[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M74FTVVHVDTMTP0EX4W924NK":3,"skill-same:01M74FTVVHVDTMTP0EX4W924NK":86},{"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":31,"commitSha":32,"skillDirectory":33,"forksCount":34,"repositoryHtmlUrl":35,"repositoryHomepageUrl":6,"ownerHtmlUrl":36,"ownerBlogUrl":37,"ownerDescription":38,"updatedAt":39,"install":40,"resources":53},"01M74FTVVHVDTMTP0EX4W924NK","work-review",null,"Reviews a single tracked issue's pushed implementation across GitHub (gh), GitLab (glab), Linear (MCP) or local issue files as a fresh, independent agent — reads the issue's requirements and the pushed diff, adversarially checks whether the work is correct and complete, and writes a verdict that routes the issue to done (accepted), changes-requested (feedback, back to the implement loop), needs-human (escalation), or blocked. It never implements or fixes anything — review only. It also applies a human's verdict on a needs-human issue. Tracker, labels and the round cap come from the committed config (.tituskirch-skills.json). Use when the user wants to review, check, or sign off one specific issue's AI work, mentions an issue awaiting review, or says things like \"review issue 42\", \"check the work on issue 42\", \"reviewe Issue 42\", or gives a verdict on an escalated issue.","TitusKirch\u002Fskills","TitusKirch","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F16943133?v=4","01M74FN1AAM5K2NH5EG292B2HZ","Titus Kirch","01M74FN1AX7TX6RAECBHE5PMNT",3,"MIT","https:\u002F\u002Fgithub.com\u002FTitusKirch\u002Fskills\u002Fblob\u002Fmain\u002Fskills\u002Fwork\u002Fwork-review\u002FSKILL.md",[],[],"2026-11-11 15:11:14",false,"---\nname: work-review\nmetadata:\n  summary: Reviews one issue's pushed work as an independent agent — verdict routes to done, changes, needs-human, or blocked.\ndescription: Reviews a single tracked issue's pushed implementation across GitHub (gh), GitLab (glab), Linear (MCP) or local issue files as a fresh, independent agent — reads the issue's requirements and the pushed diff, adversarially checks whether the work is correct and complete, and writes a verdict that routes the issue to done (accepted), changes-requested (feedback, back to the implement loop), needs-human (escalation), or blocked. It never implements or fixes anything — review only. It also applies a human's verdict on a needs-human issue. Tracker, labels and the round cap come from the committed config (.tituskirch-skills.json). Use when the user wants to review, check, or sign off one specific issue's AI work, mentions an issue awaiting review, or says things like \"review issue 42\", \"check the work on issue 42\", \"reviewe Issue 42\", or gives a verdict on an escalated issue.\nallowed-tools:\n  - Bash\n  - Read\n  - Grep\n  - Glob\n---\n\n# work-review\n\nTake **one** issue that an implementer pushed and **review it** — the stateless review-unit behind `work-review-queue`, and the deliberate counterweight to `work-implement`. A **different** agent, with **fresh context**: it did not build this and carries no \"it works because I wrote it\" bias. State lives in the issue's **lifecycle label**; the review is **read-only** and idempotent, so a crashed review just re-runs.\n\nThis skill is the **review half** of the two-loop workflow. It **never implements, edits, commits, or merges** — its only outputs are a **verdict** (a label move plus a comment): `done`, `changes-requested`, `needs human`, or `blocked`.\n\n**Opted out?** If the repo config sets `work` to `false`, all `work-*` skills are **disabled** — stop and tell the user they are turned off in `.tituskirch-skills.json`. Check `.work == false` on the resolved config before any action. A missing `jq` or config exits non-zero too, so a pass is not evidence the config was read.\n\n## Workflow\n\n### 1. Load config & resolve tracker\n\nResolve `.tituskirch-skills.json` via [`templates\u002Fresolve-config.sh`](templates\u002Fresolve-config.sh), never by reading the raw file ([REFERENCE.md](REFERENCE.md#reading-the-config) states how, missing `jq` included); the `work.*` section holds tracker, labels, `work.review.maxRounds` (default 3) and `work.feedback` — where the verdict's comment is written ([feedback recipes](REFERENCE.md#feedback-recipes); it defaults from `work.branch`, the full rule being **Feedback destination** in `work-implement`'s REFERENCE). Resolve the tracker (`work.tracker`, falling back to `issue.tracker`); reuse the `issue` catalog cache. Config + mechanics: [REFERENCE.md](REFERENCE.md).\n\n### 2. Resolve the target issue\n\n- **Explicit** — an id\u002Fnumber\u002Fkey (for example `work-review 42` or `ENG-123`).\n- **Self-select** — none given → the [selection query](REFERENCE.md#selection-query): the single highest-priority issue in `reviewRequested`. None → say so and stop.\n\n### 3. Read the state → pick the action\n\n- **`reviewRequested`** → claim it (step 4) then review the pushed work (steps 5–7). The normal path.\n- **`reviewing` assigned to this runner** → a previous review leased it and crashed; **resume** — review is read-only and idempotent, so just re-run steps 5–7 (no re-claim needed). `reviewing` held by a **different** runner → in-flight elsewhere, not ours.