Files
AGENTS/.pi/gsd/references/checkpoints.md
2026-04-24 20:00:33 +02:00

31 KiB

Plans execute autonomously. Checkpoints formalize interaction points where human verification or decisions are needed.

Core principle: the agent automates everything with CLI/API. Checkpoints are for verification and decisions, not manual work.

Golden rules:

  1. If the agent can run it, the agent runs it - Never ask user to execute CLI commands, start servers, or run builds
  2. the agent sets up the verification environment - Start dev servers, seed databases, configure env vars
  3. User only does what requires human judgment - Visual checks, UX evaluation, "does this feel right?"
  4. Secrets come from user, automation comes from the agent - Ask for API keys, then the agent uses them via CLI
  5. Auto-mode bypasses verification/decision checkpoints - When workflow._auto_chain_active or workflow.auto_advance is true in config: human-verify auto-approves, decision auto-selects first option, human-action still stops (auth gates cannot be automated)

<checkpoint_types>

## checkpoint:human-verify (Most Common - 90%)

When: the agent completed automated work, human confirms it works correctly.

Use for:

  • Visual UI checks (layout, styling, responsiveness)
  • Interactive flows (click through wizard, test user flows)
  • Functional verification (feature works as expected)
  • Audio/video playback quality
  • Animation smoothness
  • Accessibility testing

Structure:

<task type="checkpoint:human-verify" gate="blocking">
  <what-built>[What the agent automated and deployed/built]</what-built>
  <how-to-verify>
    [Exact steps to test - URLs, commands, expected behavior]
  </how-to-verify>
  <resume-signal>[How to continue - "approved", "yes", or describe issues]</resume-signal>
</task>

Example: UI Component (shows key pattern: the agent starts server BEFORE checkpoint)

<task type="auto">
  <name>Build responsive dashboard layout</name>
  <files>src/components/Dashboard.tsx, src/app/dashboard/page.tsx</files>
  <action>Create dashboard with sidebar, header, and content area. Use Tailwind responsive classes for mobile.</action>
  <verify>npm run build succeeds, no TypeScript errors</verify>
  <done>Dashboard component builds without errors</done>
</task>

<task type="auto">
  <name>Start dev server for verification</name>
  <action>Run `npm run dev` in background, wait for "ready" message, capture port</action>
  <verify>fetch http://localhost:3000 returns 200</verify>
  <done>Dev server running at http://localhost:3000</done>
</task>

<task type="checkpoint:human-verify" gate="blocking">
  <what-built>Responsive dashboard layout - dev server running at http://localhost:3000</what-built>
  <how-to-verify>
    Visit http://localhost:3000/dashboard and verify:
    1. Desktop (>1024px): Sidebar left, content right, header top
    2. Tablet (768px): Sidebar collapses to hamburger menu
    3. Mobile (375px): Single column layout, bottom nav appears
    4. No layout shift or horizontal scroll at any size
  </how-to-verify>
  <resume-signal>Type "approved" or describe layout issues</resume-signal>
</task>

Example: Xcode Build

<task type="auto">
  <name>Build macOS app with Xcode</name>
  <files>App.xcodeproj, Sources/</files>
  <action>Run `xcodebuild -project App.xcodeproj -scheme App build`. Check for compilation errors in output.</action>
  <verify>Build output contains "BUILD SUCCEEDED", no errors</verify>
  <done>App builds successfully</done>
</task>

<task type="checkpoint:human-verify" gate="blocking">
  <what-built>Built macOS app at DerivedData/Build/Products/Debug/App.app</what-built>
  <how-to-verify>
    Open App.app and test:
    - App launches without crashes
    - Menu bar icon appears
    - Preferences window opens correctly
    - No visual glitches or layout issues
  </how-to-verify>
  <resume-signal>Type "approved" or describe issues</resume-signal>
</task>
## checkpoint:decision (9%)

When: Human must make choice that affects implementation direction.

