[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M7AB01YWMD60ACWBHB4Z6ECA":3,"skill-same:01M7AB01YWMD60ACWBHB4Z6ECA":68},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":16,"categories":17,"keywords":18,"publishedAt":19,"linkOnly":20,"markdownContent":21,"markdownBody":22,"frontMatter":23,"refName":25,"commitSha":26,"skillDirectory":27,"forksCount":28,"repositoryHtmlUrl":29,"repositoryHomepageUrl":6,"ownerHtmlUrl":30,"ownerBlogUrl":31,"ownerDescription":32,"updatedAt":33,"install":34,"resources":47},"01M7AB01YWMD60ACWBHB4Z6ECA","ask-colleague",null,"Ask colleague — a *different* backend\u002Fmodel than you (e.g. a local vLLM Qwen) — to take a scoped repo task off your plate, then fold its answer back. The point isn't a stronger model; it's a second, independent mind, and that diversity is the value: `ask-colleague review` gets a candid second opinion on a diff, `ask-colleague explore` gets a fresh read of an area, `ask-colleague write` hands off a small implementation, `ask-colleague feedback` grades a finished work item (the ROI loop), and `ask-colleague clean` reaps stale\u002Fcorrupt `colleague\u002F*` branches a crashed run left behind (which can break `git fetch`). Pilot a running work item with `monitor`\u002F`guide`\u002F`stop`. Reach for it REFLEXIVELY, the way you'd lean over to the teammate at the next desk — not only when asked: before you present or open a PR on a non-trivial committed diff, run `review` for a diverse second opinion; for a fresh read of an unfamiliar area whose answer is independent of your current context, run `explore`. Both are read-only — isolated in a throwaway git worktree, zero side effects to your tree\u002Fbranch — so the reflex is always safe; the side-effecting `write --apply` \u002F `write --pr` still needs the user's go-ahead. Triggers when the user says \"ask colleague\", \"ask a colleague to review\u002Fexplore\u002Fwrite this\", \"have colleague take a look\", \"get a second opinion\", \"ask the other model\", \"rate that work item\", \"clean up a crashed colleague run\" — and still on the legacy \"outsource this\". Colleague's output is a second opinion to verify and own, never authority.","agentculture\u002Fculture-nodes","agentculture","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F277824871?v=4","01M7AAZDPCFA8RDCA27DQKKWCJ","AgentCulture","01M7AAZDQ1B6FYE24021YCDAPR",3,"Apache-2.0","https:\u002F\u002Fgithub.com\u002Fagentculture\u002Fculture-nodes\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fask-colleague\u002FSKILL.md",[],[],"2026-11-15 14:19:17",false,"---\nname: ask-colleague\ntype: command\ndescription: >\n  Ask colleague — a *different* backend\u002Fmodel than you (e.g. a local vLLM Qwen) —\n  to take a scoped repo task off your plate, then fold its answer back. The point\n  isn't a stronger model; it's a second, independent mind, and that diversity is\n  the value: `ask-colleague review` gets a candid second opinion on a diff,\n  `ask-colleague explore` gets a fresh read of an area, `ask-colleague write`\n  hands off a small implementation, `ask-colleague feedback` grades a finished\n  work item (the ROI loop), and `ask-colleague clean` reaps stale\u002Fcorrupt\n  `colleague\u002F*` branches a crashed run left behind (which can break `git fetch`).\n  Pilot a running work item with `monitor`\u002F`guide`\u002F`stop`.\n  Reach for it REFLEXIVELY, the way you'd lean over to the\n  teammate at the next desk — not only when asked: before you present or open a PR\n  on a non-trivial committed diff, run `review` for a diverse second opinion; for a\n  fresh read of an unfamiliar area whose answer is independent of your current\n  context, run `explore`. Both are read-only — isolated in a throwaway git\n  worktree, zero side effects to your tree\u002Fbranch — so the reflex is always safe; the\n  side-effecting `write --apply` \u002F `write --pr` still needs the user's go-ahead. Triggers when the\n  user says \"ask colleague\", \"ask a colleague to review\u002Fexplore\u002Fwrite this\", \"have\n  colleague take a look\", \"get a second opinion\", \"ask the other model\", \"rate that\n  work item\", \"clean up a crashed colleague run\" — and still on the legacy \"outsource this\".\n  Colleague's output is a second opinion to verify and own, never authority.\n---\n\n# ask-colleague — lean on colleague as a different mind\n\n`ask-colleague` drives the **`colleague`** CLI so a Claude agent can hand a scoped\ntask to a *different* backend (default: a local vLLM `Qwen3.8-27B` on\n`:8001`). Colleague's model is **not** assumed to be stronger than you — its\nvalue is **diversity**. A second, independent mind catches things the author's\nmind glides past, which is why **review** is the headline verb. Treat it the way\nyou'd turn to the colleague at the next desk: a quick, low-ceremony ask, not a\nheavyweight handoff.\n\nThis skill is the operator: a portable wrapper that resolves the CLI and turns\neach verb into a `colleague work`, then prints the work item's result summary.\n\n## The reflex — when to ask colleague *unprompted*\n\nDon't wait to be told. The two read-only verbs have **zero side\neffects** (throwaway worktree, can't touch your tree), so reaching for them is\nalways safe — make it a habit on tasks that fit:\n\n**GO** (reach out on your own — default to the read-only verbs):\n\n- **`review` — the standing reflex.