\n- **`needs human` + a human verdict this session** → apply the human's call: \"looks good\" → `done`; feedback → `changes-requested` (record the feedback). This is the one place a human's word resolves an escalation.\n- **anything else** (`ready` \u002F `working` \u002F `changes-requested` \u002F `done` \u002F `blocked`) → not ours; nothing to do.\n\n### 4. Claim the review (lease) — when `reviewing` is configured\n\nIf `work.labels.reviewing` resolves to a **label string**, flip `reviewRequested → reviewing` and assign the issue to the runner **before** reading the diff — on `github` and `linear`, the tracker-global claim that stops a second clone reviewing the same issue and writing a competing verdict (the review lock only proves no live reviewer **in this checkout**). **On `local` it is not tracker-global** — the store is a directory inside the checkout, so the lease reaches no further than the lock does and the competing-verdict hazard stays open; the flip still works and is still worth having within a checkout, but read **The `reviewing` lease** in REFERENCE.md before enabling `labels.reviewing` there (its recommendation is to leave it off). Honour the review single-flight lock (the **Lease & race rules** and **The single-flight lock** in `work-implement`'s REFERENCE) — take it (direct run) or run under the drain's (queue).\n\n**When `reviewing` is `false`\u002Funset (the default), skip this step entirely** — there is no lease label, so the review loop behaves exactly as before: the lock alone, reviewing straight off `reviewRequested`. The verdict at step 7 then moves the label off `reviewRequested`; when the lease is on, it moves off `reviewing`. Either way the verdict label move clears the in-flight state.\n\n### 5. Gather the work\n\n- **Requirements** — re-read the issue **body** (what was asked) and any **prior review feedback** on the issue **and** the PR (so a re-review checks the last round was addressed). Look in both regardless of `work.feedback`: it routes where feedback is _written_, never where it is read, and rounds written before the mode changed stay where they were posted.\n- **The pushed diff** — the artifact to review, scoped to _this_ issue:\n  - **A pull\u002Fmerge request present** (`worktree`) → that request's own diff, in the forge's spelling ([recipes](REFERENCE.md#fetching-the-diff)).\n  - **Neither present** (`branch:\u003Cname>`, e.g. `branch:dev`) → the issue's own commit range on the branch (the commits referencing this issue since it was last picked up). See [review-after-land](REFERENCE.md#review-after-land).\n- **Checks** — the pushed head must be green on the repo's **own** gate, and this review **establishes** that rather than inheriting it. Two sources, in this order:\n  - **The forge's checks**, but only where the head's base actually triggers them. Read which workflows the base runs _before_ reading their result — an empty or irrelevant check list is `unknown`, never green. A **draft** PR is the expected case, not an anomaly: the implement loop opens every PR as a draft and a repo gating its CI on that has deliberately run nothing yet, so this source is normally `unknown` here. CI is not waited for at this step — it is asked for at step 7, once the verdict is otherwise `done` ([marking ready, then waiting for CI](REFERENCE.md#marking-ready-then-waiting-for-ci)).\n  - **The repo's `verify`, run here.** Under `branch:\u003Cname>` there is no PR and CI commonly never ran at all, so this is the only source that exists. Run it against the pushed head in a **throwaway worktree** — the review is read-only towards the user's tree, which is not the same as running nothing ([recipe](REFERENCE.md#verifying-the-pushed-head)).\n\n  **The implementer's own green run is not a source.** It proves a tree passed _before_ the push, and on a shared branch that tree has since moved. Re-running is the whole point of a second agent: same gate, different tree, no inherited verdict.\n\n  Red, or a gate that could not be run at all → a review finding, never a pass.\n\n### 6. Review — adversarially\n\nYou are the skeptic. Judge, in this order:\n\n- **Does it do what the issue asked?** Match the diff against the body's requirements — nothing missing, nothing out of scope.\n- **Is it correct?** Look for real defects: wrong logic, edge cases, broken references, secrets, regressions.\n- **Is a human needed?** Apply the [escalation policy](REFERENCE.md#escalation-to-needs-human) — sensitive surfaces (via CODEOWNERS \u002F branch protection), ambiguous intent you cannot confirm, a change too large\u002Frisky to sign off confidently, or checks red without a clear cause.