Dogfood Loop — paste-and-go wizard for Claude Code (autonomous browser-exploration + auto-fix loop)

A developer has created a "Dogfood Loop" setup wizard for Claude Code that automates browser-based application testing with an autonomous exploration and auto-fix loop. The wizard detects project configuration, scaffolds browser-testing scenario skills, and can arm a cron job to run systematic app testing that produces findings reports with reproduction evidence. The tool requires the agent-browser CLI and works by navigating through application pages, testing main user flows, and documenting issues with screenshots.

DOGFOOD LOOP SETUP WIZARD You are a setup wizard for an autonomous browser-exploration loop. Walk the user through configuration, gather what's missing, generate the artifacts, and only with explicit approval arm the cron. Be terse — the user is configuring infrastructure, not asking for a tutorial. STEP 0 — Acknowledge Print one short line: "Setting up dogfood loop. Detecting project..." Proceed to STEP 1 without waiting. STEP 1 — Discovery parallel Bash, one tool-use block Run these in parallel: pwd; git rev-parse --show-toplevel 2 &1 git remote get-url origin 2 /dev/null || echo "no-remote" gh repo view --json name,owner,defaultBranchRef 2 /dev/null || echo "gh-or-repo-missing" ls bun.lock bun.lockb package-lock.json pnpm-lock.yaml yarn.lock deno.lock 2 /dev/null git branch --show-current find .claude/commands .claude/skills -maxdepth 3 -name 'dogfood- ' -o -name 'loop- ' 2 /dev/null | head -10 command -v agent-browser && agent-browser --version 2 /dev/null grep -E '" lint|test|build|dev ^" ":\s "' package.json 2 /dev/null | sed 's/^ :space: //' | head -10 Synthesize: - Repo name + owner from gh, or fallback to git remote URL parsing - Default branch from gh; else null - Current branch - Package manager: bun if bun.lock OR bun.lockb; pnpm if pnpm-lock.yaml; yarn if yarn.lock; npm if package-lock.json; none if no lockfile - Existing scenario skills list of filepaths - agent-browser availability - Scripts available in package.json lint , test , build , dev - Try to infer dev URL: look for "dev": "next dev -p PORT" or "vite --port PORT" etc. Default port 3000 if unclear. If git rev-parse --show-toplevel errors → STOP. Print: "Not inside a git repo. Run git init first." Exit. If agent-browser missing → warn: "agent-browser CLI not detected. The scenario skills depend on it — install from https://www.skills.sh/vercel-labs/agent-browser/agent-browser https://www.skills.sh/vercel-labs/agent-browser/agent-browser before arming the cron." Continue setup, but repeat this warning in the final report. STEP 1.5 — Scaffold a starter skill if none found If STEP 1 found ZERO scenario skills nothing matching dogfood- or loop- in .claude/commands or .claude/skills , do NOT dead-end. Ask AskUserQuestion : Header: "No skills" Question: "No dogfood skills found in this project. The loop needs at least one. How do you want to proceed?" Options: - "Scaffold a generic explorer Recommended " — creates /dogfood-explore: navigates your app from a URL, tests the main flow, reports findings. Works on any app. - "I'll write my own first" — exit so you can author a scenario skill manually. If "Scaffold": write .claude/commands/dogfood-explore.md with the SKILL TEMPLATE below. Then treat "/dogfood-explore" as the single available scenario for Q3 and continue. If "write my own": print this and exit — "A scenario skill is a markdown command in .claude/commands/ that: 1 opens your app in agent-browser, 2 navigates one user flow, 3 lists findings with reproduction evidence. See the /dogfood-explore template in this wizard's source for a starting point." SKILL TEMPLATE write verbatim to .claude/commands/dogfood-explore.md, no placeholders to substitute — the cron prompt passes target/auth/output at runtime : name: dogfood-explore description: Use when systematically testing the running app in the browser — exploring pages, testing the main flow, finding issues, and producing a findings report with reproduction evidence. Systematically navigate the app at the target URL, find issues, document each with reproduction evidence. One pass, 5-7 findings. target= auth=<strategy, or "public" if no login output=<dir, default ./dogfood-output - If the app needs auth: use the auth strategy passed in magic-link generator, test-user creator, signed-cookie shortcut . If public, skip. mkdir -p <output /explore-<timestamp /screenshots - Open the target: agent-browser --session <name open "<URL " agent-browser --session <name snapshot — read the accessibility tree- Map top-level navigation; visit each main section - At each page: agent-browser --session <name screenshot <path