[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M75S26DH1WMZBFG5ZPCVE6VG":3,"skill-same:01M75S26DH1WMZBFG5ZPCVE6VG":289},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":15,"categories":16,"keywords":17,"publishedAt":18,"linkOnly":19,"markdownContent":20,"markdownBody":21,"frontMatter":22,"refName":24,"commitSha":25,"skillDirectory":26,"forksCount":27,"repositoryHtmlUrl":28,"repositoryHomepageUrl":6,"ownerHtmlUrl":29,"ownerBlogUrl":30,"ownerDescription":6,"updatedAt":31,"install":32,"resources":45},"01M75S26DH1WMZBFG5ZPCVE6VG","code-review",null,"No-mercy code review — security (injections first), correctness, performance — of a PR, uncommitted changes or a whole project, with grounded, verified findings. Use for \"review this\", \"проверь код\", \"найди уязвимости\".","paranoica\u002Fbedrock","paranoica","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F57444931?v=4","01M75S0NENDBV4XTD2WQ53TJ9F","01M75S0NF7JWQ0SKVCWWPFJFS2",3,"MIT","https:\u002F\u002Fgithub.com\u002Fparanoica\u002Fbedrock\u002Fblob\u002Fmain\u002Fplugins\u002Fbedrock\u002Fskills\u002Fcode-review\u002FSKILL.md",[],[],"2026-11-12 13:55:43",false,"---\nname: code-review\nversion: 2.1.1\ndescription: No-mercy code review — security (injections first), correctness, performance — of a PR, uncommitted changes or a whole project, with grounded, verified findings. Use for \"review this\", \"проверь код\", \"найди уязвимости\".\n---\n\n# Code Review Skill\n\nA senior-engineer-level reviewer that doesn't soften, doesn't pad, and doesn't praise\nunless something is genuinely worth praising. Focused on **security** (extra paranoia\naround injections and memory-safety), **performance**, and **maintainability**.\n\n## Read this first, every review\n\n1. **`index.json`** — the router + state map. It tells you which files this task needs\n   (lazy loading) and where durable state lives. Read it before anything else.\n2. **`references\u002Finvariants.md`** — the compact non-negotiable core. Re-read it at the\n   start, before writing `findings.json`, and at the grounding gate (reinjection — this\n   fights context rot in long sessions; the invariants are short so it's cheap).\n\nThe full reasoning behind the invariants (anti-hallucination, anti-laziness, fix-safety,\nthe false-positive catalog) lives in **`references\u002Freview-discipline.md`** — pull it at the\ngrounding gate (Step 5c) and before applying a fix, not for the whole review. The one rule\nunder all of it: **a finding = a tainted source AND an unsafe sink in the same flow, grounded\nin a line you actually read** — never an assumed pattern, a fabricated CVE\u002Fline, or a real bug\nrationalized away.\n\n## Routing — pick a mode\n\nRead the user's request and pick ONE mode (file paths in `index.json` → `mode_routes`).\nIf unclear, ask.\n\n| User says... | Mode |\n|---|---|\n| \"review my project\", \"audit the whole codebase\", \"посмотри весь проект\" | **full-project** |\n| \"review my changes\", \"проверь что я наклепал\", \"check uncommitted\", git diff\u002Fstaged\u002Funstaged | **uncommitted** |\n| \"review PR #123\", \"проверь чужой PR\", a GitHub PR URL | **pr-review** |\n\nRead the mode's file in `modes\u002F` for its workflow, then continue here.\n\n## Universal workflow (after mode-specific setup)\n\nOnce the mode has produced **a set of files\u002Fdiff to review**, do this.\n\n### Step 0 — Preflight (environment capability, run ONCE)\nRun `scripts\u002Fpreflight.sh` (`--no-net` offline). It probes the static-analysis toolchain\n(semgrep\u002Fgitleaks\u002Fbandit\u002Feslint), `python3`\u002F`node`, OSV reachability, **and the PoC-execution\nisolation backend** (Step 5d), returning a JSON capability report with a derived `mode`\n(`full` | `degraded`) and `install_hints`.\n\n**Announce the operating mode in one line, up front.** `mode: full` → proceed silently.\n`mode: degraded` → e.g. \"Degraded: no semgrep\u002Fbandit, OSV unreachable, no PoC isolation →\nLLM review + deterministic verifier; dep-CVE scan and execution-grounding skipped, those\nfindings flagged lower-confidence.\" Then continue — **never abort.** This is the contract\nthat makes every later \"skipped\" honest. Carry the report in working memory.\n\n### Step 0b — Load learned context (suppressions + standards)\nIf `.review\u002F` exists (paths in `index.json` → `state_files`): load `suppressions.json`\n(prior FP signatures → default-suppress unless you have *new evidence* of a real unsafe\npath; test with `scripts\u002Frecord_outcome.py match …`) and `standards.md` (accepted repo\nconventions — a change violating one is itself a finding). First run has neither; they get\nbuilt at close-out (Step 6). This is how the reviewer stops crying wolf on the same pattern.\n\n### Step 1 — Detect stack + infra + versions\nFrom lockfiles and entry points, record which **languages**, **SQL\u002Fdata stores**, and\n**infra\u002Ftooling** are in play (`package.json`\u002Flockfiles → Node\u002FJS\u002FTS, `requirements`\u002F\n`pyproject`\u002F`poetry`\u002F`Pipfile` → Python, `go.mod` → Go, `*.csproj`\u002F`*.sln` → C#;\n`pom.xml`\u002F`build.gradle` → Java\u002FKotlin, `composer.json` → PHP, `Cargo.toml` → Rust;\n`Makefile`\u002F`CMakeLists`\u002F`*.c`\u002F`*.cpp` → C\u002FC++, `*.sh` → Shell, `Dockerfile`,\n`.github\u002Fworkflows`, `*.tf`, `k8s`\u002F`helm` manifests, `nginx.conf`, etc. → infra). These\nselect the modules in Step 5 via `index.json` → `stack_routes` \u002F `infra_routes`.\n\n**Record versions, not just stacks.** Pull the runtime + key framework majors from the\nlockfile. Footguns are version-specific — don't flag what the project's version already\nfixed; do flag what exists in *their* version. State the version and how you confirmed it.\n\n### Step 1b — Diff-mode extras (uncommitted \u002F pr-review only)\n- **Score the change:** `scripts\u002Fchange_risk.py --git \u003Cbase_ref> .` → 0–100 risk, per-file\n  ranking, flags `too_big` \u002F `mixed_concerns` \u002F `risky_no_tests`. Surface up front; use the\n  ranking to decide where to read deepest.\n- **Review the deletions:** read `checklists\u002Fremoved-defenses.md`, walk the diff's `-` lines.\n  A removed auth check \u002F validation \u002F timeout \u002F transaction \u002F test is a finding at the\n  severity of what it protected — unless you confirm it moved and still covers the path.\n\n### Step 1c — Build the repo map (cross-module memory)\nRun `scripts\u002Fbuild_index.py \u003Crepo_root>` once → persistent `.review\u002Findex.json` mapping\nfiles to symbols they **define** and names they **call**, plus reverse edges. Use it before\ncalling a change \"local\": `--callers \u003Csymbol>` (blast radius includes callers outside the\ndiff — go read them) and `--defs \u003Csymbol>` (resolve a sink's real definition instead of\nassuming). It's a coarse map — treat edges as *leads to read*, never findings on their own.\nAbsent\u002Fstale → say \"repo map unavailable, cross-module reasoning limited to files read\".\n\n### Step 2 — Run the tooling layer\nRun `scripts\u002Frun_static_analysis.sh \u003Ctarget> [files_list]` (pass the changed-files list in\ndiff modes so tool output lines up with the diff). Wraps semgrep, gitleaks, bandit, eslint.\nMissing tools land in `tools_skipped` — tell the user once + surface the matching preflight\ninstall hint, then continue. **Capture the JSON; don't dump it.** It's input for Step 5:\ncross-reference, drop noise, keep signal.