Use for:

  • Technology selection (which auth provider, which database)
  • Architecture decisions (monorepo vs separate repos)
  • Design choices (color scheme, layout approach)
  • Feature prioritization (which variant to build)
  • Data model decisions (schema structure)

Structure:

<task type="checkpoint:decision" gate="blocking">
  <decision>[What's being decided]</decision>
  <context>[Why this decision matters]</context>
  <options>
    <option id="option-a">
      <name>[Option name]</name>
      <pros>[Benefits]</pros>
      <cons>[Tradeoffs]</cons>
    </option>
    <option id="option-b">
      <name>[Option name]</name>
      <pros>[Benefits]</pros>
      <cons>[Tradeoffs]</cons>
    </option>
  </options>
  <resume-signal>[How to indicate choice]</resume-signal>
</task>

Example: Auth Provider Selection

<task type="checkpoint:decision" gate="blocking">
  <decision>Select authentication provider</decision>
  <context>
    Need user authentication for the app. Three solid options with different tradeoffs.
  </context>
  <options>
    <option id="supabase">
      <name>Supabase Auth</name>
      <pros>Built-in with Supabase DB we're using, generous free tier, row-level security integration</pros>
      <cons>Less customizable UI, tied to Supabase ecosystem</cons>
    </option>
    <option id="clerk">
      <name>Clerk</name>
      <pros>Beautiful pre-built UI, best developer experience, excellent docs</pros>
      <cons>Paid after 10k MAU, vendor lock-in</cons>
    </option>
    <option id="nextauth">
      <name>NextAuth.js</name>
      <pros>Free, self-hosted, maximum control, widely adopted</pros>
      <cons>More setup work, you manage security updates, UI is DIY</cons>
    </option>
  </options>
  <resume-signal>Select: supabase, clerk, or nextauth</resume-signal>
</task>

Example: Database Selection

<task type="checkpoint:decision" gate="blocking">
  <decision>Select database for user data</decision>
  <context>
    App needs persistent storage for users, sessions, and user-generated content.
    Expected scale: 10k users, 1M records first year.
  </context>
  <options>
    <option id="supabase">
      <name>Supabase (Postgres)</name>
      <pros>Full SQL, generous free tier, built-in auth, real-time subscriptions</pros>
      <cons>Vendor lock-in for real-time features, less flexible than raw Postgres</cons>
    </option>
    <option id="planetscale">
      <name>PlanetScale (MySQL)</name>
      <pros>Serverless scaling, branching workflow, excellent DX</pros>
      <cons>MySQL not Postgres, no foreign keys in free tier</cons>
    </option>
    <option id="convex">
      <name>Convex</name>
      <pros>Real-time by default, TypeScript-native, automatic caching</pros>
      <cons>Newer platform, different mental model, less SQL flexibility</cons>
    </option>
  </options>
  <resume-signal>Select: supabase, planetscale, or convex</resume-signal>
</task>
## checkpoint:human-action (1% - Rare)

When: Action has NO CLI/API and requires human-only interaction, OR the agent hit an authentication gate during automation.

Use ONLY for:

  • Authentication gates - the agent tried CLI/API but needs credentials (this is NOT a failure)
  • Email verification links (clicking email)
  • SMS 2FA codes (phone verification)
  • Manual account approvals (platform requires human review)
  • Credit card 3D Secure flows (web-based payment authorization)
  • OAuth app approvals (web-based approval)

Do NOT use for pre-planned manual work:

  • Deploying (use CLI - auth gate if needed)
  • Creating webhooks/databases (use API/CLI - auth gate if needed)
  • Running builds/tests (use Bash tool)
  • Creating files (use Write tool)

Structure:

<task type="checkpoint:human-action" gate="blocking">
  <action>[What human must do - the agent already did everything automatable]</action>
  <instructions>
    [What the agent already automated]
    [The ONE thing requiring human action]
  </instructions>
  <verification>[What the agent can check afterward]</verification>
  <resume-signal>[How to continue]</resume-signal>
</task>

Example: Email Verification

<task type="auto">
  <name>Create SendGrid account via API</name>
  <action>Use SendGrid API to create subuser account with provided email. Request verification email.</action>
  <verify>API returns 201, account created</verify>
  <done>Account created, verification email sent</done>
</task>

<task type="checkpoint:human-action" gate="blocking">
  <action>Complete email verification for SendGrid account</action>
  <instructions>
    I created the account and requested verification email.
    Check your inbox for SendGrid verification link and click it.
  </instructions>
  <verification>SendGrid API key works: curl test succeeds</verification>
  <resume-signal>Type "done" when email verified</resume-signal>
</task>

Example: Authentication Gate (Dynamic Checkpoint)

<task type="auto">
  <name>Deploy to Vercel</name>
  <files>.vercel/, vercel.json</files>
  <action>Run `vercel --yes` to deploy</action>
  <verify>vercel ls shows deployment, fetch returns 200</verify>
</task>

<!-- If vercel returns "Error: Not authenticated", the agent creates checkpoint on the fly -->

