[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M6ACCF4FQX1PFWXEDHQ8YRF9":3,"skill-same:01M6ACCF4FQX1PFWXEDHQ8YRF9":134},{"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":23,"publishedAt":47,"linkOnly":48,"markdownContent":49,"markdownBody":50,"frontMatter":51,"refName":53,"commitSha":54,"skillDirectory":55,"forksCount":56,"repositoryHtmlUrl":57,"repositoryHomepageUrl":58,"ownerHtmlUrl":59,"ownerBlogUrl":60,"ownerDescription":6,"updatedAt":61,"install":62,"resources":75},"01M6ACCF4FQX1PFWXEDHQ8YRF9","targeted-review",null,"Findings-driven targeted code review. Understands the patch via patch-explainer and surrounding code via codebase-analysis, then consults a catalog of bug patterns (~11 categories under references\u002Fcategories\u002F). Picks only categories whose diff-signals match the patch and items whose \"look for\" hints match actual code shapes, groups them by review focus (function\u002Ffile\u002Ffeature), and spawns parallel subagents — each with a tight scope and selective checklist. Use for medium-to-large patches (50-1000 LOC) as a focused alternative to whole-patch ensemble or whole-file deep review. Triggers on: \"review this patch\u002Fdiff\u002Fchange\", \"find bugs in this change\", \"scoped review\", \"what could go wrong here\", \"review using findings\", or proactive PR review.","apache\u002Fcassandra","apache","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F47359?v=4","01M6AC83E6K4FDKK7HZ8WNMQKE","The Apache Software Foundation","01M6ACBT70F584W658X9T70A5M",10153,"Apache-2.0","https:\u002F\u002Fgithub.com\u002Fapache\u002Fcassandra\u002Fblob\u002Ftrunk\u002F.claude\u002Fskills\u002Ftargeted-review\u002FSKILL.md",[18],{"categoryNo":19,"categoryName":20,"taxonomyType":21,"taxonomyTypeDesc":22},"01JBSEEDCAT000000000000007","代码审查","SCENARIO","使用场景",[24,25,26,27,28,29,30,31,32,33,34,35,36,37,38,39,40,41,42,43,44,45,46],"review","targeted review","categories","code review","scope","as","deep","signals","focus","patterns","references","bug","patch","only","subagents","parallel","analysis","deep review","code","pr","pr review","change","go","2026-11-10 10:12:30",false,"---\nname: targeted-review\nversion: \"1.0.0\"\ndescription: >\n  Findings-driven targeted code review. Understands the patch via patch-explainer and\n  surrounding code via codebase-analysis, then consults a catalog of bug patterns\n  (~11 categories under references\u002Fcategories\u002F). Picks only categories whose diff-signals\n  match the patch and items whose \"look for\" hints match actual code shapes, groups them\n  by review focus (function\u002Ffile\u002Ffeature), and spawns parallel subagents — each with a\n  tight scope and selective checklist. Use for medium-to-large patches (50-1000 LOC) as\n  a focused alternative to whole-patch ensemble or whole-file deep review. Triggers on:\n  \"review this patch\u002Fdiff\u002Fchange\", \"find bugs in this change\", \"scoped review\", \"what\n  could go wrong here\", \"review using findings\", or proactive PR review.\n---\n\n# Targeted Review\n\nA meta-review skill. It does not encode bug patterns directly — it orchestrates other\nskills and a categorized findings catalog into a tight, evidence-driven review.\n\n## What it does\n\n```\n                +------------------+\n                |   Input patch    |\n                +--------+---------+\n                         |\n             +-----------+-----------+\n             |                       |\n   +---------v---------+   +---------v-----------+\n   | patch-explainer   |   | codebase-analysis   |\n   | (sub-skill)       |   | (sub-skill, scoped) |\n   | what changed,     |   | invariants, callers,|\n   | flow, assumptions |   | parallel paths      |\n   +---------+---------+   +---------+-----------+\n             |                       |\n             +-----------+-----------+\n                         |\n         +---------------v----------------+\n         | Pick categories + items        |\n         | × 3-5 independent iterations  |\n         | (each draw surfaces different  |\n         |  items from the same catalog)  |\n         +---------------+----------------+\n                         |\n                +--------v---------+\n                | Pool & prioritize|\n                | (items from 2+   |\n                |  iterations →    |\n                |  higher priority)|\n                +--------+---------+\n                         |\n                +--------v---------+\n                | Group items by   |\n                | review focus     |\n                | (file\u002Ffunc\u002Ffeat) |\n                +--------+---------+\n                         |\n            +------------+------------+\n            |            |            |\n    +-------v---+  +-----v-----+  +---v-------+\n    | review    |  | review    |  | review    |\n    | subagent  |  | subagent  |  | subagent  |\n    | scope A   |  | scope B   |  | scope C   |\n    +-------+---+  +-----+-----+  +---+-------+\n            |            |            |\n            +------------+------------+\n                         |\n                +--------v---------+\n                | Merge & report   |\n                +------------------+\n```\n\n## When to use this skill\n\n- A patch or diff is on the table for review (single commit, multi-commit branch, or staged changes)\n- The user wants more focus than `shallow-review` provides (six fixed lenses on the whole patch) but doesn't want to commit to `deep-review`'s file-by-file deep pass\n- A `*-findings\u002F` corpus exists or the categorized catalog under `references\u002Fcategories\u002F` is sufficient — the categories alone are usable; project corpora improve coverage\n- Medium-to-large patches where dispatching focused subagents avoids overwhelming a single reviewer\n\n## When NOT to use this skill\n\n- Trivial 1-3 line changes — direct review is faster\n- Pure refactors with no behavior change — `shallow-review` symmetry pass is enough\n- The user wants a single comprehensive write-up of one file — use `deep-review` instead\n\n---\n\n## Workflow\n\nYou will go through seven phases. Phases 1a and 1b run in parallel. Phases 2-6 are sequential.\n\n### Phase 1a: Understand the patch — call the patch-explainer skill\n\nInvoke the `patch-explainer` skill on the patch. Use the Skill tool. You want patch-explainer\nto produce its standard analysis (purpose, before\u002Fafter diagrams, flow, assumptions,\nfailure modes).\n\nIf the patch is already provided as a file path, point patch-explainer at it. If you only\nhave a commit ref, ask for `git show \u003Cref>` or read it from `git diff`. For multi-commit\nbranches, it is fine to feed the cumulative diff.\n\nCapture patch-explainer's full output. You will refer to it in Phases 3-5.\n\n### Phase 1b: Understand the surroundings — call the codebase-analysis skill\n\nIn parallel with 1a, invoke the `codebase-analysis` skill in **targeted mode** on the\naffected subsystem only — not the whole repo. Pass it:\n\n- The list of files touched by the patch\n- The subsystem name if obvious (e.g., \"the journal replay subsystem\")\n- A directive to focus on: invariants, lifecycle, parallel implementations, and recent\n  bug history in the affected area\n\nYou want codebase-analysis to produce a scoped `analysis\u002Fsystem-analysis.md` that gives\nyou the context the patch sits inside — what the surrounding code expects, where parallel\npaths exist, what historical bugs touched this area.\n\nIf invoking the full skill is too heavy for a small patch, instead spawn a research\nsubagent with this prompt: \"Read these files: {touched files}. Summarize: (a) the\ninvariants the surrounding code maintains, (b) any parallel implementations or sibling\nclasses, (c) callers of the modified methods, (d) any recent bug-fix commits in this\narea. Under 400 words.\" This is a lighter alternative that retains the spirit of\ncodebase-analysis. Prefer the full skill when the area is unfamiliar or large.\n\n### Phase 2: Pick categories — 3-5 independent iterations\n\nCategory selection and item picking are speculative: you're making probabilistic judgments\nabout which patterns might apply, and a single pass will reliably miss some. Run **3-5\nindependent iterations** of Phases 2+3, each starting from scratch with the same patch\nand context but without looking back at what prior iterations chose. Treat each iteration\nlike a fresh read — you are more likely to notice different things each time.\n\nFor each iteration, read `references\u002Fcategories\u002FINDEX.md` — it lists all 11 categories\nwith their descriptions and diff signals in one file (~200 lines). For each category, ask:\ndoes the patch contain ANY of the diff signals listed?\n- If yes → mark as **load**\n- If no → skip\n\nTypically 3-7 categories load per iteration. If you would load more than 8, re-examine.\nIf fewer than 2, the patch is probably too small for this skill.\n\n### Phase 3: Pick items within each loaded category (one pass per iteration)\n\nWithin each iteration, read the **full** content of each loaded category file. For each\nfinding (`### F-NN: title` + body + `**Look for:** {hint}`), ask:\n\n1. Does the \"Look for\" hint match a code shape that **actually appears in this patch**?\n2. Does the patch context (from patch-explainer + codebase-analysis) make this finding\n   **plausibly applicable** — not just a keyword match?\n\nIf yes to both → keep the item.\nIf no → drop.\n\nBe selective: a category typically has 25-40 findings; keep 3-12 per category per\niteration. Repeat for all iterations before moving to Phase 3b.\n\n### Phase 3b: Pool and prioritize across iterations\n\nAfter all iterations are complete, merge the item lists:\n\n- An item selected in **2 or more iterations** is a strong candidate — mark it **priority**.\n- An item selected in only one iteration is still in scope — mark it **speculative**.\n- Discard nothing yet; let the subagents decide.\n\nThe iteration count next to an item is a confidence signal for the subagents, not a\ngate. A speculative item can still turn into a real bug; a priority item can still be\na false alarm.