller",
)
)
if not worktree.is_dir():
return GitOpsStatus(
available=False,
service_status="unknown",
workspace_mode="unavailable",
workspace_branch=None,
workspace_ref=None,
workspace_dirty=None,
candidate_branch=candidate_branch,
release_channel=release_channel,
release_version=release_version,
sync_status=sync_status,
details={"reason": "controller_worktree_unavailable"},
)
inside_code, inside, _ = _run_git(worktree, "rev-parse", "--is-inside-work-tree")
if inside_code != 0 or inside != "true":
return GitOpsStatus(
available=False,
service_status="unknown",
workspace_mode="invalid",
workspace_branch=None,
workspace_ref=None,
workspace_dirty=None,
candidate_branch=candidate_branch,
release_channel=release_channel,
release_version=release_version,
sync_status=sync_status,
details={"reason": "controller_worktree_invalid"},
)
branch_code, branch, _ = _run_git(worktree, "branch", "--show-current")
ref_code, ref_value, _ = _run_git(worktree, "rev-parse", "HEAD")
dirty_code, dirty_value, _ = _run_git(worktree, "status", "--porcelain=v1")
service_status = "unknown"
try:
process = subprocess.run(
["systemctl", "is-active", "aios-gitops-model-api"],
text=True,
capture_output=True,
timeout=3,
)
service_status = process.stdout.strip() or "inactive"
except (FileNotFoundError, subprocess.TimeoutExpired):
service_status = "unavailable"
branch_value = branch if branch_code == 0 and branch else None
return GitOpsStatus(
available=True,
service_status=service_status,
workspace_mode="branch" if branch_value else "detached",
workspace_branch=branch_value,
workspace_ref=ref_value if ref_code == 0 else None,
workspace_dirty=bool(dirty_value) if dirty_code == 0 else None,
candidate_branch=candidate_branch,
release_channel=release_channel,
release_version=release_version,
sync_status=sync_status,
details={
"controller_clean": dirty_code == 0 and not dirty_value,
"controller_detached": branch_code == 0 and not branch,
},
)
+264
View File
@@ -0,0 +1,264 @@
"""AIOS V2 execution orchestrator.
The orchestrator is the only component allowed to translate an approved AIOS
task into an OpenClaw command or Hermes job. OpenClaw and Hermes remain pure
execution runtimes and never own tenant policy or task lifecycle.
"""
from __future__ import annotations
from dataclasses import dataclass
from typing import Any, Mapping
from aios_hermes import HermesAdapter, HermesJob, HermesResult
from aios_openclaw import (
CommandState,
OpenClawAdapter,
OpenClawCommand,
OpenClawResult,
sign_command,
)
from .control_plane import (
AgentRecord,
ControlPlaneService,
RequestContext,
TaskRecord,
TaskState,
)
class OrchestrationError(RuntimeError):
"""AIOS failed to dispatch or reconcile a runtime execution."""
@dataclass(frozen=True)
class DispatchRecord:
tenant_id: str
task_id: str
agent_id: str
runtime: str
runtime_execution_id: str
state: str
trace_id: str
class ExecutionOrchestrator:
"""Dispatch approved AIOS tasks to runtime adapters and reconcile results."""
def __init__(
self,
control_plane: ControlPlaneService,
*,
openclaw: OpenClawAdapter | None = None,
hermes: HermesAdapter | None = None,
openclaw_signing_material_resolver=None,
openclaw_key_id: str = "default",
default_cell_id: str = "default",
default_profile_id: str = "default",
) -> None:
self.control_plane = control_plane
self.openclaw = openclaw
self.hermes = hermes
self.openclaw_signing_material_resolver = openclaw_signing_material_resolver
self.openclaw_key_id = openclaw_key_id
self.default_cell_id = default_cell_id
self.default_profile_id = default_profile_id
self._dispatches: dict[tuple[str, str], DispatchRecord] = {}
def dispatch(
self,
context: RequestContext,
*,
task_id: str,
timeout_seconds: float | None = None,
) -> DispatchRecord:
task = self.control_plane.get_task(context, task_id)
if task.state != TaskState.APPROVED:
raise OrchestrationError("only APPROVED tasks may be dispatched")
agent = self.control_plane.get_agent(context, task.agent_id)
existing = self._dispatches.get((task.tenant_id, task.task_id))
if existing is not None:
return existing
running = self.control_plane.transition_task(
context,
task_id=task.task_id,
target_state=TaskState.RUNNING,
expected_revision=task.revision,
)
runtime = agent.runtime.strip().lower()
try:
if runtime == "openclaw":
record = self._dispatch_openclaw(
context,
agent,
running,
timeout_seconds=timeout_seconds,
)
elif runtime == "hermes":
record = self._dispatch_hermes(context, agent, running)
else:
raise OrchestrationError(f"unsupported runtime: {agent.runtime}")
except Exception as exc:
current = self.control_plane.get_task(context, task.task_id)
if current.state == TaskState.RUNNING:
self.control_plane.transition_task(
context,
task_id=task.task_id,
target_state=TaskState.FAILED,
expected_revision=current.revision,
error={"code": "dispatch_failed", "message": str(exc)},
)
raise
self._dispatches[(task.tenant_id, task.task_id)] = record
return record
def _dispatch_openclaw(
self,
context: RequestContext,
agent: AgentRecord,
task: TaskRecord,
*,
timeout_seconds: float | None,
) -> DispatchRecord:
if self.openclaw is None or self.openclaw_signing_material_resolver is None:
raise OrchestrationError("OpenClaw runtime is not configured")
key_material = self.openclaw_signing_material_resolver(
task.tenant_id,
self.openclaw_key_id,
)
command = OpenClawCommand.create(
tenant_id=task.tenant_id,
task_id=task.task_id,
trace_id=context.trace_id,
action=task.action,
payload=task.payload,
idempotency_key=task.idempotency_key or task.task_id,
key_id=self.openclaw_key_id,
)
command = sign_command(command, key_material)
result = self.openclaw.execute(command, timeout_seconds=timeout_seconds)
self.reconcile_openclaw(context, result)
return DispatchRecord(
tenant_id=task.tenant_id,
task_id=task.task_id,
agent_id=agent.agent_id,
runtime="openclaw",
runtime_execution_id=command.command_id,
state=result.state.value,
trace_id=context.trace_id,
)
def _dispatch_hermes(
self,
context: RequestContext,
agent: AgentRecord,
task: TaskRecord,
) -> DispatchRecord:
if self.hermes is None:
raise OrchestrationError("Hermes runtime is not configured")
metadata = dict(agent.metadata)
job = HermesJob.create(
tenant_id=task.tenant_id,
cell_id=str(metadata.get("cell_id") or self.default_cell_id),
profile_id=str(metadata.get("profile_id") or self.default_profile_id),
task_id=task.task_id,
trace_id=context.trace_id,
action=task.action,
payload=task.payload,
idempotency_key=task.idempotency_key or task.task_id,
priority=int(metadata.get("priority", 100)),
)
queued = self.hermes.enqueue(job)
return DispatchRecord(
tenant_id=task.tenant_id,
task_id=task.task_id,
agent_id=agent.agent_id,
runtime="hermes",
runtime_execution_id=queued.job_id,
state=queued.state.value,
trace_id=context.trace_id,
)
def reconcile_openclaw(
self,
context: RequestContext,
result: OpenClawResult,
) -> TaskRecord:
task = self.control_plane.get_task(context, result.task_id)
if result.tenant_id != task.tenant_id:
raise OrchestrationError("OpenClaw result tenant mismatch")
if task.state != TaskState.RUNNING:
return task
if result.state == CommandState.COMPLETED:
return self.control_plane.transition_task(
context,
task_id=task.task_id,
target_state=TaskState.COMPLETED,
expected_revision=task.revision,
result=result.output or {},
)
if result.state in {
CommandState.FAILED,
CommandState.TIMED_OUT,
CommandState.REJECTED,
}:
return self.control_plane.transition_task(
context,
task_id=task.task_id,
target_state=TaskState.FAILED,
expected_revision=task.revision,
error=result.error or {"code": result.state.value.lower()},
)
if result.state == CommandState.CANCELLED:
return self.control_plane.transition_task(
context,
task_id=task.task_id,
target_state=TaskState.CANCELLED,
expected_revision=task.revision,
)
return task
def reconcile_hermes(
self,
context: RequestContext,
result: HermesResult,
) -> TaskRecord:
task = self.control_plane.get_task(context, result.task_id)
if result.tenant_id != task.tenant_id:
raise OrchestrationError("Hermes result tenant mismatch")
if task.state != TaskState.RUNNING:
return task
if result.state.value == "COMPLETED":
return self.control_plane.transition_task(
context,
task_id=task.task_id,
target_state=TaskState.COMPLETED,
expected_revision=task.revision,
result=result.output or {},
)
if result.state.value == "FAILED":
return self.control_plane.transition_task(
context,
task_id=task.task_id,
target_state=TaskState.FAILED,
expected_revision=task.revision,
error=result.error or {"code": "hermes_failed"},
)
if result.state.value == "CANCELLED":
return self.control_plane.transition_task(
context,
task_id=task.task_id,
target_state=TaskState.CANCELLED,
expected_revision=task.revision,
)
return task
def dispatch_snapshot(self, *, tenant_id: str) -> Mapping[str, DispatchRecord]:
return {
task_id: record
for (record_tenant, task_id), record in self._dispatches.items()
if record_tenant == tenant_id
}
+181
View File
@@ -0,0 +1,181 @@
"""Durable single-process persistence for the AIOS V2 control plane.
The candidate uses an atomic SQLite snapshot so a gateway restart does not lose
tenants, agents, task state, idempotency records, or audit events. The domain
service remains storage-agnostic; a later HA backend can replace this module
without changing API or runtime-adapter contracts.
"""
from __future__ import annotations
import json
import sqlite3
import threading
from dataclasses import asdict
from pathlib import Path
from typing import Any, Callable
from .control_plane import (
AgentRecord,
AuditEvent,
ControlPlaneService,
InMemoryControlPlaneStore,
TaskRecord,
TaskState,
TenantRecord,
)
class SQLiteSnapshotRepository:
"""Store one versioned control-plane snapshot in SQLite."""
SCHEMA_VERSION = 1
def __init__(self, path: str | Path) -> None:
self.path = Path(path).resolve()
self.path.parent.mkdir(parents=True, exist_ok=True)
self._lock = threading.RLock()
self._initialize()
def _connect(self) -> sqlite3.Connection:
connection = sqlite3.connect(self.path, timeout=30.0)
connection.execute("PRAGMA journal_mode=WAL")
connection.execute("PRAGMA synchronous=FULL")
connection.execute("PRAGMA foreign_keys=ON")
return connection
def _initialize(self) -> None:
with self._connect() as connection:
connection.execute(
"""
CREATE TABLE IF NOT EXISTS control_plane_snapshot (
singleton_id INTEGER PRIMARY KEY CHECK (singleton_id = 1),
schema_version INTEGER NOT NULL,
payload_json TEXT NOT NULL,
updated_at TEXT NOT NULL DEFAULT CURRENT_TIMESTAMP
)
"""
)
def load(self) -> dict[str, Any] | None:
with self._lock, self._connect() as connection:
row = connection.execute(
"SELECT schema_version, payload_json FROM control_plane_snapshot WHERE singleton_id = 1"
).fetchone()
if row is None:
return None
schema_version, payload_json = row
if int(schema_version) != self.SCHEMA_VERSION:
raise RuntimeError(
f"unsupported control-plane snapshot version: {schema_version}"
)
decoded = json.loads(str(payload_json))
if not isinstance(decoded, dict):
raise RuntimeError("invalid control-plane snapshot")
return decoded
def save(self, payload: dict[str, Any]) -> None:
encoded = json.dumps(
payload,
ensure_ascii=False,
sort_keys=True,
separators=(",", ":"),
default=str,
)
with self._lock, self._connect() as connection:
connection.execute("BEGIN IMMEDIATE")
connection.execute(
"""
INSERT INTO control_plane_snapshot (
singleton_id, schema_version, payload_json, updated_at
) VALUES (1, ?, ?, CURRENT_TIMESTAMP)
ON CONFLICT(singleton_id) DO UPDATE SET
schema_version = excluded.schema_version,
payload_json = excluded.payload_json,
updated_at = CURRENT_TIMESTAMP
""",
(self.SCHEMA_VERSION, encoded),
)
connection.commit()
class DurableControlPlaneService(ControlPlaneService):
"""Control-plane service that persists after every successful mutation."""
