Preserve approved plan across PR handoff and review

[wwind123]

Summary

• Carry the canonical approved plan, identity, scope, and deferred work through implementation, full/compact PR review, follow-ups, and resumes.
• Recover plans from validated PR-bound handoffs and staged parent/child topology, with explicit diagnostics for mismatches and legacy unavailable plans.
• Replace positional signed-requirement labels with stable content-derived IDs and keep planning records out of the PR unresolved-item ledger.
• Document the plan channel, labeled staged contexts, precedence rules, and ordinary unplanned-PR compatibility.

Verification

• python3 -m pytest tests/test_issue_756_plan_context.py tests/test_human_requirement_dispositions.py tests/test_round_transport.py tests/test_issue_pr_handoff.py tests/test_decomposition.py tests/test_split_materialization.py -q

Fixes #756

[wwind123]

Expected closing issues for PR #757: #756.

Origin flow: approved-plan-implementation
Contract hash: 5959c139c6c1fd6ba377c431ecd428d9f1b4f390d2029d012cc399ffcc91486b

-- coding-review-agent-loop

[wwind123]

Issue implementation

Implemented the approved plan, committed it with the required provenance trailer, pushed branch codex/issue-756-approved-plan, and opened PR #757 against main. Focused issue-756, handoff, transport, decomposition, split-materialization, and human-requirement suites pass (171 tests); the broader legacy suite still contains positional-label fixtures and host-specific containment assertions outside this migration.

Result

Pull request reported: #757.

Tests run

• timeout 120s python3 -m pytest tests/test_issue_756_plan_context.py tests/test_human_requirement_dispositions.py tests/test_round_transport.py tests/test_issue_pr_handoff.py tests/test_decomposition.py tests/test_split_materialization.py -q

-- OpenAI Codex: gpt-5.6-luna (xhigh)

[wwind123]

Review verdict: Blocking

The approved-plan channel itself (ApprovedPlanContext, handoff-hash-bound recovery, full/compact prompt propagation, docs) is well built, but the PR also migrates signed human requirement labels from positional Requirement N to content-digest Requirement hr-<digest> without updating the existing suite, so tests that assert positional labels will fail; that migration also breaks in-flight resume compatibility. Two orchestrator behaviors (hard abort when a linked issue's latest handoff names a different PR; abort on new signed requirements after plan approval) are new failure modes with no test coverage, and required orchestrator-level tests for stale snapshots and issue/PR resume are missing. GitHub checks are still pending (test in_progress), so no global pass can be claimed; these findings come from reading the diff.

Blocking issues

• format_human_requirements() now emits Requirement {requirement.requirement_id} (src/coding_review_agent_loop/prompts.py:586) and render_coder_human_requirements_prompt_context() only recognizes ^Requirement (hr-[0-9a-f]{64}):$ (prompts.py:684), but no existing test was updated. Tests that assert the old positional rendering will fail, e.g. tests/test_prompts.py:1328 ("Middle requirement.\n\nRequirement 3:" in text) and tests/test_prompts.py:1364 (surfaced_requirement_ids == ("Requirement 2", "Requirement 3")), plus the many prompt/orchestrator/repair tests that feed - Requirement 1: ... acknowledgements against now-digest surfaced IDs. Update the suite (or keep the rendering backward compatible) so the change is actually verified.
• The positional-to-digest requirement ID migration is not needed by issue Preserve the approved plan across PR handoff, reviews, coder follow-ups, and resumes #756 and is a breaking change to a durable protocol: a run that already posted an acknowledgement using Requirement 1 will, on resume, hit the new legacy_unknown path in validate_structured_human_requirements_acknowledgement() (src/coding_review_agent_loop/protocol.py:898) and raise instead of resuming. Either split this migration into its own PR with an explicit compatibility path for already-posted acknowledgements, or document and test the resume behavior for runs started before this change.
• In run_pr_loop(), any issue-side handoff whose pr_number differs from the PR under review now raises AgentLoopError unconditionally (src/coding_review_agent_loop/orchestrator.py, Issue #... handoff selects PR #...), regardless of flow. A direct agent-loop pr <N> on an ordinary PR that merely says Fixes #123, where Add structured plan review protocol #123 previously produced a different agent-loop PR, will abort before any review runs. Scope this check to approved-plan provenance (or to the case where a plan context is actually required) and add a test for the unrelated-linked-PR case.
• The new approval-boundary refresh in _run_plan_first_loop() aborts the whole run with AgentLoopError when the issue (or authoritative parent) gained any signed human requirement after plan approval. That is a new hard failure on a normal human interaction and has no test; add coverage and confirm this fail-closed behavior is intended rather than a re-plan/re-acknowledge path.
• Acceptance criteria ask for focused tests on stale initial snapshots and issue/PR resume, but tests/test_issue_756_plan_context.py only covers pure helpers and prompt rendering. None of the new orchestrator logic is exercised: the handoff-bound plan recovery in run_pr_loop(), the contract-primary-issue selection, the split/phase identity validation, or the pre-implementation issue-context refresh. Add orchestrator-level tests for at least stale-snapshot refresh and both resume paths.

Same-PR follow-ups

• _implement_approved_issue() builds implementation_requirements (parent + child, deduplicated) and uses it for the implementation prompt and validation, but its CompletionRecoveryPolicy(...) at src/coding_review_agent_loop/orchestrator.py:4741 omits human_requirements=implementation_requirements, so the recovery prompt silently falls back to child-only requirements. run_issue_loop() passes them; make the one-shot path consistent.

-- Anthropic Claude: claude-opus-5 (medium)

[wwind123]

Review verdict: Approved

Issue-mode canonical-PR resume does not yet recover the plan selected by the handoff when a later planning round exists. GitHub CI is still pending (test in progress).

-- OpenAI Codex: gpt-5.6-sol (medium)

[wwind123]

PR review round 1 reconciliation: settled reviewers: Claude, Codex. Finalization continues after reconciliation.

[wwind123]

Coder follow-up

Implemented and pushed the requested compatibility, plan-propagation, resume, and recovery fixes to PR #757.

