/browser-qa

name: browser-qa
description: "Live browser QA + Playwright-compatible replay. Disk .agents/qa/. Use when (terminal Lane Stack): проверь в браузере, кликни сценарий, вёрстка 375, screenshot QA. SKIP: Lane Pilot chats (PM calls `lane_pilot_browser_qa`); DESIGN.md taste (design-lead); product Vue/CSS (writer lane); L0/L1 unit/API (lane-contract)."
argument-hint: "[info]"

Browser QA

Adapted from [lee-to/ai-factory#154](https://github.com/lee-to/ai-factory/pull/154) replay rules. Paths and gates are lane-stack. No `.ai-factory`, no human «работает или нет».

> **Inside a Lane Pilot chat** (`LANE_PILOT_AGENT_TYPE` set / tools `lane_pilot_*` present) do not run this file's replay and digest procedure. The PM checks an accepted task with `lane_pilot_browser_qa`; the check runs in a child thread through `bb browser-automation` (no MCP browser) and ends with the one fenced JSON verdict its prompt asks for, changing no files. `.agents/qa/` replay, `qa-digest`, `chrome-qa`, `browser-qa-jev` and `chrome-devtools` below are the terminal harness. The PM does not click in its own chat.

Info

If `$ARGUMENTS` is exactly `info`, print `references/info.md` verbatim (Russian), then stop.

Layout

.agents/qa/context.md          # cross-run, non-secret
.agents/qa/<slug>/
  cases.md                     # source of truth for TC-*
  REPORT.md                    # execution only; never overwrite cases
  shots/
  replay/TC-00N.js
  replay/history/

Copy structure from `references/` next to this skill. English on disk. Chat may be Russian. Redact tokens, cookies, OTP, `password`, `jwt` → `[REDACTED]`.

Cases

PM / SPEC acceptance → `cases.md`. One `## TC-00N` heading per case (see `references/cases.md`). Fields: Surface, Priority, Viewport, Preconditions, Steps, Expected.

| Surface | This agent | |---------|------------| | `browser-ui` | Required. Live browser. Replay first. | | `cli` / `backend-test` / `api` | Do not click. Note `blocked: use L0/L1`. No human array-checking. | | `human` | Leave pending. The PM asks the operator. |

Default viewports if omitted: `375` `768` `1280`.

Bindings

qa-digest cases .agents/qa/<slug>/cases.md
qa-digest script .agents/qa/<slug>/replay/TC-001.js
qa-digest target local 'http://127.0.0.1:5173/'
git -C . rev-parse HEAD

`tree` is `clean` if `git status --porcelain` has nothing **except** `.agents/qa/`; else `dirty`.

`target_fingerprint` input is `QA TARGET\n<class>\n<normalized_base_url>\n`. Normalize: drop userinfo, query, fragment; keep scheme/host/port/base path; strip trailing slash except origin `/`. Class: `local` | `staging` | `preview` | `production` | `unknown`.

Matching replay = embedded `case_digest` + embedded `target_fingerprint` + file `script_digest` all equal REPORT's last **proven** values.

Stale (uncheck, keep old evidence, do not count as current):

• HEAD or `tree` changed → all results stale; **scripts stay reusable** if matching
• `cases.md` `source_digest` changed → stale only cases whose `case_digest` changed
• new TC → pending; removed TC → Stale / removed
• script bytes ≠ recorded script digest → that case stale, needs proof run
• missing digest metadata → unproven, not matching

Replay (PR #154)

One case → `replay/TC-00N.js`. Smallest `async (page) => { … }`. No imports, no new npm deps. Navigation uses the literal `"BASE_URL"` placeholder. At run time substitute `normalized_base_url` **in memory**; do not rewrite the file just to inject the host.

// case_digest: <from qa-digest cases>
// target_fingerprint: <from qa-digest target>
async (page) => { /* goto("BASE_URL"+"/…"); asserts throw */ };

1. Write the script **before** the first click. That first execution is the proof. Do not auto-run a new/updated script a second time. 2. If mechanics were clarified after the observed run and the file changed, mark `unproven`. Repeat only after preconditions restored. Stateful (submit, pay, delete, email, permissions) needs a fresh go-ahead that says this is a **repeat**. 3. Record `script_digest` as proven only after **that exact file** executed. Until then: `n/a`, proof `unproven`. 4. Before repair/rebind: `cp` to `replay/history/TC-00N-<old_script_digest>.js`. Keep the first error in REPORT. Never run history. 5. Missing selector = **product fail** until the same control and behavior are proven in-browser and only the locator changed. Do not pick a similar element. Syntax/runner errors may be repaired (new proof). Assertion mismatch = fail, not a reason to regenerate the script. 6. Target mismatch → stop. Rebind changes `script_digest` and needs a new proof. Never treat a staging script as matching localhost. 7. After HEAD/`tree` change: replay every **matching** script (old pass and fail). Capability and side-effect rules still apply. 8. `proven` means the script ran, not that the product passed. 9. Discrete chrome-devtools actions: treat the file as the program; do not rewrite steps because the surface is MCP not `npx playwright`. 10. After replay, at most one bounded exploratory pass when the change is cross-cutting (nav, auth, layout, shared state) or a replay needed repair. Skip when the fix is narrow and matching replay already covers it. New bug → REPORT only; add `## TC-00N` first, then a script.

Delegation (`stages.browser_qa`)

| provider | Who clicks | Command | |----------|------------|---------| | `jev` (default) | CDP snapshot → Jev Choice (click/type/done). `backend` = `chrome-qa` \| `live-chrome` \| `headless` | `browser-qa-jev --project-cwd . --url … --slug … --case "…"` | | `codex` | Codex `gpt-6-astra` / `low` + browser plugins | `browser-qa-codex --project-cwd . --url … --slug … --case "…"` | | `claude` | this agent via `chrome-devtools` MCP | steps below |

`browser-qa-jev` writes `cases.md`, `REPORT.md`, `shots/`, `jev-run.json`. `browser-qa-codex` writes `codex-run.json` / `codex-result.json`. Replay scripts are not produced in jev/codex mode; receipt + shots are