<task type="checkpoint:human-action" gate="blocking">
  <action>Authenticate Vercel CLI so I can continue deployment</action>
  <instructions>
    I tried to deploy but got authentication error.
    Run: vercel login
    This will open your browser - complete the authentication flow.
  </instructions>
  <verification>vercel whoami returns your account email</verification>
  <resume-signal>Type "done" when authenticated</resume-signal>
</task>

<!-- After authentication, the agent retries the deployment -->

<task type="auto">
  <name>Retry Vercel deployment</name>
  <action>Run `vercel --yes` (now authenticated)</action>
  <verify>vercel ls shows deployment, fetch returns 200</verify>
</task>

Key distinction: Auth gates are created dynamically when the agent encounters auth errors. NOT pre-planned - the agent automates first, asks for credentials only when blocked. </checkpoint_types>

<execution_protocol>

When the agent encounters type="checkpoint:*":

  1. Stop immediately - do not proceed to next task
  2. Display checkpoint clearly using the format below
  3. Wait for user response - do not hallucinate completion
  4. Verify if possible - check files, run tests, whatever is specified
  5. Resume execution - continue to next task only after confirmation

For checkpoint:human-verify:

╔═══════════════════════════════════════════════════════╗
║  CHECKPOINT: Verification Required                    ║
╚═══════════════════════════════════════════════════════╝

Progress: 5/8 tasks complete
Task: Responsive dashboard layout

Built: Responsive dashboard at /dashboard

How to verify:
  1. Visit: http://localhost:3000/dashboard
  2. Desktop (>1024px): Sidebar visible, content fills remaining space
  3. Tablet (768px): Sidebar collapses to icons
  4. Mobile (375px): Sidebar hidden, hamburger menu appears

────────────────────────────────────────────────────────
→ YOUR ACTION: Type "approved" or describe issues
────────────────────────────────────────────────────────

For checkpoint:decision:

╔═══════════════════════════════════════════════════════╗
║  CHECKPOINT: Decision Required                        ║
╚═══════════════════════════════════════════════════════╝

Progress: 2/6 tasks complete
Task: Select authentication provider

Decision: Which auth provider should we use?

Context: Need user authentication. Three options with different tradeoffs.

Options:
  1. supabase - Built-in with our DB, free tier
     Pros: Row-level security integration, generous free tier
     Cons: Less customizable UI, ecosystem lock-in

  2. clerk - Best DX, paid after 10k users
     Pros: Beautiful pre-built UI, excellent documentation
     Cons: Vendor lock-in, pricing at scale

  3. nextauth - Self-hosted, maximum control
     Pros: Free, no vendor lock-in, widely adopted
     Cons: More setup work, DIY security updates

────────────────────────────────────────────────────────
→ YOUR ACTION: Select supabase, clerk, or nextauth
────────────────────────────────────────────────────────

For checkpoint:human-action:

╔═══════════════════════════════════════════════════════╗
║  CHECKPOINT: Action Required                          ║
╚═══════════════════════════════════════════════════════╝

Progress: 3/8 tasks complete
Task: Deploy to Vercel

Attempted: vercel --yes
Error: Not authenticated. Please run 'vercel login'

What you need to do:
  1. Run: vercel login
  2. Complete browser authentication when it opens
  3. Return here when done

I'll verify: vercel whoami returns your account

────────────────────────────────────────────────────────
→ YOUR ACTION: Type "done" when authenticated
────────────────────────────────────────────────────────

</execution_protocol>

<authentication_gates>

Auth gate = the agent tried CLI/API, got auth error. Not a failure - a gate requiring human input to unblock.

Pattern: the agent tries automation → auth error → creates checkpoint:human-action → user authenticates → the agent retries → continues

Gate protocol:

  1. Recognize it's not a failure - missing auth is expected
  2. Stop current task - don't retry repeatedly
  3. Create checkpoint:human-action dynamically
  4. Provide exact authentication steps
  5. Verify authentication works
  6. Retry the original task
  7. Continue normally

Key distinction:

  • Pre-planned checkpoint: "I need you to do X" (wrong - the agent should automate)
  • Auth gate: "I tried to automate X but need credentials" (correct - unblocks automation)

</authentication_gates>

<automation_reference>

The rule: If it has CLI/API, the agent does it. Never ask human to perform automatable work.