def __init__(
self,
repository: SQLiteSnapshotRepository,
*,
policy_evaluator=None,
) -> None:
self.repository = repository
store = InMemoryControlPlaneStore()
self._restore(store, repository.load())
super().__init__(store, policy_evaluator=policy_evaluator)
@staticmethod
def _restore(
store: InMemoryControlPlaneStore,
snapshot: dict[str, Any] | None,
) -> None:
if not snapshot:
return
for item in snapshot.get("tenants", []):
record = TenantRecord(**item)
store.tenants[record.tenant_id] = record
for item in snapshot.get("agents", []):
normalized = dict(item)
normalized["capabilities"] = tuple(normalized.get("capabilities", ()))
record = AgentRecord(**normalized)
store.agents[record.agent_id] = record
for item in snapshot.get("tasks", []):
normalized = dict(item)
normalized["state"] = TaskState(normalized["state"])
record = TaskRecord(**normalized)
store.tasks[record.task_id] = record
store.audit_events.extend(AuditEvent(**item) for item in snapshot.get("audit", []))
for item in snapshot.get("idempotency", []):
store.idempotency[(item["tenant_id"], item["key"])] = (
item["task_id"],
item["digest"],
)
def _snapshot(self) -> dict[str, Any]:
with self.store.lock:
return {
"schema_version": self.repository.SCHEMA_VERSION,
"tenants": [asdict(item) for item in self.store.tenants.values()],
"agents": [asdict(item) for item in self.store.agents.values()],
"tasks": [
{**asdict(item), "state": item.state.value}
for item in self.store.tasks.values()
],
"audit": [asdict(item) for item in self.store.audit_events],
"idempotency": [
{
"tenant_id": tenant_id,
"key": key,
"task_id": task_id,
"digest": digest,
}
for (tenant_id, key), (task_id, digest)
in self.store.idempotency.items()
],
}
def _mutate(self, operation: Callable[[], Any]) -> Any:
with self.store.lock:
result = operation()
self.repository.save(self._snapshot())
return result
def register_tenant(self, *args, **kwargs):
return self._mutate(lambda: super(DurableControlPlaneService, self).register_tenant(*args, **kwargs))
def register_agent(self, *args, **kwargs):
return self._mutate(lambda: super(DurableControlPlaneService, self).register_agent(*args, **kwargs))
def create_task(self, *args, **kwargs):
return self._mutate(lambda: super(DurableControlPlaneService, self).create_task(*args, **kwargs))
def transition_task(self, *args, **kwargs):
return self._mutate(lambda: super(DurableControlPlaneService, self).transition_task(*args, **kwargs))
+23
View File
@@ -0,0 +1,23 @@
FROM python:3.12-slim
ENV PYTHONDONTWRITEBYTECODE=1 \
PYTHONUNBUFFERED=1 \
PYTHONPATH=/opt/aios/app \
AIOS_V2_HOST=0.0.0.0 \
AIOS_V2_PORT=18920 \
AIOS_V2_STATE_DIR=/var/lib/aios-v2
RUN groupadd --system --gid 12001 aios \
&& useradd --system --uid 12001 --gid aios --home-dir /nonexistent --shell /usr/sbin/nologin aios \
&& install -d -o aios -g aios -m 0750 /opt/aios/app /var/lib/aios-v2
WORKDIR /opt/aios/app
COPY --chown=aios:aios active-runtime/app/ /opt/aios/app/
USER aios:aios
EXPOSE 18920
HEALTHCHECK --interval=15s --timeout=5s --start-period=10s --retries=4 \
CMD python3 -c "import urllib.request; urllib.request.urlopen('http://127.0.0.1:18920/api/v2/health', timeout=3).read()"
CMD ["python3", "-m", "aios_v2.app"]
+44
View File
@@ -0,0 +1,44 @@
services:
aios-v2:
build:
context: ../../..
dockerfile: deploy/aios-v2/docker/Dockerfile
image: aios-v2:V2R1C0-candidate
init: true
restart: unless-stopped
ports:
- "127.0.0.1:${AIOS_V2_PORT:-18920}:18920"
environment:
AIOS_V2_HOST: 0.0.0.0
AIOS_V2_PORT: 18920
AIOS_V2_STATE_DIR: /var/lib/aios-v2
AIOS_OPENCLAW_URL: ${AIOS_OPENCLAW_URL:-}
AIOS_OPENCLAW_CREDENTIAL_REF: ${AIOS_OPENCLAW_CREDENTIAL_REF:-}
AIOS_OPENCLAW_SIGNING_MATERIAL_FILE: ${AIOS_OPENCLAW_SIGNING_MATERIAL_FILE:-}
AIOS_RELEASE_CHANNEL: candidate
AIOS_RELEASE_VERSION: V2R1C0-candidate
volumes:
- aios-v2-state:/var/lib/aios-v2
# Mount runtime signing material explicitly when OpenClaw is enabled.
# The host file must be mode 0600; never store it in this repository.
# - /etc/aios-v2/openclaw-signing-material:/run/secrets/openclaw-signing-material:ro
read_only: true
tmpfs:
- /tmp:rw,noexec,nosuid,nodev,size=64m
cap_drop:
- ALL
security_opt:
- no-new-privileges:true
healthcheck:
test:
- CMD
- python3
- -c
- "import urllib.request; urllib.request.urlopen('http://127.0.0.1:18920/api/v2/health', timeout=3).read()"
interval: 15s
timeout: 5s
retries: 4
start_period: 10s
volumes:
aios-v2-state:
+21
View File
@@ -0,0 +1,21 @@
# AIOS V2 candidate runtime configuration.
# Copy to /etc/aios-v2/aios-v2.env and keep ownership root:aios, mode 0640.
AIOS_V2_HOST=127.0.0.1
AIOS_V2_PORT=18920
AIOS_V2_STATE_DIR=/var/lib/aios-v2
# Optional OpenClaw runtime. Leave blank to run with OpenClaw disabled.
AIOS_OPENCLAW_URL=
AIOS_OPENCLAW_CREDENTIAL_REF=
# File path only. The referenced file must be mode 0600 or stricter and contain
# at least 32 bytes. Never place signing material in this env file or Git.
AIOS_OPENCLAW_SIGNING_MATERIAL_FILE=
# Reverse proxy must authenticate users and inject X-AIOS-* identity headers.
AIOS_TRUST_PROXY_HEADERS=true
# Candidate deployment metadata exposed through health tooling.
AIOS_RELEASE_CHANNEL=candidate
AIOS_RELEASE_VERSION=V2R1C0-candidate
+86
View File
@@ -0,0 +1,86 @@
#!/usr/bin/env bash
set -Eeuo pipefail
BASE_URL="${AIOS_V2_BASE_URL:-http://127.0.0.1:18920}"
STATE_DIR="${AIOS_V2_STATE_DIR:-/var/lib/aios-v2}"
CONTROLLER_WT="${AIOS_GITOPS_CONTROLLER_WORKTREE:-/data/aios-gitops/worktrees/controller}"
SERVICE="${AIOS_V2_SERVICE:-aios-v2}"
fail() {
echo "FAIL: $*" >&2
exit 1
}
pass() {
echo "PASS: $*"
}
command -v python3 >/dev/null || fail "python3 not installed"
if command -v systemctl >/dev/null && systemctl list-unit-files "$SERVICE.service" >/dev/null 2>&1; then
systemctl is-active --quiet "$SERVICE" || fail "$SERVICE is not active"
pass "$SERVICE is active"
fi
python3 - "$BASE_URL" <<'PY'
import json
import sys
import urllib.request
base_url = sys.argv[1].rstrip("/")
with urllib.request.urlopen(f"{base_url}/api/v2/health", timeout=5) as response:
payload = json.loads(response.read().decode("utf-8"))
if payload.get("status") not in {"ok", "healthy", "ready"}:
raise SystemExit(f"unexpected health payload: {payload}")
print(f"PASS: HTTP health status={payload.get('status')}")
PY
[[ -d "$STATE_DIR" ]] || fail "state directory missing: $STATE_DIR"
[[ -w "$STATE_DIR" ]] || fail "state directory not writable: $STATE_DIR"
pass "state directory writable"
if [[ -e "$STATE_DIR/control-plane.sqlite3" ]]; then
python3 - "$STATE_DIR/control-plane.sqlite3" <<'PY'
import sqlite3
import sys
path = sys.argv[1]
connection = sqlite3.connect(f"file:{path}?mode=ro", uri=True, timeout=5)
try:
row = connection.execute("PRAGMA quick_check").fetchone()
finally:
connection.close()
if not row or row[0] != "ok":
raise SystemExit(f"SQLite quick_check failed: {row}")
print("PASS: SQLite quick_check=ok")
PY
else
echo "WARN: control-plane.sqlite3 not created yet"
fi
if [[ -d "$CONTROLLER_WT" ]] && git -C "$CONTROLLER_WT" rev-parse --is-inside-work-tree >/dev/null 2>&1; then
branch="$(git -C "$CONTROLLER_WT" branch --show-current)"
dirty="$(git -C "$CONTROLLER_WT" status --porcelain=v1)"
[[ -z "$branch" ]] || fail "GitOps controller attached to branch: $branch"
[[ -z "$dirty" ]] || fail "GitOps controller workspace is dirty"
pass "GitOps controller detached and clean"
fi
python3 - "$STATE_DIR/executions" <<'PY'
import json
import pathlib
import sys
root = pathlib.Path(sys.argv[1])
if not root.exists():
print("WARN: execution evidence directory not created yet")
raise SystemExit(0)
for path in root.glob("*/*.json"):
value = json.loads(path.read_text(encoding="utf-8"))
for key in ("execution_id", "tenant_id", "task_id", "executor", "status"):
if not value.get(key):
raise SystemExit(f"invalid execution evidence {path}: missing {key}")
print("PASS: execution evidence files valid")
PY
pass "AIOS V2 candidate healthcheck complete"
+86
View File
@@ -0,0 +1,86 @@
#!/usr/bin/env bash
set -Eeuo pipefail
SERVICE="${AIOS_V2_SERVICE:-aios-v2}"
RELEASE_ROOT="${AIOS_V2_RELEASE_ROOT:-/opt/aios/releases}"
CURRENT_LINK="${AIOS_V2_CURRENT_LINK:-/opt/aios/current}"
STATE_DIR="${AIOS_V2_STATE_DIR:-/var/lib/aios-v2}"
BACKUP_ROOT="${AIOS_V2_BACKUP_ROOT:-/var/backups/aios-v2}"
TARGET_RELEASE="${1:-}"
