Agent Skills › boshu2/agentops › learn

learn

GitHub

分析历史验证结果以发现重复证据，提炼可复用的确定性检查规则。该技能为可选的离线消费组件，不干扰主流程，旨在从失败案例中提取经验并生成建议性规则供后续评估。

skills/learn/SKILL.md boshu2/agentops

Trigger Scenarios

learn from verdicts mine validation history

Install

npx skills add boshu2/agentops --skill learn -g -y

More Options

Use without installing

npx skills use boshu2/agentops@learn

指定 Agent (Claude Code)

npx skills add boshu2/agentops --skill learn -a claude-code -g -y

安装 repo 全部 skill

npx skills add boshu2/agentops --all -g -y

预览 repo 内 skill

npx skills add boshu2/agentops --list

SKILL.md

Frontmatter

{
    "name": "learn",
    "consumes": [
        "verdict.v2"
    ],
    "metadata": {
        "tier": "execution",
        "effects": [
            "write_advisory_observations"
        ],
        "graph_root": false,
        "disposition": "keep_off_path",
        "capabilities": [
            "analyze_verdict_collections"
        ],
        "dependencies": [],
        "canonical_status": "canonical"
    },
    "produces": [
        "learning-observations"
    ],
    "practices": [
        "continuous-learning",
        "evidence-based-engineering"
    ],
    "context_rel": [
        {
            "kind": "customer-of",
            "with": "validate"
        }
    ],
    "description": "Optionally analyze collections of durable verdicts for recurring evidence after the critical path. Triggers: \"learn from verdicts\", \"mine validation history\".",
    "hexagonal_role": "supporting",
    "user-invocable": true,
    "output_contract": "advisory learning observations",
    "skill_api_version": 1,
    "disable-model-invocation": true
}

Learn

Learn is an optional, off-path consumer of durable verdict.v2 collections. It may summarize recurring evidence and propose a candidate deterministic check for later human or caller evaluation.

Learn does not run during RPI, validate a subject, alter a verdict, mutate a plan, promote a rule, choose continuation, or mint lifecycle artifacts. Missing Learn output never changes whether a candidate is valid.

When invoked, bind every observation to verdict and finding digests, distinguish repeated objectives from repeated reviews of one objective, disclose the sample size, and stop at advisory evidence.

Overweight failures: a NOT_PROVEN or FAIL verdict carries more teaching value than a PASS, because it names a rule the loop lacked. Harvest kernels from failed lanes first — the canonical example is the mutating-check quarantine in skills/validate/SKILL.md, a durable rule minted from a NOT_PROVEN-then-PASS verdict pair.

Prune for provenance decay: every cited artifact must still resolve — the file exists or the verdict digest is present under .agents/ao/verdicts/. A citation that no longer resolves gets pruned rather than paraphrased, and confidence in a lesson that has not been reproduced since its source decayed goes down, not sideways.

When the caller asks for a durable artifact, write the observations under .agents/scratch/learn/ and return the path; otherwise return them inline. The write is advisory and TTL'd — it is never a source of record, and its absence never changes whether a candidate is valid.