# AEH Core Architecture v2 Engineering Ledger

This ledger records evidence-backed findings for the external Codex-led architecture campaign. It is not an AEH OperationRecord and is intentionally maintained outside the candidate workflow.

Findings, commands, counts, and status labels below are historical at their recorded checkpoints. They do not claim current certification unless a separately dated verification says so. For the f1518caf405f44851227cf73ea37b79267b1a098 read-only audit snapshot, consult the status and conformance documents; that audit did not rerun the evidence.

## Findings

| ID | Severity | Component | Symptom | Root cause | Files | Regression | Fix | Deterministic verification | Real E2E verification | Status |
| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
| AEH-V2-0001 | P1 | Build/runtime isolation | A detached operation launched from the development checkout temporarily made `dist/main.js` unavailable; a concurrent CLI status invocation failed with `MODULE_NOT_FOUND`. The accidental audit later failed when its Paseo transport disconnected. | The self-hosted detached operation and its build lifecycle shared the development checkout's mutable `dist` output. `npm run build` removed `dist` before recompiling, so concurrent control-plane invocations had no stable executable boundary. | `package.json`, `scripts/build.mjs`, `tests/buildHygiene.test.ts` | `tests/buildHygiene.test.ts` runs a real build while checking the entrypoint, CLI probes and retained release set. | `npm run build` compiles in a unique staging directory, publishes a complete immutable release behind an atomic `dist/current` pointer and atomic launcher, serializes publishers with a lock, and retains only the current plus two exact release directories. | `npm run typecheck && npm run build`; `tests/buildHygiene.test.ts` — 3 passed; packed dry-run — 2,146 files, 0.94 MB compressed, 9.60 MB unpacked. | Not run; prohibited on the development checkout. | FIXED_AND_RECERTIFIED (deterministic) |
| AEH-V2-0002 | P1 | Operation cancellation/lifecycle | Cancelling a detached operation with a controller PID and no registered participants exceeded the 10-second regression timeout. | Cleanup attempted a live Paseo agent discovery even when the temporary operation root had no project configuration and no evidence that Paseo was configured; the discovery command could consume its full 30-second timeout. | `src/operations/controller.ts`, `tests/operations.test.ts` | Existing detached-descendant cancellation regression now completes within its bounded test timeout. | Paseo compatibility discovery is gated on a configured Paseo orchestration provider; process-group/descendant cleanup remains unconditional. | `npx vitest run tests/operations.test.ts` — 16 passed. | Not required for this deterministic lifecycle fix. | FIXED_AND_RECERTIFIED |
| AEH-V2-0003 | P0 | Architecture contract/lifecycle boundary | The pre-v2 surface represented implementation as a binary Quick/SPEC decision and had no shared candidate, receipt, authority, runtime ownership, or project identity contract. | Architecture intent, implementation route, assurance, candidate identity, runtime terminal evidence, capability leases, context addressing, and project registration were spread across incompatible legacy shapes or absent. | `src/architecture/`, `src/agents/routingV2.ts`, `src/operations/v2Contracts.ts`, `src/operations/state.ts`, `src/security/`, `src/runtime/`, `src/context/runtimeV2.ts`, `src/projects/`, `src/control-center/` | New deterministic contract, lifecycle, routing, runtime, authority, context, registry, Control Center and certification-matrix tests cover the v2 foundation. | Added explicit four-way implementation routing plus independent assurance; candidate-bound operation receipts and success terminal gate; scoped runtime/Serena ownership; external human decision ledger; registry-backed Home and loopback token/CSRF Control Center. | `npm run typecheck`; focused v2 suites; non-system Vitest — 133 files passed, 551 tests passed, 7 skipped. | Not run against the development checkout; later candidate certification remains isolated. | FOUNDATION_IMPLEMENTED_AND_DETERMINISTICALLY_VERIFIED |
| AEH-V2-0004 | P1 | Packaged candidate boundary | Source-only tests can miss missing packaged assets or entrypoint wiring. | The architecture campaign requires the candidate artifact to be exercised from a clean consumer boundary, separate from the development checkout. | `package.json`, `dist/`, `src/entry.ts`, `src/control-center/`, `src/projects/` | Fresh temporary packed consumer was extracted and run independently. | No source checkout workflow was used; candidate was packed, extracted, linked only to the existing dependency tree, and exercised through packaged `init`, deterministic `start`, `home --once`, `control-center --once`, project registration/listing and two-project isolation. | Latest packed candidate `0.8.4` smoke passed: deterministic user turn completed; two registered projects received distinct identities; Home and Control Center bound loopback ports; temporary fixtures were cleaned and no child AEH/Paseo processes remained. | Deterministic packed-consumer smoke PASS. Real external-model/provider lanes remain explicitly unclaimed. | PACKAGED_DETERMINISTIC_SMOKE_PASS |
| AEH-V2-0005 | P0 | Candidate terminal truth | A successful operation could be accepted without candidate-bound receipts when the controller had no participant evidence; detached startup also did not bind the candidate revision before terminal evaluation. | The v2 terminal gate was present but optional when no candidate was supplied, and legacy operation bootstrap records were not normalized to a candidate identity. | `src/operations/state.ts`, `src/operations/controller.ts`, `src/workers/agentPrompt.ts`, `tests/operationV2Identity.test.ts`, `tests/operationLifecycleRegression.test.ts` | Candidate identity, revision binding, receipt provenance and success-terminal rejection are covered by deterministic lifecycle tests. | Operation records normalize a compatibility candidate, detached controllers bind the current revision, and successful terminal transitions require matching candidate-bound receipts/artifacts. | `tests/operationV2Identity.test.ts`, `tests/operationLifecycleRegression.test.ts` — PASS. | Real model/provider evidence not claimed. | FIXED_AND_RECERTIFIED (deterministic) |
| AEH-V2-0006 | P1 | Routing / feature handoff | Canonical route and assurance metadata could be treated as advisory while compatibility `mode` selected execution; delegated work had no compact normative handoff, and durable escalation could leave formal contracts carrying the old route. | Production selection, review, planning and evidence paths still keyed important decisions to Quick/SPEC compatibility or ignored route metadata. | `src/agents/routing.ts`, `src/core/run.ts`, `src/operations/change.ts`, `src/architecture/featureCapsule.ts`, `tests/routing.test.ts`, `tests/architectureV2Contracts.test.ts` | Route/assurance selection, delegated capsule validation and route-aware review/acceptance tests pass. | Canonical four-way routes and assurance drive selection and gates; delegated changes persist a validated FeatureCapsule; durable escalation rewrites route evidence and formal contract metadata. Removed the obsolete Quick/mode surface from active source, templates and tests. | Non-system Vitest — 133 files passed, 551 tests passed, 7 skipped. | Real delegated/formal model journey remains pending in an isolated candidate fixture. | FIXED_AND_RECERTIFIED (deterministic) |
| AEH-V2-0007 | P1 | Serena runtime isolation | Per-agent Serena launch paths could create duplicated processes and a writer projection could expose raw editing capability outside an explicit lease. | Production launch/configuration did not use a shared headless broker; projection and lease enforcement were not centralized. | `src/providers/serenaProxy.ts`, `src/runtime/serenaPool.ts`, `src/agents/permissions.ts`, `src/paseo/launchSpec.ts`, `tests/serenaPool.test.ts` | Pool identity, reader/writer conflict and projection tests pass. | A shared Unix-socket Serena broker starts one headless server, multiplexes readers, enforces one writer lease, hides edit tools from readers, and returns only the projected MCP surface. | Actual Serena 1.6.1 pooled MCP smoke passed with two clients, nine projected tools each and edit tools hidden from the reader; no child processes remained. | Supported-provider model execution is not claimed; Paseo executable is unavailable in this environment. | FIXED_AND_ISOLATED_PROVIDER_SMOKE_PASS |
| AEH-V2-0008 | P1 | Runtime / Control Center authority | Runtime state writes from independent owners could overwrite each other; Control Center decisions lacked a project/candidate-bound callback; normal start did not materialize the local authority surface. | Runtime persistence was last-writer-wins and the Control Center UI/server boundary had no durable decision integration. | `src/runtime/managed.ts`, `src/entry.ts`, `src/control-center/decision.ts`, `src/control-center/server.ts`, `tests/runtimeSupervisorV2.test.ts`, `tests/controlCenter.test.ts` | Independent runtime-owner merge and decision callback tests pass; packed loopback startup smoke passes. | Runtime snapshots merge under a file lock, normal non-deterministic start launches a tokenized loopback Control Center, and decision callbacks validate operation/candidate scope before writing the human decision ledger. | Non-system Vitest — PASS; packed consumer and direct loopback Control Center smoke — PASS. Browser automation was unavailable, so browser visual QA is not claimed. | Paseo/human-facing journey remains pending in an isolated candidate fixture. | FIXED_AND_RECERTIFIED (deterministic + loopback) |
| AEH-V2-0009 | P1 | Certification evidence boundary | Capability rows with `requiredEvidence` could report PASS without enforcing that evidence, allowing configuration-only checks to masquerade as model/provider certification. | Certification evaluated checks and overall status but did not resolve required evidence tokens against oracle evidence and passing checks. | `src/certification/core.ts`, `tests/certification.test.ts`, `tests/certificationMatrix.test.ts` | Missing required evidence now yields `INSUFFICIENT_EVIDENCE`/non-PASS capability results. | Certification resolves required evidence explicitly and separates deterministic contract PASS from model/provider PASS. | Certification suites and non-system Vitest — PASS. | Real model/provider lanes remain unclaimed until isolated candidate certification. | FIXED_AND_RECERTIFIED (deterministic) |
| AEH-V2-0010 | P0 | Participant terminal authority | A syntactically valid receipt from an unregistered participant could be recorded and contribute to terminal success. | Receipt recording validated the receipt/candidate shape but did not perform the registration check inside the operation-state lock. | `src/operations/state.ts`, `tests/operationV2Identity.test.ts` | Unknown participant receipt rejection is covered by the operation identity suite. | Receipt recording now rejects participants absent from the current operation while holding the state lock, so a concurrent registration race cannot bypass the gate. | Non-system Vitest — 133 files passed, 551 tests passed, 7 skipped. | Real participant runtime evidence remains pending in isolated candidate certification. | FIXED_AND_RECERTIFIED (deterministic) |
| AEH-V2-0011 | P0 | Context reference authority | A caller could self-authorize a ContextRef by supplying operation/project/candidate fields without a controller-issued grant; authorization could also issue refs from a self-consistent durable envelope that was not represented in the bound launch manifest. | ContextRef creation treated caller-provided scope as sufficient authorization and did not validate a controller-issued durable receipt. Grant issuance reread the durable ContextEnvelope without checking its verified digest or exact `(refId, artifactPath, sourceDigest)` set against the frozen S1 ContextManifest. | `src/context/runtimeV2.ts`, `src/context/authorizationV2.ts`, `tests/contextRuntimeV2.test.ts`, `tests/contextAuthorizationV2.test.ts` | Self-authorization, stale/forged scope, expiry, source drift, wrong participant/session, replay, and durable budget checks pass. Controller issue requires current S3 ownership plus a current participant read lease and rejects launch/envelope digest or exact `(refId, artifactPath, sourceDigest)` set mismatches. | Strict ContextRef authorization binds operation/revision, project, candidate, participant/generation, blueprint, policy, controller epoch, actual session, exact permitted fragment/source identity, expiry, budget, and final S1 ExecutionBinding receipt. Before constructing grants, the issuer checks the immutable launch ContextManifest digest against that binding, its envelopeDigest against verified envelope provenance, and exact unique addressable-ref tuple equality; source bytes are revalidated. Digest integrity alone is not treated as proof of controller origin. | S5 remediation: authorization red/green 1 file (4 failed / 4 passed before fix; 8 passed after); focused context/session/provenance 12 files / 125 tests; full non-system 152 files, 835 passed / 7 skipped; TypeScript/build passed; relevant deterministic system suites 3 files / 45 tests passed. Campaign Director independently reran the focused, full non-system, build/type, and relevant system gates and accepted S5 deterministic closure. | Provider/model context journey remains pending; no REAL_PROVIDER claim. | FIXED_AND_RECERTIFIED (S5 deterministic); REAL_PROVIDER pending |
| AEH-V2-0012 | P1 | Candidate identity propagation | A managed validation report could be persisted without the candidate revision, leaving review, quality, evidence and delivery artifacts unable to prove they refer to the same source. | Candidate identity existed in operation state but was not projected into ValidationReport or downstream artifacts. | `src/core/verify.ts`, `src/core/types.ts`, `src/agents/reviewLifecycle.ts`, `src/agents/qualityConvergence.ts`, `src/evidence/graph.ts`, `src/delivery/finalize.ts`, `tests/candidateIdentity.test.ts` | Managed verification now carries the current candidate revision; quality state, evidence graph and delivery result carry its digest/identity. Remediation rejects candidate drift. | Non-system Vitest — 133 files passed, 551 tests passed, 7 skipped. | Real candidate mutation/review/delivery journey remains pending in isolated candidate certification. | FIXED_AND_RECERTIFIED (deterministic) |
| AEH-V2-0013 | P0 | Execution authority enforcement | Direct, Podman, Paseo and distributed worker adapters could launch without a capability lease tied to a managed operation and candidate. | Authority preparation existed only on selected prompt paths and was optional for concrete worker executors. | `src/security/executionLease.ts`, `src/workers/direct.ts`, `src/workers/podman.ts`, `src/workers/paseo.ts`, `src/distributed/worker.ts`, `tests/workerExecutionAuthority.test.ts` | Unmanaged launches are rejected before spawn; managed direct launches propagate participant, candidate digest and capability leases; distributed jobs require and revalidate the authority envelope. | All concrete worker adapters now fail closed without managed candidate authority and persist authority on WorkerSession. | `npx tsc --noEmit`; non-system Vitest — 133 files passed, 551 tests passed, 7 skipped; focused authority suite — 4 passed. | Real provider/model launch remains pending in isolated candidate certification. | FIXED_AND_RECERTIFIED (deterministic) |
| AEH-V2-0014 | P1 | Participant plan runtime binding | The compiler produced a deterministic ParticipantPlan/ExecutionBlueprint, but wave execution still launched work units without consuming compiled participant identity or current candidate revision. | Blueprint metadata was recorded for telemetry only; executor selection remained detached from assignment identities. | `src/architecture/participantPlan.ts`, `src/agents/waveExecutor.ts`, `src/distributed/worker.ts`, `tests/architectureWorkGraph.test.ts` | Each wave resolves work units through the compiled assignment map; launch authority receives the compiled participant ID and blueprint records the managed candidate revision. | Runtime roles/tool packs remain compiled from canonical role profiles; model output cannot choose concrete agents or widen scope. | `npx tsc --noEmit`; non-system Vitest — 133 files passed, 551 tests passed, 7 skipped. | Real multi-worker participant journey remains pending in isolated candidate certification. | FIXED_AND_RECERTIFIED (deterministic) |
| AEH-V2-0015 | P1 | Participant architecture / knowledge / stack | Runtime participant identity, skills, tools, stack detection and knowledge sufficiency were previously distributed across ad hoc topology and provider paths. | There was no deterministic compiler boundary for canonical roles, capability packs, stack detector precedence, or knowledge-gap status. | `src/participants/`, `src/knowledge/index.ts`, `tests/participantsCanonical.test.ts`, `tests/participantsStackDiscovery.test.ts`, `tests/knowledgeGate.test.ts` | Canonical role/skill/tool catalogs, registry-backed stack detection and explicit KnowledgeGate statuses are covered by deterministic tests. | Added participant compiler foundations, skill/tool registries, StackDetectorRegistry and KnowledgeGate/Librarian status model; model output cannot mint runtime roles or skills. | Non-system Vitest — 133 files passed, 551 tests passed, 7 skipped. | Real external knowledge discovery remains pending in isolated candidate certification. | FIXED_AND_RECERTIFIED (deterministic) |
| AEH-V2-0016 | P1 | Candidate assembly / impact | Worker patches could be applied without one controller-owned change-set and without candidate-bound impact interpretation. | Patch application and candidate revision advancement were centralized at the wave boundary, but impact categories and review dimensions were still inferred from path heuristics. | `src/candidates/assembler.ts`, `src/agents/waveExecutor.ts`, `tests/candidateAssembler.test.ts` | Stale operation/participant/candidate bindings, patch digest drift and scope escapes are rejected; bound semantic impact claims are checked against every observed changed path; absent semantics fail closed to independent review. | Added ChangeSet, CandidateAssembler and candidate-bound wave assembly. Candidate impact now accepts only a typed `MODEL`/`HYBRID` assessment bound to changed-file evidence; no assessment yields `BLOCKED` impact and `requiresIndependentReview=true`. | Focused candidate suite — 3 passed; full non-system Vitest and `npx tsc --noEmit` are rerun after the current campaign slice. | Real multi-participant mutation/review journey remains pending in isolated candidate certification. | FIXED_AND_RECERTIFIED (deterministic boundary; provider integration pending) |
| AEH-V2-0017 | P1 | Control Center API / Paseo boundary | The UI transport exposed broad untyped data and had no bounded AEH-owned Paseo integration boundary. | Resource projections, event IDs, knowledge statuses and Paseo interactions were coupled to inline server details or left implicit. | `src/control-center/contracts.ts`, `src/control-center/server.ts`, `src/paseo/gateway.ts`, `ui/control-center/`, `tests/controlCenterApiContracts.test.ts`, `tests/paseoGateway.test.ts` | Authenticated resource projections, operation detail, event history, degraded gateway state and typed knowledge statuses pass deterministic tests. | Added versioned Control Center projections, one separately buildable React Home/Project frontend, and bounded PaseoGateway methods for lead conversation, participant timelines and capability preflight. | Non-system Vitest — 133 files passed, 551 tests passed, 7 skipped; `npx tsc --noEmit`; isolated temporary-directory frontend `npm install` + `npm run build` PASS. | Browser/UI visual QA and real Paseo journeys remain pending in isolated candidate certification. | FIXED_AND_RECERTIFIED (deterministic) |
| AEH-V2-0018 | P1 | Decision mechanism / semantic reasoning boundary | Heuristic routing, stack and issue interpretation can grow into an implicit expert system, while small semantic judgments risk being materialized as unnecessary participants. | Core-v2 had no shared typed contract for bounded model assessments, profile capability resolution, evidence-bound caching or non-authoritative provenance. | `AGENTS.md`, `src/semantic/assessment.ts`, `src/agents/routingV2.ts`, `src/core/triage.ts`, `src/issues/intake.ts`, `src/core/errors.ts`, `tests/semanticAssessment.test.ts`, `tests/issueIntake.test.ts`, `tests/implementationRoutingV2.test.ts`, `tests/triage.test.ts`, `tests/changeRouting.test.ts`, `docs/CORE_ARCHITECTURE_V2_CONFORMANCE.md` | Cheapest certified profile selection, evidence-reference validation, cache invalidation, escalation and authority-denial tests pass; route and assurance remain independently gated; issue normalization blocks without semantic planning and preserves raw evidence. | Added `SemanticAssessmentServiceV1` as an injected, bounded, read-only capability with typed assessment requirements, capability profiles, deterministic profile resolution, cache/provenance digests and telemetry hooks. Route assessment can contribute bounded meaning without making security/data/public-contract work automatically formal; typed `formalizationNeed` replaces prose-only escalation. Issue intake now accepts only validated planner/semantic output and derives triage facts from explicit labels/path evidence rather than issue prose. | Focused semantic/issue/routing suites; full non-system Vitest — 133 files passed, 564 tests passed, 7 skipped; `npx tsc --noEmit`; `npm run build`. | Real profile/provider evaluation remains pending and must run only in isolated packed candidates. | FIXED_AND_RECERTIFIED (deterministic boundary; provider integration pending) |
| AEH-V2-0019 | P1 | Frozen execution-selection boundary | Operation supervision, escalation and recovery could re-derive execution policy from a topology object inside lifecycle helpers. | `operations/supervisor.ts`, `agents/escalation.ts` and `agents/recovery.ts` accepted topology where caller-frozen selections or recovery policy were sufficient. | `src/operations/supervisor.ts`, `src/agents/escalation.ts`, `src/agents/recovery.ts`, `src/core/run.ts`, `src/operations/change.ts`, `src/audit/run.ts`, `src/agents/reviewLifecycle.ts`, `tests/supervisorBootstrapRegression.test.ts`, `tests/escalation.test.ts`, `tests/recovery.test.ts` | Supervisor, stage-selection and recovery tests pass with missing/invalid boundaries failing closed; lifecycle helpers no longer import `ResolvedAgentTopology`. | Core/change/audit resolve selections and recovery policy once; escalation consumes frozen stage/role/model selections; recovery consumes `RecoveryMap`; topology remains only at the outer resolution boundary. | Focused supervisor/escalation/recovery suite — 15 passed; `npx tsc --noEmit`. | Real supervisor/provider journey remains pending in isolated packed candidates. Reviewer discovery still uses topology at its outer caller boundary by design. | FIXED_AND_RECERTIFIED (deterministic boundary) |
| AEH-V2-0020 | P1 | Model-first issue and stack interpretation | Deterministic issue prose parsing and detector-first stack discovery could silently manufacture semantic scope, requirements or stack claims. | Issue normalization still had a deterministic plan fallback, while project discovery implicitly treated detector output as interpretation. | `src/issues/intake.ts`, `src/participants/stack.ts`, `src/core/errors.ts`, `tests/issueIntake.test.ts`, `tests/participantsStackDiscovery.test.ts` | Missing issue semantic planning blocks with a preserved raw snapshot; invalid issue assessment evidence is rejected; detector-disabled stack discovery returns `BLOCKED`; a digest-bound model assessment is accepted; explicit detector fast-path cache behavior remains covered. | Removed deterministic issue-contract fabrication and added typed normalization errors. Added objective stack evidence collection, validated `MODEL`/`HYBRID` stack assessments, evidence/assessment digests, unknown/blocked status and opt-in detector fast path. Inspection now exposes only frozen issue and explicit evidence; route selection waits for semantic normalization. | Focused issue/stack suites — 12 passed; full non-system Vitest — 133 files passed, 564 tests passed, 7 skipped; `npx tsc --noEmit`; `npm run build`. | Real semantic model execution and supported-provider stack journeys remain pending in isolated packed candidates. | FIXED_AND_RECERTIFIED (deterministic boundary; provider integration pending) |
| AEH-V2-0021 | P1 | Candidate impact / failure recovery semantics | Candidate impact and recovery classification used path/output heuristics as if they were semantic truth. | Candidate impact derived security/schema/API review dimensions from filenames; failure recovery classified provider/tool output with unbound regex interpretation. | `src/candidates/assembler.ts`, `src/agents/waveExecutor.ts`, `src/agents/recovery.ts`, `src/core/errors.ts`, `tests/candidateAssembler.test.ts`, `tests/recovery.test.ts` | Bound candidate impact claims are checked against observed ChangeSet paths; missing impact semantics require independent review; bound failure assessments are accepted; invalid evidence/digests are rejected; unknown output remains ambiguous. | Added evidence-bound `MODEL`/`HYBRID` CandidateImpact and FailureAssessment contracts. Existing planner output is reused for candidate impact without a duplicate classifier call; deterministic filename/provider markers remain conservative fallback only. | Focused candidate/recovery suites — 6 passed; full non-system Vitest — 133 files passed, 564 tests passed, 7 skipped; `npx tsc --noEmit`; `npm run build`. | Real model failure-assessment/provider journeys remain pending in isolated packed candidates. | FIXED_AND_RECERTIFIED (deterministic boundary; provider integration pending) |
| AEH-V2-0022 | P1 | KnowledgeGate/Librarian runtime integration | The canonical Librarian role was present, but the default topology advertised the Explorer output contract, and planner-wave compilation did not invoke KnowledgeGate for unknown competencies. | An unknown technology could reach participant compilation without a public-path knowledge resolution; the wrong output contract prevented a Librarian from producing a validated KnowledgePack. | `src/agents/outputContracts.ts`, `src/agents/waveExecutor.ts`, `src/architecture/participantPlan.ts`, `src/knowledge/index.ts`, `presets/agents/default.jsonc`, `tests/knowledgeGate.test.ts`, `tests/architectureWorkGraph.test.ts`, `tests/agentTopologyLayers.test.ts` | Planner competencies now pass through KnowledgeGate; unknown competencies use a bounded canonical Librarian lookup and operation-local pack digest; absent trusted lookup returns typed `KNOWLEDGE_GAP_BLOCKED`; forged pack/candidate provenance is rejected. | Added the dedicated `knowledge-pack` output contract, corrected the Librarian preset, integrated bounded lookup into planner-wave blueprint compilation, and projected validated pack references into ParticipantPlan. | Focused gate/compiler/topology/hygiene suites — 22 passed; full non-system Vitest — 133 files passed, 564 tests passed, 7 skipped; `npx tsc --noEmit`; `npm run build`. | Provider-backed Librarian/unknown-technology journey remains pending in an isolated packed candidate. | FIXED_AND_RECERTIFIED (deterministic boundary; provider integration pending) |
| AEH-V2-0023 | P1 | Frozen run execution boundary | `runTask` still carried a resolved topology through lifecycle execution and re-derived recovery/reroute choices during remediation, allowing mutable topology-shaped state to remain an implicit runtime authority. | Route, participant selections, execution catalog and recovery policy were not returned as one frozen boundary; remediation could consult topology-derived policy after execution began. | `src/core/run.ts`, `src/architecture/executionCatalog.ts`, `tests/executionCatalog.test.ts`, `tests/operationLifecycleRegression.test.ts`, `tests/recovery.test.ts` | Typecheck, build, focused lifecycle/recovery tests and the full non-system suite pass after the boundary migration; packaged user-turn smoke also passes. | `runTask` now resolves topology once at the outer boundary and consumes a `FrozenExecutionBoundaryV1` containing route, selections, compiled `ExecutionCatalog`, `RecoveryMap` and reroute bindings. Lifecycle stages no longer receive or re-derive topology. | `npm run typecheck`; `npm run build`; non-system Vitest — 133 files passed, 564 tests passed, 7 skipped; packed human black-box user-turn smoke — 1 passed. | Real supervisor/provider journey remains pending in isolated packed candidates. | FIXED_AND_RECERTIFIED (deterministic boundary; provider integration pending) |
| AEH-V2-0024 | P1 | Unknown-stack semantic grounding | Objective evidence collection preserved an unseen repository, but the semantic stack profile could not represent an interpreted language outside the built-in fast-path vocabulary; supported semantic claims could also omit grounding refs. | `ProjectLanguageV1` was a closed detector vocabulary and the generic semantic payload permitted ungrounded supported claims. | `src/participants/stackTypes.ts`, `src/participants/stack.ts`, `src/semantic/assessment.ts`, `tests/participantsStackDiscovery.test.ts`, `tests/semanticAssessment.test.ts` | An unseen Cargo/Rust-like repository remains `BLOCKED` without interpretation, accepts a digest/evidence-bound `rust` semantic profile without adding a detector, and rejects supported claims without evidence. | Semantic stack profiles now preserve arbitrary normalized language labels while built-in detectors remain optional fast paths; semantic payload validation requires evidence for supported/conflicting claims and recommendations. | `npm run typecheck`; focused stack/semantic suites — 12 passed. | Real unknown-technology model/Librarian journey remains pending in an isolated packed candidate. | FIXED_AND_RECERTIFIED (deterministic boundary; provider integration pending) |
| AEH-V2-0025 | P1 | KnowledgeGate pack-to-skill continuity | A Librarian could return a valid sufficient KnowledgePack without the optional explicit skill candidate, leaving the planner gate marked sufficient while participant compilation still lacked an operation-local competency skill. | Pack-only lookup results were not normalized through the same high-confidence candidate derivation used by cache hits and direct pack lookups. | `src/knowledge/index.ts`, `tests/knowledgeGate.test.ts`, `tests/architectureWorkGraph.test.ts` | A pack-only Librarian result derives a trusted candidate bound to the validated pack; invalid explicit candidates remain rejected. | KnowledgeGate now normalizes supplied-or-derived candidates through one validation path before returning `SUFFICIENT`. | `npm run typecheck`; focused KnowledgeGate/compiler suites — 14 passed. | Provider-backed Librarian journey remains pending in an isolated packed candidate. | FIXED_AND_RECERTIFIED (deterministic boundary; provider integration pending) |
| AEH-V2-0026 | P1 | Knowledge provenance and competency coverage | A cached or looked-up KnowledgePack could be accepted based on cache-key shape while its digest or high-confidence claims did not cover the requested competency. | Cache writes did not validate pack integrity, and resolution did not revalidate cached packs or require one trusted claim per missing competency before deriving a skill candidate. | `src/knowledge/index.ts`, `tests/knowledgeGate.test.ts` | Poisoned cached and looked-up packs are rejected; pack-only results still derive a candidate only after the same validation path. | KnowledgeGate validates pack digests on cache insertion and resolution, requires high-confidence coverage for every requested gap, and validates explicit/derived skill candidates against the pack. | `npm run typecheck`; full non-system Vitest — 133 files passed, 568 tests passed, 7 skipped; packed scenario/adversarial/concurrency/human slices — 64 passed. | Provider-backed Librarian and real model evidence remain pending in isolated packed candidates. | FIXED_AND_RECERTIFIED (deterministic boundary; provider integration pending) |
| AEH-V2-0027 | P1 | Local project and UI isolation | Symlinked repository evidence could read outside the project, a symlinked Control Center asset could escape the UI root, a symlinked registered project could appear available, and a persisted non-loopback health URL could trigger an arbitrary outbound probe. | Local-boundary checks used lexical paths or unrestricted file/fetch behavior without enforcing realpath identity and loopback-only health targets. | `src/participants/stack.ts`, `src/projects/registry.ts`, `src/control-center/server.ts`, `tests/participantsStackDiscovery.test.ts`, `tests/projectRegistry.test.ts`, `tests/controlCenter.test.ts` | Outside-target symlinks are not read/served, redirected project roots are marked moved-or-missing, malformed frontend paths fail closed, and non-loopback health records are reported unreachable without probing. | Evidence collection skips symlink content/traversal; registry availability compares realpath identity; static frontend serving rejects symlinks and outside-root targets; health registration/probing is constrained to loopback addresses. | `npm run typecheck`; focused stack/registry/Control Center suites — 26 passed; full non-system Vitest — 133 files passed, 569 tests passed, 7 skipped; `npm run build`. | Packed loopback Control Center smoke passed; external service/network lanes remain unclaimed. | FIXED_AND_RECERTIFIED (deterministic + loopback) |
| AEH-V2-0028 | P1 | Validation need versus validation execution | Planner output represented validation as free-form gate strings, so semantic “what must be demonstrated” could not be carried into the frozen execution blueprint or deterministically resolved to an approved action. | `validationGates` was advisory metadata; no typed requirement schema or resolver connected project scripts, configured validators/providers and policy to the execution boundary. | `src/architecture/validationRequirements.ts`, `src/architecture/participantPlan.ts`, `src/agents/outputContracts.ts`, `src/agents/waveExecutor.ts`, `src/core/errors.ts`, `tests/architectureWorkGraph.test.ts`, `tests/scale-governance.test.ts` | Typed validation requirements resolve to an approved project script and block when no approved action exists; planner schemas and blueprint contracts remain aligned. | Added `ValidationRequirementV1` and `ValidationResolutionV1`; the resolver chooses only project-authoritative scripts, configured commands/validators or approved providers, with optional ToolRegistry availability checks. Unresolved requirements fail closed with `VALIDATION_REQUIREMENT_BLOCKED`; model output cannot provide commands, tools, providers or credentials. | `npm run typecheck`; focused architecture/schema suite — 26 passed; full non-system Vitest — 133 files passed, 569 tests passed, 7 skipped; `npm run build`. | Real project-specific provider resolution remains pending in isolated packed candidates. | FIXED_AND_RECERTIFIED (deterministic boundary; provider integration pending) |
| AEH-V2-0029 | P0 | Parallel wave candidate lineage | Sibling workers in one wave all observed the same base candidate, but sequential integration relabelled later ChangeSets with the revision advanced by earlier siblings, and wave worktrees were materialized from HEAD + prior patches rather than from the frozen candidate. | Wave assembly used the moving `currentCandidate` for each sibling and had no explicit wave base, no derivation record and no reconciliation outcome. | `src/candidates/wave.ts`, `src/candidates/assembler.ts`, `src/candidates/direct.ts`, `src/agents/waveExecutor.ts`, `src/distributed/worker.ts`, `tests/candidateWaveLineage.test.ts`, `tests/candidateAssembler.test.ts` | Independent siblings, clean same-file rebase, unresolvable conflict, stale base, digest mismatch, superseded wave base, deterministic ordering, retry against the advanced candidate, resource-claim conflict and post-apply assembly failure are all covered. | Added frozen `WaveBaseV1`; ChangeSets now carry `baseCandidateDigest` and are never rewritten; later siblings produce explicit `WAVE_REBASE` derived ChangeSets; unresolved siblings return typed `WAVE_RECONCILIATION_REQUIRED`; wave worktrees materialize the exact candidate state and verify its digest; failed assembly rolls the workspace back to the bound candidate. Distributed jobs carry and verify the frozen base digest. | `npm run typecheck`; `npx vitest run tests/candidateWaveLineage.test.ts` — 10 passed; full non-system Vitest — 140 files passed, 1 skipped; 675 tests passed, 8 skipped. | Real multi-participant wave journey remains pending in isolated packed candidates. | FIXED_AND_RECERTIFIED (deterministic boundary; provider integration pending) |
| AEH-V2-0030 | P1 | WorkUnit logical resource claims | Path independence was treated as execution independence, so two units could mutate the same logical resource (schema, migration sequence, lockfile, environment) concurrently. | No WorkUnit-level resource model existed; the scheduler only reasoned about scope overlap and graph proximity. | `src/architecture/workGraph.ts`, `src/agents/parallelism.ts`, `src/agents/outputContracts.ts`, `src/agents/waveExecutor.ts`, `tests/architectureWorkGraph.test.ts` | Claim validation, conflict semantics, duplicate-order detection, exclusive-write serialization, shared-read co-scheduling and ordered-sequence wave ordering are covered. | Added `SHARED_READ` / `EXCLUSIVE_WRITE` / `ORDERED_SEQUENCE` claims with deterministic conflict and ordering helpers; the scheduler serializes conflicting units and enforces ordered sequences across waves; the wave integrator re-checks claims as defense in depth. | `npm run typecheck`; `npx vitest run tests/architectureWorkGraph.test.ts` — 14 passed; full non-system Vitest — 140 files passed, 1 skipped; 675 tests passed, 8 skipped. | No distributed lock system was introduced by design; cross-host claims are not claimed. | FIXED_AND_RECERTIFIED (deterministic) |
| AEH-V2-0031 | P0 | Controller fencing | A controller process whose wall-clock lease had not expired could keep mutating an operation after another controller had taken over; controller identity was self-declared. | There was no durable monotonic controller identity, no secret controller proof, and candidate/terminal/tool-action mutations did not verify controller ownership. | `src/operations/state.ts`, `src/operations/controller.ts`, `src/security/executionLease.ts`, `src/security/toolActionGate.ts`, `src/architecture/participantPlan.ts`, `src/agents/waveExecutor.ts`, `tests/security/controllerFencing.test.ts` | Zombie candidate binds, terminal transitions, execution authority, blueprints, tool actions and receipt recording are rejected; a forged token with the right epoch is rejected; candidate revision jumps are rejected. | Added durable `controller.epoch` + a secret controller token whose digest is the only durable copy; every takeover increments the epoch and mints a new token; all controlled mutations and controller-authority tool actions verify epoch and token; execution authority, blueprints and action intents carry the epoch. | `npm run typecheck`; `npx vitest run tests/security/controllerFencing.test.ts` — 8 passed; full non-system Vitest — 140 files passed, 1 skipped; 675 tests passed, 8 skipped. | Same-user filesystem forgery of operation state remains outside the threat model; a hostile same-user process can still re-claim ownership. | FIXED_AND_RECERTIFIED (deterministic) |
| AEH-V2-0032 | P0 | ToolActionGate public enforcement and reconciliation | The deterministic tool action gate existed but did not cover all public delivery mutations, and a crash between side effect and persistence could cause an unsafe retry. | The public issue auto-handoff path ran before the controller AcceptanceOracle; workspace creation was bundled into public issue/branch setup; uncertain effects lacked a complete common intent/reconciliation path. The standalone `aeh sdd handoff` CLI also advertised public delivery but invoked a managed-operation-only handoff without managed acceptance context. | src/security/toolActionGate.ts, src/security/actionPolicy.ts, src/security/gatedAction.ts, src/security/actionReconciliation.ts, src/delivery/finalize.ts, src/delivery/handoff.ts, src/issues/workflow.ts, src/operations/controller.ts, src/core/run.ts, tests/deliveryFinalize.test.ts, tests/security/actionPolicy.test.ts, tests/security/gatedAction.test.ts, tests/security/toolActionGate.test.ts, tests/security/actionReconciliation.test.ts, src/cli.ts, tests/cliS8Handoff.test.ts | Every configured requested public effect waits for the current AcceptanceOracle and policy-required S7 gate; each sensitive action persists intent, passes ToolActionGate, and persists an ActionReceipt or safe reconciled equivalent. Controller-owned Paseo execution workspaces are local resources, not delivery effects, while preserving S3 fencing and conservative HUMAN_REQUIRED reconciliation. | S8 preserves the strict acceptance boundary: default issue intake starts a managed controller run without public handoff; controller-owned disposable Paseo workspaces are LOCAL_RESOURCE_CREATION, absent from configured external effects, but remain authority-fenced and intent/receipt/reconciliation tracked. After current acceptance and policy-required S7 verification, handoff/finalization gates public GitHub issue/branch/push/PR effects and persists receipts or safe reconciled receipts. Executor uncertainty is inspected before retry; unresolved or unknown Paseo identity remains HUMAN_REQUIRED. The normal standalone `sdd handoff` command now reports `SDD_HANDOFF_UNSUPPORTED` before config loading and cannot enter direct delivery; managed accepted finalization is unchanged. | Final focused S8 action/delivery/issue/CLI command — 14 files / 198 tests; `npx tsc --noEmit` exit 0; `npm run build` exit 0 (`release-1790322187860-3232428-0507f39b`); full non-system Vitest — 162 files passed, 1 skipped; 1,014 passed, 7 skipped; scenario/concurrency system — 2 files / 17 passed; `git diff --check` clean; normative target equals HEAD. The default issue-flow fixture stubbed GitHub and pushed only to a local bare repository. CLI regression passed 2 tests; the separate Flash baseline failed 2/2 against the stale surface and passed 2/2 against the remediation. | No real GitHub/Paseo provider certification. Paseo reconciliation remains HUMAN_REQUIRED when provider identity cannot be safely confirmed. S8 was independently accepted on its deterministic gate by the Campaign Director on 2026-09-25; no S9 work had started at acceptance. | FIXED_AND_DETERMINISTICALLY_VERIFIED (S8 lead gate; external/provider lanes pending)  |
| AEH-V2-0033 | P1 | Paseo session binding identity | Paseo session identity was inferred from provider ids without a durable binding to AEH operation/participant/blueprint identity, so a reused session could be attributed to stale semantic state. | The binding was partial and had no production runtime consumer. | `src/paseo/sessionBinding.ts`, `src/workers/agentPrompt.ts`, `tests/paseo/sessionBinding.test.ts`, `tests/agentPromptStructuredResultProvenance.test.ts` | Exact-field schema, digest/path corruption, idempotent bind, conflict/stale drift, rotation, active status, participant isolation, and concurrent rotation pass; production prompt preparation binds or validates the actual provider session before dispatch. | `PaseoSessionBindingV1` now binds full canonical operation/revision, participant/generation, candidate, blueprint, policy, context/prompt, controller epoch, and actual Paseo agent identity. A different session rotates; an identity mismatch on the same live session rejects reuse rather than relabeling it. | `npx vitest run tests/paseo/sessionBinding.test.ts` — 17 passed; full non-system Vitest — 320 files, 830 passed, 7 pending; `npx tsc -p tsconfig.json --noEmit` and `npm run build` passed. | No REAL_PROVIDER session journey or certification is claimed. | FIXED_AND_RECERTIFIED (S5 deterministic binding/runtime); provider behavior pending |
| AEH-V2-0034 | P2 | Candidate self-dependency | The development checkout declared a released copy of itself as a devDependency, so an old AEH release could be resolved while developing the candidate. | `package.json` and `package-lock.json` pinned `agentic-engineering-harness@^0.6.29`. | `package.json`, `package-lock.json`, `tests/architectureHygiene.test.ts` | A hygiene test asserts no dependency/lock entry for the package's own name. | Removed the self-dependency and regenerated the lockfile; the repo-local bin shim is provided by `scripts/link-self-bin.mjs`. | `npm install --package-lock-only --offline`; `npx vitest run tests/architectureHygiene.test.ts`. | None. | FIXED_AND_RECERTIFIED (deterministic) |
| AEH-V2-0035 | P1 | Candidate impact mechanism labelling and receipt integrity | Planner-derived impact claims were labelled `HYBRID` with no unknowns, an orphaned receipt could make the gate re-execute a non-idempotent action, and a candidate revision jump could be bound without lineage. | Impact assessment hardcoded `mechanism: "HYBRID"`; `authorizeToolAction` did not inspect an intent-less receipt; `bindOperationCandidate` allowed arbitrary revision jumps. | `src/agents/waveExecutor.ts`, `src/security/toolActionGate.ts`, `src/operations/state.ts`, `tests/security/toolActionGate.test.ts`, `tests/security/controllerFencing.test.ts` | Planner impact is now `MODEL` with explicit unknowns; an intent-less receipt yields `TOOL_ACTION_RECEIPT_ORPHANED`; non-placeholder binds must advance exactly one revision. | Labeled planner claims honestly, added orphan-receipt fail-closed behavior and operation-context enforcement for receipts, and enforced revision+1 lineage. | `npx vitest run tests/security/toolActionGate.test.ts tests/security/controllerFencing.test.ts`. | `requiresIndependentReview` is still not consumed by reviewer compilation. | FIXED_AND_RECERTIFIED (deterministic) |
| AEH-V2-0036 | P2 | Provenance scenario fixture | The packed system scenario matrix expected a hand-written provenance manifest to verify while omitting the now-required BuildIdentity, so the P10 truth-property assertion failed. | The fixture predated the BuildIdentity hardening of `verifyProvenanceManifest`. | `tests/system/aehScenarioMatrix.test.ts` | The scenario matrix now supplies a valid BuildIdentity and still proves that source tampering breaks provenance. | `npx vitest run tests/system/aehScenarioMatrix.test.ts` — 8 passed; `npm run test:system-all` rerun. | None. | FIXED_AND_RECERTIFIED (deterministic) |
| AEH-V2-0037 | P0 | Frozen execution and skill identity | Participant launches and structured results lacked one complete immutable v2 identity across policy, role ceiling, blueprint, runtime binding, actual prompt/context digests, and accepted ephemeral procedure content; individual procedure steps were not required to link pack evidence. Durable generic mutation paths also allowed stale-digest in-place mutation of frozen policy and participant binding bodies through `updateOperationMetadata` callbacks. The Director also reproduced generic patching of a future executionSemanticsDigest followed by a same-digest owner no-op, which left policy/bindings stale; a separate durable-record reproduction showed a missing operationExecutionRevision could be generically upgraded to 1. | Identity was spread across optional provenance fields, and the SkillCandidate trust boundary accepted aggregate pack/competency coverage without grounded evidence references for every procedure step. The generic participant updater could add a bound participant, while its transition check reconstructed binding identity from mutable post-callback participant maps; deletion/re-keying could alias both sides. Separately, policy and binding identity needed a pre-callback snapshot to prevent same-digest body changes from crossing persistence. The transition guard omitted executionSemanticsDigest, checked revision changes only when the mutable post-callback record already had a revision, and did not snapshot scalar identity presence/value before callbacks. | `src/architecture/executionIdentity.ts`, `src/architecture/participantPlan.ts`, `src/operations/state.ts`, `src/agents/outputContracts.ts`, `src/knowledge/index.ts`, `src/workers/agentPrompt.ts`, `src/workers/runtimeSessions.ts`, `src/workers/resultGateway.ts`, `src/distributed/worker.ts`, `src/paseo/launchSpec.ts`, `src/paseo/runtimeCore.ts`, `src/paseo/sdk.ts`, `tests/operations.test.ts`; S1 identity/provenance tests. | Full policy/role/blueprint/binding and grounded SkillManifest digests are compiled, frozen, and checked at launch/result boundaries. The expanded compiler blueprint is deeply immutable; direct OpenCode/Codex, Podman/OpenCode, Paseo, and distributed launch paths bind an actual pre-dispatch runtime session and check the returned identity. Operation state captures participant IDs and each complete canonical binding body before callbacks, then compares every ID in the pre-callback/post-callback union; generic add/replace/mutate/clear attempts fail closed. Only the validated `operation.participant.execution-bound` event may create/replace one binding, while candidate assembly, takeover, and changed execution-semantics events may only clear. The complete canonical ResolvedOperationPolicy snapshot and its explicit bind/invalidation lifecycle remain enforced. Procedure-text, trust-decision-digest, participant-binding, same-digest policy/body, stale-digest in-place `runtime.sessionId`, participant delete/re-key, and durable-file-byte-preservation regressions are asserted. The guard now snapshots semantics digest and execution-revision presence/value before callbacks; only the semantics owner can change the digest, first/identical binds preserve execution revision, and replacing bound semantics advances once while clearing frozen policy and every participant execution binding. Generic events cannot synthesize a missing revision, and old records remain unsupported at execution boundaries. | Candidate/fencing prerequisite command `npx vitest run tests/candidateIdentity.test.ts tests/security/controllerFencing.test.ts` — 2 files / 12 tests passed. Before the revision/digest fix, the Flash worker ran the operation suite against the unremediated state: 1 file failed, 4 failed / 29 passed (33), with failures in the four intended bypass groups. Post-fix `npx vitest run tests/operations.test.ts` — 1 file / 33 tests passed; focused S1 — 15 files / 165 tests passed; `npx tsc --noEmit` — exit 0; `npm run build` — exit 0; `npx vitest run --exclude 'tests/system/**'` — 149 files passed, 1 skipped; 773 tests passed, 7 skipped; `git diff --check` clean. Full-suite build-hygiene tests passed with missing-source-map notices. | No REAL_PROVIDER certification. Adapter and session behavior was captured at deterministic boundaries; provider execution remains pending. | FIXED_AND_RECERTIFIED (S1 deterministic adapter boundary; provider execution pending) |
| AEH-V2-0038 | P2 | CandidateImpact provenance boundary | Candidate assembly and wave inputs still accepted a precomputed CandidateImpact object that could bypass canonical SemanticAssessmentService validation of agent/profile, Paseo session, read receipts, policy, and evidence provenance. | A legacy optional `impactAssessment` field on internal assembly/submission APIs projected caller-supplied claims into production impact output after only structural and digest checks. | `src/candidates/assembler.ts`, `src/candidates/wave.ts`, `src/agents/waveExecutor.ts`, `tests/candidateAssembler.test.ts`, `tests/candidateWaveLineage.test.ts` | Tests verify canonical service assessment against the assembled candidate and exact file receipts, rollback on invalid changed-path judgment, and a BLOCKED independent-review minimum when no runtime is supplied. | Removed the precomputed impact input from assembly and wave APIs. Only the canonical service can produce MODEL impact; absent runtime remains explicitly BLOCKED and keeps independent review required. | S2 focused suites — 14 files / 108 tests passed; `npx tsc --noEmit` — exit 0; `npm run build` — exit 0; full non-system Vitest — 151 files passed, 1 skipped; 788 tests passed, 7 skipped. | No REAL_PROVIDER execution or certification; fakes verify deterministic contracts only. | FIXED_AND_DETERMINISTICALLY_VERIFIED (S2; S4 downstream assurance consumption remains future scope) |
| AEH-V2-0039 | P1 | STACK repository evidence and proposed path boundary | STACK receipts could be stamped with another root's binding, stale content could be scanned during drift, ignored secret/runtime files could enter evidence, and model-proposed `projectSkillRoots` could contain absolute, traversing, malformed, missing, or escaping paths. | The scanner inventoried its own filesystem root without comparing canonical root/content digests and did not use Git's tracked/non-ignored file set; projected skill-root strings were copied without generic path/existence/containment validation. | `src/core/git.ts`, `src/semantic/runtime.ts`, `src/participants/stack.ts`, `src/semantic/assessment.ts`, `tests/participantsStackDiscovery.test.ts`, `tests/semanticAssessmentRuntime.test.ts` | Regressions reject cross-root substitution, pre-capture and in-assessor drift, missing STACK root digest, ignored `.env`/`.harness/cache` references, and invalid project skill-root paths; valid in-root roots remain descriptive. | Binding creation hashes canonical realpath. STACK requires and compares that root digest plus the worktree digest before scan, after evidence capture, after assessment and before return; evidence inventory uses the worktree digest's Git file set or stricter generic exclusions. Project skill roots require normalized relative syntax, existence and realpath containment and cannot grant skills/capabilities. | Focused S2 suites — 14 files / 123 tests passed; stack/semantic runtime set — 5 files / 75 tests passed; `npx tsc --noEmit` — exit 0; `npm run build` — exit 0; full non-system Vitest — 151 files passed, 1 skipped; 803 tests passed, 7 skipped. | No REAL_PROVIDER, BROWSER, or ADVERSARIAL evidence; all checks use deterministic fixtures. | FIXED_AND_DETERMINISTICALLY_VERIFIED (S2; provider evidence pending) |
| AEH-V2-0040 | P1 | High-risk semantic INTENT capability ceiling | Every semantic INTENT request with `risk: "high"` was rejected during policy validation even though the caller explicitly supports HIGH and downstream route/assurance logic supplies conservative floors. | `semanticCapabilityPolicyV1.INTENT.maxRiskClass` was STANDARD while the public caller mapped high-risk input to HIGH. | `src/semantic/assessment.ts`, `src/audit/intent.ts`, `tests/auditIntent.test.ts` | Focused regressions assert HIGH is admitted within existing request budgets and keeps deterministic CRITICAL assurance and delegation floors. | Raised only the INTENT maximum risk ceiling to HIGH. Reasoning/context/token/time bounds and deterministic routing/assurance floors are unchanged. | Focused S2 suites — 14 files / 123 tests passed; `npx tsc --noEmit` — exit 0; `npm run build` — exit 0; full non-system Vitest — 151 files passed, 1 skipped; 803 tests passed, 7 skipped. | No REAL_PROVIDER or certification claim. | FIXED_AND_DETERMINISTICALLY_VERIFIED (S2; provider evidence pending) |
| AEH-V2-0041 | P1 | CandidateImpact assurance recompilation | Actual post-assembly impact was not consumed to compile policy floors, validation requirements, actual independent Reviewer assignments, candidate-bound acceptance assertions, or evidence strength. Impact flags alone did not create independent review evidence. | CandidateImpact was provenance-bound by S2, but downstream run, wave, repair, validation, and review lifecycles did not deterministically recompile requirements against the current CandidateRevision and frozen ResolvedOperationPolicy. | `src/architecture/candidateAssurance.ts`, `src/candidates/assembler.ts`, `src/candidates/repair.ts`, `src/core/run.ts`, `src/agents/reviewLifecycle.ts`, `tests/candidateAssurance.test.ts`, `tests/candidateAssembler.test.ts`, `tests/runRepairCandidateLifecycle.test.ts`, `docs/CORE_ARCHITECTURE_V2_S4_WORKGRAPH.md` | Regressions cover policy floors, stale/tampered impact and resolver binding, absent/disallowed required validations, high-risk unknowns, reviewer eligibility/diversity, actual Reviewer session identity, candidate assertion/evidence-strength binding, and repair lifecycle recompile. | CandidateImpact now binds candidate/base/patch facts; deterministic compiler and lifecycle consume exact current impact/policy, execute required validation, assign actual independent read-only Reviewers, persist candidate-bound assertions/evidence strength, and fail closed. No S6 AcceptanceOracle or normative target change. Delegation assessment and exact profile/worktree evidence are recorded in `docs/CORE_ARCHITECTURE_V2_S4_WORKGRAPH.md`. | Focused `npx vitest run tests/candidateAssurance.test.ts tests/candidateAssembler.test.ts tests/reviewLifecycle.test.ts tests/reviewRepairCandidateLifecycle.test.ts tests/runRepairCandidateLifecycle.test.ts tests/candidateRepairProtection.test.ts tests/candidateWaveLineage.test.ts` — exit 0, 7 files / 41 tests; `npm run typecheck` — exit 0; `npm run build` — exit 0 (`release-1790299120251-2597132-058f35f2`); `npx vitest run --exclude 'tests/system/**'` — exit 0, 153 files passed, 1 skipped; 848 tests passed, 7 skipped; `git diff --check` clean. No matching S4 system suite was found; the deterministic public `runTask` integration test exercises the slice. | No REAL_PROVIDER, BROWSER, OPA_PROVIDER, ADVERSARIAL, or PACKED_E2E claim. S6 acceptance oracle and live semantic/provider behavior remain pending. Flash worker `b6f54a20-9bd9-45bd-ae6b-fea824586940` used DeepSeek V4.1 Flash Implementer in the separate `s4-candidate-assurance-flash` Paseo worktree; no child agents, commit, or push. | FIXED_AND_DETERMINISTICALLY_VERIFIED (S4 gate; S6 oracle and provider lanes pending) |
| AEH-V2-0042 | P1 | AcceptanceOracle and objective completion | Managed run/change success could proceed from validation/review success without a persisted current AcceptanceOracle disposition, managed Lead evidence ordering, or deterministic objective Definition of Done; delivery could start before managed acceptance. | S4 produced bound assertions and assurance strength, but the production path did not resolve each assertion to sufficient provenance-bound evidence, persist a controller-owned disposition, enforce the disposition before delivery, or re-evaluate the complete objective evidence set at terminalization. | `src/architecture/acceptanceOracle.ts`, `src/architecture/objectiveCompletion.ts`, `src/agents/managedLeadAcceptance.ts`, `src/core/run.ts`, `src/delivery/finalize.ts`, `src/operations/controller.ts`, `src/operations/state.ts`, `tests/acceptanceOracle.test.ts`, `tests/managedLeadAcceptance.test.ts`, `tests/objectiveCompletion.test.ts`, `tests/objectiveCompletionLifecycle.test.ts`, `tests/deliveryFinalize.test.ts` | Regressions cover validated current policy/epoch identity, takeover digest rotation and old-evidence rejection, required assertion evidence strength, actual Reviewer execution/session provenance, bound managed Lead evidence, epoch-zero DoD, invalid/stale evidence, certification omission when not required, exact persisted-oracle/objective coverage matching, terminalization, and delivery blocking without accepted current disposition. | S6 resolves VerificationRequirements from unchanged S4 assertions, builds and persists current EvidenceBundles and deterministic ACCEPTED/REJECTED controller dispositions, requests bound Lead semantic evidence where frozen policy requires it, gates delivery effects, and recomputes complete objective DoD including WorkGraph/participant accounting at terminalization. The normative target, CertificationCore responsibility, S3 authority semantics, and S4 assertion/evidence-strength contracts were not changed. | Focused S6 `npx vitest run tests/objectiveCompletion.test.ts tests/objectiveCompletionLifecycle.test.ts tests/acceptanceOracle.test.ts tests/managedLeadAcceptance.test.ts tests/candidateAssurance.test.ts tests/runRepairCandidateLifecycle.test.ts tests/reviewLifecycle.test.ts tests/deliveryFinalize.test.ts tests/operationV2Identity.test.ts tests/operationLifecycleRegression.test.ts tests/operationSupervisionArchitecture.test.ts` — 11 files / 124 tests passed; `npx tsc --noEmit` — exit 0; `npm run build` — exit 0 (`release-1790303504564-2721309-64c0f9cc`); full non-system Vitest — 157 files passed, 1 skipped; 931 passed, 7 skipped; relevant system scenario/concurrency tests — 2 files / 17 passed; `git diff --check` — clean. | No REAL_PROVIDER, BROWSER, OPA_PROVIDER, ADVERSARIAL, or complete PACKED_E2E acceptance/delivery certification. The Campaign Director independently accepted the deterministic S6 gate on 2026-09-24 after reviewing the final diff and rerunning the focused, typecheck, build, full non-system, and relevant system gates. No external certification is claimed. | FIXED_AND_DETERMINISTICALLY_VERIFIED (S6 deterministic gate; external evidence lanes pending) |