\n\n### 7. Verdict — one label move + a comment\n\nCount the review rounds first — the number of times this issue has entered `reviewRequested` ([recipe](REFERENCE.md#round-count)) — and compare to `work.review.maxRounds`. A count that could not be read is **not** zero rounds: escalate to `needs human` rather than let the loop run uncapped. Then:\n\n| Verdict                 | When                                                                                                           | Action                                                                                                                                                                                                  |\n| :---------------------- | :------------------------------------------------------------------------------------------------------------- | :------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ |\n| **`done`**              | correct, complete, low-risk                                                                                    | set `done` — accepted                                                                                                                                                                                   |\n| **`needs human`**       | correct but **risky\u002Fsensitive**, OR you cannot confidently judge, OR round ≥ `maxRounds` and still not passing | set `needs human` + comment why a human is needed                                                                                                                                                       |\n| **`changes-requested`** | fixable problems, round \u003C `maxRounds`                                                                          | **post the feedback** at the configured destination (a comment on the pull\u002Fmerge request, or an issue\u002FLinear comment referencing the commit), then set `changes-requested` — back to the implement loop |\n| **`blocked`**           | broken beyond a fixable change \u002F a hard human call                                                             | set `blocked` + comment                                                                                                                                                                                 |\n\n**A `done` is the moment CI is worth spending, so it is spent before the verdict is written.** Under `worktree` the implement loop leaves every PR a **draft** (**The draft gate** in `work-implement`'s REFERENCE), which on a repo gating its workflows on that means no CI has run. So where the review above lands on `done` and the PR is still a draft: **mark it ready for review**, then wait for the checks the base actually triggers, bounded by `work.review.timeout` (default 600s).\n\n**Two things decide what that wait reads, and both are easy to get subtly wrong.** A **`skipping` check is never a pass** — a repo gating its _jobs_ on the draft state reports one such row per gated job rather than reporting nothing, so counting them accepts a head CI never analysed; drop them and read only the jobs that answered. And the budget starts at the **later** of the un-draft and this head's own commit — a base that exists on _every_ round, including one answering a CI failure on a PR that is already ready, where nothing un-drafts and the un-draft event alone would put the deadline in the past before the first poll. That base bounds the wait only; it never decides which results count.\n\n| CI, after the un-draft                       | Verdict                                                                                                  |\n| :------------------------------------------- | :------------------------------------------------------------------------------------------------------- |\n| **green**                                    | `done` — the verdict the review already reached, now paid for                                            |\n| **red**                                      | `changes-requested` with the failure as the feedback — or `needs human` at `maxRounds`, as any red would |\n| **still running when the timeout elapses**   | `needs human` — no verdict rests on a gate that has not reported                                         |\n| **none triggered at all** (not \"none _yet_\") | `done` stands: step 5's second source already ran the repo's own `verify` against this head              |\n\n**Every other verdict leaves the draft alone, and nothing ever re-drafts.** A review that already wants changes routes to `changes-requested` without un-drafting, so no CI is spent on it; once a PR has been ready it stays ready, CI failures included, so the following rounds get CI feedback directly instead of going dark again. Recipe, and the `unknown`-versus-`no checks` distinction the fourth row turns on: [marking ready, then waiting for CI](REFERENCE.md#marking-ready-then-waiting-for-ci).\n\n**The label move always lands on the issue; the comment goes where `work.feedback` says** — the PR's thread (`pr`) or the issue's comments (`issue`), so a three-round review does not turn the requirement into a log ([recipes](REFERENCE.md#feedback-recipes)). In `pr` mode the **comment** is the primitive that carries every verdict — `gh pr comment` on GitHub, `glab mr note` on GitLab. GitHub's `gh pr review --request-changes` is an optional upgrade it **refuses on a self-authored PR**, which is the normal case wherever both loops run as one identity; **GitLab has no second verb to upgrade to at all**, so the note is the whole story there. Two fallbacks, both reported in the run: **no pull\u002Fmerge request to write to** → post to the issue; **the review call refused** → post the same body as a comment.