RESTORE_STATE="${AIOS_V2_RESTORE_STATE:-0}"
fail() {
echo "ERROR: $*" >&2
exit 1
}
[[ "${EUID}" -eq 0 ]] || fail "rollback must run as root"
[[ -n "$TARGET_RELEASE" ]] || fail "usage: $0 <release-name-or-absolute-path>"
if [[ "$TARGET_RELEASE" = /* ]]; then
TARGET_PATH="$(readlink -f "$TARGET_RELEASE")"
else
[[ "$TARGET_RELEASE" != *"/"* && "$TARGET_RELEASE" != "." && "$TARGET_RELEASE" != ".." ]] \
|| fail "invalid release name"
TARGET_PATH="$(readlink -f "$RELEASE_ROOT/$TARGET_RELEASE")"
fi
[[ -d "$TARGET_PATH/active-runtime/app" ]] || fail "target release is invalid: $TARGET_PATH"
[[ -f "$TARGET_PATH/deploy/aios-v2/healthcheck/check.sh" ]] \
|| fail "target release has no healthcheck"
PREVIOUS_PATH=""
if [[ -L "$CURRENT_LINK" ]]; then
PREVIOUS_PATH="$(readlink -f "$CURRENT_LINK")"
fi
STAMP="$(date -u +%Y%m%dT%H%M%SZ)"
BACKUP_DIR="$BACKUP_ROOT/rollback-$STAMP"
install -d -m 0750 "$BACKUP_DIR"
if [[ -d "$STATE_DIR" ]]; then
tar --xattrs --acls -C "$(dirname "$STATE_DIR")" \
-czf "$BACKUP_DIR/state-before-rollback.tar.gz" \
"$(basename "$STATE_DIR")"
fi
printf '%s\n' "$PREVIOUS_PATH" > "$BACKUP_DIR/previous-release.txt"
printf '%s\n' "$TARGET_PATH" > "$BACKUP_DIR/target-release.txt"
systemctl stop "$SERVICE"
ln -sfn "$TARGET_PATH" "${CURRENT_LINK}.next"
mv -Tf "${CURRENT_LINK}.next" "$CURRENT_LINK"
systemctl daemon-reload
systemctl start "$SERVICE"
if AIOS_V2_SERVICE="$SERVICE" \
AIOS_V2_STATE_DIR="$STATE_DIR" \
"$TARGET_PATH/deploy/aios-v2/healthcheck/check.sh"; then
echo "ROLLBACK_COMPLETE target=$TARGET_PATH backup=$BACKUP_DIR"
exit 0
fi
echo "Target release failed healthcheck; restoring previous release." >&2
systemctl stop "$SERVICE" || true
if [[ -n "$PREVIOUS_PATH" && -d "$PREVIOUS_PATH" ]]; then
ln -sfn "$PREVIOUS_PATH" "${CURRENT_LINK}.previous"
mv -Tf "${CURRENT_LINK}.previous" "$CURRENT_LINK"
fi
if [[ "$RESTORE_STATE" == "1" && -f "$BACKUP_DIR/state-before-rollback.tar.gz" ]]; then
rm -rf "$STATE_DIR"
tar --xattrs --acls -C "$(dirname "$STATE_DIR")" \
-xzf "$BACKUP_DIR/state-before-rollback.tar.gz"
fi
systemctl daemon-reload
systemctl start "$SERVICE" || true
if [[ -n "$PREVIOUS_PATH" && -x "$PREVIOUS_PATH/deploy/aios-v2/healthcheck/check.sh" ]]; then
AIOS_V2_SERVICE="$SERVICE" \
AIOS_V2_STATE_DIR="$STATE_DIR" \
"$PREVIOUS_PATH/deploy/aios-v2/healthcheck/check.sh" || true
fi
fail "rollback target failed; previous release restored; evidence=$BACKUP_DIR"
+37
View File
@@ -0,0 +1,37 @@
[Unit]
Description=AIOS V2 Control Plane Candidate
After=network-online.target
Wants=network-online.target
[Service]
Type=simple
User=aios
Group=aios
WorkingDirectory=/opt/aios/current/active-runtime/app
EnvironmentFile=-/etc/aios-v2/aios-v2.env
Environment=PYTHONUNBUFFERED=1
ExecStart=/usr/bin/python3 -m aios_v2.app
Restart=on-failure
RestartSec=3
TimeoutStartSec=30
TimeoutStopSec=30
KillSignal=SIGTERM
NoNewPrivileges=true
PrivateTmp=true
ProtectSystem=strict
ProtectHome=true
ProtectKernelTunables=true
ProtectKernelModules=true
ProtectControlGroups=true
RestrictSUIDSGID=true
RestrictRealtime=true
LockPersonality=true
MemoryDenyWriteExecute=true
ReadWritePaths=/var/lib/aios-v2
StateDirectory=aios-v2
StateDirectoryMode=0750
UMask=0027
LimitNOFILE=65536
[Install]
WantedBy=multi-user.target
+243
View File
@@ -0,0 +1,243 @@
# AIOS V2 用户需求规格
- 文档版本:`V2R1C0-candidate`
- 状态:候选版本需求基线
- 适用范围:AIOS Control Plane、Agent Router、OpenClaw Adapter、Hermes Runtime、Web Console、GitOps、审计与部署
## 1. 背景与目标
现有系统已经具备 OpenClaw、Hermes、Agent Router、GitOps、局部 WebUI、审计和多租户候选能力。AIOS V2 的目标不是推倒重写,而是将这些分散能力收敛到统一、可审计、可 Web 化、可多租户隔离的控制面。
AIOS V2 必须成为唯一控制面,负责租户、身份、Agent、策略、任务生命周期、审批、审计、Runtime 目录和执行证据。OpenClaw 与 Hermes 只承担执行职责,不得持有第二套控制面状态。
## 2. 用户角色
| 角色 | 主要诉求 |
|---|---|
| 平台管理员 | 管理租户、运行时、发布、GitOps 状态和全局策略 |
| 租户管理员 | 管理本租户 Agent、任务、审批、审计和成员权限 |
| 操作员 | 创建任务、发起执行、取消任务、查看运行状态 |
| 审批人 | 审批高风险任务,查看策略判定依据 |
| 审计员 | 查询本租户完整审计链和执行证据 |
| Runtime Worker | 领取已批准任务、心跳、返回执行结果 |
| 开发与运维人员 | 通过 GitOps 提交、测试、部署、回滚和修复 Bug |
## 3. 核心用户需求
### UR-001 唯一控制面
所有任务创建、审批、取消、状态变化、策略判定和审计必须由 AIOS Control Plane 统一处理。
验收:
- Web Console、API、Codex、OpenClaw 和 Hermes 均不得绕过 AIOS 创建或修改任务生命周期。
- OpenClaw/Hermes 的公开接口仅限执行、状态、取消、心跳和结果回传。
### UR-002 复用现有能力
系统不得重写已经稳定的 OpenClaw 执行能力、Hermes Runtime、Agent Router 或 GitOps 基础设施。
验收:
- 通过 Adapter/Boundary 接入现有能力。
- 替换控制逻辑时保留原执行接口和可迁移数据。
### UR-003 OpenClaw 与 Hermes 解耦
OpenClaw 与 Hermes 必须从策略、租户、审批和任务生命周期中逐步解耦。
验收:
- OpenClaw 仅接收已签名、已批准、带租户上下文的命令。
- Hermes 仅接收已批准的租户分区 Job。
- 二者均不得直接写 Control Plane 权威状态。
### UR-004 全面 Web 化
所有面向用户和运维的能力必须通过 Web Console 或稳定 API 访问。
Web Console 至少展示:
- Tenant、Agent、Task、Approval、Runtime、Audit、Execution Evidence、GitOps Status。
- 所有写操作由 AIOS API 完成,浏览器不得直接调用 OpenClaw/Hermes。
### UR-005 多租户隔离
所有模块必须支持租户隔离,且默认拒绝跨租户访问。
隔离范围:
- API 与会话
- Tenant/Agent/Task 数据
- OpenClaw 命令
- Hermes Queue、Cell、Profile、Worker Lease
- Memory、Cache、Artifact、SecretRef
- Audit、Execution Evidence、日志和指标
验收:
- 资源记录的 `tenant_id` 不可为空。
- 跨租户访问返回拒绝或“资源不存在”,不得泄露存在性。
- 必须有跨租户负向测试。
### UR-006 权威任务生命周期
任务必须遵循统一状态机:
```text
CREATED -> APPROVED -> RUNNING -> COMPLETED
| | |
+----------+----------+-> CANCELLED
+-> FAILED
```
验收:
- 禁止越级变化,例如 `CREATED -> COMPLETED`。
- 每次变化包含 revision、principal、trace、时间和审计事件。
### UR-007 策略与审批
所有有副作用任务必须经过策略判定;高风险任务必须支持人工审批。
验收:
- Policy Decision 包含 allow/deny、reason 和 policy version。
- 拒绝事件进入审计。
- Runtime 不得自行放宽策略。
### UR-008 幂等与并发安全
任务和 Runtime 执行必须支持幂等与乐观并发控制。
验收:
- 相同租户与幂等键、相同内容返回同一任务/执行。
- 相同幂等键不同内容返回冲突。
- revision 或 branch lease 冲突不得静默覆盖。
### UR-009 可观测与执行证据
平台必须能够确认 Codex、OpenClaw、Hermes 等执行器是否领取任务、当前进度、最近提交、测试结果和完成状态。
验收:
- 存在 claim、heartbeat、progress、result 和 evidence 记录。
- Web/API 可按租户查询。
- 执行证据不得包含明文密钥。
### UR-010 GitOps 唯一变更流程
所有代码、配置、文档和部署资产修改必须进入非 `main` 分支,通过 commit、测试、审查和受控提升。
验收:
- Controller worktree 必须 detached 且 clean。
- 本地 bare repo 与 Gitea 目标分支 OID 一致。
- 不允许直接 push `main`、直接创建生产 Tag 或绕过审批部署。
### UR-011 持久化与恢复
租户、Agent、Task、Audit、Idempotency 和 Execution Evidence 必须在服务重启后恢复。
验收:
- 候选版支持 SQLite WAL 原子快照。
- 有 restart recovery 测试。
- 生产版需进一步提供迁移、备份、恢复、HA 和容量证据。
### UR-012 审计与合规
所有控制与执行动作必须生成不可混租户的审计事件。
至少记录:
- tenant、principal、trace、action、resource、outcome、policy version、timestamp。
- OpenClaw/Hermes Runtime 生命周期。
- GitOps、部署和回滚动作。
### UR-013 密钥安全
Git、日志、Web、任务 payload 和审计记录中不得出现明文 API Key、密码、Token 或私钥。
验收:
- Git 只存 `secret_ref`、`credential_ref`、`key_id` 或受保护文件路径。
- Secret 文件权限为 `0600` 或更严格。
- Secret Scanner 默认拒绝可疑提交。
### UR-014 部署与回滚
系统必须提供 systemd、Docker/Compose、健康检查和受保护回滚入口。
验收:
- 服务以非 root 权限运行。
- 状态目录独立持久化。
- 回滚前自动备份状态。
- 默认只回滚代码,状态恢复需要显式授权。
### UR-015 文档与代码同步
任何影响用户行为、架构边界、API、数据模型、状态机、安全策略、部署或运维方式的变更,必须在同一分支同步更新 `docs/aios-v2/`。
验收:
- 架构/功能代码变化但无文档变化时,文档同步门禁失败。
- Traceability Matrix 必须指向当前代码和测试。
- 新架构决策必须增加或更新 ADR。
## 4. 非功能需求
### NFR-001 安全
- 默认拒绝、最小权限、租户隔离、身份可信链、SecretRef、审计。
### NFR-002 可靠性
- 幂等、revision、worker lease、heartbeat、重试边界、原子持久化、健康检查和回滚。
### NFR-003 可维护性
- 模块边界稳定,Control Plane 与 Runtime 解耦;代码、文档、测试和部署可追踪。
### NFR-004 可扩展性
- Store、Transport、Policy、EventPublisher、Runtime Adapter 均通过接口扩展。
### NFR-005 可运维性
- Web/API 状态、systemd/Docker、健康检查、日志、GitOps 状态、候选报告。
### NFR-006 性能候选指标
候选版上线前至少验证:
- API P95、任务调度延迟、队列等待、Worker 心跳、SQLite 锁等待、审计写入时间。
- 具体 SLO 在预发布基准测试后固化,禁止无证据承诺。
## 5. 本候选版本范围
### 已纳入代码
- Control Plane、持久化、Gateway、Orchestrator、OpenClaw Adapter、Hermes Boundary、Execution Tracking、Web Console、GitOps Status、部署和回滚。
### 尚需环境验证
- 精确测试执行证据。
- OpenClaw 真实 Runtime 联调。
- 生产数据库/队列替换或容量验证。
- 预发布性能、安全、重启和回滚演练。
- Owner 对 merge、Tag 和生产部署的批准。
## 6. 用户验收场景
1. 平台管理员创建租户并注册 OpenClaw/Hermes Agent。
2. 操作员创建任务,审批人批准。
3. AIOS Orchestrator 选择 Runtime,并创建签名命令或租户分区 Job。
4. Runtime 执行并返回结果。
5. AIOS 更新 Task、Audit 和 Execution Evidence。
6. 租户用户在 Web Console 查看完整链路。
7. 跨租户访问被拒绝。
8. 服务重启后状态恢复。
9. 候选部署健康检查通过,回滚演练成功。
+461
View File
@@ -0,0 +1,461 @@
# AIOS V2 目标架构
- 文档版本:`V2R1C0-candidate`
- 架构状态:目标架构已实现为候选代码,待预发布验证
- 适用范围:Control Plane、Agent Router、OpenClaw、Hermes、Web Console、GitOps、审计、部署与运维
## 1. 架构目标
AIOS V2 的核心目标是把分散在 OpenClaw、Hermes、Agent Router、局部 WebUI 和 GitOps 中的控制能力收敛为一个权威控制面,同时保留现有执行能力,避免推倒重写。
目标状态:
```text
用户 / Web Console / Public API
|
v
AIOS Control Plane
Tenant | Identity | Agent | Task | Policy | Approval | Audit
|
v
Execution Orchestrator / Agent Router
/ \
v v
OpenClaw Adapter Hermes Boundary
| |
v v
OpenClaw Runtime Hermes Workers
\ /
v v
Result Reconciliation + Execution Evidence