| AEH-V2-0043 | P1 | Supply-chain provenance gate | Required provenance verification did not receive the current CandidateRevision or configured packed-artifact reference, and generated evidence did not form one deterministic candidate/build/artifact/policy/builder/SBOM/signature chain before delivery. | Version-1 provenance carried file digests and optional attestations but had no mandatory exact S7 bindings; strict verification selected evidence by filename rather than the policy-selected artifact, the default SBOM scan targeted the repository root, and delivery called the gate without current candidate/artifact context. | `src/provenance/generate.ts`, `src/build/identity.ts`, `src/core/config.ts`, `src/core/types.ts`, `schemas/project.schema.json`, `src/delivery/finalize.ts`, `src/cli.ts`, `src/core/doctor.ts`, `tests/provenance.test.ts`, `tests/provenanceS7.test.ts`, `tests/deliveryFinalize.test.ts`, `tests/system/aehAdversarialE2E.test.ts` | Deterministic regressions cover missing chain elements, malformed/tampered artifacts and SBOM, wrong/stale candidate/build/policy/builder bindings, statement/predicate/subject mismatch, configured-key signing boundary, explicit retired-alias rejection, non-strict behavior, and no delivery request or commit when required evidence is missing. | Version-2 evidence binds exact current candidate, policy-selected packed artifact and digest, running BuildIdentity, supply-chain policy digest, expected builder, SBOM and in-toto/SLSA evidence, and policy-required Cosign key verification. Strict delivery checks after the persisted S6 AcceptanceOracle and before effects. Old manifest versions and `cosignKey` alias have no compatibility path. No target wording, acceptance, authority, or S8 subsystem changed. | Focused `npx vitest run tests/provenance.test.ts tests/provenanceS7.test.ts tests/architectureIntegrationContracts.test.ts tests/deliveryFinalize.test.ts --reporter=dot` — 4 files / 49 passed; typecheck exit 0; build exit 0 (`release-1790308464438-2905085-da2062aa`); full non-system — 158 files passed, 1 skipped; 965 passed, 7 skipped; S7-specific adversarial missing-manifest system case — 1 passed / 27 skipped; schema JSON and `git diff --check` clean. Three-file broader system run had 43 passed / 2 failed in unrelated terminal-operation fixtures missing required S6 objective/AcceptanceOracle evidence. | No REAL_PROVIDER signing was run because the `cosign` CLI is unavailable in this environment. The S7 fake Cosign/PATH-shim bundle is explicitly marked deterministic fixture data and cannot establish provider certification. | FIXED_AND_DETERMINISTICALLY_VERIFIED (S7 deterministic gate accepted by Campaign Director 2026-09-24; provider certification pending) |
| AEH-V2-0044 | P2 | S7 provenance schema conditions | Optional/non-strict provenance configuration could be rejected because the schema conditions applied required constraints when their relevant flags were absent. | JSON Schema properties checks do not fail when an instance property is absent, and the allOf if branches lacked required guards at the provenance object and nested policy object. | schemas/project.schema.json, src/core/config.ts, templates/project.yaml, tests/projectSchemaS7.test.ts | Structural regressions require parent and nested required guards for every policy flag; existing config-loader cases accept the template/non-strict policies and valid strict policy, and reject missing artifact or required key material. | Added parent and nested required guards to both provenance condition groups. Runtime Zod policy and template semantics are unchanged; no JSON Schema validator dependency was added. | Pre-fix Flash worktree run: 1 file failed, 6 structural guard tests failed, 15 passed. Post-fix focused S7 set: 5 files / 70 tests passed; schema-specific suite 21 passed; Campaign Director independently reran TypeScript noEmit (exit 0), build (exit 0; release-1790310426337-17-0e3b4825), full non-system suite (159 files passed, 1 skipped; 986 passed, 7 skipped), scenario matrix (8 passed), isolated adversarial missing-manifest case (1 passed / 27 skipped), schema JSON parse, and diff checks. | No JSON Schema validator is installed/resolvable. No REAL_PROVIDER signing claim. The previously recorded two broader adversarial terminal-fixture failures remain the known S6 objective-completion/AcceptanceOracle setup issue; this remediation did not change them. | FIXED_AND_DETERMINISTICALLY_VERIFIED (S7 deterministic gate accepted by Campaign Director 2026-09-24) |
| AEH-V2-0045 | P1 | Runtime supervision and Control Center projection | The initial S9 review found that managed provider leases were not consumed by a production provider/session adapter, and the Control Center projection omitted typed stage/blocker/participant/runtime detail and a cancellation control. | Generic managed lease operations and fixture UI callbacks did not establish production Paseo lifecycle ownership or durable operation decision/resume behavior. | `src/runtime/supervisorV2.ts`, `src/runtime/managed.ts`, `src/runtime/providerLifecycle.ts`, `src/paseo/runtimeCore.ts`, `src/paseo/runtimeInitialTurn.ts`, `src/operations/controller.ts`, `src/control-center/operationProjection.ts`, `src/control-center/contracts.ts`, `src/entry.ts`, `ui/control-center/src/api.ts`, `ui/control-center/src/App.tsx`, `tests/paseoProviderLeaseLifecycle.test.ts`, `tests/paseoInitialTurnBarrier.test.ts`, `tests/packed/s9ProviderLifecycleCampaign.mjs`, Control Center suites | Deterministic tests cover concurrent conflicting lease owners, durable acquire/renew/release, expiry fencing, stale controller/identity rejection, uncertain provider retention, exact-session takeover, bound Lead-generation checks, unsupported unleased fallback rejection, stale-controller stop fencing, same-controller overlapping-write exclusion, cancellation cleanup, durable semantic projection, and paired scoped decision/replay boundaries. The packed campaign kills a child owner, verifies newer-epoch takeover after exact-session idle observation, injects a provider stop failure and verifies the operation stays RUNNING with its lease retained, then completes scoped restart cancellation only after stop/re-observation and lease release. | Production operation-owned Paseo SDK materialization and participant/bound-Lead turns now use the managed provider lease API; each lifecycle action has a unique owner, while current operation/candidate/execution revision/policy/token/epoch and S1/S5 session identity remain authoritative. Operation-owned materialization fails closed if only the unleased CLI lifecycle is available. Control Center projections and paired controls remain read-only/server-routed. No normative architecture, S3 authority, S5 session/context, or S8 public-effect acceptance boundary changed. | Final remediation focused command — 13 files / 137 tests passed; `npm run typecheck`, `npm run ui:build`, and `npm run build` exited 0 (direct build `release-1790334556192-3490030-97632a32`); `tests/system/aehConcurrencyCampaign.test.ts` — 1 file / 9 tests passed (source-importing system lane); `npm pack` prepare and the final R5 packed campaign passed against `/tmp/s9-packed-final4/agentic-engineering-harness-0.8.4.tgz`, extracted to `/tmp/s9-packed-fixture-final4/package`. The final tarball digest is included in the S9 handoff report. Prior fixture browser choice/cancel results remain BROWSER_FIXTURE only. | The packed/provider calls used deterministic inspect/stop stubs, so REAL_PROVIDER lifecycle certification is unclaimed. The direct Paseo browser bridge returned `No browser automation host is connected.`; durable-operation browser choice/consume/resume/revalidation and packed product-choice certification remain open. | **SLICE_BLOCKED — Campaign Director disposition 2026-09-25.** Production lease wiring and deterministic packed lifecycle recovery pass; durable-operation browser choice/consume/resume/revalidation remains open because Paseo reported no connected browser automation host. REAL_PROVIDER lifecycle certification remains a separate open lane. |