\n\n### Step 3 — Check CVEs via OSV.dev\nRun `scripts\u002Fcheck_cves.sh`. Parses lockfiles (npm\u002Fpnpm\u002Fyarn classic+berry; poetry\u002FPipfile\u002F\nrequirements), batch-queries OSV, computes severity from the CVSS vector, carries `fixed`\nversions. If `deps_total > deps_scanned`, say the cap was hit. A lockfile parsing to zero\ndeps is a **tooling gap to report, not \"no vulnerabilities\".** Vulnerable deps go at the top\nas CRITICAL (or computed band). **If OSV is unreachable the scan is SKIPPED, not clean** —\nstate it, escalate the Step 4 web budget to the few highest-risk deps, mark those\nlower-confidence.\n\n### Step 4 — Web-augment for fresh threats\nTargeted searches **only if local data isn't enough**: new CVE classes for the framework,\nspecific pattern uncertainty, a dependency at an unusual version. **Budget: max 5 searches.**\nCache topic+date so you don't repeat.\n\n### Step 5 — LLM-driven review\nFirst study the repo's own defenses: **`references\u002Fsecurity-precision.md` §1** (auth,\nvalidation, escaping, secrets, threat model). A finding is a deviation from them, not a\nguess. Then read **`checklists\u002Ftaint-spine.md`** — the language-agnostic source→sink→sanitizer\nmodel and the finding rule. Then load checklists by relevance via `index.json` (don't read\nall of them — that just burns context):\n- `always_checklists` — injection-deep, security-general, performance, resilience. Every review.\n- `stack_routes` — the language module(s) for the detected stack(s) only.\n- `infra_routes` — the misconfig checklist(s) for detected infra only.\n- `conditional_checklists` — concurrency \u002F migrations \u002F removed-defenses \u002F privacy \u002F finops \u002F\n  architecture \u002F test-quality \u002F supply-chain \u002F llm-slop, each loaded only when the diff\n  actually triggers it.\n\nThen walk the files\u002Fdiff with the checklists open. For each file: skim for entry points;\n**build a taint map** (every public input → trace through the call graph to every sink:\nDB query, shell, file, outbound HTTP, template render, response body, eval-like construct;\nwhen a flow crosses a file boundary resolve the callee with `build_index.py --defs` and its\nblast radius with `--callers`); for each tainted flow check the matching injection rule;\n**verify defenses by reading them** (construct the input that *would* exploit, check the\ndefense stops it); check perf on loops\u002Fqueries; cross-reference tool findings and drop FPs.\n**A finding requires a tainted source AND an unsafe sink in the same flow.** The full\nfalse-positive catalog is in `references\u002Freview-discipline.md`.\n\n### Step 5b — Improvements pass (MANDATORY, even at zero bugs)\nA separate pass. Read `checklists\u002Fimprovements.md` fully (strict rules on what qualifies).\nHard cap **5–7 improvements**, categorized by IMPACT (not severity). Nothing worth saying →\nexactly one line in the user's language (\"Code clean — nothing substantial to suggest.\" \u002F\n\"Код чистый — ничего существенного не предложу.\"). Don't pad.\n\n### Step 5c — Ground every finding before it ships (MANDATORY)\nRe-read `references\u002Finvariants.md` and `references\u002Freview-discipline.md` (the verification\nsection). Then:\n1. Write findings to `findings.json`: `{\"findings\":[{\"id\",\"severity\",\"category\",\"confidence\",\n   \"file\",\"lines\":[s,e],\"quote\":\"\u003Cverbatim code>\",\"cve\":null}, …]}`.\n2. Run `python3 scripts\u002Fverify_findings.py findings.json \u003Crepo_root> \u003Cosv_output.json>`. It\n   writes a hash-bound `.review\u002Fverify-receipt.json` (binds findings + the sha256 of every\n   source file it read + the verdict counts). Put its `receipt_id` on the report's Verification\n   line as `receipt:\u003C12-hex>` — this is what makes the counts un-forgeable.\n3. Honor verdicts: `CUT_NO_FILE`\u002F`CUT_NO_QUOTE`\u002F`CUT_BAD_CVE` → delete; `RELOCATE` → keep,\n   replace lines with `actual_lines`; `NEEDS_HUMAN` → keep only as `needs verification`;\n   `VERIFIED` → ship. When it says CUT, you cut — don't rephrase it back in.\n3b. **Security findings** pass `references\u002Fsecurity-precision.md` §2–§4: a named exploit\n   path, at least 0.8 sure of exploitability for CRITICAL\u002FHIGH, and nothing from the\n   not-a-vulnerability list or against a precedent. Drop or move what fails, before the verifier.\n4. **Self-consistency** for `confidence: low` survivors: re-derive once from scratch; can't\n   reproduce → downgrade or cut.\n4b. **Counterfactual on \"safe\" verdicts:** when you called a spot safe *because of a defense*,\n   construct the specific exploit input and trace whether the defense stops it. If something\n   slips through, it's a finding after all.\n5. **Spawn the verifier subagent** for medium+ reviews (≥5 findings or any CRITICAL\u002FHIGH):\n   `Task(subagent_type:\"bedrock:verifier\", prompt: \u003Ccontents of subagents\u002Ffinding-verifier.md>\n   + the findings JSON + repo root + OSV output path)`. Fresh context → can't be anchored by\n   your reasoning. Treat its cut\u002Fdowngrade as authoritative unless you can point at the exact\n   refuting line. Gate by scope. Fallbacks: a user-registered `finding-verifier` agent by\n   name; else the deterministic verifier + self-consistency still cover grounding.\n\n### Step 5d — Execution-grounding (when safe; strongest receipt)\nFor a CRITICAL\u002FHIGH security finding whose exploitability reproduces in a few lines, an\nexecuted PoC beats reasoning. Gate on ALL: preflight `poc_execution.available: true`; the\nhypothesis is reproducible in a **self-contained snippet** (no live app\u002FDB\u002Fnetwork). Write\nthe minimal PoC with a success marker, run `scripts\u002Fpoc_runner.sh --lang py|js --code '\u003Cpoc>'\n--expect VULN_CONFIRMED` (isolated, no network, resource-limited, repo not writable).\n`confirmed: true` → cite \"Execution: reproduced\". `confirmed: false` → **does NOT prove\nsafety** (likely the PoC was wrong); keep at reasoned severity or downgrade only if the\nreasoning was shaky. `available: false` → leave the finding as reasoned. The runner executes\n**only your tiny PoC** — never the project's app\u002Ftests\u002Fbuild\u002Finstall, never a fix.\n\n**Probing the project's own functions.** The runner mounts the filesystem read-only with no\nnetwork, so a PoC may import a module of the project (put the repo on `sys.path` or use an\nabsolute `import()`) and call one of its functions with the crafted input. Use it for HIGH\nor CRITICAL correctness findings too, not only security: \"this parser drops the last\nrecord\", \"this check passes a negated phrase\". Rules: call functions, never the app's entry\npoints, servers, CLIs, tests, builds or installs; nothing that needs a network, a database or\na write outside the runner's temp folder; a module whose import itself starts work (connects,\nspawns, schedules) is out. A reproduced probe is cited as \"Execution: reproduced (project\ncode)\".\n\n### Step 6 — Write the report + close out\nOutput format is fixed: read `references\u002Freport-template.md` and follow it. **Adapt detail\nto scope** (tiny diff → conversational enumerate; PR → full template tight; big audit →\ngroup by file, count table at top). Order: vulnerable deps → CRITICAL → HIGH → MEDIUM → LOW\n→ Improvements → Summary (2–3 lines, brutally honest). Never split related issues to inflate\nthe count. No emojis, no padding, no \"I hope this helps.