\n\nReport the verdict and the reasoning. Inside a `work-review-queue` drain, return the verdict and let the drain move on.\n\n## Guardrails\n\n- **Review only — never implement.** No `Edit`\u002F`Write`\u002Fcommit\u002Fmerge\u002Fpush. If the fix is obvious, describe it in the feedback; do not apply it. The implement loop applies it.\n- **Marking a PR ready for review is this loop's, and only on an accept.** It changes no code and merges nothing — it is the verdict said in the forge's own vocabulary, which is why the read-only rule above does not forbid it. Flip it only where the review would otherwise write `done`, **never** as housekeeping, and **never back into draft** on any verdict, CI failures included ([marking ready, then waiting for CI](REFERENCE.md#marking-ready-then-waiting-for-ci)).\n- **A different agent than the implementer.** Review as a fresh skeptic; when in doubt, `needs human` — never rubber-stamp.\n- **Lease before review, when `reviewing` is configured.** Claim `reviewRequested → reviewing` + assign before reading the diff, so a second clone cannot review the same issue and write a competing verdict — the review loop's counterpart of the implement lease. With `labels.reviewing` off (the default), the review lock alone applies: today's behaviour, unchanged.\n- **Read-only and idempotent.** A crashed review re-runs; posting feedback, dedupe against feedback you already left.\n- **`done` is a real acceptance**, given by AI review (low-risk) or by a human (via `needs human`). Never set `done` on a change you could not confidently review — escalate instead.\n- **The round cap is a floor for the human, not a ceiling on quality** — at `maxRounds`, escalate to `needs human`; never quietly accept unfinished work to end the loop.\n- **Attribution-free & secret-free** — no `Generated with`\u002F🤖 line, no agent self-naming in comments; scan the diff for secrets and flag them.\n\n## Reference\n\nConfig, the selection query, the round-count recipe, the escalation policy, review-after-land and the feedback recipes — everything a run needs whatever its tracker: [REFERENCE.md](REFERENCE.md). The implement half and the shared lifecycle: `work-implement`.\n\n**One file per tracker recipe, and a repo reads exactly one.** `work.tracker` (falling back to `issue.tracker`) is a single value, so step 1 settles the branch before any recipe is opened: [`trackers\u002Fgithub.md`](trackers\u002Fgithub.md), [`trackers\u002Fgitlab.md`](trackers\u002Fgitlab.md), [`trackers\u002Flinear.md`](trackers\u002Flinear.md), [`trackers\u002Flocal.md`](trackers\u002Flocal.md).\n","# work-review\n\nTake **one** issue that an implementer pushed and **review it** — the stateless review-unit behind `work-review-queue`, and the deliberate counterweight to `work-implement`. A **different** agent, with **fresh context**: it did not build this and carries no \"it works because I wrote it\" bias. State lives in the issue's **lifecycle label**; the review is **read-only** and idempotent, so a crashed review just re-runs.\n\nThis skill is the **review half** of the two-loop workflow. It **never implements, edits, commits, or merges** — its only outputs are a **verdict** (a label move plus a comment): `done`, `changes-requested`, `needs human`, or `blocked`.\n\n**Opted out?** If the repo config sets `work` to `false`, all `work-*` skills are **disabled** — stop and tell the user they are turned off in `.tituskirch-skills.json`. Check `.work == false` on the resolved config before any action. A missing `jq` or config exits non-zero too, so a pass is not evidence the config was read.\n\n## Workflow\n\n### 1. Load config & resolve tracker\n\nResolve `.tituskirch-skills.json` via [`templates\u002Fresolve-config.sh`](templates\u002Fresolve-config.sh), never by reading the raw file ([REFERENCE.md](REFERENCE.md#reading-the-config) states how, missing `jq` included); the `work.*` section holds tracker, labels, `work.review.maxRounds` (default 3) and `work.feedback` — where the verdict's comment is written ([feedback recipes](REFERENCE.md#feedback-recipes); it defaults from `work.branch`, the full rule being **Feedback destination** in `work-implement`'s REFERENCE). Resolve the tracker (`work.tracker`, falling back to `issue.tracker`); reuse the `issue` catalog cache. Config + mechanics: [REFERENCE.md](REFERENCE.md).\n\n### 2. Resolve the target issue\n\n- **Explicit** — an id\u002Fnumber\u002Fkey (for example `work-review 42` or `ENG-123`).\n- **Self-select** — none given → the [selection query](REFERENCE.md#selection-query): the single highest-priority issue in `reviewRequested`. None → say so and stop.\n\n### 3. Read the state → pick the action\n\n- **`reviewRequested`** → claim it (step 4) then review the pushed work (steps 5–7). The normal path.\n- **`reviewing` assigned to this runner** → a previous review leased it and crashed; **resume** — review is read-only and idempotent, so just re-run steps 5–7 (no re-claim needed). `reviewing` held by a **different** runner → in-flight elsewhere, not ours.