Verify Omnigent
Use this skill to see a behavior happen in the real app and to prove a change,
not to reason about it from code. It has two parts:
- An isolated instance.
scripts/verify-env wraps
python -m dev.repro_env: a server, runner, and mock model server on private
ports, with their own config, data, Claude, and Codex directories. It never
touches ~/.omnigent, a running host daemon, or another developer server.
- A feature map. feature map lists each user-facing
feature's entry points, the tests that drive them, and the traps. A fix is
verified only when every entry point listed for its feature has proof.
Run all commands from the repository root. Put scripts/ on your path or call
feature-map/skills/verify-omnigent/scripts/verify-env directly.
Launch
-
Install dependencies, Chromium, and build the web UI for the checkout:
uv sync --frozen --extra all --group test
uv run --no-sync playwright install --with-deps chromium
pnpm install --frozen-lockfile --filter web && pnpm --filter web run build
-
Start an instance and load its paths:
verify-env start # waits until the runner is online, about 10 seconds
eval "$(verify-env paths)"
start writes VERIFY_ROOT to .omnigent/verify/current, so later shells
find the same instance. The instance stops itself after its lease (default
90 minutes; --lease SECONDS, 60 to 21600).
Ready means the server answers, the runner reports online, and the mock model
server answers. start fails with the environment's error and log path
otherwise.
Ready does not check browser launch. Use the Chromium version installed by
this checkout's Playwright package; if the environment provides it through
PLAYWRIGHT_BROWSERS_PATH, keep that path available to the test process.
--ui-skip-build reuses the existing bundle, so rebuild after frontend edits.
Native-terminal journeys also need the relevant CLI and terminal prerequisites
(such as tmux); a browser installation alone does not provide them.
In CI, the repro workflow already runs this environment. Use
python -m dev.repro_env exec -- ... there instead of starting another.
Doctor
Run verify-env doctor before the first drive, after any failed drive, and
whenever something looks off. It checks that the instance is ready, that its
supervisor is alive, that the runner and model server answer, and it warns when
the checkout has moved since launch. A warning about a moved checkout means the
instance runs old code: stop it and start a new one.
Drive
-
Open the matching file in feature map and list every
entry point for the behavior in question.
-
For each entry point, run the named test through the instance, recording on:
verify-env run -- python -m pytest <tests/...py::test_name> \
--ui-skip-build --video=on --screenshot=on \
--output="$VERIFY_EVIDENCE/<feature>"
Tests under tests/browser_ui/ need no instance:
uv run pytest <test> --browser-ui-skip-build --video=on --output=....
Tests the feature file marks "own environment" also run with plain
uv run pytest. Server/transport/component tests use plain pytest without
browser recording or build flags.
-
For an entry point with no test, hand-drive it with Playwright against
OMNIGENT_REPRO_SERVER_URL inside verify-env run -- python <script>, and
script model replies with the mock helpers named in the feature map README.
-
To reproduce a bug, run the same journey on the unfixed code first and keep
that evidence; then run it again on the fix.
A test that fails before reaching the user action has a setup failure, not a
reproduction. Check its browser error and server/runner logs. When a test starts
its own server or runner from inside an agent, check whether it inherited the
parent's OMNIGENT_RUNNER_*, RUNNER_SERVER_URL, or
OMNIGENT_PROCESS_LOG_FILE. Use an isolated child environment for that test;
do not change the controlling agent's environment. Record setup failures,
skipped variants, and unavailable credentials separately from behavior results.
Evidence
Everything under $VERIFY_EVIDENCE is the proof, one directory per feature.
Instance logs, the database, and the mock model's recorded requests stay in
$VERIFY_ENV. For each claim, record the checkout commit, installed harness versions,
feature file, entry point ID, command, and resulting artifact. Keep source
review, component tests, real process checks, browser drives, and live provider
checks distinct. A map contract pass verifies references and structure; it is
not evidence that the journeys ran. The proof standards are in the
feature map README. Mock runs prove
Omnigent's integration with Claude and Codex, not a live vendor model.
Cleanup
verify-env stop stops only this instance's supervisor, which stops its own
server, runner, and model processes and saves the model's request log. It never
deletes $VERIFY_ROOT, so the evidence survives; remove old roots under
.omnigent/verify/ or /tmp/verify-omnigent.* yourself once the evidence is
attached. Never kill Omnigent processes by name: a developer's own server and
host daemon may be running on the same machine.
Helpers
scripts/verify-env is the only helper. Its subcommands are start [--lease SECONDS], doctor, run -- COMMAND..., paths, and stop; run it with no
arguments for usage.
On macOS the helper places the instance under /tmp/verify-omnigent.* when the
checkout path is long. macOS limits Unix socket paths to 104 bytes, and the
environment's socket relay only works around long paths on Linux.
Keeping the map honest
When a reviewer says a change missed a surface, add that surface to the feature
file in the same change. The checks and the weekly upkeep job are described in
Keeping the map current.