Service CLI Reference

Service CLI/API Key Commands Auth Gate
Vercel vercel --yes, env add, --prod, ls vercel login
Railway railway init, up, variables set railway login
Fly fly launch, deploy, secrets set fly auth login
Stripe stripe + API listen, trigger, API calls API key in .env
Supabase supabase init, link, db push, gen types supabase login
Upstash upstash redis create, redis get upstash auth login
PlanetScale pscale database create, branch create pscale auth login
GitHub gh repo create, pr create, secret set gh auth login
Node npm/pnpm install, run build, test, run dev N/A
Xcode xcodebuild -project, -scheme, build, test N/A
Convex npx convex dev, deploy, env set, env get npx convex login

Environment Variable Automation

Env files: Use Write/Edit tools. Never ask human to create .env manually.

Dashboard env vars via CLI:

Platform CLI Command Example
Convex npx convex env set npx convex env set OPENAI_API_KEY sk-...
Vercel vercel env add vercel env add STRIPE_KEY production
Railway railway variables set railway variables set API_KEY=value
Fly fly secrets set fly secrets set DATABASE_URL=...
Supabase supabase secrets set supabase secrets set MY_SECRET=value

Secret collection pattern:

<!-- WRONG: Asking user to add env vars in dashboard -->
<task type="checkpoint:human-action">
  <action>Add OPENAI_API_KEY to Convex dashboard</action>
  <instructions>Go to dashboard.convex.dev → Settings → Environment Variables → Add</instructions>
</task>

<!-- RIGHT: the agent asks for value, then adds via CLI -->
<task type="checkpoint:human-action">
  <action>Provide your OpenAI API key</action>
  <instructions>
    I need your OpenAI API key for Convex backend.
    Get it from: https://platform.openai.com/api-keys
    Paste the key (starts with sk-)
  </instructions>
  <verification>I'll add it via `npx convex env set` and verify</verification>
  <resume-signal>Paste your API key</resume-signal>
</task>

<task type="auto">
  <name>Configure OpenAI key in Convex</name>
  <action>Run `npx convex env set OPENAI_API_KEY {user-provided-key}`</action>
  <verify>`npx convex env get OPENAI_API_KEY` returns the key (masked)</verify>
</task>

Dev Server Automation

Framework Start Command Ready Signal Default URL
Next.js npm run dev "Ready in" or "started server" http://localhost:3000
Vite npm run dev "ready in" http://localhost:5173
Convex npx convex dev "Convex functions ready" N/A (backend only)
Express npm start "listening on port" http://localhost:3000
Django python manage.py runserver "Starting development server" http://localhost:8000

Server lifecycle:

# Run in background, capture PID
npm run dev &
DEV_SERVER_PID=$!

# Wait for ready (max 30s) - uses fetch() for cross-platform compatibility
timeout 30 bash -c 'until node -e "fetch(\"http://localhost:3000\").then(r=>{process.exit(r.ok?0:1)}).catch(()=>process.exit(1))" 2>/dev/null; do sleep 1; done'

Port conflicts: Kill stale process (lsof -ti:3000 | xargs kill) or use alternate port (--port 3001).

Server stays running through checkpoints. Only kill when plan complete, switching to production, or port needed for different service.

CLI Installation Handling

CLI Auto-install? Command
npm/pnpm/yarn No - ask user User chooses package manager
vercel Yes npm i -g vercel
gh (GitHub) Yes brew install gh (macOS) or apt install gh (Linux)
stripe Yes npm i -g stripe
supabase Yes npm i -g supabase
convex No - use npx npx convex (no install needed)
fly Yes brew install flyctl or curl installer
railway Yes npm i -g @railway/cli

Protocol: Try command → "command not found" → auto-installable? → yes: install silently, retry → no: checkpoint asking user to install.

Pre-Checkpoint Automation Failures

Failure Response
Server won't start Check error, fix issue, retry (don't proceed to checkpoint)
Port in use Kill stale process or use alternate port
Missing dependency Run npm install, retry
Build error Fix the error first (bug, not checkpoint issue)
Auth error Create auth gate checkpoint
Network timeout Retry with backoff, then checkpoint if persistent

Never present a checkpoint with broken verification environment. If the local server isn't responding, don't ask user to "visit localhost:3000".

Cross-platform note: Use node -e "fetch('http://localhost:3000').then(r=>console.log(r.status))" instead of curl for health checks. curl is broken on Windows MSYS/Git Bash due to SSL/path mangling issues.