\"\n\n**Self-check the report before shipping it.** Run `python3 scripts\u002Fcheck_report.py \u003Creport.md>\n--receipt .review\u002Fverify-receipt.json --repo \u003Crepo_root>` (or pipe the rendered report on stdin\nwith `-`). It deterministically fails if the mandatory receipts are missing — the **Verification**\nline with its tier, the literal `verify_findings.py` counts (VERIFIED\u002FRELOCATE\u002FCUT\u002FNEEDS_HUMAN),\n**and its hash-bound `receipt:\u003C12-hex>` id** — the Coverage & limits block, the Scope\u002FStack\u002FTools\nlines, or if an emoji slipped in. With `--receipt` it goes further: it recomputes the receipt id,\nre-hashes the cited source files **from disk**, and confirms the counts equal the receipt — so a\nhand-typed Verification line cannot pass without `verify_findings.py` having actually run against\nthe real files. A non-zero exit means the grounding gate (Step 5c) left no honest receipt: fix the\nreport, don't ship around it. This turns the\n\"mandatory\" sections from prose you can quietly skip into an omission that is *visible*.\n\n**Close-out (learning loop).** When findings get resolved, log outcomes with\n`scripts\u002Frecord_outcome.py {tp|fp} …` — feeds the Brier scorecard, records FP signatures into\n`.review\u002Fsuppressions.json` and accepted conventions into `.review\u002Fstandards.md`, so the next\nreview on this repo is smarter. Commit those two; `.review\u002Findex.json` and `outcomes.jsonl`\nare rebuildable\u002Flocal. **Tune for precision first, per category** (raw SQL \u002F auth \u002F crypto\ncan run hotter than style\u002Fperf nits); when you have a labeled set, score it by **F1**, not\ncatch-rate.\n\nConfidence maps to a true-positive probability (high≈0.9, medium≈0.65, needs-verification≈0.35;\n`scripts\u002Fcalibration.py p \u003Clabel>`). Check `scripts\u002Fcalibration.py report` — the Brier score\ntells you whether \"high confidence\" actually means it, so labels stay honest over time.\n\n## Cross-skill: working with design-creator\n\nCooperate without looping (see design-creator's `tools\u002Fhandoff.md`).\n\n**Activation boundary (who answers a bare \"what do you think\" on a web page).** This skill owns\n*\"is it correct\u002Fsafe?\"* — bugs, security, perf, on **existing** code. design-creator owns\n*\"build\u002Fredesign the visual layer.\"* When the user shares frontend code and the intent is\nambiguous, this skill takes it: a review is read-only and non-destructive, and it *offers* the\ndesign handoff if findings turn out design-dominated. A request to **design\u002Fredesign\u002Fbuild** a\nUI is design-creator's, not a review. This deterministic split is what stops both skills matching\nthe same message.\n\n**Spawned BY design-creator** (prompt contains `HANDOFF=design-creator`, review-only): review\nthe listed files only; focus on fatal\u002Fsecurity bugs, frontend-safety, accessibility, CWV\u002F\nbundle\u002Frender perf, supply-chain on added deps. **Return findings only — do not auto-fix, do\nnot hand off onward** (you are the leaf, depth=1).\n\n**Run normally, hitting design\u002Ffrontend issues:** if findings are dominated by design\u002Fa11y\u002FUX,\n**suggest** — never auto-launch — handing them to the design engine with a scoped prompt. Wait\nfor the user's yes. Suggest-only is what keeps the two from ping-ponging.\n\n## When to STOP and ask\n\n- Project >500 source files with full-project requested → say it'll take many passes, ask to\n  scope (most-risky module first, or changes since tag X).\n- A file has clear malicious intent (backdoor, exfil) → stop, flag loudly, ask if known.\n- Empty diff \u002F nonexistent PR → say so, don't fabricate.\n\n## What this skill does NOT do\n\n- **Does not auto-fix.** It reports; the user can ask \"apply the fix for #3\" separately.\n- **Does not run the project's app\u002Ftests\u002Fbuild\u002Finstall.** Static by default. The one guarded\n  exception is Step 5d: a minimal self-contained PoC snippet it wrote, under isolation, only\n  when preflight reports a backend. No isolation ⇒ no execution. `scripts\u002F` are read-only\n  analyzers; `poc_runner.sh` is the one sandboxed executor: it runs your PoC, which may call\n  a project function (Step 5d), never the project's app, tests, build or install.\n- **Does not approve or merge.** Output is a report.\n","# Code Review Skill\n\nA senior-engineer-level reviewer that doesn't soften, doesn't pad, and doesn't praise\nunless something is genuinely worth praising. Focused on **security** (extra paranoia\naround injections and memory-safety), **performance**, and **maintainability**.\n\n## Read this first, every review\n\n1. **`index.json`** — the router + state map. It tells you which files this task needs\n   (lazy loading) and where durable state lives. Read it before anything else.\n2. **`references\u002Finvariants.md`** — the compact non-negotiable core. Re-read it at the\n   start, before writing `findings.json`, and at the grounding gate (reinjection — this\n   fights context rot in long sessions; the invariants are short so it's cheap).\n\nThe full reasoning behind the invariants (anti-hallucination, anti-laziness, fix-safety,\nthe false-positive catalog) lives in **`references\u002Freview-discipline.md`** — pull it at the\ngrounding gate (Step 5c) and before applying a fix, not for the whole review. The one rule\nunder all of it: **a finding = a tainted source AND an unsafe sink in the same flow, grounded\nin a line you actually read** — never an assumed pattern, a fabricated CVE\u002Fline, or a real bug\nrationalized away.\n\n## Routing — pick a mode\n\nRead the user's request and pick ONE mode (file paths in `index.json` → `mode_routes`).\nIf unclear, ask.\n\n| User says... | Mode |\n|---|---|\n| \"review my project\", \"audit the whole codebase\", \"посмотри весь проект\" | **full-project** |\n| \"review my changes\", \"проверь что я наклепал\", \"check uncommitted\", git diff\u002Fstaged\u002Funstaged | **uncommitted** |\n| \"review PR #123\", \"проверь чужой PR\", a GitHub PR URL | **pr-review** |\n\nRead the mode's file in `modes\u002F` for its workflow, then continue here.\n\n## Universal workflow (after mode-specific setup)\n\nOnce the mode has produced **a set of files\u002Fdiff to review**, do this.\n\n### Step 0 — Preflight (environment capability, run ONCE)\nRun `scripts\u002Fpreflight.sh` (`--no-net` offline). It probes the static-analysis toolchain\n(semgrep\u002Fgitleaks\u002Fbandit\u002Feslint), `python3`\u002F`node`, OSV reachability, **and the PoC-execution\nisolation backend** (Step 5d), returning a JSON capability report with a derived `mode`\n(`full` | `degraded`) and `install_hints`.\n\n**Announce the operating mode in one line, up front.** `mode: full` → proceed silently.\n`mode: degraded` → e.g. \"Degraded: no semgrep\u002Fbandit, OSV unreachable, no PoC isolation →\nLLM review + deterministic verifier; dep-CVE scan and execution-grounding skipped, those\nfindings flagged lower-confidence.\" Then continue — **never abort.** This is the contract\nthat makes every later \"skipped\" honest. Carry the report in working memory.\n\n### Step 0b — Load learned context (suppressions + standards)\nIf `.review\u002F` exists (paths in `index.json` → `state_files`): load `suppressions.json`\n(prior FP signatures → default-suppress unless you have *new evidence* of a real unsafe\npath; test with `scripts\u002Frecord_outcome.py match …`) and `standards.md` (accepted repo\nconventions — a change violating one is itself a finding). First run has neither; they get\nbuilt at close-out (Step 6). This is how the reviewer stops crying wolf on the same pattern.