| AEH-V2-0046 | P1 | Detached operation bootstrap failure finalization | A non-Git project operation was persisted as `RUNNING/preparing` and exited with a dead controller PID after strict `baseRef=main` resolution failed; no terminal event or current failure projection was written. | HEAD and the Lead-selected packed release performed config load, base-ref resolution, and portfolio sync after the RUNNING transition but before the failure-finalization catch. The exception escaped terminalization. | `src/operations/controller.ts`, `tests/operationStartupFailure.test.ts`, `docs/CORE_ARCHITECTURE_V2_S9_WORKGRAPH.md`, S9 status/conformance documents | The single regression exercises `executeOperation` and production state APIs in an initialized non-Git root. It asserts terminal `FAILED`, diagnostic, terminal event, current portfolio projection, and no participants/process handles/provider leases. Flash baseline run reproduced the escape; Luna rerun passes against the fix. | Moved the RUNNING transition and bootstrap steps inside the existing controller try/catch, preserving strict base-ref failure and controller fencing. The controller does not synthesize a Git base. | Flash: `npx vitest run tests/operationStartupFailure.test.ts` against `c036471` — exit 1, 1 failed at `resolveBaseRef` before catch. Integrated current source: `./node_modules/.bin/vitest run tests/operationStartupFailure.test.ts --reporter=dot` — 1 file / 1 test passed; `npm run typecheck` — exit 0. Director independently ran the earlier draft — 1 test passed, typecheck and build exit 0 (`release-1790359341925-17-d65e1b0f`). Fresh package SHA-256 `77a0c39f580053b67abe2ee444e6b4c2995ecdc417a66613eb56d5dc20a1782d`, selected release `release-1790359422074-34-41c17389`; packed inspection confirms the catch encloses RUNNING/config/base-ref/portfolio bootstrap. | Director owns recovery of the same durable operation after ordinary Git initialization. Real Playwright choice/consume/resume/replay/cancel/refreshed-projection evidence remains pending; no product browser acceptance is claimed. | FIXED_AND_DETERMINISTICALLY_VERIFIED (S9 R12 subgate; full S9/browser gate open) |

| AEH-V2-0047 | P1 | Operation Supervisor provider identity | The canonical Operation Supervisor could not acquire a provider lease or bind a structured result: `PASEO_PROVIDER_LEASE_PARTICIPANT_UNKNOWN`, `V2_RESULT_PROVENANCE: execution binding participant is not registered for this role`, and `CONTEXT_RUNTIME_V2_BINDING_REJECTED`. | Supervisors are tracked in `supervision.generations` and deliberately excluded from `operation.participants`, but the S9 provider-lease, result-provenance, context-authorization, and dispatch paths all resolved actor identity from the participant map only. `dispatchMaterializedAgentPrompt` also used the provider session id as the authority participant and omitted the projected prompt and bound actor labels. | `src/runtime/supervisorV2.ts`, `src/runtime/providerLifecycle.ts`, `src/paseo/runtimeCore.ts`, `src/operations/state.ts`, `src/workers/agentPrompt.ts`, `src/workers/resultGateway.ts`, `src/context/authorizationV2.ts`, `tests/paseoProviderLeaseLifecycle.test.ts`, `tests/managedLeadAcceptance.test.ts`, `tests/changeControllerOwnedDecision.test.ts` | Deterministic regressions cover supervisor lease identity, agent-record execution binding, prompt-manifest dispatch, and the full controller-owned FORMAL_SDD journey. | A supervisor actor (`supervisorAgentId`) is validated against the durable agent record for leases, execution bindings, context authorization, and result provenance; dispatch uses the materialized participant identity, the projected prompt, and complete bound labels. | Focused S9 set 25 files / 221 tests passed; `tests/changeControllerOwnedDecision.test.ts` passed; typecheck/build exited 0. | Real browser journey passed 46/46; supervisor initialization ran through the scripted provider boundary (not REAL_PROVIDER). | FIXED_AND_DETERMINISTICALLY_VERIFIED (S9 completion; independent review pending) |
| AEH-V2-0048 | P1 | Agent preset vs frozen RoleProfile ceilings | FORMAL_SDD failed at the first Explorer turn with `V2_AUTHORITY_DENIED: role 'Explorer' cannot receive capabilities outside its compiled RoleProfile ceiling: execute`; lead, librarian, planner, spec-manager, and reviewer presets had the same class of mismatch, and preset skill lists referenced repository skills absent from the seed catalog (`SKILL_MANIFEST_INVALID`). | The default preset predated the frozen role ceilings and granted shell/network/write/delegate capabilities the ceilings deny, plus repository skill ids the SkillManifest compiler cannot resolve. The effective authority was already the ceiling, so the excess only produced hard launch failures. | `presets/agents/default.jsonc`, `presets/agents/orchestration.jsonc` | The controller-owned FORMAL_SDD journey reaches the Spec Manager and suspends; topology/hygiene tests still pass. | Aligned each default-preset permission set to the compiled ceiling and replaced unresolvable repository skill ids with seeded skills. Repository-skill resolution into SkillManifest remains an open capability gap. | Focused S9 set passed; `tests/changeControllerOwnedDecision.test.ts` passed; typecheck/build exited 0. | Real browser journey passed 46/46. | FIXED_AND_DETERMINISTICALLY_VERIFIED (repository-skill resolution gap recorded) |
| AEH-V2-0049 | P1 | Multi-choice Spec Manager continuation | A second product-choice suspension in the same FORMAL_SDD Spec Manager loop failed with `DECISION_REQUEST_STATE_INVALID` because the guard required phase `spec-authoring` while the resumed turn had advanced to phase `context`. | `suspendOperationForProductChoice` computed `replacingConsumedChoice` with an over-strict phase equality instead of the durable `RESUMING` continuation state. | `src/operations/state.ts`, `tests/changeControllerOwnedDecision.test.ts` | The regression drives decision, consumption, resume, and a second suspension through the real controller. | The guard now accepts a consumed `RESUMING` continuation without a pending request. | `tests/changeControllerOwnedDecision.test.ts` passed; focused S9 set passed. | Real browser journey exercised two controller-owned suspensions and passed 46/46. | FIXED_AND_DETERMINISTICALLY_VERIFIED |
| AEH-V2-0050 | P1 | Cancellation writer fencing | Cancellation could not terminalize because it attempted `paseo stop` on authority participant identities (`participant:<digest>`) and scripted fixture sessions, producing `AEH_CANCELLATION_FENCING_REQUIRED`. | The cleanup loop treated every registered agent id as a provider session; authority identities and deterministic fixture sessions have no external process. | `src/operations/controller.ts`, `src/paseo/deterministicRuntime.ts` | `tests/operationLifecycleRegression.test.ts` still stops real registered agents; the controller-owned journey now cancels to `CANCELLED`. | Authority `participant:<16hex>` ids are skipped, deterministic session ids are recognized by prefix in any lifecycle owner, and deterministic provider leases are released without an external stop. | Focused S9 set passed; `tests/changeControllerOwnedDecision.test.ts` terminalizes CANCELLED. | Real browser journey cancelled to durable `CANCELLED` with stale-writer and stale-decision rejection; 46/46 passed. | FIXED_AND_DETERMINISTICALLY_VERIFIED |
| AEH-V2-0051 | P2 | Operation coordination lock and candidate task id | Supervisor initialization timed out acquiring its own coordination lock (`Timed out acquiring operation state lock ...coordination`), and a CHANGE operation with an explicit `--task` id failed structured-result provenance with `AEH_RESULT_PROVENANCE: task identity does not match the candidate`. | `withOperationCoordinationLock` was not re-entrant for the holding async context, and `saveOperation` always bound the operation id as the change candidate task id. | `src/operations/state.ts`, `tests/operationStartupFailure.test.ts`, `tests/changeControllerOwnedDecision.test.ts` | The controller-owned journey exercises nested context authorization and an explicit task id. | The coordination lock is re-entrant only within the holding async context (other callers still serialize on the file lock); the change candidate task id uses the payload task id when declared. | Focused S9 set passed; typecheck/build exited 0. | Real browser journey passed 46/46. | FIXED_AND_DETERMINISTICALLY_VERIFIED |
| AEH-V2-0052 | P1 | Rootless sandbox and validator isolation | Rootless/network/scope isolation and validator-command isolation were unproven: the only hardened worker path built Podman arguments, `podman`/`buildah` were absent, isolation was asserted only by argument-array unit tests, and validator commands executed as host child processes through `runShell` with no namespace boundary. | No execution provider abstracted rootless isolation; `src/security/sandbox.ts` produced Podman flags and `src/workers/agentPrompt.ts` spawned `podman run`, while `src/validators/*` ran commands directly on the host. | src/security/isolation.ts, src/validators/external.ts, src/validators/toolCommand.ts, src/validators/commands.ts, src/validators/registry.ts, src/core/verify.ts, src/core/run.ts, src/core/types.ts, src/core/config.ts, schemas/project.schema.json, src/utils/process.ts, tests/security/isolation.test.ts, tests/security/validatorIsolationFailClosed.test.ts, tests/security/providerFailClosed.test.ts, tests/system/aehIsolationCampaign.test.ts | Real rootless `bwrap 0.13.0` execution proves user/mount/PID/UTS/IPC/network namespaces, network denial by real `ENETUNREACH`, host-home/PID/env masking, workspace-only writes, read-only `/etc`, and validator repository-mutation/egress denial; missing or unsupported providers return explicit `ISOLATION_PROVIDER_UNAVAILABLE`/`ISOLATION_PROVIDER_UNSUPPORTED` blockers without running the validator. Focused suites 5 files / 30 tests passed; system lanes scenario+concurrency+isolation 3 files / 20 tests passed. | Real local OSS provider execution (bwrap) with real kernel namespaces; rootless Podman/OCI execution remains unprovisioned and fails closed. | FIXED_AND_REAL_LOCAL_PROVIDER_VERIFIED (S10; independent slice review pending) |
| AEH-V2-0053 | P1 | Candidate-bound SAST evidence | SAST was not wired: `ValidationContext` had no candidate identity, the external tool validator wrote only a `spec.id`-named raw artifact, and no SAST report was traceable to the exact current CandidateRevision. | Security validation ran without a candidate binding, and evidence artifacts were not candidate-scoped or digest-verified. | src/security/sastEvidence.ts, src/validators/external.ts, src/validators/types.ts, src/core/run.ts, tests/security/sastEvidence.test.ts, tests/packed/s10RealSastCampaign.mjs, docs/evidence/s10/real-sast-campaign.json | `SastEvidenceV1` binds candidate identity, workspace source digest, adapter, tool version, command digest, raw-output digest, findings, and isolation evidence under a candidate-scoped path; digest/tamper/stale/missing checks block; the S4 security impact path requires it for opengrep/trivy. Real local OSS Trivy v0.70.0 produced a finding-bound FAIL for r1 and a clean PASS for r2, with r1 evidence rejected as stale for r2. | Real local OSS scanner execution; hosted-provider certification remains S13. | FIXED_AND_REAL_LOCAL_PROVIDER_VERIFIED (S10; independent slice review pending) |
| AEH-V2-0054 | P1 | Contract/integration/browser/visual lane evidence and specialized capability fallthrough | No candidate-bound evidence artifact existed for CONTRACT, INTEGRATION, BROWSER, or VISUAL validation; a declared `browser-test` capability silently ran generic unit tests and could PASS with no browser; `visual` did not exist; a required Pact/OCI check was coerced from SKIP to FAIL with no explicit provider blocker; and the integration lifecycle treated an absent readiness command as ready. | Provider results were normalized into transient `ValidationCheck` details only, `capabilityAdapter` defaulted specialized capabilities to `test-execution`, and the integration normalizer defaulted readiness. | src/validation/laneEvidence.ts, src/providers/validation/registry.ts, src/providers/validation/integrationEnvironment.ts, src/providers/validation/types.ts, src/validators/registry.ts, src/validators/external.ts, src/core/run.ts, src/architecture/validationRequirements.ts, src/architecture/candidateAssurance.ts, src/architecture/acceptanceOracle.ts, src/core/doctor.ts, src/core/sdd.ts, schemas/validation-result.schema.json, tests/validation/laneEvidence.test.ts, tests/validation/failClosedProviders.test.ts, tests/system/aehContractIntegrationCampaign.test.ts, tests/fixtures/integration/http-service.mjs, docs/evidence/s11/baseline-gaps.md | `ProviderLaneEvidenceV1` binds candidate/workspace/provider/command/findings/artifact digests under a candidate-scoped path with stale/tamper/required rejection and lane separation; the S4 impact path requires lane evidence for contract/BDD/integration/browser/visual assertions; declared specialized capabilities resolve to matching adapters or fail with `UNSUPPORTED_VALIDATION_CAPABILITY`; missing Pact/OCI providers return explicit `*_PROVIDER_UNAVAILABLE` blockers; the lifecycle requires provision/readiness/test/cleanup and always attempts cleanup. Real OpenAPI/BDD/loopback-service executions produced verified candidate-bound CONTRACT/INTEGRATION evidence. Focused S11 set 11 files / 70 tests; system campaign 3 tests. | Real local deterministic executions; Pact, hosted-provider, and packed lanes remain open and fail closed. | FIXED_AND_DETERMINISTICALLY_VERIFIED (S11; independent review rejected the first submission; remediation recorded in AEH-V2-0056..0058; independently accepted 2026-09-26 by R3, checkpoint 78de9c5) |
| AEH-V2-0055 | P1 | Candidate-bound browser and visual validation through the validator path | The Playwright adapter produced only a `ValidationCheck` and raw stdout; no candidate-bound browser evidence existed, `visual` was absent, and the default command used bare `npx playwright` rather than the pinned candidate binary. | External validators ran without a candidate-scoped artifact, and no visual requirement kind/provider/evidence lane existed. | src/validators/external.ts, src/validation/laneEvidence.ts, src/architecture/candidateAssurance.ts, src/architecture/validationRequirements.ts, src/core/run.ts, tests/browser/s11-control-center-browser-visual.e2e.ts, tests/browser/fixture/controlCenterJourney.ts, scripts/s11BrowserVisualCampaign.ts, docs/evidence/s11/browser-visual-campaign.json | Required browser/visual impact validations require candidate-bound BROWSER/VISUAL evidence; the pinned candidate Playwright binary is the only default provider; missing binary/browser/candidate binding fails with explicit blockers; digest-bound screenshot artifacts are retained. `npm run test:browser-visual` ran the real Control Center journey in real Chromium through both providers (release-1790392483278-608471-43c5a8c4) and verified both evidence digests. | Real local OSS browser execution with the S9 file-scripted deterministic Paseo provider boundary; not REAL_PROVIDER, live-model, or packed certification. | FIXED_AND_REAL_LOCAL_BROWSER_VERIFIED (S11; independent review rejected the first submission; remediation recorded in AEH-V2-0056..0058; independently accepted 2026-09-26 by R3, checkpoint 78de9c5) |
| AEH-V2-0056 | P1 | Required specialized lanes and browser/visual artifact durability | The independent S11 review found that a required `browser-test`/`visual-test` requirement could be satisfied by a project script (`npm run e2e` where the script was `node -e "process.exit(0)"`) whose passing check was synthesized into BROWSER evidence by `ensureProviderLaneEvidenceFromCheckV1`, with provider identity taken from the action selector and version `unknown`; no browser ran. The same false equivalence covered contract/integration/BDD. In addition, `scripts/s11BrowserVisualCampaign.ts` ran BROWSER and VISUAL with the same Playwright `--output` directory, so the VISUAL run deleted the BROWSER screenshot attachment and post-campaign `verifyProviderLaneEvidenceV1(BROWSER)` failed with `PROVIDER_LANE_EVIDENCE_TAMPERED: screenshot artifact ... is missing`; the campaign's `record.verified` only reloaded and digest-compared the JSON instead of verifying artifacts. | `runCandidateImpactValidations` called the synthesis helper for every lane-kind PASS; lane evidence was minted from the check rather than from provider execution. Both lanes shared one Playwright output directory and cleared it at run start; the campaign never invoked `verifyProviderLaneEvidenceV1`. | src/validation/laneEvidence.ts, src/core/run.ts, src/validators/openapi.ts, src/architecture/validationRequirements.ts, src/validators/external.ts, scripts/s11BrowserVisualCampaign.ts, tests/validation/requiredLaneEvidence.test.ts, tests/validation/laneEvidence.test.ts, tests/system/aehContractIntegrationCampaign.test.ts | Adversarial `requiredLaneEvidence.test.ts` drives the real impact-validation dispatch: project-script-only browser/visual/contract/integration requirements FAIL with `PROVIDER_LANE_EVIDENCE_REQUIRED` after the script demonstrably ran and exited zero, while the OpenAPI, isolated-service lifecycle, and Playwright-provider routes PASS; lane-evidence tests pin the typed missing-evidence behavior. | `ensureProviderLaneEvidenceFromCheckV1` is deleted; `requireProviderLaneEvidenceForActionV1` loads and verifies the candidate-bound lane evidence persisted by the matching provider/adapter and fails closed with `PROVIDER_LANE_EVIDENCE_REQUIRED` otherwise. The OpenAPI adapter resolves for `contract-test` and persists its own CONTRACT evidence bound to the current candidate. Each campaign lane owns a separate durable Playwright output directory and the campaign calls `verifyProviderLaneEvidenceV1` for both lanes after the full run, failing unless both are `ok` with all artifacts present; the sanitized summary records the verified artifacts. | Focused suites (8 files / 49 tests): tests/validation/laneEvidence.test.ts, tests/validation/failClosedProviders.test.ts, tests/validation/requiredLaneEvidence.test.ts (8 tests), tests/system/aehContractIntegrationCampaign.test.ts, tests/openapi.test.ts, tests/validationProviders.test.ts, tests/candidateAssurance.test.ts, tests/acceptanceOracle.test.ts — all passed; `npm run typecheck` exit 0; full non-system 178 files passed / 1 skipped with only the known `tests/paseoSdkResolve.test.ts` mise drift (3 tests) failing; deterministic system suites 4 files / 23 tests passed. | `npm run test:browser-visual` exit 0 (release-1790455687992-64113-eb8d526a): BROWSER and VISUAL evidence both `ok=true` with a MATCH workspace binding, all artifacts present, and recomputed digests. Real local OSS browser execution with the file-scripted deterministic Paseo provider boundary; not REAL_PROVIDER. | FIXED_AND_REAL_LOCAL_BROWSER_VERIFIED (S11 remediation; independently accepted 2026-09-26 by R3, checkpoint 78de9c5) |
| AEH-V2-0057 | P2 | VISUAL lane reference and comparison binding | The independent S11 review found that `ProviderLaneEvidenceV1` had no reference/baseline identity or comparison-configuration binding, even though the visual journey used a Playwright comparator (`maxDiffPixelRatio: 0.05`) against a committed baseline; the visual lane could not prove which baseline or comparator produced its verdict. | The lane-evidence contract carried only provider/command/artifact digests; the committed baseline was an implicit input of the test source and the comparison options were not recorded or verified. | src/validation/laneEvidence.ts, src/validators/external.ts, scripts/s11BrowserVisualCampaign.ts, tests/validation/laneEvidence.test.ts, tests/validation/requiredLaneEvidence.test.ts, tests/browser/s11-control-center-browser-visual.e2e.ts | Focused lane-evidence tests pin the bare-VISUAL rejection, the bound baseline/comparison pass, and baseline tamper detection. | The repaired campaign persisted VISUAL evidence with baseline artifact digest `21ec4e27bdb2aca29a3ed672c310e35533289f4636ddfe26f93badd8772565ee` and comparison `playwright-toHaveScreenshot` `{ animations: disabled, maxDiffPixelRatio: 0.05 }`. | Minimal contract extension: the artifact kind union adds `baseline` and the evidence payload adds an optional `comparison: { tool, name?, options }`. Required VISUAL checks must declare `options.referenceBaseline` and `options.comparison`; a missing/unreadable baseline or malformed comparison returns `VISUAL_REFERENCE_BASELINE_REQUIRED`/`VISUAL_COMPARISON_CONFIG_REQUIRED` and never auto-generates a baseline from the candidate. `verifyProviderLaneEvidenceV1` requires VISUAL evidence to carry a baseline artifact plus a non-empty comparison tool. | `npm run test:browser-visual` exit 0 with both lanes `ok=true`. Not REAL_PROVIDER or cross-platform pixel certification. | FIXED_AND_REAL_LOCAL_BROWSER_VERIFIED (S11 remediation; independently accepted 2026-09-26 by R3, checkpoint 78de9c5) |
| AEH-V2-0058 | P2 | Candidate digest reproducibility, docs citations, and integration cleanup semantics | The independent S11 review found that `computeWorktreeDigest` included the untracked `docs/evidence/s11/browser-visual-campaign.json` while the campaign rewrote that file after computing the candidate digest, so every campaign run changed the candidate identity (prior dirs 682e53f/9e350e0/d662dae/3f5d878c/77df7aba), the recorded source digest no longer matched the tree, and docs cited digests matching nothing. It also found `INTEGRATION_CLEANUP_REQUIRED` emitted even when provisioning failed and no resources were created. | The campaign wrote a generated summary into the source inventory after binding the candidate; the integration normalizer gated cleanup only on the failed predicate and never distinguished zero resources from a missing cleanup mechanism. | src/core/git.ts (read-only observation), scripts/s11BrowserVisualCampaign.ts, src/providers/validation/integrationEnvironment.ts, src/providers/validation/types.ts, schemas/validation-result.schema.json, tests/validation/failClosedProviders.test.ts, docs/evidence/s11/browser-visual-campaign.rejected-pre-repair.json | A fresh post-campaign `computeWorktreeDigest` recomputation equals the recorded candidate source digest and `assertWorkspaceMatchesCandidate` returns `status: MATCH`; cleanup semantics are pinned by two new `failClosedProviders.test.ts` regressions (provision failure creates nothing and yields no cleanup blocker; provisioned resources without a cleanup mechanism yield `INTEGRATION_CLEANUP_REQUIRED`); docs now cite the exact repaired digests. | The campaign's machine summary now lives in the ignored configured evidence directory (`.harness/evidence/s11/browser-visual-campaign.json`); the campaign recomputes the worktree digest before and after the summary write and fails unless both equal the candidate source digest. The integration lifecycle records `cleanupRequired: provisioned` and emits `INTEGRATION_CLEANUP_REQUIRED` only when resources were provisioned without a cleanup command. The rejected pre-repair summary is preserved as `docs/evidence/s11/browser-visual-campaign.rejected-pre-repair.json`. | Campaign evidence: candidate `CAND-S11-BROWSER-VISUAL` r1 identity digest `986c1daadb6ed8fe030badc5e9cd8215d2a182cc63f45fdad0760b3269528d73`, workspace source digest `73c33007e15d4fdc5516d49a02dfc6cf3a9cd0d5a995778e5e3460aaa6cf465a`; BROWSER evidence digest `9c166b826992119a44cf61f064c57e34a464e059bf251aecf3275e2ffea63ac2`, VISUAL evidence digest `b9c187ddd52092a4c23816289f7616e6e41ee63f00ebb140c7811e3405193355`. | `npm run test:browser-visual` exit 0 with both lanes verified; `git diff --check` clean. No REAL_PROVIDER claim. | FIXED_AND_DETERMINISTICALLY_VERIFIED (S11 remediation; round-1 repairs verified by R2; independently accepted 2026-09-26 by R3, checkpoint 78de9c5) |
| AEH-V2-0059 | P1 | Declared BDD/gherkin route bypassed the capability provider | The R2 review found that a required declared `bdd` capability with no runner returned `FAIL` with the bare message `bdd-runner acceptance was skipped.` and no typed blocker (the normalized `SKIP` was only coerced by `required`), a configured `bdd` provider `{capability:"bdd", command:<passing runner>}` was ignored in favor of the package.json `bdd` script, and no candidate-bound CONTRACT evidence could be persisted on the declared route. | `src/validators/registry.ts` routed `case "gherkin"/"bdd"` to `runGherkinValidator`, which built a `ValidationProviderContext` without `providerSpec` or `candidate` and returned the generic `resultCheck(...)` instead of the typed `runCapabilityValidator` BDD path (`CONTRACT_PROVIDER_UNAVAILABLE` + `finalizeLaneCheck`). | src/validators/registry.ts, deleted src/validators/gherkin.ts, maturity/components.yaml, docs/component-maturity.json, tests/validation/failClosedProviders.test.ts | Three new `failClosedProviders` regressions: declared `bdd` with no runner → FAIL + `CONTRACT_PROVIDER_UNAVAILABLE` and never the bare skip message; a declared `bdd` with a configured passing runner → configured runner executes (configured marker present, package `bdd` script marker absent) and candidate-bound CONTRACT evidence verifies; a configured failing runner → FAIL. The real BDD campaign test and the BDD provider unit tests still pass. | `gherkin`/`bdd` now route through `runCapabilityValidator` with `providerSpec` and `candidate`; the superseded `runGherkinValidator`/`src/validators/gherkin.ts` adapter is deleted (no alias), and the maturity inventory migrates its adapter-source evidence to `src/validators/registry.ts`. Reqnroll and the S4 impact-path BDD routes are unchanged. | Focused S11 set 9 files / 57 tests passed; full non-system 177 files passed / 1 skipped (1161 passed / 7 skipped) with only the known `tests/paseoSdkResolve.test.ts` mise drift failing; four deterministic system suites 4 files / 23 tests passed; `npm run typecheck` exit 0. | Local deterministic BDD runners only; no REAL_PROVIDER claim. The round-2 `npm run test:browser-visual` rerun also passed (release-1790457526805-146084-3114256d). | FIXED_AND_DETERMINISTICALLY_VERIFIED (S11 R2 remediation; independently accepted 2026-09-26 by R3, checkpoint 78de9c5) |
| AEH-V2-0060 | P3 | Browser/visual lane output durability across candidates | The R2 review found the per-lane Playwright `--output` directory was not candidate-scoped, so a later campaign run against a different candidate deleted the prior candidate's screenshot artifacts and its `ProviderLaneEvidenceV1` failed verification (observed for candidate `986c1daadb6e` after a newer run). | `scripts/s11BrowserVisualCampaign.ts` removed and reused a flat `.harness/evidence/playwright-s11-output/<lane>` path for every candidate. | scripts/s11BrowserVisualCampaign.ts | The earlier `a20fd933b163` candidate evidence still verifies `ok=true` after the newer `7ad7d70b2ef1` run; the campaign still fails if either lane does not verify and still checks for digest stability before/after the summary write. | Each lane now writes to `.harness/evidence/playwright-s11-output/<lane>/<candidateId>-r<revision>-<identity12>` and only removes its own candidate-scoped directory, so later candidates cannot delete prior candidate artifacts. | Round-2 campaign: BROWSER digest `d98c455a87ddcff72214d4100fb0938628496e1407300ef3a783e3d800aef932`, VISUAL digest `409aca8d6deb4758f55b437ae4a92c18a6943b4b9fca66d7d164099c2b9b6515`, both `ok=true`, workspace `MATCH`; prior-candidate retention independently re-verified. | `npm run test:browser-visual` exit 0 (release-1790457526805-146084-3114256d); not REAL_PROVIDER or packed certification. | FIXED_AND_REAL_LOCAL_BROWSER_VERIFIED (S11 R2 remediation; independently accepted 2026-09-26 by R3, checkpoint 78de9c5) |
| AEH-V2-0131 | P1 | Candidate-bound S11 visual provider route | The required `visual-test` requirement resolved to the generic `test:browser-visual` project script, which ran a standalone campaign with a synthetic `CAND-S11-BROWSER-VISUAL` identity. Its matching source digest did not bind evidence to the outer operation candidate, so the deterministic assurance gate correctly rejected it. | `.harness/project.yaml` had no configured `visual-test` provider; standalone campaign identity was minted inside `scripts/s11BrowserVisualCampaign.ts`. | `.harness/project.yaml`, `src/validators/external.ts`, `scripts/s11BrowserVisualCampaign.ts`, `tests/validation/requiredLaneEvidence.test.ts` | Project-script-only VISUAL remains blocked with `PROVIDER_LANE_EVIDENCE_REQUIRED`; the configured visual adapter persists lane evidence with exact candidate identity, committed baseline digest and comparison configuration; same-source evidence for another candidate is rejected. | The S11 project config now resolves `visual-test` through the configured `playwright` provider and its committed baseline/comparison. The adapter passes its frozen CandidateRevision to the command environment. The campaign recognizes that provider invocation, runs only the real S11 visual Playwright journey, and returns Playwright JSON; the adapter persists and verifies VISUAL evidence against the outer identity. The standalone two-lane campaign remains an independent local evidence entrypoint. | Local focused provider tests and build/typecheck; browser launch/runtime still depends on the pinned Playwright Chromium being provisioned on the executing machine. No REAL_PROVIDER or cross-platform pixel certification is claimed. | FIXED_AND_DETERMINISTICALLY_VERIFIED (candidate-bound provider route) |
| AEH-V2-0061 | P1 | Telemetry identity, metric instruments, and advisory eval corpus | Telemetry events and spans carried `taskId`/`operationId` but no candidate/revision/identity-digest/execution-revision/policy/epoch/participant/runtime-session correlation and had no stale-attribution detector; no metric instruments or local metric export existed; `evals/corpus/` did not exist and the runner recorded no corpus identity or reproducibility check; no local export verification lane or adversarial proof that eval/telemetry output cannot influence policy, authority, acceptance, permissions, or routing existed. | The S12 audit snapshot recorded NDJSON/OTel spans and an eval runner only; identity existed in operation state but was not projected into telemetry, and eval results were unbound observations with no corpus contract. | src/telemetry/identity.ts, src/telemetry/events.ts, src/telemetry/metrics.ts, src/evals/runner.ts, src/evals/scoring.ts, src/evals/advisoryInvariant.ts, src/evals/types.ts, src/core/run.ts, src/core/verify.ts, src/workers/agentPrompt.ts, src/paseo/sessionBinding.ts, src/core/config.ts, src/core/types.ts, templates/project.yaml, package.json, package-lock.json, evals/corpus/, evals/fixtures/, evals/scenarios/, scripts/s12ObservabilityCampaign.ts, tests/telemetry/, tests/evals/corpus.test.ts, tests/evals/advisoryInvariant.test.ts, tests/helpers/otlpReceiver.ts | `TelemetryCorrelationV1` gives one canonical identity and domain-separated digest across NDJSON events, OTel spans, and metric points; `recordEvent` resolves current durable identity at emission time and marks supplied stale/unresolved identity with explicit violations instead of silently re-attributing it. Real OTel counters/histograms cover operation/participant/runtime-session/validation signals with canonical local NDJSON export. The versioned corpus records corpus/case digests and scores reproducibly. The advisory-invariant guard proves no authority decision module imports observation modules, and fabricated eval/telemetry artifacts cannot change validation, durable operation truth, acceptance, or routing. | Focused S12 set 9 files / 25 tests passed; `npm run typecheck` exit 0; `npm run build` exit 0 (`release-1790460335620-200076-486c0202`); full non-system Vitest 185 files passed / 1 failed (`tests/paseoSdkResolve.test.ts`, 3 pre-existing mise environment-drift tests) with 1,182 passed / 3 failed / 7 pending; scenario matrix + concurrency + contract/integration 3 files / 20 tests passed; isolation campaign 1 file / 3 tests passed; adversarial suite 26 passed with its two long-recorded S6 fixture failures unchanged; `npx tsx scripts/s12ObservabilityCampaign.ts` wrote `docs/evidence/s12/s12-observability-evals-campaign.json` (corpus digest `e46fa3e1996b0cd70e62eb4069ae108836264ba8fbb57987340a748d896ef71e`, reproducible metric snapshot digests, verified local OTLP/HTTP JSON payload against an in-process loopback receiver). **Superseded by AEH-V2-0062 (S12-R1):** the reviewer rejected gate B because the corpus/case digest did not bind the executable scenario harness and the campaign canonical metric digest was not reproducible across executions; the file-status total was also misstated. | Local file/NDJSON plus local loopback OTLP verification only; no hosted collector, REAL_PROVIDER, BROWSER, or PACKED_E2E claim. | SUPERSEDED by AEH-V2-0062 (S12; independently accepted 2026-09-26 by S12-REVIEW-R2, checkpoint ff1f420) |
| AEH-V2-0062 | P1 | S12 corpus scenario binding, campaign cross-execution reproducibility, and metric correlation trust | The S12-REVIEW rejected gate B: (F1) `computeEvalCorpusIdentity` digested only `corpus.json`, each `eval.yaml`, and fixture files, so the declared but never-consumed `scenarios[].scenario` harness could change while `digest`/`caseDigest` stayed `e46fa3e1…`; a scored result could change under an identical recorded corpus digest. (F2) the campaign used `mkdtemp` for the operation root, and candidate identity binds `projectId`/`worktree`, so the canonical metric digest `a91916f4…` became `724b58d4…` on an independent run; `metrics.reproducible:true` was intra-process only. (F3) docs overstated the full non-system file counts. INFO: metric recording APIs trusted a caller-supplied correlation with no mismatch detection. | F1: scenario path was declared in the manifest and parsed by the eval schema but never resolved into the identity computation. F2: `src/operations/state.ts` derives project/candidate identity from the source root path, so a random temp root changed candidate → policy → correlation → metric attributes. F3: file-status totals were copied from the earlier JSON aggregation instead of the observed 186-file total. INFO: event recording had current-identity mismatch detection but metric recording did not. | src/evals/runner.ts, src/evals/types.ts, src/evals/scoring.ts, src/telemetry/metrics.ts, src/core/run.ts, src/core/verify.ts, src/workers/agentPrompt.ts, src/paseo/sessionBinding.ts, scripts/s12ObservabilityCampaign.ts, tests/evals/corpus.test.ts, tests/telemetry/metrics.test.ts, docs/CORE_ARCHITECTURE_V2_STATUS.md, docs/CORE_ARCHITECTURE_V2_CONFORMANCE.md, docs/OBSERVABILITY.md, docs/CORE_ARCHITECTURE_V2_S12_WORKGRAPH.md | F1: `tests/evals/corpus.test.ts` appends to `evals/scenarios/scopeGovernance.ts` and asserts the case and corpus digests change while another case is unchanged, then restoring the file restores the digest; a declared-but-missing scenario now fails with `EVAL_CORPUS_SCENARIO_MISSING`; each result also records `EvalBuildIdentityV1`. F2: the campaign recreates the fixed repo-relative `.harness/s12-campaign-scenario` root per execution; six independent executions in separate processes produced identical candidate `698e397a…`, policy `b1c8f1dd…`, participant correlation `20af5760…`, base correlation `ad2d03c5…`, corpus `11e9b972…`, and canonical metric snapshot digest `838305e6…`. The bounded read-only verification then found the digest keying ambiguous (duplicates retained, corpus-dir-relative keys), so the rule was canonicalized to a unique file set keyed by normalized repo-relative path, sorted, hashed as `relativePath\0fileSha256` lines; an independent implementation of that rule now reproduces `11e9b972…` exactly. The artifact `crossExecutionFields` path was corrected to `reproducibility.canonicalMetricSnapshotDigest`. INFO: `record*Telemetry` now resolves current durable identity and records nothing on mismatch; `tests/telemetry/metrics.test.ts` asserts a stale correlation is rejected and the current one records. | F1: scenario closure binding (declared entry plus relative imports inside the project `evals/` tree; production `src/` bound by build identity). F2: fixed ignored scenario root under the repo, removed/recreated per run; artifact records the canonical digest and cross-execution fields. F3: docs corrected to the observed 186 files / 1,195 tests (1,185 passed / 3 failed / 7 pending; the 7 pending are opt-in real-provider contracts reported as skipped by some views). INFO: metric correlation verification documented in OBSERVABILITY.md, including the detached unverified boundary. | Focused S12 set 9 files / 28 tests passed; affected wiring set 13 files / 126 tests passed; `npm run typecheck` exit 0; `npm run build` exit 0; full non-system Vitest ran 186 files and 1,195 tests (1,185 passed / 3 failed / 7 pending; file statuses 185 passed / 1 failed), the only failures being the 3 pre-existing `tests/paseoSdkResolve.test.ts` mise drift tests; system scenario/concurrency/contract 3 files / 20 tests passed; isolation 1 file / 3 tests passed; adversarial 26 passed / 2 long-recorded S6 fixture failures; campaign run twice per pair in separate processes, six total executions in separate processes identical; regenerated artifact SHA-256 `285e78ff74ca9bfe64bd3ece8d4018387e4508ba2597fbad94a9871f867398af`. | Local OTLP loopback only; no REAL_PROVIDER, hosted-collector, BROWSER, or PACKED_E2E claim. The fresh independent reviewer owns re-verification and acceptance. | FIXED_AND_DETERMINISTICALLY_VERIFIED (S12-R1 repair; independently accepted 2026-09-26 by S12-REVIEW-R2, checkpoint ff1f420) |