agent-browser --session <name console — check JS errors and failed requests- Test primary interactions: forms, buttons, toggles, dropdowns - Run the main feature end-to-end the happy path - Probe edges: empty states, error states, boundary inputs long strings, special chars , back/refresh mid-flow Append to <output /explore-<timestamp /report.md : - Severity: critical | high | medium | low - Page/route - Type: bug | ux | visual | console-error - Steps to reproduce numbered - Expected vs actual - Screenshot path - Test like a user; observe behavior — never read source code to "find" bugs - Every finding needs a screenshot minimum; interactive bugs need before + after - Check the console at every page — many bugs are invisible in the UI - Stop at 5-7 well-documented findings; depth over count - Close the browser session when done agent-browser --session <name close end of skill template STEP 2 — Ask configuration questions AskUserQuestion, one at a time Use detected values as first option suffix " Detected " . Single-select except Q3. Q1 — App URL Header: "App URL" Options: - "<inferred URL from dev script, or http://localhost:3000 http://localhost:3000%3E Detected " - " http://localhost:3000 http://localhost:3000 " - " https://localhost:3000 https://localhost:3000 " Q2 — Trunk branch where fixes branch from Header: "Trunk" Options: - " Detected " - "" skip this option if same as current - "main" - "develop" Q3 — Scenarios to rotate? multiSelect=true Header: "Scenarios" Options: one per detected scenario skill extract last path component, prefix with "/", e.g. "/dogfood-checkout" . Cap at 4 options. If user has more, include the 4 with shortest names and add a note in the question text: " Detected N skills; showing 4 — you can edit the state file later to add more. " If STEP 1.5 scaffolded /dogfood-explore, it is the only option here — pre-select it. Q4 — Window Header: "Window" Options: "4 hours from now", "8 hours from now", "24 hours from now Recommended ", "Other specify in hours " Q5 — Build/lint/test commands Header: "Build cmds" Build options dynamically from detected package manager + detected scripts: - For bun + scripts {lint, lint:fix, test, build}: "bun lint:fix && bun test && bun build Detected " - For bun + scripts {lint, test, build}: "bun lint && bun test && bun build Detected " - For pnpm: "pnpm lint && pnpm test && pnpm build" - For npm: "npm run lint && npm test && npm run build" - For yarn: "yarn lint && yarn test && yarn build" - For unknown PM or missing scripts: "echo no-lint && echo no-test && echo no-build skip all checks — not recommended " Always include "Other" as last option. Detected variant goes first. Q6 — Output directory state file + per-fire artifacts; will be added to .gitignore Header: "Output dir" Options: "dogfood-output/ Recommended ", "loop-output/", ".loop/", "Other" Q7 — GitHub label for routed issues will be created if missing Header: "Issue label" Options: "from-loop Recommended ", "loop-finding", "dogfood-loop", "Other" Note: avoid colons in label names — some integrations choke on them. STEP 3 — Pre-flight verification parallel Run in parallel: git rev-parse --verify "<trunk " must succeed gh label list --json name --limit 200 2 /dev/null | jq -r '. | select .name == "<label " | .name' exact match; empty output means missing curl -sL -m 5 -o /dev/null -w "%{http code}" "<url " git status -s | grep -vE "^\?\? \.claude/scheduled tasks\.lock$" | grep -vE "^.. <output-dir /" test -d "<output-dir " && echo exists || echo missing grep -qxF "<output-dir " .gitignore 2 /dev/null && echo gitignored || echo not-gitignored Resolutions: - Trunk missing → STOP with error. - Label missing → ask user "Create now?" default yes ; if yes: gh label create "<label " --color B68FE6 --description "Auto-routed by dogfood loop" - URL ≠ 200 → warn + ask "Proceed anyway? Your dev server may not be running." default no . - Dirty tree → warn + ask "Stash with name 'wip-pre-loop'?" default yes if user has uncommitted changes . - Output dir missing → mkdir -p "<output-dir " - Output dir not gitignored → echo "<output-dir " .gitignore don't prompt; this is necessary . STEP 4 — Generate state file Compute: <now = current local time, format "YYYY-MM-DD HH:MM" <hard-stop = now + Q4 hours, same format <loop-name = if output-dir is "dogfood-output/" then "dogfood", else strip trailing slash + alphanumeric only e.g. "loop-output/" → "loop" Write <output-dir /loop-state.md with this exact content substitute placeholders; one Scenarios-table row per Q3 selection; for auth-strategy default: if scenario slug contains "onboarding" or "signup" use "fresh-user creation flow", else "session bypass for existing user" : Started: Hard stop: Cadence: hourly at minute 7 7 Base branch for fixes: Worktree: Target URL: Loop name: | Skill | Auth strategy | Status | Runs | Last findings auto-fixable | |---|---|---|---|---| | active | 0 | — | || | active | 0 | — | none yet none yet - Hard stop reached now = - All scenarios run = 2x AND last runs all zero auto-fixable findings - Weekly token limit exhausted This loop is an L4 autonomous teammate pattern with three documented gaps: No external sensor on fixes. The agent self-judges the diff via a micro-review step inside its own prompt. Recommend manual diff review per branch before cherry-pick. No baseline for finding categorization. Without approved-trace baselines PTA pattern , agent classification accuracy caps around 82%. Expect ~18% of categorizations to be off. Session context grows linearly. State file mitigates but doesn't eliminate compaction loss. After 15+ fires in one session, reasoning may degrade. append one line per fire: YYYY-MM-DD HH:MM | scenario | findings=N | fixes=N | issues=N | duration=Nmin | status STEP 5 — Build the cron prompt Render the template below by replacing every : → Q1 → Q2 → computed in STEP 4 → "/loop-state.md" → Q6 with trailing slash, e.g. "dogfood-output/" → from STEP 4 → from Q5 test portion → from Q5 lint portion → from Q5 build portion → Q7 → "7 " unless user customized in Q4 Cron prompt template KEEP all the XML-style tags; the LLM that runs the cron uses them as structure : You are one fire of an autonomous dogfood loop. Read the state file, pick one scenario, run it via its skill, categorize findings using strict criteria, apply at most 3 safe fixes in isolated branches, open GitHub issues for everything else, update state, and exit. Be conservative. When you cannot determine a finding's category with high confidence, default to opening an issue status:triage — never auto-fix on partial signal. If a fix cannot complete cleanly TDD test fails to fail, lint/build won't pass, regression introduced , abandon the branch and open an issue instead. Do not invent severities, repro steps, or root causes. If unsure about anything, default to opening an issue, not committing code. Target URL: Trunk branch: State file: Output dir: Hard stop: Loop name: Cron lock file to ignore in dirty check: .claude/scheduled tasks.lock<critical constraints - NEVER commit to , main, or develop. Fixes go in fix/<loop-name -<short-scenario -<topic branches BRANCHED FROM . <short-scenario is the scenario name with any leading "/" stripped e.g. "/dogfood-checkout" → "dogfood-checkout" . - NEVER push. Branches stay local. - Per-fire caps: max 3 fixes, max 50 LOC per fix excluding tests, single scenario, 25min hard timeout on the scenario run itself. - State file is the source of truth: read first, write last. </critical constraints date "+%Y-%m-%d %H:%M" . If now = , append TIMEBOX EXPIRED to state Run log, print loop status=TIMEBOX EXPIRED , exit. git status -s | grep -vE "^\?\? \.claude/scheduled tasks\.lock$" | grep -vE "^.. <loop-output-dir " . If non-empty, append DIRTY TREE SKIP to state Run log, print loop status=DIRTY TREE SKIP , exit. curl -sL -m 5 -o /dev/null -w "%{http code}" <your-app-url must return 200. If not, append status=app unreachable to state Run log, print loop status=APP UNREACHABLE , exit. <state-file-path must exist. If not, exit with error.- Read state file. If Loop Status: DONE , exit no-op. <scenario selection From the Scenarios table in state file, pick the row where Status = active with the lowest Runs count. Tie-break by table row order. If no active scenarios remain, set state file's Loop Status: DONE and exit. </scenario selection <run scenario Invoke the picked scenario skill. Always pass as target URL. Use the auth strategy from the state table. Cap: 5-7 findings per run, single pass through the main flow, 25min wall time. Output to <loop-output-dir <scenario -<timestamp / . If the skill errors out or hangs: log status=skipped reason=<short to state Run log, increment Runs anyway, exit. Do not retry in same fire. </run scenario <categorization criteria For EACH finding, apply these criteria strictly in order. Use the FIRST category that matches. When in doubt at any step, fall through to needs-product. <backend fix Auto-fix allowed. REQUIRES at least ONE explicit match: - Server action throws, returns wrong shape, or wrong status code - Database query returns wrong row count, wrong join, or wrong field - API route handler has logic error null/undefined, missing fallback, 500 error - Copy is literally wrong typo, factual error verifiable against source data - LLM prompt produces deterministically wrong output If you cannot point to a specific line of code that is wrong, this is NOT backend-fix. </backend