<!-- WRONG: Checkpoint with broken environment -->
<task type="checkpoint:human-verify">
  <what-built>Dashboard (server failed to start)</what-built>
  <how-to-verify>Visit http://localhost:3000...</how-to-verify>
</task>

<!-- RIGHT: Fix first, then checkpoint -->
<task type="auto">
  <name>Fix server startup issue</name>
  <action>Investigate error, fix root cause, restart server</action>
  <verify>fetch http://localhost:3000 returns 200</verify>
</task>

<task type="checkpoint:human-verify">
  <what-built>Dashboard - server running at http://localhost:3000</what-built>
  <how-to-verify>Visit http://localhost:3000/dashboard...</how-to-verify>
</task>

Automatable Quick Reference

Action Automatable? the agent does it?
Deploy to Vercel Yes (vercel) YES
Create Stripe webhook Yes (API) YES
Write .env file Yes (Write tool) YES
Create Upstash DB Yes (upstash) YES
Run tests Yes (npm test) YES
Start dev server Yes (npm run dev) YES
Add env vars to Convex Yes (npx convex env set) YES
Add env vars to Vercel Yes (vercel env add) YES
Seed database Yes (CLI/API) YES
Click email verification link No NO
Enter credit card with 3DS No NO
Complete OAuth in browser No NO
Visually verify UI looks correct No NO
Test interactive user flows No NO

</automation_reference>

<writing_guidelines>

DO:

  • Automate everything with CLI/API before checkpoint
  • Be specific: "Visit https://myapp.vercel.app" not "check deployment"
  • Number verification steps
  • State expected outcomes: "You should see X"
  • Provide context: why this checkpoint exists

DON'T:

  • Ask human to do work the agent can automate
  • Assume knowledge: "Configure the usual settings"
  • Skip steps: "Set up database" (too vague)
  • Mix multiple verifications in one checkpoint

Placement:

  • After automation completes - not before the agent does the work
  • After UI buildout - before declaring phase complete
  • Before dependent work - decisions before implementation
  • At integration points - after configuring external services

Bad placement: Before automation | Too frequent | Too late (dependent tasks already needed the result) </writing_guidelines>

Example 1: Database Setup (No Checkpoint Needed)

<task type="auto">
  <name>Create Upstash Redis database</name>
  <files>.env</files>
  <action>
    1. Run `upstash redis create myapp-cache --region us-east-1`
    2. Capture connection URL from output
    3. Write to .env: UPSTASH_REDIS_URL={url}
    4. Verify connection with test command
  </action>
  <verify>
    - upstash redis list shows database
    - .env contains UPSTASH_REDIS_URL
    - Test connection succeeds
  </verify>
  <done>Redis database created and configured</done>
</task>

<!-- NO CHECKPOINT NEEDED - the agent automated everything and verified programmatically -->

Example 2: Full Auth Flow (Single checkpoint at end)

<task type="auto">
  <name>Create user schema</name>
  <files>src/db/schema.ts</files>
  <action>Define User, Session, Account tables with Drizzle ORM</action>
  <verify>npm run db:generate succeeds</verify>
</task>

<task type="auto">
  <name>Create auth API routes</name>
  <files>src/app/api/auth/[...nextauth]/route.ts</files>
  <action>Set up NextAuth with GitHub provider, JWT strategy</action>
  <verify>TypeScript compiles, no errors</verify>
</task>

<task type="auto">
  <name>Create login UI</name>
  <files>src/app/login/page.tsx, src/components/LoginButton.tsx</files>
  <action>Create login page with GitHub OAuth button</action>
  <verify>npm run build succeeds</verify>
</task>

<task type="auto">
  <name>Start dev server for auth testing</name>
  <action>Run `npm run dev` in background, wait for ready signal</action>
  <verify>fetch http://localhost:3000 returns 200</verify>
  <done>Dev server running at http://localhost:3000</done>
</task>

<!-- ONE checkpoint at end verifies the complete flow -->
<task type="checkpoint:human-verify" gate="blocking">
  <what-built>Complete authentication flow - dev server running at http://localhost:3000</what-built>
  <how-to-verify>
    1. Visit: http://localhost:3000/login
    2. Click "Sign in with GitHub"
    3. Complete GitHub OAuth flow
    4. Verify: Redirected to /dashboard, user name displayed
    5. Refresh page: Session persists
    6. Click logout: Session cleared
  </how-to-verify>
  <resume-signal>Type "approved" or describe issues</resume-signal>
</task>

<anti_patterns>

BAD: Asking user to start dev server

<task type="checkpoint:human-verify" gate="blocking">
  <what-built>Dashboard component</what-built>
  <how-to-verify>
    1. Run: npm run dev
    2. Visit: http://localhost:3000/dashboard
    3. Check layout is correct
  </how-to-verify>
</task>

Why bad: the agent can run npm run dev. User should only visit URLs, not execute commands.