\n\n### Step 1 — Detect stack + infra + versions\nFrom lockfiles and entry points, record which **languages**, **SQL\u002Fdata stores**, and\n**infra\u002Ftooling** are in play (`package.json`\u002Flockfiles → Node\u002FJS\u002FTS, `requirements`\u002F\n`pyproject`\u002F`poetry`\u002F`Pipfile` → Python, `go.mod` → Go, `*.csproj`\u002F`*.sln` → C#;\n`pom.xml`\u002F`build.gradle` → Java\u002FKotlin, `composer.json` → PHP, `Cargo.toml` → Rust;\n`Makefile`\u002F`CMakeLists`\u002F`*.c`\u002F`*.cpp` → C\u002FC++, `*.sh` → Shell, `Dockerfile`,\n`.github\u002Fworkflows`, `*.tf`, `k8s`\u002F`helm` manifests, `nginx.conf`, etc. → infra). These\nselect the modules in Step 5 via `index.json` → `stack_routes` \u002F `infra_routes`.\n\n**Record versions, not just stacks.** Pull the runtime + key framework majors from the\nlockfile. Footguns are version-specific — don't flag what the project's version already\nfixed; do flag what exists in *their* version. State the version and how you confirmed it.\n\n### Step 1b — Diff-mode extras (uncommitted \u002F pr-review only)\n- **Score the change:** `scripts\u002Fchange_risk.py --git \u003Cbase_ref> .` → 0–100 risk, per-file\n  ranking, flags `too_big` \u002F `mixed_concerns` \u002F `risky_no_tests`. Surface up front; use the\n  ranking to decide where to read deepest.\n- **Review the deletions:** read `checklists\u002Fremoved-defenses.md`, walk the diff's `-` lines.\n  A removed auth check \u002F validation \u002F timeout \u002F transaction \u002F test is a finding at the\n  severity of what it protected — unless you confirm it moved and still covers the path.\n\n### Step 1c — Build the repo map (cross-module memory)\nRun `scripts\u002Fbuild_index.py \u003Crepo_root>` once → persistent `.review\u002Findex.json` mapping\nfiles to symbols they **define** and names they **call**, plus reverse edges. Use it before\ncalling a change \"local\": `--callers \u003Csymbol>` (blast radius includes callers outside the\ndiff — go read them) and `--defs \u003Csymbol>` (resolve a sink's real definition instead of\nassuming). It's a coarse map — treat edges as *leads to read*, never findings on their own.\nAbsent\u002Fstale → say \"repo map unavailable, cross-module reasoning limited to files read\".\n\n### Step 2 — Run the tooling layer\nRun `scripts\u002Frun_static_analysis.sh \u003Ctarget> [files_list]` (pass the changed-files list in\ndiff modes so tool output lines up with the diff). Wraps semgrep, gitleaks, bandit, eslint.\nMissing tools land in `tools_skipped` — tell the user once + surface the matching preflight\ninstall hint, then continue. **Capture the JSON; don't dump it.** It's input for Step 5:\ncross-reference, drop noise, keep signal.\n\n### Step 3 — Check CVEs via OSV.dev\nRun `scripts\u002Fcheck_cves.sh`. Parses lockfiles (npm\u002Fpnpm\u002Fyarn classic+berry; poetry\u002FPipfile\u002F\nrequirements), batch-queries OSV, computes severity from the CVSS vector, carries `fixed`\nversions. If `deps_total > deps_scanned`, say the cap was hit. A lockfile parsing to zero\ndeps is a **tooling gap to report, not \"no vulnerabilities\".** Vulnerable deps go at the top\nas CRITICAL (or computed band). **If OSV is unreachable the scan is SKIPPED, not clean** —\nstate it, escalate the Step 4 web budget to the few highest-risk deps, mark those\nlower-confidence.\n\n### Step 4 — Web-augment for fresh threats\nTargeted searches **only if local data isn't enough**: new CVE classes for the framework,\nspecific pattern uncertainty, a dependency at an unusual version. **Budget: max 5 searches.**\nCache topic+date so you don't repeat.\n\n### Step 5 — LLM-driven review\nFirst study the repo's own defenses: **`references\u002Fsecurity-precision.md` §1** (auth,\nvalidation, escaping, secrets, threat model). A finding is a deviation from them, not a\nguess. Then read **`checklists\u002Ftaint-spine.md`** — the language-agnostic source→sink→sanitizer\nmodel and the finding rule. Then load checklists by relevance via `index.json` (don't read\nall of them — that just burns context):\n- `always_checklists` — injection-deep, security-general, performance, resilience. Every review.\n- `stack_routes` — the language module(s) for the detected stack(s) only.\n- `infra_routes` — the misconfig checklist(s) for detected infra only.\n- `conditional_checklists` — concurrency \u002F migrations \u002F removed-defenses \u002F privacy \u002F finops \u002F\n  architecture \u002F test-quality \u002F supply-chain \u002F llm-slop, each loaded only when the diff\n  actually triggers it.\n\nThen walk the files\u002Fdiff with the checklists open. For each file: skim for entry points;\n**build a taint map** (every public input → trace through the call graph to every sink:\nDB query, shell, file, outbound HTTP, template render, response body, eval-like construct;\nwhen a flow crosses a file boundary resolve the callee with `build_index.py --defs` and its\nblast radius with `--callers`); for each tainted flow check the matching injection rule;\n**verify defenses by reading them** (construct the input that *would* exploit, check the\ndefense stops it); check perf on loops\u002Fqueries; cross-reference tool findings and drop FPs.\n**A finding requires a tainted source AND an unsafe sink in the same flow.** The full\nfalse-positive catalog is in `references\u002Freview-discipline.md`.\n\n### Step 5b — Improvements pass (MANDATORY, even at zero bugs)\nA separate pass. Read `checklists\u002Fimprovements.md` fully (strict rules on what qualifies).\nHard cap **5–7 improvements**, categorized by IMPACT (not severity). Nothing worth saying →\nexactly one line in the user's language (\"Code clean — nothing substantial to suggest.\" \u002F\n\"Код чистый — ничего существенного не предложу.\"). Don't pad.\n\n### Step 5c — Ground every finding before it ships (MANDATORY)\nRe-read `references\u002Finvariants.md` and `references\u002Freview-discipline.md` (the verification\nsection). Then:\n1. Write findings to `findings.json`: `{\"findings\":[{\"id\",\"severity\",\"category\",\"confidence\",\n   \"file\",\"lines\":[s,e],\"quote\":\"\u003Cverbatim code>\",\"cve\":null}, …]}`.\n2. Run `python3 scripts\u002Fverify_findings.py findings.json \u003Crepo_root> \u003Cosv_output.json>`. It\n   writes a hash-bound `.review\u002Fverify-receipt.json` (binds findings + the sha256 of every\n   source file it read + the verdict counts). Put its `receipt_id` on the report's Verification\n   line as `receipt:\u003C12-hex>` — this is what makes the counts un-forgeable.\n3. Honor verdicts: `CUT_NO_FILE`\u002F`CUT_NO_QUOTE`\u002F`CUT_BAD_CVE` → delete; `RELOCATE` → keep,\n   replace lines with `actual_lines`; `NEEDS_HUMAN` → keep only as `needs verification`;\n   `VERIFIED` → ship. When it says CUT, you cut — don't rephrase it back in.\n3b. **Security findings** pass `references\u002Fsecurity-precision.md` §2–§4: a named exploit\n   path, at least 0.8 sure of exploitability for CRITICAL\u002FHIGH, and nothing from the\n   not-a-vulnerability list or against a precedent. Drop or move what fails, before the verifier.\n4. **Self-consistency** for `confidence: low` survivors: re-derive once from scratch; can't\n   reproduce → downgrade or cut.