[ops] feat: consolidate AIOS v20 architecture

[arena-ai-coding-agent]

Summary

• generate coordination/AGENTS_STATUS.md deterministically from session journals and claims
• detect stale/finished claim inconsistencies and support --check
• integrate and harden AIOS v20 identity → trust → policy → audit foundation
• add bounded runtime lifecycle, monotonic heartbeat TTL, and fail-closed action budgets
• align all CodeQL steps on one SHA-pinned v4 action
• document CI baseline and OpenHands Cloud V1 review findings

Validation

• pytest --noconftest tests/test_generate_agents_status.py tests/kernel -q — 15 passed
• ruff check scripts/generate_agents_status.py tests/test_generate_agents_status.py aios_core/kernel aios_core/runtime tests/kernel
• python scripts/generate_agents_status.py --check
• python scripts/check_module_size_budget.py --strict — 0 contract errors
• python scripts/verify_supply_chain_pins.py — 0 findings
• CodeQL workflow parsed and all three action refs verified equal

Known blockers / follow-ups

• Docker baseline: Trivy fails on pinned prom/alertmanager:v0.33.1; v0.34.0 exists, but registry TLS was unavailable in the sandbox, so no unverified digest or ignore was committed.
• OpenHands PR openhands: strengthen role agent prompts #243 remains under another ACTIVE claim. Isolated review found TaskStatus.QA collection failure and incorrect Cloud V1 start-task/conversation ID mapping; details posted on that PR.
• Full repository pytest was not run locally because this sandbox lacks the full production dependency environment; GitHub CI is expected to provide the complete matrix.

[arena-ai-coding-agent]

CI follow-up on dependency closure (41525538, inventory a693f6b9):

• CodeQL v4: PASS.
• Core gate / targeted tests / secrets / supply-chain / docs / Android: PASS.
• Validation now passes dependency installation and generated inventory, and reaches the full test run. The former 8 collection errors (pandas, ccxt) are closed.
• Local Python 3.11 confirmation: full pytest --collect-only rc=0; the 8 previously blocked test files run 36/36 PASS; dependency contract has 0 errors; pip check clean.

Remaining full-suite baseline after successful collection: 15 failures + 19 fixture errors, grouped as hard-coded /root/AIOS paths in a non-root runner, live Kraken/TLS calls, benchmark fixture/plugin gaps, and missing legacy top-level modules. These are unrelated to the dependency/kernel diff and need separate hermeticity fixes by scope.

Other known external checks: repository-wide ruff still points only to aios_core/accounting_reporter.py (PR #240); Alertmanager Trivy pin remains the documented v0.33.1 baseline.

[arena-ai-coding-agent]

Architecture consolidation is now in head 4bacaf96: aios_core/kernel (PDP), runtime enforcement, execution (PEP), supervisor, and canonical aios_core/architecture/ArchitectureRuntime. 36 architecture tests pass; canonical map: docs/AIOS_V20_ARCHITECTURE.md.

[arena-ai-coding-agent]

Architecture stage 2 landed in cb0d536: one-shot ApprovalGate for configured high-risk capabilities plus persistent hash-chained ArchitectureAuditStore with task_id:action_id correlation. Gate order is PDP → approval → lifecycle/heartbeat/budget → PEP. Replay/rejection fail closed; tampering is detected. Architecture regression: 40 passed, ruff/compile clean.

[arena-ai-coding-agent]

Architecture stage 3 landed in da975fe: SupervisorRuntimeExecutor maps each selected role to a scoped SpecialistInvocation and calls only ArchitectureRuntime.execute. No supervisor path receives CapabilityEngine directly. Parallel observations are lock-protected; policy denial and missing invocation stop dependent graph batches fail-closed. Architecture regression: 43 passed.

[arena-ai-coding-agent]

Architecture stage 4 landed in 536a725: supervisor specialist calls now require owner-attributed, issuer-recorded, task/role/agent/capability-scoped delegation grants with timezone-aware expiry and revocation. Validation occurs immediately before governed Action creation. Expiry, revocation, task mismatch, and capability escalation fail closed. Architecture regression: 45 passed.

[arena-ai-coding-agent]

10-iteration parallel security batch landed in 9337609: (1) canonical HMAC signer, (2) authenticated approval commands, (3) timestamp+nonce replay guard, (4) signed delegation artifacts, (5) scope-attenuating delegation chains, (6) idempotency ledger, (7) capability owner/risk registry, (8) audit query/export, (9) architecture health snapshot, (10) integrated security profile. Unified architecture regression: 48 passed; ruff/compile/size budget clean.

[arena-ai-coding-agent]

Governed OpenHands boundary landed in 78c1683. openhands_cloud_run is now an ArchitectureRuntime capability: Policy → explicit Approval → lifecycle/heartbeat/budget → ExecutionKernel → ContourService → hash-audit. FakeContour tests prove policy deny and pending approval produce 0 Cloud calls; approved retry with the same action ID produces exactly 1 call. The legacy HTTP router is documented as non-production until migrated. Real Cloud was not invoked: OPENHANDS_CLOUD_API_KEY is absent and paid execution still requires explicit consent.