Run agent evolution

run-agent-evolution

Drive the genetic loop for <id-or-slug>. State is derived from evolution.yaml field presence — re-read between phases.

0. Dispatch

Resolve the argument:

• Pure integer → match the leading <id> under .agents/evolutions/<id>-*/.
• String → match a slug fragment; on multi-match pick the lowest <id> and note in chat.

Read evolution.yaml (validated against evolution.schema.json). Branch:

• Directory missing → "No evolution found for <arg>. Run /new-agent-evolution <title> to capture one." Stop.
• applied set → terminal. "Evolution <id> already applied (winner: <variant_id>)." Stop.
• winner already set → "Winner already picked (v<n>). Run /apply-agent-evolution <id> to land it." Stop.
• Otherwise → continue.

Each iteration is one generation — a batch of up to budget.parallel variants planned, executed, recorded together. The generation boundary is where the loop learns.

1. Resume sweep

On entry, reset any running variants without a result.json to pending and re-dispatch them. Variants with result.json already on disk → ingest immediately and mark evaluated.

2. Stop conditions (first match wins)

• len(variants) >= budget.max_variants → stop, pick winner.
• budget.max_minutes set and elapsed ≥ it → stop, pick winner.
• budget.plateau_generations set and top score has not improved over the last N gens (need ≥ 2 full gens) → stop, pick winner.
• gen >= 2 and zero eligible variants → abort; gate spec is likely broken.

3. Plan the next batch

Compute gen = max(variant.generation) + 1 (or 1). batch_size = min(parallel, max_variants - len(variants)).

Generation 1 — seed for diversity. For each seed, pick a distinct dimension to vary (algorithm, data structure, library, prompt style, control flow). Write a one-line falsifiable hypothesis and a 5–15 line approach concrete enough that a sub-agent can implement it without re-deriving. parents: [].

Generation 2+ — mutate, cross, explore. Read survivors (status evaluated, all gates passing, sorted by the composite score from §5). Mix per your judgment: mutations (parents: [v_n]), crossovers (parents: [v_a, v_b]), and one or two explores (fresh dimension; parents: []). When all prior variants failed gates, do not propagate them — diagnose and seed fresh.

Append each new variant to evolution.yaml.variants[] with id: v<n>, generation, parents, hypothesis, approach, status: pending. Add a one-line entry to EVOLUTION.md §Variants: - v<n> (gen <gen>, parents <ids or "—">) — <hypothesis>.

4. Materialize and dispatch

For each pending variant in this generation:

1. Workspace at .agents/evolutions/<id>-<slug>/variants/<vid>/workspace/. Per scope.kind: worktree → git worktree add <path> HEAD (fail loudly on dirty repo); files → copy each scope.include path preserving structure.
2. Set status: running, workspace: <path>. Save yaml once before dispatch.

Spawn one sub-agent per variant in one message (parallel Agent tool calls). Each prompt is self-contained: workspace absolute path (sub-agent's cwd), the §Brief section of EVOLUTION.md verbatim, hypothesis, approach, every gate (G<id>: <description> | cmd: <cmd>), every rubric axis (R<id>: <description> | direction | cmd | extract), and the output contract — write result.json one directory above the workspace (i.e. .agents/evolutions/<id>-<slug>/variants/<vid>/result.json), validated against result.schema.json:

{
  "status": "evaluated",
  "gates": [{ "id": "G1", "passes": true, "output": "tsdown ✓" }],
  "rubric": [{ "id": "R1", "value": 98.6 }]
}

Hard rules for the sub-agent: implement the approach in the workspace; run gates in declared order; if any gate fails, write status: "evaluated" with that gate's passes: false and omit rubric; if all pass, run rubric measures, extract per extract, write status: "evaluated" with both arrays. Don't modify files outside the workspace. Don't commit. Don't push. The chat reply is advisory — result.json is the contract.

When sub-agents finish, read each result.json. Missing file → status: errored, retry once. Malformed JSON → errored, retry once. Gate failures and killed are honest signals, never retried. Set status, gates, rubric on the matching variant; bump updated_at; save.

5. Score, refresh, maybe stop

After the batch, recompute §Results in EVOLUTION.md from evolution.yaml. The yaml stores raw inputs only — composite scores and ranks are derived on read.

Eligible variants: status == evaluated AND every gates[].passes == true. For each rubric axis, rank eligible variants by direction (best → 1, worst → n; average ranks for ties). Normalize each rank to [0, 1] with (n - rank) / (n - 1) (n = 1 → 1.0). Composite = weighted mean of normalized ranks across axes (default weight 1.0). Render ## Results (after gen <N>) as a Markdown table with columns Rank | Variant | Gen | Parents | Score | <R1..Rk> | Notes, sorted by composite descending (tie-break on lower id); append failed variants below with score — and Notes showing failed gate ids (G2 ✗) or non-evaluated status.

If no stop condition fired, loop back to §2.

Pick the winner (when stopping). Pick rank 1 from §5; tie-break on lower id, then shorter parents[]. Write to evolution.yaml:

winner:
  variant_id: v<n>
  rationale: |
    <one paragraph — composite score, axis-by-axis comparison vs runner-up,
    any caveats from harness or wall-clock cap.>

Refresh §TL;DR in EVOLUTION.md (Winner, Coverage, Next: /apply-agent-evolution <id>).

If no eligible winner exists, don't write winner. Stop with a chat sentence naming the failing gates and recommending the user revise or relax them.

6. Hand off

End with one sentence: generations run, what changed, next pointer. On a clean stop with a winner: /apply-agent-evolution <id>. On no-winner: revise gates and re-run /run-agent-evolution <id>.

Invariants. Every variant has a unique id and a generation strictly greater than its parents'. parents[] references only ids already in variants[]. winner.variant_id is set iff a stop condition fired with at least one eligible variant.