Auto

/auto — Autonomous Execution Loop

Inspired by Karpathy's autoresearch ratcheting pattern and self-improving agent loops.

Usage

/auto                         # run autonomous loop on all pending stories
/auto --group B               # run on a specific parallel group
/auto --dry-run               # plan iterations without executing

Prerequisites

• specs/stories/ approved by human (run /spec first)
• specs/design/ approved by human (run /design first)
• .claude/program.md exists with instructions + constraints + stopping criteria

Agent Delegation

/auto is the orchestrator — it does NOT implement code itself. It delegates via the Agent tool:

• Implementation: Spawn implementer agent (or agent team for parallel groups)
• Review: Spawn code-reviewer + security-reviewer agents concurrently
• Testing: Spawn test-engineer agent for E2E tests (after all stories pass)

The /auto skill itself runs the ratchet gate (bash verification commands) and manages state files directly.

The Loop

Read program.md → Pick next story → Implement → Test → Review → Keep/Revert → Learn → Repeat

For each iteration:

1. Read .claude/program.md — check instructions, constraints, stopping criteria.
2. Read specs/state/learned-rules.md — inject lessons into the implementer's prompt.
3. Pick the next story from dependency graph (respect group ordering).
4. Record coverage baseline before implementation:

   uv run pytest --cov=src --cov-report=term-missing -q 2>/dev/null | grep "^TOTAL" | awk '{print $NF}' > /tmp/coverage-before.txt
   

5. Spawn implementer agent via Agent tool with this prompt:

   Implement story [ID]. Read `.claude/skills/code-gen/SKILL.md` for quality principles.
   Read `specs/state/learned-rules.md` — do NOT repeat these past mistakes: [paste rules].
   Write code to backend/ and frontend/. Write tests. Run verification.
   

6. Ratchet gate — run the full verification:

   uv run pytest -x -q --cov=src --cov-report=term-missing
   uv run ruff check . && uv run mypy src/
   npm test -- --coverage && npm run lint && npm run typecheck
   

7. Coverage ratchet — compare before/after:

   uv run pytest --cov=src --cov-report=term-missing -q 2>/dev/null | grep "^TOTAL" | awk '{print $NF}' > /tmp/coverage-after.txt
   # If coverage-after < coverage-before → FAIL
   

8. Decision:
   • PASS → git commit, append to specs/state/iteration-log.md, move to next story.
   • FAIL → enter self-healing loop (see below).
9. Update .claude/program.md Current Focus section with current story and iteration count.
10. Check stopping criteria — if met, stop and report to human.

After all stories in a group pass:

11. Spawn code-reviewer + security-reviewer concurrently via Agent tool (same pattern as /review skill).
12. BLOCK findings → enter self-healing loop (count toward retry budget).
13. All clear → commit group, update iteration log with group summary.

Self-Healing Loop

When a ratchet gate or review fails, do NOT immediately revert. Instead:

Step 1: Diagnose

Read the error output and classify into one of these categories:

Step 2: Targeted fix

Apply the fix strategy for the diagnosed category. Do NOT rewrite the whole story — make the minimal change to pass the gate.

Diagnose error → Classify category → Apply minimal fix → Re-run gate → Pass? → Commit
                                                                        → Fail? → Retry (max 3 total)
                                                                                  → 3rd fail? → Revert + Learn

Step 3: Revert only as last resort

After 3 failed fix attempts on the same error:

1. git checkout -- . (revert all changes for this story)
2. Append full failure details to specs/state/failures.md
3. Extract a defensive rule to specs/state/learned-rules.md
4. Escalate to human by updating program.md with BLOCKED: [story-id] — [reason]

Failure-Driven Learning

When to extract a rule

Extract a defensive rule into specs/state/learned-rules.md when:

• The same error type appears 2+ times in failures.md (pattern detected)
• A self-healing fix succeeds after a failure (capture what worked)
• A story is BLOCKED after 3 retries (capture what to avoid)

Rule format

## Rule N: [short title]
- **Source:** Iteration X, Story [ID]
- **Pattern:** [what went wrong — be specific]
- **Rule:** [what to do instead — imperative, actionable]
- **Applied in:** [iteration where it was first used, or "pending"]

How rules are injected

At step 2 of every iteration, read learned-rules.md and include ALL rules in the implementer agent's prompt as a "do NOT repeat" checklist. The implementer must acknowledge each rule before writing code.

Rule categories

Rules accumulate in these categories — they are NEVER deleted:

• Import rules — dependency ordering, layer compliance, circular import prevention
• Type rules — Pydantic model patterns, TypeScript interface patterns
• Test rules — fixture requirements, mock boundaries, coverage patterns
• Runtime rules — error handling patterns, edge cases discovered
• Architecture rules — layer boundary violations, file placement

Self-Improvement

The scaffold improves itself over time through three mechanisms:

1. Learned rules accumulate (per-project)

Each project builds its own learned-rules.md. Over time, the implementer gets better at this specific codebase because recurring mistakes are explicitly prevented.

2. Review feedback tightens (per-project)

The code-reviewer reads learned-rules.md and adds new rules when it finds violations. This creates a feedback loop: failures → rules → prevention → fewer failures.

3. Gotchas evolve (cross-project)

When you start a new project with this scaffold, review the learned-rules.md from previous projects. Promote recurring rules to the relevant SKILL.md Gotchas section — this makes them permanent and project-independent.

Project failures → learned-rules.md (project) → Gotchas in SKILL.md (scaffold) → All future projects benefit

State Files

All in specs/state/ — persistent across iterations:

• iteration-log.md — every iteration: story, action, result, duration, coverage, commit
• learned-rules.md — defensive rules extracted from failures (injected into future prompts)
• failures.md — raw failure details for pattern analysis
• coverage-baseline.txt — last known good coverage % (updated on each PASS commit)

See references/state-schema.md for format.

Ratcheting Rules

• Only improvements persist. Failed iterations revert completely (after self-healing attempts).
• Coverage never drops. If a commit would lower coverage, it's rejected.
• Tests never break. If existing tests fail after a change, revert.
• Lessons accumulate. learned-rules.md grows monotonically — rules are never deleted.
• Self-heal first, revert last. Always try to fix the error before reverting the entire story.

Stopping Criteria (from program.md)

The loop stops when ANY of these are true:

• All stories complete and tests pass.
• A story fails 3 consecutive iterations (escalate).
• Architecture violation that hooks can't fix.
• Coverage drops below threshold.
• Max iterations exceeded.

Gotchas

• Not reading program.md each iteration. Constraints change mid-run. Always re-read.
• Retrying the same approach. After a failure, read learned-rules.md and try a DIFFERENT strategy. Same code = same failure.
• Reverting too eagerly. Try self-healing first. Revert is expensive — it throws away all work for the story. A lint fix or type annotation takes seconds.
• Reverting too broadly. Only revert the current story's files, not the entire repo. Use git stash or targeted git checkout.
• Ignoring the failure log. If the same error appears 2+ times in failures.md, it's a pattern — extract a defensive rule.
• Autonomous drift. If the agent starts making changes not in any story, the constraints in program.md should catch it. If not, add a constraint.
• No human check-in. Cap at 50 iterations. The human needs to review progress periodically.
• Not injecting learned rules into the prompt. If the implementer doesn't receive the rules, it'll repeat the same mistakes. Always paste them into the agent spawn prompt.