\n4b. **Counterfactual on \"safe\" verdicts:** when you called a spot safe *because of a defense*,\n   construct the specific exploit input and trace whether the defense stops it. If something\n   slips through, it's a finding after all.\n5. **Spawn the verifier subagent** for medium+ reviews (≥5 findings or any CRITICAL\u002FHIGH):\n   `Task(subagent_type:\"bedrock:verifier\", prompt: \u003Ccontents of subagents\u002Ffinding-verifier.md>\n   + the findings JSON + repo root + OSV output path)`. Fresh context → can't be anchored by\n   your reasoning. Treat its cut\u002Fdowngrade as authoritative unless you can point at the exact\n   refuting line. Gate by scope. Fallbacks: a user-registered `finding-verifier` agent by\n   name; else the deterministic verifier + self-consistency still cover grounding.\n\n### Step 5d — Execution-grounding (when safe; strongest receipt)\nFor a CRITICAL\u002FHIGH security finding whose exploitability reproduces in a few lines, an\nexecuted PoC beats reasoning. Gate on ALL: preflight `poc_execution.available: true`; the\nhypothesis is reproducible in a **self-contained snippet** (no live app\u002FDB\u002Fnetwork). Write\nthe minimal PoC with a success marker, run `scripts\u002Fpoc_runner.sh --lang py|js --code '\u003Cpoc>'\n--expect VULN_CONFIRMED` (isolated, no network, resource-limited, repo not writable).\n`confirmed: true` → cite \"Execution: reproduced\". `confirmed: false` → **does NOT prove\nsafety** (likely the PoC was wrong); keep at reasoned severity or downgrade only if the\nreasoning was shaky. `available: false` → leave the finding as reasoned. The runner executes\n**only your tiny PoC** — never the project's app\u002Ftests\u002Fbuild\u002Finstall, never a fix.\n\n**Probing the project's own functions.** The runner mounts the filesystem read-only with no\nnetwork, so a PoC may import a module of the project (put the repo on `sys.path` or use an\nabsolute `import()`) and call one of its functions with the crafted input. Use it for HIGH\nor CRITICAL correctness findings too, not only security: \"this parser drops the last\nrecord\", \"this check passes a negated phrase\". Rules: call functions, never the app's entry\npoints, servers, CLIs, tests, builds or installs; nothing that needs a network, a database or\na write outside the runner's temp folder; a module whose import itself starts work (connects,\nspawns, schedules) is out. A reproduced probe is cited as \"Execution: reproduced (project\ncode)\".\n\n### Step 6 — Write the report + close out\nOutput format is fixed: read `references\u002Freport-template.md` and follow it. **Adapt detail\nto scope** (tiny diff → conversational enumerate; PR → full template tight; big audit →\ngroup by file, count table at top). Order: vulnerable deps → CRITICAL → HIGH → MEDIUM → LOW\n→ Improvements → Summary (2–3 lines, brutally honest). Never split related issues to inflate\nthe count. No emojis, no padding, no \"I hope this helps.\"\n\n**Self-check the report before shipping it.** Run `python3 scripts\u002Fcheck_report.py \u003Creport.md>\n--receipt .review\u002Fverify-receipt.json --repo \u003Crepo_root>` (or pipe the rendered report on stdin\nwith `-`). It deterministically fails if the mandatory receipts are missing — the **Verification**\nline with its tier, the literal `verify_findings.py` counts (VERIFIED\u002FRELOCATE\u002FCUT\u002FNEEDS_HUMAN),\n**and its hash-bound `receipt:\u003C12-hex>` id** — the Coverage & limits block, the Scope\u002FStack\u002FTools\nlines, or if an emoji slipped in. With `--receipt` it goes further: it recomputes the receipt id,\nre-hashes the cited source files **from disk**, and confirms the counts equal the receipt — so a\nhand-typed Verification line cannot pass without `verify_findings.py` having actually run against\nthe real files. A non-zero exit means the grounding gate (Step 5c) left no honest receipt: fix the\nreport, don't ship around it. This turns the\n\"mandatory\" sections from prose you can quietly skip into an omission that is *visible*.\n\n**Close-out (learning loop).** When findings get resolved, log outcomes with\n`scripts\u002Frecord_outcome.py {tp|fp} …` — feeds the Brier scorecard, records FP signatures into\n`.review\u002Fsuppressions.json` and accepted conventions into `.review\u002Fstandards.md`, so the next\nreview on this repo is smarter. Commit those two; `.review\u002Findex.json` and `outcomes.jsonl`\nare rebuildable\u002Flocal. **Tune for precision first, per category** (raw SQL \u002F auth \u002F crypto\ncan run hotter than style\u002Fperf nits); when you have a labeled set, score it by **F1**, not\ncatch-rate.\n\nConfidence maps to a true-positive probability (high≈0.9, medium≈0.65, needs-verification≈0.35;\n`scripts\u002Fcalibration.py p \u003Clabel>`). Check `scripts\u002Fcalibration.py report` — the Brier score\ntells you whether \"high confidence\" actually means it, so labels stay honest over time.\n\n## Cross-skill: working with design-creator\n\nCooperate without looping (see design-creator's `tools\u002Fhandoff.md`).\n\n**Activation boundary (who answers a bare \"what do you think\" on a web page).** This skill owns\n*\"is it correct\u002Fsafe?\"* — bugs, security, perf, on **existing** code. design-creator owns\n*\"build\u002Fredesign the visual layer.\"* When the user shares frontend code and the intent is\nambiguous, this skill takes it: a review is read-only and non-destructive, and it *offers* the\ndesign handoff if findings turn out design-dominated. A request to **design\u002Fredesign\u002Fbuild** a\nUI is design-creator's, not a review. This deterministic split is what stops both skills matching\nthe same message.\n\n**Spawned BY design-creator** (prompt contains `HANDOFF=design-creator`, review-only): review\nthe listed files only; focus on fatal\u002Fsecurity bugs, frontend-safety, accessibility, CWV\u002F\nbundle\u002Frender perf, supply-chain on added deps. **Return findings only — do not auto-fix, do\nnot hand off onward** (you are the leaf, depth=1).\n\n**Run normally, hitting design\u002Ffrontend issues:** if findings are dominated by design\u002Fa11y\u002FUX,\n**suggest** — never auto-launch — handing them to the design engine with a scoped prompt. Wait\nfor the user's yes. Suggest-only is what keeps the two from ping-ponging.\n\n## When to STOP and ask\n\n- Project >500 source files with full-project requested → say it'll take many passes, ask to\n  scope (most-risky module first, or changes since tag X).\n- A file has clear malicious intent (backdoor, exfil) → stop, flag loudly, ask if known.\n- Empty diff \u002F nonexistent PR → say so, don't fabricate.\n\n## What this skill does NOT do\n\n- **Does not auto-fix.** It reports; the user can ask \"apply the fix for #3\" separately.\n- **Does not run the project's app\u002Ftests\u002Fbuild\u002Finstall.** Static by default. The one guarded\n  exception is Step 5d: a minimal self-contained PoC snippet it wrote, under isolation, only\n  when preflight reports a backend. No isolation ⇒ no execution. `scripts\u002F` are read-only\n  analyzers; `poc_runner.sh` is the one sandboxed executor: it runs your PoC, which may call\n  a project function (Step 5d), never the project's app, tests, build or install.\n- **Does not approve or merge.** Output is a report.