** You (or a subagent) just produced a\n  non-trivial *committed* diff and are about to present it or open a PR. Get a\n  diverse second opinion *first*. A different mind catches what the author's mind\n  glides past — that's the whole point. A real review on a local model takes tens\n  of seconds to a few minutes (it reads the diff + touched files), so **kick it\n  off in the background and keep working** rather than blocking on it; fold the\n  findings in when it lands.\n- **`explore`** — you need to understand an unfamiliar area whose answer is\n  **independent of your current context**, and you could be doing something else\n  meanwhile. Fan it out, fold the findings back.\n- A scoped, **verifiable** subtask where a fresh pass raises your confidence and\n  you can cheaply check the result.\n\n**NO-GO** (just do it yourself):\n\n- Work that needs *your* accumulated context, the user's intent, or cross-cutting\n  design judgment — a context-free second mind will drift, not help.\n- Anything **outward-facing or destructive** without a user nod: `write --apply` \u002F\n  `write --pr`, posting, deleting. The read-only verbs are the unprompted reflex;\n  side-effecting ones are not.\n- Trivial work that's faster to just do (a one-line edit) — the work item + fold-back\n  costs more than the edit.\n- Output you can't verify cheaply — if you can't check it, diversity is just noise.\n\n**Guardrails (always):**\n\n- **One-glance readiness.** `colleague whoami` names the live work engine +\n  model; if it reports `mock` or you're unsure the server is up, run `colleague\n  doctor --probe`. Don't burn time on a dead or no-op backend.\n- **Second opinion, not authority.** colleague is a *different* mind, not a\n  stronger one. Weigh its findings, verify its claims, own the decision. Diversity\n  is the value; verification is the price.\n- **Close the loop.** Occasionally `ask-colleague feedback last --rating N` so the\n  ROI of asking colleague for this *kind* of task is measurable — and you learn\n  when to stop.\n\n## How to run\n\nThe entry point is `scripts\u002Fask-colleague.sh`. Invoke it from the repo you want\ncolleague to work on:\n\n```bash\nbash .claude\u002Fskills\u002Fask-colleague\u002Fscripts\u002Fask-colleague.sh \u003Cverb> \"\u003Ctext>\" [options]\n```\n\nIt resolves the CLI portably — an installed `colleague` on `PATH` (the normal\ncase), falling back to `uv run colleague` when inside the colleague checkout,\nelse an install hint.\n\n### Verbs\n\n| Verb | What it does | Side effects |\n|------|--------------|--------------|\n| `explore \"\u003Cquestion or area>\"` | Read-only investigation of the repo; the model reads and reports findings. | **None** to your working tree \u002F branch — runs in a throwaway worktree at HEAD; writes only a gradable run artifact under the gitignored `.colleague\u002F` bookkeeping dir. |\n| `review \"\u003Cwhat to focus on>\" [--base main]` | A diverse second opinion on the **committed** diff (`\u003Cbase>...HEAD`). | **None** to your working tree \u002F branch — throwaway worktree, committed changes only; writes only a gradable run artifact under the gitignored `.colleague\u002F` bookkeeping dir. |\n| `write \"\u003Ctask>\" [--apply\\|--pr]` | Implement a change. **Previews by default** (throwaway worktree, prints the would-be diff); `--apply` lands a work branch in place; `--pr` pushes + opens a PR. | **None** to your working tree \u002F branch by default (preview); a `colleague\u002F\u003Cid>` work branch \u002F PR only with `--apply` \u002F `--pr`. |\n| `plan \"\u003Ctask>\"` | Colleague **PLANS** a complex task: it proposes a spec, then a split plan, then fans the waves out to a subagent-colleague workforce (`colleague plan run … --yes`). The *inverse of `\u002Fthink`* — same arc, but colleague is the planning mind, not Claude. Needs a live backend. | Runs `colleague plan` in `--repo` (not a throwaway worktree): the workforce stage spawns isolated subagent worktrees and can land branches — treat like `write --apply` (gets a user nod). |\n| `feedback \u003Cid\\|last> [--rating N]` | **Grade a finished work item** (the ROI loop). With `--rating N` (1–5, plus `--notes`) it records feedback; without, it shows the work item's existing feedback. `last` resolves the most recent work item in `--repo`. | Writes `.colleague\u002F\u003Cid>.feedback.json` only when `--rating` is given; read-only otherwise. |\n| `resume \u003Ctask-id\\|last> [--detach]` | **Resume a cut run** — a timed-out \u002F SIGTERM'd \u002F budget-exhausted work item picked back up from its persisted artifact (`colleague work --continue`, lineage on `TaskResult.continued_from`). `last` resolves the most recent work item in `--repo`. `--detach` runs it under `setsid`\u002F`nohup` and returns at once (not `--background`, which drops the continue id — colleague#418). | Continues the ORIGINAL run's `colleague\u002F\u003Cid>` work branch; never touches your tree \u002F branch. |\n| `clean [--dry-run]` | **Reap what a crashed run left behind** (#162): stale\u002Fcorrupt `colleague\u002F*` branches + orphaned 0-byte `.colleague\u002F` artifacts that can wedge `git fetch`. Scoped strictly to `colleague\u002F*` (never touches an unrelated branch); conservative with `.git\u002Fobjects` (reports 0-byte loose objects + suggests `git prune`, never deletes them). A thin pass-through to `colleague clean`. | Deletes corrupt `colleague\u002F\u003Cid>` refs + 0-byte `.colleague\u002F` artifacts in `--repo`; `--dry-run` changes nothing. |\n\n### Options\n\n| Option | Meaning |\n|--------|---------|\n| `--repo PATH` | Target repo (default: `.