Verified Agent Skill record

codex-qa

QA the omo Codex Light edition (lazycodex / packages/omo-codex) in strict isolation using an isolated CODEX_HOME and a local mock model so only the plugin is exercised and the user's real ~/.codex is never touched. The first-party method drives the real codex app-server against the isolated home plus the mock, and proves a plugin hook fired by asserting hook/started and hook/completed notifications. It also covers isolated install verification, per-component hook probes, a tmux TUI smoke, and runtime log observation.

Tasks

Run scripts/app-server-drive.sh --plugin to prove a plugin hook fires in a live Codex turn, asserting hook/started and hook/completed notifications for sessionStart and userPromptSubmit
Run scripts/install-verify.sh --self-test to verify the local omo build installs into an isolated CODEX_HOME with config.toml enabling omo@sisyphuslabs and real ~/.codex unchanged
Run scripts/hook-unit-probe.sh --self-test to deterministically test a single component's hook logic, such as ultrawork injecting <ultrawork-mode> on an ulw UserPromptSubmit event
Run scripts/tui-smoke.sh --self-test to smoke-test the real codex TUI under tmux, confirming it boots, renders, and survives without early exit

Inputs

CODEX_HOME environment variable pointing to an isolated sandbox directory
Local mock model provider serving Responses SSE (started by cqa_start_mock)
Local omo Codex Light edition build under packages/omo-codex

Outputs

Captured JSON or pane output written under .omo/evidence/<YYYYMMDD>-<slug>/
Assertion results for hook/started and hook/completed notifications
Shasum comparison of ~/.codex/config.toml before and after proving the real home stayed clean

Limitations and checks

The interactive codex is a shell function that injects --profile quotio; bash scripts bypass it and get the real binary, so never rely on the interactive alias
There is no /debugging command in Codex; runtime observation is limited to the app-server notification stream, RUST_LOG, logs SQLite under CODEX_HOME, the TUI's /debug-config, and codex debug subcommands
tmux capture-pane degrades color and CJK width, so TUI visual QA should instead use the xterm.js web terminal via script/qa/web-terminal-visual-qa.mjs
Confirm bash scripts/lib/common.sh --self-check passes, verifying deps are present and the isolated CODEX_HOME is created inside a sandbox and auto-removed on exit
Shasum ~/.codex/config.toml before and after every QA run and assert it is unchanged
Verify that app-server-drive.sh --plugin produces hook/completed assertions for sessionStart and userPromptSubmit
Ensure captured evidence is written under .omo/evidence/<YYYYMMDD>-<slug>/, as no evidence file means the QA did not happen
Confirm the mock model serves the Responses SSE and no real model API call was made