AGENT SKILL

asset-placement-gate

Routes a proposed skill, plugin, or agent to its correct home — forge-harness (FH) meta-skill, project-local agent, or drop — by applying a 4-criteria meta-skill bar followed by a project-local value test.

License 许可
MIT
最近更新
2026-11-15
Stars
★ 33
Forks
1

01 前言 Front Matter

name
asset-placement-gate
model
sonnet
description
Routes a proposed skill, plugin, or agent to its correct home — forge-harness (FH) meta-skill, project-local agent, or drop — by applying a 4-criteria meta-skill bar followed by a project-local value test.
allowed-tools
[ "Read", "Grep", "Glob", "Bash" ]
user-invocable
true

03 SKILL.md 正文

asset-placement-gate

Routes new skills, agents, and plugins to the correct location when proposed.

Triggers

  • "Should I make this an FH skill?"
  • "Where should I put this agent?"
  • When a new FH asset creation is proposed
  • /asset-placement-gate {asset name or description}

Natural Language Triggers (example user phrases)

When unsure where to place a new asset or skill:

Example phrase Intent
"Should I put this in a separate file?" Asset necessity + placement decision
"Can this pattern be shared across projects?" FH cross-project value assessment
"Is this only for our project, or can others use it too?" cross-project vs local routing
"What if I turn this into a skill and use it across projects?" FH 4-criteria trigger
"Where should I put this agent if I want to share it?" Placement routing needed
"Should I extract this to a separate file, or leave it?" Asset necessity + placement decision
"I'd like to manage this as a shared resource" Cross-project shared management review
"Decide whether this stays local or is available to other teams too" cross-project vs local routing
"I don't know where to save this" Placement routing needed
"Should this become a shared asset?" FH 4-criteria trigger

Execution Order (summary)

  1. Request full file path from user (or accept natural language description)
  2. Load asset content via Read (if path provided) 2.5. Step 0.5 mechanical overlap pre-scan (grounds criterion ④) + Step 0.6 official-corpora check (grounds criterion ③)
  3. Evaluate Step 1 4-criteria in order (LLM makes the judgment, ④ gated on the Step 0.5 scan, ③ informed by Step 0.6)
  4. ① + ④ both pass + at least one of ②③ passes → output "FH suitable" Otherwise, proceed to Step 2 local assessment → if fails, output "Project-local agent or no asset needed"

Step 0. Parse Input + Fetch Asset

Immediately after trigger, acquire asset content in the following order.

  1. Path provided (contains .md or starts with /): Read the file.

    • If read fails: "File not found. Please verify the path or provide a direct description." — then stop.
  2. Description provided (natural language): Check whether all 3 fields can be identified:

    • Purpose: What this asset does (1 line)
    • Trigger: In what situation is it invoked?
    • Expected callers: Which project / which users will use it?

    All 3 identifiable → use as asset description. Any 1 unclear → stop with this question:

    The asset description is insufficient. Please provide:

    1. Purpose of this asset (one line)
    2. In what situation is it invoked?
    3. Which project/users will primarily use it?
  3. No input: Stop with the question below.

    Which asset should I evaluate? Enter a file path (e.g., .claude/agents/jira-create.md) or a description.

After acquiring the asset content, run Step 0.5 (mechanical overlap pre-scan) and Step 0.6 (official-corpora check) before the judged Step 1.

Step 0.5. Mechanical Overlap Pre-Scan (grounds criterion ④)

Criterion ④ ("no overlap with existing FH skills") is otherwise an LLM recall judgment with no ground truth — a duplicate skill with a novel name passes because the judge has no enumerated list to check against (judge-robustness swarm, 2026-06-13). Ground it mechanically first:

Run this in Bash from the FH repo root (the pipeline below is grep | grep -v | grep -c, which the Grep tool cannot express).

# LEG A — enumerate existing skill names + descriptions (grounds the judged comparison).
# This leg is ALSO the liveness check for leg B: a dead scan and a genuine no-collision both
# produce collision-count 0, so the enumerate count must be read FIRST.
ROSTER="$(grep -riE 'name:|description:' plugins/fh-meta/skills/*/SKILL.md plugins/fh-commons/skills/*/SKILL.md)"
ROSTER_N="$(printf '%s' "$ROSTER" | grep -c . )"
echo "roster_entries=$ROSTER_N"
if [ "$ROSTER_N" -eq 0 ]; then
  echo "SCAN_DEAD — enumerate leg returned 0 rows (wrong cwd / missing plugins tree / glob did not match)."
  echo "Criterion ④ = UNDETERMINED. Do NOT read this as 'no collision'. Fix the cwd and re-run."
  # fail-closed: stop here, do not run leg B, do not pass ④
fi
# LEG B — hard-collision check (run ONLY when ROSTER_N > 0).
# WHOLE proposed name or a WHOLE trigger phrase reused verbatim.
# grep -wF (whole-word, fixed-string) on the full strings — NOT -E on tokens (a shared common
# word like "review" is not a collision). Exclude the asset's own file (self-match = false hit).
SELF="plugins/fh-meta/skills/PROPOSED_NAME/SKILL.md"
grep -rwF -e "PROPOSED_FULL_NAME" -e "FULL_TRIGGER_PHRASE_1" -e "FULL_TRIGGER_PHRASE_2" \
     plugins/*/skills/*/SKILL.md | grep -v "$SELF" | grep -c .

