unclejobs-ai/second-claude-code/evolve

Resolve runtime paths before file or script operations.

Iron Law

Every evolve check is hand-authored by the maintainer — never by the optimized model class — and it scores structural conformance of the prompt asset, not runtime behavior. The maintainer's merge review of `winner.diff` is the final behavioral gate.

Red Flags

• "I'll write a check for what the asset should do" → STOP. The agent must NEVER author the assertion. The maintainer hand-authors every regex; evolve only surfaces which asset is weak. A model-authored check is self-deception — the optimizer games a benchmark it invented.
• "This asset feels weak, evolve it" → STOP. evolve only targets assets with a real, recurring logged gate failure (recurrence ≥ 3). No provenance, no target.
• "Lower min_delta so it promotes" → STOP. Generated suites pin min_delta to the 0.02 floor; evolve refuses to run below it.
• "Score went up, so it's better" → STOP. The score only proves the asset matches the maintainer's regex. Behavior is verified by reading winner.diff before merge.
• "It scored higher, merge it" → STOP. Winners land on the isolated codex/loop-* branch; merging to main is a manual maintainer decision.

Evolve

The ouroboros maintainer loop. It closes the self-improvement ring: PDCA runs → Check gate fails → failure logged → evolve harvest surfaces the weak asset → the maintainer authors a structural check → the unmodified loop engine evolves the asset on an isolated branch → the maintainer merges → the next PDCA run starts stronger.

evolve is a target-selection + check-authoring layer on top of `loop`. It runs no optimizer of its own and relaxes no loop gate.

When to Use

• The same PDCA gate (sources_min_5, min_two_reviewers, …) keeps failing across runs and the root cause is the producing prompt asset, not the individual run.
• The maintainer can express the desired structure as static regexes.
• A fixed loop engine + promotion gates already exist (they do).

v1 Scope — what can actually promote

The loop's mutations are five fixed text substitutions: should→must, can→must, a legacy command-prefix rewrite to /scc:, trailing-whitespace trim, blank-line collapse. None add content. A candidate beats baseline only when the maintainer's regex flips fail→pass because a substitution fired — i.e. the asset still holds a should / can / legacy-prefix token the regex wants rewritten.

A check requiring new content the asset lacks (a missing heading, an added section) can never be satisfied by v1 mutations — every candidate stays byte-identical and the run ends min_delta_not_met. Those need manual editing or a future mutation set. Author assertions inside the promotable space.

Subcommands

Workflow

1. node scripts/evolve-runner.mjs list-failures — which assets recur. Provenance only, never a check.
2. Decide the structural property, inside the v1 promotable space above. The maintainer authors the regex. The agent must not invent it.
3. node scripts/evolve-runner.mjs harvest <id> --assertion '/scc:' — a normalization check (the legacy long command prefix is rewritten to the canonical /scc: namespace). Writes a validated suite with the check inline (base64) so the worktree needs no extra file.
4. Commit scripts/evolve-scorer.mjs and the target asset, then node scripts/evolve-runner.mjs run evolve-<id> — runs the loop on an isolated branch; prior ratified checks ride along as weight-0 holdouts. (Refuses if either is missing from HEAD.)
5. Read .captures/loop-<runId>/winner.diff. Merge to main manually only if genuinely better.

Hard Gates

• Maintainer authorship is a procedural control, not a code gate. The runner stamps check_author: "maintainer" and refuses an empty --assertion, but cannot tell a human regex from a model one. The "never model-authored" guarantee rests on this skill plus your review — never let the agent invent the assertion.
• Generated min_delta is pinned to 0.02; evolve run refuses below it. (Invoking loop-runner.mjs run evolve-* directly bypasses this — always use evolve run.)
• Targets stay in the v1 scope (skills/**/SKILL.md, agents/*.md, commands/*.md, templates/*.md), restricted to a safe charset so they cannot carry shell metacharacters into the generated command.
• Inherited loop gates (syntax, manifest, contract-smoke, allowed-targets, critic-output) apply unchanged — evolve edits no engine code. evolve does not validate a suite's case-command safety (a baseline loop property); only run suites evolve itself generated.

State

• Provenance ledger: <plugin-data>/evolve/failures.jsonl (local-only)
• Asset attribution map: config/evolve-asset-map.json
• Generated suites: benchmarks/loop/evolve-*.json
• Run artifacts: .captures/loop-<run_id>/ (owned by the loop engine)

Gotchas

• Slash-only. No prompt-detect.mjs routing. A self-mutating loop must never trigger autonomously, and is never daemon-scheduled.
• v1 scores structure, not behavior, bounded by the loop's fixed mutations. Calling it "failure reproduction" would be false — it is structural-conformance evolution against a maintainer-authored check.
• Do not author the assertion from a model's guess. Do not merge a winner without reading winner.diff.