`). |\n| `--base BRANCH` | Base for the `review` diff (default: `main`). |\n| `--engine NAME` | Backend plugin (default: `$COLLEAGUE_ENGINE` or `vllm-openai`). |\n| `--model NAME` | Model (default: `$COLLEAGUE_MODEL` or `unsloth\u002FQwen3.8-27B-NVFP4`). |\n| `--base-url URL` | OpenAI base URL (default: `$COLLEAGUE_BASE_URL` or `http:\u002F\u002Flocalhost:8001\u002Fv1`). |\n| `--role NAME` | Typed subagent role for the run (`explorer`, `reviewer`, `validator`, `planner`, `writer`). Since #416 a top-level `--role explorer` runs at thinking effort **`low`** (off selectable via `--effort off`); other top-level roles keep the acting seat's effort. |\n| `--effort RUNG` | **Thinking effort for the acting seat** (#416): `off` \\| `low` \\| `medium` \\| `high` \\| `xhigh` \\| `default`. Unset = colleague's own table (acting seat `medium`). `off` sends `enable_thinking:false` — the measured win for small, well-specified briefs (same brief: off 24 s \u002F xhigh 88 s \u002F medium 129 s, all correct — `docs\u002Fevidence\u002F2026-08-22-per-seat-thinking-effort-416-results.md`); `xhigh` for open-ended judgement; `default` = the kill-switch (send nothing, the pre-#416 wire). Exported as `COLLEAGUE_CORTEX_REASONING_EFFORT` + `COLLEAGUE_WORKER_REASONING_EFFORT`; validated before the run. |\n| `--seat-effort S=R[,S=R]` | Per-seat override for the other seats (`cortex`, `worker`, `deepthink`, `senses`, `evaluator`, `design`) — e.g. `--seat-effort senses=off,deepthink=xhigh`. Exported as `COLLEAGUE_\u003CSEAT>_REASONING_EFFORT`. Children keep their role table (writer\u002Fplanner medium, reviewer\u002Fvalidator low, explorer off) unless the parent overrides per delegation. |\n| `--detach` | (`resume`) run detached and return at once; pilot with `monitor` \u002F `guide` \u002F `stop` once the log names the new flight id. |\n| `--max-steps N` | Loop step budget (default: 20). `explore`\u002F`review` select colleague's own native **`explore`**\u002F**`review`** mode profile (`colleague\u002Fprofiles.py`, applied via `colleague work --mode`) instead of a wrapper-side override — today that profile defaults to 30, since read-only mapping fans out across more files. An explicit `--max-steps N` always overrides the profile's default, in either direction. If the resolved `colleague` predates `--mode` (a stale install on `PATH`), the wrapper falls back to the old caller-side `--max-steps 30` + reserved-steps behavior so it keeps working. |\n| `--apply` | (`write`) apply the change in place (work branch) instead of previewing. |\n| `--allow-dirty` | (`write`) allow running on a dirty tree (only matters with `--apply` \u002F `--pr`). |\n| `--pr` | (`write`) push + open a PR instead of a local work branch (implies `--apply`). |\n| `--rating N` | (`feedback`) record a 1–5 quality rating for the work item. |\n| `--notes \"...\"` | (`feedback`) free-text notes stored with the rating. |\n| `--by NAME` | (`feedback`) who is grading (default: colleague's resolved identity). |\n| `--dry-run` | (`clean`) report what would be reaped without changing anything. |\n| `--json` | (any verb) machine-readable output: stdout carries **only** the result JSON, every diagnostic\u002Fdigest line goes to stderr. |\n\nThe result printed to stdout is the work item's `TaskResult.summary` (plus\n`changed_files` \u002F work branch for `write`), parsed from `colleague work\n--json`. Per-step progress streams to stderr while it runs. Pass `--json` to get\nthe raw `TaskResult` on stdout instead of the human digest (the drive verbs emit\nthe normalized `TaskResult`; `feedback` \u002F `clean` forward `--json` to colleague),\nkeeping stdout valid JSON for a machine consumer while diagnostics stay on stderr.\n\n## When to reach for which verb\n\n- **review** — the standing use. You wrote (or an agent wrote) a change and you\n  want a candid, independent pass over the *committed* diff before you trust it.\n  Treat the output as a second opinion to weigh, not a verdict.\n- **explore** — you want a fresh, unbiased read of an unfamiliar area (\"how does\n  X work here?\") without anchoring on your own assumptions.\n- **write** — a small, well-scoped implementation you're happy to delegate. It\n  **previews by default** (runs in a throwaway worktree and prints the would-be\n  diff without touching your tree); pass `--apply` to land it on a\n  `colleague\u002F\u003Cid>` work branch you can inspect, merge, or discard, or `--pr` to\n  open a PR.\n- **feedback** — *after* colleague finishes a work item, close the loop: record how\n  good it was. Every work item's artifact already carries always-on **stats** (elapsed\n  time, tokens read\u002Fgenerated, tools used, bytes written, reasoning-vs-answer\n  sizes); `feedback` adds a 1–5 quality grade. Stats say what it *cost*, feedback\n  says how *good* it was — together they let you compute the **ROI of asking\n  colleague** and decide whether to ask again (and which backend). Grade the most\n  recent work item with `ask-colleague feedback last --rating 4 --notes \"…\"`.\n- **clean** — recovery, not routine. A crashed \u002F interrupted `write --apply` can\n  leave a dangling `colleague\u002F\u003Cid>` branch pointing at half-written (0-byte)\n  objects that **breaks `git fetch` \u002F `git pull`**. Run `ask-colleague clean`\n  (or `colleague clean`) to reap it — start with `--dry-run` to see what it would\n  remove. It only ever touches `colleague\u002F*` refs and `.colleague\u002F` artifacts.