fleet-harness

Diagnose over-abstraction in a package, then consolidate files by domain and deduplicate functional logic.

git worktree 생성과 제거 라이프사이클을 하나의 절차로 통합해 실행하는 스킬입니다.

Per-carrier request-block contracts and dispatch composition rules for carrier_dispatch requests. Load before the first dispatch of the session; skip reloading if already in context.

Host-owned Fleet Plan authoring and mutation workflow. Load before the first host plan_write call in a session and skip reloading when already in context. Use it for the required Plan template, optional Nimitz Plan assurance, Genesis Plan-driven handoff, host completion marking after artifact inspection, and Plan-state verification boundaries.

Drive a headless real-browser end-to-end test or diagnosis of the Fleet Console web UI with agent-browser. Use for Console SPA behavior such as blank screens, terminal rendering, Theater or Operation interactions, modal and keyboard boundaries, responsive layout, browser errors, HTTP requests, WebSockets, or final browser verification after a Console change. Do not use for Electron lifecycle, native menus or dialogs, Desktop security policy, packaging, or installed-app behavior; use desktop-e2e instead.

Use the coordinated Fleet protocol mode for multi-carrier or parallel ownership work.

Use the normal Fleet protocol mode for bounded operational work without downward-guard triggers.

Use the risk-controlled Fleet protocol mode for irreversible, structural, multi-module, or prompt-policy work.

End-to-end PR lifecycle on sbluemin/fleet-harness — commit staged work, open a PR, add its PR-number changelog fragment in a second commit, await Codex review, adversarially adjudicate context-incomplete feedback against frozen product intent, roll back review-driven scope drift, and auto-merge only after a final context audit. Supersedes the former pr-creates and pr-review-fixes skills.

릴리스 버전을 업데이트하고 changelog fragment를 컴파일하는 절차를 정의합니다.

Load before reading or interpreting any Fleet Wiki entry or raw source, before any wiki_* tool call, before staging a Fleet Wiki entry (wiki-create or wiki-update), or before adjudicating a wiki_patch_queue entry. If this skill cannot be loaded, do not interpret Wiki content, call Wiki tools, stage Fleet Wiki entries, or adjudicate patches. Defines Fleet Wiki trust, routing, ACL, and approval policy; the host performs all Fleet Wiki operations directly; load once per session and skip reloading if already in context.

Register a Fleet Wiki PRD entry that captures cognitive-debt-resolving decision history. NOT a code-grounded document. NOT a planning document. Records the WHY of a decision so future readers do not have to reverse-engineer it from code and git log.

Periodically audit the whole fleet-console visual surface (core client styles plus every fleet-plugins CSS) for design drift — elements that visually stick out from the skeleton — by sweeping the code with concrete detector patterns, measuring the live product across all three themes, classifying findings against the design charter (channel inversion / chroma jump / theme invariance / control-grammar drift), routing material redesigns through product-proposal and pure conformance fixes to direct approval, then implementing token-first with the design contract test co-updated and shipping via pr-workflow. Use for a recurring product-wide design health check, after a large visual feature lands, or whenever the user reports that parts of the UI feel disjoint or dated; not for building a new visual feature from scratch (product-proposal) or fixing a single known CSS bug (direct fix).

Run and diagnose headed end-to-end tests of the Fleet Console Electron Desktop shell. Use for entry-to-Console handoff, BrowserWindow security and navigation policy, native lifecycle, menus, dialogs, tray behavior, Desktop-owned sidecar and logs, reload or relaunch behavior, packaged artifacts, signing boundaries, or platform-specific Desktop verification. Use console-e2e for behavior wholly inside the Console SPA after handoff.

Resolve decision-shaped blocking gaps one at a time — Context Confidence gate failures during an active protocol, or pre-engagement requirements ambiguity routed by the Command Integrity Standing Order.

Use the compact Fleet protocol mode for simple, reversible, single-surface work.

After substantive work, run a compounding-engineering learning-harvest loop — retrospect the task, extract reusable learnings under a three-condition gate (recurrence, cost, generality), route each to its single best persistent home (an AGENTS.md rule, a new or edited skill, a test, a wiki-history ADR, persistent memory, or a skill's gotchas), encode it as "symptom -> action -> why", and prune weak or stale rules. Propose candidates for the Admiral of the Navy's approval and encode only approved ones, so the next task starts from a higher baseline. Use at the end of substantive work, when friction is detected, or on a periodic pruning pass. This is a meta-skill that produces other skills and assets, not product code.

From a PO/PD lens, take a fleet-console product-improvement or new-feature request, measure the current behavior live through the console-e2e skill to pin it as fact, separate the user's named solution (surface) from the real Job-to-be-done (essence), derive UX options scored on fixed trade-off axes, then build an interactive mock in the product's real design tokens via the frontend-design skill, and report a decision-ready proposal. Use whenever a fleet-console feature request, issue, or complaint is vague — or names a specific fix that needs essence-checking — and must become a decision-ready proposal grounded in measurement and a visual mock rather than guesswork. Implementation is out of scope (hand off to genesis/pr-workflow).

별도 워크트리 또는 현재 체크아웃된 브랜치의 토픽 브랜치를 canary 기준으로 리베이스하는 절차를 정의합니다. 기본적으로 충돌은 양측(both-sides) 의도를 통합해 스스로 해결하고 사후 검증하며, 해결 불가하거나 검증 실패 시에만 에스컬레이션합니다.

AGENTS.md 파일이 있는 모든 디렉토리에 CLAUDE.md 심볼릭 링크를 생성하는 절차를 정의합니다.

현재 세션의 캐리어 오케스트레이션 실사용량을 실측 집계하고, 제독 단독 수행 가정 대비 In/Output 토큰·컨텍스트 윈도우 절감 효과를 정량 보고하는 스킬입니다.