菜单
Sub-loader for domain specialists (agent infra, terminal, harness, perf, security, speed). Each entry captures the WHY and decision tree only — not implementation details. Trigger via `/specialists` or whenever a decision in one of these domains is on the table.
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
| name | specialists |
| description | Sub-loader for domain specialists (agent infra, terminal, harness, perf, security, speed). Each entry captures the WHY and decision tree only — not implementation details. Trigger via `/specialists` or whenever a decision in one of these domains is on the table. |
Index of domain specialists. Each file answers why to pick an approach and when — never how. Implementation details (APIs, constants, syntax) stay in code and library docs (context7).
| Specialist | When to load |
|---|---|
| mcp-design | Building MCP servers, tool granularity, MCP vs in-process |
| prompt-caching | Structuring prompts for reuse, cache breakpoints, invalidation reasons |
| agent-memory | What to persist across sessions, retrieval triggers, memory shape |
| tool-routing | Eager vs deferred tool loading, large catalogs, dispatch accuracy |
| Specialist | When to load |
|---|---|
| ratatui-arch | Structuring a ratatui app, state ownership, blocking work |
| ansi-pty | Escape sequences, capability detection, PTY semantics |
| terminal-input | Kitty protocol, mouse, bracketed paste, raw mode safety |
| Specialist | When to load |
|---|---|
| agent-sandbox | Running model-generated commands, fs/net scoping, blast radius |
| harness-hooks | Pre/Post tool hooks, SessionStart, enforcing invariants |
| parallel-subagents | Fan-out decisions, context isolation, when sequential wins |
| context-mgmt | Compaction, deferred tools, summarization triggers |
| Specialist | When to load |
|---|---|
| profiling | Picking a profiler, baselines, measurement discipline |
| cache-locality | Data layout in hot loops, SoA/AoS, false sharing |
| allocator | Arena/slab/system choice, lifetime-based reasoning |
| contention | Lock-free vs mutex vs sharding, RwLock pitfalls |
| Specialist | When to load |
|---|---|
| supply-chain | New deps, lockfile policy, audit cadence |
| secrets | Storing credentials, rotation, blast radius |
| capability-sandbox | Permission model for tools/plugins/agents |
| input-validation | Trust boundaries, parse-don't-validate |
| Specialist | When to load |
|---|---|
| startup-latency | CLI cold start, lazy init, daemonization |
| tail-latency | p99/p999 SLOs, fan-out, hedging |
| streaming-batch | Stream vs micro-batch vs full batch, backpressure |
Every specialist file follows:
No code, no constants, no syntax. If a file accumulates implementation details, refactor — that content belongs in the relevant library doc (use context7) or the source itself.
File a machine self-diagnostic report as a tagged GitHub issue. When a tool, daemon, hook, skill, or agent hangs, errors, returns garbage, or behaves wrong, open a `machine-report`-labeled issue on the machine plugin repo so every plugin user's reports funnel to the maintainer. Requires `gh` logged in. Write-side only — it logs, it does not fix. Trigger: "/report", "report a problem", "this tool is broken", "file a machine report", "log this hang/error/hiccup".
Consume the machine self-diagnostic backlog. Lists the open `machine-report` GitHub issues on the machine repo, ranks them by severity then age, fixes the single highest-priority one at its root cause, verifies green, then closes that issue. The act side of /report (which only writes). Machine-repo-only skill — not shipped to other installs. Trigger: "/resolve", "resolve a report", "triage the reports", "fix the top report", "work the reports backlog".
Second-AI review via the Codex CLI at the drill's single review point — the consolidated review of the finished diff, alongside the persona panel, before a landing is proposed. Advisory only: it returns notes and concerns for the driver and user to weigh, it never gates, blocks, merges, or edits. Skips cleanly when codex is not installed. Trigger via "/codex-review", "codex review", "second opinion from codex", or invoked by the drill skill.
The drill — the orchestrator's default driver mode. Runs grill-first: the drill and the user refine a request one question at a time until the user calls it a valid plan. From there the job runs autonomously — a plan subagent writes a markdown brief, a miner implements it on a git-fs branch inside the orchestrator's worktree, and the gate iterates until the build is green — with one consolidated advisory review (personas + codex) at the end. The drill surfaces only to land the work into main, plus on any blocker it cannot resolve. The live roster lives in the hub (claims + board), not on disk. Trigger via "/drill", "drill mode", "orchestrator mode", "background this", "spawn an agent for this", "drive this".
Promote crystallized brainstorm findings into board tickets. After a brainstorm (drill grill) settles one or more well-scoped features, this turns each into a card on the current cwd's board so the drill's plan step can iterate per ticket. The bridge between brainstorm and plan — it creates tickets, it does not plan or implement. Trigger: "/promote", "promote these", "promote to tickets", "make tickets", "turn findings into cards".
Run the single ongoing chat that resolves the open plan and implementation questions parallel agents raise. Plan/implement agents post questions to mesh and wait; the questioneer aggregates them into one prioritized queue, presents each to you for a decision, and writes the answer back so the waiting agent proceeds. It routes decisions — it does not plan, implement, or dispatch. Trigger: "/questioneer", "questioneer", "answer the questions", "resolve open questions", "what are the agents blocked on".
