fix(codex): count explicit structured selections as human turns (2.6.44)

[gadzbi123]

Summary

• mint the existing HUMAN_TURN audit event after answered Codex request_user_input calls
• keep typed UserPromptSubmit handling unchanged
• pin the generated PostToolUse matcher and regenerate all distributions

This fixes approval and interview gates rejecting a structured selection until the user types a second acknowledgment.

Testing

• bun test tests/unit/t150-codex-packaging.test.ts tests/unit/t132-hooks-doc-count-sync.test.ts tests/unit/t68-version-changelog-sync.test.ts
• bun scripts/package.ts --check
• full unit runner: 138/140 files passed; t140 and t142 could not load the optional @anthropic-ai/claude-agent-sdk dependency in this checkout

[apackeer]

Thanks, Kacper. This addresses a real Codex gap: structured request_user_input selections still do not mint the human-turn evidence that typed prompts do, so the direction remains useful.

This path needs to preserve the framework's responsibility split: the LLM supplies the knowledge and context needed to present the question, the human supplies judgment by making a selection, and deterministic code verifies and records that decision. The current wiring collapses the last two responsibilities: PostToolUse proves only that the tool returned, but the adapter ignores tool_response and appends HUMAN_TURN unconditionally as though a human had decided.

Codex can return {} in headless mode, and #681 documents cancelled or automatically resolved questions arriving as completed-looking responses. Because HUMAN_TURN authorizes approval and interview gates, the adapter must require an explicit, nonempty human selection and reject empty, cancelled, dismissed, timed-out, auto-resolved, error, and malformed responses.

Adapter-level tests should cover those cases, a genuine selection, duplicate delivery minting exactly once, no active workflow state, and one real approval/interview path. #692 overlaps the downstream non-answer checks but does not add this Codex registration, so the two pieces need to land coherently.

The upgrade note also needs to tell Codex users to re-run the hook-trust step. There is no need to rebase this stale branch or resolve its old version slot; maintainers can port the corrected change and handle regeneration and versioning.

[leandrodamascena]

Maintainer port pushed at 804910036.

Codex structured selections now mint human-presence evidence without turning non-decisions into authority:

• Added a request_user_input PostToolUse registration using the existing record-human-turn target.
• The adapter accepts only the real Codex JSON-encoded response contract: answers: Record<questionId, { answers: string[] }>.
• Empty objects, malformed JSON, unsupported response shapes, blank answers, cancellation boilerplate, timeout/error/auto-resolved metadata, and partial multi-answer cancellations fail closed.
• Values such as Abort count only when they exactly match an option presented by that question.
• Question IDs remain opaque; an ID such as error is not mistaken for failure metadata.
• Duplicate Codex delivery emits exactly one HUMAN_TURN.
• Projects without active workflow state remain untouched.
• Typed UserPromptSubmit behavior remains unchanged.
• Codex hook wiring and trust hashes were regenerated.
• Upgrade guidance now explicitly requires regenerating and replacing the Codex hook-trust entries.
• The stale contribution was ported directly onto current v2 and versioned as 2.6.44.

Validation:

• Post-rebase focused slice: 93 tests, 1,794 assertions, 0 failures.
• Codex adapter, packaging, trust, and version slice: 50 tests, 362 assertions, 0 failures.
• Package parity, coverage registry, typecheck, lint, and git diff --check passed.
• The full deterministic profiles retained only the existing unrelated local timeout/baseline failures.
• Independent final review found no remaining issues.

The PR is ready for CI and final approval.

[leandrodamascena]

The stale contribution has been safely ported onto current v2 at 804910036. Only explicit Codex structured selections mint HUMAN_TURN; non-decisions fail closed, duplicate delivery is idempotent, and trust/package validation is green. Approved pending CI.