```
设计结论:
- AIOS 是唯一控制面。
- Agent Router 是调度平面,不是第二控制面。
- OpenClaw 是执行适配与 Runtime。
- Hermes 是队列、Worker、Profile 与执行 Runtime。
- Web Console 只调用 AIOS API。
- GitOps 是代码、配置、文档和发布变更的唯一入口。
## 2. 从现有架构到目标架构
### 2.1 现有能力
现有系统已经具备:
- OpenClaw 执行与 Fleet 能力。
- Hermes Worker、Profile、队列和工作区能力。
- Agent Router 路由经验和部分调度接口。
- GitOps 分支、提交、审批、回滚和审计能力。
- Tenant Context、Tenant Router、Memory Scope、SecretRef、Message Ledger 和审计候选实现。
- 局部 WebUI、Gateway 和自动任务领取能力。
### 2.2 现有问题
- 控制逻辑分散,可能存在多个任务事实源。
- OpenClaw/Hermes 内部仍可能保存租户、策略或任务状态。
- 候选能力分散在多个分支,主线缺少权威实现。
- WebUI 与 API 入口不统一。
- 分支名称、代码、测试和部署证据之间缺少映射。
- 多租户隔离多为逻辑约束,尚缺全链路生产证据。
### 2.3 演进原则
- 先加边界和适配器,再迁移控制权。
- 先复用接口,再替换内部实现。
- 先在候选分支验证,再受控提升。
- 不允许同时存在两套权威任务生命周期。
- 不以“大分支存在”代替验收证据。
## 3. 逻辑分层
### 3.1 Web 与 API 层
职责:
- 提供 Tenant、Agent、Task、Approval、Runtime、Audit、Execution Evidence 和 GitOps 页面/API。
- 依赖可信身份代理注入 `X-AIOS-*` 身份头。
- 不允许浏览器直接访问 OpenClaw 或 Hermes。
- 不允许客户端自行选择或覆盖 `tenant_id`。
候选实现:
- `active-runtime/app/aios_v2/app.py`
- `active-runtime/app/aios_v2/gateway.py`
- `active-runtime/app/aios_v2/console.py`
### 3.2 AIOS Control Plane
权威职责:
- Tenant Registry
- Agent Registry
- Task Lifecycle
- Policy Decision
- Approval
- Audit Event
- Runtime Inventory
- Execution Evidence
- Deployment/GitOps Status
不负责:
- OpenClaw 内部执行细节。
- Hermes Worker 内部执行细节。
- 直接持有 Runtime 凭据明文。
候选实现:
- `active-runtime/app/aios_v2/control_plane.py`
- `active-runtime/app/aios_v2/persistence.py`
- `active-runtime/app/aios_v2/execution_tracking.py`
- `active-runtime/app/aios_v2/gitops_status.py`
### 3.3 Execution Orchestrator / Agent Router
职责:
- 只接受 `APPROVED` 任务。
- 根据 Agent Registry 和 Policy Result 选择 Runtime。
- 生成 OpenClawCommand 或 HermesJob。
- 传播 tenant、task、trace、idempotency 和 policy version。
- 将 Runtime Result 回写 AIOS Task。
不负责:
- 创建租户。
- 修改策略。
- 绕过审批。
- 保存第二套任务状态机。
候选实现:
- `active-runtime/app/aios_v2/orchestrator.py`
### 3.4 OpenClaw Adapter
职责:
- 验证签名命令。
- 执行、查询、取消、健康检查和回调。
- 支持幂等、超时和租户隔离。
- 把 Runtime 生命周期转换为 AIOS 审计事件。
不负责:
- Tenant、Policy、Approval、Task Lifecycle。
- 向浏览器暴露公网控制接口。
候选实现:
- `active-runtime/app/aios_openclaw/contract.py`
- `active-runtime/app/aios_openclaw/adapter.py`
- `active-runtime/app/aios_openclaw/transport.py`
- `active-runtime/app/aios_openclaw/health.py`
- `active-runtime/app/aios_openclaw/callback.py`
### 3.5 Hermes Runtime Boundary
职责:
- Tenant/Cell/Profile 分区队列。
- Worker claim、start、heartbeat、complete、fail、cancel。
- Lease 超时和自动回队列。
- Result 回传和 Runtime Health。
不负责:
- Policy、Approval、Tenant Registry、权威 Task Lifecycle。
候选实现:
- `active-runtime/app/aios_hermes/contract.py`
- `active-runtime/app/aios_hermes/adapter.py`
- `active-runtime/app/aios_hermes/worker.py`
### 3.6 Persistence
候选实现使用 SQLite WAL 原子快照,持久化:
- Tenant
- Agent
- Task
- Audit
- Idempotency
- Execution Evidence 文件
生产演进:
- Store 接口保持稳定。
- 可替换为 PostgreSQL、对象存储或分布式事件存储。
- 生产替换必须提供迁移、备份、恢复、HA、容量和租户隔离证据。
## 4. 多租户架构
### 4.1 权威租户上下文
每次请求和执行至少传播:
```text
tenant_id
principal_id
session_id
trace_id
task_id
policy_version
```
Runtime 还传播:
```text
command_id / job_id
cell_id
profile_id
idempotency_key
worker_id
```
### 4.2 隔离层级
| 层级 | 隔离要求 |
|---|---|
| HTTP/API | 可信身份头、租户不可由浏览器覆盖 |
| Control Plane | 所有资源带 `tenant_id`,跨租户默认拒绝 |
| OpenClaw | 签名命令绑定 tenant/task/trace |
| Hermes | tenant/cell/profile 三层分区 |
| Storage | 候选为显式租户字段;生产需 RLS 或物理隔离策略 |
| Evidence/Audit | 仅按当前租户查询 |
| Cache/Memory | key 必须包含 tenant namespace |
| Secret | 只保存引用,不保存明文 |
| Logs/Metrics | 必须包含 tenant/task/trace 标签,并限制查询范围 |
### 4.3 失败行为
- 跨租户资源访问返回拒绝或“未找到”。
- 不泄露其他租户资源存在性。
- 缺少 tenant context 时 fail closed。
- Runtime 返回的 tenant/task/trace 不匹配时拒绝回写。
## 5. 任务与执行数据流
### 5.1 任务创建与审批
```text
User/API
|
v
Trusted Identity
|
v
AIOS Gateway
|
v
Policy Gate
|
v
Task CREATED
|
v
Approval
|
v
Task APPROVED
```
### 5.2 OpenClaw 执行
```text
APPROVED Task
|
v
Execution Orchestrator
|
v
Signed OpenClawCommand
|
v
OpenClaw Adapter
|
v
Runtime Result
|
v
Task Reconciliation + Audit + Evidence
```
### 5.3 Hermes 执行
```text
APPROVED Task
|
v
Execution Orchestrator
|
v
Tenant/Cell/Profile HermesJob
|
v
Worker Claim + Lease + Heartbeat
|
v
HermesResult
|
v
Task Reconciliation + Audit + Evidence
```
## 6. 安全架构
### 6.1 身份边界
- Public API 位于可信反向代理之后。
- 反向代理完成登录、会话、MFA 和身份头注入。
- 候选应用拒绝未认证身份断言。
- 生产不得允许客户端直接伪造 `X-AIOS-*`。
### 6.2 Runtime 命令完整性
OpenClawCommand 使用规范化序列化和 HMAC-SHA256:
- 签名覆盖 tenant、task、trace、action、payload、idempotency 和时间。
- 支持 key ID、到期时间和时钟偏差。
- Runtime 密钥通过受保护文件或外部 Secret Provider 解析。
### 6.3 Secret 边界
Git 允许:
```text
secret_ref
credential_ref
key_id
受保护文件路径
```
Git 禁止:
```text
API Key
密码
Token
私钥
签名材料
```
### 6.4 Web 安全
候选应用设置:
- no-store
- nosniff
- DENY frame
- no-referrer
- CSP
- 请求体大小限制
生产还需验证:
- CSRF/CORS
- Rate Limit
- Session Timeout
- CSP Nonce/Hash
- WAF 和反向代理访问策略
## 7. 可靠性架构
### 7.1 幂等
- Control Plane:`tenant_id + idempotency_key + request_digest`。
- OpenClaw:`tenant_id + idempotency_key + signature`。
- Hermes:`tenant_id + idempotency_key + job_digest`。
### 7.2 并发
- Task 使用 revision。
- Hermes 使用 revision + worker lease。
- GitOps 使用 detached controller + branch lease。
- 持久化使用 SQLite `BEGIN IMMEDIATE` 和原子快照。
### 7.3 恢复
- Controller restart:detached guard。
- AIOS restart:SQLite restore。
- Worker crash:lease 到期后 Job 回到 QUEUED。
- Release failure:健康检查失败自动恢复 previous release。
## 8. 可观测架构
统一关联键:
```text
tenant_id
principal_id
trace_id
task_id
agent_id
command_id / job_id
worker_id
commit
```
Web Console 展示:
- Tasks
- Agents
- Runtime Health
- Audit
- Execution Evidence
- GitOps Status
生产还需接入:
- 结构化日志
- Metrics
- Distributed Tracing
- Alerting
- SLO Dashboard
## 9. 部署架构
支持两种候选部署:
### systemd
- 非 root `aios` 用户。
- `ProtectSystem=strict`、`NoNewPrivileges` 等安全选项。
- 状态目录 `/var/lib/aios-v2`。
### Docker/Compose
- 非 root UID/GID。
- Read-only root filesystem。
- Drop all capabilities。
- State volume。
- HTTP healthcheck。
## 10. 架构约束
以下行为视为架构违规:
1. OpenClaw/Hermes 直接创建或批准 AIOS Task。
2. Web Console 直接调用 Runtime。
3. 新模块不携带 `tenant_id`。
4. 新增第二套任务状态机或租户事实源。
5. 在 Git 中提交明文凭据。
6. 绕过 GitOps 修改生产代码/配置。
7. 架构或功能变化未同步更新 `docs/aios-v2/`。
8. 未通过测试和预发布门禁直接提升候选版本。
## 11. 目标架构完成判定
目标架构只有在以下证据齐全时才视为完成:
- 文档与代码映射完整。
- Candidate tests 对精确 commit 执行通过。
- OpenClaw/Hermes 实际 Runtime 联调通过。
- 多租户负向测试通过。
- 重启恢复和回滚演练通过。
- GitOps 本地/Gitea OID 一致。
- Owner 批准 merge、Tag 和生产部署。
+313
View File
@@ -0,0 +1,313 @@
# AIOS V2 迁移与交付策略
- 文档版本:`V2R1C0-candidate`
- 交付原则:增量迁移、能力复用、控制权收敛、GitOps 唯一变更入口
## 1. 迁移目标
AIOS V2 不通过一次性重写替换现有系统,而是在保留 OpenClaw、Hermes、Agent Router、GitOps 和既有运行能力的前提下,将租户、策略、任务、审批、审计和发布治理收敛到 AIOS Control Plane。
迁移完成的标志不是“新代码存在”,而是:
- AIOS 成为唯一权威控制面;
- OpenClaw/Hermes 不再保存第二套控制状态;
- 所有控制功能 Web/API 化;
- 所有模块通过租户隔离、测试和预发布门禁;
- 本地 Git、Gitea、测试、部署和回滚证据可追踪。
## 2. 为什么采用增量迁移
### 2.1 避免能力回退
OpenClaw、Hermes 和 Agent Router 已经包含大量运行经验、边界处理和执行能力。整体重写会扩大风险面,并导致业务能力、操作流程和历史数据回退。
### 2.2 保持可回滚
每个迁移步骤必须:
- 独立分支;
- 小范围提交;
- 独立测试;
- 可单独回滚;
- 不要求立即关闭旧执行路径。
### 2.3 先收敛控制权,再优化内部实现
迁移优先级:
```text
唯一入口
-> 权威租户/任务/策略
-> Runtime Adapter
-> 数据层生产化
-> Web 与可观测
-> 性能与 HA
```
## 3. 分阶段迁移路线
### Phase 0:GitOps 与架构事实源
目标:建立可靠变更链和权威设计基线。
已完成候选能力:
- Controller detached guard;
- 避免 fetch 到已检出本地分支;
- Gitea SSH 镜像;
- `docs/aios-v2/` 架构事实源;
- 候选分支、部署和回滚资产。
退出条件:
- Controller detached 且 clean;
- 本地/Gitea OID 一致;
- 架构与需求文档纳入版本。
### Phase 1:Control Plane 权威化
目标:将 Tenant、Agent、Task、Policy、Approval 和 Audit 收敛到 AIOS。
候选实现:
- `aios_v2/control_plane.py`
- `aios_v2/persistence.py`
- `aios_v2/gateway.py`
- `aios_v2/app.py`
退出条件:
- 任务状态机统一;
- 幂等、revision、审计和跨租户拒绝通过测试;
- 服务重启后权威状态恢复。
### Phase 2:OpenClaw 解耦
目标:OpenClaw 从控制逻辑转为执行适配器。
步骤:
1. 定义签名命令合同;
2. 引入 Adapter;
3. 将直接调用迁移到 AIOS Orchestrator;
4. Runtime 只返回执行结果和健康状态;
5. 关闭或网络隔离旧公共控制入口。
退出条件:
- 未批准任务无法生成有效命令;
- 命令绑定 tenant/task/trace/idempotency;
- 旧控制接口冻结并记录替代路径。
### Phase 3:Hermes 解耦
目标:Hermes 只承担租户分区队列和 Worker Runtime。
步骤:
1. 定义 HermesJob/HermesResult;
2. 强制 tenant/cell/profile 分区;
3. 引入 claim/lease/heartbeat;
4. 将结果统一回写 AIOS;
5. 移除 Hermes 内部策略和任务事实源。
退出条件:
- Worker 无权创建 AIOS Task;
- 租约过期后 Job 可恢复;
- 跨租户 claim/read/result 被拒绝。
### Phase 4:Web 与运维统一
目标:所有用户与运维功能通过 AIOS Web/API。
范围:
- Tenant、Agent、Task、Runtime、Audit、Execution Evidence、GitOps;
- systemd/Docker/Compose;
- healthcheck、rollback、release report。
退出条件:
- 浏览器不直接调用 Runtime;
- 生产身份代理和权限模型通过验证;
- 部署、重启和回滚演练通过。
### Phase 5:生产数据层与 HA
目标:将候选 SQLite/内存运行状态替换或强化为生产级后端。
要求:
- 数据迁移;
- RLS 或物理隔离;
