[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M6DNH107TQMJJSQBJ5PZSWKN":3,"skill-same:01M6DNH107TQMJJSQBJ5PZSWKN":64},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":9,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":15,"categories":16,"keywords":17,"publishedAt":38,"linkOnly":39,"markdownContent":40,"markdownBody":41,"frontMatter":42,"refName":43,"commitSha":44,"skillDirectory":45,"forksCount":46,"repositoryHtmlUrl":47,"repositoryHomepageUrl":6,"ownerHtmlUrl":48,"ownerBlogUrl":6,"ownerDescription":6,"updatedAt":49,"install":50,"resources":63},"01M6DNH107TQMJJSQBJ5PZSWKN","pi-delegation",null,"Delegate read-only work through Pi classified workflows or visible Pi Zellij workers using authenticated gpt-6.1-sol. Use for parallel investigation, independent verification or multi-agent synthesis.","0xgleb\u002Fdotconfig","0xgleb","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F39841057?v=4","01M6DNDD83BRMM1SH4R4C6K8GY","01M6DNDDCPH33FEFB6JN6Y2GVD",2,"MIT","https:\u002F\u002Fgithub.com\u002F0xgleb\u002Fdotconfig\u002Fblob\u002Fmaster\u002Fai\u002Fskills\u002Fpi-delegation\u002FSKILL.md",[],[18,19,20,21,22,23,24,25,26,27,28,29,30,31,32,33,34,35,36,37],"delegation","pi delegation","pi","sol","read","multi","zellij","only","read-only","for","workflows","gpt","workers","use","investigation","delegate","gpt-6","work","verification","parallel","2026-11-10 10:12:30",false,"---\nname: pi-delegation\ndescription: Delegate read-only work through Pi classified workflows or visible Pi Zellij workers using authenticated gpt-6.1-sol. Use for parallel investigation, independent verification or multi-agent synthesis.\n---\n\n# Pi delegation\n\nUse the least powerful delegation mode that reliably fits the task. An explicit\nuser choice always wins.\n\n## Choose a mode\n\nWork directly when one agent can complete the task without meaningful\nindependent investigation.\n\nUse Zellij workers when:\n\n- there are only a few independent tasks;\n- each worker can return a direct textual result;\n- no dependency graph, retry loop, shared budget, or automatic synthesis is\n  needed;\n- visible panes are useful to the user.\n\nUse a classified dynamic workflow when:\n\n- tasks depend on earlier results;\n- the work needs fan-out followed by verification or synthesis;\n- competing approaches should be generated and judged;\n- bounded loops, retries, checkpoints, or cancellation are useful;\n- a task-specific program describes the desired behavior better than a fixed\n  schema.\n\nNever use tmux.\n\nIndependent delegation must not block the human foreground. If the parent does\nnot need a result before it can continue useful work, start the classified\nworkflow with `background: true` (or use an isolated visible worker), return\ncontrol immediately, and keep processing human prompts. Await a foreground\nworkflow only when its result is a genuine dependency of the very next parent\naction.\n\n## Current model and harness\n\nUse Pi's classified `workflow` tool with authenticated\n`openai-codex\u002Fgpt-6.1-sol` for every agent lane, including editing criticism,\ninspection, verification and synthesis. Keep execution within Pi; do not launch\nexternal model CLIs or subscription wrappers. If the required model\u002Ftool is\nunavailable, retain the exact incomplete state rather than switching model,\nprovider or harness without owner direction.\n\n## Shared safety\n\n- Give every worker a self-contained task, scope, expected result, and stopping\n  condition.\n- Never access, search, list, expose, or return credential or secret-bearing\n  files.\n- Broad searches must explicitly exclude credential-shaped paths.\n- Start parallel investigation read-only.\n- Never allow parallel agents to modify the same working tree. Agent-initiated\n  fan-out remains read-only; separate mutating worktrees require the owner or\n  another agent launched by the owner to be working on that repository\n  concurrently.\n- Keep agent-initiated workers read-only in the session's assigned checkout.