Track and surface sub-agent dispatch calibration ratios. Extract from JSONL transcript or read existing ledger. Per-agent clean/blocked/errored/partial trends.
中文简介正在准备中。当前显示原文说明。
Evidence-gated synthesis over recent operating evidence. Appends admitted lessons and a dated synthesis with propose-only improvement candidates.
中文简介正在准备中。当前显示原文说明。
实测评分 未实测
查看安装方法安装前可以查看交给 AI 代理的指令内容。
/dream is the canonical writer for lessons and Dream synthesis. It never
applies an improvement, promotes its own proposal, edits system instructions,
or treats an agent-authored claim as verification.
vault/memory/runtime/LESSONS.jsonl.vault/memory/DREAM_LOG.md.memory/ fallback is valid only for minimal installations.runtime/DREAM_LOG.md copy is not an input or output./close and /nightly invoke this protocol instead of carrying a competing
inline writer./dream
/dream --window nightly
/dream --set-status <candidate-id> <approved|merged|rejected|blocked> \
--operator-evidence "<direct operator reference>"
proposed -> approved always requires the operator's exact decision. Never
infer approval from silence, an automated status, another agent, or prior prose.
Read, in order:
created_at.NIGHTLY_CAPTURE_CANDIDATES.jsonl rows from the prior fire as
discovery and pressure signals only.Record exact source references. A capsule or Session Log entry can establish that a claim occurred; it cannot prove the claim true.
Append a lesson only when one is true:
Two retellings of one event are not independent. Discard one-off observations, unverified causal stories, and self-confirmation.
Each admitted lesson has:
lesson_id: next max numeric lNNN suffix plus one;type: procedural | calibration | strategic | failure-mode | doctrine;confidence: 0.0..1.0, capped at 0.85 for one mechanical event;source_events: at least two refs unless an allowed exception applies;created_at: the configured local YYYY-MM-DD;admission: one of:
{kind: "two-source"};{kind: "operator-outcome", operator_ref, outcome_ref};{kind: "mechanical-red-green", red_exit, green_exit: 0, artifact_ref}.Serialize UTF-8 JSON, parse the complete JSONL file after append, and confirm the new line round-trips. Historical duplicate ids do not authorize a new collision.
When evidence names a repeatable system gap, append this inside the same dated Dream section:
### ni-YYYYMMDD-NN - short name
- **status:** proposed
- **evidence:** source-ref-1; source-ref-2
- **failure class:** concise class
- **proposed_change:** smallest testable change
- **mutation proof:** break condition -> named red; restore -> named green
- **operator gate:** approve / reject / revise
- **operator evidence:** pending
Generated candidates always start with the literal field status: proposed.
Any generated approved,
merged, or self-ratifying wording is a hard failure.
Append one ## YYYY-MM-DD (...) section every time synthesis runs, even with
zero admitted lessons. It must contain nonempty forms of all eight labels:
A quiet run writes a minimal zero-lesson entry and appends no Lessons row. Never overwrite an earlier entry.
A nightly fire uses the snapshot created by nightly-pass begin. Explicit
/dream creates an operating-system temp snapshot first:
node daemons/nightly-contracts.mjs snapshot --root <aigent-root> \
--out <os-temp>/dream-contract-before.json
After writing:
node daemons/nightly-contracts.mjs dream --root <aigent-root> \
--date <YYYY-MM-DD> --snapshot <os-temp>/dream-contract-before.json
Only DREAM_CONTRACT PASS is green. The validator proves byte-prefix
preservation, parses JSONL, derives ids, validates admission, masks fenced
examples, requires the eight sections and current configured date, and rejects
a newly generated candidate that does not start proposed.
Return:
DREAM PASS date=<YYYY-MM-DD> sources=<N> lessons=<N> candidates=<N>
还没有评价。欢迎分享你的使用感受。
无需登录即可发表。欢迎分享使用感受、提问,或说明在哪些环境下无法运行。
还没有评论。来写第一条吧。
Track and surface sub-agent dispatch calibration ratios. Extract from JSONL transcript or read existing ledger. Per-agent clean/blocked/errored/partial trends.
中文简介正在准备中。当前显示原文说明。
Compose the AIgent's body state by reading existing signals. Lazy compute. Never polled. Reports the five pressures + recommended reflex.
中文简介正在准备中。当前显示原文说明。
Generate a structured delegation brief for any task
中文简介正在准备中。当前显示原文说明。
Quarterly skill that walks the skill-index against the actual SKILL.md files on disk, reports drift (skills indexed that no longer exist, skills on disk that aren't enrolled, skills whose descriptions have changed since enrollment). Hardens the auto-enrollment pipeline against silent rot.
中文简介正在准备中。当前显示原文说明。
When a [CADDY] hint fires and the principal is confused why, /caddy-explain walks the scoring transparently — which trigger matched, what points it earned, what the threshold was, what other skills were close. Pairs with the legibility thesis. Caddy's decisions should be inspectable on demand.
中文简介正在准备中。当前显示原文说明。
Enroll a skill into Caddy's index so it can be surfaced when the right prompt arrives. Use when a new skill is created, copied in, or when a skill needs updated triggers. Usage — "/caddy-enroll <path to SKILL.md>".
中文简介正在准备中。当前显示原文说明。