- 备份恢复;
- HA 与容量;
- 多实例并发;
- 灾难恢复目标。
该阶段不得改变上层合同,避免 Runtime 和 Web 再次耦合存储实现。
## 4. 分支与提交策略
### 4.1 分支类型
| 类型 | 用途 |
|---|---|
| `task/*` | 可执行任务合同与主体实现 |
| `patch/*` | 独立修复或窄范围增强 |
| `integration/*` | 多模块候选收敛 |
| `result/*` | 自动执行结果与证据 |
| `release/*` | 受控发布候选 |
### 4.2 规则
- 禁止直接修改 `main`;
- 一个提交应可独立审阅和回滚;
- 大文件和大范围变更应拆分;
- 所有行为变化必须同步文档和测试;
- 生产 Tag、部署和回滚需要 Owner 批准。
## 5. GitOps 交付流程
```text
需求/架构更新
|
v
任务分支 + 文档 + 测试
|
v
代码提交 + Secret Gate
|
v
本地 bare repo 同步
|
v
Gitea 镜像与 OID 校验
|
v
Candidate 集成
|
v
自动测试与预发布
|
v
Owner 审批
|
v
Merge / Tag / Production Deploy
```
### 5.1 同步状态
GitOps 返回必须区分:
- local sync;
- Gitea sync;
- branch OID;
- workspace mode;
- dirty state。
仅返回 `common_git_dir synced` 不足以证明 Web 远端可见。
## 6. 发布门禁
### Gate A:设计完整性
- 用户需求、目标架构、模块契约、ADR 和追踪矩阵同步;
- 无冲突标记、断链或过期路径。
### Gate B:代码质量
- 包可导入;
- 静态编译通过;
- Secret Scanner 通过;
- 无已知越权旁路。
### Gate C:测试
- 单元、集成、多租户负向、重启恢复、Runtime 边界测试通过;
- 记录 commit、命令、exit code、pass/fail 和 artifact digest。
### Gate D:预发布
- systemd 或 Compose 启动成功;
- 健康检查通过;
- OpenClaw/Hermes 真实联调通过;
- 备份、重启和回滚演练通过;
- 日志中无敏感信息。
### Gate E:生产批准
- Owner 批准 merge、Tag 和生产部署;
- 发布窗口、回滚负责人和观察指标明确;
- 生产变更通过 GitOps 记录。
## 7. 回滚策略
### 7.1 代码回滚
- 使用 `/opt/aios/current` 原子 symlink 切换;
- 回滚前备份状态;
- 健康检查失败自动恢复 previous release。
### 7.2 状态回滚
默认不自动回滚状态,避免覆盖候选运行期间新增的 Task、Audit 和 Evidence。
状态恢复只有在:
- 明确确认 schema 不兼容;
- 数据备份可验证;
- Owner 显式授权;
时启用。
### 7.3 Runtime 回滚
OpenClaw/Hermes Adapter 采用稳定合同。若候选执行层失败,可恢复旧 Runtime 版本,但不得恢复旧控制权旁路。
## 8. Bug 修复流程
生产或预发布 Bug:
1. 创建 `patch/bugfix/*`;
2. 添加可复现测试;
3. 修复代码;
4. 更新相关需求/架构/契约/ADR;
5. 运行候选验收;
6. 进入新的 release candidate;
7. Owner 批准部署。
禁止在生产目录直接编辑文件后不回写 Git。
## 9. 文档同步规则
以下变化必须更新文档:
- 用户行为或页面;
- API 路径、Header、请求/响应;
- 数据模型、状态机、策略;
- OpenClaw/Hermes/Router 边界;
- 多租户与安全模型;
- 部署、健康检查、回滚;
- 测试或发布门禁。
对应关系:
- 需求变化:更新 `01-USER-REQUIREMENTS.md`;
- 架构变化:更新 `02-TARGET-ARCHITECTURE.md` 和 ADR;
- 迁移/发布变化:更新本文件;
- 合同变化:更新 `04-MODULE-CONTRACTS.md`;
- 代码/测试路径变化:更新 `05-TRACEABILITY-MATRIX.md`。
## 10. 当前候选后续步骤
1. 完成文档同步门禁和 Acceptance Runner;
2. 在精确 Candidate commit 执行测试;
3. 推送 Gitea 并校验 OID;
4. 部署隔离预发布环境;
5. 联调真实 OpenClaw/Hermes;
6. 执行重启、性能、安全和回滚演练;
7. 生成最终 evidence;
8. 请求 Owner 的 merge/Tag/production approval。
+502
View File
@@ -0,0 +1,502 @@
# AIOS V2 模块契约
- 文档版本:`V2R1C0-candidate`
- 状态:候选实现契约基线
- 原则:Control Plane 权威、Runtime 执行、全链路租户隔离、所有状态可审计
## 1. 通用上下文契约
所有控制与执行请求必须携带以下上下文:
| 字段 | 必填 | 说明 |
|---|---:|---|
| `tenant_id` | 是 | 权威租户标识 |
| `principal_id` | 控制请求是 | 已认证主体 |
| `trace_id` | 是 | 全链路追踪 |
| `task_id` | 执行阶段是 | AIOS 权威任务 |
| `session_id` | Web/交互请求建议 | 用户会话 |
| `policy_version` | 控制请求是 | 策略版本 |
| `idempotency_key` | 创建/执行是 | 幂等键 |
规则:
- 缺少必填字段时 fail closed。
- 客户端不得覆盖由可信代理或 AIOS 生成的 `tenant_id`。
- Runtime 返回的 tenant/task/trace 必须与请求完全匹配。
- 所有持久化键、缓存键和队列分区必须包含租户命名空间。
## 2. Tenant Contract
### 2.1 TenantRecord
```text
tenant_id
display_name
status
metadata
created_at
updated_at
```
约束:
- `tenant_id` 不可为空且不可变。
- `status` 至少支持 `active` 和受控停用状态。
- Tenant 只由 AIOS Control Plane 创建和维护。
- Runtime 不得创建租户。
实现:`active-runtime/app/aios_v2/control_plane.py`
## 3. Agent Contract
### 3.1 AgentRecord
```text
agent_id
tenant_id
display_name
runtime
capabilities
enabled
metadata
created_at
updated_at
```
约束:
- Agent 属于唯一租户。
- `runtime` 由 Registry 决定,可取 `openclaw`、`hermes` 或后续注册 Runtime。
- `capabilities` 是允许调度的动作集合。
- `metadata` 可存 `cell_id`、`profile_id`、priority 等非敏感引用。
- 不得在 Agent metadata 中保存明文凭据。
## 4. Task Contract
### 4.1 TaskRecord
```text
task_id
tenant_id
agent_id
action
payload
state
revision
idempotency_key
request_digest
result
error
created_by
approved_by
created_at
updated_at
```
### 4.2 状态机
```text
CREATED -> APPROVED -> RUNNING -> COMPLETED
| | |
+----------+----------+-> CANCELLED
+-> FAILED
```
允许转换:
| 当前 | 允许目标 |
|---|---|
| CREATED | APPROVED, CANCELLED |
| APPROVED | RUNNING, CANCELLED |
| RUNNING | COMPLETED, FAILED, CANCELLED |
| COMPLETED | 无 |
| FAILED | 无 |
| CANCELLED | 无 |
约束:
- 每次转换必须校验 `revision`。
- COMPLETED 必须带 `result`。
- FAILED 必须带 `error`。
- 每次转换生成 AuditEvent。
- Runtime 不得直接写 Task Store,只能通过 Orchestrator reconcile。
## 5. Policy Contract
### 5.1 PolicyDecision
```text
allowed
reason
policy_version
```
约束:
- Policy 默认拒绝。
- Runtime 不得覆盖 PolicyDecision。
- 拒绝必须审计。
- 高风险动作必须经过 Approval。
## 6. Audit Contract
### 6.1 AuditEvent
```text
event_id
tenant_id
principal_id
trace_id
action
resource_type
resource_id
outcome
policy_version
details
created_at
```
约束:
- 按租户查询。
- `details` 不得包含明文凭据。
- OpenClaw/Hermes 生命周期事件必须映射到统一 AuditEvent。
- 生产版需增加不可变外部固化和完整性验证。
## 7. OpenClaw Contract
### 7.1 OpenClawCommand
```text
contract_version
command_id
tenant_id
task_id
trace_id
action
payload
idempotency_key
created_at
expires_at
key_id
signature
```
签名覆盖:
- command_id
- tenant_id
- task_id
- trace_id
- action
- payload
- idempotency_key
- created_at / expires_at
- contract_version / key_id
约束:
- 使用规范化 JSON 和 HMAC-SHA256。
- 只接受已批准任务生成的命令。
- 过期、未来时间超差、签名错误命令必须拒绝。
- `key_id` 只引用运行时材料,不包含材料本身。
### 7.2 OpenClawResult
```text
command_id
tenant_id
task_id
trace_id
state
output
error
started_at
finished_at
runtime_id
attempts
```
状态:
```text
ACCEPTED
RUNNING
COMPLETED
FAILED
CANCELLED
TIMED_OUT
REJECTED
```
约束:
- Adapter 对外只接受终态结果用于 reconcile。
- 身份字段不匹配时拒绝回写。
- 结果通过 Orchestrator 更新 AIOS Task。
### 7.3 RuntimeHealth
```text
runtime
runtime_id
status
version
heartbeat_at
active_tasks
workspace_isolated
credential_ref_configured
details
```
健康策略:
- heartbeat 不得超时。
- workspace isolation 与 credential reference 必须符合策略。
- 不安全 Runtime 应 quarantine,不得继续接收新任务。
实现:`active-runtime/app/aios_openclaw/`
## 8. Hermes Contract
### 8.1 HermesJob
```text
job_id
tenant_id
cell_id
profile_id
task_id
trace_id
action
payload
idempotency_key
priority
state
revision
assigned_worker_id
lease_expires_at
created_at
updated_at
```
状态机:
```text
QUEUED -> CLAIMED -> RUNNING -> COMPLETED
| | |
+---------+----------+-> CANCELLED
+-> FAILED
```
约束:
- `tenant_id + cell_id + profile_id` 形成执行分区。
- claim 只能领取同分区 Job。
- Worker ID 不得跨分区复用。
- Lease 到期 Job 回到 QUEUED。
- revision 冲突不得覆盖。
### 8.2 HermesResult
```text
job_id
tenant_id
task_id
trace_id
state
worker_id
output
error
started_at
finished_at
```
约束:
- COMPLETED 必须有 output。
- FAILED 必须有 error。
- 只允许 AIOS Orchestrator reconcile。
### 8.3 HermesHealth
```text
runtime_id
status
version
heartbeat_at
active_workers
active_jobs
queued_jobs
tenant_partitions
profile_isolation_enforced
details
```
实现:`active-runtime/app/aios_hermes/`
## 9. Execution Evidence Contract
### 9.1 ExecutionEvidence
```text
execution_id
tenant_id
task_id
agent_id
executor
branch
status
progress_completed
progress_total
current_step
last_commit
test_status
test_summary
result_summary
claimed_at
heartbeat_at
completed_at
```
状态建议:
```text
claimed
running
blocked
complete
failed
cancelled
```
约束:
- 按租户存储和查询。
- 每个阻塞必须记录原因、责任方和解除条件。
- 每次结果必须记录 branch、commit、测试命令和状态。
- 不得包含 Secret 或完整敏感 payload。
## 10. GitOps Status Contract
```text
available
service_status
workspace_mode
workspace_branch
workspace_ref
workspace_dirty
candidate_branch
release_channel
release_version
sync_status
details
```
约束:
- Web 响应不得包含 remote URL、凭据或 stderr。
- Controller 必须 `detached + clean`。
- `sync_status` 必须区分 local 和 Gitea 镜像状态。
## 11. Web/API Contract
### 11.1 身份头
可信代理注入:
```text
X-AIOS-Authenticated
X-AIOS-Tenant-ID
X-AIOS-Principal-ID
X-AIOS-Roles
X-AIOS-Trace-ID
X-AIOS-Session-ID
X-AIOS-Policy-Version
```
生产约束:
- 代理必须删除客户端同名头后重新注入。
- 应用不得直接暴露公网端口。
### 11.2 核心 API
```text
GET /api/v2/health
POST /api/v2/tenants
GET /api/v2/tenants/current
POST /api/v2/agents
GET /api/v2/agents
POST /api/v2/tasks
GET /api/v2/tasks
GET /api/v2/tasks/{id}
POST /api/v2/tasks/{id}/approve
POST /api/v2/tasks/{id}/start
POST /api/v2/tasks/{id}/complete
POST /api/v2/tasks/{id}/fail
POST /api/v2/tasks/{id}/cancel
POST /api/v2/tasks/{id}/dispatch
GET /api/v2/audit
GET /api/v2/runtime
GET /api/v2/executions
GET /api/v2/gitops
POST /api/v2/hermes/results/reconcile
```
## 12. Persistence Contract
候选版:
- SQLite WAL
- `BEGIN IMMEDIATE`
- 原子 snapshot
- schema version
- restart restore
持久化范围:
- tenants
- agents
- tasks
- audit
- idempotency
生产替换要求:
- 不改变 Domain/API Contract。
- 提供 migration、backup、restore、HA、RLS/隔离和容量测试。
## 13. Deployment Contract
### systemd
- 非 root `aios` 用户。
- 状态目录 `/var/lib/aios-v2`。
- 最小权限和文件系统保护。