\n- **`needs human` + a human verdict this session** → apply the human's call: \"looks good\" → `done`; feedback → `changes-requested` (record the feedback). This is the one place a human's word resolves an escalation.\n- **anything else** (`ready` \u002F `working` \u002F `changes-requested` \u002F `done` \u002F `blocked`) → not ours; nothing to do.\n\n### 4. Claim the review (lease) — when `reviewing` is configured\n\nIf `work.labels.reviewing` resolves to a **label string**, flip `reviewRequested → reviewing` and assign the issue to the runner **before** reading the diff — on `github` and `linear`, the tracker-global claim that stops a second clone reviewing the same issue and writing a competing verdict (the review lock only proves no live reviewer **in this checkout**). **On `local` it is not tracker-global** — the store is a directory inside the checkout, so the lease reaches no further than the lock does and the competing-verdict hazard stays open; the flip still works and is still worth having within a checkout, but read **The `reviewing` lease** in REFERENCE.md before enabling `labels.reviewing` there (its recommendation is to leave it off). Honour the review single-flight lock (the **Lease & race rules** and **The single-flight lock** in `work-implement`'s REFERENCE) — take it (direct run) or run under the drain's (queue).\n\n**When `reviewing` is `false`\u002Funset (the default), skip this step entirely** — there is no lease label, so the review loop behaves exactly as before: the lock alone, reviewing straight off `reviewRequested`. The verdict at step 7 then moves the label off `reviewRequested`; when the lease is on, it moves off `reviewing`. Either way the verdict label move clears the in-flight state.\n\n### 5. Gather the work\n\n- **Requirements** — re-read the issue **body** (what was asked) and any **prior review feedback** on the issue **and** the PR (so a re-review checks the last round was addressed). Look in both regardless of `work.feedback`: it routes where feedback is _written_, never where it is read, and rounds written before the mode changed stay where they were posted.\n- **The pushed diff** — the artifact to review, scoped to _this_ issue:\n  - **A pull\u002Fmerge request present** (`worktree`) → that request's own diff, in the forge's spelling ([recipes](REFERENCE.md#fetching-the-diff)).\n  - **Neither present** (`branch:\u003Cname>`, e.g. `branch:dev`) → the issue's own commit range on the branch (the commits referencing this issue since it was last picked up). See [review-after-land](REFERENCE.md#review-after-land).\n- **Checks** — the pushed head must be green on the repo's **own** gate, and this review **establishes** that rather than inheriting it. Two sources, in this order:\n  - **The forge's checks**, but only where the head's base actually triggers them. Read which workflows the base runs _before_ reading their result — an empty or irrelevant check list is `unknown`, never green. A **draft** PR is the expected case, not an anomaly: the implement loop opens every PR as a draft and a repo gating its CI on that has deliberately run nothing yet, so this source is normally `unknown` here. CI is not waited for at this step — it is asked for at step 7, once the verdict is otherwise `done` ([marking ready, then waiting for CI](REFERENCE.md#marking-ready-then-waiting-for-ci)).\n  - **The repo's `verify`, run here.** Under `branch:\u003Cname>` there is no PR and CI commonly never ran at all, so this is the only source that exists. Run it against the pushed head in a **throwaway worktree** — the review is read-only towards the user's tree, which is not the same as running nothing ([recipe](REFERENCE.md#verifying-the-pushed-head)).\n\n  **The implementer's own green run is not a source.** It proves a tree passed _before_ the push, and on a shared branch that tree has since moved. Re-running is the whole point of a second agent: same gate, different tree, no inherited verdict.\n\n  Red, or a gate that could not be run at all → a review finding, never a pass.\n\n### 6. Review — adversarially\n\nYou are the skeptic. Judge, in this order:\n\n- **Does it do what the issue asked?** Match the diff against the body's requirements — nothing missing, nothing out of scope.\n- **Is it correct?** Look for real defects: wrong logic, edge cases, broken references, secrets, regressions.\n- **Is a human needed?** Apply the [escalation policy](REFERENCE.md#escalation-to-needs-human) — sensitive surfaces (via CODEOWNERS \u002F branch protection), ambiguous intent you cannot confirm, a change too large\u002Frisky to sign off confidently, or checks red without a clear cause.