plan: Codex native review protocol for agent-harness#924

plans/decision-interim-president-ratification-20260904.md

@@ -138,3 +138,4 @@ closed by construction. Four converged PRs (Consiliency/agent-harness#765, #768,
 | Consiliency/agent-harness#797 (issue; PR pending) | plan / production_code | plan candidate `7245a6259ddb54ce61e3af878d9c2dee607a1678`; implementation review and landing pending | Bounded A2 assumption reconciliation under the coordinator's single-writer slot. The complete plan and sole manifest registration received fresh Codex, Claude Opus 5/high, Gemini and Grok AGREE in round 1; artifact SHA256 `8b67426e8ff0f0cf87cea3b11e7a14d9ca0056c3d3fedc6dc3ee092f22dd6a52`. Canonical-repository authority, exact packet bindings, heartbeat-only/null model and silence deadlines, and all cleanup receipts validated. Existing two-node RED is the typed OPEN/CLOSED and 0.5.0/0.7.14 mismatch, not missing observations. Scope preserves historical RUNTIME metadata and all existing authority prefixes; no historical phase reacceptance, HARDEN execution, release or landing is claimed. At implementation `505c852fff551bc10544214cc2305a1884531f7c`, declared automation passed 296 tests with 2 skips and 1 explicitly named A3 deselection; the full suite failed with 11 failed, 6698 passed and 721 skipped (receipt SHA256 `c194e840b9158fc8376d054076c20c2322ff73065f9f105d604c12f4ff952062`). Ten failures match retained baseline causes; the remaining unchanged ownership sentinel reproduces in baseline and candidate source exports lacking local build metadata (disposition SHA256 `0814c6e839d02b6c612843e1337c59a23d8529c740d81b015b919d3aaabc3388`). The detailed plan records failed verification, not completion. Exact-candidate implementation review, required CI and landing remain pending; these results do not attest a later integrated candidate. |
 | Consiliency/agent-harness#981 | plan / production_code | exact final candidate recorded in the PR body; implementation review and CI pending | Bounded agent-harness#797 A2 maintenance under the standing operator instruction and this interim note. Prior plan/registration received four fresh vendor AGREE votes; these do not approve implementation. Fresh automation on f7c4932b: 296 passed, 2 skipped, 1 explicit A3 deselection; source505c full remains failed (11 failed / 6698 passed / 721 skipped), with retained cause comparisons and metadata-layout limits. Historical RUNTIME grounding, all prior authority prefixes and the own-plan failed-verification event remain unchanged. Landing requires four fresh scoped implementation approvals, verified broker/isolation/cleanup evidence and green required CI on the final candidate. No president ruling, HARDEN retry, whole-phase completion, native-filled A3 observation or release is claimed. |
 | Consiliency/agent-harness#974 | production_code | `84a85885` (merge commit recorded on landing) | THREE-vendor board, NOT four: the codex seat was UNAVAILABLE for the entire review (subscription usage limit, "try again at Sep 28th, 2026"), recorded as an absence and NOT substituted with another reviewer. Usable standing verdicts from three distinct vendors: xAI/grok AGREE (r1, and again at r3 after a delta), Google/agy AGREE (r2 — its r1 and r2 runs ERRORed on the broker's chunk-acknowledgement contract and left no standing verdict, so it was re-run rather than carried; filed as agent-harness#983), Anthropic/claude native DISAGREE at r1 and again at r3, each on a real blocking defect, each fixed, closed by a single-seat delta. Three rounds under a cap announced before round 1, plus one dissent-closing delta. FOUR consumers of the moved clamp were missed by the author's symbol-keyed affected-test selection and found by review: `fixtures.DEFAULT_SEAT_EFFORT_ARGS`, `test_grok_leg_argv_is_headless_plain_with_reasoning_effort`, `scripts/verify_harden_evidence.py` `ARGV_GROK_EFFORT` (a LIVE fail-closed HARDEN gate that rejected the token the change emits — the only production-affecting one), and `test_phase_loop_launcher.py:1538` (a would-red assertion hidden behind the CI-deselected `dotfiles_integration` marker, found independently by the grok and native seats). A unit-time drift guard was added so the standalone verifier's grammar cannot silently lag the clamp again, proven non-vacuous by mutation. Exact-head CI green and required contexts present; note that earlier green readings on this PR belonged to a superseded sha while it was CONFLICTING with main (GitHub does not run pull_request checks on a conflicted PR), filed as agent-harness#984. No release, HARDEN execution, or phase completion is claimed. |
+| Consiliency/agent-harness#986 | plan | exact final candidate recorded in the PR body; plan review and required CI pending | Planning-only Codex/OpenAI native-cell slice of EC-REVIEWTRUTH-18 (agent-harness#924), under the standing operator instruction and this interim note. Initial broker draft at `9483cb72` passed commands-only native manifest, dispatch, ownership and history validation; receipt SHA256 `cb9e6c2c2acce7f4ab8fc6530fc848745b6ae785e5fbc449535c7cdf1ad57a2b`. Round 4 on the earlier `29887e24` did not converge: Scope A Codex/Gemini AGREE and Grok PARTIALLY AGREE; Scope B Codex/Gemini/Grok AGREE; both Opus seats ended without a verdict after explicit HTTP 429 errors and authorized cleanup. All eight records are retained publicly on agent-harness#924; no absence or earlier vote transfers to this candidate. The two bounded plan clarifications and this conditional ledger append require fresh four-vendor scoped plan convergence, joint disposition, exact-candidate native validation and green required CI before landing. Historical manifest registration remains unchanged. CONFORM/HARDEN/SCHED, REVIEWTRUTH SL-0/SL-1 and EC-REVIEWTRUTH-15 execution gates remain closed; no EC-18 waiver, HARDEN retry, implementation, native operational qualification, president ruling, phase completion, release or issue closure is claimed. |