feat: align hook lifecycle with upstream Codex (v1.5.0)

[sf-jin-ku]

Summary

Upstream Codex moved on since v1.3.0; three behaviors here were still written against the old runtime.

• Restore the SessionEnd hook. Codex 0.146+ dispatches SessionEnd for root sessions, so the teardown removed in v1.3.0 as non-dispatched now runs again: it reaps background jobs whose process died and clears this session's current-session marker instead of leaving it to age out after seven days. Teardown stays inside Codex's few-second budget (timeout: 3) — it never kills or waits on live processes, so detached jobs keep running and the UserPromptSubmit sweeper still covers them.
• Stop requiring [features].plugin_hooks. Upstream retired that flag (Stage::Removed), and native plugin hooks now ride on [features].hooks alone. Setup requires only hooks = true and strips a leftover plugin_hooks line from ~/.codex/config.toml, the same way it already upgrades the legacy codex_hooks alias.
• Correct the review skills' description of request_user_input. Omitting [tools] experimental_request_user_input leaves the tool enabled; only an explicit false hides it in an interactive thread. The conditional ask is unchanged — it still depends on the thread actually having a question tool.

Upgrade notes

• Existing users' ~/.codex/config.toml gets its stale plugin_hooks = true line stripped on the next $cc:setup run.
• A Codex restart is needed after the hooks/hooks.json change for the new SessionEnd registration to take effect.

Tests

• tests/hooks.test.mjs: SessionEnd reaps a dead-PID job to failed, leaves a live job running, and drops the session marker; a SessionEnd from a different session leaves the active marker alone. Hook-name allowlist widened to the current 12 HOOK_EVENT_NAMES.
• tests/codex-config.test.mjs: ensureNativePluginHooksEnabled requires only hooks and strips plugin_hooks; nativePluginHooksStatus reports from hooks alone.
• Existing plugin_hooks = true assertions across install/installer/e2e/skills-contract tests inverted to doesNotMatch.
• npm run check passes (lint, typecheck, unit, integration, e2e).

Note: the SessionEnd entry copies statusMessage from SessionStart; if upstream ignores it for that event it is simply inert.

🤖 Generated with Claude Code

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-08-30T00:49:18.217131Z	17848c2	Manual request

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" or "@codex security review".

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]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: e8da74826c

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".

[chatgpt-codex-connector]

Synchronize the lockfile with the release version

This release bumps package.json to 1.5.0, but package-lock.json still declares 1.4.2 in both its top-level version and packages[""] entry. As a result, a standard npm install --package-lock-only immediately rewrites the tracked lockfile, leaving release metadata inconsistent; update those lockfile entries as part of the version bump.

Useful? React with 👍 / 👎.

[upwind-code-us]

Upwind Code Scan - ✅ Proceed with Deployment

0 newly introduced vulnerabilities · 0 resolved · 0 total in this PR vs main

View full analysis in Upwind Console

Scan completed in 18s

Scan history (4 scans)

Commit	Scanned at	New	Resolved	Net
e8da748	2026-08-30 00:19 UTC	0	0	0
17848c2	2026-08-30 00:29 UTC	0	0	0
5068eb8	2026-08-30 01:13 UTC	0	0	0
df064f9 <	2026-08-30 01:39 UTC	0	0	0

Last scanned: df064f9 · 2026-08-30 01:39 UTC

[upwind-code-us]

Upwind IaC Scan - ✅ Proceed with Deployment

0 newly introduced misconfigurations · 0 resolved · 0 total in this PR vs main

View full analysis in Upwind Console →

Scan completed in 2s

Scan history (5 scans)

Commit	Scanned at	New	Resolved	Net
e8da748	2026-08-30 00:19 UTC	0	0	0
e8da748	2026-08-30 00:19 UTC	—	—	—
17848c2	2026-08-30 00:29 UTC	0	0	0
5068eb8	2026-08-30 01:13 UTC	0	0	0
df064f9 <	2026-08-30 01:39 UTC	0	0	0

Last scanned: df064f9 · 2026-08-30 01:39 UTC

[sf-jin-ku]

@codex review

[chatgpt-codex-connector]

Codex Review: Didn't find any major issues. Bravo.

Reviewed commit: 17848c2ac1

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".

[sf-jin-ku]

Manual dispatch verification (Codex 0.149.1)

SessionEnd dispatch was verified empirically, not just from upstream source. Probe: the plugin installed into a throwaway CODEX_HOME from a local marketplace, its hooks/hooks.json commands swapped for a logger, all discovered hooks trusted via hooks/list → config/batchWrite hooks.state (upstream's own recipe), then one real codex exec run.

discovered hooks: sessionStart=untrusted, sessionEnd=untrusted, userPromptSubmit=untrusted, stop=untrusted
=== FIRED ===
SessionStart      {"session_id":"01a05035-…","transcript_path":"…"}
UserPromptSubmit  {"session_id":"01a05035-…","turn_id":"…"}
Stop              {"session_id":"01a05035-…","turn_id":"…"}
SessionEnd        {"session_id":"01a05035-…","transcript_path":"…"}

SessionEnd fires — in codex exec, not only the TUI — and SessionEndRequest carries session_id and a non-optional cwd (codex-rs/hooks/src/events/session_end.rs:27), which is exactly what handleSessionEnd reads.

Upgrade caveat this turned up

Codex trusts hooks per hook, keyed cc@sendbird:hooks/hooks.json:<event>:<group>:<index> with a trusted_hash in [hooks.state]. The hash is computed per hook, and the index is per event, so adding SessionEnd does not invalidate the existing session_start / stop / user_prompt_submit entries. But the new session_end hook has no entry, so on upgrade it lands untrusted and silently inert until the user approves it in the TUI hooks browser. Worth calling out in release notes; I did not make setup self-trust its own hook, since that is plausibly the exact thing the trust system exists to prevent.

Test coverage, stated honestly

• Handler logic: unit-tested with synthetic hook stdin, mutation-checked.
• Registration: contract-tested against upstream HOOK_EVENT_NAMES.
• Real dispatch: not covered by the repo's e2e suite. E2E runs real codex exec, but a fresh CODEX_HOME has no [hooks.state] trust entries, so no plugin hook has ever actually fired there — those tests assert config.toml contents only. The verification above was manual.

A trust-aware dispatch e2e is feasible with the recipe above; it is new scope and costs a real model call per run, so it is not in this PR.

[sf-jin-ku]

Follow-up to the note above: the dispatch gap is now closed in the suite, not just manually.

tests/e2e/codex-skills-e2e.test.mjs gains a native hook dispatch case that installs the plugin, trusts its hooks through hooks/list → config/batchWrite hooks.state, runs one real codex exec against the existing mock provider, and asserts the workspace state directory survives without a current-session.json. Only setCurrentSession (SessionStart) creates that directory and only clearCurrentSession (SessionEnd) removes the marker, so both halves must have run; if hooks never dispatch, the directory is missing and the test fails. CLAUDE_COMPANION_SESSION_ID is cleared first, since a nested session id would make SessionStart skip the marker and fake a pass.

CI on codex-cli 0.151.0: 23 passed, 0 failed, 0 skipped — so this runs for real in CI rather than skipping out.

So the earlier "real dispatch is not covered by e2e" line is now stale: it is covered. No model tokens are spent — the suite's mock provider serves the turn.