\n\n### Phase 4: Group items by review focus\n\nNow you have a pooled list of selected items across categories and iterations. Group them\nby **review focus** — what part of the change a single reviewer should examine. A focus\nis one of:\n\n- A specific function or method (`MyClass.replay()`)\n- A specific file or small group (`FooSerializer.java` and its sibling `FooDeserializer.java`)\n- A feature, behavior, or concept (\"the new fsync coordination\", \"the version-gated dispatch path\")\n- A specific cross-cutting concern in the patch (\"any place we read the new `dirty` flag\")\n\nAim for 2-5 foci per patch. Each focus should have **5-15 checklist items** drawn from\nacross the categories — not too few (subagent has nothing to do) and not too many\n(subagent gets overwhelmed). Include each item's priority annotation (priority \u002F\nspeculative) so the subagent knows where to focus first.\n\nIf multiple foci would receive the same checklist item, that's fine — the same pattern\ncan apply at multiple sites. Tag each item with the category it came from so the\nsubagent knows the rationale.\n\n### Phase 5: Spawn focused review subagents\n\nFor each focus, spawn one review subagent in parallel using the Agent tool. Use the\ntemplate in `references\u002Fsubagent-prompt-template.md`. The subagent prompt must contain:\n\n- The patch (path, or pasted, or ref)\n- A 100-word summary of patch-explainer's findings (relevant slice only — not the full output)\n- A 100-word summary of codebase-analysis's findings (relevant slice only)\n- The focus statement (what part of the change to review)\n- The checklist items, each with: category, priority annotation, finding text, and\n  look-for hint\n- Instructions on how to gather evidence (read source, grep, check parallel paths) and\n  what confidence to assign\n\nEach subagent should produce: a list of findings, with location, confidence, and what\nevidence was checked.\n\nRun all subagents in a single tool call (parallel). They are independent.\n\n### Phase 6: Merge and report\n\nWhen all subagents return, merge findings:\n\n1. **Deduplicate** — same code location + similar reasoning → one finding, both subagents\n   credited\n2. **Rank** by confidence — High first, then Medium, then Low\n3. **Cross-reinforce** — if a finding was raised by multiple subagents, or if multiple\n   findings cluster on the same lifecycle\u002Fstate, boost confidence\n4. **Three-point check** each surviving finding:\n   - The code construct actually exists in the diff (not inferred from absence alone)\n   - The bug is plausible given the visible context (not speculative)\n   - The finding is actionable (the reader can tell what to change)\n\nUse the format in `references\u002Freport-format.md`.\n\n---\n\n## Reference files\n\n| Reference | When to read |\n|---|---|\n| `references\u002Fsubagent-prompt-template.md` | Phase 5, when spawning each review subagent |\n| `references\u002Freport-format.md` | Phase 6, when merging and presenting findings |\n| `references\u002Fcategories\u002FINDEX.md` | Phase 2 — single-file scan of all category descriptions + diff signals |\n| `references\u002Fcategories\u002F\u003Cname>.md` | Phase 3 — full body of each loaded category, picked from INDEX |\n\nThe categories in `references\u002Fcategories\u002F` are project-agnostic patterns mined from real\nbug fixes across distributed-systems projects (Cassandra, Kafka, Iceberg). They describe\ngeneralized code-level shapes, not project-specific issues.\n\n## How this differs from related skills\n\n| Skill | Scope | Pattern source | Subagent dispatch |\n|---|---|---|---|\n| `shallow-review` (cassandra-edge-case-explorer) | Whole patch | 6 fixed lenses (~100 items total) | One subagent per lens, fixed |\n| `deep-review` | User-named files | 444-pattern catalog | One subagent per file, full catalog |\n| `targeted-review` (this) | Whole patch, focus-decomposed | ~11 categories (300+ findings) — selectively loaded | One subagent per focus, selectively populated checklist |\n| `mega-review` | Large branches | Orchestrates the above | Multi-pass |\n\nUse `targeted-review` when:\n- The patch spans 50-1000 LOC\n- You want to spend tokens on **what matters for THIS patch**, not on a fixed catalog\n- You have a `*-findings\u002F` corpus or are reviewing systems-level code that benefits from\n  the curated categories\n\n## Pitfalls and guardrails\n\n- **Don't skip Phase 1.** Calling patch-explainer and codebase-analysis is what lets you\n  pick categories and items intelligently. Without that grounding, you'll loose-match\n  on keywords and overwhelm subagents.\n- **Don't look at prior iterations while picking.** The whole value of multiple iterations\n  is independent draws. If you anchor on the first iteration's selections, you get one\n  draw repeated instead of several independent ones.\n- **Don't load every category \"just in case\".** The whole value is selection. If you\n  match more than 8 categories in an iteration, you've probably been too generous.\n- **Don't pass the full catalog to subagents.** Each subagent gets only the items\n  selected for its focus. Passing all items defeats the purpose.\n- **Don't substitute findings for understanding.