skill-tournament
Apply the evolve campaign to one fixed skill identity against a benchmark built and qualified for it.
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 skill-tournament技能说明
name: skill-tournament description: Apply the evolve campaign to one fixed skill identity against a benchmark built and qualified for it. disable-model-invocation: true
Require: skill, the fixed skill identity being evolved; surface, its
declared mutable surface, which belongs to the campaign and its candidates;
policy, the frozen search policy, promotion rule and margin; bound, the
campaign's budget, which the benchmark's own allocation is never drawn from;
and sources, rigor, standard, pinned target_configuration and preregistered
calibration_policy. Supply benchmark package/git workspace, repository-relative records, durable external
evidence_export, evidence-store root, native mechanism, access_policy,
qualification/reference-audit allocation and separate benchmark bounds.
Missing benchmark inputs stop before dependent work with named gaps.
One skill improves against one benchmark built and qualified for it — freeze: "Fix the identity before any candidate exists and forbid every later call from touching it." Both halves are workflows, not calls: each opens its own frame under this one, and the ticket tree is the call tree.
tickets.py frame-open <run> --goal-file <tournament-goal> --bound <bound> --workflow skill-tournament
Build the benchmark. Invoke benchmaker with target=skill, the
skill's declared observable outcome as outcome, and all benchmark inputs above,
including configuration/policy and workspace and records/export. The construction standard
must resolve in benchmaker's own scope; a tournament-private name cannot cross
this public boundary. Open its frame under this one and execute its body:
tickets.py frame-open <run> --parent <frame> --goal-file <benchmark-goal> --workflow benchmaker
Enter the campaign only with revision-bound VALID qualification, CALIBRATED development evidence meeting the declared policy, an eligible frozen git revision, and final-evaluation evidence satisfying required rigor. Pending, INVALID, UNVERIFIED, OUT_OF_BAND or legacy controls-only results stop here with their evidence and gaps. Keep final score/drift separate from development calibration. The eligible benchmark revision stays fixed for the campaign.
Spend the campaign. Invoke evolve under this frame the same way, with
target=skill, the skill's current fixed result/evidence as incumbent,
the benchmark's qualified revision plus policy as evaluation,
writer=orch-do, mutation_scope=surface, and bound. Its final score
card names the final incumbent and the one benchmark revision every
candidate was scored against.
Never: mutate skill — the benchmark is built for it, never by changing it;
generate, score or compare a candidate here, since both halves own that;
change the benchmark or the policy inside the campaign; restate or call
evolve's verification, search or selection internals; let a benchmaker frame
targeting benchmaker invoke evolve; or activate a selected result — that
requires a separate authorized integration.
Return: tickets.py frame-close <run> <frame> --done <check> over the
campaign's final score card and the fixed benchmark revision it cites.