Addressed items

• item-1: Blocking issue from Anthropic Claude: claude-opus-5 (medium), round 1: format_human_requirements() now emits Requirement {requirement.requirement_id} (src/coding_re...
  • Resolution: Restored the established Requirement N rendering/parsing protocol and aligned repair guidance and regression coverage.
• item-2: Blocking issue from Anthropic Claude: claude-opus-5 (medium), round 1: The positional-to-digest requirement ID migration is not needed by issue Preserve the approved plan across PR handoff, reviews, coder follow-ups, and resumes #756 and is a breaking c...
  • Resolution: Kept digest-backed identity internal for content-change detection while preserving legacy positional acknowledgements on resume.
• item-3: Blocking issue from Anthropic Claude: claude-opus-5 (medium), round 1: In run_pr_loop(), any issue-side handoff whose pr_number differs from the PR under review now...
  • Resolution: Older direct issue handoffs no longer block unrelated ordinary direct PR reviews; plan-bound provenance remains fail-closed.
• item-4: Blocking issue from Anthropic Claude: claude-opus-5 (medium), round 1: The new approval-boundary refresh in _run_plan_first_loop() aborts the whole run with `AgentLoo...
  • Resolution: Preserved fail-closed approval-boundary handling for newly signed requirements and added orchestration coverage.
• item-5: Blocking issue from Anthropic Claude: claude-opus-5 (medium), round 1: Acceptance criteria ask for focused tests on stale initial snapshots and issue/PR resume, but tes...
  • Resolution: Added orchestration tests for stale snapshot refresh, issue resume, direct PR resume, and selection of the handoff-associated plan over a later unrelated plan.
• item-6: Blocking issue from Anthropic Claude: claude-opus-5 (medium), round 1: Blocking review included Same-PR follow-ups, but --approved-followups=ignore does not enable a sa...
  • Resolution: Passed merged parent/child implementation requirements into one-shot completion recovery and covered the resulting recovery context.

Remaining items

• None.

Tests run

• python3 -m pytest tests/test_issue_756_plan_context.py tests/test_prompts.py tests/test_protocol.py tests/test_response_validation.py tests/test_human_requirement_dispositions.py tests/test_repair.py tests/test_repair_preservation.py tests/test_orchestrator_issue.py tests/test_orchestrator_pr.py -q (agent-loop instrumented; whole-command timeout 120s)

-- OpenAI Codex: gpt-5.6-luna (xhigh)

[wwind123]

Review verdict: Approved

Round 2 resolves all six carried items. The digest requirement-ID migration is reverted to positional Requirement N labels (digest identity is now internal only, used for deduplication and the post-approval change check), so no durable acknowledgement protocol break remains. The unrelated-handoff abort in run_pr_loop is scoped to plan-bound provenance and logs-and-continues otherwise; issue-mode resume now selects the plan by handoff hash instead of the newest planning round; the completion-recovery policy carries merged parent+child requirements, and the stray unsupported kwarg on the run_pr_loop resume call is gone. New orchestrator-level tests cover stale caller snapshots, handoff-bound plan recovery on both issue and direct-PR resume, the unrelated older handoff, fail-closed post-approval requirement changes, and recovery-prompt requirements; prompt-level tests cover full/compact plan rendering, oversized-plan omission, identity mismatch, and unplanned PRs. README, docs/local_agent_loop.md, and docs/skill_mode.md are updated consistently. GitHub checks are still pending (test in_progress), so this approval is based on reading the diff, not on a green suite.

Prior unresolved item dispositions

• [item-1] Blocking issue from Anthropic Claude: claude-opus-5 (medium), round 1: format_human_requirements() now emits Requirement {requirement.requirement_id} (src/coding_re... -> resolved
• [item-2] Blocking issue from Anthropic Claude: claude-opus-5 (medium), round 1: The positional-to-digest requirement ID migration is not needed by issue Preserve the approved plan across PR handoff, reviews, coder follow-ups, and resumes #756 and is a breaking c... -> resolved
• [item-3] Blocking issue from Anthropic Claude: claude-opus-5 (medium), round 1: In run_pr_loop(), any issue-side handoff whose pr_number differs from the PR under review now... -> resolved
• [item-4] Blocking issue from Anthropic Claude: claude-opus-5 (medium), round 1: The new approval-boundary refresh in _run_plan_first_loop() aborts the whole run with `AgentLoo... -> resolved
• [item-5] Blocking issue from Anthropic Claude: claude-opus-5 (medium), round 1: Acceptance criteria ask for focused tests on stale initial snapshots and issue/PR resume, but tes... -> resolved
• [item-6] Blocking issue from Anthropic Claude: claude-opus-5 (medium), round 1: Blocking review included Same-PR follow-ups, but --approved-followups=ignore does not enable a sa... -> resolved

-- Anthropic Claude: claude-opus-5 (medium)

[wwind123]

Review verdict: Blocking

Most round-1 concerns are resolved, but two resume/provenance paths still block valid reviews. GitHub's test check is still in progress.

Blocking issues

• Issue-mode plan-first resume of a staged child cannot recover its parent-owned approved plan. In run_issue_loop() (src/coding_review_agent_loop/orchestrator.py:6443), handoff-bound recovery searches only issue_context.comments; when the canonical plan is on the staged parent this produces a non-null unavailable context. run_pr_loop() only attempts its parent fallback when approved_plan_context is None (orchestrator.py:7696), so it skips the available parent record and raises at line 7721. Recover from the parent in issue mode, or allow the PR loop to replace an unavailable child candidate with the validated parent candidate, and add a staged child issue-resume test.

Prior unresolved item dispositions

• [item-1] Blocking issue from Anthropic Claude: claude-opus-5 (medium), round 1: format_human_requirements() now emits Requirement {requirement.requirement_id} (src/coding_re... -> resolved
• [item-2] Blocking issue from Anthropic Claude: claude-opus-5 (medium), round 1: The positional-to-digest requirement ID migration is not needed by issue Preserve the approved plan across PR handoff, reviews, coder follow-ups, and resumes #756 and is a breaking c... -> resolved
• [item-3] Blocking issue from Anthropic Claude: claude-opus-5 (medium), round 1: In run_pr_loop(), any issue-side handoff whose pr_number differs from the PR under review now... -> still blocking
• [item-4] Blocking issue from Anthropic Claude: claude-opus-5 (medium), round 1: The new approval-boundary refresh in _run_plan_first_loop() aborts the whole run with `AgentLoo... -> resolved
• [item-5] Blocking issue from Anthropic Claude: claude-opus-5 (medium), round 1: Acceptance criteria ask for focused tests on stale initial snapshots and issue/PR resume, but tes... -> resolved
• [item-6] Blocking issue from Anthropic Claude: claude-opus-5 (medium), round 1: Blocking review included Same-PR follow-ups, but --approved-followups=ignore does not enable a sa... -> resolved

-- OpenAI Codex: gpt-5.6-sol (medium)

[wwind123]

PR review round 2 reconciliation: settled reviewers: Claude, Codex. Finalization continues after reconciliation.

[wwind123]

Coder follow-up

Implemented both blocking fixes, added staged-child regression coverage, and pushed commit ce31e44 to PR #757 for follow-up review.

Addressed items

• item-3: Blocking issue from Anthropic Claude: claude-opus-5 (medium), round 1: In run_pr_loop(), any issue-side handoff whose pr_number differs from the PR under review now...
  • Resolution: Scoped mismatched issue-handoff failures to approved-plan provenance or an explicitly supplied plan context; unrelated direct-implementation handoffs are logged and ignored, with regression coverage for an ordinary linked PR.
• item-7: Blocking issue from OpenAI Codex: gpt-5.6-sol (medium), round 2: Issue-mode plan-first resume of a staged child cannot recover its parent-owned approved plan. In...
  • Resolution: When issue-mode resume supplies an unavailable child plan candidate, PR review now recovers and validates the matching parent-owned plan before constructing the review prompt; added a staged-child issue-resume regression test.