** A finding that \"matches a code shape\"\n  is a hypothesis to verify — the subagent confirms or refutes by reading the code, not\n  by trusting the pattern. Mark uncertain findings as Low confidence.\n- **The checklist is the menu, not the order.** A subagent should also report bugs it\n  notices that aren't on the checklist — the checklist primes attention, it doesn't cap\n  it.\n","# Targeted Review\n\nA meta-review skill. It does not encode bug patterns directly — it orchestrates other\nskills and a categorized findings catalog into a tight, evidence-driven review.\n\n## What it does\n\n```\n                +------------------+\n                |   Input patch    |\n                +--------+---------+\n                         |\n             +-----------+-----------+\n             |                       |\n   +---------v---------+   +---------v-----------+\n   | patch-explainer   |   | codebase-analysis   |\n   | (sub-skill)       |   | (sub-skill, scoped) |\n   | what changed,     |   | invariants, callers,|\n   | flow, assumptions |   | parallel paths      |\n   +---------+---------+   +---------+-----------+\n             |                       |\n             +-----------+-----------+\n                         |\n         +---------------v----------------+\n         | Pick categories + items        |\n         | × 3-5 independent iterations  |\n         | (each draw surfaces different  |\n         |  items from the same catalog)  |\n         +---------------+----------------+\n                         |\n                +--------v---------+\n                | Pool & prioritize|\n                | (items from 2+   |\n                |  iterations →    |\n                |  higher priority)|\n                +--------+---------+\n                         |\n                +--------v---------+\n                | Group items by   |\n                | review focus     |\n                | (file\u002Ffunc\u002Ffeat) |\n                +--------+---------+\n                         |\n            +------------+------------+\n            |            |            |\n    +-------v---+  +-----v-----+  +---v-------+\n    | review    |  | review    |  | review    |\n    | subagent  |  | subagent  |  | subagent  |\n    | scope A   |  | scope B   |  | scope C   |\n    +-------+---+  +-----+-----+  +---+-------+\n            |            |            |\n            +------------+------------+\n                         |\n                +--------v---------+\n                | Merge & report   |\n                +------------------+\n```\n\n## When to use this skill\n\n- A patch or diff is on the table for review (single commit, multi-commit branch, or staged changes)\n- The user wants more focus than `shallow-review` provides (six fixed lenses on the whole patch) but doesn't want to commit to `deep-review`'s file-by-file deep pass\n- A `*-findings\u002F` corpus exists or the categorized catalog under `references\u002Fcategories\u002F` is sufficient — the categories alone are usable; project corpora improve coverage\n- Medium-to-large patches where dispatching focused subagents avoids overwhelming a single reviewer\n\n## When NOT to use this skill\n\n- Trivial 1-3 line changes — direct review is faster\n- Pure refactors with no behavior change — `shallow-review` symmetry pass is enough\n- The user wants a single comprehensive write-up of one file — use `deep-review` instead\n\n---\n\n## Workflow\n\nYou will go through seven phases. Phases 1a and 1b run in parallel. Phases 2-6 are sequential.\n\n### Phase 1a: Understand the patch — call the patch-explainer skill\n\nInvoke the `patch-explainer` skill on the patch. Use the Skill tool. You want patch-explainer\nto produce its standard analysis (purpose, before\u002Fafter diagrams, flow, assumptions,\nfailure modes).\n\nIf the patch is already provided as a file path, point patch-explainer at it. If you only\nhave a commit ref, ask for `git show \u003Cref>` or read it from `git diff`. For multi-commit\nbranches, it is fine to feed the cumulative diff.\n\nCapture patch-explainer's full output. You will refer to it in Phases 3-5.\n\n### Phase 1b: Understand the surroundings — call the codebase-analysis skill\n\nIn parallel with 1a, invoke the `codebase-analysis` skill in **targeted mode** on the\naffected subsystem only — not the whole repo. Pass it:\n\n- The list of files touched by the patch\n- The subsystem name if obvious (e.g., \"the journal replay subsystem\")\n- A directive to focus on: invariants, lifecycle, parallel implementations, and recent\n  bug history in the affected area\n\nYou want codebase-analysis to produce a scoped `analysis\u002Fsystem-analysis.md` that gives\nyou the context the patch sits inside — what the surrounding code expects, where parallel\npaths exist, what historical bugs touched this area.\n\nIf invoking the full skill is too heavy for a small patch, instead spawn a research\nsubagent with this prompt: \"Read these files: {touched files}. Summarize: (a) the\ninvariants the surrounding code maintains, (b) any parallel implementations or sibling\nclasses, (c) callers of the modified methods, (d) any recent bug-fix commits in this\narea. Under 400 words.