\n\n### 7. Verdict — one label move + a comment\n\nCount the review rounds first — the number of times this issue has entered `reviewRequested` ([recipe](REFERENCE.md#round-count)) — and compare to `work.review.maxRounds`. A count that could not be read is **not** zero rounds: escalate to `needs human` rather than let the loop run uncapped. Then:\n\n| Verdict                 | When                                                                                                           | Action                                                                                                                                                                                                  |\n| :---------------------- | :------------------------------------------------------------------------------------------------------------- | :------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ |\n| **`done`**              | correct, complete, low-risk                                                                                    | set `done` — accepted                                                                                                                                                                                   |\n| **`needs human`**       | correct but **risky\u002Fsensitive**, OR you cannot confidently judge, OR round ≥ `maxRounds` and still not passing | set `needs human` + comment why a human is needed                                                                                                                                                       |\n| **`changes-requested`** | fixable problems, round \u003C `maxRounds`                                                                          | **post the feedback** at the configured destination (a comment on the pull\u002Fmerge request, or an issue\u002FLinear comment referencing the commit), then set `changes-requested` — back to the implement loop |\n| **`blocked`**           | broken beyond a fixable change \u002F a hard human call                                                             | set `blocked` + comment                                                                                                                                                                                 |\n\n**A `done` is the moment CI is worth spending, so it is spent before the verdict is written.** Under `worktree` the implement loop leaves every PR a **draft** (**The draft gate** in `work-implement`'s REFERENCE), which on a repo gating its workflows on that means no CI has run. So where the review above lands on `done` and the PR is still a draft: **mark it ready for review**, then wait for the checks the base actually triggers, bounded by `work.review.timeout` (default 600s).\n\n**Two things decide what that wait reads, and both are easy to get subtly wrong.** A **`skipping` check is never a pass** — a repo gating its _jobs_ on the draft state reports one such row per gated job rather than reporting nothing, so counting them accepts a head CI never analysed; drop them and read only the jobs that answered. And the budget starts at the **later** of the un-draft and this head's own commit — a base that exists on _every_ round, including one answering a CI failure on a PR that is already ready, where nothing un-drafts and the un-draft event alone would put the deadline in the past before the first poll. That base bounds the wait only; it never decides which results count.\n\n| CI, after the un-draft                       | Verdict                                                                                                  |\n| :------------------------------------------- | :------------------------------------------------------------------------------------------------------- |\n| **green**                                    | `done` — the verdict the review already reached, now paid for                                            |\n| **red**                                      | `changes-requested` with the failure as the feedback — or `needs human` at `maxRounds`, as any red would |\n| **still running when the timeout elapses**   | `needs human` — no verdict rests on a gate that has not reported                                         |\n| **none triggered at all** (not \"none _yet_\") | `done` stands: step 5's second source already ran the repo's own `verify` against this head              |\n\n**Every other verdict leaves the draft alone, and nothing ever re-drafts.** A review that already wants changes routes to `changes-requested` without un-drafting, so no CI is spent on it; once a PR has been ready it stays ready, CI failures included, so the following rounds get CI feedback directly instead of going dark again. Recipe, and the `unknown`-versus-`no checks` distinction the fourth row turns on: [marking ready, then waiting for CI](REFERENCE.md#marking-ready-then-waiting-for-ci).\n\n**The label move always lands on the issue; the comment goes where `work.feedback` says** — the PR's thread (`pr`) or the issue's comments (`issue`), so a three-round review does not turn the requirement into a log ([recipes](REFERENCE.md#feedback-recipes)). In `pr` mode the **comment** is the primitive that carries every verdict — `gh pr comment` on GitHub, `glab mr note` on GitLab. GitHub's `gh pr review --request-changes` is an optional upgrade it **refuses on a self-authored PR**, which is the normal case wherever both loops run as one identity; **GitLab has no second verb to upgrade to at all**, so the note is the whole story there. Two fallbacks, both reported in the run: **no pull\u002Fmerge request to write to** → post to the issue; **the review call refused** → post the same body as a comment.\n\nReport the verdict and the reasoning. Inside a `work-review-queue` drain, return the verdict and let the drain move on.