跳至正文
OpenAgentSkill
浏览已开放 · 目录记录日期 2026-11-21 · 自动更新和账户服务暂时关闭。
营销与销售

evolve

Evolve one target through bounded candidate generations against one frozen evaluation. Manual-only campaign.

DanMcInerney · 已有静态检查记录

安装与使用

运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。

npx skills add DanMcInerney/orchflows --skill evolve

技能说明


name: evolve description: Evolve one target through bounded candidate generations against one frozen evaluation. Manual-only campaign. disable-model-invocation: true

Require: target, the identity being evolved; incumbent, its fixed starting result/evidence identity; evaluation, the frozen evaluation identity — mode, criteria, promotion rule, margin and search policy — or none when one must be designed first; writer, the skill each candidate is written through; bound, the campaign's budget; and mutation_scope, the candidate workspace, which is the campaign call's alone.

tickets.py frame-open <run> --goal-file <campaign-goal> --bound <bound> --workflow evolve

Design the evaluation only where evaluation is none — a supplied frozen identity is already the campaign's, and every call below reads that instead:

tickets.py do <run> --standard orch-code --parent <frame>
  --goal-file <eval-goal> --bound "<= 40 tool calls"

Its goal: one candidate-blind evaluation for target — freeze: "Fix the identity before any candidate exists and forbid every later call from touching it." — identity, mode, scoring criteria, required admission and regression criteria, artifact-evidence adapter, promotion rule, margin and search policy — written into that call's ## Report and nowhere inside mutation_scope. In judged mode the accepted design owns the judge brief, criteria, aggregation and adapter; in benchmark mode the qualified benchmark and its runner own them.

Admit the incumbent: one judge --standard orch-code whose typed artifact line is incumbent's fixed evidence identity, verdicts over the frozen evaluation's required admission and regression criteria. A covered PASS is what permits generation to open, and nothing else does.

Generations, until the promotion rule and margin are met over the final incumbent's score card or bound is spent. Each generation is prose, not machinery: one do --standard orch-code --isolation required per candidate, written through writer inside mutation_scope and handed the eligibility findings verbatim; one judge scoring the incumbent and every eligible candidate blind against the frozen evaluation; search_plan.py advance selecting the next search-policy/v1 cases. The judge's verdict is the loop's only exit condition.

Close the campaign: one judge --standard orch-code over the final incumbent — one score card naming it and the admitted result/evidence behind it.

Never: rank an ineligible candidate; keep a candidate lacking PASS on every required admission criterion — kill it, a score never compensates; re-execute or substitute admitted evidence; expose protected evidence; call benchmaker; activate a selected candidate; add a closing wrapper; take an archive member as anything but an exploration parent; or unfreeze a campaign constant — a changed constant starts a new campaign and re-evaluates every retained candidate rather than continuing this one.

Return: tickets.py frame-close <run> <frame> --done <check> over that final score card and the admitted evidence it cites.