### Docker
- 非 root UID/GID。
- read-only root filesystem。
- drop all capabilities。
- 独立状态 volume。
### Rollback
- 回滚前备份状态。
- 默认只回滚代码。
- 状态恢复需显式授权。
- 回滚后必须执行 healthcheck。
## 14. 契约变更规则
任何以下变化必须在同一提交或同一分支更新本文档:
- 字段、状态、Header、API、错误语义;
- 租户或安全约束;
- OpenClaw/Hermes 边界;
- Persistence 或 Deployment Contract;
- Evidence/Audit 格式。
不兼容变化必须新增 ADR 和版本号,不得静默修改。
+113
View File
@@ -0,0 +1,113 @@
# AIOS V2 需求—架构—代码—测试—部署追踪矩阵
- 文档版本:`V2R1C0-candidate`
- 目的:确保每项用户需求都能追踪到架构设计、代码实现、自动测试、部署资产和验收命令
## 1. 追踪规则
每条需求至少具备:
1. 需求编号;
2. 架构章节;
3. 实现路径;
4. 测试路径;
5. 部署/运维资产;
6. 验收状态。
状态定义:
- `Implemented`:代码已进入 Candidate 分支;
- `Test-defined`:测试代码已进入 Candidate 分支;
- `Test-verified`:精确 commit 的测试已运行并保存证据;
- `Pre-production verified`:预发布部署、健康检查和回滚演练已通过;
- `Production released`:Owner 批准并完成 merge/tag/deploy。
## 2. 核心追踪矩阵
| 需求 | 设计依据 | 代码实现 | 测试 | 部署/运维 | 当前状态 |
|---|---|---|---|---|---|
| UR-001 AIOS 唯一控制面 | `02-TARGET-ARCHITECTURE.md` §3.2/§3.3;ADR-001 | `aios_v2/control_plane.py`、`aios_v2/orchestrator.py`、`aios_v2/app.py` | `tests/aios_v2_candidate/test_candidate_flow.py` | `deploy/aios-v2/systemd/aios-v2.service` | Implemented / Test-defined |
| UR-002 复用现有能力 | `02-TARGET-ARCHITECTURE.md` §2;`03-MIGRATION-AND-DELIVERY.md` §2 | `aios_openclaw/`、`aios_hermes/` | OpenClaw/Hermes flow tests | Candidate branch + adapter boundaries | Implemented / Test-defined |
| UR-003 OpenClaw/Hermes 解耦 | `02-TARGET-ARCHITECTURE.md` §3.4/§3.5;ADR-002 | `aios_openclaw/adapter.py`、`aios_hermes/adapter.py` | runtime boundary tests in candidate flow suite | systemd/Docker runtime configuration | Implemented / Test-defined |
| UR-004 全面 Web 化 | `02-TARGET-ARCHITECTURE.md` §3.1/§8 | `aios_v2/console.py`、`aios_v2/gateway.py`、`aios_v2/app.py` | API/console smoke test pending | Docker/systemd + reverse proxy requirement | Implemented / Test pending |
| UR-005 多租户隔离 | `02-TARGET-ARCHITECTURE.md` §4;ADR-004 | `control_plane.py`、`orchestrator.py`、`aios_openclaw/*`、`aios_hermes/*`、`execution_tracking.py` | `test_cross_tenant_task_read_is_denied` | tenant-scoped state/evidence directories | Implemented / Test-defined |
| UR-006 权威任务生命周期 | `04-MODULE-CONTRACTS.md` §4 | `control_plane.py` | OpenClaw/Hermes full-flow tests | persisted state + audit | Implemented / Test-defined |
| UR-007 策略与审批 | `04-MODULE-CONTRACTS.md` §5 | `control_plane.py`、`gateway.py` | role/policy negative tests需补充 | reverse proxy identity + audit | Implemented / Partial test |
| UR-008 幂等与并发安全 | `02-TARGET-ARCHITECTURE.md` §7 | `control_plane.py`、`aios_openclaw/adapter.py`、`aios_hermes/adapter.py`、`worker.py` | flow + restart tests;并发专项测试需补充 | GitOps detached guard + SQLite WAL | Implemented / Partial test |
| UR-009 执行证据 | `04-MODULE-CONTRACTS.md` §9 | `execution_tracking.py` | `test_execution_tracking_persists_claim_heartbeat_and_result` | `/var/lib/aios-v2/executions` | Implemented / Test-defined |
| UR-010 GitOps 唯一变更流程 | `03-MIGRATION-AND-DELIVERY.md` §4/§5;ADR-003 | GitOps Runtime、`gitops_status.py` | healthcheck 的 detached/clean 检查 | `deploy/aios-v2/healthcheck/check.sh` | Implemented / Runtime verified partially |
| UR-011 持久化与恢复 | `02-TARGET-ARCHITECTURE.md` §3.6 | `persistence.py` | `test_durable_service_recovers_state_after_restart` | state dir + backup/rollback | Implemented / Test-defined |
| UR-012 审计与合规 | `04-MODULE-CONTRACTS.md` §6 | `control_plane.py`、`aios_openclaw/callback.py`、`app.py` | flow test + audit assertions需补充 | Console audit view | Implemented / Partial test |
| UR-013 密钥安全 | `02-TARGET-ARCHITECTURE.md` §6;ADR-006 | `FileSigningMaterialResolver`、credential refs、Secret Gate | permission/secret-scan test需补充 | env template + external protected file | Implemented / Partial test |
| UR-014 部署与回滚 | `03-MIGRATION-AND-DELIVERY.md` §7 | deployment scripts/assets | acceptance runner + healthcheck | `deploy/aios-v2/` | Implemented / Pre-prod pending |
| UR-015 文档与代码同步 | `03-MIGRATION-AND-DELIVERY.md` §9;ADR-005 | `tools/aios_v2_docs_code_check.py` | checker self-test / acceptance runner | CI/Gitea Actions 可调用 | Implemented after checker commit |
## 3. API 追踪
| API | 需求 | 实现 | 测试状态 |
|---|---|---|---|
| `GET /api/v2/health` | UR-004, UR-014 | `gateway.py` | healthcheck 定义完成 |
| `POST /api/v2/tenants` | UR-001, UR-005 | `gateway.py` / `control_plane.py` | 需增加 HTTP 集成测试 |
| `POST /api/v2/agents` | UR-001, UR-005 | `gateway.py` / `control_plane.py` | Domain flow 已覆盖 |
| `POST /api/v2/tasks` | UR-006, UR-008 | `gateway.py` / `control_plane.py` | Domain flow 已覆盖 |
| `POST /api/v2/tasks/{id}/approve` | UR-006, UR-007 | `gateway.py` | Domain flow 已覆盖 |
| `POST /api/v2/tasks/{id}/dispatch` | UR-001, UR-003 | `app.py` / `orchestrator.py` | OpenClaw/Hermes flow 已覆盖 |
| `GET /api/v2/runtime` | UR-004, UR-009 | `app.py` | HTTP smoke test待执行 |
| `GET /api/v2/executions` | UR-009 | `app.py` / `execution_tracking.py` | Evidence persistence test已定义 |
| `GET /api/v2/gitops` | UR-010 | `app.py` / `gitops_status.py` | Runtime/healthcheck验证待执行 |
| `POST /api/v2/hermes/results/reconcile` | UR-003, UR-006 | `app.py` / `orchestrator.py` | Hermes flow已覆盖 Domain 路径 |
## 4. 运行时契约追踪
| 契约 | 设计 | 代码 | 测试 |
|---|---|---|---|
| OpenClawCommand | `04-MODULE-CONTRACTS.md` §7.1 | `aios_openclaw/contract.py` | OpenClaw full-flow;签名负向测试需补充 |
| OpenClawResult | §7.2 | `aios_openclaw/contract.py` | OpenClaw full-flow |
| RuntimeHealth | §7.3 | `aios_openclaw/contract.py`、`health.py` | health policy测试需补充 |
| HermesJob | §8.1 | `aios_hermes/contract.py` | Hermes full-flow |
| HermesResult | §8.2 | `aios_hermes/contract.py` | Hermes full-flow |
| HermesHealth | §8.3 | `aios_hermes/contract.py`、`adapter.py` | healthcheck/worker test待执行 |
| ExecutionEvidence | §9 | `execution_tracking.py` | evidence persistence test |
| GitOpsStatus | §10 | `gitops_status.py` | healthcheck detached/clean检查 |
## 5. 部署追踪
| 能力 | 文件 | 验收命令 |
|---|---|---|
| systemd | `deploy/aios-v2/systemd/aios-v2.service` | `systemctl status aios-v2` |
| 环境模板 | `deploy/aios-v2/env/aios-v2.env.example` | `systemd-analyze verify` + 文件权限检查 |
| Docker | `deploy/aios-v2/docker/Dockerfile` | `docker build` |
| Compose | `deploy/aios-v2/docker/compose.yaml` | `docker compose config` / `up` |
| 健康检查 | `deploy/aios-v2/healthcheck/check.sh` | 直接执行 |
| 回滚 | `deploy/aios-v2/rollback/rollback.sh` | 预发布回滚演练 |
| 全量验收 | `deploy/aios-v2/acceptance/run-all.sh` | 直接执行 |
## 6. 文档覆盖追踪
| 文档 | 覆盖内容 | 代码/变更触发条件 |
|---|---|---|
| `README.md` | 事实源索引、状态语言、实现地图 | 新模块/目录/版本 |
| `01-USER-REQUIREMENTS.md` | 用户需求、NFR、验收场景 | 用户行为、功能范围、验收变化 |
| `02-TARGET-ARCHITECTURE.md` | 分层、边界、数据流、安全、租户 | 架构边界、数据流、运行时变化 |
| `03-MIGRATION-AND-DELIVERY.md` | 迁移、GitOps、Gate、回滚 | 发布流程、迁移阶段、门禁变化 |
| `04-MODULE-CONTRACTS.md` | 字段、API、状态机、错误语义 | Contract/API/模型变化 |
| `05-TRACEABILITY-MATRIX.md` | 需求到代码测试部署映射 | 代码、测试、部署路径变化 |
| `adr/*` | 关键不可逆或跨模块决策 | 新架构决策或旧决策替换 |
## 7. 当前未完成验证项
以下项目代码或测试定义已存在,但尚不能标记为 `Test-verified` 或 `Pre-production verified`:
1. 精确 Candidate commit 的测试执行结果;
2. HTTP API 与 Web Console 冒烟测试;
3. 真实 OpenClaw Runtime 签名、取消、超时和健康联调;
4. Hermes 长任务心跳、Worker 崩溃和 Lease 恢复演练;
5. 多实例 SQLite/并发容量边界;
6. systemd、Docker、Compose、restart 和 rollback 演练;
7. local/Gitea Candidate OID 一致证明;
8. 生产身份代理、防伪 Header、CSRF/CORS/Rate Limit 验证;
9. 文档—代码同步检查器在 CI/Acceptance 中执行通过。
## 8. 更新要求
任何代码、API、部署或测试路径变化,必须同步更新本矩阵。验收脚本会检查本文档引用的关键路径存在,并检查架构/功能代码变化是否同时包含 `docs/aios-v2/` 变化。
+98
View File
@@ -0,0 +1,98 @@
# AIOS V2 Architecture and Delivery Index
- Version line: `V2R1C0-candidate`
- Candidate branch: `integration/aios-v2r1-cto-production-candidate-20260720`
- Architecture owner: AIOS V2 CTO
- Status: implementation-complete candidate; runtime and production verification pending
## 1. Purpose
This directory is the single source of truth for AIOS V2 requirements, target architecture, module boundaries, migration strategy, release gates and architectural decisions.
The design is based on the following non-negotiable principles:
1. Do not rewrite the existing platform.
2. Reuse OpenClaw, Hermes and Agent Router capabilities.
3. Decouple OpenClaw and Hermes from control-plane ownership.
4. AIOS is the only control plane.
5. All user-facing and operational capabilities are Web/API accessible.
6. Every module enforces multi-tenant isolation.
7. Every change flows through Git branches, commits, tests, review and controlled promotion.