GOOD: the agent starts server, user visits

<task type="auto">
  <name>Start dev server</name>
  <action>Run `npm run dev` in background</action>
  <verify>fetch http://localhost:3000 returns 200</verify>
</task>

<task type="checkpoint:human-verify" gate="blocking">
  <what-built>Dashboard at http://localhost:3000/dashboard (server running)</what-built>
  <how-to-verify>
    Visit http://localhost:3000/dashboard and verify:
    1. Layout matches design
    2. No console errors
  </how-to-verify>
</task>

BAD: Asking human to deploy / GOOD: the agent automates

<!-- BAD: Asking user to deploy via dashboard -->
<task type="checkpoint:human-action" gate="blocking">
  <action>Deploy to Vercel</action>
  <instructions>Visit vercel.com/new → Import repo → Click Deploy → Copy URL</instructions>
</task>

<!-- GOOD: the agent deploys, user verifies -->
<task type="auto">
  <name>Deploy to Vercel</name>
  <action>Run `vercel --yes`. Capture URL.</action>
  <verify>vercel ls shows deployment, fetch returns 200</verify>
</task>

<task type="checkpoint:human-verify">
  <what-built>Deployed to {url}</what-built>
  <how-to-verify>Visit {url}, check homepage loads</how-to-verify>
  <resume-signal>Type "approved"</resume-signal>
</task>

BAD: Too many checkpoints / GOOD: Single checkpoint

<!-- BAD: Checkpoint after every task -->
<task type="auto">Create schema</task>
<task type="checkpoint:human-verify">Check schema</task>
<task type="auto">Create API route</task>
<task type="checkpoint:human-verify">Check API</task>
<task type="auto">Create UI form</task>
<task type="checkpoint:human-verify">Check form</task>

<!-- GOOD: One checkpoint at end -->
<task type="auto">Create schema</task>
<task type="auto">Create API route</task>
<task type="auto">Create UI form</task>

<task type="checkpoint:human-verify">
  <what-built>Complete auth flow (schema + API + UI)</what-built>
  <how-to-verify>Test full flow: register, login, access protected page</how-to-verify>
  <resume-signal>Type "approved"</resume-signal>
</task>

BAD: Vague verification / GOOD: Specific steps

<!-- BAD -->
<task type="checkpoint:human-verify">
  <what-built>Dashboard</what-built>
  <how-to-verify>Check it works</how-to-verify>
</task>

<!-- GOOD -->
<task type="checkpoint:human-verify">
  <what-built>Responsive dashboard - server running at http://localhost:3000</what-built>
  <how-to-verify>
    Visit http://localhost:3000/dashboard and verify:
    1. Desktop (>1024px): Sidebar visible, content area fills remaining space
    2. Tablet (768px): Sidebar collapses to icons
    3. Mobile (375px): Sidebar hidden, hamburger menu in header
    4. No horizontal scroll at any size
  </how-to-verify>
  <resume-signal>Type "approved" or describe layout issues</resume-signal>
</task>

BAD: Asking user to run CLI commands

<task type="checkpoint:human-action">
  <action>Run database migrations</action>
  <instructions>Run: npx prisma migrate deploy && npx prisma db seed</instructions>
</task>

Why bad: the agent can run these commands. User should never execute CLI commands.

BAD: Asking user to copy values between services

<task type="checkpoint:human-action">
  <action>Configure webhook URL in Stripe</action>
  <instructions>Copy deployment URL → Stripe Dashboard → Webhooks → Add endpoint → Copy secret → Add to .env</instructions>
</task>

Why bad: Stripe has an API. the agent should create the webhook via API and write to .env directly.

</anti_patterns>

Checkpoints formalize human-in-the-loop points for verification and decisions, not manual work.

The golden rule: If the agent CAN automate it, the agent MUST automate it.

Checkpoint priority:

  1. checkpoint:human-verify (90%) - the agent automated everything, human confirms visual/functional correctness
  2. checkpoint:decision (9%) - Human makes architectural/technology choices
  3. checkpoint:human-action (1%) - Truly unavoidable manual steps with no API/CLI

When NOT to use checkpoints:

  • Things the agent can verify programmatically (tests, builds)
  • File operations (the agent can read files)
  • Code correctness (tests and static analysis)
  • Anything automatable via CLI/API