[Air000000]

Scope

P1-A/P1-B implementation under the approved plan:

docs/superpowers/plans/2026-09-20-labpilot-p1-result-aware-iteration.md

PR remains Draft until the final P1-B review/full-CI gate.

P1-A implementation review — ACCEPTED

Reviewed head:

e3116ecb7f5799e550317a02ea35d868ffbb23b0

Remote evidence:

• GitHub CI #300: SUCCESS
• local-equivalent CI gate: SUCCESS
• production image/build-context hygiene: SUCCESS
• local ./scripts/check_ci.sh: 277 passed, 1 skipped, 5 deselected

Task 1 — AgentRun result provenance

ACCEPTED.

• Alembic head 0013_result_aware_agent_runs
• nullable source_batch_summary_revision_id
• RESTRICT FK to exact BatchSummaryRevision
• initial Agent compatibility preserved

Task 2 — ResultContext + ContextCompiler

ACCEPTED.

• Agent-owned bounded ResultContext
• exact summary/result facts participate in canonical request hash
• initial request/hash semantics preserved when ResultContext absent
• 128 KiB ResultContext bound under existing 256 KiB request cap

Task 3 — exact result binding adapter

ACCEPTED.

• same BatchRow FOR UPDATE linearizes latest-at-create
• latest check + AgentRun insert happen in same UoW
• exact pinned summary load; no later-latest substitution
• no raw Artifact/stdout/stderr/raw invalid metric promotion
• agent_runtime -> results_validation dependency remains unchanged

Non-blocking note:
the lock-race test currently uses short sleeps to order waiters. It is stable in local/CI evidence; harden test synchronization only if flakiness appears.

Task 4 — Follow-up AgentRun service/API

ACCEPTED.

• optional Follow-up ports preserve all initial Agent construction paths
• follow-up creation fails closed when runtime ports absent
• local-only + remote provider blocks before ResultContext load/provider call
• exact ResultContext reaches allowed provider
• stop(goal_satisfied) completes without Proposal/Version/Batch writes
• follow-up propose and persisted propose replay fail closed until P1-B
• production composition uses one SqlAgentRepository and one shared SqlResultAwareAgentAdapter

P1-B Task 5 — ACCEPTED

Reviewed head:

fdb9738614c36ffa29d995ecc7593bb5366fcb65

Evidence:

• GitHub CI #302: SUCCESS
• local proposal service tests: 24 passed
• proposal loop regression: 8 passed
• Agent unit regression: 88 passed
• architecture guard: 12 passed
• local full gate: 282 passed, 1 skipped, 5 deselected

Verified boundaries:

• initial create_agent_revision() semantics unchanged
• follow-up Agent revision stays in the existing Experiment
• Experiment FOR UPDATE serializes revision-number allocation
• no second revision allocator introduced
• no new Experiment is created by follow-up revision creation
• exact replay requires same payload + Project + Experiment
• cross-Experiment replay of one origin AgentStep is rejected
• origin-step unique-conflict recovery returns only the exact winning revision
• Task 6 remains unimplemented at this reviewed head

Next authorized step

P1-B Task 6 only:

Follow-up ProposeDecision
→ load exact pinned ResultContext
→ derive Experiment from ResultContext
→ create_follow_up_agent_revision()
→ link ProposalRevision to AgentRun
→ proposal_created

Do not verify, approve, materialize ExperimentVersion V2, or create/start a Batch in Task 6.

Explicitly excluded

• historical Result Tool / P1-C
• controlled code modification / Worktree
• automatic Batch start
• new Agent decision kind
• UI redesign
• Slice 2 hardening expansion

P1-B Task 6 — ACCEPTED

Reviewed implementation head:

ae079be252fd62eff415ef040af4228ab45e2d61

Verified boundaries:

• fresh Follow-up propose derives Experiment only from exact pinned ResultContext
• proposal Snapshot mismatch fails before Proposal write
• follow-up materialization uses same-Experiment Proposal service
• crash after committed R2 replays the same origin AgentStep
• retry links the same R2; no R3 is created
• provider is not called again during persisted-proposal replay
• no automatic verification, approval, ExperimentVersion, or Batch is created

Task 6 is included in later successful CI #310 on the current PR head.

P1-B Task 7 — ACCEPTED

Current reviewed head:

ab73e2c5a96a9b661b418ebf59d1169c23afad11

Remote evidence:

• GitHub CI #310: SUCCESS
• PR feat: add P1 result-aware experiment iteration #9 current head matches the reviewed head
• production image/build-context hygiene passed as part of CI

Tranche acceptance proves:

Experiment E
→ R1
→ V1
→ B1
→ exact S1
→ Follow-up AgentRun A2
→ R2 in same E
→ explicit deterministic verification
→ explicit human approval
→ V2 in same E
→ no automatic Batch for V2

Evidence boundary also proves:

• selected valid metric evidence and validation report identity reach ResultContext
• excluded Trial/reason and Trial conditions reach ResultContext
• raw file sentinel does not reach provider-visible context
• post-summary raw MetricObservation mutation does not override or leak past the exact pinned summary

Stop-path acceptance proves no R2/V2/Batch writes when the Agent returns goal_satisfied.

Next authorized step

Task 8 only: focused architecture/data-boundary/regression gate.

No production implementation changes are expected in Task 8. Any failure must be classified before code is changed.

Task 9 full repository gate / closeout remains not started.

Task 8 — ACCEPTED

Reviewed head:

e8caaf668fed2b5e5878393fa0c47c001ba218e9

Remote evidence:

• GitHub CI #312: SUCCESS
• production image/build-context hygiene: SUCCESS
• PR head matches reviewed head

Task 8 focused regression gates all passed:

• architecture
• migration
• Agent unit/integration
• same-Experiment Proposal service
• initial Slice 1 API/proposal loop
• Result-Aware API
• Result Plane
• P1 tranche acceptance
• OpenCode constructor compatibility
• Ruff / mypy

The only Task 8 code change was test-only concurrency hardening:

• replaced timer-based lock ordering with an explicit test barrier after real BatchRow FOR UPDATE acquisition
• focused race repeated 5x successfully
• no production code changed

Verified invariants remain:

• agent_runtime -> results_validation dependency unchanged
• Agent does not auto-verify
• Agent does not auto-approve
• Agent does not auto-create/start Batch
• initial Proposal path unchanged
• Follow-up Experiment authority comes from exact pinned ResultContext

Next authorized step

Task 9 only: final repository-equivalent gate, closeout/status documentation, final CI/hygiene verification, then PR merge-readiness review.

Do not reopen global architecture.