(Replace the PROPOSED_* / FULL_TRIGGER_* tokens with the literal strings under evaluation.)

Surface roster_entries + collision count + nearest existing skill(s) — the roster count is part of the report, not a private step: collision=0 is only meaningful next to a non-zero roster_entries (not-found ≠ zero). Criterion ④ then passes only if roster_entries > 0 AND 0 whole-name/ whole-trigger collision AND the judged ≤90%-overlap check agrees. roster_entries = 0 is SCAN_DEAD → ④ UNDETERMINED (fail-closed), never a pass. A verbatim whole-name or whole-trigger reuse is a hard ④ fail regardless of the LLM judgment. Honest scope: the grep grounds literal name/trigger reuse only — a post-cutoff duplicate with a paraphrased trigger is invisible to both the judge (cutoff) and the grep (literal); that residual leans on the judged half fed the enumerated descriptions above (grounded comparison, not pure memory), not on full mechanization. A shared common word is a judged-review flag, not a hard fail.

External grounding (2026-09-14 frontier signal, citation verified live before inclusion): Saha, Faghih & Feizi, "Under the Hood of SKILL.md: Semantic Supply-chain Attacks on AI Agent Skill Registry" (arXiv:2605.11418, May 2026 — https://arxiv.org/abs/2605.11418). Using real registry mechanisms (not a hypothetical), it measures SKILL.md-only attacks that manipulate which skill gets discovered (adversarial phrasing wins embedding-retrieval visibility: up to 86% pairwise win rate, 80% Top-10 placement) and selected (description-only framing biases the agent toward a functionally-equivalent adversarial variant in 77.6% of paired trials on average). This is external confirmation that the paraphrase/semantic gap named just above — the grep grounds literal reuse only — sits on a measured admission-time attack surface, not a theoretical edge case. This note does not close the residual: no new mechanized check is added here, and none should be added by an autonomous pass — a steel-quench Wave scoped to adversarial (not merely accidental) trigger/description framing is a candidate follow-up, but the gate-design decision belongs to human review (§Mechanization Boundary — judgment stays with evolution).

🟥 The paper's fourth stage is the one aimed at THIS gate, and the paragraph above left it out (caught 2026-09-14 by reading the abstract directly rather than re-reading the citing paragraph — the omission is invisible to a re-read, which is the whole argument for opening the source). Verbatim: "In Governance, semantic evasion strategies cause malicious skills to avoid a blocking verdict in 36.5%-100% of cases." Discovery and selection describe which skill gets picked; governance describes how a blocking verdict gets dodged — and an admission gate is exactly a blocking-verdict surface, so that band is a measurement of this gate's own evasion ceiling, not of a neighbouring concern. Read it against what Step 0.5 actually holds: the grep leg is literal, so semantic evasion passes it by construction, and the judged leg is a model reading descriptions — the same modality the 36.5%-100% was measured on. The honest reading is narrow in both directions: that band is a different registry, a different threat model (deliberately adversarial submissions, where this gate's common case is accidental duplication), and an upper end of 100% means some configuration blocked nothing at all, not that this gate does. What it does establish is that "semantic paraphrase evades a literal check" is measured, not speculative, on a governance surface. Still no mechanized check is added — for the reason the previous paragraph gives, unchanged.

Step 0.6. Tier-0/1 Official-Corpora Check (grounds criterion ③)

Besides the FH roster (Step 0.5), check the proposal against the official corpora: platform built-ins, claude-plugins-official, and the Claude Cookbook pattern list (platform.claude.com/cookbook — agent patterns · tools · RAG · evals · skills). A hit here is a judged ③ flag, not an automatic ③ fail: it obliges the proposal to state its governance increment over the official pattern (no-reinvention rule — an official pattern that fully covers the need outranks a net-new build; a pattern the proposal governs on top of does not fail ③). Offline fallback: note cookbook: unchecked in the Step 3 routing output rather than silently skipping. (Provenance: tracks/_audit/session_2026_07_25_claude5-context-rules-sister.md.)

Done When

Each condition declares its check class (mandatory-pass / measured / judged); every judged condition names its adversarial pairing — no judge-only path.