| AEH-V2-0063 | P1 | Packed certification bootstrap (S13) | `runExternalSelfDogfood` failed with `npm pack did not return JSON.` on this environment, and the bound `CandidateRevision` recorded only the harness build identity, not the build identity of the freshly packed artifact. | npm 10.9.8 runs the package `prepare` script during `npm pack` even with `--ignore-scripts`, so vite/build output precedes the JSON array on stdout; `parsePackFilename` required pure JSON. `runExternalSelfDogfood` also never read the installed artifact's `dist/current` + `build-identity.json`. | `src/certification/bootstrap.ts`, `tests/certificationJourneys.test.ts`, `tests/packed/s13RealProviderCertification.mjs` | `tests/certificationJourneys.test.ts` proves prepare-noise tolerance and unsafe/missing-artifact rejection; the S13 campaign's `startup`, `project-home`, and `multi-project` lanes exercised the repaired path end-to-end against freshly packed candidates. | `parsePackFilename` scans from the end for the JSON array line and parses from there; after fixture install the bootstrap reads `node_modules/agentic-engineering-harness/dist/current` and the release `build-identity.json` and records `packedBuildRelease`, `packedBuildDigest`, `packedBuildGitSha`, `packedBuildPackageVersion` in `CandidateRevision.metadata`. No alias or fallback path was added. | `npx vitest run tests/certification.test.ts tests/certificationMatrix.test.ts tests/certificationJourneys.test.ts` — 3 files / 28 tests passed; `npm run typecheck` exit 0. | S13 real-provider campaign: startup/project-home/multi-project ACCEPTED with real `codex-cli 0.156.1` receipts and packed artifact digest binding. Independent slice review pending. | FIXED_AND_REAL_PROVIDER_EXERCISED (S13; independent review pending) |
| AEH-V2-0064 | P1 | Semantic INTENT budget versus real providers (S13) | The packed candidate's `aeh intent` failed with `SEMANTIC_ASSESSMENT_UNAVAILABLE` (`exit=124, status=timeout`) although the Paseo semantic assessor agent actually launched and completed its turn. | `classifyEngineeringIntentWithSemanticAssessment` hardcodes `deadlineMs: 10_000`, and `PaseoSemanticAssessmentRunnerV1` derives the Paseo agent launch timeout from it (`timeoutSeconds = deadlineMs/1000`), which no cold real provider can meet. | `src/audit/intent.ts`, `src/semantic/runtime.ts`, `docs/evidence/s13/lanes/informational.json` | The S13 `informational` lane deterministically fails closed with missing `intent`/`answer`/`model.evidence` evidence; the product repair (provider-appropriate or configurable semantic deadline) is a blocked S13 row with an exact required action. | S13 intentionally did not change the accepted S2 semantic budget; the product defect is recorded rather than silently reinterpreted. | `docs/evidence/s13/lanes/informational.json` oracle FAIL with real actor receipt `SEMANTIC_ASSESSMENT_UNAVAILABLE ... exit=124, status=timeout`. | A real Paseo `opencode/opencode-go/mimo-v2.6-flash` assessor agent was observed launched for the fixture in a standalone attempt while the packed command had already timed out. | FIXED_AND_REAL_PROVIDER_VERIFIED (S13 round 2) |
| AEH-V2-0065 | P2 | Default AgentTopology workhorse model drift (S13) | The shipped default `workhorse` alias names `MiMo-V2.6-Flash`, which is absent from the local Paseo `opencode` catalog (the catalog id is `opencode-go/mimo-v2.6-flash`), so default-configuration participant journeys cannot launch here. | Case-sensitive model-id drift between the shipped preset and the locally available provider catalog. | `presets/agents/default.jsonc`, `presets/agents/orchestration.jsonc`, `docs/evidence/s13/s13-real-provider-certification.json` | S13 records the exact catalog ids observed through the Paseo daemon; the packed campaign fixture used a supported per-project `models.workhorse` override to exercise the real local model path. | No default topology change was made in S13 because the shipped alias is a product configuration decision; the campaign certifies the override explicitly and records the default-config blocker with a required user action. | S13 `toolchain` facts record the Paseo provider/model catalog observation without printing credentials. | S13 `informational` lane launched the real local model through the override before the unrelated INTENT deadline defect aborted the product command. | FIXED_AND_REAL_PROVIDER_VERIFIED (S13 round 2) |
| AEH-V2-0066 | P2 | Local provider provisioning drift (S13) | The previously recorded S10 Trivy binary path `/tmp/opencode/s10-tools/trivy/0.70.0/trivy` no longer exists, and `opengrep`, `podman`, `buildah`, `crun`, `opa`, and `cosign` are absent, so the S10 REAL_PROVIDER lane and OPA/OCI/signing lanes cannot be re-executed in this session. | User-local tool provisioning under `/tmp` is not persistent and no system packages provide these providers. | `docs/evidence/s13/s13-real-provider-certification.json`, `docs/CORE_ARCHITECTURE_V2_S13_WORKGRAPH.md` | S13 records the exact absent/present tool inventory (`toolchain.tools`) as evidence instead of substituting fixtures; affected capability rows are BLOCKED with exact required actions. | No product change; provider absence remains fail-closed (`ISOLATION_PROVIDER_UNAVAILABLE`, `CONTRACT_PROVIDER_UNAVAILABLE`, `SAST_EVIDENCE` blockers unchanged). | Tool inventory captured by bounded `command -v`-equivalent checks. | Not applicable; provider absent. | OPEN — provision or authorize providers before re-running S10/OPA/OCI/signing lanes |

| AEH-V2-0067 | P1 | Real Paseo semantic assessor completion (S13 round 2) | With the deadline fix, the packed `intent` command still failed: the Paseo wait never observed the completed turn (`SEMANTIC_ASSESSMENT_UNAVAILABLE exit=0, status=idle` with empty output), the SDK timeline fetch rejected `direction: "backward"` (`@getpaseo/client` 0.9.1 expects `tail`), the model reply was fenced JSON that `JSON.parse` rejected, the prompt did not include the assessment JSON Schema, and the topology `variant: max` launched max-reasoning turns for LIGHT assessments (2–6+ minutes). | The native wait was purely event-driven with no polling fallback and its idle-acceptance guard never fired when the provider emitted no updates; timeline fetches used a retired direction value and swallowed the error; the semantic runner required a bare JSON envelope; the assessor prompt named the schema without shipping it; the thinking option blindly inherited the topology variant. | `src/paseo/native.ts`, `src/paseo/sdk.ts`, `src/semantic/runtime.ts`, `src/semantic/assessment.ts`, `src/audit/intent.ts`, `src/core/triage.ts`, `src/issues/intake.ts`, `tests/paseoNative.test.ts`, `tests/paseoAtomicSdkTurn.test.ts`, `tests/semanticAssessment.test.ts`, `tests/semanticAssessmentRuntime.test.ts`, `tests/packed/s13RealProviderCertification.mjs`, `docs/evidence/s13/lanes/informational.json` | New regressions: canonical-snapshot polling resolves a completed turn with no subscription updates; fenced/prose-wrapped assessment payloads are extracted while non-JSON still fails; the reasoning-class bound maps LIGHT/STANDARD/DEEP to low/medium/high with the topology as the upper bound. Focused suites passed. | `semanticModelDeadlineMsV1 = 300_000` with the request schema ceiling raised to match; `waitForPaseoAgentHandle` polls the canonical snapshot every 2 s as a bounded fallback; timeline fetches use the supported `direction: "tail"`; `parseSemanticAssessmentOutputV1` extracts exactly one typed JSON object; the launch prompt carries `outputJsonSchema`; `boundSemanticThinkingOptionV1` caps thinking by the request reasoning class. No authority, acceptance, or routing semantics changed. | `npx vitest run tests/semanticAssessment.test.ts tests/semanticAssessmentRuntime.test.ts tests/paseoNative.test.ts tests/paseoAtomicSdkTurn.test.ts tests/auditIntent.test.ts` — all passed; `npm run typecheck` exit 0. | S13 round-2 `informational` lane ACCEPTED with a real Paseo session (`opencode-go/mimo-v2.6-flash`, `thinkingOptionId: "low"`, `status: idle`) and durable INTENT assessment cache. | FIXED_AND_REAL_PROVIDER_VERIFIED (S13 round 2) |