\n\n## Guardrails\n\n- **Review only — never implement.** No `Edit`\u002F`Write`\u002Fcommit\u002Fmerge\u002Fpush. If the fix is obvious, describe it in the feedback; do not apply it. The implement loop applies it.\n- **Marking a PR ready for review is this loop's, and only on an accept.** It changes no code and merges nothing — it is the verdict said in the forge's own vocabulary, which is why the read-only rule above does not forbid it. Flip it only where the review would otherwise write `done`, **never** as housekeeping, and **never back into draft** on any verdict, CI failures included ([marking ready, then waiting for CI](REFERENCE.md#marking-ready-then-waiting-for-ci)).\n- **A different agent than the implementer.** Review as a fresh skeptic; when in doubt, `needs human` — never rubber-stamp.\n- **Lease before review, when `reviewing` is configured.** Claim `reviewRequested → reviewing` + assign before reading the diff, so a second clone cannot review the same issue and write a competing verdict — the review loop's counterpart of the implement lease. With `labels.reviewing` off (the default), the review lock alone applies: today's behaviour, unchanged.\n- **Read-only and idempotent.** A crashed review re-runs; posting feedback, dedupe against feedback you already left.\n- **`done` is a real acceptance**, given by AI review (low-risk) or by a human (via `needs human`). Never set `done` on a change you could not confidently review — escalate instead.\n- **The round cap is a floor for the human, not a ceiling on quality** — at `maxRounds`, escalate to `needs human`; never quietly accept unfinished work to end the loop.\n- **Attribution-free & secret-free** — no `Generated with`\u002F🤖 line, no agent self-naming in comments; scan the diff for secrets and flag them.\n\n## Reference\n\nConfig, the selection query, the round-count recipe, the escalation policy, review-after-land and the feedback recipes — everything a run needs whatever its tracker: [REFERENCE.md](REFERENCE.md). The implement half and the shared lifecycle: `work-implement`.\n\n**One file per tracker recipe, and a repo reads exactly one.** `work.tracker` (falling back to `issue.tracker`) is a single value, so step 1 settles the branch before any recipe is opened: [`trackers\u002Fgithub.md`](trackers\u002Fgithub.md), [`trackers\u002Fgitlab.md`](trackers\u002Fgitlab.md), [`trackers\u002Flinear.md`](trackers\u002Flinear.md), [`trackers\u002Flocal.md`](trackers\u002Flocal.md).\n",{"name":5,"metadata":24,"description":7,"allowed-tools":26},{"summary":25},"Reviews one issue's pushed work as an independent agent — verdict routes to done, changes, needs-human, or blocked.",[27,28,29,30],"Bash","Read","Grep","Glob","main","0343b3844cdfe00ebabe556f2a63c0515208e92c","skills\u002Fwork\u002Fwork-review",0,"https:\u002F\u002Fgithub.com\u002FTitusKirch\u002Fskills","https:\u002F\u002Fgithub.com\u002FTitusKirch","https:\u002F\u002Fkirch.dev\u002F","💼 Self-employed Web Developer\r\n🎓 B. Sc. Computer Science","2026-11-02 00:46:17",{"source":8,"skillName":5,"commands":41,"prompt":51,"zipAvailable":52},[42,45,48],{"tool":43,"command":44},"npx","npx skills add TitusKirch\u002Fskills --skill work-review",{"tool":46,"command":47},"bunx","bunx skills add TitusKirch\u002Fskills --skill work-review",{"tool":49,"command":50},"pnpm","pnpm dlx skills add TitusKirch\u002Fskills --skill work-review","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002FTitusKirch\u002Fskills\u002Fblob\u002Fmain\u002Fskills\u002Fwork\u002Fwork-review\u002FSKILL.md",true,[54,60,64,70,74,78,82],{"path":55,"resourceType":56,"resourceTypeDesc":57,"sizeBytes":58,"sourceHtmlUrl":59},"evals\u002Fevals.json","REFERENCE","参考文档",6100,"https:\u002F\u002Fgithub.com\u002FTitusKirch\u002Fskills\u002Fblob\u002Fmain\u002Fskills\u002Fwork\u002Fwork-review\u002Fevals\u002Fevals.json",{"path":61,"resourceType":56,"resourceTypeDesc":57,"sizeBytes":62,"sourceHtmlUrl":63},"REFERENCE.md",82146,"https:\u002F\u002Fgithub.com\u002FTitusKirch\u002Fskills\u002Fblob\u002Fmain\u002Fskills\u002Fwork\u002Fwork-review\u002FREFERENCE.md",{"path":65,"resourceType":66,"resourceTypeDesc":67,"sizeBytes":68,"sourceHtmlUrl":69},"templates\u002Fresolve-config.sh","SCRIPT","脚本",2402,"https:\u002F\u002Fgithub.com\u002FTitusKirch\u002Fskills\u002Fblob\u002Fmain\u002Fskills\u002Fwork\u002Fwork-review\u002Ftemplates\u002Fresolve-config.sh",{"path":71,"resourceType":56,"resourceTypeDesc":57,"sizeBytes":72,"sourceHtmlUrl":73},"trackers\u002Fgithub.md",590,"https:\u002F\u002Fgithub.com\u002FTitusKirch\u002Fskills\u002Fblob\u002Fmain\u002Fskills\u002Fwork\u002Fwork-review\u002Ftrackers\u002Fgithub.md",{"path":75,"resourceType":56,"resourceTypeDesc":57,"sizeBytes":76,"sourceHtmlUrl":77},"trackers\u002Fgitlab.md",641,"https:\u002F\u002Fgithub.com\u002FTitusKirch\u002Fskills\u002Fblob\u002Fmain\u002Fskills\u002Fwork\u002Fwork-review\u002Ftrackers\u002Fgitlab.md",{"path":79,"resourceType":56,"resourceTypeDesc":57,"sizeBytes":80,"sourceHtmlUrl":81},"trackers\u002Flinear.md",6534,"https:\u002F\u002Fgithub.com\u002FTitusKirch\u002Fskills\u002Fblob\u002Fmain\u002Fskills\u002Fwork\u002Fwork-review\u002Ftrackers\u002Flinear.md",{"path":83,"resourceType":56,"resourceTypeDesc":57,"sizeBytes":84,"sourceHtmlUrl":85},"trackers\u002Flocal.md",2098,"https:\u002F\u002Fgithub.com\u002FTitusKirch\u002Fskills\u002Fblob\u002Fmain\u002Fskills\u002Fwork\u002Fwork-review\u002Ftrackers\u002Flocal.md",[87,94,101,108,115,122],{"skillNo":88,"name":89,"nameZh":6,"description":90,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":91,"categories":92,"keywords":93,"publishedAt":19},"01M74FV6SYWA6MEBCX3EY64484","gamma","The