\n  Worktrees are permitted ONLY when the owner or another agent launched by the\n  owner is concurrently working on the same repository as an agent. Isolation\n  alone, dirty local state, ordinary setup, review, builds, and agent-initiated\n  fan-out never justify worktrees.\n- When that owner-authorized concurrency requires a worktree, use only a stable\n  repository-local role slot such as `.worktrees\u002Fsecondary`,\n  `.worktrees\u002Ftertiary`, `.tmp\u002Fworktrees\u002Fsecondary`, or\n  `.tmp\u002Fworktrees\u002Ftertiary`. Never create PR-, ticket-, branch-, timestamp-, or\n  task-named worktree directories.\n- A disposable one-off worktree is an exceptional temporary resource. Put it\n  under `.tmp\u002Fworktrees\u002F\u003Crole-slot>`, record its exact provenance, and make the\n  clanker or agent that created it remove the Git worktree registration, its\n  generated outputs, and the directory immediately after success, failure, or\n  cancellation. If cleanup cannot safely finish, persist the exact path and\n  blocker and make cleanup the first resumed action. Never transfer this disk\n  debt silently to the owner or another agent.\n- A pre-existing owner-managed role slot may be reused and retained. A worktree\n  created by the current clanker or agent must be removed before completion\n  unless the owner explicitly asks to keep it.\n- Treat classifier decisions as policy. Do not evade a block by rewording,\n  obfuscating, or switching tools.\n- Delegation is not a sandbox. Keep scope and capabilities minimal.\n- Report failed, blocked, timed-out, or incomplete work explicitly.\n\n## Zellij workers\n\nRequire an existing Zellij session. Do not silently create a detached session.\nWhen the user explicitly asks to spawn an agent, focusing its new tab or pane is\nallowed. For agent-initiated background delegation, snapshot the exact active\ntab and pane IDs before launch and restore both before returning control. Verify\nthe restoration from structured Zellij state; if exact restoration is\nunavailable, use a classified background workflow instead.\n\nCreate or reuse a tab named `pi-workers` under that focus contract. Start each\nworker in a named pane with an ephemeral non-interactive Pi process:\n\n`pi --print --no-session --tools read,grep,find,ls`\n\nCapture the pane ID returned by Zellij. Use structured pane state to determine\nwhen the process exits and preserve its exit status. Collect the final\nplain-text output with `dump-screen --full`.\n\nPolling, harvesting, and closing background workers must preserve the user's\ncurrent focus and must never send keys to the user's active pane.\n\nHarvest each completed pane's output and exit status promptly, then close the\npane automatically so finished workers do not linger. Keep failed panes visible\nonly long enough to capture diagnostics, then close them too unless the user\nexplicitly asks to preserve worker panes.\n\nIf a pane exits unsuccessfully or produces no usable result, mark that worker as\nfailed. Never infer or invent its answer.\n\nVerify worker claims before using them in consequential edits or conclusions.\nSynthesize centrally in the parent.\n\n## Classified dynamic workflows\n\nWrite task-specific JavaScript for the `workflow` tool. The program may compose\nagents, parallel work, verification, selection, checkpoints, and bounded loops.\nDo not force the request into a predefined workflow template.\n\nEvery workflow must specify the smallest sufficient:\n\n- total agent limit;\n- concurrency limit;\n- per-agent and whole-workflow timeouts (classified children require at least\n  180 seconds per agent so spawn classification, child execution, tool-result\n  classification, and return classification all fit);\n- retry limit;\n- total token budget.\n\nCall agents as `agent(\"focused task\", { cwd?, tools?, model?, thinking? })`.\nPass agent promises directly to `parallel`, for example\n`await parallel([agent(\"first task\"), agent(\"second task\")])`. Every call must\nprovide a focused task, working directory when it differs from the parent, tool\ncapabilities, and model or thinking level when the default is not appropriate.