evolve

Evolve

Treat this skill as the single-agent abstraction of ASI-Evolve-dev. Preserve the system's core idea: learn from prior knowledge, design the next candidate, run the evaluator, analyze the outcome, and feed the lesson back into the next round.

Preserve the operating model

• Keep the four-stage loop intact: learn, design, experiment, analyze.
• Use the bundled wrappers in scripts/ for stateful operations. Let the agent make orchestration decisions; do not recreate the repository's pipeline stack.
• The agent itself must choose mutations, run experiments, inspect outputs, and write round analyses step by step.
• Keep two memory systems alive for the whole run:
  • cognition store for approved external research, paper takeaways, and search-derived insights that may help future rounds
  • experiment database for every candidate, score, analysis, lineage, best snapshot, and lessons learned from actual experiments
• Treat .evolve_runs/<run-name>/ as the source of truth for the active run.

Start with preflight

• Draft or update the run spec before any mutate or evaluate step.
• Align with the user on:
  • objective
  • core score
  • secondary metrics
  • evaluation command or script
  • evaluation timeout
  • success criteria
  • stop conditions and round budget
  • writable file scope and primary targets
  • sampling algorithm
  • island feature dimensions when sampling.algorithm=island
  • cognition source mode
• Inspect the evaluator before confirmation. If the command or script is vague, pause and resolve it before continuing.
• Even when the user already gave a detailed task description, you must still produce a concrete preflight plan/approach summary before any evolve round starts.
• Require an explicit evaluator timeout during preflight. Do not treat timeout as an implicit default.
• Confirm that the evaluator path you will use has timeout handling. The outer evolve-eval run timeout is mandatory, and the evaluator command or script should also accept or honor the configured timeout when it can hang internally.
• Confirm that the evaluator can load the materialized candidate path that evolve-eval run produces. The default step artifact is steps/<step-name>/code with no forced extension or filename convention.
• If you choose the sampling algorithm yourself during preflight, tell the user explicitly which algorithm you picked and why.
• If you choose island, also tell the user the default feature semantics: complexity=len(code) and diversity=code-difference heuristic over stored programs, and mention that they can override the feature list before confirmation.
• Refresh the preflight artifacts with scripts/evolve-brief normalize.
• Keep approval.confirmed=false until the user explicitly approves the preflight summary.
• Only flip approval.confirmed after the user says the plan is confirmed or approved. Never self-confirm because the request seemed detailed or complete.
• Refuse to run evolve commands that mutate files, execute the evaluator, or write the final summary before confirmation.