| AEH-V2-0068 | P0 | AUDIT detached-operation bootstrap policy (S13 round 3) | Every real `aeh operation start audit` failed before any reviewer turn: the materialized semantic supervisor was rejected with `EXECUTION_POLICY_STALE: durable ResolvedOperationPolicy does not match the current execution inputs.` | The controller bound the audit bootstrap policy as `NO_AGENT`/`NONE` while the audit TaskContract declares `DELEGATED`/risk-based assurance, so structured-result launch identity compilation could never match the frozen policy. | `src/operations/controller.ts`, `tests/governedOperationBootstrap.test.ts` | `tests/governedOperationBootstrap.test.ts` proves an AUDIT operation executes with a `DELEGATED`/`STANDARD` policy and matching intent before the audit runner starts. | Audit bootstrap binds the same route/assurance the audit contract declares (`DELEGATED`, `CRITICAL` for high risk else `STANDARD`) and patches the durable intent. The packed real-provider campaign then materialized and initialized a real `gpt-6-luna` supervisor and real reviewer sessions. | `npx vitest run tests/governedOperationBootstrap.test.ts` — 2 passed; full S13 focused set — 21 files / 144 tests passed; `npm run typecheck`; `npm run build`. | Packed governed-operation campaign: audit lane advanced through supervisor initialization, reviewer runs, validation and into consolidation against real Paseo sessions. | FIXED_AND_REAL_PROVIDER_EXERCISED (S13 round 3; consolidation residency defect recorded as AEH-V2-0074) |
| AEH-V2-0069 | P0 | Workspace candidate clears the frozen policy (S13 round 3) | Real packed CHANGE operations failed with `PASEO_PROVIDER_LEASE_IDENTITY_INCOMPLETE: current candidate, operation execution revision, and frozen policy are required.` | Binding the execution-worktree CandidateRevision clears `resolvedOperationPolicy`, and the controller did not rebind the bootstrap policy before participant launch. | `src/operations/controller.ts`, `tests/governedOperationBootstrap.test.ts` | `tests/governedOperationBootstrap.test.ts` proves the change runner observes a frozen policy bound to the current workspace candidate revision and digest. | After binding a workspace candidate the controller rebinds the bootstrap policy with the resolved route/assurance before any participant launch. | `npx vitest run tests/governedOperationBootstrap.test.ts` — 2 passed; `npm run typecheck`. | Packed change lane reached workspace creation, candidate binding, stack assessment and implementation staging with real sessions. | FIXED_AND_REAL_PROVIDER_EXERCISED (S13 round 3) |
| AEH-V2-0070 | P0 | Operation-bound semantic assessor lease (S13 round 3) | Real packed CHANGE operations failed with `PASEO_PROVIDER_LEASE_CONTEXT_MISMATCH: operation-bound Paseo work requires matching participant or bound Lead generation labels.` at the STACK assessment. | `launchManagedPaseoAgent` routed any prompt/operation/role launch through the managed foreground participant lifecycle, which requires a participant or Lead lease identity; the Semantic Assessor is intentionally not a participant. | `src/paseo/runtimeInitialTurn.ts`, `tests/managedForegroundAssessmentLease.test.ts` | `tests/managedForegroundAssessmentLease.test.ts` proves `aeh.kind=semantic-assessment` launches on the unleased assessor lifecycle while worker launches still materialize through the managed lifecycle. | Controller-side semantic assessments are explicitly excluded from the managed foreground-turn lifecycle; participant launches are unchanged. | `npx vitest run tests/managedForegroundAssessmentLease.test.ts` — 2 passed. | Packed change lane executed the real STACK assessor to completion after the fix. | FIXED_AND_REAL_PROVIDER_EXERCISED (S13 round 3) |
| AEH-V2-0071 | P1 | Structured-result contract-repair continuation identity (S13 round 3) | The bounded serialization repair turn inside a managed operation failed with the same provider-lease context mismatch before reaching the model. | The repair continuation called `continueManagedPaseoAgent` without the frozen execution identity labels, so the operation-bound lease could not be established. | `src/workers/agentPrompt.ts`, `tests/reviewerSerializationRecovery.test.ts` | The existing repair regression now asserts the repair turn carries `aeh.operation` and `aeh.participant` execution labels. | The contract-repair continuation propagates the same execution labels as the first turn. | `npx vitest run tests/reviewerSerializationRecovery.test.ts` — 6 passed. | Real supervisor initialization turn completed and triggered the repair path in the packed audit lane. | FIXED_AND_REAL_PROVIDER_EXERCISED (S13 round 3) |
| AEH-V2-0072 | P1 | Supervisor event skill seed completeness (S13 round 3) | Real audit consolidation failed with `SKILL_MANIFEST_INVALID: selected skill lacks procedure content: finding-dedup, audit-consolidation-protocol.` | `supervisorEventSkills` referenced `finding-dedup`, `audit-consolidation-protocol`, `verification-planning` and `recovery-classifier`, but the built-in skill seed contained none of them. | `src/participants/skills.ts`, `tests/supervisorPromptPolicy.test.ts` | New regression asserts every supervisor event skill resolves to seed procedure content. | Added cross-cutting seed skills with procedures and competencies for all four event skills. | `npx vitest run tests/supervisorPromptPolicy.test.ts` — 5 passed; focused S13 set — 144 passed. | Packed audit lane proceeded past consolidation skill compilation. | FIXED_AND_REAL_PROVIDER_EXERCISED (S13 round 3) |
| AEH-V2-0073 | P1 | Remaining semantic request deadlines (S13 round 3) | The packed change lane's STACK assessment timed out at 45 s (`SEMANTIC_ASSESSMENT_UNAVAILABLE ... exit=124, status=timeout`); candidate-impact and failure-classification constructors also hardcoded 45 s / 10 s. | Round 2 raised `semanticModelDeadlineMsV1` to 300 s but only the INTENT/ROUTE/ISSUE constructors consumed it; STACK, CANDIDATE_IMPACT and FAILURE still hardcoded provider-inappropriate deadlines. | `src/participants/stack.ts`, `src/candidates/assembler.ts`, `src/agents/recovery.ts` | The packed change lane reached implementation staging after the fix; existing semantic suites remain green. | All semantic request constructors now use the shared provider-appropriate `semanticModelDeadlineMsV1` ceiling (300 s). | Focused S13 set — 21 files / 144 tests passed; `npm run typecheck`; `npm run build`. | Real Paseo `opencode-go/mimo-v2.6-flash` STACK assessor completed within the repaired deadline. | FIXED_AND_REAL_PROVIDER_EXERCISED (S13 round 3) |
| AEH-V2-0074 | P0 | Supervisor event resume residency (S13 round 3, OPEN) | The packed audit lane reached consolidation and failed with `EXECUTION_BINDING_STALE: ContextManifest changed after session materialization.` while resuming the active supervisor session with the consolidation prompt. | `resolveExecutionBinding`/`assertResultProvenanceMatchesExecution` require a resumed BOUND session to reuse the exact ContextManifest/PromptManifest from materialization, so a persistent supervisor cannot execute a second semantic event with new context. | `src/workers/agentPrompt.ts`, `src/operations/supervisor.ts`, `docs/evidence/s13/governed-lanes/audit.json` | S13 records the exact failure and the blocked rows rather than reinterpreting the binding. | OPEN — no product change in S13 round 3; the required repair is a defined supervisor event/resume binding model (new participant generation per event, or an authorized context-continuation binding) that preserves candidate/policy/lease identity. | Packed audit lane durable trace shows `operation.supervisor.initialized` followed by the consolidation `EXECUTION_BINDING_STALE` failure. | Real supervisor/reviewer sessions executed; the operation remains non-accepted and the row is reported blocked. | OPEN — blocks audit, repair variants, context-handoff and the remaining FULL_OPERATION rows |
| AEH-V2-0075 | P1 | Packed candidate workspace stability and invalid semantic payloads (S13 round 3, OPEN) | One packed change run failed `CANDIDATE_WORKSPACE_MISMATCH` at the `executeIsolatedCandidateMutation` baseline assert although the archived worktree later recomputed to the exact bound r2 digest; a later run failed `STACK_ASSESSMENT_INVALID: STACK project skill root 'AGENTS.md' does not resolve to an existing directory.` | A transient non-ignored workspace mutation between candidate binding and baseline assertion (provider/session scratch is the leading hypothesis) changed the digest; the semantic service has no bounded retry for schema-valid-but-semantically-invalid model payloads. | `src/candidates/direct.ts`, `src/participants/stack.ts`, `docs/evidence/s13/governed-lanes/change-direct.json` | S13 records both exact failures; no compensating ignore rules or retry fakes were added. | SUPERSEDED (round 5) — split into AEH-V2-0081 (provider-scratch exclusion from candidate identity and ChangeSet scope) and AEH-V2-0083 (bounded non-authoritative payload retry). No separate work remains under 0075. | Digest recomputation of the retained failed worktree matched the bound r2 digest exactly; both failure records are preserved in the campaign evidence. | Real stack assessor executed; change row remains non-accepted. | SUPERSEDED by AEH-V2-0081 + AEH-V2-0083 (round 5 reconciliation) |
| AEH-V2-0076 | P0 | `aeh issue import` outside a managed operation (S13 round 3; fixed in round 5, see the merged AEH-V2-0076 row) | The real GitHub effect campaign's `aeh issue import` failed with `ISSUE_NORMALIZATION_BLOCKED: canonical Planner failed` whose cause is `V2_AUTHORITY_REQUIRED: bounded worker execution requires a managed operation.` | `normalizeIssueWithPlanner` launches the Planner through `executeAgentPrompt` with `requireExecutionAuthority: true`, but the CLI provides no managed operation context; the operation-bound import path is not wired. | `src/issues/intake.ts`, `src/cli.ts`, `docs/evidence/s13/github-lanes/issue-driven.json` | The deterministic half (real issue snapshot digest and the drift gate) is certified; the operation-bound import is reported blocked with the exact error. | OPEN — required repair is to route issue import through the managed operation/controller path (or provide controller-issued authority) before planner launch. | Real `gh` issue #1 snapshot digest and `ISSUE_DRIFT` rejection captured against the disposable private repository. | Real GitHub snapshot + drift gate PASS; import blocked; `issue-driven` reported PARTIAL. | SUPERSEDED_BY_ROUND_5_FIX (historical round-3 checkpoint; see the merged AEH-V2-0076 row) |
| AEH-V2-0077 | P1 (historical) | GitHub deletion scope (S13 round 3; superseded by Round-10 user disposition) | The Round-3 campaign's deletion attempt failed HTTP 403 because the authenticated token lacked `delete_repo`. | GitHub deletion requires a scope absent from the campaign credential; Round-10 disposition makes manual user cleanup non-blocking. | `docs/evidence/s13/s13-github-effects.json` | Exact 403 remains historical evidence; no token or credential was persisted. | Round 10: `USER_CLEANUP_NON_BLOCKING`; user will delete the private disposable repo manually. Do not retry deletion, request `delete_repo`, broaden credentials, or block S13 solely on repository existence. | Round-3 deletion failure preserved; Round-10 created no repository and made no deletion request. | The previously created private repository may remain until user cleanup. | USER_CLEANUP_NON_BLOCKING (Round-10 user disposition) |
| AEH-V2-0078 | P0 | Supervisor event resume binding (S13 round 4) | `consolidateWithOperationSupervisor` could not resume the initialized supervisor session: `EXECUTION_BINDING_STALE: ContextManifest changed after session materialization`. | `assertResultProvenanceMatchesExecution` required a resumed BOUND session to reuse the exact materialization manifests and participant identity, so a persistent supervisor could not execute a second semantic event with new context. | `src/workers/agentPrompt.ts`, `src/operations/supervisor.ts`, `tests/supervisorEventContinuation.test.ts` | Explicit `continueBoundSession` turns keep the generation's frozen binding and patch event evidence; a continuation with a changed participant, candidate, execution revision, generation, blueprint, policy, epoch or runtime session is still rejected. | Added a continuation mode that skips only the turn-scoped manifest equality checks while keeping every binding identity check, plus a durable delivered-turn record (session, participant, event, prompt digest, binding digest) in the consolidation artifact. | `npx vitest run tests/supervisorEventContinuation.test.ts` — 4 passed; packed audit lane reached consolidation with real sessions. | `docs/evidence/s13/governed-lanes/audit.json` shows the audit lane passing supervisor initialization, reviewer runs and validation before the remaining round-4 stall (AEH-V2-0082). | FIXED_AND_PACKED_EXERCISED (S13 round 4) |
| AEH-V2-0079 | P0 | Provider session workspace identity for isolated mutation (S13 round 4) | The DIRECT implementer ran inside the operation's bound candidate worktree instead of the disposable isolated tree, so its edits escaped the isolation boundary and assembly failed on a mutated base candidate. | `compilePaseoAgentLaunchSpec` always set `workspaceId` from the operation, and Paseo bind the session to the workspace cwd when a workspace id is present. | `src/paseo/launchSpec.ts`, `tests/paseoLaunchSpec.test.ts` | A session only claims the operation workspace when the launch root equals the durable operation workspace root; isolated launches carry no `aeh.workspace.id` and run in the isolated cwd. | `npx vitest run tests/paseoLaunchSpec.test.ts` — 6 passed (round 5 adds the unloadable-operation regression). | Unit-exercised only. The round-4 packed `change-direct.json` records 0 receipts and no assembled ChangeSet, so no packed evidence supports a workspace-isolation claim; round-5 rerun is the evidence gate. | FIXED_UNIT_REGRESSED (round 4); packed evidence pending round-5 rerun |
| AEH-V2-0080 | P1 | Operation artifact root for isolated DIRECT runs (S13 round 4) | `finalizeOperationSession` rejected a completed isolated participant with `V2_RECEIPT_REJECTED: participant artifact ... could not be read`. | Transcripts were persisted relative to the isolated execution root but accepted structured results are stored under the operation state root, so the receipt path resolved in the wrong tree. | `src/workers/agentPrompt.ts`, `tests/reviewerSerializationRecovery.test.ts` | `operationArtifactRoot(root)` resolves the control root when present; transcripts and receipts use it consistently. | Exported helper regression plus the existing reviewer serialization suite — round 5 adds the isolated state-root receipt persistence regression (8 tests). | The round-4 `change-direct.json` records 0 participant receipts, so the earlier packed claim was not supported; round 5 completes the fix by routing every finalization operation-state read/write through the authoritative state root and binding receipts to the controller-issued participant identity. | FIXED_INCOMPLETE_IN_ROUND_4 (R4-F3) — COMPLETED_IN_ROUND_5; packed evidence pending rerun |
| AEH-V2-0081 | P1 | Provider-generated scratch and candidate identity (S13 round 4) | Serena/Graphify scratch created inside the workspace changed the candidate digest between binding and the DIRECT baseline assert, and provider scratch appeared as ChangeSet files that escaped scope (`PARTICIPANT_PLAN_INVALID`). | Candidate/repository digests and the isolated DIRECT diff enumerated all untracked non-ignored files, including provider-owned scratch. | `src/core/git.ts`, `src/candidates/direct.ts`, `src/candidates/identity.ts`, `tests/candidateIdentity.test.ts`, `tests/directCandidateLifecycle.test.ts` | Untracked `.serena`/`graphify-out` scratch is excluded from candidate/repository identity and from ChangeSet pathspecs; tracked files always contribute, and mismatch errors now carry a bounded digest-input inventory. | `npx vitest run tests/candidateIdentity.test.ts tests/directCandidateLifecycle.test.ts` — 11 passed in round 5 after adding the legitimate-untracked-source, unknown-untracked-file, intent-to-add and DIRECT ChangeSet regressions. | The round-4 artifact contains no scratch-mismatch error but also no assembled ChangeSet, so the packed claim is limited to "no scratch escape recorded"; the intent-to-add filter hardening is new in round 5. | FIXED_AND_UNIT_HARDENED (round 4 + round 5); packed evidence pending rerun |
| AEH-V2-0082 | P0 | Audit supervisor event stall (S13 round 4, OPEN) | The packed audit lane passes reviewer/validation work but never reaches a terminal report: the supervisor session is repeatedly woken by the watchdog while the consolidation event is in flight, and the operation stalls until the 15-minute lane budget. | Watchdog wake policy does not treat an in-flight supervisor semantic event as a reason to suppress or serialize wake dispatches; the consolidation wait is bounded only by the worker timeout. | `src/operations/controller.ts`, `src/operations/state.ts`, `src/workers/agentPrompt.ts`, `src/workers/resultGateway.ts`, `src/paseo/runtimeCore.ts`, `tests/governedOperationBootstrap.test.ts` | Round-5 reproduction with `S13_GOV_KEEP=1`, per-stage timestamps and `operation.watchdog.*` trace capture established the owning defect: the audit report was persisted and the consolidation stage completed, but the detached controller exited without a terminal state. Root cause: `return terminalizeOperation(...)` without `await` bypassed the controller catch, so the successful-terminalization gate rejection (`V2_TERMINAL_GATE_REJECTED: participant participant:<digest> has no terminal receipt`) killed the process and left the operation RUNNING; the receipt gate also rejected legitimate persistent-supervision and controller-authority identities. Repairs: `await` on every terminal branch; bounded-work receipts bind to the controller-issued participant identity while the Paseo session id remains receipt provenance; persistent supervisor generations and read-only Semantic Assessor registrations are exempt from the bounded-work receipt loop; Semantic Assessor launches no longer register work participants. | `tests/governedOperationBootstrap.test.ts` — successful audit with a bounded receipt plus exempt identities, and gate rejection becomes durable FAILED (5 tests). | Round-5 `governed-lanes/audit.json` rerun with `S13_GOV_KEEP=1` trace capture and a terminal audit report. | FIXED_AND_PACKED_EXERCISED (S13 round 5) — packed audit terminal SUCCEEDED, oracle PASS, `checkoutUntouched: true` (`governed-lanes/audit.json` SHA-256 `84517c3dd6445f9a…`) |
| AEH-V2-0083 | P1 | Semantic assessor completion and bounded retries (S13 round 4) | Real assessor turns returned empty/unparseable replies (`SEMANTIC_ASSESSMENT_UNAVAILABLE ... structured result`), schema-invalid payloads (missing `judgment.unknowns`, invented evidence refs, wrong judgment type), and one hung 300 s turn. | The assessor launch path did not expose the completed OpenCode `assistant_message` timeline text; the semantic service had no bounded retry for invalid/hung non-authoritative output and no corrective re-ask. | `src/paseo/sdk.ts`, `src/paseo/runtimeInitialTurn.ts`, `src/paseo/runtimeCore.ts`, `src/semantic/assessment.ts`, `src/semantic/runtime.ts`, `src/participants/stack.ts`, `tests/semanticAssessment.test.ts`, `tests/paseoAtomicSdkTurn.test.ts`, `tests/managedForegroundAssessmentLease.test.ts`, `tests/participantsStackDiscovery.test.ts` | Assessor turns run as managed atomic turns without a writer lease, capture the reply via the canonical timeline fallback, and retry once with a bounded repair note naming the expected judgment type and allowed evidence refs; a hung timeout retries once; other provider/authority failures still fail closed. | `npx vitest run tests/semanticAssessment.test.ts tests/paseoAtomicSdkTurn.test.ts tests/managedForegroundAssessmentLease.test.ts tests/participantsStackDiscovery.test.ts` — 52 passed. | Packed change lane completed real ROUTE, STACK and CANDIDATE_IMPACT assessments; residual model-output flakiness remains a real-provider observation. | FIXED_AND_PACKED_EXERCISED (S13 round 4); residual model-output flakiness remains reported per lane |
| AEH-V2-0084 | P0 | Controller terminalization escape (S13 round 5, FIXED) | The packed audit lane persisted its report and completed consolidation but never reached a terminal operation state; the detached controller exited with no terminal write and no error trace. | `return terminalizeOperation(...)` without `await` in the audit/change/run branches bypassed the controller `catch`, so any rejection inside terminalization (for example the successful-terminalization gate) became an unhandled top-level rejection and killed the detached process. | `src/operations/controller.ts`, `tests/governedOperationBootstrap.test.ts` | A terminalization gate rejection now produces a durable `FAILED` terminal state with the exact gate error instead of a silent exit. | `tests/governedOperationBootstrap.test.ts` — "converts a terminalization gate rejection into a durable FAILED instead of escaping the controller"; packed audit rerun reached terminal `SUCCEEDED`. | Packed `governed-lanes/audit.json` (round 5) terminal `SUCCEEDED` with trace capture. | FIXED_AND_PACKED_EXERCISED (S13 round 5) |
| AEH-V2-0085 | P0 | Bounded-work participant identity split and read-only registration (S13 round 5, FIXED) | The successful-terminalization gate required a receipt for every registered participant, but one bounded work unit could be registered twice (controller-issued authority identity and Paseo session identity), and read-only Semantic Assessor turns registered as work participants; the packed audit therefore could not terminalize SUCCEEDED. | Finalization wrote the receipt against the runtime session id while the execution binding and authority lived on the deterministic participant id; `materializeAgentPrompt`, the result gateway and session-running markers upserted session participants; managed assessor launches registered work participants. | `src/operations/state.ts`, `src/workers/agentPrompt.ts`, `src/workers/resultGateway.ts`, `src/paseo/runtimeCore.ts`, `tests/reviewerSerializationRecovery.test.ts`, `tests/governedOperationBootstrap.test.ts` | One durable bounded-work identity per participant (controller-issued id when present, Paseo session id carried as the receipt `sessionId`); registered-only updates never upsert; persistent supervisor generations and Semantic Assessor identities are exempt from the bounded-work receipt loop; authority-bound and semantic-assessment launches no longer register second work participants. | `tests/reviewerSerializationRecovery.test.ts` (8, including the isolated state-root receipt regression); `tests/governedOperationBootstrap.test.ts` (audit terminalizes with one bounded receipt plus exempt identities). | Packed round-5 audit record shows exactly one bounded participant with a candidate-bound receipt and a DRAINING supervisor generation at terminal. | FIXED_AND_PACKED_EXERCISED (S13 round 5) |
| AEH-V2-0086 | P1 | Intent-to-add provider-scratch smuggling and missing untracked-identity regressions (S13 round 5, FIXED) | The provider-scratch filter covered only the untracked list; `git add -N` entries surface under `git ls-files --cached`, so provider scratch could be smuggled into a candidate digest, and the reviewer-mandated negative regressions were missing. | Unconditional untracked prefix filtering without intent-to-add classification. | `src/core/git.ts`, `tests/candidateIdentity.test.ts`, `tests/directCandidateLifecycle.test.ts` | Deterministic porcelain-v2 intent-to-add classification excludes provider-generated intent-to-add entries while keeping tracked provider files and legitimate intent-to-add source files contributing; unknown untracked files are never silently ignored. | `tests/candidateIdentity.test.ts` (5) and `tests/directCandidateLifecycle.test.ts` (5). | Not applicable (unit lane). | FIXED_AND_UNIT_REGRESSED (S13 round 5) |
| AEH-V2-0087 | P1 | Fail-open operation-workspace guard (S13 round 5, FIXED) | `compilePaseoAgentLaunchSpec` claimed the operation workspace when the durable operation could not be loaded (`!durableWorkspaceRoot`), so an unprovable operation identity could bind a session to a workspace it did not own. | Fail-open boolean plus `path.resolve`-only comparison. | `src/paseo/launchSpec.ts`, `tests/paseoLaunchSpec.test.ts` | A workspace is claimed only under positive proof: a durable operation whose canonical resolved workspace root equals the canonical resolved launch root; otherwise no workspace binding. | `tests/paseoLaunchSpec.test.ts` — unloadable-operation regression (6 tests). | Not applicable (unit lane). | FIXED_AND_UNIT_REGRESSED (S13 round 5) |
| AEH-V2-0088 | P1 | Structured-output boundary hardening (S13 round 5, FIXED) | Hung-turn retry used a message regex; completion extraction could collect nested tool payloads as assistant text; timeline refetch errors were swallowed; `direction: "tail"` was unverified. | Missing typed timeout classification and canonical assistant-message selection. | `src/semantic/runtime.ts`, `src/semantic/assessment.ts`, `src/paseo/sdk.ts`, `tests/semanticAssessment.test.ts`, `tests/semanticAssessmentRuntime.test.ts`, `tests/paseoAtomicSdkTurn.test.ts` | Typed `details.timeout` classification from the runner drives the single bounded retry; `extractLastAssistantText` selects only the canonical last assistant entry's top-level text/content; timeline refetch failures are traced; `direction: "tail"` verified against the installed `@getpaseo/protocol@0.9.1` enum (`tail`/`before`/`after`); schema validation remains authoritative and retries stay at 2. | `tests/semanticAssessment.test.ts` (15), `tests/semanticAssessmentRuntime.test.ts` (5), `tests/paseoAtomicSdkTurn.test.ts` (8). | Not applicable (unit lane). | FIXED_AND_UNIT_REGRESSED (S13 round 5) |
| AEH-V2-0089 | P0 | Resumable launch identity context closure (S13 round 5, FIXED) | The full suite failed `tests/changeControllerOwnedDecision.test.ts`: a controller-owned product-choice suspension/resume could not launch its next turn with `CONTEXT_RUNTIME_V2_BINDING_REJECTED: context access is closed for a terminal participant execution.` | The first round-5 receipt-identity model recorded the receipt against the controller-issued launch identity and `recordParticipantReceipt` flipped that participant to COMPLETED, closing the context runtime execution that a later resumed turn reuses. | `src/operations/state.ts`, `src/workers/agentPrompt.ts`, `tests/changeControllerOwnedDecision.test.ts`, `tests/reviewerSerializationRecovery.test.ts` | `recordParticipantReceipt` no longer flips `participant:<digest>` launch identities to a terminal status (per-turn terminal status stays on the runtime session participant); the successful-terminal gate accepts a receipt bound by `participantId` or by the runtime `sessionId`; finalization keeps the runtime session participant lifecycle and stores the launch identity as the receipt participant with the session id as provenance. | `tests/changeControllerOwnedDecision.test.ts` passes in isolation (9s) and in the 16-file focused set (153 tests). | Full-suite run that exposed the regression (`27 failed / 1273 passed / 7 skipped`) followed by focused re-runs green. | FIXED_AND_UNIT_REGRESSED (S13 round 5) |
| AEH-V2-0076 | P0 | Managed issue import outside a managed operation (S13 rounds 3–6; merged row) | The real GitHub effect campaign's `aeh issue import` failed with `ISSUE_NORMALIZATION_BLOCKED: canonical Planner failed` (`V2_AUTHORITY_REQUIRED: bounded worker execution requires a managed operation.`). | `normalizeIssueWithPlanner` launched the Planner through `executeAgentPrompt` with `requireExecutionAuthority: true` while the CLI provided no managed operation context. | `src/issues/intake.ts`, `src/issues/workflow.ts`, `src/operations/controller.ts`, `src/operations/state.ts`, `src/cli.ts`, `tests/governedOperationBootstrap.test.ts` | Managed controller intake with controller-issued Planner authority; deterministic intake terminal gate; explicit candidate-lineage advancement. | `aeh issue import` now runs inside a controller-owned managed change operation (`ChangeOperationPayload.issueIntake`): the controller binds a deterministic DELEGATED/STANDARD normalization policy against the bound candidate and epoch, the canonical Planner launches with controller-issued `ExecutionAuthorityV1`, the authored contract and frozen snapshot persist under the control root, and a deterministic terminal gate verifies the contract digest and the bounded Planner receipt. No `requireExecutionAuthority:false`, ambient authority, or GitHub special-casing. | `tests/governedOperationBootstrap.test.ts` managed-intake regression. | Round-6 real-provider verification against the current build in the authorized disposable repository stable clone: terminal `SUCCEEDED` (`phase=issue-intake-finished`), bounded Planner receipt `participant:c5e8b2888a5fb145`/session `73ff250b-…`, sealed `GH-1` contract digest `a46bccc0…`, frozen snapshot `bd777eed…` (`docs/evidence/s13/issue-import-round6.json`, SHA-256 `de71fac2982856ad…`). The semantic route is model-selected (round 5 `FORMAL_SDD`, round 6 `DELEGATED`). | FIXED_AND_REAL_PROVIDER_VERIFIED (round 6) |
| AEH-V2-0090 | P1 | CandidateImpact review-dimension contract (S13 round 6, FIXED) | The real packed change lane blocked in candidate-assurance recompilation before any participant launched: `UNRESOLVED_VALIDATION_REQUIREMENT: impact-review-absence-of-accompanying-tests …`, because the MODEL `CANDIDATE_IMPACT` judgment accepted free-form `reviewDimensions` that the deterministic compiler mapped through a default `unit-test` rule into validation requirements no project could anticipatorily provision; assurance was BLOCKED, so the reviewers that the AcceptanceOracle then required never ran. | `semanticJudgmentSchema.reviewDimensions` was `z.array(z.string())`; `reviewDimensionRule` fabricated a concrete `unit-test` obligation for every unknown string. | `src/semantic/assessment.ts`, `src/architecture/candidateAssurance.ts`, `tests/packed/s13GovernedOperationCampaign.mjs` | `reviewDimensions` is constrained to the canonical typed vocabulary (`candidateReviewDimensionValues`, each with one deterministic validation-kind mapping and assurance floor); the assessment contract cannot invent a validation kind, and canonical dimensions without a provider still fail closed. The campaign fixture also provisions real `test`/`contract-test` validations. | `tests/semanticAssessment.test.ts` — non-canonical dimension rejected, every canonical dimension accepted; `tests/candidateAssurance.test.ts` unchanged green. | Round-6 packed change lane: the typed dimensions (`behavior.correctness`, `public API`) resolved and their real configured validations PASSed with compilation `READY` and one compiled Reviewer assignment (observed in the live run; the durable final artifact is the provider-limitation run). | FIXED_AND_UNIT_REGRESSED (S13 round 6) |
| AEH-V2-0091 | P1 | Deterministic contract requirement→validator traceability (S13 round 6, FIXED) | After AEH-V2-0090, the packed change lane failed `evidence.requirement-coverage`: `AC-1 lacks PASS evidence for: a declared validator.` | `createRoutedContract` emitted `validators: []` for every acceptance-derived requirement, so the evidence-completeness gate could never be satisfied for a deterministic DIRECT contract. | `src/core/contract.ts`, `tests/changeRouting.test.ts` | Requirements that do not declare their own approved validators are deterministically bound to the project's configured validation command and validator ids (the model cannot select commands; an unconfigured project keeps the fail-closed empty set). | `tests/changeRouting.test.ts` — configured validators bound; explicit requirement validators preserved. | Round-6 packed change lane: the evidence graph then passed (`evidence.requirement-coverage` absent from the failure set) after the repair (live-run observation). | FIXED_AND_UNIT_REGRESSED (S13 round 6) |
| AEH-V2-0092 | P1 | Provider-enforced assessor structured output, bounded retry and failure diagnostics (S13 round 6; round-5 R5-F3) | The operation-scoped semantic assessor ran through materialize+send with `outputSchema` stripped, so the installed Paseo/OpenCode stack never applied provider-enforced JSON output; STACK composed up to four real model turns and rejected replies were not captured. | `src/paseo/runtimeInitialTurn.ts` stripped `outputSchema` for `aeh.kind=semantic-assessment`; the semantic service and the STACK correction loop each owned a retry. | `src/paseo/runtimeInitialTurn.ts`, `src/semantic/assessment.ts`, `src/semantic/runtime.ts`, `src/participants/stack.ts` | Assessor turns use the session-creating provider-enforced channel (assessment prompt as initial prompt plus the unchanged schema) with no AEH result sink/MCP/lease/participant registration; the service exposes a bounded `attemptBudget` and STACK caps each call at one turn (total ≤2, single owner); rejected replies persist bounded `lengthBytes`+`sha256`+200-char fingerprint and assessor session id in failure details plus a `semantic.assessor.reply.rejected` trace; extraction adds a bounded stray-trailing-brace fallback without repairing fields. | `tests/managedForegroundAssessmentLease.test.ts`, `tests/semanticAssessment.test.ts`, `tests/semanticAssessmentRuntime.test.ts`, `tests/participantsStackDiscovery.test.ts` (27). | Round-6 real providers: `audit` ACCEPTED (real `opencode/opencode-go/mimo-v2.6-flash` reviewer session), certification `informational` ACCEPTED (schema-conformant INTENT assessment persisted through the repaired creation path). The provisional `REAL_PROVIDER_CONTRACT_LIMITATION` (malformed mimo JSON) was withdrawn by the round-6 amendment (AEH-V2-0096/0097: permission projection + uncertified model); the post-amendment rerun reached implementation/review and now fails on AEH-V2-0099/0100 deterministic wiring. | FIXED_AND_REAL_PROVIDER_EXERCISED (S13 round 6); provisional `REAL_PROVIDER_CONTRACT_LIMITATION` withdrawn (review-round6 R6-F6) |
| AEH-V2-0093 | P1 | Clean-build runtime asset resolution (S13 round 6; round-5 R5-F2) | A clean build had no `dist/releases/templates`, so `init` failed `ENOENT`; the checkout only worked because stale pre-v2 shared assets from 2026-09-22 remained, and `npm pack` shipped them inside every packed candidate. | `src/core/init.ts` resolved `templates/*` two levels above the compiled module (`dist/releases/templates`) instead of the running release root. | `src/core/init.ts`, `tests/buildHygiene.test.ts`, `scripts/build.mjs` (unchanged asset publication) | `init` resolves runtime assets through the release-aware `PACKAGE_ROOT`; a clean build and packed candidates use the release's own assets. The stale shared `dist/releases/{templates,presets,policies,schemas,maturity,skills,scripts,package.json}` leftovers were removed. | `tests/buildHygiene.test.ts` — fresh-build `init` with shared leftovers deleted copies the release's own template digests; packed `init` without shared leftovers uses the packed release's assets; a missing release template still fails closed. Pre-fix reproduction: exact ENOENT `dist/releases/templates/project.yaml`. | Packed certification lanes and every round-6 `init` executed from the release's own assets. | FIXED_AND_UNIT_REGRESSED (S13 round 6) |
| AEH-V2-0094 | P3 | Racy descendant-cancel test fixture (S13 round 6; round-5 R5-F5 TEST_DEFECT) | `tests/operations.test.ts` "cancels detached descendants that are outside the controller process group" failed once per 19 current executions (0/19 at base): the fixture spawned the descendant writer before `seed()` bound the candidate, so the writer's `descendant.pid` raced `assertWorkspaceSourceDigest`. | The pid file was written inside the candidate digest root. | `tests/operations.test.ts` | The fixture writes the descendant pid file outside the digest root; the product gate is unchanged. | Descendant-cancel test 8/8 isolated executions green; `tests/operations.test.ts` fully green. | Not applicable (test-defect repair; the product gate correctly failed closed on the mutated workspace before). | FIXED_AND_UNIT_REGRESSED (S13 round 6) |
| AEH-V2-0095 | P3 | Truthfulness/harness reconciliation defects (S13 round 6; round-5 R5-F1/F4/F5/F6/F7/F10) | Round-5 GitHub artifacts were rewritten while the docs cited round-3 digests and called `authority` "UNCHANGED"; the governed-campaign issue import recorded a 420 s null empty-output timeout; the ledger carried contradictory `AEH-V2-0076` rows; the "earlier round-5 DELEGATED run" and the cleanup claims were not evidence-bound; the full-suite counts/classifications were inaccurate. | Last-invocation-only summaries, stale digest citations, KEEP=1 cleanup skip, unmerged ledger rows. | `docs/CORE_ARCHITECTURE_V2_S13_WORKGRAPH.md`, `docs/CORE_ARCHITECTURE_V2_STATUS.md`, `docs/CORE_ARCHITECTURE_V2_CONFORMANCE.md`, `docs/ENGINEERING_LEDGER.md`, `tests/packed/s13GithubEffectCampaign.mjs`, `tests/packed/s13GovernedOperationCampaign.mjs`, `tests/packed/s13RealPaseoLifecycleCampaign.mjs` | Docs corrected in place with round-6 digests and explicit withdrawals of unsupported claims; the ledger `AEH-V2-0076` rows merged into one truthful row; the GitHub import bound raised 420 s→900 s and every campaign round stamp parameterized (`S13_ROUND`, default 6); round-5 workspaces `wks_1a8379328b475b5e`, `wks_effa47345587b823`, `wks_7fcfc35e1d1bdcd5` plus all round-6 lane workspaces and `/tmp` staging archived/removed. | Round-6 reruns (`authority`/`issue-driven` digests above; managed import `SUCCEEDED`); cleanup verified by an empty S13 workspace query. | Re-run rows recorded with current build-bound digests; no credential persisted. | FIXED_AND_RE_RECORDED (S13 round 6) |
| AEH-V2-0096 | P1 | Semantic Assessor permission projection suppressed provider structured output (S13 round 6 amendment, FIXED) | Real AEH-mediated assessor turns never reached OpenCode's structured channel: with the wildcard deny the model wrote free text (or the turn stalled to its 300 s deadline) and the operation failed at `SEMANTIC_ASSESSMENT_INVALID`. | `buildOpenCodePermission` set `permission['*']='deny'` for the Semantic Assessor; OpenCode implements `format: json_schema` through an injected `StructuredOutput` tool, which the wildcard also denied (case-sensitive key). | `src/agents/permissions.ts`, `tests/semanticStructuredOutputCapability.test.ts` | `permission.StructuredOutput='allow'` under the retained repository/tool denials (probe: wildcard-deny fails; wildcard-deny + `StructuredOutput:'allow'` succeeds canonically; lowercase fails; explicit tool-family denies also succeed). | Permission-projection regression asserts `{'*':'deny', StructuredOutput:'allow'}` for the assessor and all real tools denied. | Durable real-provider evidence: the wildcard permission probe (4 variants) and the per-model probe matrix in `docs/evidence/s13/structured-output-probe.json` (`a72260fb…`) show wildcard-deny failing, wildcard-deny + `StructuredOutput:"allow"` succeeding canonically, lowercase keys failing, and the certified models producing schema-valid canonical output. Caveat (review-round6 R6-F1): the AEH-mediated post-fix probe (11.3 s, typed STACK judgment, digest `5afdd632…`, session `51014e17…`) is docs-only and not preserved as evidence; the post-amendment change capture's real `opencode-go/gpt-6-luna` assessor preflights corroborate the repaired path. | FIXED_AND_REAL_PROVIDER_PROBE_VERIFIED (S13 round 6 amendment; AEH-mediated probe record not preserved) |
| AEH-V2-0097 | P1 | Runtime-level structured-output flag treated as per-model proof (S13 round 6 amendment, FIXED) | The assessor shipped on `opencode-go/mimo-v2.6-flash`, which even with the channel open captured schema-invalid structured arguments (judgment as an encoded string, 3/3 full-schema trials), so assessments failed or stalled regardless of retries. | `resolveSemanticAssessor` required `runtimeCapabilities.structuredOutput === true` (runtime-level) with no per-model certification. | `src/semantic/structuredOutput.ts`, `src/semantic/assessment.ts`, `presets/agents/default.jsonc`, `tests/semanticStructuredOutputCapability.test.ts` | Deterministic certified capability registry with canonical levels (`TEXTUAL_JSON < JSON_OBJECT < SCHEMA_BOUND_TOOL_CALL < JSON_SCHEMA_VALIDATED < JSON_SCHEMA_CONSTRAINED`), required `SCHEMA_BOUND_TOOL_CALL`, probe-verified certified models (`opencode-go/gpt-6-luna`, `deepseek-v4.1-flash`, `kimi-k3`, `glm-5.3`), shipped assessor default `@structured-assessor` = `opencode-go/gpt-6-luna`, and fail-closed resolution that refuses uncertified models without silent fallback. | Registry/level tests, fail-closed selection tests for mimo and unknown models, certified default resolution. | Bounded probe matrix (`docs/evidence/s13/structured-output-probe.json`): gpt-6-luna 4/4 canonical schema-valid, deepseek/kimi 2/2, glm 2/2 (slow), mimo 0/3 schema-valid. | FIXED_AND_REAL_PROVIDER_VERIFIED (S13 round 6 amendment); mimo-only `REAL_PROVIDER_CONTRACT_LIMITATION` documented |
| AEH-V2-0098 | P2 | Semantic Assessor output discipline (S13 round 6 amendment, FIXED) | Probe case C showed a compact discipline materially improves structured tool-call compliance (mimo 2/3 vs 1/3), and the assessor prompt lacked the canonical discipline clauses. | The assessor system prompt did not state complete required keys, empty-array/empty-record conventions, discriminator discipline, or nested-object encoding. | `src/semantic/runtime.ts`, `tests/semanticStructuredOutputCapability.test.ts` | Canonical discipline added to `SEMANTIC_ASSESSOR_SYSTEM_PROMPT` (exactly one JSON object, required keys, `[]`/`{}`, only supplied evidenceRefs, exact discriminator, nested objects never encoded strings, no comments/trailing commas, concise). Error reduction only; deterministic schema/evidence validation remains the gate. | Prompt-discipline clause regression. | Probe case C evidence plus capable-model verification (discipline does not regress certified models). | FIXED_AND_UNIT_REGRESSED (S13 round 6 amendment) |
| AEH-V2-0099 | P1 | Base contract assertions lack a resolved validation path in the AcceptanceOracle (S13 round 6 amendment discovery, OPEN) | After the structured-output repairs, the rerun reached the AcceptanceOracle and rejected with `VERIFICATION_VALIDATION_PATH_MISSING: assertion 'AC-1' has no resolved candidate-bound validation requirement` even though `command.fixture-greeting` and `contract-test` checks PASSed and `evidence.requirement-coverage` PASSed. | `resolveVerificationRequirementsV1` maps an assertion to `compilation.validationRequirements` by requirement refs/acceptanceRefs; for DIRECT the compilation's base validation requirements are empty (`waveResult?.plan?.validationRequirements ?? []`) and generated impact requirements never reference the contract assertion, so AC-1 resolves to no validation requirement. | `src/architecture/validationRequirements.ts`, `src/core/run.ts` | ROUND 7 FIXED: `contractValidationRequirementsV1` compiles each frozen contract requirement's declared validators (configured `command.<id>` commands and configured validator ids) into exactly one candidate-bound ValidationRequirement whose id is the deterministic check id and whose refs name every declaring requirement; `resolveConfigured` recognizes the `command.<id>` convention; the executed `candidate.assurance.validation.*` checks are merged into the report on every assurance recompilation. Unknown check ids are skipped and id collisions fail closed; no oracle gate changed. | `tests/acceptanceValidationPath.test.ts` (8) + the `runTask` producer regression in `tests/runRepairCandidateLifecycle.test.ts`. | Round-7 packed lane run 6: AcceptanceOracle `ACCEPTED` with `AC-1` covered and persisted validation items for `command.fixture-greeting` and `contract-test`; round-6 late terminal record `ffb42c0e838a7a74…` preserved as the pre-fix blocker. | FIXED_AND_REAL_PROVIDER_EXERCISED (S13 round 7); terminal `SUCCEEDED` blocked by AEH-V2-0106 |
| AEH-V2-0100 | P1 | Reviewer evidence matcher requires binding and artifact on one participant record (S13 round 6 amendment discovery, OPEN) | The same rerun's review lifecycle PASSed with a real independent Reviewer (`candidate.assurance.reviewer.0.reviewer` PASS, reviewer session COMPLETED with a persisted result artifact and a SUCCEEDED reviewer receipt), yet the AcceptanceOracle reported 0 independent reviewers for every dimension. | `currentReviewEvidence` requires a single participant record that simultaneously holds `executionBinding.runtime.sessionId` and `resultArtifact`; finalization stores the execution binding on the controller-issued authority identity (`participant:<digest>`) and the result artifact on the runtime session participant. | `src/operations/state.ts` | ROUND 7 FIXED: `recordParticipantReceipt` persists the receipt's accepted result artifact on the receipt participant identity (the controller-issued `participant:<digest>`) while keeping launch identities non-terminal and the runtime session id as receipt provenance; the oracle's single-record predicate is unchanged and now reads one coherent chain. | `tests/acceptanceReviewProvenance.test.ts` (7: producer persistence + accepted oracle, split-record rejection, session-only rejection, stale-candidate rejection, resumed/rebound identity preservation, multi-round selection). | Round-7 packed lane runs 1/4/6: reviewer identity carries binding + artifact; oracle counts exactly the bound reviewer. Late terminal record `ffb42c0e838a7a74…` preserved as the pre-fix blocker. | FIXED_AND_REAL_PROVIDER_EXERCISED (S13 round 7); terminal `SUCCEEDED` blocked by AEH-V2-0106 |

