v4.9.0 ships -- per-edit test feedback now lives in Codex CLI

[mailnike]

Hi all,

Shipping a meaningful refresh today. v4.9.0 adds a PostToolUse hook so tailtest fires after every apply_patch instead of waiting until end of turn.

What's new in v4.9.0:

• PostToolUse hook fires per-edit, queues qualified source files, and surfaces them to Codex as mid-turn context. Matches how the Claude Code and Cursor variants behave.
• Stop hook stays as a turn-end safety net for files touched outside apply_patch (shell redirects, build steps).
• Shared lib/scanner.py so Stop and PostToolUse use the same mtime walker + patch parser.
• 400 tests now (was 380; +20 PostToolUse-specific tests).

Also in this session, v4.8.0 shipped earlier:

• Dropped the deprecated [features].codex_hooks = true flag from install docs. Codex 0.129.0+ has hooks GA, on by default.
• Added .codex-plugin/plugin.json + .agents/plugins/marketplace.json so you can register the plugin via codex plugin marketplace add avansaber/tailtest-codex.
• 5 skill descriptions polished to surface better in Codex's @-mention picker.

Honest caveat on the marketplace install: Codex CLI's plugin_hooks feature is still listed as "under development" in codex features list. That means plugin-bundled hooks don't auto-fire yet after marketplace add. You still need to run init.sh per project for hooks to actually wire up. Once upstream plugin_hooks stabilizes, init.sh retires and marketplace install becomes one-step.

Requirements:

• Codex CLI 0.129.0 or newer (hooks GA, on by default)
• macOS or Linux

Upgrade path:

cd ~/.codex/plugins/tailtest
git pull
# init.sh is idempotent; re-run if you want to refresh the per-project .codex/hooks.json
cd <your-project>
bash ~/.codex/plugins/tailtest/scripts/init.sh

Release notes (full):

• https://github.com/avansaber/tailtest-codex/releases/tag/v4.9.0
• https://github.com/avansaber/tailtest-codex/releases/tag/v4.8.0

Happy to answer questions on this thread.

[pramodavansaber]

Update from a few hours after this announcement: the same rule layer that v4.9.0 hooks into per-edit just caught a real bug in the wild.

While shaking out V14.9, I ran tailtest's adversarial mode against 8 fresh Python utility / parser libraries (boltons, xmltodict, python-dotenv, userpath, nest_asyncio, anyio, ciso8601, inflect). 24 minutes of compute, ~89 scenarios. One real bug surfaced:

jaraco/inflect -- engine().number_to_words(n) for n >= 10^36 leaks IndexError: list index out of range instead of the documented NumOutOfRangeError. Filed at jaraco/inflect#242.

Mentioning here because it's a concrete demo of what the R15 adversarial categories shipped with v4.7.0+ (and now hooked into per-edit by v4.9.0) actually do. Same R1-R15 rules across all four plugins; same 8 adversarial categories (boundary, format/injection, type confusion, concurrent state, time/locale, partial failures, resource exhaustion, off-by-one). The inflect find sits in the resource-exhaustion category: 'documented error for out-of-range inputs' as a probe.

Full write-up on the case-studies page: https://tailtest.com/case-studies