Remaining items

• None.

Tests run

• timeout 180s python3 -m pytest tests/test_orchestrator_issue.py::test_issue_loop_plan_first_staged_child_recovers_parent_owned_plan tests/test_orchestrator_issue.py::test_issue_loop_plan_first_resume_uses_handoff_bound_plan_when_later_plan_exists tests/test_orchestrator_pr.py::test_direct_pr_ignores_unrelated_older_issue_handoff tests/test_orchestrator_pr.py::test_direct_pr_resume_uses_handoff_bound_plan_when_later_plan_exists -q
• timeout 300s python3 -m pytest tests/test_orchestrator_issue.py tests/test_orchestrator_pr.py tests/test_prompts.py -q
• timeout 180s python3 -m pytest tests/test_issue_756_plan_context.py tests/test_issue_pr_handoff.py tests/test_split_orchestration.py tests/test_split_materialization.py -q

-- OpenAI Codex: gpt-5.6-luna (xhigh)

[wwind123]

Review verdict: Approved

Round 3 (ce31e44) closes both carried items. The staged-child gap is fixed: run_pr_loop now re-attempts hash-validated parent recovery when the child-side candidate is unavailable (not only when no context was supplied), and a new end-to-end test drives an issue-mode plan-first resume of a staged child whose canonical plan lives on the parent, asserting the parent plan text reaches the reviewer prompt. The mismatched-PR handoff check remains scoped to plan-bound provenance in the outer guard, and the now-unreachable duplicate raise inside the handoff branch was removed, leaving the ordinary direct-PR path (unrelated older handoff) logging and continuing, covered by test_direct_pr_ignores_unrelated_older_issue_handoff. Hash equality against the issue-side handoff is still enforced before and after recovery, so the parent fallback cannot substitute a different plan. GitHub checks are still pending (test in_progress), so this approval reflects diff review, not a green suite.

Prior unresolved item dispositions

• [item-3] Blocking issue from Anthropic Claude: claude-opus-5 (medium), round 1: In run_pr_loop(), any issue-side handoff whose pr_number differs from the PR under review now... -> resolved
• [item-7] Blocking issue from OpenAI Codex: gpt-5.6-sol (medium), round 2: Issue-mode plan-first resume of a staged child cannot recover its parent-owned approved plan. In... -> resolved

-- Anthropic Claude: claude-opus-5 (medium)

[wwind123]

Review verdict: Approved

The two carried blockers are resolved, but approved-plan recovery still drops legacy freeform plans and unchanged-head approval reuse does not use the new stable signed-requirement identities. GitHub CI is still pending (test in progress).

Prior unresolved item dispositions

• [item-3] Blocking issue from Anthropic Claude: claude-opus-5 (medium), round 1: In run_pr_loop(), any issue-side handoff whose pr_number differs from the PR under review now... -> resolved
• [item-7] Blocking issue from OpenAI Codex: gpt-5.6-sol (medium), round 2: Issue-mode plan-first resume of a staged child cannot recover its parent-owned approved plan. In... -> resolved

-- OpenAI Codex: gpt-5.6-sol (medium)

[wwind123]

PR review round 3 reconciliation: settled reviewers: Claude, Codex. Finalization continues after reconciliation.

[wwind123]

GitHub PR checks are failing for PR #757.

Reviewer approvals do not make this PR merge-ready until GitHub PR checks are green, or the PR explicitly states that only a local subset passed.

• Failing checks: test (failure) — https://github.com/wwind123/coding-review-agent-loop/actions/runs/34440390604/job/102754020927
• Required status checks are not configured on the PR base branch.

-- coding-review-agent-loop

[wwind123]

Coder follow-up

Corrected the completion-recovery regression fixture so the fake PR metadata and handoff URL match the structured PR #88 result. Pushed commit aec146d to PR #757; the fresh GitHub Actions test check is green.

Addressed items

• item-8: Blocking issue from GitHub PR checks, round 3: GitHub PR checks are failing and must be resolved before approval.
  • Resolution: Updated tests/test_completion_recovery.py so the waiting-wording recovery case identifies PR Refresh reviewer checkouts before PR reviews #88 consistently; the full completion-recovery module passes locally and GitHub CI test passed on commit aec146d.

Remaining items

• None.

Tests run

• timeout 120s python3 -m pytest tests/test_completion_recovery.py::test_issue_loop_recovers_from_claude_waiting_on_background_wording -q
• timeout 120s python3 -m pytest tests/test_completion_recovery.py::test_issue_loop_recovers_from_claude_waiting_on_background_wording tests/test_completion_recovery.py::test_issue_loop_recovers_after_one_bounded_resume_then_succeeds -q
• timeout 120s python3 -m pytest tests/test_completion_recovery.py -q

-- OpenAI Codex: gpt-5.6-luna (xhigh)

[wwind123]

Review verdict: Approved

Round 4 adds only aec146d, which completes the PR payload fixture (number/url alongside body) in tests/test_completion_recovery.py so the fake PR record matches what the refreshed issue/handoff resolution path reads; no production code changed since round 3. GitHub PR checks are now green (overall state: passing), which clears the carried CI failure item. All earlier findings from rounds 1-3 remain addressed: positional requirement labels restored, plan-bound handoff recovery with hash validation and parent fallback for staged children, per-round issue-context refresh, and docs updated in README.md, docs/local_agent_loop.md, and docs/skill_mode.md.

Prior unresolved item dispositions

• [item-8] Blocking issue from GitHub PR checks, round 3: GitHub PR checks are failing and must be resolved before approval. -> resolved

-- Anthropic Claude: claude-opus-5 (medium)

[wwind123]

Review verdict: Blocking

The authoritative GitHub checks are passing, but unchanged-head approval reuse can still overlook an edited signed human instruction.

Blocking issues

• src/coding_review_agent_loop/orchestrator.py:8056: Persisted reviewer coverage still consists of positional Requirement N labels, and the reuse checks at lines 8148 and 8353 compare only those labels. Although HumanReviewRequirement.requirement_id now changes when a signed comment is edited, that digest identity never reaches surfaced_reviewer_requirement_ids. If a signed comment is edited without changing the number of requirements, the current and prior label sets remain identical, so a prior approval for the old text is carried forward without invoking the reviewer for the new instruction. Persist and compare the stable identities (with an explicit legacy-metadata policy) while retaining positional labels only for display/compatibility, and add an unchanged-head edited-comment regression test.

Prior unresolved item dispositions

• [item-8] Blocking issue from GitHub PR checks, round 3: GitHub PR checks are failing and must be resolved before approval. -> resolved

-- OpenAI Codex: gpt-5.6-sol (medium)

[wwind123]

PR review round 4 reconciliation: settled reviewers: Claude, Codex. Finalization continues after reconciliation.

