Add autoloop — autonomous improvement loop (karpathy/autoresearch)

[5uck1ess]

Summary

Autonomous codebase improvement loop inspired by karpathy/autoresearch. Three files, zero new Go code — YAML workflow executed by the existing Go engine.

How it works

Each cycle is a discrete, auditable experiment:

1. Baseline — run metric command, record starting number + direction
2. Audit — analyze codebase, pick single highest-impact hypothesis
3. Fix — agent makes the recommended change (minimal, focused)
4. Measure — run the SAME metric command again
5. Compare — IMPROVED or REGRESSED based on direction
6. Keep (if improved) — git commit, update scratchpad, loop to audit
7. Revert (if regressed) — git checkout ., update scratchpad, loop to audit
8. Report — final summary with kept/reverted counts and net improvement

Key design decisions

• Measurement is deterministic — same Bash command in baseline and measure steps
• Keep/discard is a branch decision — not agent judgment
• Scratchpad prevents repeating failures — .devkit/scratchpads/current.md
• No new Go code — YAML workflow + existing engine handles everything
• Skill asks questions first — objective, metric, direction, iterations, scope

Files

• workflows/autoloop.yml — 8-step workflow with branch-loop pattern
• skills/autoloop/SKILL.md — triggers on "autoloop", gathers inputs
• commands/autoloop.md — slash command

Test plan

• CI passes (validate-counts: 23 commands, 17 skills, 15 workflows)
• YAML parses correctly via devkit workflow autoloop
• Run on devkit itself: autoloop "improve test coverage" with go test -cover ./...