\" This is a lighter alternative that retains the spirit of\ncodebase-analysis. Prefer the full skill when the area is unfamiliar or large.\n\n### Phase 2: Pick categories — 3-5 independent iterations\n\nCategory selection and item picking are speculative: you're making probabilistic judgments\nabout which patterns might apply, and a single pass will reliably miss some. Run **3-5\nindependent iterations** of Phases 2+3, each starting from scratch with the same patch\nand context but without looking back at what prior iterations chose. Treat each iteration\nlike a fresh read — you are more likely to notice different things each time.\n\nFor each iteration, read `references\u002Fcategories\u002FINDEX.md` — it lists all 11 categories\nwith their descriptions and diff signals in one file (~200 lines). For each category, ask:\ndoes the patch contain ANY of the diff signals listed?\n- If yes → mark as **load**\n- If no → skip\n\nTypically 3-7 categories load per iteration. If you would load more than 8, re-examine.\nIf fewer than 2, the patch is probably too small for this skill.\n\n### Phase 3: Pick items within each loaded category (one pass per iteration)\n\nWithin each iteration, read the **full** content of each loaded category file. For each\nfinding (`### F-NN: title` + body + `**Look for:** {hint}`), ask:\n\n1. Does the \"Look for\" hint match a code shape that **actually appears in this patch**?\n2. Does the patch context (from patch-explainer + codebase-analysis) make this finding\n   **plausibly applicable** — not just a keyword match?\n\nIf yes to both → keep the item.\nIf no → drop.\n\nBe selective: a category typically has 25-40 findings; keep 3-12 per category per\niteration. Repeat for all iterations before moving to Phase 3b.\n\n### Phase 3b: Pool and prioritize across iterations\n\nAfter all iterations are complete, merge the item lists:\n\n- An item selected in **2 or more iterations** is a strong candidate — mark it **priority**.\n- An item selected in only one iteration is still in scope — mark it **speculative**.\n- Discard nothing yet; let the subagents decide.\n\nThe iteration count next to an item is a confidence signal for the subagents, not a\ngate. A speculative item can still turn into a real bug; a priority item can still be\na false alarm.\n\n### Phase 4: Group items by review focus\n\nNow you have a pooled list of selected items across categories and iterations. Group them\nby **review focus** — what part of the change a single reviewer should examine. A focus\nis one of:\n\n- A specific function or method (`MyClass.replay()`)\n- A specific file or small group (`FooSerializer.java` and its sibling `FooDeserializer.java`)\n- A feature, behavior, or concept (\"the new fsync coordination\", \"the version-gated dispatch path\")\n- A specific cross-cutting concern in the patch (\"any place we read the new `dirty` flag\")\n\nAim for 2-5 foci per patch. Each focus should have **5-15 checklist items** drawn from\nacross the categories — not too few (subagent has nothing to do) and not too many\n(subagent gets overwhelmed). Include each item's priority annotation (priority \u002F\nspeculative) so the subagent knows where to focus first.\n\nIf multiple foci would receive the same checklist item, that's fine — the same pattern\ncan apply at multiple sites. Tag each item with the category it came from so the\nsubagent knows the rationale.\n\n### Phase 5: Spawn focused review subagents\n\nFor each focus, spawn one review subagent in parallel using the Agent tool. Use the\ntemplate in `references\u002Fsubagent-prompt-template.md`. The subagent prompt must contain:\n\n- The patch (path, or pasted, or ref)\n- A 100-word summary of patch-explainer's findings (relevant slice only — not the full output)\n- A 100-word summary of codebase-analysis's findings (relevant slice only)\n- The focus statement (what part of the change to review)\n- The checklist items, each with: category, priority annotation, finding text, and\n  look-for hint\n- Instructions on how to gather evidence (read source, grep, check parallel paths) and\n  what confidence to assign\n\nEach subagent should produce: a list of findings, with location, confidence, and what\nevidence was checked.\n\nRun all subagents in a single tool call (parallel). They are independent.\n\n### Phase 6: Merge and report\n\nWhen all subagents return, merge findings:\n\n1. **Deduplicate** — same code location + similar reasoning → one finding, both subagents\n   credited\n2. **Rank** by confidence — High first, then Medium, then Low\n3. **Cross-reinforce** — if a finding was raised by multiple subagents, or if multiple\n   findings cluster on the same lifecycle\u002Fstate, boost confidence\n4. **Three-point check** each surviving finding:\n   - The code construct actually exists in the diff (not inferred from absence alone)\n   - The bug is plausible given the visible context (not speculative)\n   - The finding is actionable (the reader can tell what to change)\n\nUse the format in `references\u002Freport-format.md`.