\n",{"name":5,"version":23,"description":7},"2.1.1","main","ff167da1e0a6cfc5bdb942eb3e109845f4f88d23","plugins\u002Fbedrock\u002Fskills\u002Fcode-review",1,"https:\u002F\u002Fgithub.com\u002Fparanoica\u002Fbedrock","https:\u002F\u002Fgithub.com\u002Fparanoica","https:\u002F\u002Fgamesense.pub\u002F","2026-11-08 01:01:15",{"source":8,"skillName":5,"commands":33,"prompt":43,"zipAvailable":44},[34,37,40],{"tool":35,"command":36},"npx","npx skills add paranoica\u002Fbedrock --skill code-review",{"tool":38,"command":39},"bunx","bunx skills add paranoica\u002Fbedrock --skill code-review",{"tool":41,"command":42},"pnpm","pnpm dlx skills add paranoica\u002Fbedrock --skill code-review","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Fparanoica\u002Fbedrock\u002Fblob\u002Fmain\u002Fplugins\u002Fbedrock\u002Fskills\u002Fcode-review\u002FSKILL.md",true,[46,52,56,60,64,68,72,76,80,84,88,92,95,99,103,107,111,115,119,123,127,131,135,139,143,147,151,155,159,163,167,171,175,179,183,187,191,195,199,203,207,211,215,219,223,227,231,235,239,245,249,253,257,261,265,269,273,277,281,285],{"path":47,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":50,"sourceHtmlUrl":51},"checklists\u002Farchitecture.md","REFERENCE","参考文档",3627,"https:\u002F\u002Fgithub.com\u002Fparanoica\u002Fbedrock\u002Fblob\u002Fmain\u002Fplugins\u002Fbedrock\u002Fskills\u002Fcode-review\u002Fchecklists\u002Farchitecture.md",{"path":53,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":54,"sourceHtmlUrl":55},"checklists\u002Fconcurrency-and-data-integrity.md",4567,"https:\u002F\u002Fgithub.com\u002Fparanoica\u002Fbedrock\u002Fblob\u002Fmain\u002Fplugins\u002Fbedrock\u002Fskills\u002Fcode-review\u002Fchecklists\u002Fconcurrency-and-data-integrity.md",{"path":57,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":58,"sourceHtmlUrl":59},"checklists\u002Ffinops.md",3313,"https:\u002F\u002Fgithub.com\u002Fparanoica\u002Fbedrock\u002Fblob\u002Fmain\u002Fplugins\u002Fbedrock\u002Fskills\u002Fcode-review\u002Fchecklists\u002Ffinops.md",{"path":61,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":62,"sourceHtmlUrl":63},"checklists\u002Fimprovements.md",11720,"https:\u002F\u002Fgithub.com\u002Fparanoica\u002Fbedrock\u002Fblob\u002Fmain\u002Fplugins\u002Fbedrock\u002Fskills\u002Fcode-review\u002Fchecklists\u002Fimprovements.md",{"path":65,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":66,"sourceHtmlUrl":67},"checklists\u002Finfra\u002Fci-cd.md",5073,"https:\u002F\u002Fgithub.com\u002Fparanoica\u002Fbedrock\u002Fblob\u002Fmain\u002Fplugins\u002Fbedrock\u002Fskills\u002Fcode-review\u002Fchecklists\u002Finfra\u002Fci-cd.md",{"path":69,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":70,"sourceHtmlUrl":71},"checklists\u002Finfra\u002Fcloud-aws.md",3868,"https:\u002F\u002Fgithub.com\u002Fparanoica\u002Fbedrock\u002Fblob\u002Fmain\u002Fplugins\u002Fbedrock\u002Fskills\u002Fcode-review\u002Fchecklists\u002Finfra\u002Fcloud-aws.md",{"path":73,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":74,"sourceHtmlUrl":75},"checklists\u002Finfra\u002Fdocker.md",3548,"https:\u002F\u002Fgithub.com\u002Fparanoica\u002Fbedrock\u002Fblob\u002Fmain\u002Fplugins\u002Fbedrock\u002Fskills\u002Fcode-review\u002Fchecklists\u002Finfra\u002Fdocker.md",{"path":77,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":78,"sourceHtmlUrl":79},"checklists\u002Finfra\u002Fgraphql.md",3235,"https:\u002F\u002Fgithub.com\u002Fparanoica\u002Fbedrock\u002Fblob\u002Fmain\u002Fplugins\u002Fbedrock\u002Fskills\u002Fcode-review\u002Fchecklists\u002Finfra\u002Fgraphql.md",{"path":81,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":82,"sourceHtmlUrl":83},"checklists\u002Finfra\u002Fkubernetes.md",3682,"https:\u002F\u002Fgithub.com\u002Fparanoica\u002Fbedrock\u002Fblob\u002Fmain\u002Fplugins\u002Fbedrock\u002Fskills\u002Fcode-review\u002Fchecklists\u002Finfra\u002Fkubernetes.md",{"path":85,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":86,"sourceHtmlUrl":87},"checklists\u002Finfra\u002Fpostgres-mysql.md",5330,"https:\u002F\u002Fgithub.com\u002Fparanoica\u002Fbedrock\u002Fblob\u002Fmain\u002Fplugins\u002Fbedrock\u002Fskills\u002Fcode-review\u002Fchecklists\u002Finfra\u002Fpostgres-mysql.md",{"path":89,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":90,"sourceHtmlUrl":91},"checklists\u002Finfra\u002Fredis.md",4009,"https:\u002F\u002Fgithub.com\u002Fparanoica\u002Fbedrock\u002Fblob\u002Fmain\u002Fplugins\u002Fbedrock\u002Fskills\u002Fcode-review\u002Fchecklists\u002Finfra\u002Fredis.md",{"path":93,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":50,"sourceHtmlUrl":94},"checklists\u002Finfra\u002Fterraform.md","https:\u002F\u002Fgithub.com\u002Fparanoica\u002Fbedrock\u002Fblob\u002Fmain\u002Fplugins\u002Fbedrock\u002Fskills\u002Fcode-review\u002Fchecklists\u002Finfra\u002Fterraform.md",{"path":96,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":97,"sourceHtmlUrl":98},"checklists\u002Finfra\u002Fweb-servers.md",6220,"https:\u002F\u002Fgithub.com\u002Fparanoica\u002Fbedrock\u002Fblob\u002Fmain\u002Fplugins\u002Fbedrock\u002Fskills\u002Fcode-review\u002Fchecklists\u002Finfra\u002Fweb-servers.md",{"path":100,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":101,"sourceHtmlUrl":102},"checklists\u002Finjection-deep.md",13747,"https:\u002F\u002Fgithub.com\u002Fparanoica\u002Fbedrock\u002Fblob\u002Fmain\u002Fplugins\u002Fbedrock\u002Fskills\u002Fcode-review\u002Fchecklists\u002Finjection-deep.md",{"path":104,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":105,"sourceHtmlUrl":106},"checklists\u002Fjavascript-typescript.md",4743,"https:\u002F\u002Fgithub.com\u002Fparanoica\u002Fbedrock\u002Fblob\u002Fmain\u002Fplugins\u002Fbedrock\u002Fskills\u002Fcode-review\u002Fchecklists\u002Fjavascript-typescript.md",{"path":108,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":109,"sourceHtmlUrl":110},"checklists\u002Flang\u002Fc.md",4729,"https:\u002F\u002Fgithub.com\u002Fparanoica\u002Fbedrock\u002Fblob\u002Fmain\u002Fplugins\u002Fbedrock\u002Fskills\u002Fcode-review\u002Fchecklists\u002Flang\u002Fc.md",{"path":112,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":113,"sourceHtmlUrl":114},"checklists\u002Flang\u002FCONTRACT.md",2414,"https:\u002F\u002Fgithub.com\u002Fparanoica\u002Fbedrock\u002Fblob\u002Fmain\u002Fplugins\u002Fbedrock\u002Fskills\u002Fcode-review\u002Fchecklists\u002Flang\u002FCONTRACT.md",{"path":116,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":117,"sourceHtmlUrl":118},"checklists\u002Flang\u002Fcpp.md",3484,"https:\u002F\u002Fgithub.com\u002Fparanoica\u002Fbedrock\u002Fblob\u002Fmain\u002Fplugins\u002Fbedrock\u002Fskills\u002Fcode-review\u002Fchecklists\u002Flang\u002Fcpp.md",{"path":120,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":121,"sourceHtmlUrl":122},"checklists\u002Flang\u002Fcsharp.md",4483,"https:\u002F\u002Fgithub.com\u002Fparanoica\u002Fbedrock\u002Fblob\u002Fmain\u002Fplugins\u002Fbedrock\u002Fskills\u002Fcode-review\u002Fchecklists\u002Flang\u002Fcsharp.md",{"path":124,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":125,"sourceHtmlUrl":126},"checklists\u002Flang\u002Fdart.md",4046,"https:\u002F\u002Fgithub.com\u002Fparanoica\u002Fbedrock\u002Fblob\u002Fmain\u002Fplugins\u002Fbedrock\u002Fskills\u002Fcode-review\u002Fchecklists\u002Flang\u002Fdart.md",{"path":128,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":129,"sourceHtmlUrl":130},"checklists\u002Flang\u002Fgo.md",6268,"https:\u002F\u002Fgithub.com\u002Fparanoica\u002Fbedrock\u002Fblob\u002Fmain\u002Fplugins\u002Fbedrock\u002Fskills\u002Fcode-review\u002Fchecklists\u002Flang\u002Fgo.md",{"path":132,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":133,"sourceHtmlUrl":134},"checklists\u002Flang\u002Fjava