All steps 0–3 completed
  (mandatory-pass — each step's output block is present; a skipped step is a FAIL, not a default pass)
+ Step 0.5 scan is LIVE: roster_entries > 0
  (measured: the reported roster_entries count. 0 = SCAN_DEAD → criterion ④ UNDETERMINED and this
   skill is NOT done — a dead scan and a genuine no-collision both read 0 collisions, so the
   liveness number is what separates them)
+ Step 3 routing result output (location: FH meta-skill / local agent / drop)
  (judged — adversarial pairing: `fh-meta:challenger` re-argues the case for the destination that was
   NOT chosen, citing the 4 criteria; a routing verdict that survives the opposite case passes, an
   unopposed one does not)
+ Next action specified (write SKILL.md / create .claude/agents/ / none)
  (mandatory-pass — a literal next-action string from that enum; blank or "TBD" does not satisfy it)

Step 1. FH Meta-Skill 4-Criteria Evaluation

# Criterion Evaluation Question
① Cross-project value Is this asset equally useful in other projects without depending on a specific project?
② Orchestration / judgment layer Is it just a list of MCP/Bash calls, or a judgment layer that synthesizes multiple signals?
③ Not replaceable by built-ins Can this be equally achieved with direct MCP calls or basic bash? (If yes, fails this criterion.) A Step 0.6 official-pattern hit is a judged flag: fails ③ only if the pattern fully covers the need with no governance increment
④ No overlap with existing FH skills Step 0.5 mechanical scan = 0 name/trigger collision AND judged ≤90% overlap. Non-zero collision → hard fail.

FH suitable → ① + ④ both pass + at least one of ②③ passes. Fail → ① or ④ fails → immediate fail. Or both ②③ fail → proceed to Step 2.


Step 2. Project-Local Agent Value Assessment

For assets that failed Step 1:

# Criterion Evaluation Question
A Project-specific knowledge Does it encode paths, conventions, or domain rules specific to a project?
B Repeated workflow pattern Is it a workflow performed repeatedly within that project?
C Differentiation from built-ins Does it provide local-context-based convenience such as automatic convention application or step integration?

Local agent suitable → A satisfied + at least one of B·C. → Recommend creating {project}/.claude/agents/{name}.md.

Drop → A not satisfied or both B·C not satisfied. → Equivalent to built-in capability. No asset needed.


Step 3. Routing Result Output

[asset-placement-gate verdict]

Asset: {asset name}
Description: {one-line summary}

── FH 4-criteria ──────────────────
① Cross-project value     : O / X
② Orchestration/judgment  : O / X
③ Not replaceable         : O / X
④ No overlap with existing : O / X  ← required gate with ① (immediate fail if either fails)
→ FH suitable: Pass / Fail  (① + ④ required + at least one of ②③)

── Project-local assessment ──────  (if FH failed)
A. Project-specific knowledge : O / X
B. Repeated workflow pattern  : O / X
C. Differentiation from built-ins : O / X
→ Local agent: Suitable / Drop

── Conclusion ────────────────────
Location: [FH meta-skill | {project} local agent | Drop]
Next action: [Write SKILL.md | Create .claude/agents/ | None]

04 同仓库更多 Skills

chrono-meta
chrono-meta/forge-harness

todo-orchestrator

TODO CLI 프로젝트의 기능 추가·버그 수정·리팩터링·테스트 보강을 개발자와 QA 에이전트로 조율한다. todo 앱 관련 코드 작업 요청에 사용한다. 다시 실행, 재실행, 수정, 보완, 이전 결과 개선, QA만 다시 같은 후속 요청에도 사용한다. 단순 질문에는 사용하지 않는다.

★ 33
chrono-meta
chrono-meta/forge-harness

todo-dev

TODO CLI(todo/cli.py)에 명령을 추가하거나 버그를 고치고 테스트를 작성하는 방법. todo 앱 코드 수정·기능 추가·테스트 작성·리팩터링 작업에 사용한다. 수정, 보완, 다시 구현 요청에도 사용한다.

★ 33
chrono-meta
chrono-meta/forge-harness

todo-dev-team

TODO CLI(todo/cli.py) 개발 작업을 개발자·테스터·리뷰어 팀으로 진행하는 오케스트레이터. 명령 추가, 버그 수정, 테스트 작성, 리팩터링, 커밋·푸시 요청에 사용한다. 다시 실행, 재실행, 수정, 보완, 리뷰 반영, 테스트만 다시 같은 후속 요청에도 사용한다. 단순 질문에는 쓰지 않는다.

★ 33
chrono-meta
chrono-meta/forge-harness

todo-dev

TODO CLI 개발 작업(명령·기능 추가, 버그 수정, 테스트 작성, 리뷰, 커밋·push)을 개발자·테스터·리뷰어 에이전트 팀으로 진행하는 오케스트레이터. todo CLI, todo.json, todo/cli.py 관련 구현·수정·테스트 요청에 사용한다. 재실행, 수정, 보완, 이전 결과 개선, 테스트만 다시, 리뷰만 다시 요청에도 사용한다. 단순 질문에는 쓰지 않는다.

★ 33
chrono-meta
chrono-meta/forge-harness

weekly-summary-write

정리된 회의록으로 주간 회의 요약 문서를 작성하는 형식과 문체 규칙. 주간 요약 작성, 요약 문서 수정·보완, 검토 의견 반영 시 사용한다.

★ 33
chrono-meta
chrono-meta/forge-harness

weekly-meeting-summary

주간 회의록을 정리하고 요약 문서를 만드는 전체 작업을 조율하는 오케스트레이터. 회의록 요약, 주간 회의 정리, 주간 보고용 요약 요청 시 사용한다. 다시 실행, 재실행, 요약 수정·보완, 이전 결과를 바탕으로 한 개선, 정리만/검토만 다시 같은 후속 요청에도 사용한다.

★ 33