Mutate¶

mutate converts observations into candidate edits. It is required and may change only paths permitted by both editable_roots and the recipe's mutable surface.

Contract¶

class MutateOperator:
    def mutate(self, checkout, observation, ctx) -> MutateResult: ...

The result contains changed paths, notes, and usage.

Library operators¶

Operator	Mutation strategy
`aevolve`	distill recent task observations into prompt, skill, memory, or tool updates
`ahe`	make one testable agent-harness change from current evidence
`gepa`	reflect on configured components and propose a component edit
`hyperagents`	self-referential mutation that may co-evolve target and selected operator policy

Runner configuration¶

operator chooses the mutation strategy. config.runner chooses how its editing agent is launched.

operators:
  mutate:
    operator: gepa
    timeout_s: 3600
    config:
      runner: harbor
      expose_gate_data: false
      agent: codex
      model: gpt-5.4
      environment: docker
      image: evolve-mutate-codex:20260818-codex0146
      editable_roots: [target]
      max_retries: 1

runner: local executes a trusted host command.
runner: harbor creates an isolated writable copy and transactionally imports allowed returned files.
editable_roots does not expand surface.include.
expose_gate_data defaults to false; keep it false for disjoint mutation and protected evaluation.

Operator-specific path keys include components, prompt_path, skills_dir, and memory_dir.

Artifacts¶

Standard outputs include:

runs/gen-N/mutate/changed.json
runs/gen-N/mutate/patch.diff
runs/gen-N/mutate/surface-check.json
runs/gen-N/mutate/rationale.md
runs/gen-N/mutate/usage.json

Harbor runs additionally retain prompt, command, trial, artifact manifest, jobs, and tasks under runs/gen-N/mutate/harbor/.

See Mutate operator execution for runner semantics, artifact handoffs, authentication, and custom Harbor adapters.