## 2. Document map
| Document | Purpose |
|---|---|
| [01-USER-REQUIREMENTS.md](01-USER-REQUIREMENTS.md) | Product goals, user requirements, non-functional requirements and acceptance criteria |
| [02-TARGET-ARCHITECTURE.md](02-TARGET-ARCHITECTURE.md) | Target logical architecture, control boundaries, data flow, security and tenancy model |
| [03-MIGRATION-AND-DELIVERY.md](03-MIGRATION-AND-DELIVERY.md) | Incremental migration plan, reuse strategy, GitOps workflow and release gates |
| [04-MODULE-CONTRACTS.md](04-MODULE-CONTRACTS.md) | Tenant context, task lifecycle, OpenClaw, Hermes, audit and execution-evidence contracts |
| [05-TRACEABILITY-MATRIX.md](05-TRACEABILITY-MATRIX.md) | Requirement-to-design-to-code-to-test mapping |
| [adr/README.md](adr/README.md) | Architecture decision records |
## 3. Current implementation map
| Architecture capability | Candidate implementation |
|---|---|
| Tenant, agent, task, policy and audit control plane | `active-runtime/app/aios_v2/control_plane.py` |
| Durable control-plane state | `active-runtime/app/aios_v2/persistence.py` |
| Web/API gateway | `active-runtime/app/aios_v2/gateway.py` |
| Unified application entrypoint | `active-runtime/app/aios_v2/app.py` |
| Runtime dispatch and reconciliation | `active-runtime/app/aios_v2/orchestrator.py` |
| Agent execution evidence | `active-runtime/app/aios_v2/execution_tracking.py` |
| GitOps runtime visibility | `active-runtime/app/aios_v2/gitops_status.py` |
| Web Console | `active-runtime/app/aios_v2/console.py` |
| OpenClaw execution boundary | `active-runtime/app/aios_openclaw/` |
| Hermes runtime boundary | `active-runtime/app/aios_hermes/` |
| Candidate tests | `tests/aios_v2_candidate/` |
| Deployment, health and rollback | `deploy/aios-v2/` |
## 4. Status vocabulary
The project separates code availability from release readiness:
- **Designed**: documented boundary and acceptance criteria exist.
- **Implemented**: code is committed on the candidate branch.
- **Test-defined**: automated tests exist in Git.
- **Test-verified**: tests were executed for the exact commit and evidence was recorded.
- **Pre-production verified**: deployed in an isolated environment with health, restart and rollback evidence.
- **Production released**: owner-approved merge/tag/deployment completed.
A branch or file existing does not, by itself, mean a capability is production-ready.
## 5. Design summary
```text
Browser / Web Console / Public API
|
v
AIOS Control Plane
Tenant | Agent | Task | Policy | Audit
|
v
Execution Orchestrator / Router
/ \
v v
OpenClaw Adapter Hermes Boundary
| |
v v
OpenClaw Runtime Hermes Workers
\
v
Result Reconciliation + Evidence
```
The control plane owns authoritative identity, policy, approval, task lifecycle and audit state. OpenClaw and Hermes receive only tenant-scoped, approved execution work and return runtime results.
## 6. Release rule
No candidate may be promoted until:
- requirement-to-code traceability is complete;
- multi-tenant negative tests pass;
- controller workspace is detached and clean;
- local Git and Gitea branch OIDs match;
- persistence restart recovery passes;
- healthcheck and rollback pass in pre-production;
- no plaintext secret enters Git;
- owner approval is recorded for merge, tag and production deployment.
@@ -0,0 +1,39 @@
# ADR-001: AIOS 作为唯一控制面
- Status: Accepted
- Date: 2026-07-20
## Context
现有系统包含多个执行和调度能力,若每个 Runtime 保存自己的任务、租户和策略状态,会形成多个事实源。
## Decision
AIOS Control Plane 是唯一权威控制面,负责:
- Tenant
- Agent
- Task
- Policy
- Approval
- Audit
- Execution Evidence
OpenClaw、Hermes、Agent Router 只能执行或路由,不拥有第二套控制状态。
## Consequences
Positive:
- 状态唯一;
- 审计完整;
- Web/API 统一。
Negative:
- Runtime 接入需要 Adapter。
- 迁移需要阶段性边界收敛。
## Compliance
任何新 Runtime 必须通过 AIOS Orchestrator 接入。
@@ -0,0 +1,46 @@
# ADR-002: OpenClaw/Hermes 与控制面解耦
- Status: Accepted
- Date: 2026-07-20
## Context
OpenClaw 和 Hermes 已具备执行能力,但历史上可能承担部分控制职责。
## Decision
引入边界层:
```text
AIOS Control Plane
|
v
Adapter / Boundary
|
v
Runtime
```
OpenClaw:
- 执行签名命令;
- 返回结果;
- 提供健康状态。
Hermes:
- 队列;
- Worker;
- Lease;
- Heartbeat;
- 执行结果。
## Consequences
- Runtime 可独立演进。
- 控制逻辑不会分裂。
- 需要维护稳定 Contract。
## Migration
先接 Adapter,再迁移旧调用入口,最后关闭控制旁路。
@@ -0,0 +1,44 @@
# ADR-003: 所有变更必须通过 GitOps
- Status: Accepted
- Date: 2026-07-20
## Context
直接在生产目录、长期 worktree 或主分支上修改代码,会造成不可追踪、不可回滚和分支冲突。
## Decision
所有代码、配置、文档、测试、部署和回滚资产必须通过以下流程:
```text
non-main branch -> commit -> local sync -> Gitea sync -> test -> review -> controlled promotion
```
Controller worktree 必须长期保持 detached 且 clean。每个写任务使用独立工作区或受控写入流程;相同分支写入必须串行。
## Required evidence
每个任务必须记录:
- branch
- commit
- changed files
- test command and result
- blocker and resolution
- local/Gitea sync status
- rollback note
## Prohibited
- 直接 push `main`
- 未审批创建生产 Tag
- 在生产目录手工修改后不回写 Git
- fetch 到已被 worktree 检出的本地 branch ref
- 在 Git 中提交明文密钥
## Consequences
优点:可审计、可回滚、可并行治理。
成本:紧急修复也必须形成 patch branch 和证据。
@@ -0,0 +1,43 @@
# ADR-005: 文档与代码同步门禁
- Status: Accepted
- Date: 2026-07-20
## Context
架构、功能和部署变化如果不同步文档,会导致版本无法审计,后续 Agent 无法判断真实设计。
## Decision
所有影响以下内容的变更必须同步 `docs/aios-v2/`:
- 用户行为;
- API;
- 数据模型;
- 状态机;
- Runtime Boundary;
- 安全策略;
- 部署流程;
- 测试门禁。
## Gate
验收必须检查:
1. 文档引用代码路径存在;
2. API 文档有实现映射;
3. Contract 字段有模型映射;
4. Traceability Matrix 已更新;
5. ADR 与架构变化一致。
## Consequences
优点:
- 架构长期可维护;
- Agent 执行有明确上下文;
- Bug 修复有历史依据。
成本:
- 小功能也需要更新对应文档。
+39
View File
@@ -0,0 +1,39 @@
# AIOS V2 Architecture Decision Records
本目录记录影响多个模块、长期演进、兼容性、安全或交付流程的关键架构决策。
## ADR 列表
| ADR | 决策 |
|---|---|
| [ADR-001](ADR-001-single-control-plane.md) | AIOS 作为唯一控制面 |
| [ADR-002](ADR-002-openclaw-hermes-decoupling.md) | OpenClaw/Hermes 与控制面解耦 |
| [ADR-003](ADR-003-gitops-only-change-flow.md) | 所有变更必须通过 GitOps |
| [ADR-004](ADR-004-multi-tenant-isolation.md) | 多租户隔离默认拒绝 |
| [ADR-005](ADR-005-document-code-sync-gate.md) | 文档与代码同步门禁 |
| [ADR-006](ADR-006-candidate-persistence-strategy.md) | 候选持久化采用可替换 SQLite Snapshot |
## ADR 状态
- `Proposed`:提出但未采纳。
- `Accepted`:当前有效决策。
- `Superseded`:被新 ADR 替代。
- `Deprecated`:保留历史但禁止新实现采用。
## 何时必须新增或更新 ADR
以下变化必须新增或更新 ADR:
1. 控制面与 Runtime 权限边界变化;
2. 多租户隔离模型变化;
3. API、数据模型或状态机出现不兼容变化;
4. 持久化、消息、密钥、审计或部署方案发生跨模块变化;
5. GitOps、发布、回滚或文档门禁变化;
6. 替换 OpenClaw、Hermes、Agent Router 或核心基础设施。
ADR 变化必须同步更新:
- `02-TARGET-ARCHITECTURE.md`
- `04-MODULE-CONTRACTS.md`
- `05-TRACEABILITY-MATRIX.md`
- 对应代码、测试和部署资产
@@ -0,0 +1,47 @@
# TASK-AIOS-CONTROL-PLANE-GATEWAY-V1
## Owner
Codex
## Goal
建立 AIOS V2 唯一控制面 Gateway v1。
## Architecture Boundary
AIOS Control Plane:
- Tenant Registry
- Agent Registry
- Task Lifecycle
- Policy Gate
- Approval
- Audit Event
不允许:
- OpenClaw 直接管理任务生命周期
- Hermes 直接暴露控制 API
- Web UI 绕过 AIOS API
## Deliverables
1. Control Plane API skeleton
2. Tenant context middleware
3. Task state machine
4. Agent registration contract
5. Audit event interface
## Acceptance
- 所有 API 携带 tenant_id
- 未授权跨租户访问拒绝
- Task create/update/cancel 可审计
- GitOps commit 可追踪
- pytest + integration test 通过
## Migration
复用:
- Agent Router
- OpenClaw adapter
- Hermes runtime
不重写执行层。
@@ -0,0 +1,9 @@
{
"type": "branch",
"branch": "integration/aios-v2r1-cto-production-candidate-20260720",
"base": "task/codex/AIOS-V2-CONTROL-PLANE-GATEWAY-V1-20260720",
"created_by": "chatgpt-gitops-bot",
"created_at": "2026-07-20T01:47:01Z",
"direct_main_push": false,
"owner_approval_required": true
}
@@ -0,0 +1,9 @@
{
"type": "branch",
"branch": "task/codex/AIOS-V2-CONTROL-PLANE-GATEWAY-V1-20260720",
"base": "main",
"created_by": "chatgpt-gitops-bot",
"created_at": "2026-07-19T17:06:04Z",
"direct_main_push": false,
"owner_approval_required": true
}
@@ -0,0 +1,9 @@
{
"type": "branch",
"branch": "task/codex/AIOS-V2R1-PRODUCTION-SUPERVISOR-20260720",
"base": "integration/aios-v2r1-cto-production-candidate-20260720",
"created_by": "chatgpt-gitops-bot",
"created_at": "2026-07-20T03:49:13Z",
"direct_main_push": false,
"owner_approval_required": true
}
@@ -0,0 +1,39 @@
apiVersion: aios.v2/v1
kind: AgentTask
metadata:
name: codex-production-supervisor-task
owner: AIOS-Control-Plane
spec:
executor: codex
role: architecture-supervisor
objective: >-
Supervise AIOS V2R1 production candidate deployment, resolve blocking issues,
review patches, and maintain architecture/document consistency.