.md",4963,"https:\u002F\u002Fgithub.com\u002Fparanoica\u002Fbedrock\u002Fblob\u002Fmain\u002Fplugins\u002Fbedrock\u002Fskills\u002Fcode-review\u002Fchecklists\u002Flang\u002Fjava.md",{"path":136,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":137,"sourceHtmlUrl":138},"checklists\u002Flang\u002Fkotlin.md",2651,"https:\u002F\u002Fgithub.com\u002Fparanoica\u002Fbedrock\u002Fblob\u002Fmain\u002Fplugins\u002Fbedrock\u002Fskills\u002Fcode-review\u002Fchecklists\u002Flang\u002Fkotlin.md",{"path":140,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":141,"sourceHtmlUrl":142},"checklists\u002Flang\u002Fphp.md",4760,"https:\u002F\u002Fgithub.com\u002Fparanoica\u002Fbedrock\u002Fblob\u002Fmain\u002Fplugins\u002Fbedrock\u002Fskills\u002Fcode-review\u002Fchecklists\u002Flang\u002Fphp.md",{"path":144,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":145,"sourceHtmlUrl":146},"checklists\u002Flang\u002Frust.md",4269,"https:\u002F\u002Fgithub.com\u002Fparanoica\u002Fbedrock\u002Fblob\u002Fmain\u002Fplugins\u002Fbedrock\u002Fskills\u002Fcode-review\u002Fchecklists\u002Flang\u002Frust.md",{"path":148,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":149,"sourceHtmlUrl":150},"checklists\u002Flang\u002Fshell.md",3297,"https:\u002F\u002Fgithub.com\u002Fparanoica\u002Fbedrock\u002Fblob\u002Fmain\u002Fplugins\u002Fbedrock\u002Fskills\u002Fcode-review\u002Fchecklists\u002Flang\u002Fshell.md",{"path":152,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":153,"sourceHtmlUrl":154},"checklists\u002Flang\u002Fswift.md",3869,"https:\u002F\u002Fgithub.com\u002Fparanoica\u002Fbedrock\u002Fblob\u002Fmain\u002Fplugins\u002Fbedrock\u002Fskills\u002Fcode-review\u002Fchecklists\u002Flang\u002Fswift.md",{"path":156,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":157,"sourceHtmlUrl":158},"checklists\u002Fllm-slop.md",3826,"https:\u002F\u002Fgithub.com\u002Fparanoica\u002Fbedrock\u002Fblob\u002Fmain\u002Fplugins\u002Fbedrock\u002Fskills\u002Fcode-review\u002Fchecklists\u002Fllm-slop.md",{"path":160,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":161,"sourceHtmlUrl":162},"checklists\u002Fmigrations-and-compat.md",4062,"https:\u002F\u002Fgithub.com\u002Fparanoica\u002Fbedrock\u002Fblob\u002Fmain\u002Fplugins\u002Fbedrock\u002Fskills\u002Fcode-review\u002Fchecklists\u002Fmigrations-and-compat.md",{"path":164,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":165,"sourceHtmlUrl":166},"checklists\u002Fperformance.md",7394,"https:\u002F\u002Fgithub.com\u002Fparanoica\u002Fbedrock\u002Fblob\u002Fmain\u002Fplugins\u002Fbedrock\u002Fskills\u002Fcode-review\u002Fchecklists\u002Fperformance.md",{"path":168,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":169,"sourceHtmlUrl":170},"checklists\u002Fprivacy-and-data-flow.md",3532,"https:\u002F\u002Fgithub.com\u002Fparanoica\u002Fbedrock\u002Fblob\u002Fmain\u002Fplugins\u002Fbedrock\u002Fskills\u002Fcode-review\u002Fchecklists\u002Fprivacy-and-data-flow.md",{"path":172,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":173,"sourceHtmlUrl":174},"checklists\u002Fpython.md",6694,"https:\u002F\u002Fgithub.com\u002Fparanoica\u002Fbedrock\u002Fblob\u002Fmain\u002Fplugins\u002Fbedrock\u002Fskills\u002Fcode-review\u002Fchecklists\u002Fpython.md",{"path":176,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":177,"sourceHtmlUrl":178},"checklists\u002Fremoved-defenses.md",3598,"https:\u002F\u002Fgithub.com\u002Fparanoica\u002Fbedrock\u002Fblob\u002Fmain\u002Fplugins\u002Fbedrock\u002Fskills\u002Fcode-review\u002Fchecklists\u002Fremoved-defenses.md",{"path":180,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":181,"sourceHtmlUrl":182},"checklists\u002Fresilience.md",3736,"https:\u002F\u002Fgithub.com\u002Fparanoica\u002Fbedrock\u002Fblob\u002Fmain\u002Fplugins\u002Fbedrock\u002Fskills\u002Fcode-review\u002Fchecklists\u002Fresilience.md",{"path":184,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":185,"sourceHtmlUrl":186},"checklists\u002Fsecurity-general.md",8907,"https:\u002F\u002Fgithub.com\u002Fparanoica\u002Fbedrock\u002Fblob\u002Fmain\u002Fplugins\u002Fbedrock\u002Fskills\u002Fcode-review\u002Fchecklists\u002Fsecurity-general.md",{"path":188,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":189,"sourceHtmlUrl":190},"checklists\u002Fsql-and-data.md",7126,"https:\u002F\u002Fgithub.com\u002Fparanoica\u002Fbedrock\u002Fblob\u002Fmain\u002Fplugins\u002Fbedrock\u002Fskills\u002Fcode-review\u002Fchecklists\u002Fsql-and-data.md",{"path":192,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":193,"sourceHtmlUrl":194},"checklists\u002Fsupply-chain.md",3393,"https:\u002F\u002Fgithub.com\u002Fparanoica\u002Fbedrock\u002Fblob\u002Fmain\u002Fplugins\u002Fbedrock\u002Fskills\u002Fcode-review\u002Fchecklists\u002Fsupply-chain.md",{"path":196,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":197,"sourceHtmlUrl":198},"checklists\u002Ftaint-spine.md",7278,"https:\u002F\u002Fgithub.com\u002Fparanoica\u002Fbedrock\u002Fblob\u002Fmain\u002Fplugins\u002Fbedrock\u002Fskills\u002Fcode-review\u002Fchecklists\u002Ftaint-spine.md",{"path":200,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":201,"sourceHtmlUrl":202},"checklists\u002Ftest-quality.md",3396,"https:\u002F\u002Fgithub.com\u002Fparanoica\u002Fbedrock\u002Fblob\u002Fmain\u002Fplugins\u002Fbedrock\u002Fskills\u002Fcode-review\u002Fchecklists\u002Ftest-quality.md",{"path":204,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":205,"sourceHtmlUrl":206},"index.json",6507,"https:\u002F\u002Fgithub.com\u002Fparanoica\u002Fbedrock\u002Fblob\u002Fmain\u002Fplugins\u002Fbedrock\u002Fskills\u002Fcode-review\u002Findex.json",{"path":208,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":209,"sourceHtmlUrl":210},"modes\u002Ffull-project.md",2891,"https:\u002F\u002Fgithub.com\u002Fparanoica\u002Fbedrock\u002Fblob\u002Fmain\u002Fplugins\u002Fbedrock\u002Fskills\u002Fcode-review\u002Fmodes\u002Ffull-project.md",{"path":212,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":213,"sourceHtmlUrl":214},"modes\u002Fpr-review.md",2885,"https:\u002F\u002Fgithub.com\u002Fparanoica\u002Fbedrock\u002Fblob\u002Fmain\u002Fplugins\u002Fbedrock\u002Fskills\u002Fcode-review\u002Fmodes\u002Fpr-review.md",{"path":216,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":217,"sourceHtmlUrl":218},"modes\u002Funcommitted.md",2094,"https:\u002F\u002Fgithub.com\u002Fparanoica\u002Fbedrock\u002Fblob\u002Fmain\u002Fplugins\u002Fbedrock\u002Fskills\u002Fcode-review\u002Fmodes\u002Funcommitted.md",{"path":220,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":221,"sourceHtmlUrl":222},"references\u002Finvariants.md",4523,"https:\u002F\u002Fgithub.com\u002Fparanoica\u002Fbedrock\u002Fblob\u002Fmain\u002Fplugins\u002Fbedrock\u002Fskills\u002Fcode-review\u002Freferences\u002Finvariants.md",{"path":224,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":225,"sourceHtmlUrl":226},"references\u002Freport-template.md",5969,"https:\u002F\u002Fgithub.com\u002Fparanoica\u002Fbedrock\u002Fblob\u002Fmain\u002Fplugins\u002Fbedrock\u002Fskills\u002Fcode-review\u002Freferences\u002Freport-template.md",{"path":228,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":229,"sourceHtmlUrl":230},"references\u002Freview-discipline.md",7684,"https:\u002F\u002Fgithub.com\u002Fparanoica\u002Fbedrock\u002Fblob\u002Fmain\u002Fplugins\u002Fbedrock\u002Fskills\u002Fcode-review\u002Freferences\u002Freview-discipline.md",{"path":232,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