owner
6ffd7fe19c
Apply the approved Implementation Baseline 1.1 architecture revisions onto the current mainline, verified faithful to the delivered change bundle: - Add aios_baseline: signed owner-delegation-grant.v1 / operational-epoch.v1 / task-principal.v1 auth chain, runtime contract + deployment-truth evaluator, declarative operation registry (effect-class + fail-closed dispatch), lineage ledger, release/recipe/saga ledger, bounded optimization controller, authenticated HTTP transport, resident worker loop and systemd host probe. - Extend existing Job authority (aios_gitops_job_controller/executor) with durable intent/receipt/reconcile and fail-closed admission; harden task_plane/store heartbeat CAS and terminal-state fencing; make reconciliation_controller fail-closed on incomplete/drift evidence. - Add isolated-validation profile templates (deploy/baseline-v1.1, systemd .example), source change/rollback tooling (tools/baseline_v1_1), the AT-01..AT-23 acceptance matrix, M0-M5 runbook, installation/rollback docs and the baseline_v1_1 test suite. - Keep legacy write authorities untouched; isolated-validation profile only. Staged blobs verified byte-identical to the delivered source-changes bundle. Cross-platform baseline tests pass; systemd/fcntl-dependent paths require the declared Linux x86_64 + CPython 3.12 + systemd runtime (site-bound).
6.0 KiB
AIOS Implementation Baseline 1.1 Implementation Plan
For agentic workers: use the approved user architecture and the task interfaces below. Implement continuously, with failure-first tests and scoped review. No production access is part of this package-build task.
Goal: deliver source changes, runnable integration/acceptance, safe migration/package tools and precise deployment instructions for the seven final architecture revisions. Architecture: extend the existing Task/Job/Release authorities through explicit adapters. RuntimeContract describes stable structure; deployment targets are dynamic Release-owned records. Historical authorization wire formats remain frozen. New Grant and OperationalEpoch are separate objects. Generic operation execution persists intent before effects and uses receiver fencing and reconciliation. Tech stack: Python 3.11+, SQLite/Git/systemd, existing cryptography and pytest; no new online services or middleware. Spec: ARCHITECTURE_BASELINE_1_1.md (generated from the approved baseline and current user's seven revisions).
Global constraints:
- The imported ZIP SHA256 is fd28f8f065873c5600f17de0a536cae6f4bc18567687fd3f178d67ad8c1f931a; local snapshot commit is an import identity, not remote main.
- Legacy global writes remain false. No arbitrary remote shell, generated production trust roots, direct production policy DB writes or implicit authority cutover.
- Tenant and subject come from authenticated service context. All durable identity links are tenant-scoped. Existing TaskPlaneStore remains the task authority for the explicit standalone profile; existing official Task API integrations use adapters and activation evidence, not silent replacement.
- Every code task: write behavioral tests, observe expected failure, implement, run covering regression, record output. Package inspection and deployment preflight are read-only by default.
Task 1: Existing Job reliability
Files: aios_gitops_job_controller.py, aios_gitops_job_executor.py, tests/baseline_v1_1/test_job_safety.py and affected existing fixtures. Deliver: finite leases/state checks, recoverable reconcile claims/epoch CAS, universal dry-run, complete operation receipt/readback. Unknown effects must not blindly repeat; DB failure latches stop further effects until explicit successful reconciliation. Preserve old campaign authorization limits. Tests: dry-run mutator never called; expired and terminal lease refused; observer crash/reordering; duplicate execution; intent commit failure; external success followed by receipt storage failure.
Task 2: Signed authorization chain
Files: aios_baseline/authorization.py; tests/baseline_v1_1/test_authorization.py. Deliver: signed owner-delegation-grant.v1, new operational-epoch.v1 and task-principal.v1 with root grant digest/generation, persistent grant/revocation state, expiry/UNTIL_REVOKED, narrowing, Ed25519 verified trusted keys. Preserve old v1/v2 modules. Tests: signature tampering, fake parent Mapping, cross-tenant/subject/task/scope, revoked generation, expired proof auto-refresh under active grant, restart persistence.
Task 3: Truth and operation contracts
Files: aios_baseline/runtime_contract.py, aios_baseline/operations.py, reconciliation_controller.py; tests/baseline_v1_1/test_contracts.py. Public interfaces: validate_contract(document)->dict; evaluate_deployment(contract,target,observation,provenance,now)->dict; OperationDefinition.from_dict(data); OperationRegistry(definitions).get(operation_id). Deliver: structural contract rejects dynamic desired; dynamic target/observation/provenance validation and UNKNOWN, effect class plus orthogonal readback/retry/reversibility policy. No arbitrary command execution in registry.
Task 4: Task loop and identity linkage
Files: aios_baseline/lineage.py, service.py, worker.py, http_api.py, host.py; existing task_plane through adapters; tests/baseline_v1_1/test_loop.py. Deliver: use TaskPlaneStore for tasks/claims; authenticate tenant-scoped thin Action; persistent correlation graph; registry-based fallback eligibility at ready_at+600; bounded subprocess host-status operation; cancellation and evidence; no special main-agent branch. Worker heartbeat/transport are real. Tests: authenticated HTTP submit->claim->real local diagnostic->receipt->result; tenant denial; duplicate submission; restart and fallback; lineage referential constraints.
Task 5: Release and optimization
Files: aios_baseline/release.py, optimization.py; tests/baseline_v1_1/test_release_optimization.py. Deliver: Release-owned targets, deployment/acceptance IDs and source lineage, recipe/receipt adapters to governed operations; bounded OptimizationEnvelope, proposal/task-only execution, objective quality floors. No direct policy mutation. Tests: acceptance wrong deployment rejected; restartable saga; stale target CAS; forbidden optimization; quality regression stops promotion.
Task 6: Packaging, activation and acceptance
Files: scripts/baseline_v1_1/, deploy/baseline_v1_1/, docs/implementation-baseline-1.1/*; tests/baseline_v1_1/test_package_tools.py. Deliver: manifest with before/after hashes, patch/full source packages, conflict-safe stage/apply/rollback, authority and credential preflight, sanitized config examples, Linux systemd profile, full AT-01..23 matrix with actual statuses. AT-22 uses subprocess crash plus durable separate effect sink; physical power failure is marked online-required. AT-23 exercises SQLite and receipt failures with fail-closed admission. Tests: clean staging, wrong baseline/tamper refusal, repeated apply, rollback conflict refusal, credential path validation; run integrated regression and public-interface suites.
Rulings:
- Effect class is primary dispatch metadata, with independent retry/readback/rollback properties, since categories overlap.
- Historical owner-delegation-epoch.v2 is not relabeled or reinterpreted. New proof schema refers to the new Grant; migration verifies original bytes.
- Live activation requires observed official authority bindings and valid installed signing identities. Packaging does not claim production activation.