| AEH-V2-0101 | P1 | Objective completion vs resumable controller-issued identities (S13 round 7 lane discovery, FIXED) | The round-7 lane's AcceptanceOracle `ACCEPTED` but objective completion failed closed with `PARTICIPANT_INCOMPLETE` for the controller-issued Implementer/Reviewer identities: their receipts are settled `SUCCEEDED` and their session participants are terminal, but `recordParticipantReceipt` deliberately leaves the launch identities non-terminal (AEH-V2-0089) and objective completion read only the live status. | Objective completion consumed the raw participant status while the successful-terminal gate already consumes the durable candidate-bound receipt; the two completion boundaries disagreed about the same bounded work. | `src/core/run.ts` | `objectiveParticipantAccountingV1` derives the completion status for in-flight participants from the same candidate-bound settled receipt the successful-terminal gate uses (by `participantId` or session provenance); no receipt → live status, stale candidate → live status, so in-flight work still fails closed. | `tests/objectiveParticipantAccounting.test.ts` (6). | Round-7 packed lane run 6: reviewer identity closed `COMPLETED` from its receipt; the remaining blocker is the assembly-lineage gap (AEH-V2-0106). | FIXED_AND_UNIT_REGRESSED (S13 round 7) |
| AEH-V2-0102 | P1 | Supervisor consolidation provenance echo (S13 round 7 lane discovery, FIXED) | A multi-round review lane failed with `AEH_OPERATION_SUPERVISOR_PROVENANCE`: the persistent supervisor's consolidation continuation returned a superseded finding id from an earlier round (`expected=reviewer:REVIEW-<current>`; `received=<stale>,<current>`). | The supervisor session is a deliberate continuation carrying earlier rounds in context; the exact-source provenance gate correctly rejects a consolidation that adds a stale id, but no bounded correction existed on the consolidation turn (the serialization repair exists only on first-dispatch turns). | `src/operations/supervisor.ts` | One bounded correction turn re-states the exact required id set and forbids earlier-round ids; the strict exact-set check remains and a second violation is a hard failure. | `tests/supervisorConsolidationProvenance.test.ts` (provenance cases). | Round-7 packed lane run 2 recorded the exact mismatch; the correction is exercised deterministically. | FIXED_AND_UNIT_REGRESSED (S13 round 7) |
| AEH-V2-0103 | P1 | Duplicate review evidence identities across review rounds (S13 round 7 lane discovery, FIXED) | A multi-round review lane's AcceptanceOracle rejected with `EVIDENCE_ITEM_INVALID` for `review:reviewer:<dimension>:<assertion>` duplicated: the report carried reviewer checks from several rounds (an earlier round bound a superseded candidate) and the bundle emitted one item per matching check with the same deterministic evidence id. | `currentReviewEvidence` treated every matching check as current evidence and did not reconcile round/candidate binding, so a stale round both collided with the current item and could not count. | `src/architecture/acceptanceOracle.ts` (evidence producer) | Exactly one check per (reviewer, dimension, assertion) is selected: the latest round that binds the current candidate/policy/impact wins; with no current round the latest round is reported and fails closed. The strict evaluation function is unchanged. | `tests/acceptanceReviewProvenance.test.ts` (multi-round single current item + ACCEPTED; superseded-only single FAIL + REJECTED). | Round-7 packed lane run 3 recorded the duplicate blocker; run 6 then produced a single current review item and `ACCEPTED`. | FIXED_AND_REAL_PROVIDER_EXERCISED (S13 round 7) |
| AEH-V2-0104 | P1 | Superseded review-round check poisoned the merged report status (S13 round 7 lane discovery, FIXED) | A remediated multi-round lane's AcceptanceOracle `ACCEPTED` but `report.status` stayed `FAIL` on the stale round-0 `candidate.assurance.reviewer.0.reviewer` check (the round-1 check PASSed), so objective completion and delivery were skipped and the run terminalized FAILED. | `runReviewRound` pushed a per-round check id and never reconciled earlier rounds; `mergeChecks` dedupes by id, so round-specific ids accumulated and any superseded FAIL kept the report FAIL. | `src/agents/reviewLifecycle.ts` | `replaceSupersededReviewerChecksV1` replaces a reviewer's earlier-round checks with the current round's check when the round advances; the lifecycle's quality history still records the earlier round. | `tests/reviewLifecycle.test.ts` (3). | Round-7 packed lane run 4 recorded the exact stale check; run 6's report was PASS with the current round only. | FIXED_AND_REAL_PROVIDER_EXERCISED (S13 round 7) |
| AEH-V2-0105 | P1 | Schema-invalid supervisor consolidation with no bounded correction (S13 round 7 lane discovery, FIXED) | A lane failed with `AEH_OPERATION_SUPERVISOR_CONTRACT` (`summary` missing; `missingEvidence` entries were objects) on the consolidation continuation; the bounded serialization repair applies only to first-dispatch turns, so the invalid response failed the operation without any correction. | Continuation turns (`executeAgentPrompt` + `continueBoundSession`) bypass the serialization repair wrapper; the consolidation parse is strict by design. | `src/operations/supervisor.ts` | The single bounded consolidation correction now also covers a contract-invalid response by re-stating the frozen supervisor contract and the exact required id set; the schema check still fails closed and a second failure is a hard error (total ≤2 turns per consolidation). Trace `operation.supervisor.consolidation-correction` records the failure/ids. | `tests/supervisorConsolidationProvenance.test.ts` (contract cases). | Round-7 packed lane run 5 recorded the exact schema failure; the correction is bounded and documented. | FIXED_AND_UNIT_REGRESSED (S13 round 7) |
| AEH-V2-0106 | P1 | Assembly-lineage receipt/candidate gap (S13 round 7 lane discovery; round 8 FIXED) | The round-7 final lane's AcceptanceOracle `ACCEPTED` but objective completion failed closed with one `PARTICIPANT_INCOMPLETE`: the Implementer's controller-issued identity has a settled `SUCCEEDED` receipt bound to candidate r2 (the candidate current during the isolated EXECUTING turn), while the current candidate is r3, the assembled result of that turn's ChangeSet. `evaluateTerminalGate` enforces the same equality (`CANDIDATE_MISMATCH`), so a passed objective completion would still fail the successful-terminal gate. | Receipts are recorded during participant finalization (before assembly) and bind the then-current candidate; only ASSEMBLING advances candidate truth (TARGET 8.1/COMPLETION_GATE), and no producer reconciled the work's receipt with the assembled candidate it produced. | `src/operations/v2Contracts.ts`, `src/operations/state.ts`, `src/operations/participantAccounting.ts`, `src/candidates/binding.ts`, `src/core/run.ts` | ROUND 8 FIXED: every successful candidate transition records a deterministic `CandidateAssemblyReceiptV1` (base candidate, assembled candidate, ChangeSet digest/patch digest, producer identity, and — when a settled `SUCCEEDED` source receipt exists — its exact id and canonical digest; explicit rebases record the observed source base). `resolveCandidateLineageReceiptV1` walks the durable assembly-receipt chain from the current candidate and accepts a work receipt only when the chain names that exact receipt id/digest and the receipt's observed candidate is on the walked ancestry; the receipt itself is never rewritten. `objectiveParticipantAccountingV1` and the successful-terminal receipt decision consume the same resolver, while `evaluateTerminalGate` remains byte-identical and is still invoked against each receipt's own binding; a receipt with no lineage, a broken chain, a divergent source candidate or a tampered assembly digest stays fail-closed. The completion snapshot check recomputes the same deterministic accounting. AEH-V2-0089 resumability and the one-coherent-chain invariant are preserved. | `tests/objectiveParticipantAccounting.test.ts` (12: lineage accepted, multi-hop repair chain, wrong source receipt rejected, divergent candidate rejected, broken/tampered chain rejected, resumable identity preserved, plus the round-7 gap reproduction), `tests/assemblyLineageCompletion.test.ts` (3: producer persistence + terminal gate acceptance, no-lineage rejection, resumed identity), `tests/objectiveCompletionLifecycle.test.ts` (lineage terminalization + fabricated snapshot rejection). | Round-8 packed `governed-lanes/change-direct.json`: operation `CHANGE-20260928T013113Z-5de819ea` terminal **`SUCCEEDED`** at revision 84 (513 s, inside the 900 s bound), AcceptanceOracle `ACCEPTED`, objective completion complete with zero blockers and every participant snapshot `COMPLETED`; the persisted operation record (sha256 `8e19971981b536f2…`, `governed-lanes/change-direct-round8-operation.json`) contains `assembly:…:r3` binding base r2 to the Implementer receipt `receipt:participant:429ce2201a833daa:…`. | FIXED_AND_REAL_PROVIDER_EXERCISED (S13 round 8) |
| AEH-V2-0107 | P1 | FORMAL_SDD escalation left the frozen policy on the superseded route/assurance (S13 round 8 lane discovery, FIXED) | The first round-8 `change-delegated` packed lane routed FORMAL_SDD from durable explorer/planner evidence and terminalized `FAILED`: the Spec Manager launch rejected with `EXECUTION_POLICY_STALE: durable ResolvedOperationPolicy does not match the current execution inputs.` The operation intent said `FORMAL_SDD`/`ELEVATED` while the frozen policy still said `DELEGATED`/`NONE`. | `runChangeOperation`'s escalation block patched intent and re-ran the OpenSpec preflight but never advanced execution semantics or rebound the frozen policy, so every later participant identity compiled against the superseded policy (spec-authoring never started). | `src/operations/change.ts` | ROUND 8 FIXED: `rebindEscalatedChangePolicy` deterministically advances the operation execution revision (establishing the pre-escalation semantics baseline first when no digest exists), clears the superseded policy through `bindOperationExecutionSemantics`, and rebinds a recompiled policy with the escalated route/assurance for the current candidate and epoch, before any FORMAL_SDD participant launches. Repeating the identical escalation is idempotent; no gate or policy guard was weakened. | `tests/specEscalationPolicyRebind.test.ts` (3: absent-digest baseline + rebind, existing-baseline single advance, fail-closed without a current policy). | Live packed lane evidence: operation `CHANGE-20260928T014341Z-226f0cf0` (route FORMAL_SDD/ELEVATED, frozen policy DELEGATED/NONE) failed at `runSpecManagerUntilReady` with the exact `EXECUTION_POLICY_STALE`; the superseding rerun routed DELEGATED (model triage variance) and did not re-exercise the escalation path. | FIXED_AND_UNIT_REGRESSED (S13 round 8; live re-exercise of the escalation path remains for round 9) |
| AEH-V2-0108 | P1 | Audit consolidation hangs without a bounded supervisor-turn deadline (S13 round 8 discovery; fixed round 9) | The round-8 build-bound `audit` lane reached `validating: COMPLETED` and `reviewing: COMPLETED` (02:12:01Z), fired one `operation.supervisor.consolidation-correction`, and then stayed `consolidating: RUNNING` past the 900 s bound and an explicit 600 s late wait (terminal `RUNNING` at revision 50 after 1,522,618 ms). The watchdog traced `harness.paseo.operation.watchdog.supervisor-busy` every ~7 s for 20+ minutes but only observed; no product deadline ended the stuck supervisor continuation. The campaign cancel was correctly rejected without a current HumanDecision. | The persistent supervisor's consolidation continuation turn has no bounded deadline; the watchdog observes a busy supervisor but cannot terminalize a stalled provider turn. | `src/operations/supervisor.ts`, `src/operations/controller.ts` (watchdog), `src/operations/state.ts` (terminalization) | ROUND 9 FIXED: every persistent-supervisor continuation turn uses a hard 300 s ceiling; timeout produces AEH_OPERATION_SUPERVISOR_TURN_TIMEOUT, stops the session, marks the generation FAILED, and propagates into the operation failure path. The final-build audit lane terminalized SUCCEEDED/PASS in 230.784 s; the timeout outcome itself is unit-regressed but was not hit in that real lane. | Reproduction: `governed-lanes/audit.json` (round-8 sweep) stage trace, `timeout: true`, `controllerBeforeCancel.alive: true`, one `operation.supervisor.consolidation-correction`, 151 `watchdog.supervisor-busy` traces. | `governed-lanes/audit.json` sha256 `004baabdb0e18e52afb4790951d14eb4f369755bf88677e8009fa59afe876ecb`. | FIXED_AND_BUILD_BOUND_REAL_PROVIDER_EXERCISED (S13 round 9) |
| AEH-V2-0109 | P3 | Governed-campaign route scoping and lane-artifact retention defects (R7-F2/F4, S13 round 8, FIXED) | The `change-direct` campaign lane asserted `change.delegated-route`/`delegated.feature-capsule`, which a DIRECT lane can never satisfy, so a fully accepted trunk still reported `result: FAIL`; lane envelopes had no checkout proof and no workspace query/archive evidence; the acceptance artifact digest was not independently recomputable. | Route-agnostic verification mapping; per-lane checkout/cleanup fields absent; acceptance artifact bytes discarded with the fixture. | `tests/packed/s13GovernedOperationCampaign.mjs` | ROUND 8 FIXED (route-scoped aggregate and lane-level assembly receipt projection completed round 9 after independent review R8-F1/R8-F2): verification checks and rows are scoped to the observed route (DIRECT asserts only direct-change; non-DIRECT asserts delegated-change), every lane envelope now carries `checkoutProof` (tracked digest before/after), `workspaceInventory` and archive/query results (with `S13_GOV_ARCHIVE_EVIDENCE=1` under `KEEP=1`), the acceptance artifact bytes are retained with a recomputable digest, and `candidateAssemblyReceipts` are projected into the operation evidence. No product gate changed. | The round-8 `change-direct` artifact carries `checkoutProof.checkoutUntouched: true`, four archived workspaces with `remaining: []`, and `artifacts.acceptance.raw` whose sha256 recomputes to `4263478a961036d2…`. | `docs/evidence/s13/review-round7.json` R7-F2/R7-F4; round-8 lane artifacts. | FIXED_IN_CAMPAIGN_SOURCE; final direct lane exercised the new projections but failed product validation/terminal acceptance (S13 round 9) |
| AEH-V2-0110 | P1 | DELEGATED planning failure and unexplained FORMAL_SDD route (S13 round 8 discovery; R9-F2, OPEN) | Round 8 failed in `planning` without an Implementer or ChangeSet. Round 9's retained Planner artifact says `formalizationNeed=NONE` while the operation route is FORMAL_SDD; its Explorer payload was not retained. | Exact route predicate still not isolated; the Round-11 run proves the route legitimately escalates from durable explorer/planner evidence while the Planner's own `formalizationNeed` is `NONE`, so the escalation must be driven by a different deterministic constraint over that evidence set. | `src/core/run.ts` (planner waves), `src/agents/waveExecutor.ts`, `src/operations/change.ts`, `tests/packed/s13GovernedOperationCampaign.mjs` | ROUND 10 PARTIAL: the campaign persists both Explorer/Planner structured payloads and `harness.change.triage-escalated` events; the round-10 run stayed DELEGATED with no escalation event. ROUND 11 ADVANCE: operation `CHANGE-20260928T183103Z-226f0cf0` escalated to FORMAL_SDD exactly once — `harness.change.triage-escalated` at `2026-09-28T18:35:09.505Z`, reason "SPEC escalation came from durable explorer/planner evidence", explorer artifact `…7e6fc809…json` sha256 `4816557c809ff72838b4e3267519bbe30f2009de14383601933016a8f60487c8`, planner artifact payload `cba6db51dc87601b80bc10e02d0c82d3693326aa0f69ff27c090c927d60c401b` (`formalizationNeed=NONE`), route diagnostics `plannerAssessmentMechanism: MODEL` / `routeGateMechanism: DETERMINISTIC`. The route trigger now has durable provenance of the exact model judgments; no accepted DELEGATED completion exists. Keep OPEN until the deterministic escalation constraint is identified and a DELEGATED lane reaches accepted completion. | Round-11 lane `docs/evidence/s13/governed-lanes/round-11/run-20260928183023575-c38a80/change-delegated.json` sha256 `5e9b18a100fa4d5dd2b4da0810b8cac7ab2ce3c309ea18531c9511e82bdf044b`; prior Round-9 diagnosis `e5b33d9148ea745acf159a5014c1188d5231934f47621f279a3699de4c4d0b27`. ROUND 12 REPAIRED IN SOURCE: `specEscalationConstraintV1` escalates only on the Planner's typed `formalizationNeed: REQUIRED` and the `harness.change.triage-escalated` event records the constraint name, Planner artifact/reason/evidence refs and Explorer finding statuses; the live refactor fixture stayed DELEGATED with zero escalation events across round-12 runs 4–11 (run 2 recorded exactly one legitimate Planner-REQUIRED escalation). Regressions: `tests/changeRouting.test.ts`. Round-12 also repaired the delegated pipeline defects this lane exposed (contract-before-planning, formalization ordering, provisional-policy supersede, known-competency skill selection, explicit launch permissions, single-command resolution, approval-stop failure); live lanes `run-r12-delegated-8/10/11` sha256 `a2a0ad38e53a1bdd40f1aafedf2449ed9ea261096a1c0826d64c2c1799d34b3a` / `9ef78df6cf5b7c8bc5fe9377ac6a8711a1bef164cdaf877989a4102315ff35b7` / `360fea66a73308d2d96d476c751aaf0108d156332807c68da0e9cf5575132e5d`. | FIXED AND REAL E2E VERIFIED (Round 17): the final-freeze `change-delegated` lane `r17-delegated-1` ran route DELEGATED with 0 escalation events to terminal `SUCCEEDED` and an `ACCEPTED` AcceptanceOracle (`docs/evidence/s13/governed-lanes/round-17/r17-delegated-1/change-delegated.json`); the constraint was already isolated in Round 12 (`PLANNER_FORMALIZATION_NEED_REQUIRED`) |

