메뉴
Extend @tangle-network/agent-eval internals: campaigns, scorecards, trace capture, backend integrity, held-out gates, analysts, auto-PR loops, RL bridge, and eval release checks.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
Raise the ceiling, don't climb toward it. When a target is nearly hit or a metric has plateaued, question the target itself: name the binding constraint, separate the physics floor from the assumed floor, design the regime change that makes 10x reachable, commit through the valley. Triggers: 'breakout', '10x this', 'raise the ceiling', 'why are we capped', 'think way bigger', 'this metric is a cage'.
Prove that a merged change is actually live and production behavior is measured against the deployed artifact. Use after merge, before claiming deployed/live/validated, for Cloudflare Pages/Workers deploys, cache claims, perf claims, and release closeout.
Diagnose and improve eval harnesses built on @tangle-network/agent-eval or similar trace-first systems. Use when success/failure rates, scorecard deltas, judge outputs, promotion gates, benchmark rows, or improvement campaigns may be contaminated by harness bugs, evaluator drift, routing/auth failures, missing traces, or invalid baselines.
Goal-pursuit loop: measure, diagnose, experiment, verify, compare, and iterate against a measurable target. Triggers: evolve, optimize, make this better, push to target.
Split one messy experiment branch into N clean, independent branches — one logical change each, built from the merge-base, disjoint change-sets, each an isolated reviewable PR. Triggers: 'finalize', 'split this branch', 'atomize this', 'make reviewable PRs', 'decompose this mess'.
Read repo/session state and choose the single next skill: exploit (evolve/polish), explore (pursue/meta-harness/breakout), bootstrap (eval-agent), diagnose (diagnose/eval-harness-diagnose), or step back (reflect). One dispatch, then exit.
| name | agent-eval |
| description | Extend @tangle-network/agent-eval internals: campaigns, scorecards, trace capture, backend integrity, held-out gates, analysts, auto-PR loops, RL bridge, and eval release checks. |
Use this only when changing or extending @tangle-network/agent-eval internals.
If the task is product-side adoption, use the relevant adoption skill instead.
references/full-reference.md only when you need the full primitive table or historical bug-class catalog.assertRealBackend / equivalent checks before interpreting results.AgentProfile canonical; do not invent local profile shapes.HeldOutGate depends on correct split tags; mistagging invalidates the verdict.runEvalCampaign plus scorecard recording and integrity checks.Report the source files read, exact exported primitives used, integrity checks preserved, and tests run. If a result depends on trace or scorecard data, include the artifact path and row count.
eval-agent when building a judge component rather than editing the eval package.harden when the change touches trust boundaries, credential capture, or release checks.