\n\n## Thinking effort and resuming (#416)\n\nColleague resolves a **per-seat thinking effort** where each seat is built,\nnever per turn — the acting seat defaults to `medium`, deepthink `xhigh`,\nsenses\u002FTalker `off`, children by role (writer\u002Fplanner `medium`,\nreviewer\u002Fvalidator `low`, explorer `off`). The wrapper exposes the operator\noverrides as `--effort` (acting seat) and `--seat-effort` (any seat); a typo\nfails fast. Rule of thumb from the measurements: **`--effort off` for small,\nwell-specified briefs** (5× faster, same result), **leave the default for\nordinary work**, **`--effort xhigh` for open-ended judgement** (review of a\nsubtle diff, a plan). Effort does not rescue a module-sized brief — split the\nrequest instead (#415).\n\nA run cut by a timeout, SIGTERM or an exhausted budget is **resumable**:\n`ask-colleague resume \u003Ctask-id|last>` continues it from its artifact on the same\nwork branch; `--detach` keeps your shell free. Prefer resume over re-dispatch —\nthe continuation carries the prior steps and lineage.\n\n## Piloting a flight\n\nDispatch a drive with `--watch` (on `explore`, `review`, or `write`) to make the\nwork item watchable. While it runs you can:\n\n- **`ask-colleague monitor \u003Ctask-id>`** — watch the flight's live feed\n- **`ask-colleague guide \u003Ctask-id> \"\u003Cmessage>\"`** — send mid-flight guidance\n- **`ask-colleague stop \u003Ctask-id>`** — cooperatively ask the flight to stop\n\nControl is applied at the running loop's next turn boundary, so guidance and\nstop requests take effect on the next iteration rather than interrupting mid-step.\n\n## Hard rules (do not violate)\n\n- **explore and review are read-only.** They run in a throwaway `git worktree`\n  at HEAD, so a stray write can't reach your working tree or branch; the prompts\n  also tell the model not to modify anything. Don't route a change-making task\n  through them — use `write`.\n- **`write` previews by default; applying refuses a dirty tree.** A preview runs\n  in an isolated worktree and never touches your tree, so it is safe even when\n  dirty. `--apply` \u002F `--pr` (the in-place path) refuses a dirty tree unless you\n  pass `--allow-dirty` — this guards the dirty-tree hazard: committing\n  *uncommitted* edits onto the work branch and leaving you there. Commit or\n  stash first before applying. `--allow-dirty` is propagated to the runtime,\n  which since colleague#149 enforces the same guard directly (a bare\n  `colleague work`\u002F`drive` also refuses uncommitted *tracked* changes).\n- **Colleague's output is a second opinion, not authority.** The backend may be a\n  smaller\u002Fdifferent model; weigh its findings, verify its claims, and own the\n  decision yourself.\n\n## Honest limits\n\n- Read-only is enforced by **worktree isolation + prompt constraint**, not a\n  sandbox — the loop always exposes `write_file`\u002F`run_command`, so the model can\n  still run arbitrary *read-only* commands.\n- `review` covers **committed** changes only (`\u003Cbase>...HEAD`). To review\n  uncommitted work, commit it first.\n- The default backend is whatever single model is running locally; a multi-model\n  fleet (different model per verb) is separate infrastructure.\n- **Every verb writes bookkeeping under `.colleague\u002F`** (run artifacts for\n  explore\u002Freview\u002Fwrite; feedback records; the `last_work` pointer) — none of it\n  in your tracked tree, but in a repo that does **not** already gitignore\n  `.colleague\u002F` it shows up as untracked files. **Add `.colleague\u002F` to your\n  `.gitignore`** (keep `!\u002F.colleague\u002Fcommands\u002F` if you commit command templates).\n- **A crashed run can wedge `git fetch`.** A `write --apply` interrupted\n  mid-commit can leave a dangling `colleague\u002F\u003Cid>` branch + 0-byte artifacts;\n  `ask-colleague clean` recovers it. A SIGKILL\u002FOOM *during* the commit can still\n  corrupt git objects (git\u002Ffilesystem durability, not the skill's to guarantee)\n  — which is exactly what `clean` is for.\n\n## Provenance\n\nThis is a **first-party** colleague skill — colleague is its origin. See\n`docs\u002Fskill-sources.md` for the consumer's per-repo skill ledger. The `cite,\ndon't import` policy holds: downstream repos copy it, they don't symlink or\ndepend on it.\n","# ask-colleague — lean on colleague as a different mind\n\n`ask-colleague` drives the **`colleague`** CLI so a Claude agent can hand a scoped\ntask to a *different* backend (default: a local vLLM `Qwen3.8-27B` on\n`:8001`). Colleague's model is **not** assumed to be stronger than you — its\nvalue is **diversity**. A second, independent mind catches things the author's\nmind glides past, which is why **review** is the headline verb. Treat it the way\nyou'd turn to the colleague at the next desk: a quick, low-ceremony ask, not a\nheavyweight handoff.\n\nThis skill is the operator: a portable wrapper that resolves the CLI and turns\neach verb into a `colleague work`, then prints the work item's result summary.\n\n## The reflex — when to ask colleague *unprompted*\n\nDon't wait to be told. The two read-only verbs have **zero side\neffects** (throwaway worktree, can't touch your tree), so reaching for them is\nalways safe — make it a habit on tasks that fit:\n\n**GO** (reach out on your own — default to the read-only verbs):\n\n- **`review` — the standing reflex.** You (or a subagent) just produced a\n  non-trivial *committed* diff and are about to present it or open a PR. Get a\n  diverse second opinion *first*. A different mind catches what the author's mind\n  glides past — that's the whole point. A real review on a local model takes tens\n  of seconds to a few minutes (it reads the diff + touched files), so **kick it\n  off in the background and keep working** rather than blocking on it; fold the\n  findings in when it lands.