miaodx/roboclaws/eval-evolution

Eval Evolution

Use this Skill for maintainer-owned optimization campaigns driven by frozen eval evidence. roboclaws.evals remains the control plane. The optimizer and robot trials are distinct OpenAI Agents SDK agents; a provider profile selects their models and never changes the agent engine.

Prepare

Freeze an eval_evolution_campaign_v1 JSON manifest before the first optimizer call. Bind one target kind and target, the baseline commit and target digest, mutable paths, optimizer and robot provider/model identities, paired training suites, the orchestrator-only sealed confirmation reference, quality and minimum-improvement policy, runtime identity, and explicit turn/trial/token/ cost/time/retry ceilings. The budget block also declares positive optimizer_call_tokens, optimizer_call_cost_usd, robot_attempt_tokens, and robot_attempt_cost_usd reservations. Each reservation is a frozen maximum for one optimizer run or one robot attempt, not a post-hoc usage target.

Skill campaigns target exactly one skills/<name>/SKILL.md. Keep static-full as the baseline. no-skill is a non-promotable negative control. Do not mix Skill and MCP changes.

Run

Inspect the blocked preflight without provider execution:

just agent::eval evolve campaign=<campaign.json>

Run the frozen campaign only after provider/runtime readiness is established:

just agent::eval evolve campaign=<campaign.json> live_execution=run

The optimizer can only read the declared target, read sanitized feedback, and submit one hypothesis plus patch. It has no shell, filesystem, git, network, eval-launch, commit, or publication tool. Host-owned validation creates one content-addressed full baseline snapshot and runs deterministic gates before paired robot trials.

Review

Reject missing or mismatched identity, incomplete paired evidence, privacy or checker failures, trajectory/terminal violations, quality regressions, neutral rewrites, and no-skill. Efficiency ranks only candidates that meet the frozen quality and minimum-improvement rule. At most one training winner reaches one sealed confirmation. Its evidence never returns to the optimizer.

The orchestrator reserves the full optimizer-run or robot-attempt token and cost maximum before provider execution and passes that limit into the Agents SDK model settings. Insufficient remaining campaign capacity is inconclusive without starting the provider. A model without catalog output pricing is also blocked before execution because its cost ceiling cannot be derived safely. Behavior/provider failures do not retry. A separately recorded classified infrastructure attempt is allowed only when the frozen campaign permits one retry.

Promote

Promotion requires an accepted selection report and a digest-bound eval_evolution_promotion_manifest_v1 with maintainer_approved=true:

just agent::eval evolve-promote \
  report=<selection-report.json> \
  manifest=<maintainer-approved.json> \
  live_execution=run

Stop before creating that approval unless a maintainer has reviewed the exact candidate, paired training, sealed confirmation, limitations, and digests. Promotion applies only the reviewed patch. It does not commit, change defaults, or publish a baseline/catalog artifact.

MCP description candidates additionally require the Phase 2 structural text-only proof. Never run MCP behavior candidates live until the Phase 3 malicious isolation proof has passed on the selected placement.