القائمة
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.
التثبيت باستخدام Codex أو Claude انسخ هذا Prompt والصقه في Codex أو Claude أو مساعد آخر ليراجع صفحة Skill ويثبّتها لك.
استنادا إلى تصنيف SOC المهني
Profile and improve performance for newly written or modified performance-critical code in the Irreden Engine. Use whenever the current change touches a system tick function, a render pipeline stage, a shader, audio/video processing, math hot paths, or anywhere on the per-frame critical path. Also use when the user says "optimize", "profile", "this is slow", "find the hotspot", "check perf", "benchmark this", or "why is this frame slow". Should run BEFORE simplify when the change is performance-relevant — optimize may add profile blocks or comments that simplify would otherwise consider extraneous. Reports CPU profiler findings and, where GPU profiling infrastructure exists, GPU timer query results. Files an issue if the necessary profiling infrastructure is missing.
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.
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.
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.
| 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.