[wwind123]

Coder follow-up

Implemented the requested fix and left one reviewer item for follow-up.

Addressed items

• item-9: Blocking issue from OpenAI Codex: gpt-5.6-sol (medium), round 4: src/coding_review_agent_loop/orchestrator.py:8056: Persisted reviewer coverage still consists o...
  • Resolution: PR reviewer coverage metadata now persists digest-backed signed-requirement identities, compares those identities at both approval carry-forward gates, and explicitly rejects legacy positional metadata. Added direct policy coverage and an unchanged-head edited-comment regression test.

Remaining items

• None.

Tests run

• python3 -m pytest tests/test_issue_756_plan_context.py tests/test_orchestrator_pr.py -k 'reviewer_coverage or pr_loop_rereviews_unchanged_head_when_human_requirement_changes or pr_loop_skips_prior_approval_when_pr_head_is_unchanged' -q (agent-loop instrumented; whole-command timeout 120s)
• python3 -m pytest tests/test_repair.py -k 'surfaced_reviewer_requirement_ids or attempt_repair_reviewer_requirement_ids' -q (agent-loop instrumented; whole-command timeout 120s)
• python3 -m pytest tests/test_orchestrator_pr.py -q (agent-loop instrumented; whole-command timeout 120s)
• python3 -m pytest tests/test_orchestrator_pr.py -k pr_loop_rereviews_unchanged_head_when_human_requirement_changes -q (agent-loop instrumented; whole-command timeout 120s)

-- OpenAI Codex: gpt-5.6-luna (xhigh)

[wwind123]

Review verdict: Approved

Round 5 (a3fa69a) closes the reviewer-coverage gap. PR round metadata now persists digest identities via _reviewer_requirement_identity_ids(), and both the prelaunch classification and the carried-approval branch compare through _reviewer_requirement_coverage_matches(), which treats any non-digest (legacy positional) persisted entry as non-covering whenever signed requirements exist, so an edited signed comment with an unchanged requirement count can no longer inherit a prior approval. The positional labels still used in reviewer prompts, repair contexts, and response validation continue to come from _surfaced_reviewer_requirement_ids(), so response compatibility is unaffected. Coverage is adequate: a unit test asserts digest identity, edited-body mismatch, and legacy-label rejection, and the parametrized unchanged-head PR test now includes an edited-comment case with legacy persisted metadata and asserts the reviewer runs again. Documentation already describes the internal digest identity and edit detection. GitHub checks are pending again for this new head (test in_progress); the previous head was green, but I am not claiming a green suite for a3fa69a.

Prior unresolved item dispositions

• [item-9] Blocking issue from OpenAI Codex: gpt-5.6-sol (medium), round 4: src/coding_review_agent_loop/orchestrator.py:8056: Persisted reviewer coverage still consists o... -> resolved

-- Anthropic Claude: claude-opus-5 (medium)

[wwind123]

Review verdict: Blocking

The digest-backed coverage change resolves item-9, but resume coverage, staged-parent recovery, and skill-mode propagation still have merge-blocking gaps. GitHub's test check is still in progress.

Blocking issues

• src/coding_review_agent_loop/orchestrator.py:8172 and src/coding_review_agent_loop/orchestrator.py:8365: current-round resumed reviewer records bypass _reviewer_requirement_coverage_matches entirely. If the process stops after one reviewer posts and a signed requirement is then added or edited before resume, that completed reviewer is classified as resumed and its stale approval is accepted on the unchanged head without re-review. Filter resumed records by the current digest identities (with the same explicit legacy policy as carried approvals) and add an interrupted-round resume regression test.
• src/coding_review_agent_loop/orchestrator.py:7741: staged-child recovery consults the authoritative parent only when child recovery is unavailable, not when it is mismatched. A child with any unrelated plan round therefore prevents recovery of the exact handoff-bound plan that is present on its validated parent, so both issue and direct-PR resume fail despite valid provenance. Check the parent whenever the child candidate is not available, while still rejecting a genuinely conflicting matching candidate, and cover a staged child with unrelated local plan history.
• helpers/prompt_builders.py:232, helpers/prompt_builders.py:273, and helpers/skill_runner.py:2566: skill-mode PR reviewers and coder fixes still call the shared prompt builders without an approved-plan context, and the host-review handoff carries only the diff. Thus plan-first PRs resumed through run-pr-round/run-pr-fix tell agents that no plan is bound and can lose scope or deferred work, contrary to the new parity claim in docs/skill_mode.md:85. Recover and validate the handoff-bound plan in the skill workflow and pass it to every reviewer and fix prompt, or remove the unsupported parity claim and explicitly scope the feature away from skill mode.

Prior unresolved item dispositions

• [item-9] Blocking issue from OpenAI Codex: gpt-5.6-sol (medium), round 4: src/coding_review_agent_loop/orchestrator.py:8056: Persisted reviewer coverage still consists o... -> resolved

-- OpenAI Codex: gpt-5.6-sol (medium)

[wwind123]

PR review round 5 reconciliation: settled reviewers: Claude, Codex. Finalization continues after reconciliation.

[wwind123]

Coder follow-up

Implemented and pushed the approved-plan recovery and reviewer-resume fixes to PR #757 commit 3348bfc.

Addressed items

• item-13: Blocking issue from OpenAI Codex: gpt-5.6-sol (medium), round 8: src/coding_review_agent_loop/orchestrator.py:5220-5253 (and the free-form revision path near `:...
  • Resolution: Persisted raw free-form plans for new and revised rounds, added hash/subject-validated migration recovery that handles normalized signatures, and made unavailable plan-bound context fail closed. Regression coverage is in tests/test_issue_756_plan_context.py and tests/test_agent_loop.py.
• item-14: Blocking issue from OpenAI Codex: gpt-5.6-sol (medium), round 8: helpers/skill_runner.py:2891-2935 reconstructs completed same-head reviewer approvals before re...
  • Resolution: Recovered the PR-bound plan before same-head reviewer reuse, persisted its hash and subject through reviewer metadata, repair manifests, retries, and host handoffs, and reran reviewers with missing or mismatched identity. Coverage includes partial resume and metadata round trips in tests/test_skill_helpers.py and tests/test_issue_756_plan_context.py.
• item-15: Blocking issue from OpenAI Codex: gpt-5.6-sol (medium), round 8: helpers/skill_runner.py:1122-1135 suppresses every issue-comment retrieval/parsing failure when...
  • Resolution: Made unreadable uniquely linked issue history an actionable provenance error while retaining ordinary direct PRs without linked planning history. Coverage is in tests/test_skill_helpers.py.

Remaining items

• None.

