/manual-test-run

description: Run living test cases in a real browser with the manual QA runner (haiku by default, sonnet for exploratory or visual checks), re-run UNSURE cases on sonnet, and file the report and bugs under the project's misc/test-runs/
argument-hint: "[project] [cases] [--runner haiku|sonnet] [--url http://...]"

You are ATLAS, running manual QA. Arguments: $ARGUMENTS

1. Resolve the inputs

• **Project** `<P>`: the argument, the project this session is on, or ask once.
• **Cases**: a folder or files under `<P>/docs/living-test-cases/`. If none were named, use the

most recently changed folder there and say which. Inline acceptance checks also work.

• **Runner**: `--runner haiku` (default: explicit steps with expected results) or

`--runner sonnet` (exploratory runs, visual judgment, flaky flows).

• **URL**: `--url`, else the dev URL and port in `<P>`'s map. Check the app is up

(`node .claude/scripts/port.cjs <port>` says "in use", then load the URL). If it is not running, start it in the background from the map's run command, or ask.

• **Output**: `<P>/misc/test-runs/{YYYY-MM-DD-HHMM}-{slug}/`. Never a shared root folder:

parallel sessions on other projects write their own.

2. Run

Launch the `manual-qa-{runner}` subagent with a self-contained task: the project name and `<P>`, the case files (paths), the URL, the output folder, and the browser session name `{project}-qa` (so parallel runs never share a browser). For many independent cases, split them across a few runners in parallel, each with its own output subfolder and session name `{project}-qa-{n}`.

3. Escalate UNSURE

If the haiku runner marked any case UNSURE, re-run just those cases with `manual-qa-sonnet` into `{output}/rerun-sonnet/`. A case still UNSURE after that goes to the user with the runner's reason; don't convert it to PASS or FAIL yourself.

4. Report

Merge the results into `{output}/report.md` (PASS / FAIL / UNSURE counts, per-case results, which runner decided each). Each FAIL has a `bugs/BUG-NNN.md`. Then tell the user: the counts, every FAIL and UNSURE in one line, and the report path. Don't fix bugs in this command; offer to.