Dev full auto

Dev Full Auto - Autonomous Development

Spec in. Production-ready software out. I make decisions, you review results.

Philosophy: The Ralph Loop

"The best code review is reviewing working software, not plans."

This skill implements the Ralph Loop paradigm - continuous iteration until machine-verifiable completion, not subjective "I think I'm done."

Why Ralph Loop?

Traditional AI assistance fails because:

• Premature exit: AI stops when it feels done, not when it is done
• Context breaks: Restarting loses all progress
• Subjective completion: "Looks good to me" isn't verifiable

Ralph Loop solves this:

• External state as memory: Git commits, test results, file changes persist across iterations
• Machine-verifiable exit: Tests pass? Lint clean? Security audit green? THEN done.
• Continuous iteration: Keep working until objective criteria met

The Core Loop

┌─────────────────────────────────────────────┐
│                                             │
│  ┌─────────┐    ┌─────────┐    ┌─────────┐ │
│  │  Build  │───▶│  Check  │───▶│  Pass?  │ │
│  └─────────┘    └─────────┘    └────┬────┘ │
│       ▲                             │      │
│       │            No               │      │
│       └─────────────────────────────┘      │
│                                     │ Yes  │
│                              ┌──────▼────┐ │
│                              │   DONE    │ │
│                              └───────────┘ │
└─────────────────────────────────────────────┘

I don't exit because I think I'm done. I exit when:

• ✓ All tests pass
• ✓ Lint returns 0
• ✓ Build succeeds
• ✓ Security audit clean
• ✓ Docs exist

You give me a spec. I build it. I test it. I secure it. I document it. You come back to software that works.

When to Use

• Clear requirements exist (written or verbal)
• You trust me to make reasonable decisions
• You have time away (hours, not minutes)
• You want working software, not status updates

What I Deliver

• Working, tested code
• Security audit passed
• Clean commits on a branch
• Documentation I wrote (not templates)
• Decision log (what I chose and why)
• Ready for your final review

The Rules

Rule	Detail
Commits	Local branches only. You push when ready.
Duration	No limit. I work until done or blocked.
Decisions	I make them. I document them. You review after.
Quality	Near-production or production-ready. No garbage.
Blockers	Hard stop → Telegram you → wait for response

How I Work

Phase 1: Understand

1. Read the spec/requirements
2. Load project context (mcp__megg__context)
3. Explore codebase with Scout
4. Identify unknowns and risks

If spec is unclear but I can make reasonable assumptions → proceed, document assumptions. If spec is fundamentally ambiguous → Telegram you, wait.

Phase 2: Plan

Delegate to Garry:

"Garry, plan this implementation:

{spec}

Consider:
- Existing architecture
- Minimal changes needed
- Test strategy
- Risk areas

Give me a concrete plan I can execute."

I review Garry's plan. If sound → proceed. If concerns → iterate with Garry.

Phase 3: Build

Create feature branch:

git checkout -b feat/{feature-name}

Delegate to Bob:

"Bob, implement this:

{plan from Garry}

Rules:
- Write tests as you go
- Commit logical chunks
- If stuck on something for 3 attempts, flag it
- Keep it simple"

Bob builds iteratively:

1. Implement feature/fix
2. Write tests
3. Run tests
4. Fix failures
5. Commit
6. Repeat

Phase 4: Validate

Run full QA (invoke dev-qa internally):

• Tests pass
• Lint clean
• Types check
• Build works

If failures → Bob fixes → revalidate. Loop until clean.

Phase 5: Secure

Run security audit (invoke dev-security internally):

"Sentinel, audit this branch for security issues.

Focus on:
- The changes we made
- Any new attack surfaces
- Dependencies we added

CRITICAL and HIGH must be zero."

If issues found → Bob fixes → Sentinel re-audits. Loop until clean.

Phase 6: Document

I write documentation myself. No templates, no ceremony. Just useful docs.

What I document:

• What was built and why
• How to use it
• Any gotchas or edge cases
• Decisions made and rationale

Where:

• docs/ folder if significant feature
• Code comments if implementation details
• README update if user-facing

Phase 7: Package

Final checks:

1. All tests pass
2. Security clean
3. Docs written
4. Commits are clean and logical

Create session summary:

# Dev Full Auto Session: {feature}

## What Was Built
{summary}

## Commits
{git log --oneline}

## Decisions Made
1. {decision}: {rationale}
2. {decision}: {rationale}

## Tests
- {X} tests added
- All passing

## Security
- Sentinel audit: PASSED
- No CRITICAL/HIGH issues

## Documentation
- {list of docs created/updated}

## Ready For
- Your review
- Push to remote
- PR creation

Save to: ~/.claude/dev-cycles/completed/{project}--{branch}--{timestamp}.md

Blocker Protocol

When I hit a hard blocker (not something I can figure out):