\n- **`explore`** — you need to understand an unfamiliar area whose answer is\n  **independent of your current context**, and you could be doing something else\n  meanwhile. Fan it out, fold the findings back.\n- A scoped, **verifiable** subtask where a fresh pass raises your confidence and\n  you can cheaply check the result.\n\n**NO-GO** (just do it yourself):\n\n- Work that needs *your* accumulated context, the user's intent, or cross-cutting\n  design judgment — a context-free second mind will drift, not help.\n- Anything **outward-facing or destructive** without a user nod: `write --apply` \u002F\n  `write --pr`, posting, deleting. The read-only verbs are the unprompted reflex;\n  side-effecting ones are not.\n- Trivial work that's faster to just do (a one-line edit) — the work item + fold-back\n  costs more than the edit.\n- Output you can't verify cheaply — if you can't check it, diversity is just noise.\n\n**Guardrails (always):**\n\n- **One-glance readiness.** `colleague whoami` names the live work engine +\n  model; if it reports `mock` or you're unsure the server is up, run `colleague\n  doctor --probe`. Don't burn time on a dead or no-op backend.\n- **Second opinion, not authority.** colleague is a *different* mind, not a\n  stronger one. Weigh its findings, verify its claims, own the decision. Diversity\n  is the value; verification is the price.\n- **Close the loop.** Occasionally `ask-colleague feedback last --rating N` so the\n  ROI of asking colleague for this *kind* of task is measurable — and you learn\n  when to stop.\n\n## How to run\n\nThe entry point is `scripts\u002Fask-colleague.sh`. Invoke it from the repo you want\ncolleague to work on:\n\n```bash\nbash .claude\u002Fskills\u002Fask-colleague\u002Fscripts\u002Fask-colleague.sh \u003Cverb> \"\u003Ctext>\" [options]\n```\n\nIt resolves the CLI portably — an installed `colleague` on `PATH` (the normal\ncase), falling back to `uv run colleague` when inside the colleague checkout,\nelse an install hint.\n\n### Verbs\n\n| Verb | What it does | Side effects |\n|------|--------------|--------------|\n| `explore \"\u003Cquestion or area>\"` | Read-only investigation of the repo; the model reads and reports findings. | **None** to your working tree \u002F branch — runs in a throwaway worktree at HEAD; writes only a gradable run artifact under the gitignored `.colleague\u002F` bookkeeping dir. |\n| `review \"\u003Cwhat to focus on>\" [--base main]` | A diverse second opinion on the **committed** diff (`\u003Cbase>...HEAD`). | **None** to your working tree \u002F branch — throwaway worktree, committed changes only; writes only a gradable run artifact under the gitignored `.colleague\u002F` bookkeeping dir. |\n| `write \"\u003Ctask>\" [--apply\\|--pr]` | Implement a change. **Previews by default** (throwaway worktree, prints the would-be diff); `--apply` lands a work branch in place; `--pr` pushes + opens a PR. | **None** to your working tree \u002F branch by default (preview); a `colleague\u002F\u003Cid>` work branch \u002F PR only with `--apply` \u002F `--pr`. |\n| `plan \"\u003Ctask>\"` | Colleague **PLANS** a complex task: it proposes a spec, then a split plan, then fans the waves out to a subagent-colleague workforce (`colleague plan run … --yes`). The *inverse of `\u002Fthink`* — same arc, but colleague is the planning mind, not Claude. Needs a live backend. | Runs `colleague plan` in `--repo` (not a throwaway worktree): the workforce stage spawns isolated subagent worktrees and can land branches — treat like `write --apply` (gets a user nod). |\n| `feedback \u003Cid\\|last> [--rating N]` | **Grade a finished work item** (the ROI loop). With `--rating N` (1–5, plus `--notes`) it records feedback; without, it shows the work item's existing feedback. `last` resolves the most recent work item in `--repo`. | Writes `.colleague\u002F\u003Cid>.feedback.json` only when `--rating` is given; read-only otherwise. |\n| `resume \u003Ctask-id\\|last> [--detach]` | **Resume a cut run** — a timed-out \u002F SIGTERM'd \u002F budget-exhausted work item picked back up from its persisted artifact (`colleague work --continue`, lineage on `TaskResult.continued_from`). `last` resolves the most recent work item in `--repo`. `--detach` runs it under `setsid`\u002F`nohup` and returns at once (not `--background`, which drops the continue id — colleague#418). | Continues the ORIGINAL run's `colleague\u002F\u003Cid>` work branch; never touches your tree \u002F branch. |\n| `clean [--dry-run]` | **Reap what a crashed run left behind** (#162): stale\u002Fcorrupt `colleague\u002F*` branches + orphaned 0-byte `.colleague\u002F` artifacts that can wedge `git fetch`. Scoped strictly to `colleague\u002F*` (never touches an unrelated branch); conservative with `.git\u002Fobjects` (reports 0-byte loose objects + suggests `git prune`, never deletes them). A thin pass-through to `colleague clean`. | Deletes corrupt `colleague\u002F\u003Cid>` refs + 0-byte `.colleague\u002F` artifacts in `--repo`; `--dry-run` changes nothing. |\n\n### Options\n\n| Option | Meaning |\n|--------|---------|\n| `--repo PATH` | Target repo (default: `.