\nSet `model: \"openai-codex\u002Fgpt-6.1-sol\"` for every lane. Discovery, review,\nverification and synthesis use the same current model; role labels do not select\na different tier or imply additional authority.\n\nAssume every agent spawn, tool action, tool result, and returned result will be\nclassified. Handle a blocked boundary as a typed failure. Do not expose blocked\nresult content to later workflow steps.\n\nUse read-only fan-out. Apply edits centrally unless the owner or another agent\nlaunched by the owner is concurrently working on the same repository and the\nmutating workers have separate repository-approved worktrees.\n\nFor high-impact mutations, insert an explicit checkpoint. Non-interactive\ncheckpoints deny rather than approve automatically.\n\nKeep workflow results compact and structured. Preserve evidence, uncertainty,\nfailures, and provenance needed by the synthesizer without forwarding\nunnecessary conversation or tool-result content.\n\n## Verification patterns\n\nPrefer the smallest pattern that fits:\n\n- fan out independent investigation, then synthesize;\n- generate candidates, filter them, then verify the survivor;\n- have one agent produce and another independently challenge;\n- classify inputs, then dispatch only the relevant specialist;\n- loop only when each iteration has a measurable stopping condition.\n\nDo not delegate microtasks whose coordination cost exceeds their value.\n\n## Completion\n\nBefore presenting the result:\n\n- distinguish confirmed results from agent claims;\n- verify consequential claims against source material or tests;\n- identify workers that failed, timed out, or were blocked;\n- state when budgets prevented complete coverage;\n- preserve Zellij panes for inspection;\n- never claim a workflow completed work it did not perform.\n","# Pi delegation\n\nUse the least powerful delegation mode that reliably fits the task. An explicit\nuser choice always wins.\n\n## Choose a mode\n\nWork directly when one agent can complete the task without meaningful\nindependent investigation.\n\nUse Zellij workers when:\n\n- there are only a few independent tasks;\n- each worker can return a direct textual result;\n- no dependency graph, retry loop, shared budget, or automatic synthesis is\n  needed;\n- visible panes are useful to the user.\n\nUse a classified dynamic workflow when:\n\n- tasks depend on earlier results;\n- the work needs fan-out followed by verification or synthesis;\n- competing approaches should be generated and judged;\n- bounded loops, retries, checkpoints, or cancellation are useful;\n- a task-specific program describes the desired behavior better than a fixed\n  schema.\n\nNever use tmux.\n\nIndependent delegation must not block the human foreground. If the parent does\nnot need a result before it can continue useful work, start the classified\nworkflow with `background: true` (or use an isolated visible worker), return\ncontrol immediately, and keep processing human prompts. Await a foreground\nworkflow only when its result is a genuine dependency of the very next parent\naction.\n\n## Current model and harness\n\nUse Pi's classified `workflow` tool with authenticated\n`openai-codex\u002Fgpt-6.1-sol` for every agent lane, including editing criticism,\ninspection, verification and synthesis. Keep execution within Pi; do not launch\nexternal model CLIs or subscription wrappers. If the required model\u002Ftool is\nunavailable, retain the exact incomplete state rather than switching model,\nprovider or harness without owner direction.\n\n## Shared safety\n\n- Give every worker a self-contained task, scope, expected result, and stopping\n  condition.\n- Never access, search, list, expose, or return credential or secret-bearing\n  files.\n- Broad searches must explicitly exclude credential-shaped paths.\n- Start parallel investigation read-only.\n- Never allow parallel agents to modify the same working tree. Agent-initiated\n  fan-out remains read-only; separate mutating worktrees require the owner or\n  another agent launched by the owner to be working on that repository\n  concurrently.\n- Keep agent-initiated workers read-only in the session's assigned checkout.