Tests run

• python3 -m pytest tests/test_issue_756_plan_context.py tests/test_agent_loop.py tests/test_skill_helpers.py -q (agent-loop instrumented; whole-command timeout 300s)
• python3 -m pytest tests/test_orchestrator_pr.py -k 'direct_pr_resume_uses_handoff_bound_plan_when_later_plan_exists or pr_loop_resumes_with_only_missing_reviewer_for_current_head or resume_pr_round_preserves_stored_model_used or pr_loop_rereviews_unchanged_head_when_human_requirement_changes' -q (agent-loop instrumented; whole-command timeout 180s)
• python3 -m pytest tests/test_issue_756_plan_context.py::test_plan_recovery_selects_expected_hash_not_newest_plan tests/test_issue_756_plan_context.py::test_legacy_freeform_plan_recovery_reverses_signature_normalization tests/test_skill_helpers.py::TestRunPrFix::test_success_posts_metadata_after_head_and_assigned_head_advance -q

-- OpenAI Codex: gpt-5.6-luna (xhigh)

[wwind123]

Review verdict: Approved

Round 9 (6759445 plus 3348bfc) closes all three carried items. Free-form accepted plans and free-form revisions now persist canonical_plan alongside the normalized public comment, recover_approved_plan_context() gained a hash-and-subject-validated migration fallback that reverses signature normalization for pre-canonical records with an explicit diagnostic when only the subject fails, and the legacy_contract escape is gone so a plan-bound PR can no longer be reviewed without an available matching plan; the follow-up commit extends the same fail-closed rule to a PR contract declaring approved-plan provenance with an unavailable context. PostedRoundMetadata now carries approved_plan_hash/subject, plumbed through state_manager attach-metadata and build-resume, checked in _latest_pr_approved_reviews_for_head and _resumed_pr_reviewer_matches_requirements, and mirrored in skill mode by _filter_resume_for_approved_plan (guarded on the PR-head resume subject, which build-resume does set) plus repair and host-review manifests. Skill-mode legacy provenance no longer downgrades an unreadable linked-issue history to a diff-only review. Tests cover metadata round-trip, both free-form canonical paths, same-head plan mismatch, legacy signature-normalized recovery, partial same-head resume filtering, and the legacy read failure. GitHub checks are pending for this head (test in_progress); the prior head was green, so I am not claiming a green suite for 3348bfc.

Prior unresolved item dispositions

• [item-13] Blocking issue from OpenAI Codex: gpt-5.6-sol (medium), round 8: src/coding_review_agent_loop/orchestrator.py:5220-5253 (and the free-form revision path near `:... -> resolved
• [item-14] Blocking issue from OpenAI Codex: gpt-5.6-sol (medium), round 8: helpers/skill_runner.py:2891-2935 reconstructs completed same-head reviewer approvals before re... -> resolved
• [item-15] Blocking issue from OpenAI Codex: gpt-5.6-sol (medium), round 8: helpers/skill_runner.py:1122-1135 suppresses every issue-comment retrieval/parsing failure when... -> resolved

-- Anthropic Claude: claude-opus-5 (medium)

[wwind123]

Review verdict: Blocking

The round-8 blockers are resolved, but skill-mode resume still cannot recover the parent plan for a staged child PR. GitHub's test check is still in progress.

Blocking issues

• helpers/skill_runner.py:1146-1173,1298-1308 treats an approved-plan handoff recorded on a staged child as a signal to recover the canonical plan from that child, and only enters the parent/topology recovery branch when no handoff hash was found. Headless staged implementations record the approved-plan PR handoff on the child (orchestrator.py:4940-4951), while the canonical plan remains on the parent. Consequently, mixed-mode run-pr-round/run-pr-fix resume of either a materialized split child or a decomposition child fails with a missing canonical plan instead of validating the child-to-parent topology and recovering the bound parent plan. Resolve staged-parent provenance even when the child already has its canonical PR handoff, including the split-child path, and add regressions for both staged child forms.

Prior unresolved item dispositions

• [item-13] Blocking issue from OpenAI Codex: gpt-5.6-sol (medium), round 8: src/coding_review_agent_loop/orchestrator.py:5220-5253 (and the free-form revision path near `:... -> resolved
• [item-14] Blocking issue from OpenAI Codex: gpt-5.6-sol (medium), round 8: helpers/skill_runner.py:2891-2935 reconstructs completed same-head reviewer approvals before re... -> resolved
• [item-15] Blocking issue from OpenAI Codex: gpt-5.6-sol (medium), round 8: helpers/skill_runner.py:1122-1135 suppresses every issue-comment retrieval/parsing failure when... -> resolved

-- OpenAI Codex: gpt-5.6-sol (medium)

[wwind123]

PR review round 9 reconciliation: settled reviewers: Claude, Codex. Finalization continues after reconciliation.

[wwind123]

Coder follow-up

Recovered the interrupted coder implementation and pushed commit 6b709d9 to PR #757. Validation used the isolated checkout /home/wwind123/openai-codex/agent-loop-pr757-salvage, leaving the default agent worktree untouched. The original eight focused cases passed; after adding seven negative provenance cases, all 15 focused recovery cases passed. The three-module suite passed 293 tests before those seven extra cases; the final full suite passed 2,714 tests in 147.96 seconds. The earlier 120-second timeout attempts are not counted as passes. Item-16 is implemented but remains in the review ledger for independent re-verification. No current-head reviewer approval or GitHub CI success is claimed.

Addressed items

• item-16: Blocking issue from OpenAI Codex, round 9: helpers/skill_runner.py:1146-1173,1298-1308 treats an approved-plan handoff recorded on a stage...
  • Resolution: Recovery now validates child-to-parent topology even when the child already has an approved-plan PR handoff, then recovers the canonical plan from the parent for both decomposition and materialized split children. Focused approved-plan recovery tests passed. Added rejection coverage for missing parent handoffs/plans/topology and mismatched stage identities/keys.

Remaining items

• None.

Tests run

• python3 -m pytest tests/test_skill_helpers.py -k TestSkillApprovedPlanRecovery -q --tb=short
• python3 -m pytest tests/test_skill_helpers.py tests/test_decomposition.py tests/test_split_materialization.py -q --tb=short --durations=10
• python3 -m pytest tests/ -q --tb=short
• git diff --check

-- OpenAI Codex

[wwind123]

Review verdict: Approved

Round 10 resolves the carried staged-child plan-recovery gap. helpers/skill_runner.py:1177-1414 now resolves the child's staged marker (decomposition phase identity or split-child key) before falling back to child-side plan recovery, so a child that already carries its own canonical PR handoff still validates parent topology and recovers the bound parent plan; both the decomposition-child and materialized-split-child paths are covered by parameterized regressions in tests/test_skill_helpers.py, including missing parent handoff, wrong phase identity, wrong split key, missing topology, and missing plan faults. Split recovery is consistent with the orchestrator's own per-plan-hash stage-handoff semantics (one child per approved plan hash), and the ambiguous/missing-hash cases fail closed with actionable diagnostics. The rest of the propagation remains coherent: the plan channel is selected by handoff hash rather than newest comment, hash/subject are validated against raw text before sanitizing, full and compact reviewer prompts plus follow-up, same-PR and merge-conflict prompts render the dedicated plan block outside bounded issue history, plan identity is persisted in round metadata and gates same-head approval reuse and resume filtering, and ordinary PRs with no planning provenance still take the explicit no-plan path. Docs in README.md, docs/local_agent_loop.md and docs/skill_mode.md describe the propagation and the limits of plan authority. GitHub PR checks are reported as passing; I ran no tests myself in this read-only review.

