菜单
Read durable reflect projections, group repeated signals by skill, and emit validated handoffs to skill-lab improve when the evidence clears configured floors.
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。
Governed runtime for agent skills: discover and install portable skills, run bounded skill graphs with explicit authority, and inspect signed receipts for what happened.
Audit a sealed runx receipt for governance, comparing the authority a run exercised against what it was granted, and flag over-reach, ungated mutation, unrecorded refusals, or exposed secret material.
Verify the receipt ids behind a normalized authority-usage summary, compare that evidence with granted scopes, and propose the narrowest grant the evidence supports.
Answer a cross-run audit question against the receipt ledger, returning matched receipts and a chain-verification result.
Review receipts and harness failures to propose bounded skill improvements.
Read Runx's native receipt history and skill catalog, then return deterministic run outcomes, catalog test coverage, and governance follow-ups without model-authored metrics.
| name | reflect-digest |
| description | Read durable reflect projections, group repeated signals by skill, and emit validated handoffs to skill-lab improve when the evidence clears configured floors. |
| runx | {"category":"authoring"} |
Turn repeated post-run reflection signals into bounded improvement work without re-diagnosing one receipt or drafting a parallel pull-request artifact.
The graph reads durable reflect events through data-store, admits explicit
replay projections when supplied, applies deterministic confidence and support
floors, then asks one bounded agent act to describe at most one improvement per
skill. A final deterministic step verifies every cited receipt belongs to that
skill's group and emits a skill-lab improve handoff.
The skill does not write a package, open a pull request, publish, or invoke the mutating improvement runner. The handoff preserves the target, objective, receipt evidence, and non-goals for later governed execution.
reflect_projections: explicit projections for replay; otherwise durable
state is read from data_source_ref and state_resource.skill_filter, since: optional read bounds.min_support: minimum number of admitted projections for one skill.min_confidence: minimum confidence for each admitted projection.proposals: validated skill-specific improvement opportunities.skill_lab_handoffs: executable request packets naming skill-lab, runner
improve, target directory, objective, primary receipt, evidence summary,
and supporting receipt ids.