\n  Worktrees are permitted ONLY when the owner or another agent launched by the\n  owner is concurrently working on the same repository as an agent. Isolation\n  alone, dirty local state, ordinary setup, review, builds, and agent-initiated\n  fan-out never justify worktrees.\n- When that owner-authorized concurrency requires a worktree, use only a stable\n  repository-local role slot such as `.worktrees\u002Fsecondary`,\n  `.worktrees\u002Ftertiary`, `.tmp\u002Fworktrees\u002Fsecondary`, or\n  `.tmp\u002Fworktrees\u002Ftertiary`. Never create PR-, ticket-, branch-, timestamp-, or\n  task-named worktree directories.\n- A disposable one-off worktree is an exceptional temporary resource. Put it\n  under `.tmp\u002Fworktrees\u002F\u003Crole-slot>`, record its exact provenance, and make the\n  clanker or agent that created it remove the Git worktree registration, its\n  generated outputs, and the directory immediately after success, failure, or\n  cancellation. If cleanup cannot safely finish, persist the exact path and\n  blocker and make cleanup the first resumed action. Never transfer this disk\n  debt silently to the owner or another agent.\n- A pre-existing owner-managed role slot may be reused and retained. A worktree\n  created by the current clanker or agent must be removed before completion\n  unless the owner explicitly asks to keep it.\n- Treat classifier decisions as policy. Do not evade a block by rewording,\n  obfuscating, or switching tools.\n- Delegation is not a sandbox. Keep scope and capabilities minimal.\n- Report failed, blocked, timed-out, or incomplete work explicitly.\n\n## Zellij workers\n\nRequire an existing Zellij session. Do not silently create a detached session.\nWhen the user explicitly asks to spawn an agent, focusing its new tab or pane is\nallowed. For agent-initiated background delegation, snapshot the exact active\ntab and pane IDs before launch and restore both before returning control. Verify\nthe restoration from structured Zellij state; if exact restoration is\nunavailable, use a classified background workflow instead.\n\nCreate or reuse a tab named `pi-workers` under that focus contract. Start each\nworker in a named pane with an ephemeral non-interactive Pi process:\n\n`pi --print --no-session --tools read,grep,find,ls`\n\nCapture the pane ID returned by Zellij. Use structured pane state to determine\nwhen the process exits and preserve its exit status. Collect the final\nplain-text output with `dump-screen --full`.\n\nPolling, harvesting, and closing background workers must preserve the user's\ncurrent focus and must never send keys to the user's active pane.\n\nHarvest each completed pane's output and exit status promptly, then close the\npane automatically so finished workers do not linger. Keep failed panes visible\nonly long enough to capture diagnostics, then close them too unless the user\nexplicitly asks to preserve worker panes.\n\nIf a pane exits unsuccessfully or produces no usable result, mark that worker as\nfailed. Never infer or invent its answer.\n\nVerify worker claims before using them in consequential edits or conclusions.\nSynthesize centrally in the parent.\n\n## Classified dynamic workflows\n\nWrite task-specific JavaScript for the `workflow` tool. The program may compose\nagents, parallel work, verification, selection, checkpoints, and bounded loops.\nDo not force the request into a predefined workflow template.\n\nEvery workflow must specify the smallest sufficient:\n\n- total agent limit;\n- concurrency limit;\n- per-agent and whole-workflow timeouts (classified children require at least\n  180 seconds per agent so spawn classification, child execution, tool-result\n  classification, and return classification all fit);\n- retry limit;\n- total token budget.\n\nCall agents as `agent(\"focused task\", { cwd?, tools?, model?, thinking? })`.\nPass agent promises directly to `parallel`, for example\n`await parallel([agent(\"first task\"), agent(\"second task\")])`. Every call must\nprovide a focused task, working directory when it differs from the parent, tool\ncapabilities, and model or thinking level when the default is not appropriate.