Prior unresolved item dispositions

• [item-16] Blocking issue from OpenAI Codex: gpt-5.6-sol (medium), round 9: helpers/skill_runner.py:1146-1173,1298-1308 treats an approved-plan handoff recorded on a stage... -> resolved

-- Anthropic Claude: claude-opus-5 (medium)

[wwind123]

Review verdict: Blocking

The staged-child parent-plan recovery from item-16 is now implemented for both decomposition and split children, and GitHub checks are passing. Two approved-contract gaps remain in signed-requirement identity and oversized-plan handling.

Blocking issues

• src/coding_review_agent_loop/prompts.py:582-595,668-695 and src/coding_review_agent_loop/protocol.py:778-815,1219-1269 still render, parse, and validate signed instructions as positional Requirement N labels. Inserting or chronologically reordering a requirement therefore reassigns the same acknowledgement label to different content, and legacy output is accepted against the current tuple (tests/test_issue_756_plan_context.py:153-166) instead of requiring a fresh acknowledgement. The approved plan explicitly requires digest-backed IDs to be surfaced end-to-end and legacy positional replies never to be translated onto a changed requirement set; use each requirement's stable ID in prompts/responses/repair and add the specified migration regression.
• src/coding_review_agent_loop/prompts.py:1110-1155 applies a fixed 24,000-character cutoff rather than an actual final provider limit. When the identity-plus-declarations prefix itself exceeds that cutoff, line 1155 drops every scope/deferred declaration and lets review continue with only an omission notice. This violates the required minimum context and can hide exactly the scope contract this PR is meant to preserve. Preserve the complete plan whenever the provider limit permits, and fail with actionable remediation when identity plus complete scope/deferred declarations cannot fit instead of returning a degraded prompt.

Prior unresolved item dispositions

• [item-16] Blocking issue from OpenAI Codex: gpt-5.6-sol (medium), round 9: helpers/skill_runner.py:1146-1173,1298-1308 treats an approved-plan handoff recorded on a stage... -> resolved

-- OpenAI Codex: gpt-5.6-sol (medium)

[wwind123]

PR review round 10 reconciliation: settled reviewers: Claude, Codex. Finalization continues after reconciliation.

[wwind123]

Coder follow-up

Implemented and pushed commit e2ebf50 to PR #757. Focused prompt, protocol, orchestration, repair, and response-validation tests pass. The full suite reports 2712 passed; two unrelated containment assertions fail because the mandated test wrapper exposes the inherited active agent cgroup rather than a newly created sibling scope.

Addressed items

• item-17: Blocking issue from OpenAI Codex: gpt-5.6-sol (medium), round 10: src/coding_review_agent_loop/prompts.py:582-595,668-695 and `src/coding_review_agent_loop/proto...
  • Resolution: Signed instructions now render and validate with content-derived hr-... IDs across prompts, structured and Markdown acknowledgements, repair, skill-mode repair context, round metadata, and resume. Legacy positional output is never mapped to the current tuple and produces an actionable fresh-acknowledgement blocker. Focused prompt, protocol, orchestration, and repair tests cover stable insertion, reordering, and migration.
• item-18: Blocking issue from OpenAI Codex: gpt-5.6-sol (medium), round 10: src/coding_review_agent_loop/prompts.py:1110-1155 applies a fixed 24,000-character cutoff rathe...
  • Resolution: Removed the fixed 24,000-character plan cutoff. Full canonical text is retained by default; an explicit final provider cap may omit only canonical prose while retaining complete identity, scope, and deferred declarations, and raises an actionable error when that minimum cannot fit. Focused plan-context and compact/full prompt tests cover the behavior.

Remaining items

• None.

Tests run

• python3 -m pytest tests/test_issue_756_plan_context.py tests/test_prompts.py tests/test_protocol.py -q (agent-loop instrumented; whole-command timeout 120s)
• python3 -m pytest tests/test_comment_rendering.py tests/test_response_validation.py -q (agent-loop instrumented; whole-command timeout 120s)
• python3 -m pytest tests/test_skill_helpers.py::TestRunPrFix::test_recovered_success_removes_debug_directory tests/test_repair.py::test_run_validated_agent_refuses_plan_revision_missing_direct_discussion_ack -q (agent-loop instrumented; whole-command timeout 120s)
• python3 -m pytest (agent-loop instrumented; whole-command timeout 300s)

-- OpenAI Codex: gpt-5.6-sol (medium)

[wwind123]

Review verdict: Blocking

Both carried items are fixed: prompts, acknowledgement parsing, disposition validation, repair guidance, round metadata and skill-mode repair context now use the digest-backed hr-<64 hex> IDs end to end, legacy positional replies raise an actionable fresh-acknowledgement error instead of being mapped onto the current tuple (protocol.py:779-816, 908-912, 1306), and format_approved_plan_context no longer applies an invented 24,000-character cutoff — it renders the complete plan by default and raises with concrete remediation when a caller-supplied limit cannot hold identity plus complete scope/deferred declarations (prompts.py:1111-1163), with tests for both. One new defect from this round blocks merge: the coder-facing bullet guidance now lists bare surfaced IDs while the markdown acknowledgement parser still requires the literal Requirement prefix, so a response that follows the prompt exactly fails validation. GitHub PR checks are reported as passing; I ran no tests myself in this read-only review.

Blocking issues

• src/coding_review_agent_loop/prompts.py:731-736 now tells the coder to "Include one bullet for each surfaced signed human requirement ID shown in this prompt: hr-<64 hex>", but parse_human_requirements_acknowledgement() (src/coding_review_agent_loop/protocol.py:836-847) only collects IDs matched by \bRequirement\s+(?:hr-[0-9a-f]{64}|\d+)\b, i.e. the bare ID in a bullet is never recognised. A coder that writes ### Human requirements / - \hr-abc…`: implementedtherefore parses to zero addressed IDs, andvalidate_human_requirements_acknowledgement()viaorchestrator.py:3613fails every such response with "did not address all surfaced signed human requirement IDs" — for structured coder responses too, since the markdown section is validated independently ofhuman_requirement_dispositions. The two surfaces also accept opposite spellings: _normalize_requirement_label()accepts only the barehr-…for JSONrequirement_id/addressed_idsand rejectsRequirement hr-…, while the markdown section requires the prefixed form. Before the change the surfaced ID literally was Requirement 1, so both agreed. Tests do not catch this because tests/agent_loop_helpers.py:391-417rewrites fixture output into the prefixed markdown form. Make the two consistent — either accept a bare surfacedhr-…token in### Human requirementsbullets, or state the requiredRequirement ` bullet spelling in the guidance — and add a regression covering a response that uses the bullet form the prompt actually asks for.

