本文へ移動
cccskills
無料GitHub で公開

handoff-fusion-engineering

Prepare a traceable local Fusion engineering review draft with source identities, artifacts and unresolved assurance gates. Not for sending, publication or release approval.

インストール方法を見る

含まれるファイル(2)

  • SKILL.md3.0 KB
  • agents/openai.yaml221 B

SKILL.md(原文)

インストールする前に、エージェントに与えられる指示の中身を確認できます。

Use fusion_handoff_prepare to assemble the user's review evidence, and fusion_handoff_inspect to check an existing draft against current source and stored receipt state. Read the handoff contract when composing structured checks or interpreting freshness failures.

The accepted fields are title, plan_ids, document_ids, requirements, assumptions, reviewers, checks and external_evidence. Only the title is required; missing requirements, sources and checks stay explicit. Select actual producing plans to retain artifact and pending/uncertain job evidence. Do not invent unsupported job_ids, solver-result fields or approval flags.

Discover the active operation schemas and inspect the actual response before choosing check pointers. Automated handoff checks are limited to document.inspect, parameters.list, geometry.measure, geometry.check and cam.inspect. Numeric ranges require a value pointer and an exact observed unit pointer/string; there is no implicit unit conversion. Use identity assertions alongside array positions when the measured parameter or entity matters. Do not select credential fields or infer equality from redaction placeholders; rejected or redaction-altered comparisons cannot establish a result. Fixture values and response shapes do not qualify a live Fusion check.

Link requirements and reviewer IDs explicitly. Use manual procedures for unavailable solver, simulation, drawing, ECAD or engineering review work; they remain not_run. External evidence references and reported hashes are unfetched, unverified metadata. Reviewer assignments do not authenticate reviewers, notify them or record approval.

Known freshness_gaps, malformed coverage, changed source states and missing fingerprints prevent current use of dependent checks. A linked artifact must match its exact producing plan, document, preparation/completion source and implementation contract; an old-source or legacy receipt remains historical. Keep failed, pending and unknown plan outcomes visible. Do not rewrite a stored draft, drop unresolved evidence or refresh a stale assertion into a pass. Collect new observations into a new draft when the task calls for current evidence, keeping the original for comparison.

Return the local draft reference, its evidence scope and the unresolved decisions with their practical impact. A met criterion, screenshot, generated toolpath, setup sheet or provider success is not engineering, machine-collision or regulatory proof. These tools do not send the draft, publish artifacts, update enterprise systems or advance a release. Such actions require their own destination authority and qualified adapter; preparing an already authorized local draft does not require repetitive permission requests.

レビュー

まだレビューはありません。使ってみた感想をお寄せください。

同じリポジトリのスキル

概要と使いどころ

Analyze reviewed Grafana log evidence for event patterns, error-rate changes, and enrolled security markers. Use when a user needs bounded log analytics for an incident window, error-log comparison, anomaly triage, operational debugging, or a cautious check of approved security signals.

日本語の概要は準備中です。原文の説明を表示しています。

openai/community-plugins142026年10月5日 更新

Run a read-only App Store-readiness preflight for a declared iOS target without claiming approval or submitting anything.

日本語の概要は準備中です。原文の説明を表示しています。

openai/community-plugins142026年10月5日 更新

Audit whether a declared local review evaluation can support fair quality claims. Use only when the user explicitly asks to audit review-eval validity or use the bundled synthetic fixture.

日本語の概要は準備中です。原文の説明を表示しています。

openai/community-plugins142026年10月5日 更新

Deterministically score comparable imported review-run lanes with root-cause metrics and uncertainty. Use only when the user explicitly asks to benchmark normalized local review runs or use the bundled synthetic fixture.

日本語の概要は準備中です。原文の説明を表示しています。

openai/community-plugins142026年10月5日 更新

Stage and verify Fusion exports, drawing PDFs, flat-pattern DXF, viewport images and local renders. Use for engineering deliverables, not automatic publication or release.

日本語の概要は準備中です。原文の説明を表示しています。

openai/community-plugins142026年10月5日 更新

Prepare, execute and verify authorized Fusion CAD changes using resolved entities, explicit units and state-bound plans. Use for reviewed parameter, sketch, feature, assembly or configuration operations.

日本語の概要は準備中です。原文の説明を表示しています。

openai/community-plugins142026年10月5日 更新

openai のスキルをすべて見る

このスキルの問題を報告する