[ViperJuice]

This draft records the bounded Codex/OpenAI native-review protocol plan for agent-harness#924. It changes only the detailed plan, its historical manifest registration and one conditional ledger row for agent-harness#986. It does not implement the protocol or admit execution.

Candidate: d427bbe1d01e9c092a912782a5e78aeed705f67c, tree 26add9f77c9ff2b26ed56b7ce1f3fb805a2a4ba4, based on main 939b9b3a15e5b8faf60df6f57c18cf2966719cb8. Current plan SHA-256: ee1a9576a66055efbb50e30ef5e077654f85154b3bb58d705b93a4f33c79e627.

The planned cell covers an already-composed, unambiguous single Codex seat under EC-REVIEWTRUTH-18, using existing emit/fill/invoke bindings. It keeps native seats unsupported under heartbeat-only, preserves legacy Claude behavior and explicit child/ambiguous-host refusals, separates request-side posture from actual enforcement, and retains the documented missing-fill/floor-only residual. Other cells and agent-harness#924 remain open.

The manifest remains committed, with planning_only=true and execution_admitted=false. Its original plan digest is historical registration evidence; all earlier row bytes/order, lifecycle events and authority history remain unchanged. This PR does not reseal the roadmap or rewrite historical RUNTIME metadata.

Review is incomplete. The preceding 29887e24 plan underwent four-vendor review in two complete scopes. Scope A: Codex/Gemini AGREE, Grok PARTIALLY AGREE, Opus unavailable/no verdict. Scope B: Codex/Gemini/Grok AGREE, Opus unavailable/no verdict. Both Opus calls recorded explicit HTTP 429 rate-limit errors. Their adapters were orderly canceled through verified owned drivers; cleanup completed, but the strict validators failed on the non-OK seats. No quota response, cancellation or older vote counts as approval. All reports and the full joint disposition are public on agent-harness#924: Scope A record, Scope A findings, Scope B record, Scope B Grok.

This candidate adds only the agreed bounded clarifications: the Codex Policy tests explicitly inject their markers, while retained omitted-env arms still exercise actual per-call ambient detection; static eligibility resolution names the existing no-live-probe matrix and precise PATH/auth tripwires, without changing policy precedence or later composition checks. All runtime/test bytes remain unchanged. Compared with the previous review candidate, only the plan and the new conditional ledger row differ; the original manifest registration is unchanged. These new bytes do not yet have fresh review approval.

Validation: commands-only native plan-publication validation passed on this three-path candidate. Manifest, dispatch literals, exact source/ownership/history preservation and diff checks passed. All 1,538 tracked files and 212 qualified runtime pins stayed unchanged. Receipt SHA-256 f0a71df10586ff5b98ac0166fb1941ff770d4861b05825c17124f74e58f66900; native log 6f02345bcb180dacd6fcfb48ffdcf5aeae6fe4aa7d9ab7a827a4ff7659050a71. The future implementation automation.suite_command remains unexecuted: its tests and implementation are gated. No RED lane, production code, runtime-test or native-provider qualification is claimed. Exact-head CI run 35768715858 is green on d427bbe1: Python 3.10/3.11/3.12, standalone-wheel Gate A and the suite gate all passed. Workflow-conditioned ai-offload/main-red skips are retained as skips. The terminal observation SHA-256 is eab5f0d6ac252daa64a98b15e4879e60c2a839e8e1a1bded5ee1db39b7600233. Fresh four-vendor review remains pending; CI and earlier votes do not supply review approval.

Landing uses the standing interim president-ratification note only while it remains active. The initial two-path draft was broker-published as agent-harness#986 at 9483cb72. This candidate adds exactly one conditional agent-harness#986 ledger row while preserving every previous ledger byte and the historical manifest registration. It is the actual three-path plan/registration/ledger candidate for fresh review; the row does not claim convergence or landing. Four-vendor plan convergence, exact-candidate evidence and required green CI remain required; no president ruling is claimed.

CONFORM/HARDEN/SCHED, REVIEWTRUTH SL-0/SL-1 and EC-REVIEWTRUTH-15 execution requirements remain in force. The prior HARDEN refusal remains controlling, and the EC-14 early exception supplies no EC-18 waiver. No automatic bounded-policy fallback, reviewer substitution, implementation admission, release or issue closure is requested by this draft.

Source-only R5 A/B preparation is unlaunched; all four vendors must review both assigned scopes and a joint disposition must cover their boundary before any whole-plan convergence claim. No earlier vote transfers.

Current panel selection (operator update, 2026-09-22): Fable is restored as the canonical panel default, claude-fable-5-1 at max effort through the subscription TUI. Successor drivers use the runtime default; historical Opus requests, refusals and reports remain intact. The first bound Fable operation recorded a real selected-model response after reset; that establishes availability, not a review verdict. Heartbeat-only and four distinct vendors remain required. The two prepared plan scopes remain unlaunched. Read-only Git merge-tree inspection confirms append conflicts with agent-harness#978 in plans/manifest.json and the interim-ratification ledger. After agent-harness#978 settles, integrate this plan onto current main by preserving its current raw prefix and appending only the owned delta, then validate and freshly seal/review the resulting candidate. This avoids spending a board on bytes known to need integration. Existing exact-head CI remains green; implementation gates remain closed.

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}

[ViperJuice]

Triage (2026-09-25): left with codex. This is a plan-only draft for agent-harness#924 (host×seat routing) and has been idle since 2026-09-23. Please confirm whether it is still wanted. Note that the panel vendor fallback-chain work now starting (a [panel] preference order plus min_distinct_vendors) interacts with the host×seat routing rule, so the two designs should agree.