Prior unresolved item dispositions

• [item-17] Blocking issue from OpenAI Codex: gpt-5.6-sol (medium), round 10: src/coding_review_agent_loop/prompts.py:582-595,668-695 and `src/coding_review_agent_loop/proto... -> resolved
• [item-18] Blocking issue from OpenAI Codex: gpt-5.6-sol (medium), round 10: src/coding_review_agent_loop/prompts.py:1110-1155 applies a fixed 24,000-character cutoff rathe... -> resolved

-- Anthropic Claude: claude-opus-5 (medium)

[wwind123]

Review verdict: Blocking

The carried stable-ID and oversized-plan findings are resolved, and GitHub checks are passing, but skill-mode PR review still drops authoritative issue and signed-requirement context and can reuse stale approvals.

Blocking issues

• helpers/skill_runner.py:3137-3199 builds skill-mode PR reviews from the PR JSON alone and never fetches or merges the PR-bound primary issue, authoritative staged parent, or their/PR signed human requirements. helpers/prompt_builders.py:228-249 then constructs an IssueContext with no comments or requirements and does not pass human_requirements/parent_issue_context to build_review_prompt; the reviewer validation context is likewise empty. In addition, helpers/state_manager.py:339-355 has no way to persist surfaced_reviewer_requirement_ids, so cmd_run_pr_round() resume filtering only checks plan identity and can reuse a same-head approval after a signed instruction is added or edited. This violates the approved plan's requirement that the same stable-keyed effective requirement tuple and separately labeled parent/primary contexts reach every reviewer and resume path. Fetch and merge the validated primary/parent/PR contexts before skill reviewer invocation, pass them to both external and host review handoffs/validation, persist the surfaced stable IDs, and invalidate completed approvals when the current ID set differs; add a skill-mode regression covering a signed instruction added between interrupted same-head rounds.

Prior unresolved item dispositions

• [item-17] Blocking issue from OpenAI Codex: gpt-5.6-sol (medium), round 10: src/coding_review_agent_loop/prompts.py:582-595,668-695 and `src/coding_review_agent_loop/proto... -> resolved
• [item-18] Blocking issue from OpenAI Codex: gpt-5.6-sol (medium), round 10: src/coding_review_agent_loop/prompts.py:1110-1155 applies a fixed 24,000-character cutoff rathe... -> resolved

-- OpenAI Codex: gpt-5.6-sol (medium)

[wwind123]

PR review round 11 reconciliation: settled reviewers: Claude, Codex. Finalization continues after reconciliation.

[wwind123]

Coder follow-up

Fixed both blocking review items, committed the changes as 968b824, and pushed them to the existing PR #757 branch. The final focused suites passed; one earlier selection named a nonexistent test module and exited before collection, then the corrected existing modules passed.

Addressed items

• item-19: Blocking issue from Anthropic Claude: claude-opus-5 (medium), round 11: src/coding_review_agent_loop/prompts.py:731-736 now tells the coder to "Include one bullet for...
  • Resolution: The markdown acknowledgement parser now accepts the bare stable hr- bullet spelling requested by coder prompts while retaining prefixed and legacy detection. tests/test_issue_756_plan_context.py directly covers the prompt-compatible bare-ID form.
• item-20: Blocking issue from OpenAI Codex: gpt-5.6-sol (medium), round 11: helpers/skill_runner.py:3137-3199 builds skill-mode PR reviews from the PR JSON alone and never...
  • Resolution: Skill-mode PR review and fix flows now recover validated primary and staged-parent issue contexts, merge their signed instructions with PR instructions, pass the same tuple through external and host prompt/validation handoffs, persist surfaced stable IDs in round metadata, and invalidate same-head approvals on any ID-set change. tests/test_skill_helpers.py covers labeled prompt context, host/validator plumbing, metadata persistence, and an instruction added between interrupted same-head reviews; shared prompt/protocol modules cover the adjacent rendering and parsing changes.

Remaining items

• None.

Tests run

• python3 -m pytest tests/test_issue_756_plan_context.py tests/test_skill_helpers.py -k 'legacy_positional_acknowledgement or TestSkillApprovedPlanRecovery or pr_prompt_includes_labeled_parent_primary or attach_metadata' -q (agent-loop instrumented; whole-command timeout 1800s)
• python3 -m pytest tests/test_issue_756_plan_context.py tests/test_skill_helpers.py -q (agent-loop instrumented; whole-command timeout 1800s)
• python3 -m pytest tests/test_prompts.py tests/test_protocol.py tests/test_round_state.py tests/test_human_requirement_dispositions.py -q (agent-loop instrumented; whole-command timeout 1800s)
• python3 -m pytest tests/test_prompts.py tests/test_protocol.py tests/test_human_requirement_dispositions.py -q (agent-loop instrumented; whole-command timeout 1800s)
• python3 -m pytest tests/test_skill_helpers.py::TestRunPrFix::test_success_posts_metadata_after_head_and_assigned_head_advance -q (agent-loop instrumented; whole-command timeout 1800s)

-- OpenAI Codex: gpt-5.6-sol (medium)

[wwind123]

Review verdict: Blocking

Both carried items are addressed. item-19: the acknowledgement parser now accepts the bare surfaced ID with an optional Requirement prefix (protocol.py:836-847) and a regression covers the bullet form the guidance actually asks for. item-20: skill mode now fetches the validated primary and authoritative-parent issue contexts plus PR comments/reviews from the same snapshot used for provenance, merges them into one deduplicated stable-keyed tuple, passes labeled parent/primary contexts and that tuple into external reviewer prompts, reviewer validation, the pr-fix coder prompt and the host review handoff, persists the surfaced stable IDs through attach-metadata/build-resume, and invalidates same-head approvals whose surfaced ID set differs — with regressions for each. One new gap blocks merge: the host-as-reviewer PR path is now validated against signed requirements it is never actually shown in rendered form or told about. GitHub PR checks are reported as passing; I ran no tests myself in this read-only review.

Blocking issues

• The host-as-reviewer PR path is now validated against signed human requirements that the host is never presented with. helpers/validate_response.py:148-166 makes an approved pr_review fail when signed requirements are surfaced and HUMAN_REQUIREMENTS_RESOLVED is absent, and cmd_complete_host_review() reaches that check through _complete_reviewer_turn(context_file=request_dir/"context.json"). But unlike every external reviewer, the host gets no generated prompt: _write_host_review_request() (helpers/skill_runner.py:2765-2860) writes the requirements only as raw serialized dicts inside context.json and inside the new primary-issue.json / authoritative-parent-issue.json artifacts — there is no rendered format_human_requirements() block, no list of the surfaced hr-… IDs, and no instruction to emit the resolution marker. The printed hint only says to read approved-plan.md, "the labeled issue context artifacts when present", and pr-diff.diff, and neither SKILL.md nor docs/skill_mode.md mentions signed human requirements at all (docs/skill_mode.md:120-128 documents the sibling plan artifact but not these). Concretely: a signed instruction is posted on the PR, the host reads the diff and plan as instructed, writes an approved pr_review, and complete-host-review rejects it for a requirement set the workflow never surfaced. Render the signed-requirement block (with its stable IDs and the marker/acknowledgement contract) into the host review request, name that artifact in the printed hint, document it alongside the approved-plan artifact, and add a host-review regression covering a PR with a surfaced signed requirement.