responsibilities:
- review OpenClaw deployment evidence
- review Hermes runtime issues
- analyze failed tests and blockers
- approve or reject patch direction
- verify docs/code/traceability consistency
- require ADR updates for architecture changes
- maintain final release readiness assessment
prohibited:
- bypass GitOps process
- directly apply unreviewed production changes
- weaken tenant isolation or security gates
- approve missing evidence
required_checks:
- requirement traceability
- architecture boundary compliance
- test evidence
- secret safety
- rollback readiness
- document synchronization
evidence_required:
- claim
- heartbeat
- progress
- blocker
- review_result
- commit
- final_recommendation
@@ -0,0 +1,35 @@
apiVersion: aios.v2/v1
kind: AgentTask
metadata:
name: openclaw-production-deploy-task
owner: AIOS-Control-Plane
spec:
executor: openclaw
role: deployment-operator
objective: >-
Deploy AIOS V2R1 candidate runtime, execute health checks, integrate OpenClaw
runtime boundary, and fix low-risk runtime/deployment defects.
responsibilities:
- install systemd or Docker candidate deployment
- validate environment configuration
- execute healthcheck and smoke tests
- verify OpenClaw runtime connectivity
- fix small deployment/runtime defects through patch branches
- update ExecutionEvidence after every step
prohibited:
- modify AIOS control-plane architecture
- bypass GitOps workflow
- commit secrets
- directly edit production without a recorded change
evidence_required:
- claim
- heartbeat
- progress
- blocker
- commit
- test_status
- result
escalation:
- architecture change -> Codex supervisor
- security issue -> AIOS owner
- blocked deployment -> Codex supervisor
@@ -0,0 +1,256 @@
from __future__ import annotations
import sys
import tempfile
import unittest
from pathlib import Path
REPO_ROOT = Path(__file__).resolve().parents[2]
APP_ROOT = REPO_ROOT / "active-runtime" / "app"
if str(APP_ROOT) not in sys.path:
sys.path.insert(0, str(APP_ROOT))
from aios_hermes import HermesAdapter, HermesJobState, HermesWorker
from aios_openclaw import (
CommandState,
InProcessRuntimeTransport,
OpenClawAdapter,
)
from aios_v2 import (
ControlPlaneService,
ExecutionEvidence,
ExecutionOrchestrator,
ExecutionTracker,
RequestContext,
TaskState,
TenantIsolationError,
)
from aios_v2.persistence import DurableControlPlaneService, SQLiteSnapshotRepository
SIGNING_MATERIAL = b"candidate-test-signing-material-0001"
def context(
tenant_id: str,
principal: str,
*roles: str,
) -> RequestContext:
return RequestContext.create(
tenant_id=tenant_id,
principal_id=principal,
roles=roles,
trace_id=f"trace-{tenant_id}-{principal}",
)
def bootstrap(service: ControlPlaneService, tenant_id: str = "tenant-a"):
platform = context(tenant_id, "platform", "platform_admin")
service.register_tenant(
platform,
tenant_id=tenant_id,
display_name=tenant_id,
)
operator = context(
tenant_id,
"operator",
"tenant_admin",
"operator",
"approver",
"auditor",
"runtime",
)
return platform, operator
class CandidateFlowTests(unittest.TestCase):
def test_openclaw_full_flow_reconciles_completed_task(self) -> None:
service = ControlPlaneService()
_, operator = bootstrap(service)
agent = service.register_agent(
operator,
agent_id="openclaw-agent",
display_name="OpenClaw",
runtime="openclaw",
capabilities=("echo",),
)
task = service.create_task(
operator,
agent_id=agent.agent_id,
action="echo",
payload={"message": "hello"},
idempotency_key="openclaw-flow-1",
)
approved = service.transition_task(
operator,
task_id=task.task_id,
target_state=TaskState.APPROVED,
expected_revision=task.revision,
)
transport = InProcessRuntimeTransport(
lambda command: {
"state": CommandState.COMPLETED.value,
"output": {"echo": command.payload["message"]},
}
)
resolver = lambda tenant_id, key_id: SIGNING_MATERIAL
adapter = OpenClawAdapter(transport, resolver)
orchestrator = ExecutionOrchestrator(
service,
openclaw=adapter,
openclaw_signing_material_resolver=resolver,
)
dispatch = orchestrator.dispatch(operator, task_id=approved.task_id)
completed = service.get_task(operator, task.task_id)
self.assertEqual(dispatch.runtime, "openclaw")
self.assertEqual(dispatch.state, CommandState.COMPLETED.value)
self.assertEqual(completed.state, TaskState.COMPLETED)
self.assertEqual(completed.result, {"echo": "hello"})
def test_hermes_full_flow_reconciles_worker_result(self) -> None:
service = ControlPlaneService()
_, operator = bootstrap(service)
agent = service.register_agent(
operator,
agent_id="hermes-agent",
display_name="Hermes",
runtime="hermes",
capabilities=("transform",),
metadata={"cell_id": "cell-a", "profile_id": "profile-a"},
)
task = service.create_task(
operator,
agent_id=agent.agent_id,
action="transform",
payload={"value": 7},
idempotency_key="hermes-flow-1",
)
approved = service.transition_task(
operator,
task_id=task.task_id,
target_state=TaskState.APPROVED,
expected_revision=task.revision,
)
hermes = HermesAdapter(default_lease_seconds=10)
orchestrator = ExecutionOrchestrator(service, hermes=hermes)
dispatch = orchestrator.dispatch(operator, task_id=approved.task_id)
self.assertEqual(dispatch.runtime, "hermes")
self.assertEqual(dispatch.state, HermesJobState.QUEUED.value)
worker = HermesWorker(
hermes,
lambda job: {"transformed": int(job.payload["value"]) * 2},
worker_id="worker-a",
tenant_id="tenant-a",
cell_id="cell-a",
profile_id="profile-a",
heartbeat_seconds=0.05,
)
run = worker.run_once()
self.assertTrue(run.claimed)
self.assertEqual(run.state, HermesJobState.COMPLETED.value)
result = hermes.result(dispatch.runtime_execution_id, tenant_id="tenant-a")
completed = orchestrator.reconcile_hermes(operator, result)
self.assertEqual(completed.state, TaskState.COMPLETED)
self.assertEqual(completed.result, {"transformed": 14})
def test_cross_tenant_task_read_is_denied(self) -> None:
service = ControlPlaneService()
_, tenant_a = bootstrap(service, "tenant-a")
platform_b = context("tenant-b", "platform", "platform_admin")
service.register_tenant(
platform_b,
tenant_id="tenant-b",
display_name="Tenant B",
)
tenant_b = context("tenant-b", "auditor", "auditor")
agent = service.register_agent(
tenant_a,
agent_id="agent-a",
display_name="Agent A",
runtime="hermes",
capabilities=("work",),
)
task = service.create_task(
tenant_a,
agent_id=agent.agent_id,
action="work",
payload={},
idempotency_key="tenant-isolation-1",
)
with self.assertRaises(TenantIsolationError):
service.get_task(tenant_b, task.task_id)
def test_durable_service_recovers_state_after_restart(self) -> None:
with tempfile.TemporaryDirectory() as temporary:
path = Path(temporary) / "state.sqlite3"
service = DurableControlPlaneService(SQLiteSnapshotRepository(path))
_, operator = bootstrap(service)
agent = service.register_agent(
operator,
agent_id="persistent-agent",
display_name="Persistent",
runtime="hermes",
capabilities=("persist",),
)
task = service.create_task(
operator,
agent_id=agent.agent_id,
action="persist",
payload={"ok": True},
idempotency_key="persist-1",
)
recovered = DurableControlPlaneService(SQLiteSnapshotRepository(path))
restored_task = recovered.get_task(operator, task.task_id)
self.assertEqual(restored_task.payload, {"ok": True})
self.assertEqual(restored_task.state, TaskState.CREATED)
def test_execution_tracking_persists_claim_heartbeat_and_result(self) -> None:
with tempfile.TemporaryDirectory() as temporary:
tracker = ExecutionTracker(temporary)
claimed = tracker.claim(
ExecutionEvidence(
execution_id="exec-1",
tenant_id="tenant-a",
task_id="task-1",
agent_id="agent-1",
executor="codex",
branch="task/codex/example",
progress_total=3,
)
)
self.assertEqual(claimed.status, "claimed")
tracker.heartbeat(
tenant_id="tenant-a",
execution_id="exec-1",
status="running",
current_step="tests",
progress_completed=2,
last_commit="abc123",
)
tracker.complete(
tenant_id="tenant-a",
execution_id="exec-1",
status="complete",
test_status="passed",
test_summary={"passed": 5, "failed": 0},
)
restored = ExecutionTracker(temporary).get(
tenant_id="tenant-a",
execution_id="exec-1",
)
self.assertEqual(restored.status, "complete")
self.assertEqual(restored.progress_completed, 3)
self.assertEqual(restored.test_summary["passed"], 5)
if __name__ == "__main__":
unittest.main()