`). |\n| `--base BRANCH` | Base for the `review` diff (default: `main`). |\n| `--engine NAME` | Backend plugin (default: `$COLLEAGUE_ENGINE` or `vllm-openai`). |\n| `--model NAME` | Model (default: `$COLLEAGUE_MODEL` or `unsloth\u002FQwen3.8-27B-NVFP4`). |\n| `--base-url URL` | OpenAI base URL (default: `$COLLEAGUE_BASE_URL` or `http:\u002F\u002Flocalhost:8001\u002Fv1`). |\n| `--role NAME` | Typed subagent role for the run (`explorer`, `reviewer`, `validator`, `planner`, `writer`). Since #416 a top-level `--role explorer` runs at thinking effort **`low`** (off selectable via `--effort off`); other top-level roles keep the acting seat's effort. |\n| `--effort RUNG` | **Thinking effort for the acting seat** (#416): `off` \\| `low` \\| `medium` \\| `high` \\| `xhigh` \\| `default`. Unset = colleague's own table (acting seat `medium`). `off` sends `enable_thinking:false` — the measured win for small, well-specified briefs (same brief: off 24 s \u002F xhigh 88 s \u002F medium 129 s, all correct — `docs\u002Fevidence\u002F2026-08-22-per-seat-thinking-effort-416-results.md`); `xhigh` for open-ended judgement; `default` = the kill-switch (send nothing, the pre-#416 wire). Exported as `COLLEAGUE_CORTEX_REASONING_EFFORT` + `COLLEAGUE_WORKER_REASONING_EFFORT`; validated before the run. |\n| `--seat-effort S=R[,S=R]` | Per-seat override for the other seats (`cortex`, `worker`, `deepthink`, `senses`, `evaluator`, `design`) — e.g. `--seat-effort senses=off,deepthink=xhigh`. Exported as `COLLEAGUE_\u003CSEAT>_REASONING_EFFORT`. Children keep their role table (writer\u002Fplanner medium, reviewer\u002Fvalidator low, explorer off) unless the parent overrides per delegation. |\n| `--detach` | (`resume`) run detached and return at once; pilot with `monitor` \u002F `guide` \u002F `stop` once the log names the new flight id. |\n| `--max-steps N` | Loop step budget (default: 20). `explore`\u002F`review` select colleague's own native **`explore`**\u002F**`review`** mode profile (`colleague\u002Fprofiles.py`, applied via `colleague work --mode`) instead of a wrapper-side override — today that profile defaults to 30, since read-only mapping fans out across more files. An explicit `--max-steps N` always overrides the profile's default, in either direction. If the resolved `colleague` predates `--mode` (a stale install on `PATH`), the wrapper falls back to the old caller-side `--max-steps 30` + reserved-steps behavior so it keeps working. |\n| `--apply` | (`write`) apply the change in place (work branch) instead of previewing. |\n| `--allow-dirty` | (`write`) allow running on a dirty tree (only matters with `--apply` \u002F `--pr`). |\n| `--pr` | (`write`) push + open a PR instead of a local work branch (implies `--apply`). |\n| `--rating N` | (`feedback`) record a 1–5 quality rating for the work item. |\n| `--notes \"...\"` | (`feedback`) free-text notes stored with the rating. |\n| `--by NAME` | (`feedback`) who is grading (default: colleague's resolved identity). |\n| `--dry-run` | (`clean`) report what would be reaped without changing anything. |\n| `--json` | (any verb) machine-readable output: stdout carries **only** the result JSON, every diagnostic\u002Fdigest line goes to stderr. |\n\nThe result printed to stdout is the work item's `TaskResult.summary` (plus\n`changed_files` \u002F work branch for `write`), parsed from `colleague work\n--json`. Per-step progress streams to stderr while it runs. Pass `--json` to get\nthe raw `TaskResult` on stdout instead of the human digest (the drive verbs emit\nthe normalized `TaskResult`; `feedback` \u002F `clean` forward `--json` to colleague),\nkeeping stdout valid JSON for a machine consumer while diagnostics stay on stderr.\n\n## When to reach for which verb\n\n- **review** — the standing use. You wrote (or an agent wrote) a change and you\n  want a candid, independent pass over the *committed* diff before you trust it.\n  Treat the output as a second opinion to weigh, not a verdict.\n- **explore** — you want a fresh, unbiased read of an unfamiliar area (\"how does\n  X work here?\") without anchoring on your own assumptions.\n- **write** — a small, well-scoped implementation you're happy to delegate. It\n  **previews by default** (runs in a throwaway worktree and prints the would-be\n  diff without touching your tree); pass `--apply` to land it on a\n  `colleague\u002F\u003Cid>` work branch you can inspect, merge, or discard, or `--pr` to\n  open a PR.\n- **feedback** — *after* colleague finishes a work item, close the loop: record how\n  good it was. Every work item's artifact already carries always-on **stats** (elapsed\n  time, tokens read\u002Fgenerated, tools used, bytes written, reasoning-vs-answer\n  sizes); `feedback` adds a 1–5 quality grade. Stats say what it *cost*, feedback\n  says how *good* it was — together they let you compute the **ROI of asking\n  colleague** and decide whether to ask again (and which backend). Grade the most\n  recent work item with `ask-colleague feedback last --rating 4 --notes \"…\"`.\n- **clean** — recovery, not routine. A crashed \u002F interrupted `write --apply` can\n  leave a dangling `colleague\u002F\u003Cid>` branch pointing at half-written (0-byte)\n  objects that **breaks `git fetch` \u002F `git pull`**. Run `ask-colleague clean`\n  (or `colleague clean`) to reap it — start with `--dry-run` to see what it would\n  remove. It only ever touches `colleague\u002F*` refs and `.colleague\u002F` artifacts.\n\n## Thinking effort and resuming (#416)\n\nColleague resolves a **per-seat thinking effort** where each seat is built,\nnever per turn — the acting seat defaults to `medium`, deepthink `xhigh`,\nsenses\u002FTalker `off`, children by role (writer\u002Fplanner `medium`,\nreviewer\u002Fvalidator `low`, explorer `off`). The wrapper exposes the operator\noverrides as `--effort` (acting seat) and `--seat-effort` (any seat); a typo\nfails fast. Rule of thumb from the measurements: **`--effort off` for small,\nwell-specified briefs** (5× faster, same result), **leave the default for\nordinary work**, **`--effort xhigh` for open-ended judgement** (review of a\nsubtle diff, a plan). Effort does not rescue a module-sized brief — split the\nrequest instead (#415).