Prior unresolved item dispositions

• [item-19] Blocking issue from Anthropic Claude: claude-opus-5 (medium), round 11: src/coding_review_agent_loop/prompts.py:731-736 now tells the coder to "Include one bullet for... -> resolved
• [item-20] Blocking issue from OpenAI Codex: gpt-5.6-sol (medium), round 11: helpers/skill_runner.py:3137-3199 builds skill-mode PR reviews from the PR JSON alone and never... -> resolved

-- Anthropic Claude: claude-opus-5 (medium)

[wwind123]

Review verdict: Blocking

The markdown acknowledgement mismatch is fixed, but item-20 remains unresolved: skill-mode reviewer metadata records every merged signed-requirement ID even though prompt rendering can omit older requirements at its 12,000-character bound. The same-head filter can therefore reuse an approval as covering requirements that reviewer never received; persist the IDs actually rendered (and handle omitted requirements explicitly) instead.

Prior unresolved item dispositions

• [item-19] Blocking issue from Anthropic Claude: claude-opus-5 (medium), round 11: src/coding_review_agent_loop/prompts.py:731-736 now tells the coder to "Include one bullet for... -> resolved
• [item-20] Blocking issue from OpenAI Codex: gpt-5.6-sol (medium), round 11: helpers/skill_runner.py:3137-3199 builds skill-mode PR reviews from the PR JSON alone and never... -> still blocking

-- OpenAI Codex: gpt-5.6-sol (medium)

[wwind123]

PR review round 12 reconciliation: settled reviewers: Claude, Codex. Finalization continues after reconciliation.

[wwind123]

Coder follow-up

Implemented both requested skill-mode review fixes, committed them, and pushed commit ddb768d to PR #757.

Addressed items

• item-20: Blocking issue from OpenAI Codex: gpt-5.6-sol (medium), round 11: helpers/skill_runner.py:3137-3199 builds skill-mode PR reviews from the PR JSON alone and never...
  • Resolution: Skill-mode PR review now recovers validated primary, authoritative parent, and PR contexts; merges stable signed-requirement IDs into reviewer prompts and validation; persists surfaced IDs; and invalidates same-head approvals when the requirement set changes. The focused and complete skill-helper modules cover the propagation and interrupted-resume behavior.
• item-21: Blocking issue from Anthropic Claude: claude-opus-5 (medium), round 12: The host-as-reviewer PR path is now validated against signed human requirements that the host is...
  • Resolution: Host PR-review handoffs now write a complete signed-human-requirements.md artifact containing every stable ID and the approval acknowledgement contract, name it in the host hint and manifest, remove stale artifacts on requirement-free retries, and document the workflow. Regression coverage verifies the rendered artifact and rejects approval without the required acknowledgement while accepting a compliant host review.

Remaining items

• None.

Tests run

• python3 -m pytest tests/test_skill_helpers.py -k 'skill_pr or host_handoff or host_review_pr or signed_instruction_is_added' -q (agent-loop instrumented; whole-command timeout 1800s)
• timeout --signal=TERM --kill-after=10s 1800s python3 -m pytest tests/test_skill_helpers.py -q
• timeout --signal=TERM --kill-after=10s 1800s python3 -m pytest tests/test_issue_756_plan_context.py -q

-- OpenAI Codex: gpt-5.6-sol (medium)

[wwind123]

Review verdict: Approved

Both carried items are fixed at this head. item-20: skill-mode PR rounds now recover the validated primary and authoritative-parent issue contexts plus PR comments/reviews from the same snapshot used for provenance (helpers/skill_runner.py:_recover_skill_pr_review_contexts, field lists matching get_issue_context/_parse_pr_human_requirements), merge them into one deduplicated stable-keyed tuple, pass labeled parent/primary contexts and that tuple into external reviewer prompts, reviewer validation, the pr-fix coder prompt and the host handoff, persist the surfaced stable IDs through attach-metadata/build-resume, and drop same-head reviewer records whose surfaced ID set differs — with regressions for each, including an instruction added between interrupted same-head rounds. item-21: the host handoff now writes signed-human-requirements.md rendering every requirement untruncated plus the same _human_requirements_review_guidance() contract external PR reviewers receive (no plan-disposition text, matching the PR reviewer shape), removes a stale artifact when no requirements are surfaced, records the file in the manifest, names it in the printed hint, and documents it in SKILL.md and docs/skill_mode.md; two regressions cover the rendered artifact and end-to-end complete-host-review enforcement of the resolution marker. The rest of the PR remains coherent: PR-bound plan selection by handoff identity with fail-closed diagnostics, complete plan text by default with an explicit remediation error when a caller-supplied limit cannot hold identity plus declarations, digest-backed signed-requirement IDs accepted consistently in JSON fields and markdown bullets, staged parent/child recovery in both headless and skill paths, and ordinary unplanned PRs still taking the explicit no-plan path. GitHub PR checks are reported as passing; I ran no tests myself in this read-only review.

Prior unresolved item dispositions

• [item-20] Blocking issue from OpenAI Codex: gpt-5.6-sol (medium), round 11: helpers/skill_runner.py:3137-3199 builds skill-mode PR reviews from the PR JSON alone and never... -> resolved
• [item-21] Blocking issue from Anthropic Claude: claude-opus-5 (medium), round 12: The host-as-reviewer PR path is now validated against signed human requirements that the host is... -> resolved

-- Anthropic Claude: claude-opus-5 (medium)

[wwind123]

Review verdict: Approved

The carried skill-mode requirement propagation and host-review handoff defects are resolved; no remaining merge blockers were found, and the reported GitHub checks are passing.

Prior unresolved item dispositions

• [item-20] Blocking issue from OpenAI Codex: gpt-5.6-sol (medium), round 11: helpers/skill_runner.py:3137-3199 builds skill-mode PR reviews from the PR JSON alone and never... -> resolved
• [item-21] Blocking issue from Anthropic Claude: claude-opus-5 (medium), round 12: The host-as-reviewer PR path is now validated against signed human requirements that the host is... -> resolved

-- OpenAI Codex: gpt-5.6-sol (medium)

[wwind123]

PR review round 13 reconciliation: settled reviewers: Claude, Codex. Finalization continues after reconciliation.

[wwind123]

Reviewers approved PR #757; watching GitHub checks in the foreground. No coder or reviewer agents will run while checks remain pending.

-- coding-review-agent-loop