third fixture skill — carries the worklock, plan and tldr blocks, and names the resolver without hosting the config block.","https:\u002F\u002Fgithub.com\u002FTitusKirch\u002Fskills\u002Fblob\u002Fmain\u002Ftest\u002Ffixtures\u002Fregistry\u002Fskills\u002Fwork\u002Fgamma\u002FSKILL.md",[],[],{"skillNo":95,"name":96,"nameZh":6,"description":97,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":98,"categories":99,"keywords":100,"publishedAt":19},"01M74FV4K7PNY8GTJH2M8ZTJGT","beta","The second fixture skill — carries the tagged blocks but no config block.","https:\u002F\u002Fgithub.com\u002FTitusKirch\u002Fskills\u002Fblob\u002Fmain\u002Ftest\u002Ffixtures\u002Fregistry\u002Fskills\u002Frepo\u002Fbeta\u002FSKILL.md",[],[],{"skillNo":102,"name":103,"nameZh":6,"description":104,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":105,"categories":106,"keywords":107,"publishedAt":19},"01M74FV2JYBQ3VRP180AZEVQFC","alpha","The first fixture skill — carries the config block and the resolver.","https:\u002F\u002Fgithub.com\u002FTitusKirch\u002Fskills\u002Fblob\u002Fmain\u002Ftest\u002Ffixtures\u002Fregistry\u002Fskills\u002Frepo\u002Falpha\u002FSKILL.md",[],[],{"skillNo":109,"name":110,"nameZh":6,"description":111,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":112,"categories":113,"keywords":114,"publishedAt":19},"01M74FTNQ1RCJTZ929GNJ2FBX5","work-review-queue","Drains a repo's queue of issues awaiting AI review across GitHub (gh), GitLab (glab), Linear (MCP) or local issue files — selects every issue in review, then reviews each with a fresh, independent agent by delegating to work-review, routing each to done, changes-requested (back to the implement loop), needs-human, or blocked. Starts by reconciling issues whose PR a human merged or closed out-of-band. Honours a per-run cap, single-flight-locked with a lock separate from the implement loop's, so review and implement drains run concurrently. Use when the user wants to review, drain, or auto-review the pushed AI work, run the review loop, says things like \"review the queue\", \"reviewe die Issues\", \"drain the review queue\", or starts a repeating driver.","https:\u002F\u002Fgithub.com\u002FTitusKirch\u002Fskills\u002Fblob\u002Fmain\u002Fskills\u002Fwork\u002Fwork-review-queue\u002FSKILL.md",[],[],{"skillNo":116,"name":117,"nameZh":6,"description":118,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":119,"categories":120,"keywords":121,"publishedAt":19},"01M74FTFKF2GBCKMKKHN5QJW0G","work-implement","Implements a single tracked issue across GitHub (gh), GitLab (glab), Linear (MCP) or local issue files — claims it via the lifecycle label, implements on a branch (fresh from ready, or re-work from changes-requested after review feedback), runs the repo's checks, commits and PUSHES, and hands the issue to the review loop by advancing the label to review. It never reviews its own work — the separate work-review skill does that. Tracker, label lifecycle and branch strategy come from the committed config (.tituskirch-skills.json). Use when the user wants to work, implement, action or pick up one specific issue or ticket, mentions an ai-ready issue, or says things like \"work issue 42\", \"arbeite Issue 42 ab\", \"implementiere Ticket X\", or \"address the review feedback on issue Y\".","https:\u002F\u002Fgithub.com\u002FTitusKirch\u002Fskills\u002Fblob\u002Fmain\u002Fskills\u002Fwork\u002Fwork-implement\u002FSKILL.md",[],[],{"skillNo":123,"name":124,"nameZh":6,"description":125,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":126,"categories":127,"keywords":128,"publishedAt":19},"01M74FT8BQ53RZDHVCGTGDWHV8","work-implement-queue","Drains a repo's queue of implementable issues across GitHub (gh), GitLab (glab), Linear (MCP) or local issue files — selects every issue that is ready or has changes requested, defers any whose prerequisite has not landed, orders the rest by priority (and, on a shared branch, by dependency), then implements each one to a pushed, reviewable state by delegating to work-implement. It hands each issue to the review loop (label review); the separate work-review-queue reviews them. Starts by reclaiming issues an earlier run crashed mid-implementation. Honours a per-run cap, runs sequentially or in parallel per config, single-flight-locked. Use when the user wants to batch-process, drain, or auto-implement the ready issues, run the implement loop, says things like \"work the issues\", \"arbeite die Issues ab\", \"drain the queue\", or starts a repeating driver.","https:\u002F\u002Fgithub.com\u002FTitusKirch\u002Fskills\u002Fblob\u002Fmain\u002Fskills\u002Fwork\u002Fwork-implement-queue\u002FSKILL.md",[],[]]