\n\n---\n\n## Reference files\n\n| Reference | When to read |\n|---|---|\n| `references\u002Fsubagent-prompt-template.md` | Phase 5, when spawning each review subagent |\n| `references\u002Freport-format.md` | Phase 6, when merging and presenting findings |\n| `references\u002Fcategories\u002FINDEX.md` | Phase 2 — single-file scan of all category descriptions + diff signals |\n| `references\u002Fcategories\u002F\u003Cname>.md` | Phase 3 — full body of each loaded category, picked from INDEX |\n\nThe categories in `references\u002Fcategories\u002F` are project-agnostic patterns mined from real\nbug fixes across distributed-systems projects (Cassandra, Kafka, Iceberg). They describe\ngeneralized code-level shapes, not project-specific issues.\n\n## How this differs from related skills\n\n| Skill | Scope | Pattern source | Subagent dispatch |\n|---|---|---|---|\n| `shallow-review` (cassandra-edge-case-explorer) | Whole patch | 6 fixed lenses (~100 items total) | One subagent per lens, fixed |\n| `deep-review` | User-named files | 444-pattern catalog | One subagent per file, full catalog |\n| `targeted-review` (this) | Whole patch, focus-decomposed | ~11 categories (300+ findings) — selectively loaded | One subagent per focus, selectively populated checklist |\n| `mega-review` | Large branches | Orchestrates the above | Multi-pass |\n\nUse `targeted-review` when:\n- The patch spans 50-1000 LOC\n- You want to spend tokens on **what matters for THIS patch**, not on a fixed catalog\n- You have a `*-findings\u002F` corpus or are reviewing systems-level code that benefits from\n  the curated categories\n\n## Pitfalls and guardrails\n\n- **Don't skip Phase 1.** Calling patch-explainer and codebase-analysis is what lets you\n  pick categories and items intelligently. Without that grounding, you'll loose-match\n  on keywords and overwhelm subagents.\n- **Don't look at prior iterations while picking.** The whole value of multiple iterations\n  is independent draws. If you anchor on the first iteration's selections, you get one\n  draw repeated instead of several independent ones.\n- **Don't load every category \"just in case\".** The whole value is selection. If you\n  match more than 8 categories in an iteration, you've probably been too generous.\n- **Don't pass the full catalog to subagents.** Each subagent gets only the items\n  selected for its focus. Passing all items defeats the purpose.\n- **Don't substitute findings for understanding.** A finding that \"matches a code shape\"\n  is a hypothesis to verify — the subagent confirms or refutes by reading the code, not\n  by trusting the pattern. Mark uncertain findings as Low confidence.\n- **The checklist is the menu, not the order.** A subagent should also report bugs it\n  notices that aren't on the checklist — the checklist primes attention, it doesn't cap\n  it.\n",{"name":5,"version":52,"description":7},"1.0.0","trunk","ae4e78c0ece616138963019388061efa4208f67a",".claude\u002Fskills\u002Ftargeted-review",4066,"https:\u002F\u002Fgithub.com\u002Fapache\u002Fcassandra","https:\u002F\u002Fcassandra.apache.org\u002F","https:\u002F\u002Fgithub.com\u002Fapache","https:\u002F\u002Fwww.apache.org\u002F","2026-11-14 00:43:44",{"source":8,"skillName":5,"commands":63,"prompt":73,"zipAvailable":74},[64,67,70],{"tool":65,"command":66},"npx","npx skills add apache\u002Fcassandra --skill targeted-review",{"tool":68,"command":69},"bunx","bunx skills add apache\u002Fcassandra --skill targeted-review",{"tool":71,"command":72},"pnpm","pnpm dlx skills add apache\u002Fcassandra --skill targeted-review","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Fapache\u002Fcassandra\u002Fblob\u002Ftrunk\u002F.claude\u002Fskills\u002Ftargeted-review\u002FSKILL.md",true,[76,82,86,90,94,98,102,106,110,114,118,122,126,130],{"path":77,"resourceType":78,"resourceTypeDesc":79,"sizeBytes":80,"sourceHtmlUrl":81},"references\u002Fcategories\u002Fapi-contracts-and-completeness.md","REFERENCE","参考文档",22824,"https:\u002F\u002Fgithub.com\u002Fapache\u002Fcassandra\u002Fblob\u002Ftrunk\u002F.claude\u002Fskills\u002Ftargeted-review\u002Freferences\u002Fcategories\u002Fapi-contracts-and-completeness.md",{"path":83,"resourceType":78,"resourceTypeDesc":79,"sizeBytes":84,"sourceHtmlUrl":85},"references\u002Fcategories\u002Fboundaries-and-numbers.md",21164,"https:\u002F\u002Fgithub.com\u002Fapache\u002Fcassandra\u002Fblob\u002Ftrunk\u002F.claude\u002Fskills\u002Ftargeted-review\u002Freferences\u002Fcategories\u002Fboundaries-and-numbers.md",{"path":87,"resourceType":78,"resourceTypeDesc":79,"sizeBytes":88,"sourceHtmlUrl":89},"references\u002Fcategories\u002Fconcurrency-and-locking.md",17433,"https:\u002F\u002Fgithub.com\u002Fapache\u002Fcassandra\u002Fblob\u002Ftrunk\u002F.claude\u002Fskills\u002Ftargeted-review\u002Freferences\u002Fcategories\u002Fconcurrency-and-locking.md",{"path":91,"resourceType":78,"resourceTypeDesc":79,"sizeBytes":92,"sourceHtmlUrl":93},"references\u002Fcategories\u002Fconditions-and-predicates.md",18327,"https:\u002F\u002Fgithub.com\u002Fapache\u002Fcassandra\u002Fblob\u002Ftrunk\u002F.claude\u002Fskills\u002Ftargeted-review\u002Freferences\u002Fcategories\u002Fconditions-and-predicates.md",{"path":95