Conversation
Two 2026-09-06 prompt audits, deduplicated. The load-bearing fix: proof-harness-router.md offered demoting a post-admission candidate to referent_unresolved after an exhausted retry budget, contradicting loop-engineering.md (demotion is pre-admission only) and the validator (referent_unresolved is a candidate-fusion decision, blocked from closing); following it would break the enforced in_scope_count == len(verification-sheet) identity. The router now states the loop-engineering rule. Also: tool-firewall.md drops the false "tracked v0.16.x follow-up" claim (validate_planner_ingest ships and fires since v0.23.0) and the "built in a subsequent phase" fossil on quarantine_extract.py; dangling review-round/spec markers ((R1)..(R25), section 3.12/3.15.8, R5, Guardrail A) removed from five references while B1/section 3.15.6 stay (live in validator error strings, AST-verified); skill-maintainer build guidance moved to a new AGENTS.md Skill Maintenance section; migration-relative phrasing rewritten as present-state rules; the skill.json routing description drops the behavioral clause the body already mandates. Doc accuracy only: no validator, vocabulary, or workflow-semantics change. Surfaces regenerated; consensus record added; bump to 0.32.2. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01EUo5YkQZdCTkn2bBfY9hou
|
@codex review |
Codex Review SummaryThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: dfb6ae2eef
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| The five-agent official-docs recheck was run and recorded 2026-09-06 (feature-matrix refresh | ||
| date, `docs/authoritative-sources.md`); this change alters no agent envelope, packaging, or | ||
| adopted feature, so that one-day-old recheck stands as current. |
There was a problem hiding this comment.
Because this commit substantively changes runtime-loaded references and the skill's routing metadata, reusing the v0.32.1 survey does not satisfy the required same-pass refresh. Lines 53–55 acknowledge that neither docs/authoritative-sources.md nor skill-core/references/agent-feature-matrix.md was updated for v0.32.2, leaving this release without its mandated dated sources, per-agent decisions, and repeatable feature-matrix record. Perform and record the refresh for this change rather than carrying forward the prior release's entry.
AGENTS.md reference: AGENTS.md:L87-L95
Useful? React with 👍 / 👎.
P1: reusing the v0.32.1 survey does not satisfy the AGENTS.md same-pass refresh requirement for a substantive reference/metadata change. Re-ran the five-agent official-docs recheck (research date 2026-09-07): all five rows n/a/unchanged; new dated entry in docs/authoritative-sources.md, feature-matrix refresh date and per-row recheck dates bumped, consensus record corrected. One advisory observation recorded (Gemini CLI announces future replacement by Antigravity CLI for unpaid tiers; .agents/skills aliasing unchanged). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01EUo5YkQZdCTkn2bBfY9hou
|
Round 1 fix (365ae54): accepted the P1 — re-ran the five-agent official-docs recheck for this pass (research date 2026-09-07) instead of carrying the v0.32.1 entry forward. New dated entry with per-agent sources and n/a-unchanged decisions in docs/authoritative-sources.md; agent-feature-matrix refresh date and per-row recheck dates bumped; consensus record corrected. make verify green. |
|
@codex review |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 365ae54983
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| reproduction rate as `sound_with_limits` limit metadata (never as the pass/fail), and stay | ||
| `blocked_pending_calibration` without TSan confirmation. See `references/proof-harness-router.md` Retry/Stop | ||
| and §3.12 nondeterminism handling. | ||
| and the nondeterminism handling in `references/promoted-replay-contract.md`. |
There was a problem hiding this comment.
For the unpinned-race path described above, this sends the operator to promoted-replay-contract.md for “nondeterminism handling,” but that file only defines generic replay fields and fixed-promotion transcript rules; it never discusses bounded stress loops, reproduction-rate metadata, or blocked_pending_calibration (a repository-wide search finds that handling only in this file). The cleanup therefore replaces a dangling section number with a misleading target, preventing the promised detail from being loaded on demand; remove the pointer or link/add the actual guidance.
AGENTS.md reference: AGENTS.md:L96-L101
Useful? React with 👍 / 👎.
1 participant
Applies the deduplicated High/Medium findings from two 2026-09-06 prompt audits (repo audit + installed-skills audit). One line per finding:
proof-harness-router.mdoffered demoting a post-admission candidate toreferent_unresolvedafter two failed proof attempts; that contradictsloop-engineering.md(demotion is pre-admission only; an admitted candidate keeps its in-scope row and closes viaclosure-bundles.json) and the validator, which makesreferent_unresolvedacandidate-fusion.jsondecision value and blocks it from closing. Following the router would break the enforcedin_scope_count == len(verification-sheet.json)reconciliation identity — a finding would silently vanish, the exact failure the Finding-Completeness gate exists to prevent. The router now states the loop-engineering rule (the side backed by code).tool-firewall.mdclaimed the read-side planner-ingest link "remain[s] a tracked v0.16.x follow-up";validate_planner_ingestships and fires since v0.23.0 (repo at 0.32.x). Rewritten to state the enforced behavior. (Both audits flagged this; applied once.)tool-firewall.md"built in a subsequent phase — not this reference" fossil onscripts/quarantine_extract.py, which ships today; parenthetical removed.discovery-intake.md(10 sites, every rule preserved). (Both audits; applied once.)sanitizer-selection.md(pointer now namesreferences/promoted-replay-contract.md), §3.15.8 + R5 intool-firewall.md, "§3.15.8 replay fields" incontrolled-fields.md, "Guardrail A" insurface-tool-inventory.md. KeptB1and§3.15.6: AST-verified they appear verbatim in validator runtime error strings, so they are live vocabulary mapping gate output to its reference, per the audit's own keep-list criterion.discovery-orchestrator.md,discovery-specialist-router.md("Deterministic Validator Candidates"), andshadow-skill-bootstrap.mdinto a newAGENTS.md"Skill Maintenance" section.discovery-specialist-router.md("Specialist Reviewer Redesign" → "Specialist Reviewers") andartifact-contract.md("preserve the prior semantics" dropped).skill.jsonrouting description drops the trailing behavioral imperative ("validate it alongside self-found candidates") that the canonical body andAGENTS.mdalready mandate; both routing signals kept.Doc accuracy only — no validator, vocabulary, or workflow-semantics change.
make syncregenerated all four adapter packages;make verifygreen (797 unit + 9 integration). Version bumped to 0.32.2 with CHANGELOG entry and consensus record (reviews/prompt-audit-contract-drift-consensus.md) per the 0.32.1 precedent.proof-harness-router.mdsize budgets raised 57→59 lines with comments citing the record.🤖 Generated with Claude Code
https://claude.ai/code/session_01EUo5YkQZdCTkn2bBfY9hou