\nSet `model: \"openai-codex\u002Fgpt-6.1-sol\"` for every lane. Discovery, review,\nverification and synthesis use the same current model; role labels do not select\na different tier or imply additional authority.\n\nAssume every agent spawn, tool action, tool result, and returned result will be\nclassified. Handle a blocked boundary as a typed failure. Do not expose blocked\nresult content to later workflow steps.\n\nUse read-only fan-out. Apply edits centrally unless the owner or another agent\nlaunched by the owner is concurrently working on the same repository and the\nmutating workers have separate repository-approved worktrees.\n\nFor high-impact mutations, insert an explicit checkpoint. Non-interactive\ncheckpoints deny rather than approve automatically.\n\nKeep workflow results compact and structured. Preserve evidence, uncertainty,\nfailures, and provenance needed by the synthesizer without forwarding\nunnecessary conversation or tool-result content.\n\n## Verification patterns\n\nPrefer the smallest pattern that fits:\n\n- fan out independent investigation, then synthesize;\n- generate candidates, filter them, then verify the survivor;\n- have one agent produce and another independently challenge;\n- classify inputs, then dispatch only the relevant specialist;\n- loop only when each iteration has a measurable stopping condition.\n\nDo not delegate microtasks whose coordination cost exceeds their value.\n\n## Completion\n\nBefore presenting the result:\n\n- distinguish confirmed results from agent claims;\n- verify consequential claims against source material or tests;\n- identify workers that failed, timed out, or were blocked;\n- state when budgets prevented complete coverage;\n- preserve Zellij panes for inspection;\n- never claim a workflow completed work it did not perform.\n",{"name":5,"description":7},"master","43de034413ae8878259e4a93be6aed03add6ddda","ai\u002Fskills\u002Fpi-delegation",0,"https:\u002F\u002Fgithub.com\u002F0xgleb\u002Fdotconfig","https:\u002F\u002Fgithub.com\u002F0xgleb","2026-11-10 07:35:39",{"source":8,"skillName":5,"commands":51,"prompt":61,"zipAvailable":62},[52,55,58],{"tool":53,"command":54},"npx","npx skills add 0xgleb\u002Fdotconfig --skill pi-delegation",{"tool":56,"command":57},"bunx","bunx skills add 0xgleb\u002Fdotconfig --skill pi-delegation",{"tool":59,"command":60},"pnpm","pnpm dlx skills add 0xgleb\u002Fdotconfig --skill pi-delegation","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002F0xgleb\u002Fdotconfig\u002Fblob\u002Fmaster\u002Fai\u002Fskills\u002Fpi-delegation\u002FSKILL.md",true,[],[65,87,112,137,163,180],{"skillNo":66,"name":67,"nameZh":6,"description":68,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":9,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":69,"categories":70,"keywords":71,"publishedAt":38},"01M6DNKHP7NV7KR552KFZ6EW3R","worktree","Manage worktrees only for concurrent work by the owner or agents launched by the owner on the same repository, or for explicitly requested inventory and cleanup; never create one for isolation alone.","https:\u002F\u002Fgithub.com\u002F0xgleb\u002Fdotconfig\u002Fblob\u002Fmaster\u002Fai\u002Fskills\u002Fworktree\u002FSKILL.md",[],[67,72,35,73,74,75,76,77,78,79,25,80,81,82,83,84,27,85,86],"create","never","worktrees","manage","inventory","repository","fix","git","cleanup","cd","Bash","isolation","ls","mkdir","true",{"skillNo":88,"name":89,"nameZh":6,"description":90,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":9,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":91,"categories":92,"keywords":93,"publishedAt":38},"01M6DNKEGPR8F3BA95ZMWQ48V9","unslop","Rewrite draft prose before owner-facing replies, docs, issues, PR text, handoffs, or Telegram reports to remove AI tells while preserving facts, links, tone, and authority; auto-trigger on externally visible prose, but never on code, commands, logs, tool output, quoted sources, or user