\n\nA run cut by a timeout, SIGTERM or an exhausted budget is **resumable**:\n`ask-colleague resume \u003Ctask-id|last>` continues it from its artifact on the same\nwork branch; `--detach` keeps your shell free. Prefer resume over re-dispatch —\nthe continuation carries the prior steps and lineage.\n\n## Piloting a flight\n\nDispatch a drive with `--watch` (on `explore`, `review`, or `write`) to make the\nwork item watchable. While it runs you can:\n\n- **`ask-colleague monitor \u003Ctask-id>`** — watch the flight's live feed\n- **`ask-colleague guide \u003Ctask-id> \"\u003Cmessage>\"`** — send mid-flight guidance\n- **`ask-colleague stop \u003Ctask-id>`** — cooperatively ask the flight to stop\n\nControl is applied at the running loop's next turn boundary, so guidance and\nstop requests take effect on the next iteration rather than interrupting mid-step.\n\n## Hard rules (do not violate)\n\n- **explore and review are read-only.** They run in a throwaway `git worktree`\n  at HEAD, so a stray write can't reach your working tree or branch; the prompts\n  also tell the model not to modify anything. Don't route a change-making task\n  through them — use `write`.\n- **`write` previews by default; applying refuses a dirty tree.** A preview runs\n  in an isolated worktree and never touches your tree, so it is safe even when\n  dirty. `--apply` \u002F `--pr` (the in-place path) refuses a dirty tree unless you\n  pass `--allow-dirty` — this guards the dirty-tree hazard: committing\n  *uncommitted* edits onto the work branch and leaving you there. Commit or\n  stash first before applying. `--allow-dirty` is propagated to the runtime,\n  which since colleague#149 enforces the same guard directly (a bare\n  `colleague work`\u002F`drive` also refuses uncommitted *tracked* changes).\n- **Colleague's output is a second opinion, not authority.** The backend may be a\n  smaller\u002Fdifferent model; weigh its findings, verify its claims, and own the\n  decision yourself.\n\n## Honest limits\n\n- Read-only is enforced by **worktree isolation + prompt constraint**, not a\n  sandbox — the loop always exposes `write_file`\u002F`run_command`, so the model can\n  still run arbitrary *read-only* commands.\n- `review` covers **committed** changes only (`\u003Cbase>...HEAD`). To review\n  uncommitted work, commit it first.\n- The default backend is whatever single model is running locally; a multi-model\n  fleet (different model per verb) is separate infrastructure.\n- **Every verb writes bookkeeping under `.colleague\u002F`** (run artifacts for\n  explore\u002Freview\u002Fwrite; feedback records; the `last_work` pointer) — none of it\n  in your tracked tree, but in a repo that does **not** already gitignore\n  `.colleague\u002F` it shows up as untracked files. **Add `.colleague\u002F` to your\n  `.gitignore`** (keep `!\u002F.colleague\u002Fcommands\u002F` if you commit command templates).\n- **A crashed run can wedge `git fetch`.** A `write --apply` interrupted\n  mid-commit can leave a dangling `colleague\u002F\u003Cid>` branch + 0-byte artifacts;\n  `ask-colleague clean` recovers it. A SIGKILL\u002FOOM *during* the commit can still\n  corrupt git objects (git\u002Ffilesystem durability, not the skill's to guarantee)\n  — which is exactly what `clean` is for.\n\n## Provenance\n\nThis is a **first-party** colleague skill — colleague is its origin. See\n`docs\u002Fskill-sources.md` for the consumer's per-repo skill ledger. The `cite,\ndon't import` policy holds: downstream repos copy it, they don't symlink or\ndepend on it.\n",{"name":5,"type":24,"description":7},"command","main","a2056e6191a4e106faf886773efed2f44ff120b2",".claude\u002Fskills\u002Fask-colleague",0,"https:\u002F\u002Fgithub.com\u002Fagentculture\u002Fculture-nodes","https:\u002F\u002Fgithub.com\u002Fagentculture","https:\u002F\u002Fagentculture.org","Agent first OSS software","2026-09-29 08:50:58",{"source":8,"skillName":5,"commands":35,"prompt":45,"zipAvailable":46},[36,39,42],{"tool":37,"command":38},"npx","npx skills add agentculture\u002Fculture-nodes --skill ask-colleague",{"tool":40,"command":41},"bunx","bunx skills add agentculture\u002Fculture-nodes --skill ask-colleague",{"tool":43,"command":44},"pnpm","pnpm dlx skills add agentculture\u002Fculture-nodes --skill ask-colleague","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Fagentculture\u002Fculture-nodes\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fask-colleague\u002FSKILL.md",true,[48,54,58,62],{"path":49,"resourceType":50,"resourceTypeDesc":51,"sizeBytes":52,"sourceHtmlUrl":53},"prompts\u002Fexplore.md","REFERENCE","参考文档",2111,"https:\u002F\u002Fgithub.com\u002Fagentculture\u002Fculture-nodes\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fask-colleague\u002Fprompts\u002Fexplore.md",{"path":55,"resourceType":50,"resourceTypeDesc":51,"sizeBytes":56,"sourceHtmlUrl":57},"prompts\u002Freview.md",2019,"https:\u002F\u002Fgithub.com\u002Fagentculture\u002Fculture-nodes\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fask-colleague\u002Fprompts\u002Freview.md",{"path":59,"resourceType":50,"resourceTypeDesc":51,"sizeBytes":60,"sourceHtmlUrl":61},"prompts\u002Fwrite.md",1086,"https:\u002F\u002Fgithub.com\u002Fagentculture\u002Fculture-nodes\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fask-colleague\u002Fprompts\u002Fwrite.md",{"path":63,"resourceType":64,"resourceTypeDesc":65,"sizeBytes":66,"sourceHtmlUrl":67},"scripts\u002Fask-colleague.sh","SCRIPT","脚本",48893,"https:\u002F\u002Fgithub.com\u002Fagentculture\u002Fculture-nodes\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fask-colleague\u002Fscripts\u002Fask-colleague.sh",[69,76,81,86,93,100],{"skillNo":70,"name":71,"nameZh":6,"description":72,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":73,"categories":74,"keywords":75,"publishedAt":19},"01M7AB6FGWR0P3SPYFX2W801D5","sample","Fixture