ralph-dogfooding（拉尔夫狗食）

Ralph Dogfooding Loop

Autonomous loop that explores the app with Playwright MCP, captures evidence (screenshots, repro steps), dedupes against Linear, creates/updates issues with reproducible details, and appends each run to a progress log. Uses only Playwright MCP for the browser (no agent-browser); the loop runs without human approval.

Architecture

0. If no AGENTS.md → create-agents-md skill, then stop
1. Read AGENTS.md + .ralph/lessons.md + .ralph/dogfood-progress
2. Pull open issues from Linear (dedupe set)
3. Explore core routes via Playwright MCP (navigate, snapshot, screenshot)
4. Document each issue with evidence (screenshot sequence + report)
5. Create/update Linear issues with evidence
6. Append iteration summary to .ralph/dogfood-progress

Workflow (Playwright MCP + Linear)

1. Initialize

• Target URL and artifact dir come from the Ralph script (.ralph/dogfood-artifacts/iteration-N/; .ralph/ is added to .gitignore by the script).
• Create {ARTIFACTS}/screenshots, copy templates/dogfood-report-template.md to {ARTIFACTS}/report.md, fill header (Date, App URL, Session, Scope).
• Use Playwright MCP: browser_navigate to target URL, then browser_snapshot to understand the page. Lock the browser tab before interactions; unlock when done.

2. Authenticate (if needed)

If the app requires login: use browser_snapshot to get element refs, then browser_fill / browser_click to submit credentials. Wait for navigation/load between steps. For OTP/codes, ask the user once, then enter via browser.

3. Orient

• browser_snapshot (interactive) to see structure and refs.
• browser_take_screenshot (full page or viewport) for the initial state; save to {ARTIFACTS}/screenshots/initial.png.
• Identify main nav and plan which routes to hit (e.g. /, /login, /dashboard, /settings).

4. Explore

Read references/issue-taxonomy.md at session start — severity levels, categories, and the exploration checklist.

• Work through the app systematically: main nav → each section → interactive elements (buttons, forms, links).
• At each page: browser_snapshot, browser_take_screenshot for {page-name}.png, and check console/errors if the MCP exposes them.
• Test forms (submit, validation), navigation (back, deep links), empty/loading/error states.
• Spend more time on core flows; go deeper where you find clusters of issues.

5. Document issues (repro-first)

Explore and document in one pass. When you find an issue, document it immediately, then continue.

Evidence by issue type:

• Interactive/behavioral (needs steps to reproduce): Take a screenshot before the action, perform the action (click, fill, etc.), take a screenshot after. Save as issue-{NNN}-step-1.png, issue-{NNN}-step-2.png, issue-{NNN}-result.png. Write numbered repro steps in the report; each step references its screenshot. Playwright MCP may not offer video recording — use screenshot sequences as the standard.
• Static (typos, visual glitches on load): One annotated/view screenshot is enough. In the report set Repro Video to N/A.

For every issue:

• Append to the report immediately (use the template block: Severity, Category, URL, Description, Repro Steps with screenshot refs).
• Increment issue counter (ISSUE-001, ISSUE-002, …).
• Save screenshots under {ARTIFACTS}/screenshots/ with consistent names.

6. Linear + progress

• Use the linear-cli skill for all Linear operations (list issues, create/update issues, add comments). For issue descriptions and comment bodies with markdown, use --description-file and --body-file per the skill. Include artifact paths in the issue body or comments. Every issue: URL, environment, expected vs actual, deterministic repro steps.
• Dedupe: check open Linear issues (title + route + expected/actual) so you don’t duplicate.
• Append one entry to .ralph/dogfood-progress.md: iteration, routes tested, issue IDs, artifact paths. See references/progress-format.md.

7. Wrap up

• Update the report summary counts to match the issues.
• If no new findings and no issue updates this iteration, output <promise>NO_NEW_FINDINGS</promise> so the script can exit.

Reference guide

Pre-flight checklist

1. Playwright MCP in Codex config (~/.codex/config.toml; install script adds it).
2. linear-cli skill loaded (e.g. npx skills add https://github.com/schpet/linear-cli --skill linear-cli); linear on PATH.
3. Project root path and target URL (from Ralph script).
4. Linear team key and artifact dir.
5. Dedupe policy (title + route + expected/actual).
6. Stop criteria: NO_NEW_FINDINGS sentinel.

Setup

• Script: scripts/ralph-dogfooding.sh — run with --project, --iterations, optional --url.
• Progress log: .ralph/dogfood-progress.md (script creates .ralph/ and adds it to .gitignore). See progress-format.
• Lessons: Maintain .ralph/lessons.md for recurring failure patterns and tool constraints.

Evidence policy

• At least one screenshot per issue; for interactive bugs, a step-by-step screenshot sequence.
• Attach artifacts to Linear issues using the linear-cli skill (include artifact paths in description/comment; use CLI attachment commands if available).
• Every issue: exact URL, environment, expected vs actual, deterministic repro steps. Do not block on video; screenshot sequences are the standard.

Guidance

• Repro first. Match evidence to issue type: interactive → step screenshots; static → single screenshot.
• Append to the report as you go. Do not batch issues for the end.
• Never read the target app’s source code. Test as a user; all findings from the browser.
• Check the console when the MCP allows it; many issues show only as JS or network errors.
• Test like a user. Common workflows, realistic data, systematic navigation.
• Do not use agent-browser or agent-use. This skill uses only Playwright MCP and the linear-cli skill for Linear.