,"resourceType":78,"resourceTypeDesc":79,"sizeBytes":96,"sourceHtmlUrl":97},"references\u002Fcategories\u002FINDEX.md",17634,"https:\u002F\u002Fgithub.com\u002Fapache\u002Fcassandra\u002Fblob\u002Ftrunk\u002F.claude\u002Fskills\u002Ftargeted-review\u002Freferences\u002Fcategories\u002FINDEX.md",{"path":99,"resourceType":78,"resourceTypeDesc":79,"sizeBytes":100,"sourceHtmlUrl":101},"references\u002Fcategories\u002Fio-and-crash-safety.md",13646,"https:\u002F\u002Fgithub.com\u002Fapache\u002Fcassandra\u002Fblob\u002Ftrunk\u002F.claude\u002Fskills\u002Ftargeted-review\u002Freferences\u002Fcategories\u002Fio-and-crash-safety.md",{"path":103,"resourceType":78,"resourceTypeDesc":79,"sizeBytes":104,"sourceHtmlUrl":105},"references\u002Fcategories\u002Flifecycle-and-ordering.md",19757,"https:\u002F\u002Fgithub.com\u002Fapache\u002Fcassandra\u002Fblob\u002Ftrunk\u002F.claude\u002Fskills\u002Ftargeted-review\u002Freferences\u002Fcategories\u002Flifecycle-and-ordering.md",{"path":107,"resourceType":78,"resourceTypeDesc":79,"sizeBytes":108,"sourceHtmlUrl":109},"references\u002Fcategories\u002Fnull-and-type-safety.md",16039,"https:\u002F\u002Fgithub.com\u002Fapache\u002Fcassandra\u002Fblob\u002Ftrunk\u002F.claude\u002Fskills\u002Ftargeted-review\u002Freferences\u002Fcategories\u002Fnull-and-type-safety.md",{"path":111,"resourceType":78,"resourceTypeDesc":79,"sizeBytes":112,"sourceHtmlUrl":113},"references\u002Fcategories\u002Frefactor-aftermath.md",12613,"https:\u002F\u002Fgithub.com\u002Fapache\u002Fcassandra\u002Fblob\u002Ftrunk\u002F.claude\u002Fskills\u002Ftargeted-review\u002Freferences\u002Fcategories\u002Frefactor-aftermath.md",{"path":115,"resourceType":78,"resourceTypeDesc":79,"sizeBytes":116,"sourceHtmlUrl":117},"references\u002Fcategories\u002Fserialization-and-versioning.md",18659,"https:\u002F\u002Fgithub.com\u002Fapache\u002Fcassandra\u002Fblob\u002Ftrunk\u002F.claude\u002Fskills\u002Ftargeted-review\u002Freferences\u002Fcategories\u002Fserialization-and-versioning.md",{"path":119,"resourceType":78,"resourceTypeDesc":79,"sizeBytes":120,"sourceHtmlUrl":121},"references\u002Fcategories\u002Fstate-and-resource-cleanup.md",16967,"https:\u002F\u002Fgithub.com\u002Fapache\u002Fcassandra\u002Fblob\u002Ftrunk\u002F.claude\u002Fskills\u002Ftargeted-review\u002Freferences\u002Fcategories\u002Fstate-and-resource-cleanup.md",{"path":123,"resourceType":78,"resourceTypeDesc":79,"sizeBytes":124,"sourceHtmlUrl":125},"references\u002Fcategories\u002Fvalidation-and-input-handling.md",18657,"https:\u002F\u002Fgithub.com\u002Fapache\u002Fcassandra\u002Fblob\u002Ftrunk\u002F.claude\u002Fskills\u002Ftargeted-review\u002Freferences\u002Fcategories\u002Fvalidation-and-input-handling.md",{"path":127,"resourceType":78,"resourceTypeDesc":79,"sizeBytes":128,"sourceHtmlUrl":129},"references\u002Freport-format.md",3968,"https:\u002F\u002Fgithub.com\u002Fapache\u002Fcassandra\u002Fblob\u002Ftrunk\u002F.claude\u002Fskills\u002Ftargeted-review\u002Freferences\u002Freport-format.md",{"path":131,"resourceType":78,"resourceTypeDesc":79,"sizeBytes":132,"sourceHtmlUrl":133},"references\u002Fsubagent-prompt-template.md",4726,"https:\u002F\u002Fgithub.com\u002Fapache\u002Fcassandra\u002Fblob\u002Ftrunk\u002F.claude\u002Fskills\u002Ftargeted-review\u002Freferences\u002Fsubagent-prompt-template.md",[135,164,190,217,236,254],{"skillNo":136,"name":137,"nameZh":6,"description":138,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":139,"categories":140,"keywords":144,"publishedAt":47},"01M6ACCP3HAMTRVENVB5ZT9TX1","write-reproducer","Write minimal, reliable bug reproducers (repros) for any project. Use when: (1) a bug report or issue needs a test that reliably demonstrates the failure, (2) converting a described failure scenario into runnable code, (3) minimizing an existing complex test to its essential trigger conditions. Covers the full repro workflow: failure characterization, scope selection, writing the repro, verifying it fails for the right reason, and minimizing to the smallest possible trigger.","https:\u002F\u002Fgithub.com\u002Fapache\u002Fcassandra\u002Fblob\u002Ftrunk\u002F.claude\u002Fskills\u002Fwrite-reproducer\u002FSKILL.md",[141],{"categoryNo":142,"categoryName":143,"taxonomyType":21,"taxonomyTypeDesc":22},"01JBSEEDCAT000000000000008","测试与质量",[145,146,147,148,149,150,151,42,152,153,154,35,155,156,157,158,159,160,161,28,162,163],"write reproducer","reproducer","write","writing","workflow","test","bug report","report","repro","minimal","issue","selection","project","full","essential","scenario","trigger","for","when",{"skillNo":165,"name":166,"nameZh":6,"description":167,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":168,"categories":169,"keywords":170,"publishedAt":47},"01M6ACCJXC5SQSQ2K3APRC6GQV","tla-plus","Create, run, and verify TLA+ and PlusCal formal specifications. Use for modeling distributed systems, protocols, concurrent algorithms, state machines. Can compose specs from code, find divergences between spec and implementation, spot concurrency bugs and invariant violations. Use when asked to \"write a TLA+ spec\", \"model check\", \"verify protocol\", \"find race