菜单
Take an approved architect plan and file it as the fleet expects: umbrella issue labeled fleet:epic, one child fleet:task per phase, per-ticket plan files at ~/.fleet/plans/issue-<N>.md, and post-filing stack validation.
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。
基于 SOC 职业分类
Review an open GitHub PR on the Irreden Engine repo and post a structured review. Use when the user says "review PR <N>", "review #<N>", "review the last PR", "check my PR", "review any new PRs", "review the PR queue", "check for new PRs to review", "review the open PRs", or otherwise asks for a code review — and also invoke automatically from a persistent reviewer-loop session when `gh pr list` surfaces a PR that has not yet been reviewed by this fleet. Designed to be invoked by a dedicated reviewer agent (dispatched into a pool worktree) whose job is to look at another agent's work with a fresh context. Writes a structured review covering ownership, ECS invariants, allocation hot paths, naming, tests, and project-specific smells, then posts it as a PR comment.
Polish the dirty working tree before committing — catch Irreden-Engine smells that a reviewer would otherwise flag. Use whenever you're about to commit, after addressing review feedback, after amending a commit, or whenever the user says "simplify", "clean up", "polish", "self-review", or "review my changes". Also auto-invoked by commit-and-push before the commit message is drafted. The skill finds per-entity getComponent in tick functions, allocation in hot loops, ECS naming convention slips, opportunities to reuse existing helpers, dead code, debug logs left behind, tautological comments, triple-nested loops over voxel volumes, renderer leaks from creation code, CPU-side SDF grid evaluation, linear-search in save/load hot paths, and stale or drifting CLAUDE.md / role / skill docs. Dispatches a parallel reuse-detection subagent fan-out so the reuse pass runs concurrently with the main checks. Applies safe fixes inline and reports anything that needs human judgment. Saves a review-fix-rereview round-trip.
Stage, commit, push a feature branch, and open a GitHub PR against master for the Irreden Engine repo. Use whenever the user says "commit", "commit my changes", "commit and push", "open a PR", "make a PR", "wrap up this chunk", or otherwise indicates the current slice of work is ready for review. The skill assumes the parallel-agent workflow where work happens on short-lived feature branches, another agent reviews the PR, and the user merges via the GitHub UI. NEVER commit directly to master.
Process the backlog of `fleet:needs-<host>-smoke` PRs on the current host. Builds a representative demo cohort against current `origin/master`, runs each with `--auto-screenshot`, and on green batch-swaps `fleet:needs-<host>-smoke` → `fleet:verified-<host>` on every eligible merged PR. On red (build break), files / opens a fix PR for the offender and holds labels on PRs that touch the implicated source paths until the fix lands. On partial (per-demo runtime hang or crash but build green), sweeps the rest, then fix-forwards the offending demo this session (issue only as the documented fallback). Runs on Linux, macOS, AND native Windows (MSYS2 mingw64) — Windows is now the primary validation host. Use when the human cues "platform-catchup", "catch up smoke tests", or "catch up Windows builds". Skill is **cue-only**, never auto-run — multi-target builds and ~40-demo runs are expensive, the human chooses when to spend them.
Build, run, and visually evaluate a rendering demo in an automated loop. Captures screenshots across a demo's configured shot list (zoom, camera offset, render-mode combinations), reads each image, diagnoses rendering issues against a topic-indexed reference (trixel / SDF shapes, lighting, backend parity), applies fixes, and repeats. Use whenever iterating on render pipeline bugs, shape alignment, lighting, parity drift, or any visual regression in the engine.
Automated pass/fail render-regression harness for Irreden Engine demos. Wraps build → run → capture → compare → report: builds a demo, runs it with `--auto-screenshot`, then compares each shot against a committed reference image using a stdlib-only Python comparator. Reports per-shot pass/fail with diff images written for mismatches. Use after any render pipeline change to catch visual regressions automatically, or as a sanity pass on `master` before cutting releases.
| name | file-epic |
| description | Take an approved architect plan and file it as the fleet expects: umbrella issue labeled fleet:epic, one child fleet:task per phase, per-ticket plan files at ~/.fleet/plans/issue-<N>.md, and post-filing stack validation. |
The flow lives in docs/agents/skills/file-epic.md.
Read it first, then apply the engine deltas below. This wrapper carries
deltas only — see docs/design/skill-sharing.md
for why.
Placement (per #1312): this skill is repo-tracked as a project skill in each repo's
.claude/skills/file-epic/. It loads by cwd, exactly likecommit-and-push/review-pr. The shared flow is now single-sourced indocs/agents/skills/file-epic.md; only the deltas below are per-repo, so the engine and game wrappers can no longer drift on the flow itself.
| Delta key | Engine value |
|---|---|
| repo | jakildev/IrredenEngine (engine); pass --repo game for jakildev/irreden |
| epic label | fleet:epic |
| task label | fleet:task |
| architect plans dir | ~/.claude/plans/<slug>.md |
| plans dir | ~/.fleet/plans/issue-<N>.md |
| repo-side plan path | <repo>/.fleet/plans/issue-<N>.md — umbrella plan committed via the step-6.5 docs PR; each child plan committed as the first commit of its own impl PR (#1932); no T-<NNN> rename |
| validate-stack command | fleet-validate-stack <umbrella> (add --repo game for the game repo) |
| title area vocabulary | engine, render, engine/voxel, game, etc. (the same scope vocabulary commit-and-push uses) |
fleet-validate-stack helper (shipped #1317) is the validate-stack
command; it auto-discovers children and fails loudly on a malformed
body. It is the belt that the step-5 hand-filing convention is the
suspenders for.## Plan comments (flow step 6) must meet
PLANNING-PROTOCOL.md
step-2 rigor — verified current state / confirmed repro, one picked
approach, sibling + in-flight reconciliation — not a restated phase line
(#1456). The engine's fleet-queue-ingest plan gate (#1932 PR2) keys on the
## Plan comment, so a child is queue-ready as soon as its comment is posted;
its plan file lands cross-host as the first commit of the child's impl PR. The
step-6.5 docs PR carries only the umbrella plan (+ steward ledger) — land it
promptly so the steward reads it from master.jakildev/IrredenEngine.
The cross-repo info-isolation rule (see
docs/agents/CLAUDE-BASELINE.md
§"Cross-repo information isolation") means engine child issues must not
reference the private game repo by name or feature; scrub a game-authored
plan before filing engine children.