菜单
Track and surface sub-agent dispatch calibration ratios. Extract from JSONL transcript or read existing ledger. Per-agent clean/blocked/errored/partial trends.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
The capsule verb, invoked explicitly — reconcile from the live memory, write a resume-ready capsule, then stop. Mid-session checkpoint or completion capsule; a rolling best-effort version of the same write also runs automatically on every Stop event via daemons/stop-capsule-writer.mjs.
The resume verb, invoked explicitly — load the newest valid capsule by created_at, re-ground against live memory, then ACT on waiting_on. Mid-session or on explicit ask; the post-clear boot fires the same procedure automatically via daemons/resume-verb.mjs.
Run the volunteering ledger before declaring work done. Forces surfacing of what changed/didn't, what was noticed-but-not-fixed, what was guessed, and what tradeoffs were made on the principal's behalf. The standard - the principal should never be surprised later by something the agent knew at the time.
Capture a confident agent claim (Phase 1) or resolve one as right/wrong (Phase 2) in vault/memory/TRUST_DECAY.md. Builds the longitudinal record of agent calibration over time. Pairs with /honesty-check (which should reduce Phase 1 frequency) and /diagnose (which surfaces causes that often reveal Phase 2 events).
One evidence-gated pass that sequences /dream, /reconcile, /sweep-now, and /digest over your own vault, with a /system-check gate and a ledger-review pass over the measurement layer (Honesty Ledger, Trust Decay, Failure Modes). Stays exactly as human-gated as the legs it calls -- stages, never auto-applies. Skips a leg gracefully when there is nothing to do; every leg that runs leaves an evidence line.
Dispatch Hestia to run all 3 vault sweeps (DELEGATION_TRACKER stale items, HEAT_INDEX dormant flips, broken wikilinks) and append the result to HESTIA_SWEEP_LOG.
| name | agent-fitness |
| agent | none |
| description | Track and surface sub-agent dispatch calibration ratios. Extract from JSONL transcript or read existing ledger. Per-agent clean/blocked/errored/partial trends. |
| allowed-tools | Bash, Read |
| user-invocable | true |
| triggers | ["agent fitness","agent calibration","agent stats","sub-agent stats","which agents work","/agent-fitness","agent dispatch trends"] |
Three modes:
/agent-fitness extract — runs python3 daemons/agent-fitness-extract.py to scan the latest JSONL transcript and append new dispatch rows to memory/AGENT_FITNESS.md. Idempotent (dedup by session_id + tool_use_id).
/agent-fitness (no args) — runs python3 daemons/agent-fitness-report.py. Surfaces:
c7 b1 e0 p0)/agent-fitness --days 7 — same as above but windowed to last N days. Useful for "is this agent failing more lately?"
Calibration ratio per agent: clean / (total − errored). The denominator excludes hard errors (infrastructure/tool failures) and only counts agent-level outcomes.
Trend signal: any agent with >=2 blocks in the window surfaces a repeat-blocker callout. The classifier was hardened in v0.5.1 to catch real Lyra-style block patterns ("BLOCKED:", "are denied", "blocker-found" etc).