conditions\", or \"formal verification\".","https:\u002F\u002Fgithub.com\u002Fapache\u002Fcassandra\u002Fblob\u002Ftrunk\u002F.claude\u002Fskills\u002Ftla-plus\u002FSKILL.md",[],[171,172,173,174,175,42,176,177,178,179,180,181,147,182,183,184,185,186,187,162,188,189,163],"tla plus","spec","distributed","create","run","state","protocol","verification","protocols","concurrency","model","use","compose","algorithms","modeling","state machines","TLA+","bugs","model check",{"skillNo":191,"name":192,"nameZh":6,"description":193,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":194,"categories":195,"keywords":204,"publishedAt":47},"01M6ACCCAA36RCWAA93YDDRBX8","shallow-review","Shallow (quick) ensemble bug-finding review using 6 specialist agents in parallel. Each specialist reviews the same patch through a different lens: Logic & Types, Boundaries & I\u002FO, Concurrency & State, Resources & Serialization, Absence Analysis, and API Completeness. Findings are merged and deduplicated. Best for: quick first-pass review of patches, triage of diffs, broad surface-level bug scan. For deeper file-focused review with full pattern catalogs and codebase investigation, use the deep-review skill instead.","https:\u002F\u002Fgithub.com\u002Fapache\u002Fcassandra\u002Fblob\u002Ftrunk\u002F.claude\u002Fskills\u002Fshallow-review\u002FSKILL.md",[196,199],{"categoryNo":197,"categoryName":198,"taxonomyType":21,"taxonomyTypeDesc":22},"01JBSEEDCAT000000000000010","后端与接口",{"categoryNo":200,"categoryName":201,"taxonomyType":202,"taxonomyTypeDesc":203},"01JBSEEDCAT000000000000030","开发工程师","PROFESSION","适用职业",[24,205,206,207,176,180,40,208,39,36,35,209,210,211,30,158,212,213,214,215,216,162],"shallow review","diffs","triage","api","scan","deeper","pattern","resources","lens","types","quick","reviews",{"skillNo":218,"name":219,"nameZh":6,"description":220,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":221,"categories":222,"keywords":224,"publishedAt":47},"01M6ACC9R6QCJN390ERFPN8VA0","patch-explainer","Deep code analysis with ASCII visualizations showing structure, flow, and state transitions. Use when analyzing patches\u002Fdiffs, explaining classes or subsystems, understanding code architecture, reviewing changes for inconsistencies, or when asked to visualize how code works. Provides before\u002Fafter diagrams, data\u002Fcontrol flow, concurrency analysis, assumptions, and failure modes. Triggers on explain this patch\u002Fcode\u002Fclass, how does X work, show me the flow, visualize this change, code review requests, or proactive analysis during PR reviews.","https:\u002F\u002Fgithub.com\u002Fapache\u002Fcassandra\u002Fblob\u002Ftrunk\u002F.claude\u002Fskills\u002Fpatch-explainer\u002FSKILL.md",[223],{"categoryNo":19,"categoryName":20,"taxonomyType":21,"taxonomyTypeDesc":22},[225,226,36,27,24,206,227,45,228,43,42,176,40,229,230,231,232,233,234,30,235,216,163],"patch explainer","explainer","visualize","architecture","show me","work","diagrams","show","me","structure","code analysis",{"skillNo":237,"name":238,"nameZh":6,"description":239,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":240,"categories":241,"keywords":243,"publishedAt":47},"01M6ACC7FFFSHENM9P5SGMWGQJ","mega-review","Multi-pass deep code review for large patches (1000+ LOC) that maximizes real bug detection. Orchestrates targeted-review, shallow-review, and deep-review in parallel across all HIGH and MEDIUM risk files and commits: understands the feature holistically, splits by file and commit, runs deep review on every HIGH\u002FMEDIUM file, targeted review across the same scope, and per-commit shallow review, followed by cross-cut consistency checks. Use when: reviewing a large patch (feature branch, multi-commit, or single large diff, 1000+ LOC), doing a thorough pre-merge review, or when shallower reviews miss bugs due to patch size. Triggers on: \"review this branch\", \"review these commits\", \"review this feature\", \"mega review\", \"thorough review\", \"full review of X commits\", or when the user specifies a commit range or feature for review.","https:\u002F\u002Fgithub.com\u002Fapache\u002Fcassandra\u002Fblob\u002Ftrunk\u002F.claude\u002Fskills\u002Fmega-review\u002FSKILL.md",[242],{"categoryNo":19,"categoryName":20,"taxonomyType":21,"taxonomyTypeDesc":22},[24,244,27,245,246,42,41,247,36,35,248,249,30,158,250,28,251,216,252,253,163],"mega review","commit","merge","branch","files","across","cut","full review","pre-merge","merge review",{"skillNo":255,"name":256,"nameZh":6,"description":257,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":258,"categories":259,"keywords":260,"publishedAt":47},"01M6ACC5GEW6PB6RYG58VJ58SQ","heatmap","Use git heatmap analysis to identify high-churn files and lines as candidates for thorough review or bug hunting. Works for PR reviews, security audits, bug hunts, or any code analysis task.","https:\u002F\u002Fgithub.com\u002Fapache\u002Fcassandra\u002Fblob\u002Ftrunk\u002F.claude\u002Fskills\u002Fheatmap\u002FSKILL.md",[],[256,24,261,43,42,262,40,182,35,248,263,29,235,264,265,266,216,267,162],"security","git","task","churn","heatmap analysis","hunting","high"]