":233,"sourceHtmlUrl":234},"references\u002Fsecurity-precision.md",4221,"https:\u002F\u002Fgithub.com\u002Fparanoica\u002Fbedrock\u002Fblob\u002Fmain\u002Fplugins\u002Fbedrock\u002Fskills\u002Fcode-review\u002Freferences\u002Fsecurity-precision.md",{"path":236,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":237,"sourceHtmlUrl":238},"references\u002Fseverity-rubric.md",3656,"https:\u002F\u002Fgithub.com\u002Fparanoica\u002Fbedrock\u002Fblob\u002Fmain\u002Fplugins\u002Fbedrock\u002Fskills\u002Fcode-review\u002Freferences\u002Fseverity-rubric.md",{"path":240,"resourceType":241,"resourceTypeDesc":242,"sizeBytes":243,"sourceHtmlUrl":244},"scripts\u002Fbuild_index.py","SCRIPT","脚本",11201,"https:\u002F\u002Fgithub.com\u002Fparanoica\u002Fbedrock\u002Fblob\u002Fmain\u002Fplugins\u002Fbedrock\u002Fskills\u002Fcode-review\u002Fscripts\u002Fbuild_index.py",{"path":246,"resourceType":241,"resourceTypeDesc":242,"sizeBytes":247,"sourceHtmlUrl":248},"scripts\u002Fcalibration.py",3158,"https:\u002F\u002Fgithub.com\u002Fparanoica\u002Fbedrock\u002Fblob\u002Fmain\u002Fplugins\u002Fbedrock\u002Fskills\u002Fcode-review\u002Fscripts\u002Fcalibration.py",{"path":250,"resourceType":241,"resourceTypeDesc":242,"sizeBytes":251,"sourceHtmlUrl":252},"scripts\u002Fchange_risk.py",7237,"https:\u002F\u002Fgithub.com\u002Fparanoica\u002Fbedrock\u002Fblob\u002Fmain\u002Fplugins\u002Fbedrock\u002Fskills\u002Fcode-review\u002Fscripts\u002Fchange_risk.py",{"path":254,"resourceType":241,"resourceTypeDesc":242,"sizeBytes":255,"sourceHtmlUrl":256},"scripts\u002Fcheck_cves.py",17151,"https:\u002F\u002Fgithub.com\u002Fparanoica\u002Fbedrock\u002Fblob\u002Fmain\u002Fplugins\u002Fbedrock\u002Fskills\u002Fcode-review\u002Fscripts\u002Fcheck_cves.py",{"path":258,"resourceType":241,"resourceTypeDesc":242,"sizeBytes":259,"sourceHtmlUrl":260},"scripts\u002Fcheck_cves.sh",438,"https:\u002F\u002Fgithub.com\u002Fparanoica\u002Fbedrock\u002Fblob\u002Fmain\u002Fplugins\u002Fbedrock\u002Fskills\u002Fcode-review\u002Fscripts\u002Fcheck_cves.sh",{"path":262,"resourceType":241,"resourceTypeDesc":242,"sizeBytes":263,"sourceHtmlUrl":264},"scripts\u002Fcheck_report.py",9100,"https:\u002F\u002Fgithub.com\u002Fparanoica\u002Fbedrock\u002Fblob\u002Fmain\u002Fplugins\u002Fbedrock\u002Fskills\u002Fcode-review\u002Fscripts\u002Fcheck_report.py",{"path":266,"resourceType":241,"resourceTypeDesc":242,"sizeBytes":267,"sourceHtmlUrl":268},"scripts\u002Fpoc_runner.sh",5030,"https:\u002F\u002Fgithub.com\u002Fparanoica\u002Fbedrock\u002Fblob\u002Fmain\u002Fplugins\u002Fbedrock\u002Fskills\u002Fcode-review\u002Fscripts\u002Fpoc_runner.sh",{"path":270,"resourceType":241,"resourceTypeDesc":242,"sizeBytes":271,"sourceHtmlUrl":272},"scripts\u002Fpreflight.sh",4209,"https:\u002F\u002Fgithub.com\u002Fparanoica\u002Fbedrock\u002Fblob\u002Fmain\u002Fplugins\u002Fbedrock\u002Fskills\u002Fcode-review\u002Fscripts\u002Fpreflight.sh",{"path":274,"resourceType":241,"resourceTypeDesc":242,"sizeBytes":275,"sourceHtmlUrl":276},"scripts\u002Frecord_outcome.py",5946,"https:\u002F\u002Fgithub.com\u002Fparanoica\u002Fbedrock\u002Fblob\u002Fmain\u002Fplugins\u002Fbedrock\u002Fskills\u002Fcode-review\u002Fscripts\u002Frecord_outcome.py",{"path":278,"resourceType":241,"resourceTypeDesc":242,"sizeBytes":279,"sourceHtmlUrl":280},"scripts\u002Frun_static_analysis.sh",6204,"https:\u002F\u002Fgithub.com\u002Fparanoica\u002Fbedrock\u002Fblob\u002Fmain\u002Fplugins\u002Fbedrock\u002Fskills\u002Fcode-review\u002Fscripts\u002Frun_static_analysis.sh",{"path":282,"resourceType":241,"resourceTypeDesc":242,"sizeBytes":283,"sourceHtmlUrl":284},"scripts\u002Fverify_findings.py",8484,"https:\u002F\u002Fgithub.com\u002Fparanoica\u002Fbedrock\u002Fblob\u002Fmain\u002Fplugins\u002Fbedrock\u002Fskills\u002Fcode-review\u002Fscripts\u002Fverify_findings.py",{"path":286,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":287,"sourceHtmlUrl":288},"subagents\u002Ffinding-verifier.md",4210,"https:\u002F\u002Fgithub.com\u002Fparanoica\u002Fbedrock\u002Fblob\u002Fmain\u002Fplugins\u002Fbedrock\u002Fskills\u002Fcode-review\u002Fsubagents\u002Ffinding-verifier.md",[290,297,304,311,318,325],{"skillNo":291,"name":292,"nameZh":6,"description":293,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":294,"categories":295,"keywords":296,"publishedAt":18},"01M75S4P9CEEETSAKSEJKA76KG","writing","Write or rewrite skills, agent definitions, briefs and rule files by the D013 rules, then check them. Use for \"напиши скилл\", \"поправь агента\", \"write a skill\", \"edit this brief\".","https:\u002F\u002Fgithub.com\u002Fparanoica\u002Fbedrock\u002Fblob\u002Fmain\u002Fplugins\u002Fbedrock\u002Fskills\u002Fwriting\u002FSKILL.md",[],[],{"skillNo":298,"name":299,"nameZh":6,"description":300,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":301,"categories":302,"keywords":303,"publishedAt":18},"01M75S4KZRNVBQP46Q33X62R35","triage","Bug triage — reproduce each report on the current version, group duplicates, set priority, decide fix, defer or decline. Use for \"разбери баги\", \"что с репортами QA\", \"triage\", \"new bug reports\".","https:\u002F\u002Fgithub.com\u002Fparanoica\u002Fbedrock\u002Fblob\u002Fmain\u002Fplugins\u002Fbedrock\u002Fskills\u002Ftriage\u002FSKILL.md",[],[],{"skillNo":305,"name":306,"nameZh":6,"description":307,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":308,"categories":309,"keywords":310,"publishedAt":18},"01M75S4H283C8XRHFM6E7TQHV4","start","Start a project with bedrock — a new one by interview, or an existing one by onboarding; writes the base, AGENTS.md, workspace settings and verified commands. Use for \"новый проект\", \"подключи проект\", \"start a project\".","https:\u002F\u002Fgithub.com\u002Fparanoica\u002Fbedrock\u002Fblob\u002Fmain\u002Fplugins\u002Fbedrock\u002Fskills\u002Fstart\u002FSKILL.md",[],[],{"skillNo":312,"name":313,"nameZh":6,"description":314,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":315,"categories":316,"keywords":317,"publishedAt":18},"01M75S4EVW60KBZJGZWEZJYXY7","review","Review gate for a finished block — deterministic checks, code review, plan-compliance review, silent fixes, rounds until clean. Use for \"проверь\", \"ревью\", \"review this\", before any PR.","https:\u002F\u002Fgithub.com\u002Fparanoica\u002Fbedrock\u002Fblob\u002Fmain\u002Fplugins\u002Fbedrock\u002Fskills\u002Freview\u002FSKILL.md",[],[],{"skillNo":319,"name":320,"nameZh":6,"description":321,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":322,"categories":323,"keywords":324,"publishedAt":18},"01M75S4BY5EMAFTJECKJ6ZC5EP","retro","Weekly retro — read the week's logs, agents, reviews and failures, find what repeats, tune the system, plan a realistic next week. Use for \"ретро\", \"итоги недели\", \"retro\", \"how did the week go\".","https:\u002F\u002Fgithub.com\u002Fparanoica\u002Fbedrock\u002Fblob\u002Fmain\u002Fplugins\u002Fbedrock\u002Fskills\u002Fretro\u002FSKILL.md",[],[],{"skillNo":326,"name":327,"nameZh":6,"description":328,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":329,"categories":330,"keywords":331,"publishedAt":18},"01M75S47TW93TAZ04HD75JJ98K","night","Autonomous work until morning — a pre-sleep check, the allowed list, auto-continue with fuses, a watchdog, a morning report. Use for \"автономно до утра\", \"поработай ночью\", \"work overnight\".","https:\u002F\u002Fgithub.com\u002Fparanoica\u002Fbedrock\u002Fblob\u002Fmain\u002Fplugins\u002Fbedrock\u002Fskills\u002Fnight\u002FSKILL.md",[],[]]