skill proving the .qwen\u002Fskills leg carries at least one skill.","https:\u002F\u002Fgithub.com\u002Fagentculture\u002Fculture-nodes\u002Fblob\u002Fmain\u002Fadapters\u002Fqwen\u002Ftests\u002Ffixtures\u002Fconfig-repo-no-culture-yaml\u002F.qwen\u002Fskills\u002Fsample\u002FSKILL.md",[],[],{"skillNo":77,"name":71,"nameZh":6,"description":72,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":78,"categories":79,"keywords":80,"publishedAt":19},"01M7AB6E9PFCDZE193CGPM0K35","https:\u002F\u002Fgithub.com\u002Fagentculture\u002Fculture-nodes\u002Fblob\u002Fmain\u002Fadapters\u002Fqwen\u002Ftests\u002Ffixtures\u002Fconfig-repo-no-agents-md\u002F.qwen\u002Fskills\u002Fsample\u002FSKILL.md",[],[],{"skillNo":82,"name":71,"nameZh":6,"description":72,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":83,"categories":84,"keywords":85,"publishedAt":19},"01M7AB6CJ3G1ZZJD63NEZZK1JH","https:\u002F\u002Fgithub.com\u002Fagentculture\u002Fculture-nodes\u002Fblob\u002Fmain\u002Fadapters\u002Fqwen\u002Ftests\u002Ffixtures\u002Fconfig-repo-complete\u002F.qwen\u002Fskills\u002Fsample\u002FSKILL.md",[],[],{"skillNo":87,"name":88,"nameZh":6,"description":89,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":90,"categories":91,"keywords":92,"publishedAt":19},"01M7AB64WWKV9H2J7JDV5KXGPA","think","Think a vague feature idea into a buildable spec by working backwards (the idea→spec leg; drives the `devague` CLI). Start from the announcement (\"pretend it shipped\"), capture and classify claims, interrogate them with honesty conditions and hard questions, park open vagueness as a first-class object, and export a spec only once the frame *converges*. Use when the user says \"think this through\", \"spec this\", \"work backwards\", \"turn this idea into a spec\", \"announcement frame\", or \"devague\", or when a feature request is too vague to build yet. Explore scope first when the idea touches an existing codebase — the sibling `scope` skill is the optional opening leg ahead of `think`. Once a spec exports, hand off to the sibling `spec-to-plan` skill to turn it into a plan. Authored and maintained in agentculture\u002Fdevague (origin = devague); guildmaster pulls this skill from here and broadcasts it to the AgentCulture mesh — it is NOT vendored from guildmaster like the other skills here.","https:\u002F\u002Fgithub.com\u002Fagentculture\u002Fculture-nodes\u002Fblob\u002Fmain\u002F.qwen\u002Fskills\u002Fthink\u002FSKILL.md",[],[],{"skillNo":94,"name":95,"nameZh":6,"description":96,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":97,"categories":98,"keywords":99,"publishedAt":19},"01M7AB5X0FTZ081AFH3BTGQKVA","summarize-delivery","Close the loop after an `assign-to-workforce` run by turning what actually happened into an accountability artifact — planned versus actual delivery, mid-work decisions, plan drift, evidence-backed delivery claims, and remaining work. The plan the user confirmed is the contract; this skill records where execution obeyed it, where it changed, and what is genuinely safe to claim as delivered. Runs on complete, partial, AND failed runs — failure is reported faithfully, never smoothed over. Use when the user says \"summarize delivery\", \"delivery summary\", \"wrap up\", \"close the loop\", \"what did we actually ship\", or \"plan versus actual\", or after `assign-to-workforce` merges (or fails to merge) a plan's waves. Authored and maintained in agentculture\u002Fdevague (origin = devague); guildmaster pulls this skill from here and broadcasts it to the AgentCulture mesh — it is NOT vendored from guildmaster like the inbound skills here.","https:\u002F\u002Fgithub.com\u002Fagentculture\u002Fculture-nodes\u002Fblob\u002Fmain\u002F.qwen\u002Fskills\u002Fsummarize-delivery\u002FSKILL.md",[],[],{"skillNo":101,"name":102,"nameZh":6,"description":103,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":104,"categories":105,"keywords":106,"publishedAt":19},"01M7AB5NA0QY9NN076N24NJPJQ","spec-to-plan","Turn a converged devague spec into a buildable plan by working forwards (the spec→plan leg; drives the `devague plan` CLI group). Seed a plan from a converged frame, add tasks that collectively cover every coverage target (the frame's confirmed claims + honesty conditions), give each task acceptance criteria and an honest dependency order, park genuine unknowns as first-class risks, and export a plan only once it *converges*. Use when the user says \"spec to plan\", \"stp\", \"turn this spec into a plan\", \"plan this spec\", \"make a build plan\", or after the `think` skill exports a spec. Authored and maintained in agentculture\u002Fdevague (origin = devague); guildmaster pulls this skill from here and broadcasts it to the AgentCulture mesh — it is NOT vendored from guildmaster like the other skills here.","https:\u002F\u002Fgithub.com\u002Fagentculture\u002Fculture-nodes\u002Fblob\u002Fmain\u002F.qwen\u002Fskills\u002Fspec-to-plan\u002FSKILL.md",[],[]]