Codex Stop hook needs operator-dialogue fallback

Context

Friction surfaced live on 2026-06-28 while trying to end an operator-directed Codex turn as operator dialogue/planning. Claude/Grace can stop on this path; Codex kept receiving the Stop-hook no-hold reminder instead.

This is not a request to weaken §no_hold_state. It is the hook's own friction→gold clause firing: when a human operator is actively in dialogue, the hook must provide a stop option that is externally confirmable even when the Codex Stop payload cannot directly expose the prompt text.

The Problem

The current Codex Stop-hook path can fail closed into autonomous/no-hold even for a live operator turn.

Live V-B-A from the current session:

• Claude Stop audit logs show repeated ALLOW operatorInLoop=true entries for operator-dialogue turns.
• Codex Stop audit logs for the same period show BLOCK ... promptSource=transcript_path, operatorInLoop=false, including the operator-dialogue attempt.
• Running the current Codex helper against the live Codex transcript returned userLength: 0, assistantLength: 0, promptSource: transcript_path, operatorInLoop: false.
• Inspecting the live Codex transcript shape showed messages under record.payload (payload.type="message", payload.role="user"). The Codex JSONL helpers currently inspect only root/message/item-shaped records, so the transcript exists but the prompt and assistant text are not extracted.

That explains today's failure, but the operator highlighted the wider contract: even if a Stop payload or transcript cannot expose the prompt directly, operator dialogue still needs a safe stop path. Otherwise the hook becomes a leash during live human dialogue, exactly the false-positive class the friction→gold clause says to turn into substrate.

The Architectural Reality

• Shared semantics already live in ai/scripts/lifecycle/stopHookDecision.mjs: live operator dialogue is the only voluntary allow; autonomous [WAKE] / stop-hook continuations must continue driving.
• Harness adapters are still split across .claude/hooks/laneStateStopHook.mjs and .codex/hooks/codex-lane-state-stop.mjs.
• Codex has .codex/hooks/codex-context.mjs on UserPromptSubmit, which is a plausible external prompt-class capture point before Stop runs.
• There is no shared ai/agent/hooks substrate yet. Current parity is shared decision primitives only, not shared payload normalization, prompt provenance, or adapter contract.
• #13796 is the broad generic-adapter design surface. This ticket is the narrow bug leaf needed now: operator-dialogue stop fallback + live Codex transcript-schema coverage.

The Fix

Implement a Codex Stop-hook fallback that preserves the no-hold invariant while allowing operator dialogue when it is externally confirmable.

Expected shape:

1. Teach Codex transcript extraction to parse the real Codex JSONL schema where message records live under record.payload.
2. Treat synthetic hook records such as <hook_prompt ...> and <turn_aborted> as hook noise for operator-dialogue classification, so they do not hide the preceding human operator prompt.
3. Add a fallback confirmation path for cases where Stop cannot see a prompt directly: the UserPromptSubmit hook should persist a short-lived same-turn prompt-class/nonce record that the Stop hook can read. The fallback must be external to the assistant final text; a self-declared final-response marker is not enough.
4. Keep autonomous [WAKE], stop_hook_active, handoff-to-autonomous, and missing-evidence cases blocked/continued. The fallback is for confirmed operator dialogue, not for self-authorized idle.

Contract Ledger Matrix

Decision Record impact

Aligned with #13796's generic-by-default harness adapter direction and #14304's v13.2 architecture-quality work. No ADR change required for the narrow bug fix; a later shared ai/agent/hooks extraction may need a Decision Record if it establishes a new cross-harness substrate boundary.

Acceptance Criteria

• Codex JSONL extraction supports real response_item.payload message records for both user and assistant text.
• Codex operator-dialogue classification skips synthetic hook/noise user records such as <hook_prompt ...> and <turn_aborted> when finding the human prompt for the current turn.
• Codex has an externally confirmable operator-dialogue fallback when Stop cannot directly see prompt text, preferably via a same-turn prompt-class/nonce written by UserPromptSubmit and read by Stop.
• The fallback does not accept assistant self-declaration as authority; it must be tied to a hook/prompt provenance record.
• Regression tests cover: real Codex transcript fixture, direct operator prompt allow, missing-Stop-prompt fallback allow, [WAKE] block, stop_hook_active block, handoff-to-autonomous block, and malformed transcript fail-open behavior.
• Claude behavior remains unchanged or has matching shared-helper coverage if extraction is factored.

Out of Scope

• Rewriting the full harness adapter surface from #13796.
• Relaxing §no_hold_state for autonomous turns.
• Adding a broad ai/agent/hooks domain unless the narrow fix proves the shared helper is needed immediately.
• Changing PR merge / issue close human-authority gates.

Avoided Traps

• Self-declared operator dialogue: rejected. The assistant final text cannot authorize its own stop.
• PromptSource-none equals permanent block: rejected. It is safe for autonomous uncertainty, but live operator dialogue needs an external fallback path.
• Broad architecture rewrite first: rejected. Fix the live parity bug, then feed any shared-substrate extraction into #13796 / #14304.

Related

• #14034 — prior operator-dialogue wording/hint work; closed, but did not cover the live Codex transcript-schema/fallback failure.
• #14093 — adjacent operator-gated closeout tension; related but not equivalent.
• #13796 — broad generic harness adapter surface.
• #13623 / #13822 — no-hold and value-floor design surfaces.
• #14304 — v13.2 Agent OS architecture-quality umbrella.

Live latest-open sweep: checked latest 20 open issues at 2026-06-28T21:51:14Z; no equivalent found. Nearby non-duplicates checked: #14093, #13796, #13623, #13822. Exact GitHub searches for Codex Stop hook operator dialogue and promptSource transcript_path operatorInLoop found prior closed/adjacent work but no open equivalent.

A2A in-flight claim sweep: checked latest 30 A2A messages across read states at 2026-06-28T21:51:14Z; no competing [lane-claim] / [lane-intent] on this exact fallback scope found.

Origin Session ID: 3990502e-346a-47d6-8376-490e4802829c

Handoff Retrieval Hints: query_raw_memories("operator dialogue stop option missing prompt visible Codex Stop hook friction gold"); query_raw_memories("Codex Claude Stop hook operator dialogue lane-state parity fail-open promptSource none"); current live evidence centered on Codex session 019f0f79-7af4-7a70-893a-dd58e03337d6.