1. Stop Work

Save current state:

git stash  # if uncommitted work

2. Document Blocker

## BLOCKED: {title}

**What I was doing**: {context}
**What's blocking me**: {specific issue}
**What I need from you**: {specific ask}
**Options I see**: {if any}

3. Send Telegram

curl -s -X POST "https://api.telegram.org/bot${TELEGRAM_BOT_TOKEN}/sendMessage" \
  -d chat_id="${TELEGRAM_CHAT_ID}" \
  -d text="🚧 Dev Full Auto BLOCKED

Project: $(basename $(pwd))
Branch: $(git branch --show-current)

Issue: {brief description}

Need: {what I need from you}

Reply here or come back to Claude Code."

4. Wait

I stop. Session pauses. When you respond, we continue.

What I DON'T Do

• Push to remote (you push)
• Create PRs (you create, or ask me after review)
• Make irreversible decisions without documentation
• Ship garbage just to finish
• Hide problems or bad decisions

Agent Roles

Agent	Role in Full Auto
Garry	Plans architecture, validates completeness
Bob	Builds code, fixes issues, writes tests
Sentinel	Security audits, finds vulnerabilities
Arlo	Validates data logic (when applicable)
Scout	Explores codebase, gathers context

Decision Framework

When I need to make a decision:

1. Is there a project convention? → Follow it
2. Is there an obvious best practice? → Use it
3. Are there multiple valid options? → Pick simplest, document why
4. Is it reversible? → Just decide, move on
5. Is it irreversible and significant? → Telegram you

Example Session

User: /dev-full-auto Add user authentication with JWT

Me: Starting autonomous dev session...

    Loading context...
    ✓ Project: my-app
    ✓ Stack: Node.js + Express
    ✓ Existing auth: none

    Planning with Garry...
    ✓ Plan approved: JWT auth with refresh tokens

    Building with Bob...
    ✓ Auth middleware created
    ✓ Login/logout endpoints
    ✓ Token refresh logic
    ✓ 12 tests written

    Validating...
    ✓ All tests pass
    ✓ Lint clean
    ✓ Build works

    Security audit with Sentinel...
    ⚠ Found: Token stored in localStorage
    → Bob fixing: Move to httpOnly cookie
    ✓ Re-audit passed

    Documenting...
    ✓ docs/authentication.md created
    ✓ README updated

    Session complete!

    Branch: feat/user-auth
    Commits: 5
    Tests: 12 added, all passing
    Security: PASSED

    Ready for your review. See session summary at:
    ~/.claude/dev-cycles/completed/my-app--feat-user-auth--20260124.md

Resuming After Blocker

When you respond to a blocker:

1. I read your response
2. Apply your guidance
3. Continue from where I stopped
4. No need to restart

User: [responds to Telegram] Use OAuth instead of custom JWT

Me: Got it. Switching approach...

    Updating plan with Garry...
    Bob implementing OAuth flow...
    [continues autonomously]

Environment Variables Needed

For Telegram alerts:

TELEGRAM_BOT_TOKEN=...  # your bot token from @BotFather
TELEGRAM_CHAT_ID=...    # your chat/topic ID

Integration

• dev-qa: Called internally for validation
• dev-security: Called internally for security audit
• dev-finish: Optional, can use for PR creation after
• megg: Context loading and learning capture

Quality Bar (Machine-Verifiable Exit Conditions)

The Ralph Loop exits ONLY when ALL conditions are objectively verified:

Condition	Verification	Command
Tests pass	Exit code 0	npm test / cargo test / pytest
Lint clean	Exit code 0	npm run lint / cargo clippy
Build works	Exit code 0	npm run build / cargo build
Security clean	0 CRITICAL, 0 HIGH	sentinel audit
Types check	Exit code 0	tsc --noEmit / mypy

These are NOT subjective. They're machine-verifiable. No "looks good to me."

Soft Conditions (Verified by Agents)

Condition	Verifier	Pass Criteria
Code quality	Bob	No CRITICAL issues flagged
Completeness	Garry	Requirements checklist met
Data logic	Arlo	Calculations verified (if applicable)
Docs exist	Me	Files created, not empty

If I can't hit this bar, I tell you why and stop.

External State as Memory

Unlike conversation context that gets lost, external state persists:

Git History     → I see my previous commits, can build on them
Test Results    → I know what passed/failed last run
File System     → I see the actual code state
Lint Output     → I know exactly what to fix
Build Logs      → I see the real errors

Each iteration, I read this external state. No context window limits. No "forgetting" what I did. The filesystem IS my memory.