| AEH-V2-0111 | P1 | FORMAL_SDD Spec Manager result cannot be accepted against the prepared CandidateRevision (S13 round 9, OPEN) | Final round-9 formal lane ended `SPEC_MANAGER_RESULT_ARTIFACT_MISSING`; the session-reported `CANDIDATE_WORKSPACE_MISMATCH` remains unverified by a direct tool error receipt. | Source review found a deterministic pre-result mutation risk: old `prepareOpenSpecChange` called `openspec new change` in the frozen candidate workspace. This can change the workspace digest, but it is not yet proven to be the cause of the Round-9 agent-reported mismatch. | `src/operations/change.ts`, `src/spec/openspec.ts`, `src/candidates/identity.ts`, `src/workers/resultGateway.ts` | ROUND 10 PARTIAL: OpenSpec preparation now returns a canonical target without writing placeholders; the controller persists READY content only after the structured result passes. `openSpecBridge` regression proves the frozen candidate digest remains unchanged; `candidateIdentity` reports exact expected/observed digests on mismatch. ROUND 11: the live FORMAL_SDD path was re-exercised to a READY structured result — Spec Manager `participant:137e533e857cbb1c`, `status: READY`, `validationReady: true`, artifact sha256 `0f85afd2a935593c858758154b20d71709a3dd73b45cf713eb262abf3e222d6e`, payload sha256 `66ba276626051254cdee367689774dd01c678e5b279551dc76fdf14651bb5256` — and no `CANDIDATE_WORKSPACE_MISMATCH` occurred. The READY content was then rejected by OpenSpec compilation for missing delta sections, a distinct deterministic defect recorded as AEH-V2-0115. No schema, authority, identity, retry, or validation gate changed; keep OPEN pending an accepted live FORMAL_SDD result. | Round-9 corrected diagnostic `e5b33d9148ea745acf159a5014c1188d5231934f47621f279a3699de4c4d0b27`; Round-10 regressions `tests/openSpecBridge.test.ts`, `tests/candidateIdentity.test.ts`; Round-11 lane `5e9b18a100fa4d5dd2b4da0810b8cac7ab2ce3c309ea18531c9511e82bdf044b`; Round-10 delegated lane `3daada863167bf49ed98a852364491647a221e903e722e025ff42b474e6c3871`. | ROUND 15: the live `change-formal` lane reached canonical persistence and successful OpenSpec compilation (runs 2–3), so the original READY-acceptance risk is exercised; the row is FIXED AND REAL E2E VERIFIED (Round 17): the final-freeze `change-formal` lane `r17-formal-3` reached terminal `SUCCEEDED` with the canonical per-capability delta, sealed contract and an `ACCEPTED` AcceptanceOracle (`docs/evidence/s13/governed-lanes/round-17/r17-formal-3/change-formal.json`) |
| AEH-V2-0112 | P1 | Governed campaign archives a live CandidateRevision workspace after lane timeout (S13 round 9, FIXED AND EXERCISED) | The Round-9 direct lane archived four workspaces at timeout while its controller remained alive, then the controller failed after the worktree disappeared. | Campaign cleanup previously did not gate archival on durable terminal state and controller exit. | `tests/packed/s13GovernedOperationCampaign.mjs`, `tests/packed/s13GovernedCampaignPolicy.mjs` | ROUND 10 FIXED: `workspaceArchiveDecision` returns eligible only for a terminal operation and exited controller; the regression covers terminal/running × alive/exited combinations. The packed direct lane completed `SUCCEEDED` in 562.596 s, its bound candidate workspace digest matched before cleanup, and four workspaces archived with none remaining. A separate post-campaign read confirms the controller exited and the operation remained SUCCEEDED; no live candidate workspace was removed. | Round-10 direct lane `f2d8b9e80067237a9240c6069d13a63d8ad85c417944eff4d7af809d0fdf7979`; post-campaign operation read `c80bdbf812312eb86cecf6aa04d2bb12888d3632cfedeebcbb4ee96031413a55`; cleanup predicate regression `node --test tests/packed/s13GovernedCampaignPolicy.test.mjs`. | FIXED_AND_PACKED_REAL_PROVIDER_EXERCISED (S13 round 10) |
| AEH-V2-0113 | P1 | Round-9 GitHub campaign created a repository without fail-closed pre-authorization (AEH_PRODUCT_DEFECT / EXTERNAL_EFFECT_POLICY_VIOLATION) | The round-9 campaign used optional `S13_GH_REUSE_REPO`, generated a repo name when absent, and directly invoked `gh repo create` before any authorization gate. Private repo `JamesMorales04/aeh-s13-round3-20260928032308-4317d7` was created; its later HTTP 403 deletion failure is not a certification blocker. | Missing explicit target input plus a generated-name fallback and ungated external create path. | `tests/packed/s13GithubEffectCampaign.mjs`, `tests/s13GithubCampaignPolicy.test.ts` | ROUND 10 PRODUCT DEFECT FIXED: require a pre-authorized private OWNER/REPO input before staging or any external command; no repo-create/delete command remains. **ROUND-11 CORRECTION (AEH-V2-0114):** the Round-10 attempt-3/4 authority and operation-bound issue-import reruns are NOT_ACCEPTED (tainted) — they depended on campaign-fabricated HumanDecisions and unauthorized issues #2/#3 (see AEH-V2-0114). Deliveries: attempt-4 summary `s13-github-effects-round-10-attempt-4.json` `6cf03b9d…` and its lanes are retained as historical bytes only. (B) `USER_CLEANUP_NON_BLOCKING`: the user will delete the Round-9 repo manually; no retry or auth broadening. Taint the Round-9 and Round-10 authority/issue-import claims. | Round-9 source evidence `s13-github-effects.json` SHA-256 `ba4ad03b4c6ee8da8505f56dc85c08648ee79a0284bf854e80ee02b988f414e8`; Round-10 attempt artifacts `6cf03b9d73a9b41a7d7deb4d97b7c735ac56dd819e9f62dcc0c469550238c878`, `9fee87ab565ba51e6e6cd515dbff76ea3899a56d04ecb1ebd81d78ea29e4a4c5`, `68270e9ddc08866b0c825ac5e0ea2ca1ced75ddd8b28d4ff0afdd7701a122b98` (tainted). | FIXED_FOR_REPOSITORY_TARGET_GATE; Round-10 authority/import reruns NOT_ACCEPTED (AEH-V2-0114); manual repository cleanup remains USER_CLEANUP_NON_BLOCKING. |
| AEH-V2-0114 | P1 | Campaign self-minted HumanDecision approvals for external effects (S13 round 10, `AEH_PRODUCT_DEFECT / EXTERNAL_EFFECT_POLICY_VIOLATION`, HISTORICAL_POLICY_VIOLATION_CONFIRMED, repaired round 11) | `tests/packed/s13GithubEffectCampaign.mjs` fabricated its own `APPROVE` HumanDecisions (`actorId: "human:s13-github"`, invented reason) for both `github.issue.create` (lines ~172-201) and `git.push` (lines ~385-402). A durable pre-effect user response on the Round-10 lead agent (`69a65987-7d0e-4a9e-8cac-75cbb1188ebb`, `persistence.metadata.asyncQuestions[0].resolution: "Use only pre-authorized existing private repository JamesMorales04/aeh-s13-round3-20260927035743-6bb5da and existing issue #1 … Do not create a repo or issue."`) explicitly forbade creating an issue. Issues #2 (intent `action-intent:4ad874ff…` 16:01:50.692Z / receipt `action-receipt:d3d0aa63…` 16:01:52.323Z) and #3 (`action-intent:a728311b…` 16:08:26.221Z / receipt `action-receipt:a215a04b…` 16:08:27.317Z) were created anyway, and the Round-10 authority branch `aeh-s13-authority-20260928165409` was pushed under the same fabricated approval. ActionIntent + ActionReceipt never prove consent; the gate only proves the gate ran. | The campaign treated the ledger `record()` API as an authorization source and supplied both the actor and the approval content itself, so no external user authorization was consumed before the external effects. The campaign fixture configuration could therefore widen human authorization. The Round-10 "clean/user-authorized" documentation claims were unsupported. | `tests/packed/s13GithubEffectCampaign.mjs`, `tests/packed/s13ExternalEffectAuthorization.mjs`, `tests/s13GithubCampaignPolicy.test.ts`, `docs/evidence/s13/github-effect-authorization.json` | ROUND 11 REPAIRED (tests-only; product untouched): all campaign-fabricated approvals removed. The campaign now consumes an externally supplied, scoped user authorization artifact (`S13_GH_AUTHORIZATION_FILE`, prospective `docs/evidence/s13/github-effect-authorization.json` amended with machine-readable `scope.resourceNamePrefix`/`scope.allowedEffects`) through a typed loader that verifies kind/status/owner/PRIVATE_ONLY/effect coverage/resource namespace and derives the only `APPROVE` from an unforgeable validated object, bound to the frozen allowed-effect policy and the exact ActionIntent effect digest; a missing, unreadable, invalid, out-of-scope or effect-uncovering artifact fails with a typed error before any `gh` command, staging directory or approval. The default lane reuses the user-authorized existing issue #1; governed issue creation requires explicit `github.issue.create` coverage. Deterministic regressions: `tests/s13GithubCampaignPolicy.test.ts` (9) covers absence/unreadable/out-of-scope/uncovered failures with no `gh` invocation and empty staging, the positive consumption path, unbranded-approval rejection, and a static no-fabricated-approval assertion. Residual risk: the artifact is user-supplied data; the deterministic loader is the trust gate and cannot itself prove human origin — external provenance is the user-delivered prospective authorization and the artifact digest is recorded in every lane. | Issues #2/#3 remain in place (do not edit/delete) and repository deletion is not retried; leftover cleanup is `USER_CLEANUP_NON_BLOCKING`. Machine record `docs/evidence/s13/round10-0114-authorization-resolution.json` (determination `HISTORICAL_POLICY_VIOLATION_CONFIRMED`) and `docs/evidence/s13/review-round10.json` R10-F1/F2/F3. | REPAIRED_IN_CAMPAIGN_SOURCE_AND_REGRESSED (S13 round 11); Round-10 authority/issue-import rows remain NOT_ACCEPTED (tainted); a user-authorized Round-11 rerun is recorded separately. |
| AEH-V2-0115 | P1 | FORMAL_SDD READY structured content is not validated for OpenSpec canonicality (S13 round 11 discovery, OPEN) | The Round-11 `change-delegated` lane escalated to FORMAL_SDD and the Spec Manager returned `status: READY`, `validationReady: true` with `artifacts.specs` as a flat requirements document (`### Requirement:` / `#### Scenario:`) without OpenSpec delta headers. The controller persisted that content and the operation terminalized `FAILED` inside `compileOpenSpecChange` with `OpenSpec change '<id>' is not valid and cannot be compiled into AEH normative artifacts: … No delta sections found. Add headers such as "## ADDED Requirements" … Change must have at least one delta`. | READY acceptance validated the typed result shape but not the canonical OpenSpec change layout required by the compiler; the failure surfaced only after persistence, as a generic compile error instead of a typed pre-persistence rejection, and the spec-authoring contract does not require delta-formatted `specs` content. | `src/spec/openspec.ts` (canonical validation + `compileOpenSpecChange`), `src/operations/change.ts` (READY acceptance/persistence), `src/agents/outputContracts.ts` (spec-authoring contract) | ROUND 11 OPEN — required repair: validate READY content against the canonical layout (per-capability `specs/<capability>/spec.md` with `## ADDED/MODIFIED/REMOVED/RENAMED Requirements` deltas and at least one `#### Scenario:` per requirement) before persistence/compilation, fail closed with a typed `SPEC_MANAGER_CONTENT_NOT_CANONICAL` error naming the exact artifact, and align the Spec Manager contract so its `specs` output uses delta headers. No gate may be weakened. **ROUND 12 REPAIRED IN SOURCE:** `validateOpenSpecAuthoringContentCanonicalityV1` rejects a non-canonical READY result before any controller-owned write with `SPEC_MANAGER_CONTENT_NOT_CANONICAL` naming `openspec/changes/<change>/specs/<capability>/spec.md (artifacts.specs[i])`; the Spec Manager contract now carries one canonical `{ capability, content }` delta per capability and persists to per-capability `specs/<capability>/spec.md`. Regressions: `tests/specManagerCanonicality.test.ts` (6, including a frozen-controller run and the escalation-path regression). The live canonical FORMAL_SDD rerun is still pending. | Lane `docs/evidence/s13/governed-lanes/round-11/run-20260928183023575-c38a80/change-delegated.json` sha256 `5e9b18a100fa4d5dd2b4da0810b8cac7ab2ce3c309ea18531c9511e82bdf044b`; operation `CHANGE-20260928T183103Z-226f0cf0` terminal FAILED revision 80; Spec Manager artifact `0f85afd2a935593c858758154b20d71709a3dd73b45cf713eb262abf3e222d6e` / payload `66ba276626051254cdee367689774dd01c678e5b279551dc76fdf14651bb5256`. | ROUND 15: the gate now also enforces the OpenSpec strict normative keyword (`SHALL`/`MUST`) in every requirement body before persistence and the Spec Manager prompt states it (`tests/specManagerCanonicality.test.ts` 8 tests); the live `change-formal` lane runs 1–3 completed spec-authoring, canonical persistence and OpenSpec compilation. The row is FIXED AND REAL E2E VERIFIED (Round 17): `r17-formal-3` terminal `SUCCEEDED` with canonical spec authoring/compilation and an `ACCEPTED` AcceptanceOracle on the final freeze. |
| AEH-V2-0116 | P1 | Paseo-managed OpenCode participant sessions stop at the provider approval prompt despite an explicit AEH permission projection (S13 round 12 discovery, OPEN, `SLICE_BLOCKED_DRIVER`) | The round-12 `change-delegated` lane reached planning COMPLETED, participant compilation, a real Implementer session and a PASS validation report, but the Implementer session stopped with Paseo status `permission` and produced no ChangeSet; the agent was created with the deterministic projection (`edit`/`bash` allow) and `featureValues.auto_accept: true` (agent `3c313277-0aff-4c96-b1de-2a6ca5535017`, session `ses_f16171776ffeulK7VpcQPsYnjb`, run `run-r12-delegated-11`), and the daemon snapshot retained no `pendingPermissions` entry. Earlier (`run-r12-delegated-10`) the same stop was silently recorded as a successful empty turn. | The provider approval prompt is not a `pendingPermissions` tool-permission entry that `auto_accept` clears; the daemon does not retain the rejected prompt identity, so the exact requested permission is unknown. | `src/paseo/runtimeCore.ts` (`sdkExitCode`, stop detail), `src/paseo/native.ts` (`isTerminalStatus`), `src/paseo/launchSpec.ts` + `src/paseo/sdk.ts` (`featureValues`), `src/agents/waveExecutor.ts` (explicit permission projection) | ROUND 12 PARTIAL: approval stops are terminal failures with an explicit stderr detail (`tests/paseoRuntime.test.ts`), `featureValues` is plumbed and `auto_accept` is enabled only when the projected permissions contain no operator `ask` (`tests/paseoLaunchSpec.test.ts`), and the compiled tool ceiling projects explicit read/write/gitShell/network permissions (`tests/participantRuntimeBinding.test.ts`). The daemon-level approval mechanism remains unresolved. | Lanes `docs/evidence/s13/governed-lanes/round-12/run-r12-delegated-10/change-delegated.json` sha256 `9ef78df6cf5b7c8bc5fe9377ac6a8711a1bef164cdaf877989a4102315ff35b7`; `run-r12-delegated-11` sha256 `360fea66a73308d2d96d476c751aaf0108d156332807c68da0e9cf5575132e5d`; summary `9b23266356ae1593aad76575ad58ed55b49e7de0cdd0d76a09ea57059496f009`. | ROUND 13: DIAGNOSED AND AEH-OWNED PARTS REPAIRED (D1 scoped `external_directory` projection for the frozen launch root + git metadata roots; D2 pending-approval payload captured from `waitForFinish.final.pendingPermissions` into the failed-turn detail; D3 Codex `providerOptions` sandbox projection `approval_policy: never` + `sandbox_mode` from the compiled ceiling; D4 Paseo parent handle suppressed for isolated launches because Paseo children inherit the parent workspace/cwd). Paseo behavior documented as provider-specific: `auto_accept` is deterministically disabled whenever a `toolPolicy` is present (`opencode-agent.js:2425`), and child agents inherit the parent workspace/cwd (probe6). Live lane `run-20260928225446983-c5ff6c`: Implementer COMPLETED with a validated ChangeSet and r2->r3 assembly receipt, then remediation failed on the new AEH-V2-0118. Evidence `docs/evidence/s13/round13-0116-diagnosis.json`; regressions `tests/permissions.test.ts`, `tests/paseoLaunchSpec.test.ts`, `tests/paseoRuntime.test.ts`. |
| AEH-V2-0117 | P1 | Paseo child agents inherit the parent workspace/cwd, so isolated wave-task sessions executed in the operation candidate workspace (S13 round 13 discovery, `AEH_PRODUCT_DEFECT`, REPAIRED) | The round-13 DELEGATED lane launched the Implementer with cwd = the isolated task worktree under /tmp, but the Paseo agent record, session timeline and workspace watcher all show the session ran in the operation candidate worktree `/home/james/.paseo/worktrees/1yo7ez85/aeh-change-20260928t224304z-e75534ab` and mutated it (`src/format.mjs` added there, `src/greeting.mjs` modified), tripping `CANDIDATE_WORKSPACE_MISMATCH` in `recordParticipantReceipt`. The launch spec passed the supervisor as the Paseo `parent`; probe6 proved a child created with an explicit cwd is relocated to the parent's workspace and cwd. | Isolated task launches carried the Paseo parent handle even though Paseo uses it for workspace/cwd inheritance. | `src/paseo/launchSpec.ts` (`paseoParentAgentId` is omitted when the frozen launch root differs from the durable operation workspace root; `aeh.parent-agent`/`AEH_PARENT_AGENT_ID` correlation is retained), `src/workers/agentPrompt.ts` (Paseo parent option uses `spec.paseoParentAgentId`) | `tests/paseoLaunchSpec.test.ts` — isolated launch keeps the correlation label but drops the Paseo parent handle; workspace launch keeps both. | Live lane `run-20260928225446983-c5ff6c`: the Implementer then edited only its task worktree and COMPLETED with a validated ChangeSet and assembly receipt r2->r3. | FIXED_AND_REAL_PROVIDER_EXERCISED (S13 round 13) |
| AEH-V2-0118 | P1 | DELEGATED remediation fails after a PASS validation report and terminalizes FAILED with no durable error (S13 round 13 discovery, CLOSED — R17-F9) | Lane `run-20260928225446983-c5ff6c`: after Implementer COMPLETED, assembly receipt r2->r3 and a configured validation report with `command.fixture-greeting` and `contract-test` both PASS, the `implementation` stage FAILED, remediation started two Repairer sessions that returned `exitCode 0` with no changes (their transcripts diagnose a `contract-test` validator-id association blocker between the frozen contract and `.harness/project.yaml` and state no allowed implementation-only repair exists), and the operation terminalized FAILED at revision 122 (`operation.terminal`) with `error: null` and no stage message. | The repair trigger for an already-PASS validation report and the terminal failure reason are not durably persisted; the repairers' diagnosis of the `contract-test` wiring is unverified by a tool receipt. | `src/core/run.ts` (remediation trigger), `src/agents/recovery.ts`, `src/architecture/validationRequirements.ts`, `tests/packed/s13GovernedOperationCampaign.mjs` | Round-13 campaign harness now persists per-stage messages and failed-participant artifacts (`stageMessages`, `artifacts.failedParticipants`), which is how the repairer transcripts were retained. | Lane `docs/evidence/s13/governed-lanes/round-13/run-20260928225446983-c5ff6c/change-delegated.json`; fixture operation `CHANGE-20260928T225531Z-e75534ab` revision 122. | ROUND 15 PROGRESS: terminal FAILED operations now persist the owning failing checks (`operationFailureDetail`/`validationFailureDetail`, stage messages on implementation/remediation/review), so `error: null` terminals are eliminated; contract-derived base validation requirements and unresolvable plan-requirement handling were completed; live lanes now reach `review=ACCEPTED`. **ROUND 17 CLOSURE:** the closure condition is met — `change-delegated` reached terminal `SUCCEEDED` in Round-15 `run-20260929-r15-delegated-9` (oracle `ACCEPTED`, complete objective accounting) and again on the final freeze in Round-17 `r17-delegated-1` (route DELEGATED, 0 escalations, terminal `SUCCEEDED` rev 132, oracle `ACCEPTED`). No separate remediation-signal defect remains. |
| AEH-V2-0119 | P1 | Reviewer provider-stop is classified as remediable implementation debt, loops through no-change Repairers, and terminalizes without a typed error (S13 round 14/15, `AEH_PRODUCT_DEFECT`, REPAIRED) | Round-14 run 4 and the Round-15 lane: a Reviewer Paseo session stopped on an unapproved `external_directory` permission prompt (`/root/*` in run 2); `runReviewer` emitted a synthetic critical `SYSTEM_FAILURE` finding, `detectHumanException` returns `humanRequired: false` for it, the quality gate treated it as debt, Repairers produced no ChangeSet, and the loop repeated until a bounded terminal FAILED. Round-13's control-root read projection (`external_directory` allow for the launch root, git metadata roots and — for write-denied roles only — the operation control root) covered the assigned control root but not arbitrary out-of-scope model reads. | A reviewer failure is system evidence, never implementation debt; the loop must be bounded and the terminal cause typed and durable. | `src/agents/reviewLifecycle.ts` (`ReviewerRoundFailureV1`, `runReviewRound` failure short-circuit before supervisor consolidation, `reviewerProviderFailureResult`/`systemFailureResult`, `agent.system-failure` check, reviewer prompt boundary + changed-files/requirements context), `src/core/run.ts` (quality-state-tolerant review summary); retained Round-14 pieces: `escalation.maxRounds` budget, control-root external read scope in `src/paseo/launchSpec.ts` | `tests/reviewRepairCandidateLifecycle.test.ts` (5: provider stop → one Reviewer turn, zero Repairers, typed `SYSTEM_FAILURE`, quality state; mutation attempt; SYSTEM_FAILURE finding; remediation success path; fail-closed repairer authority), `tests/reviewLifecycle.test.ts` (7), `tests/paseoLaunchSpec.test.ts` (10) | Round-15 lane `run-20260929-r15-delegated-2`: the Reviewer stop produced a durable reviewer error and an immediate typed terminal (the summary crash in that run was AEH-V2-0120, fixed); `run-20260929-r15-delegated-3` reached review COMPLETED/ACCEPTED with no reviewer stop. | REPAIRED_AND_REAL_PROVIDER_EXERCISED (S13 round 15); no gate weakened, no PASS coerced |
| AEH-V2-0120 | P1 | Supervisor generation bound to a prior candidate cannot deliver consolidation turns; a terminal review failure returned no quality state and crashed the run summary (S13 round 14/15, `AEH_PRODUCT_DEFECT`, REPAIRED) | Round-14 run 1 failed with `AEH_RESULT_STALE_CANDIDATE` when the supervisor's structured-result channel was bound to a candidate superseded by assembly; Round-15 run 2's reviewer-stop terminal path crashed `runTask` with `TypeError: Cannot read properties of undefined (reading 'debtScore')` because `review.qualityHistory.at(-1)` was undefined. | A generation's structured-result binding must match the current candidate before a continuation; every terminal lifecycle result must carry a quality state. | `src/operations/supervisor.ts` (candidate-drift rotation with positive current-policy proof, bounded consolidation correction with delivered-turn evidence, turn timeout), `src/agents/reviewLifecycle.ts` (`systemFailureResult` records a quality state), `src/core/run.ts` (defensive summary fallback) | `tests/supervisorConsolidationProvenance.test.ts` (10), `tests/reviewRepairCandidateLifecycle.test.ts` (quality-state assertion) | Round-14 run 1 reproduced the stale channel; Round-15 run 3's review consolidation completed under the rotated generation; no `error: null` terminal remains | REPAIRED_AND_REAL_PROVIDER_EXERCISED (S13 round 15) |
| AEH-V2-0121 | P1 | Headless DELEGATED operations cannot satisfy Lead semantic acceptance: no bound Lead generation exists (`ACCEPTANCE_LEAD_REQUIRED`) and the governed fixture could not resolve an `architecture` impact validation (S13 round 15, `HARNESS/FIXTURE_GAP`, REPAIRED IN CAMPAIGN) | Round-15 run 1: the model classified the internal refactor as an `architecture` impact dimension; the canonical dimension→kind mapping produced `impact-review-architecture`, the fixture declared no architecture validator or project script, and candidate assurance failed closed. Run 3 (after the fixture fix) reached `review=ACCEPTED`, then failed closed with `ACCEPTANCE_LEAD_REQUIRED: managed operation has no current bound Lead generation` — a non-interactive `operation start` has no `PASEO_AGENT_ID`, so no Lead generation can supply the required independent Lead semantic evidence. | The fail-closed resolver and the Lead-acceptance oracle are correct; the certification fixture was under-provisioned for a canonical dimension it can provoke, and the headless lane did not emulate the interactive Lead entry. | `tests/packed/s13GovernedOperationCampaign.mjs` (real `scripts/architecture.mjs` architecture validation registered as `check:architecture`; per-lane real Paseo Lead agent created with `aeh.provider` label and bound via `PASEO_AGENT_ID`, deleted after the lane) | Fixture architecture script exercised by lanes 2/3; the Lead binding is exercised by the next lane run (product `bindOperationLead` + `requestManagedLeadAcceptance` paths unchanged) | Round-15 artifacts `governed-lanes/round-15/run-20260929-r15-delegated-{1,2,3}/change-delegated.json` | REPAIRED_IN_CAMPAIGN; no oracle/validator gate changed; `delegated-change` terminal SUCCEEDED in run 9 (implementation-lead result; independent review pending) |
| AEH-V2-0122 | P1 | Managed Lead acceptance requested an AEH structured-result turn for an interactive Lead session with no result channel (S13 round 15 discovery, `AEH_PRODUCT_DEFECT`, REPAIRED) | Round-15 run 5: the real Paseo Lead bound via `PASEO_AGENT_ID` failed the acceptance continuation with `AEH_RESULT_CHANNEL_STATE: no structured result channel is bound to agent '<lead>'` before the provider ran. Interactive leads created by `aeh start` are not worker launches and carry no `aeh-result` sink/tool policy or channel binding. | `requestManagedLeadAcceptance` passed `leadOutputJsonSchema` to `continueManagedPaseoAgent`, which activates an AEH structured-result turn and requires a bound channel. | `src/agents/managedLeadAcceptance.ts` (textual continuation; the typed schema parse, exact assertion coverage, verdict and unresolved checks remain authoritative) | `tests/managedLeadAcceptance.test.ts` (no schema turn requested; prompt contract and evidence projection asserted) | Round-15 run 5 reproduced; runs 6–9 executed the lead continuation successfully | FIXED_AND_REAL_PROVIDER_EXERCISED (lead evidence PASS in run 9) |
| AEH-V2-0123 | P1 | Resolved base (plan-authored) validation requirements were never executed, so the AcceptanceOracle saw dangling requirements; the Lead prompt did not project resolved validation evidence (S13 round 15 discovery, `AEH_PRODUCT_DEFECT`, REPAIRED) | Round-15 run 6: candidate assurance compiled plan requirements `architecture-boundary` (kind architecture → `check:architecture`) and `fixture-greeting` (kind command → the single configured command) plus contract-derived `command.fixture-greeting`/`contract-test`, but the report only carried checks from `verifyTask` and the impact executor, so `validationForRequirement` found no check for the two plan ids and the oracle rejected `VERIFICATION_VALIDATION_EVIDENCE_INSUFFICIENT`. Run 7 then failed `VERIFICATION_LEAD_EVIDENCE_INSUFFICIENT` because the Lead could not map `impact-acceptance-behavior-correctness` to its unit-test evidence from opaque check ids. | `runCandidateImpactValidations` only executed impact requirements; plan-authored requirement ids do not necessarily coincide with configured check ids, so no execution path existed for them. | `src/core/run.ts` (`requirements` parameter; assurance recompilation executes every resolved base requirement without a matching report check — check id, `candidate-impact-<id>`, `candidate.assurance.validation.<id>` or `details.requirementId` — reusing already-executed configured checks; no duplicate evidence for one requirement), `src/agents/managedLeadAcceptance.ts` (per-assertion `validationRequirement(kind)=oracleCheckStatus` projection) | `tests/acceptanceValidationPath.test.ts` (unmatched plan requirement executes and the oracle accepts exactly one check), `tests/managedLeadAcceptance.test.ts` (evidence projection) | Round-15 runs 6/7 reproduced; runs 8–9 reached `VERIFICATION_VALIDATION_EVIDENCE_INSUFFICIENT`-free acceptance with PASS validation evidence for every assertion | FIXED_AND_REAL_PROVIDER_EXERCISED (run 9 oracle ACCEPTED) |
| AEH-V2-0124 | P1 | Non-producing ancestor-bound work was not completion evidence and a retried launch identity's transient `FAILED` shadowed its settled receipt, blocking objective completion (S13 round 15 discovery, `AEH_PRODUCT_DEFECT`, REPAIRED) | Round-15 run 8: AcceptanceOracle `ACCEPTED` and review `ACCEPTED`, then objective completion failed `PARTICIPANT_INCOMPLETE` ×2. Explorer/Planner bound the pre-assembly candidate r2 (no ChangeSet, so no assembly receipt names them), the Reviewer launch identity retained a transient live `FAILED` from a retried attempt beside its settled receipt, and `objectiveParticipantAccountingV1` returned the live status before consulting lineage-proven receipts; `terminalReceiptDecision` likewise only accepted current-bound or assembly-named receipts. | The lineage resolver only accepted the assembly-named producer receipt or the current candidate; accounting short-circuited terminal live statuses; the terminal gate evaluated only `ASSEMBLY` receipts against their own binding. | `src/operations/v2Contracts.ts` (`ANCESTOR` resolution for settled `SUCCEEDED` receipts of known non-producing roles — Explorer, Planner, Reviewer, Librarian, Spec Manager — whose bound candidate lies on the deterministic assembly ancestry; unknown/absent roles and Implementer/Repairer stay strict), `src/operations/participantAccounting.ts` (lineage-proven settled receipt wins over a transient live status), `src/operations/state.ts` (`terminalReceiptDecision` evaluates `ANCESTOR` receipts against the candidate the work observed; `evaluateTerminalGate` unchanged) | `tests/objectiveParticipantAccounting.test.ts` (live-FAILED Explorer identity completes through its ancestor receipt; a producer receipt not named by the assembly stays rejected; terminal-gate basis for the ancestor binding), `tests/assemblyLineageCompletion.test.ts` | Round-15 run 8 reproduced; run 9 terminal `SUCCEEDED` with complete objective accounting and `error: null` | FIXED_AND_REAL_PROVIDER_EXERCISED (run 9 terminal SUCCEEDED) |
| AEH-V2-0125 | P2 | Formal change-path Planner/Explorer prompts did not require the `AEH_RESULT_JSON=` marker (S13 round 15 formal run 2, `AEH_PRODUCT_DEFECT`, REPAIRED) | `change-formal` run 2 compiled the authored OpenSpec change and then failed planning `PLANNER_RESULT_ARTIFACT_MISSING: agent completed without an accepted structured result artifact`; the codex Planner turn returned no valid native JSON and no marker (`NO_MARKER`). The formal-path planner prompt (and explorer prompt) omitted the explicit marker line that every other structured prompt requires. | Prompt completeness, not a contract defect; the durable handoff contract and schemas were correct and stayed unchanged. | `src/operations/change.ts` (Explorer and Planner prompts now require exactly one `AEH_RESULT_JSON=` final line; a turn without it keeps the typed `EXPLORER_RESULT_ARTIFACT_MISSING`/`PLANNER_RESULT_ARTIFACT_MISSING` rejection) | No dedicated unit regression (prompt text); exercised by the next formal lane runs (run 3 progressed past the marker to a workspace-identity rejection) | `governed-lanes/round-15/run-20260929-r15-formal-2/change-formal.json` | FIXED_IN_SOURCE; formal row remains NOT_ACCEPTED on AEH-V2-0126 |
| AEH-V2-0126 | P1 | FORMAL_SDD persists the authored OpenSpec change into the frozen candidate workspace without a deterministic candidate-lineage advance, so the Planner result is rejected `CANDIDATE_WORKSPACE_MISMATCH` (S13 round 15 formal run 3, `AEH_PRODUCT_DEFECT`, OPEN) | `change-formal` run 3: spec-authoring and spec-compilation COMPLETED; planning then failed `CANDIDATE_WORKSPACE_MISMATCH: workspace does not materialize CandidateRevision candidate:…:r2 (expectedSourceDigest=07ef82b3…, observedSourceDigest=d7cf7bed…)`. Persisting `openspec/changes/<change>/…` writes non-ignored files into the candidate workspace after r2 was bound, changing its digest before the Planner's structured result is accepted. | The formal path persists controller-owned authoring content after the candidate binding but does not advance/rebind the candidate (or record an assembly receipt) for that deterministic authoring transition; the issue-intake path already performs an explicit lineage advance for authoring writes. | `src/operations/change.ts` (formal spec-authoring/compilation ordering and candidate binding), `src/spec/openspec.ts` (persistence), `src/candidates/assembler.ts` / `src/operations/state.ts` (candidate advancement/assembly receipt if used) | None yet — required: a focused regression proving the formal authoring persistence advances the candidate deterministically (or records the authoring assembly receipt) and that the Planner result binds the advanced candidate; then rerun `change-formal` to a terminal result | `governed-lanes/round-15/run-20260929-r15-formal-3/change-formal.json` sha256 `6f134a4e7346df0646e2dada8c502d14f08b837a554a4ccafd3dca0c162fbbd1` | FIXED AND REAL E2E VERIFIED (Round 16 repaired the provisional bootstrap policy after the controller-authoring candidate advance; Round 17 final-freeze `r17-formal-3` reached terminal `SUCCEEDED` with `formal.spec-authoring`/`formal.spec-compilation` PASS and an `ACCEPTED` AcceptanceOracle) |
| AEH-V2-0127 | P2 | Delivery effects were not enumerable as an authorized subset: enabling `finalizeOnAcceptance` always requested `github.pull-request.create`, so an accepted delivery could not be finalized within an external authorization that covers only `git.push`/`github.branch.create` (S13 round 17, `AEH_PRODUCT_DEFECT`, FIXED) | `configuredExternalEffects` hardcoded `git.push` + `github.pull-request.create` when finalization was enabled, and `finalizeAcceptedIssue` unconditionally attempted the pull-request gate, so the exact-effect authorization chain (artifact → frozen policy → ActionIntent → ToolActionGate → effect → receipt) could not express a push-only delivery. | `src/security/actionPolicy.ts`, `src/delivery/finalize.ts`, `src/core/types.ts`, `src/core/config.ts`, `schemas/project.schema.json` | `tests/security/actionPolicy.test.ts` (push-only effect/requirement enumeration), `tests/deliveryFinalize.test.ts` (push-only finalization with exactly the commit+push intents and no `/pulls` request; 12 tests) | `delivery.github.pullRequests?: boolean` (default true) is added to the project config, types and schema; when false the frozen policy omits `github.pull-request.create` and accepted finalization returns a push-only `FINALIZED` after the gated push (TARGET 11: requested and authorized effects are enumerated in frozen policy). | `npm run typecheck`; focused 24 files / 205 tests; canonical non-system suite 1,347 passed / 3 pre-existing `paseoSdkResolve` failures | Real GitHub lane `r17-gh-4` `delivery`: accepted candidate → authorized intent → denied `HUMAN_REQUIRED` → external-authorization-derived APPROVE → branch create + commit + push → `FINALIZED` → remote ref observed, 3 intents/3 receipts, 0 pull requests, issue #1 untouched (`docs/evidence/s13/github-lanes/round-17/r17-gh-4/delivery.json`) | FIXED_AND_REAL_E2E_VERIFIED |

| AEH-V2-0128 | P1 | Change routing / execution-policy reconciliation | A bounded change request whose concrete scope lists two files (or any 1 < files ≤ 5 with a DIRECT semantic recommendation) fails closed with `AEH_OPERATION_SUPERVISOR_UNAVAILABLE: initialization failed after 2 bounded attempt(s): EXECUTION_POLICY_STALE: durable ResolvedOperationPolicy does not match the current execution inputs.` in the round-18 `multi-worker` lane, before any participant runs. | Two delegation rules disagree: `triageChangeWithSemanticAssessment` only raises a DIRECT recommendation via `files.length === 0 || nonConcrete || files.length > 5 || cross-module || scopeClarity LOW || decompositionNeed || coordinationNeed`, while the deterministic contract floor (`createRoutedContract` → `triageChange` → `resolveImplementationRoute`) raises DELEGATED whenever `expectedWorkUnits = files.length > 1`. `runChangeOperation` selects its branch from the operation triage route (DIRECT) but seals a contract whose `routing.route` is the strengthened DELEGATED route; `runTask` then initializes the supervisor against the contract route while the durable policy/bootstrapped semantics still say DIRECT, so the identity compilation fails stale instead of reconciling or rejecting with a typed route-conflict error. | `src/core/triage.ts` (`triageChangeWithSemanticAssessment` delegation floor), `src/agents/routingV2.ts` (`resolveImplementationRoute` `workUnits > 1` branch), `src/core/contract.ts` (`createRoutedContract` deterministic floor), `src/operations/change.ts` (DIRECT/DELEGATED branch selection and sealed-contract route reconciliation) | None yet — required: a focused regression proving the sealed contract route cannot exceed the operation execution semantics without a deterministic rebind (as `rebindEscalatedChangePolicy` does for FORMAL_SDD), or that the two delegation floors are defined once; then rerun the multi-unit change lane. | Round-18 reproduction: `S13_ROUND=18 S13_RUN_ID=r18-mw-exp1 S13_GOV_LANES=multi-worker` against build digest `ca1d8f56…` (source `e3dad34`); the sealed worktree contract records `route: DELEGATED` with evidence "The bounded work has multiple coordination units…" while the operation intent and durable policy record `DIRECT/NONE`; lane artifact `docs/evidence/s13/governed-lanes/round-18/r18-mw-exp1/multi-worker.json`. | Not applicable (fail-closed before any provider participant; 2 real sessions were the lane Lead and the materialized supervisor only). | **FIXED_AND_REAL_PROVIDER_EXERCISED (R20-F2 reconciliation).** Repair: the single deterministic delegation floor `requiresDelegatedPlanningV1` (`src/agents/routingV2.ts`) is consumed by both `resolveImplementationRoute` and the semantic `triageChangeWithSemanticAssessment`, and `src/core/triage.ts` no longer equates a concrete file count with work-unit count (`expectedWorkUnits: Math.max(files.length,1)` removed); the sealed contract route can no longer exceed the operation's semantic route. Regressions (`npx vitest run tests/triage.test.ts tests/changeRouting.test.ts` 2 files / 15 tests passed). Live verification: governed `change-multifile` lane `r19-gov-3` on freeze F1 (buildDigest `408b178d…`) terminal **SUCCEEDED** rev 81, route DIRECT, sealed-contract route check PASS, AcceptanceOracle `ACCEPTED`, 516 s, workspace accounting 5/5/0; lane sha256 `712fa5c17297f7f1655616d8e2a027b99b249242492058259a69b35f5bdbad38` (see Round-19 WorkGraph; the earlier "live lane run pending" status was stale). |