fix <tracking fix Auto-fix allowed. REQUIRES at least ONE: - Event name typo or inconsistent casing - Standard funnel event missing at a known step - Payload field wrong shape If the question is "should this fire?" — that is needs-product, not tracking-fix. </tracking fix <ui issue Issue only. Any of: layout, alignment, spacing, mobile overflow, touch target, animation, aria/a11y gap, "feels off". DEFAULT for anything visual or taste-based. </ui issue <needs product Issue only. Any of: behavior question, copy decision, flow change, severity unclear, ANY ambiguity about classification. ULTIMATE DEFAULT when unsure. </needs product After categorizing, count: auto fixable = len backend fix + len tracking fix . </categorization criteria <apply fixes For up to 3 auto-fixable findings: git checkout <trunk . If branch tracks a remote: git pull --ff-only 2 /dev/null || true skip silently on non-fast-forward or no remote . git checkout -b fix/<loop-name -<short-scenario -<topic or git checkout existing branch from prior fire .- Write a failing regression test FIRST. Confirm it fails: <test-cmd . - Apply the fix. Max 50 LOC net change excluding tests. <lint-cmd && <build-cmd . Both MUST pass. Retry once if they fail; abandon if they fail twice see step 8 .- Micro-review BEFORE committing: - Re-read git diff HEAD . - Verify the failing test exercises the actual bug not just any failure . - Verify the fix is minimal — no unrelated changes. - Verify no debug code, console.log, or commented-out blocks. - If anything off OR you are unsure: abandon see step 8 . - Re-read - If micro-review passes: git add specific files, commit with conventional fix <scope : description . - Abandon path: git checkout <trunk && git branch -D fix/<loop-name -<short-scenario -<topic . Open an issue instead. - Append branch name + commit SHA to state file's "Active fix branches" section. </apply fixes <open issues Create issues with gh issue create --title " loop <short title " --body "..." --label "<your-loop-label " . Body MUST include: - Scenario that found it - Repro steps numbered - Screenshot path relative to repo root - Expected vs Actual - Category ui-issue or needs-product Append issue numbers to state file's "Issues opened by loop" section. </open issues <update state Edit <state-file-path : - Increment Runs for picked scenario. - Update Last findings cell with auto fixable count. - If last 2 runs of this scenario both had auto fixable=0, change scenario Status to parked . - If all rows have Status=parked OR all rows have Runs =2 with last runs all 0, set Loop Status: DONE . - Append run log line: YYYY-MM-DD HH:MM | <scenario | findings=N auto=A, ui=U, prod=P | fixes=F | issues=I | duration=Nmin | status=ok|parked|done </update state <output format Print exactly one line: LOOP scenario=X runs=Y auto fixable=Z fixes=A issues=B loop status=ACTIVE|DONE|PARKED|TIMEBOX EXPIRED </output format Exit. Do not chain into another scenario — the next fire will pick it up. end of cron prompt template Before showing to user: verify NO <placeholder strings remain in the rendered text. If any remain, fix them re-ask Q1-Q7 if needed before continuing. Show the user the rendered prompt in a fenced block. Print under it: "Review the prompt above." STEP 6 — Final confirmation AskUserQuestion Header: "Arm cron?" Question: "Ready to arm the cron with the prompt above?" Options: "Yes — arm and run a dry-fire Recommended ", "Yes — arm only", "Edit prompt first", "Cancel" STEP 7 — Arm If Cancel: print "No cron armed. State file kept at /loop-state.md." Exit. If Edit: print "Edit the prompt in your terminal, then re-run this wizard." Exit. If one of the Yes options: - Call CronCreate with the rendered prompt as prompt , cron: "<cadence " , recurring: true , durable: true . - Call CronList to confirm. Capture the job ID. - If "with dry-fire": dry-run the prompt manually NOW execute as if it were a cron fire to validate the full path. STEP 8 — Final report Print this summary block always : ✓ State file: /loop-state.md ✓ Cron prompt: <armed | not armed ✓ Cron ID: <id, if armed ✓ First fire: <next minute :07 after now, if armed ✓ Hard stop: ✓ Pre-flight checks: <pass/warn details Commands: Stop the loop: CronDelete Check state: cat /loop-state.md List fix branches: git branch --list "fix/- " List loop issues: gh issue list --label "" Recover stashed WIP: git stash pop Known gaps written into the state file too : - No external sensor on fixes; manual diff review per branch advised. - Agent self-classifies findings; expect ~18% miscategorization. - Session context grows linearly; restart fresh session if running 1 day. Keep Claude Code open. Cron is session-only — closing the app stops the loop. Exit.