text.","https:\u002F\u002Fgithub.com\u002F0xgleb\u002Fdotconfig\u002Fblob\u002Fmaster\u002Fai\u002Fskills\u002Funslop\u002FSKILL.md",[],[89,94,73,95,96,97,98,99,100,101,102,103,104,105,106,107,108,109,110,111],"draft","pr","remove","trigger","tool","issues","auto","rewrite","code","tone","telegram","sources","links","authority","logs","facts","output","replies",{"skillNo":113,"name":114,"nameZh":6,"description":115,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":9,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":116,"categories":117,"keywords":118,"publishedAt":38},"01M6DNK9VFNJYKB6GMA15WJ3HZ","threat-model-first","Use at the START of a change that moves money, changes balances or custody, touches access control or signing, deserializes untrusted input, adds an external or on-chain integration, or wires an LLM tool. Design-time threat modeling that turns abuse cases into the first failing tests; it precedes the post-hoc security-review pass.","https:\u002F\u002Fgithub.com\u002F0xgleb\u002Fdotconfig\u002Fblob\u002Fmaster\u002Fai\u002Fskills\u002Fthreat-model-first\u002FSKILL.md",[],[119,120,121,122,123,124,125,126,127,128,129,82,130,131,132,133,22,134,135,136],"model","threat model","threat","first","threat model first","chain","time","security-review","design","forge","access","cargo audit","review","change","audit","write","edit","modeling",{"skillNo":138,"name":139,"nameZh":6,"description":140,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":9,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":141,"categories":142,"keywords":148,"publishedAt":38},"01M6DNK6M72QAMH2TYVMPYJGDM","test-inspector","Auto-run for new or changed retained behavior, even without test changes, and for test, fixture, snapshot or mock changes; verify real happy-path and error-path coverage, expose false-confidence assertions and over-mocking, but do not perform general production-code correctness review.","https:\u002F\u002Fgithub.com\u002F0xgleb\u002Fdotconfig\u002Fblob\u002Fmaster\u002Fai\u002Fskills\u002Ftest-inspector\u002FSKILL.md",[143],{"categoryNo":144,"categoryName":145,"taxonomyType":146,"taxonomyTypeDesc":147},"01JBSEEDCAT000000000000008","测试与质量","SCENARIO","使用场景",[149,150,151,152,100,153,82,102,131,154,22,155,156,157,158,159,160,79,161,162],"test","test inspector","fixture","do","false","run","coverage","error","number","URL","mktemp","snapshot","but","path",{"skillNo":164,"name":165,"nameZh":6,"description":166,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":9,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":167,"categories":168,"keywords":169,"publishedAt":38},"01M6DNK3BPN73D003FFADEWJP5","strong-typing-inspector","Lightweight review — flags primitives used where a domain type already exists (e.g. String for a USD field when Usd exists) and missed newtype opportunities.","https:\u002F\u002Fgithub.com\u002F0xgleb\u002Fdotconfig\u002Fblob\u002Fmaster\u002Fai\u002Fskills\u002Fstrong-typing-inspector\u002FSKILL.md",[],[170,171,149,172,95,173,22,174,175,79,176,177,82,178,131,179],"strong typing inspector","type","missed","domain","flags","grep","usd","gh","primitives","newtype",{"skillNo":181,"name":182,"nameZh":6,"description":183,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":9,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":184,"categories":185,"keywords":189,"publishedAt":38},"01M6DNJZHG2RK3MC4VPB9KWPYF","spec-review","Critique existing specifications and design proposals, including whether their approach or addition is justified, and refine them through adversarial Pi review and authorized author revisions. Not for code review, implementation-plan sequencing, creating a design from scratch or minor copy edits.","https:\u002F\u002Fgithub.com\u002F0xgleb\u002Fdotconfig\u002Fblob\u002Fmaster\u002Fai\u002Fskills\u002Fspec-review\u002FSKILL.md",[186],{"categoryNo":187,"categoryName":188,"taxonomyType":146,"taxonomyTypeDesc":147},"01JBSEEDCAT000000000000007","代码审查",[190,191,131,192,193,194,20,195,127,196,102,27,197,198,199,200,201,202,203,204],"spec review","spec","plan","critique","code review","refine","implementation","sequencing","specifications","proposals","scratch","Existing","author","From Scratch","adversarial"]