| AEH-V2-0129 | P1 | Multi-worker waves cannot run parallel units: `workflow.planning.worktreeIsolation` is declared but never consumed, so same-wave implementers share one candidate workspace and collide on the provider write lease (S13 round 18 multi-worker lane, `AEH_PRODUCT_DEFECT`, OPEN) | Round-18 `multi-worker` run `r18-mw-3` (operation `CHANGE-20260929T151732Z-248e53a3`): the frozen Planner decomposed the two-requirement request into exactly two work units (`farewell-module`, `greeting-constant`, one wave, `harness.plan.ready` workUnits=2 waves=1), then both Implementer launches failed and the operation terminalized `FAILED` with `PASEO_PROVIDER_LEASE_TAKEOVER_BLOCKED: provider lease lease:512ab495-… belongs to another or stale lifecycle without a current-session takeover binding.` The conflict predicate matches leases on the same provider/project/canonicalRoot/workspaceId, so the second parallel unit attempted to take over the first unit's write lease in the shared workspace. | `workflow.planning.worktreeIsolation` exists in `src/core/types.ts` and `src/core/config.ts` (and defaults to `true` in `templates/project.yaml`) but no product code reads it; the wave executor launches every same-wave unit in the operation workspace, so parallel units are neither isolated nor serialized and the provider lease lifecycle fails closed on the second launch. | `src/agents/waveExecutor.ts` (per-unit workspace materialization/serialization), `src/core/run.ts` (executePlannerWaves wiring), `src/paseo/launchSpec.ts`/`src/runtime/providerLifecycle.ts` (workspace-keyed lease identity) | None yet — required: honor `worktreeIsolation` by materializing a per-unit candidate worktree with an explicit wave-base rebase and assembly receipt (or serialize same-workspace units deterministically with a single writer lease); then rerun the multi-worker lane to a terminal result with >=2 wave units and barrier evidence. | Round-18 lane `docs/evidence/s13/governed-lanes/round-18/r18-mw-3/multi-worker.json` (checkoutUntouched true; task-dag check PASS, barrier/assembly not reached); the same harness with a single-unit decomposition (`r18-mw-2`) reached terminal `SUCCEEDED`/oracle `ACCEPTED`, isolating the defect to parallel-unit execution. | Not applicable (fail-closed before any implementer session). | **ROUND 19 FIXED AND REAL E2E VERIFIED (multi-worker).** Repair: per-unit isolated launches now derive a deterministic provider write-lease workspace key (`providerLeaseWorkspaceKeyV1`, `src/runtime/providerLifecycle.ts`, consumed by `withProviderSessionLease` in `src/paseo/runtimeCore.ts`) so same-wave units no longer collide on `operation.workspaceId`; `workflow.planning.worktreeIsolation` is now consumed (`waveConcurrencyV1` in `src/agents/waveExecutor.ts`: true → per-unit candidate worktrees with bounded concurrency, false → deterministic serialization of same-workspace writers) and recorded in `harness.plan.ready`. Regressions (`npm run typecheck` exit 0; `npx vitest run tests/paseoProviderLeaseLifecycle.test.ts tests/parallelism.test.ts` 2 files / 18 tests passed). Live lane: `r19-mw-1` on freeze F1 (`1183af86…`, build `408b178d…`) — terminal `SUCCEEDED` rev 145, `workUnits=2 waves=1 conflicts=0`, wave barrier PASS (8 checks, 0 failed waves), 2 per-unit assembly receipts, 7 real sessions, `checkoutUntouched: true`, workspace accounting 6/6/0 `accounted: true`; lane sha256 `0108aa2dbb342d0a183956dfc6af9fe5347bf655acf343af032f007965c608d7` (`docs/evidence/s13/governed-lanes/round-19/r19-mw-1/multi-worker.json`). `distributed-execution` rerun on the same freeze records its own result in the Round-19 WorkGraph. |
| AEH-V2-0130 | P1 | Campaign resource-lifecycle accumulation and missing terminal/recovery reconciliation (campaign safety amendment 2026-09-29; `AEH_PRODUCT_DEFECT`, OPEN) | Long-running AEH/Paseo campaigns accumulate resources after their owning operation is terminal: at the campaign census, 96 stale fixture provider/MCP processes (59 named by R18-F1 plus 37 older deleted-cwd orphans) were still alive, 48 fixture Paseo workspaces and 507 staging directories remained, and the Paseo store holds 415 campaign-owned agent records (only 64 could be archived through the daemon; the rest are dormant store records the current daemon no longer manages). Provider sessions/processes survived their operations' terminal states and host restarts. | Investigation required. Working hypothesis (not assumed): AEH has no durable operation→provider/resource ownership record and no terminal/recovery reconciliation/release path; campaign cleanup is ad hoc (PID/glob based) and Paseo-managed sessions survive as dormant records. Classify the actual owner from evidence before repair; do not assume a Paseo heap leak. | `src/operations/`, `src/runtime/providerLifecycle.ts`, `src/paseo/`, `tests/packed/` campaign harnesses | REQUIRED ARCHITECTURE: durable operation ownership → durable provider/resource ownership → terminal/recovery reconciliation → safe release/archive → verification. Recovery/restart must distinguish live owned resource, terminal-operation orphan and unknown/unowned resource; only terminal-operation resources with proven AEH ownership may be automatically reclaimed; never garbage-collect unrelated user Paseo sessions. Add policy-driven concurrency/resource backpressure when providers can spawn without an operation/campaign ceiling (no machine-specific memory constant in Core). | REQUIRED TEST: a bounded repeated-operation resource-stability campaign with trend/bounded-resource assertions (no monotonic growth of active agents, workspaces, provider sessions/processes), required before final self-hosting acceptance; S14's disposable self-hosting composite must include resource cleanup/reconciliation. | Campaign census `docs/evidence/s13/round-20-resource-census.json`; round-19 cleanup inventories (`docs/evidence/s13/round-19/round19-cleanup-*.json`, 96 processes / 48 workspaces / 507 staging); R18-F1. | OPEN (durable architecture still required) — harness terminality gate FIXED (R20-F1). ROUND 21 EVIDENCE: the bounded repeated-operation stability pass `r21-resource-1` (2 iterations of `change-direct` in a shared isolated root, `S13_GOV_KEEP=1`) produced truthful **RESOURCE_ACCUMULATION_OBSERVED** evidence: the terminal SUCCEEDED iteration left 4 owned workspaces + 4 idle managed agents + 28 fixture provider/MCP child processes beyond its terminal state (ceilings are 0) and the second iteration grew the residue (6/6/29), with `resource-stability.no-monotonic-growth` FAIL; the harness reconciliation terminality gate (`resource-stability.reconciliation-terminal-only`) PASSed — no non-terminal-owned resource was archived or terminated (AEH-V2-0112). Summary `docs/evidence/s13/s13-resource-stability-round-21-r21-resource-1.json` sha256 `8b67fc12258b426a26ac19a55f845d67061b16abc01cdb425bb32cc57ec42bb0`. Round-21 additionally reclaimed 17 orphaned Paseo-managed `opencode serve` processes (≈5.3 GB RSS) proven AEH-owned by `/proc/<pid>/environ` (`PASEO_AGENT_ID` + `AEH_CONTROL_ROOT`) whose fixture staging was already removed, plus the two `error`-state assessor agents of the aborted `r21-dist-2`/`r21-dist-4` attempts (archiving released their servers); preserved and reported: servers/agents of the two non-terminal `r20-dist-2`/`r20-dist-4` operations, 10 unknown-owner opencode servers (no managed agent id; rule 6), and one 5.4 GB unknown-owner server. ROUND 20 EVIDENCE: source review confirms `terminalizeOperation` (`src/operations/controller.ts:787`) performs no resource reconciliation for SUCCEEDED/FAILED (only the CANCELLED path reconciles, `:640-730`); per-turn `runWithOperationProviderLease` release is sound; supervisor generations are archived but participant/lead agent records and operation workspaces have no product-side archive path; no recovery orphan sweep exists. Runtime evidence: 7 terminal-operation candidate workspaces + 4 idle supervisor generations survived their terminal operations until hand-archived (archiving the idle agents released 8 provider children); stalled `r20-dist-2`/`r20-dist-4` operations kept live controllers, 19+19 orphan child processes and 6 workspaces with no reconciliation. Runnable bounded test delivered: `tests/packed/s13ResourceStabilityCampaign.mjs` (trend assertions T1–T5; reconciliation baseline return); bounded pass handed to Round 21. Round-20 artifacts: `docs/evidence/s13/round-20/resource-reconciliation.json`, `distributed-2-orphan-reconciliation.json`, `round20-cleanup-after.json`. **Round-23 addition (STILL OPEN, carried to S14):** the r23-gov-8 first repair attempt stopped on an `external_directory` prompt after the model read a stale sibling isolated-root path (`/tmp/aeh-direct-…`) still referenced by a shared Serena symbol index — stale provider/server caches survive failed attempts and can poison later attempts; the round-23 census (`docs/evidence/s13/round-23/round23-cleanup-after.json`, sha256 `bba410db…`) and reconciliation (`docs/evidence/s13/round-23/resource-reconciliation.json`, sha256 `93164a99…`) record 13 servers ≈4.9 GB preserved with 5 owned by preserved non-terminal operations and 8 of unknown ownership, plus the dead-controller evidence for the preserved non-terminal r22/r23 operations. **FINAL CLOSURE (2026-09-30):** the closure slice implemented the required architecture in the product and re-verified it: durable per-operation resource registry (`.harness/operations/<id>/resources.json`) + deterministic terminal/recovery reconciliation with a durable receipt (`resource-reconciliation.json`), exact ownership proof (operation bindings, product-minted `aeh.operation` labels, provider lease lifecycle, managed process handles, exact session inspection, exact staging paths), `LIVE_OWNED`/`TERMINAL_ORPHAN`/`UNKNOWN_OR_UNOWNED` classification, idempotent `alreadyReconciled`, terminalization reconciliation after the terminal transition, restart/terminal-entry recovery and a startup sweep, a registered/terminated liveness monitor, immediate exact-id release of pre-operation triage assessor sessions, and generic policy ceilings (`RESOURCE_CEILING_EXHAUSTED`, defaults 64 owned resources / 16 concurrent provider sessions per operation; no machine-specific constants). REAL_PROVIDER evidence on closure candidate buildDigest `d266692df02c95bf432ac6c9578e127fb16584aa2cc18d469d2c844ed3b0a2ab`: `closure-resource-6` two real governed iterations terminal SUCCEEDED with product receipts `cleanupComplete: true` and census baseline 0/0/0 (`docs/evidence/closure/resource-stability-product.json` sha256 `94ed891de6e1d72b4039f7a733b55e6625a94a48df61fd5f0288147c5a157f14`, `campaignSideCleanup: NONE`, `preservedIntact: true`, `postReconciliationNoGrowth: true`; five truthful failed runs retained as superseded history); `closure-disposable-1` terminal SUCCEEDED / oracle ACCEPTED / rows PASS with checkout byte-identical and 13/13 receipt dispositions (`docs/evidence/closure/disposable-self-modification-campaign.json` `17e9b926…`); lifecycle cancel/recovery 2/2 PASS (`docs/evidence/closure/paseo-lifecycle-cancel-recovery.json` `ead63493…`); focused regressions A–J plus K/L/M 12/12 and closure focused set 43 suites / 201 tests (`docs/evidence/closure/focused-regressions.json` `0d670ab4…`); S14 composite regenerated `ready: true`, 21/21 items, 0 failed artifacts, policy digest `1be97df3…`, report `9b9e2134…`, byte-reproducible. **Status: FIXED_AND_CLOSED (product-owned terminal/recovery reconciliation; real-provider exercised).** |
| AEH-V2-0131 | P1 | distributed execution / worker session protocol | `r20-dist-4` claimed and ran a real distributed job but the worker receipt was FAIL: `EXECUTION_BINDING_SESSION_UNSUPPORTED: transport 'inherit' has no approved pre-prompt session materialization boundary`. | The distributed job froze the local wave selection (`transport: inherit`), which only resolves in-process and has no meaning in the detached worker; the two-phase worker session protocol has approved idle-session boundaries for `direct`/`podman` (and an explicit in-authority-domain Paseo path), while the worker clone has no operation state root. | `src/distributed/types.ts`, `src/distributed/worker.ts`, `src/workers/agentPrompt.ts` (`prepareRuntimeSession`, `prepareAgentExecutionIdentity`, `resolveStructuredResultProvenance`), `src/workers/directProcess.ts` | FIXED: dispatch resolves/freezes a concrete worker transport at job creation (`resolveDistributedWorkerTransport`; inherited Paseo orchestration resolves to the provider idle-session boundary `direct` with the resolution recorded on the job envelope) and fails closed `DISTRIBUTED_EXECUTION_TRANSPORT_UNSUPPORTED` for `none`/unknown before submission; the worker re-validates fail-closed and never calls `prepareRuntimeSession` with `inherit`. The worker stays stateless with respect to controller authority (a draft that ran it under the operation context was withdrawn after the product correctly fenced it with `V2_CONTROLLER_FENCED`); the controller converts the settled remote result into the candidate-bound participant receipt. `prepareAgentExecutionIdentity`/`resolveStructuredResultProvenance` validate a propagated worker launch against the controller-issued authority + frozen blueprint when the durable operation record is absent. The ephemeral direct-worker home projects only the exact provider auth file (OpenCode `auth.json`/Codex `auth.json`, 0600, removed with the turn), mirroring the certified `CodexAgentProvider`. | Focused regressions: `tests/distributedExecutionIdentity.test.ts` (11 tests: inherited transport resolution + no `inherit` in the frozen job + controller-recorded receipt + stateless worker authority + forged unresolved/mismatched transport rejection), `tests/agentPromptStructuredResultProvenance.test.ts` (24), `tests/directCandidateLifecycle.test.ts`, `tests/workerExecutionAuthority.test.ts`; `npx tsc --noEmit` exit 0. | REAL_PROVIDER: `distributed-execution` lane `r21-dist-14` terminal **SUCCEEDED** (rev 137), AcceptanceOracle `ACCEPTED`, oracle all PASS — `distributed-execution.lease` prepared=2/released=2/completed=2, `distributed-execution.worker-receipt` PASS with real session `ses_f0faef79affe4CUexX9g65Hk4T` and 1 changed file, 5 terminal participant receipts; packed artifact `61e008f3…` from packed build `release-1790737862108-997061-89156943` (buildDigest `599ed363…`), source `e3dad34`, `checkoutUntouched: true`. Lane `docs/evidence/s13/governed-lanes/round-21/r21-dist-14/distributed.json` sha256 `69b735d04a75f1d37795e6d930854a20d2da7822caf9f44b5b2e59268e1042d0`; summary sha256 `8c7c30d795351cc471d31d54e7b95eda5e013d555f9ed2abea6d0215a67e1b75`. | FIXED_AND_REAL_PROVIDER_EXERCISED (S13 round 21). |
| AEH-V2-0132 | P2 | Bounded participant provider turns have no operation-level deadline; a stalled Reviewer turn hangs the operation while the watchdog suppresses supervisor wake (S13 rounds 22/23, `AEH_PRODUCT_DEFECT`, OPEN) | `r22-gov-3` (F2-final `repair` lane) reached remediation COMPLETED and then stayed `review` RUNNING from 07:40:05Z to the 1,800 s bound with the reviewer session settling only late and no accepted structured result; `r23-gov-6` reproduced it: the reviewer session (`68d21387…`, opencode-go/mimo-v2.6-flash) ran a 24+ minute deliberation loop, 117+ `operation.watchdog.supervisor-suppressed-active-children` traces reported 1 active runtime participant / 5 unresolved, and the operation terminalized FAILED only when the campaign aborted it (`candidate.assurance.reviewer.0.reviewer: Assigned Reviewer 'reviewer' did not return valid structured …`). | The persistent-supervisor continuation has a 300 s ceiling (AEH-V2-0108) but a bounded worker/reviewer turn has no participant-level deadline; `watchdog` observes `activeRuntimeParticipants>0` and refuses to wake the supervisor, so a single stalled participant turn deadlocks the operation until an external bound. | `src/operations/controller.ts` (watchdog), `src/workers/agentPrompt.ts` (provider turn wait), `src/agents/reviewLifecycle.ts` (reviewer round), `src/operations/state.ts` | Required repair: a bounded per-participant/provider-turn deadline (or explicit provider-quiescence reconciliation) that classifies a stalled turn as a typed failure and either re-dispatches within the frozen remediation budget or terminalizes truthfully; the watchdog must distinguish "children making progress" from "one turn stalled past its deadline". No gate may be weakened. | `docs/evidence/s13/governed-lanes/round-22/r22-gov-3/repair.json` (sha256 `61b588fb…`), `docs/evidence/s13/governed-lanes/round-23/r23-gov-6/repair.json`, `docs/evidence/s13/round-23/r23-gov-6-repair-aborted-attempt.json` | OBSERVATION_OPEN — intermittent provider/model behavior; does not invalidate terminal rows because terminalization remains fail-closed. Carried to S14. |
| AEH-V2-0133 | P2 | A superseded failed repair-attempt identity blocks objective completion after a successful retry (S13 round 23, `AEH_PRODUCT_DEFECT`, OPEN) | `r23-gov-8` completed the full forcing geometry (repair packet attempt 1 `FAIL` → attempt 2 `PASS`, `review=ACCEPTED`, `delivery=COMPLETED`, AcceptanceOracle `ACCEPTED`) but terminalized FAILED with `OPERATION_FAILED: objective.completion: Objective Definition of Done is blocked: PARTICIPANT_INCOMPLETE.` for `participant:f77f3e8636d523a3`, the controller-issued identity of the first repair attempt whose provider turn stopped on a permission prompt and produced no result; the retry identity settled `COMPLETED` and the repair loop itself recorded `harness.repair.finish PASS`. | Objective completion requires every required participant identity to be complete and does not resolve an abandoned attempt identity through the successful retry that superseded it (unlike AEH-V2-0124, which reconciles a transient FAILED against a settled receipt for the same identity). | `src/operations/participantAccounting.ts`, `src/operations/state.ts`, `src/operations/v2Contracts.ts` | Required repair: bounded work units whose retry/reroute attempt settled successfully must not be blocked by a superseded attempt identity that never produced a result (lineage/attempt-chain-aware accounting), while a unit with no successful attempt still fails closed. Required regression: failed attempt + successful retry → objective completion complete; only-failed attempts → blocked. | `docs/evidence/s13/governed-lanes/round-23/r23-gov-8/repair.json`; `docs/evidence/s13/s13-governed-operations-round-23-r23-gov-8.json` | OBSERVATION_OPEN — fail-closed today (no wrong PASS), but blocks legitimate PASSes after transient provider stops. Carried to S14. |
| AEH-V2-0134 | P3 | Intermittent `AEH_RESULT_PROVENANCE` failure reading an accepted Explorer result in a multi-worker operation (S13 round 22, OPEN OBSERVATION) | `r22-gov-1` multi-worker lane terminalized FAILED at rev 43 in discovery with `AEH_RESULT_PROVENANCE: result turn revision or contract does not match the active channel.` at `readVerifiedAcceptedArtifact` (`workers/resultGateway`) while reading the Explorer accepted artifact; the durable artifact exists (`…/explorer/0001-0001-c6a404e0-…json`, 16,065 bytes, sha256 `f4099dc4…`, turnId `0001-0001-c6a404e0-…`, sequence 1, revision 1). The immediate same-source rerun `r22-gov-2` PASSed. | Not reproduced deterministically; the channel/revision equality check is correct-by-design and must not be weakened. Recorded as an open observation with exact evidence per the campaign instruction; no fix created. | `src/workers/resultGateway.ts`, `src/workers/agentPrompt.ts` | If it recurs: capture the channel binding and the result turn revision in the lane artifact and establish whether the active channel was rotated by a retry before the result was read. | `docs/evidence/s13/governed-lanes/round-22/r22-gov-1/multi-worker.json`; `docs/evidence/s13/s13-governed-operations-round-22-r22-gov-1.json` (sha256 `522f40bf…`); rerun `r22-gov-2` PASS (`4d8d8a9b…`) | OPEN OBSERVATION — carried to S14; no gate change. |
| AEH-V2-0135 | P2 | S14 bounded stability campaign confirms terminal-owned residue and campaign-side reconciliation while product-side terminal reconciliation remains absent (AEH-V2-0130 progress; S14, `AEH_PRODUCT_DEFECT`, OPEN) | Two real governed operations (S14 `s14-resource-2`, iterations 1/2) terminalized `SUCCEEDED`/`ACCEPTED` and each left raw terminal-owned residue (iteration 1: 5 workspaces / 5 agents / 29 provider-MCP processes on the first run, 3/3/28 on the corrected run; iteration 2: 4/4/29) within the declared per-operation ceilings (12/12/96). The campaign-side terminality-gated reconciliation archived/terminated all of it and returned the baseline to 0/0/0 in every iteration (`baselineReturned: true`, `preservedIntact: true`, `postReconciliationNoGrowth: true`), so active resources do not accumulate — but no product path performs this. | Same root cause as AEH-V2-0130: no durable operation→resource ownership record and no terminal/recovery reconciliation in the product; the S14 evidence proves the verification half only. | `src/operations/`, `src/runtime/providerLifecycle.ts`, `tests/packed/s14ResourceStabilityCampaign.mjs` | Required architecture unchanged from AEH-V2-0130: durable operation ownership → durable provider/resource ownership → terminal/recovery reconciliation → safe release/archive → verification; classify live-owned / terminal-orphan / unknown and never GC unrelated user sessions; add policy-driven concurrency/resource backpressure (no machine-specific memory constant in Core). | `docs/evidence/s14/resource-stability-campaign.json` sha256 `090dace900895ec2a98fef1c20840fcff28fe6242f0e64186ba81f2bf0b43fb0` (classification `OPEN_PRODUCT_DEFECT_EVIDENCED`); superseded first run `docs/evidence/s14/resource-stability-campaign.superseded-s14-resource-1.json` sha256 `e17ab9ad…`. | **FIXED_AND_CLOSED (2026-09-30 final closure)** — the same product repair as AEH-V2-0130 now makes terminalization itself return owned resources to baseline without campaign-side cleanup: closure `closure-resource-6` on candidate `d266692d…` ran two real bounded `change-direct` iterations, both terminal SUCCEEDED with product receipts `cleanupComplete: true` and census 0/0/0 (`docs/evidence/closure/resource-stability-product.json` `94ed891d…`, `campaignSideCleanup: NONE`); the disposable campaign and lifecycle lanes confirm the same product path (`docs/evidence/closure/`). Superseded closure failures (five) are preserved as truthful history. |


### Round-10 reviewer findings R9-F5…F8

| Finding | Classification | Disposition |
| --- | --- | --- |
| R9-F5 — Conformance audit-pack release citation was stale | DETERMINISTIC | FIXED: the audit artifact binds `release-1790567348887-372966-19891328` with build digest `8e949b0113ddd90ae0573826b8ecc348915ae2c65fd6a77c99f82ba9b6a9ab84`; conformance now cites that release. |
| R9-F6 — lifecycle/GitHub summaries used stale round labels; GitHub rows lacked packed-build binding | DETERMINISTIC | FIXED AND ROUND-10 EXERCISED: `S13_ROUND` is required, summary/lane outputs are round/run-specific and immutable, GitHub summary/per-lane evidence binds installed packed artifact/build identity. Existing Round-9 artifact bytes remain unchanged. Round-10 lifecycle run `s13-real-paseo-lifecycle-round-10-lifecycle-round10-1.json` is explicitly round 10, binds the current build, and passes cancel/recovery. Round-10 attempt-4 GitHub lanes bind build digest `0fc216b8ebd40b942c19556b2f1e7d6c44be802bb27142d41deeab2cc7ee71ac` but are `NOT_ACCEPTED (tainted)` per AEH-V2-0114. |
| R9-F7 — timed-out direct controller's post-campaign terminal observation was not persisted | DETERMINISTIC | FIXED AND EXERCISED: Round-10 direct post-read artifact records operation `CHANGE-20260928T163444Z-5de819ea` as `SUCCEEDED`, controller PID exited, and candidate worktree absent only after its recorded archive. SHA-256 `c80bdbf812312eb86cecf6aa04d2bb12888d3632cfedeebcbb4ee96031413a55`. |
| R9-F8 — Round-9 W35 and frozen capability plan could be misread as current compliant row status | DETERMINISTIC | FIXED IN WORKGRAPH/STATUS/CONFORMANCE: W35 excludes tainted Round-9 authority/import PASS claims, the frozen 22-row availability plan is explicitly distinguished from current evidence dispositions, and the complete current disposition remains `SLICE_BLOCKED`. |

Round-10 reviewer dispositions and evidence/resource inventory: `docs/evidence/s13/round10-review-dispositions.json` — its original SHA-256 `cdf2c2f40e9f9693dae83e3bd4f0fe378f17231103f94ffbf31f0fe443a04393` was independently reproduced by the Round-10 review; Round 11 corrected the tainted `round10ExternalEffectsAuthorization`/R9-F1A `cleanLane` statements in place and the file's current digest is recorded in the Round-11 WorkGraph. Round-10 independent review: `docs/evidence/s13/review-round10.json` SHA-256 `ab31f436327641c4e8ec326ac0f4b5e13ac1c83b9f49e7f7947b60f18ccd0d44` (verdict `SLICE_BLOCKED`; R10-F1…F9).

## CI alignment campaign — PR #64 (2026-09-30–2026-10-01)

The Core Architecture v2 CI alignment campaign reconstructed PR #64 and baseline run 36782724704 at head bbe2bc6624d2b1a51f001f6dec566786590b1296. Test-level causes, classifications, owner-layer repairs, lane coverage, and local validation results are recorded in [CORE_ARCHITECTURE_V2_CI_ALIGNMENT.md](CORE_ARCHITECTURE_V2_CI_ALIGNMENT.md). The repairs update stale Paseo/terminalization/context fixtures, add deterministic suite taxonomy and explicit rootless/provider/package/supply-chain lanes, and bind strict supply-chain evidence to the real packed tarball. Historical checkpoint 325605ebf842d098aa0fdbe556c433c3aac9cc0b/run 36803071091 passed 11 jobs and failed packaged-consumer at P1 and rootless-isolation at E2. At the pre-amendment snapshot on 2026-10-01, PR #64 head was 48276ea842fa3af475eb32632d995b18f755f17c; run 36817138377 passed 12/13 jobs. The rootless hosted lane passed 3/3 on Ubuntu 22.04/bwrap 0.6.1. Packaged-consumer passed pack, install, CLI/MCP, init, toolchain, Podman setup, and policy sync, then failed at P3 because the clean consumer lacked paseo and @getpaseo/client for the session-scoped Semantic Assessor; refusing CLI fallback preserved native identity/permissions. The typed P3 test-harness repair uses the existing deterministic INTENT fixture and passed a direct built-CLI smoke. As of this snapshot, it was the next checkpoint candidate and had no hosted result. Package output capture and Podman provisioning were verified repaired by run 368171. The Ubuntu 24.04 AppArmor capability restriction remains the best-supported explanation for the prior rootless failure; run 368171 confirmed the Ubuntu 22.04 environment repair. Core-v2 daemon, provider, AcceptanceOracle, objective-completion, authority/retrieval, reversibility, and artifact-identity gates remain fail-closed. The normative TARGET was not changed.

Before restart, unrestricted local typecheck/build/release, deterministic/system/human/adversarial/scenario, bwrap (3/3), provider-contracts, strict full-stack, Pact, package consumer, and exact packed-artifact supply-chain validation passed as recorded in the inventory. In the restarted managed shell, rootless isolation fails at bwrap network namespace setup with NETLINK_ROUTE: Operation not permitted and packaged CLI execution can return empty captured child stdout; release:check typecheck/build pass but its deterministic stage is blocked by listen EPERM and child-process capture behavior. Those local results are environment-bound and do not replace hosted evidence. The complete full-stack and supply-chain evidence uploads remain configured. At the pre-amendment snapshot, run 368171 had passed OCI integration, full-stack, supply-chain, test/release, scenario, system reliability, adversarial, human, provider contracts, validation provider, Pact, and rootless-isolation 3/3; packaged-consumer was the sole failing job. The typed P3 fixture repair was the next checkpoint candidate and had no hosted result as of that snapshot; fresh independent review and green checks on the amended SHA remain required. This section records CI evidence, not architecture certification or merge approval.

## CI alignment update — run 36822314115 (2026-10-01)

PR #64 head `ff2b9f279fcf6c974f626875dc99248881613a79` was checked through PR merge SHA `64e3f963f7bf70e287da13f0f9aefb13e26c0450` and completed with 12/13 GitHub Actions jobs passing. The P3 packaged-consumer and rootless-isolation repairs are confirmed by this run; test/release failed at F5, `tests/security/isolation.test.ts` provider detection. The test invokes production capability discovery with ambient PATH, while the Ubuntu 24.04.5 runner image supplies Podman 4.9.3; detection can await `podman info` for up to 20 seconds against Vitest’s 5-second default. The hosted log identifies the timed-out detection call, not the exact subprocess, so the provider-probe cause is supported by the runner inventory and production call path rather than a provider-specific trace. The test now uses a temporary root and empty PATH, asserts explicit missing providers, marks the suite sequential, restores PATH and cache in `finally`, removes the temp root, and passes locally 10/10 in 417ms. Classification: `ENVIRONMENT_COUPLED_TEST`. The repair is unpushed and has no current hosted result. Runs 368030 and 368171 remain historical; no TARGET or product gate changed. Full evidence and status matrix: [CI Alignment Inventory](CORE_ARCHITECTURE_V2_CI_ALIGNMENT.md).

## Audit-snapshot reconciliation

This section corrects claims that the read-only architecture audit found stale or overbroad at commit f1518caf405f44851227cf73ea37b79267b1a098. It is a reconciliation note, not a roadmap. Current subsystem state, dependencies, and next gates are maintained in [Core Architecture v2 Status](CORE_ARCHITECTURE_V2_STATUS.md); evidence-lane boundaries are maintained in [Core Architecture v2 Conformance](CORE_ARCHITECTURE_V2_CONFORMANCE.md).

- **Repairer public path:** the earlier claim that candidate-mutation helpers had no production callers and that the live repair path mutated in place is superseded. Repairer candidate mutation and assembly support have production use. Remaining gaps concern complete frozen Repairer assignment and broader operation-state, impact, validation, review, and acceptance integration.
- **StructuredResult provenance:** candidate, generation, and epoch checks exist; blueprint and policy fields are optional in some paths. Production launchers still omit required full identity digests. Provenance is incomplete, not wholly absent or complete.
- **ExecutionCatalog:** the catalog exists and is used. Its policy and tool-authorization fields are not fully enforced. Do not describe the catalog itself as absent.
- **Operation and role policy:** ResolvedOperationPolicy and RoleInvocationPolicy are absent in the audited snapshot. They are distinct from ExecutionCatalog, which exists with incomplete policy/tool enforcement.
- **Semantic assessment:** typed assessment contracts and helpers exist, but the audited public paths do not construct the service consistently. Public semantic wiring remains open.
- **Paseo session binding:** a digest-bound binding module and tests exist, but the production Paseo runtime does not call it.
- **CandidateImpact:** the finding that impact did not recompile reviewer, validation, and acceptance requirements describes the audited snapshot and is superseded by S4 / AEH-V2-0041. S6 now implements the production AcceptanceOracle and objective-completion path; its deterministic gate was accepted by the Campaign Director on 2026-09-24. External certification remains separate.
- **Human decisions and cancellation (S3 remediation, 2026-09-24):** V2 HumanDecisions are consumed for exact policy-required action authorization, current-scope cancellation, and paired-session product choice. The Spec Manager BLOCKED boundary creates a typed, evidence-bound DecisionRequest and durable continuation/checkpoint; the Control Center App submits only the offered choice identity, and the controller consumes it once, revalidates current bindings, resumes SPEC_AUTHORING, and invalidates dependent execution semantics when requirements change. Campaign Director review found that consumed-choice recovery and checkpoint loading did not compare the complete saved/current binding. The Luna-owned follow-up now persists the original selected-decision binding, checks it against the ledger receipt and current operation, validates checkpoint bindings against both continuation and current operation, permits only the single authorized requirement revision under the same candidate/epoch, and never reuses a stale consumed choice. After takeover it validates the old receipt and unchanged candidate/revision, then opens a fresh current-binding HUMAN_REQUIRED request; intact unanswered requests can be safely reissued after policy/epoch drift only with unchanged candidate and execution revision. The bounded crash window after semantics persistence is recovered by the controller’s bootstrap policy binder before continuation load; the regression proves missing policy fails the load, bootstrap binds a current deterministic policy, and current checkpoint/resume validation then succeeds. Deterministic regressions cover revision, policy, epoch, digest-recomputed checkpoint tampering, same-epoch crash recovery, and takeover reconfirmation. Controller-owned mutations and active terminal transitions require current epoch plus controller token, while Lead acknowledgement binds the current Lead actor, epoch, and operation revision. Cancellation establishes and rechecks its fence before cleanup. Final S3 deterministic evidence is recorded in Conformance; browser, live OPA, real-provider, and packed adversarial/provider-fault certification remain separate pending lanes.
- **Acceptance and delivery:** validation, review/quality, and AcceptanceOracle are distinct. S6 persists a current controller-owned disposition and enforces it before delivery; where frozen policy requires managed Lead evidence, that evidence is collected before the disposition. S8 was independently accepted by the Campaign Director; external-provider/packed journey certification remains separate and pending.
- **Handoff:** the audited stale CLI path and pre-acceptance public auto-handoff finding are superseded by S8: normal issue execution enters the managed controller path, accepted finalization owns requested public effects, and standalone `sdd handoff` fails explicitly before project loading because it lacks managed acceptance context. Uncertain Paseo workspace identity still fails closed as HUMAN_REQUIRED. Deterministic fixtures do not establish live-provider behavior or complete packed reconciliation; the Campaign Director accepted S8 on its deterministic gate on 2026-09-25.
- **Packed campaign wording:** some campaign tests import src/** directly. Earlier descriptions of the campaign as entirely packed are overbroad; a complete freshly packed product journey is not established.
- **Serena evidence:** the recorded Serena smoke covers the tested MCP broker path only. It does not certify unused provider lease/release lifecycle or real-model execution.
- **Validation counts:** dated command totals below and in findings are historical checkpoint evidence. They are not results rerun during the f1518ca audit or this S0 work, and they are not current certification claims.
- **Delivery claim:** the PR body in src/delivery/finalize.ts says the lead-acceptance workflow completed before PR creation. S6 now enforces the current AcceptanceOracle disposition before effects and collects managed Lead evidence first when policy requires it; the PR wording remains policy-conditional and does not establish Campaign Director acceptance or external-provider certification.
- **Working-tree note:** the historical campaign note about uncommitted implementation/test changes described an earlier point in the campaign. At the f1518caf405f44851227cf73ea37b79267b1a098 audit snapshot, git status was clean. The fsmonitor warning did not prevent branch status from being read.

## Historical campaign notes

The following items preserve campaign history and must not be read as current certification or roadmap:

- The branch campaign was based on the repository state described in the recorded findings. No history rewrite or destructive Git recovery was performed.
- The accidental operation AUDIT-20260922T221454Z-7aa6fa6b was cancelled after inspection and reached terminal FAILED; its workspace was the development checkout and it created no additional Git worktree.
- A historical integrator checkpoint recorded an aggregate isolated packed system campaign and a packed Home/Control Center smoke. The cited campaign included tests that imported source directly; these records do not prove all-packed journeys or current certification.
- Historical environment notes reported no Paseo executable, browser automation executable/browser, or real model/provider journey at that checkpoint. Serena's local MCP broker smoke was provider-process evidence for that path, not model certification.
- AEH workflows are not used to govern changes to this development checkout. The S0 documentation work did not run AEH.

## Last green implementation checkpoint CI verification — PR #64, run 36824682741 (2026-10-01)

At the time of run [36824682741](https://github.com/JamesMorales04/agentic-engineering-harness/actions/runs/36824682741), PR #64 was OPEN at head `d21d6b5db6b36b8387af7687d9488c4f95229de1`. GitHub completed 13/13 jobs successfully on that implementation/workflow checkpoint. The matrix for that SHA is: test/release, scenario-matrix, system reliability campaign, adversarial-system, human-e2e, provider-contracts, validation provider contracts, real Pact provider verification, full-stack-contract, supply-chain, OCI integration environment contract, packaged-consumer, and rootless-isolation — all PASS. This hosted run verifies the final F5 deterministic provider-detection fixture repair and the earlier P3 packaged-consumer/rootless lane repairs. Runs 368030, 368171, and 368223 remain historical evidence, including their truthful failures and subsequent owning-layer repairs.

Architecture acceptance is separately supported by `docs/evidence/s14/review-s14.json` and `docs/evidence/closure/review-closure.json` (both `SLICE_ACCEPTED`); the closure evidence records AEH-V2-0130/0135 `FIXED/CLOSED`. The closure report includes an umbrella S1–S12 statement, but the explicit S10 WorkGraph still requires fresh independent re-acceptance and no S10-specific accepted review artifact was found. STATUS therefore keeps S10 open; do not treat the umbrella summary or S14/closure slice acceptance as evidence that S10 passed or as overall merge readiness. A later status-doc candidate `963cc3010dc5b372a282da4f37859a8ce6a27796` passed all 13 required jobs in run [36832161124](https://github.com/JamesMorales04/agentic-engineering-harness/actions/runs/36832161124) and was independently accepted. Exact-SHA results for subsequent campaign snapshots are attached to the current PR head. Ordinary PR review or branch-protection controls are outside this evidence. S10 fresh re-acceptance remains the separate merge gate; no merge-readiness claim is made here.


## Current S10 acceptance record — 2026-10-01

The earlier CI-checkpoint paragraph records the pre-reacceptance state: S10’s independent review artifact had not yet been added at that point. The separate fresh reviewer record is now present at [review-s10-final.json](evidence/s10/review-s10-final.json), SHA-256 `f940aa6a72587bccf800ceade409de9b22c175af21fa83b2b99cd63e3362a9cc`; verdict **`SLICE_ACCEPTED`**, reviewed head `87144f77a0c99088d385e848925643b5439422ea`. The reviewer found no material S10 product-contract change and no unknown source-impact classifications. This closes only the S10 acceptance-artifact gap; the named unrelated post-v2 findings above remain open/nonblocking and are not resolved by this record.

Fresh evidence: hosted rootless-isolation job `110435254361` in run [36881830507](https://github.com/JamesMorales04/agentic-engineering-harness/actions/runs/36881830507) passed the real bwrap campaign 3/3 on the reviewed SHA; the full run passed 13/13 jobs. Candidate-bound Trivy 0.70.0 evidence is [real-sast-campaign-85877349558477e21386d6e076393d15e2a90e26-release-1790866583369-387000-cc8194fb.json](evidence/s10/real-sast-campaign-85877349558477e21386d6e076393d15e2a90e26-release-1790866583369-387000-cc8194fb.json), SHA-256 `36a871fd9c8fcafdbb4c116b1d2974fa385ab17e510ba74bdbbf6a430b8e3641`; it binds a clean (`dirty:false`) build to implementation SHA `85877349558477e21386d6e076393d15e2a90e26`, proves r1 finding FAIL/r2 clean PASS, and rejects stale and tampered evidence. The prior canonical SAST artifact remains untouched; S14/closure composite inputs are unchanged, so neither composite was regenerated.
