[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M77XNACF0FPJ4EC11MZ97BER":3,"skill-same:01M77XNACF0FPJ4EC11MZ97BER":61},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":16,"categories":17,"keywords":18,"publishedAt":19,"linkOnly":20,"markdownContent":21,"markdownBody":22,"frontMatter":23,"refName":24,"commitSha":25,"skillDirectory":26,"forksCount":27,"repositoryHtmlUrl":28,"repositoryHomepageUrl":6,"ownerHtmlUrl":29,"ownerBlogUrl":6,"ownerDescription":6,"updatedAt":30,"install":31,"resources":44},"01M77XNACF0FPJ4EC11MZ97BER","vs-octocode",null,"Use when the user asks how GitHub projects solve a problem, wants prior art, external code examples, ecosystem patterns, or a landscape comparison across parallel projects.","vltansky\u002Fvs","vltansky","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F5851280?v=4","01M77XJAE230PB81N52XJHT5A1","Vlad Tansky","01M77XJAET1CMBAZ2TWTT4ZQT8",0,"MIT","https:\u002F\u002Fgithub.com\u002Fvltansky\u002Fvs\u002Fblob\u002Fmaster\u002Fskills\u002Fvs-octocode\u002FSKILL.md",[],[],"2026-11-15 14:19:17",false,"---\nname: vs-octocode\ndescription: \"Use when the user asks how GitHub projects solve a problem, wants prior art, external code examples, ecosystem patterns, or a landscape comparison across parallel projects.\"\n---\n\n# Octocode Research\n\nAnswer the user's question by looking outward at real GitHub projects. Find how external projects solve the same kind of problem, then synthesize the patterns with citations.\n\nThis skill is for GitHub-backed prior art, examples, and landscape comparison. It is not a local call-graph tracer, rename blast-radius checker, or PR reviewer. If the user asks \"where is this defined in this repo?\", use normal local code search or a dedicated local research skill instead.\n\n## Output Goal\n\nChoose one output mode from the user's wording:\n\n- **Prior-art answer:** compact synthesis of patterns, examples, tradeoffs, and takeaways.\n- **Landscape report:** structured comparison across multiple parallel projects on fixed axes.\n\nDo not turn the result into an RFC unless the user asks for one. Hand off to `\u002Fvs-rfc-research` when they want a decision document.\n\nFor a prior-art answer, stay in compact Markdown and give the user:\n\n1. What patterns exist across external projects\n2. Which examples are most relevant\n3. What tradeoffs those examples reveal\n4. What is worth borrowing, avoiding, or researching next\n\nFor a landscape report, produce a matrix-centered map of projects in the same space. Do not recommend a winner or propose copying everything the \"best\" project does; porting specific ideas belongs in `\u002Fvs-steal`, and architectural pivots belong in `\u002Fvs-shape-it`.\n\n## Tool Source\n\nThe vs plugin includes `octocode` in its plugin `.mcp.json`. Use its typed MCP tools directly when available.\n\nIf they are not active in this session, fall back to the Octocode CLI\n(`npx -y octocode@latest tools \u003Cname> --queries '\u003Cjson>' --compact`) and note in\nthe report that the plugin MCP server was not loaded. Check the interface once\nwith `npx -y octocode@latest context --minimal`; before the first raw CLI call\nto a tool, read its exact schema with\n`npx -y octocode@latest tools \u003Cname> --scheme --json --compact`. See the\n[shared Octocode access ladder](..\u002Fvs-internal-shared\u002Freferences\u002Foctocode-access.md).\n\nDo not load or invoke Octocode's prompt, skill, or orchestration workflows (for example, `reviewPR`). This skill owns the research plan and execution loop; wrapping it in another workflow can add an approval gate or route the task into local code review.\n\nIf host or repository policy requires a research subagent, delegate bounded probes to that subagent and have it call the typed tools directly. The policy changes who executes the probes, not who owns the plan; do not nest another research orchestration workflow.\n\nIf neither MCP nor the CLI can reach Octocode, say so. Stop rather than pretending web snippets or local files are equivalent evidence.\n\n## Phase 1: Frame The Question\n\nTranslate the user's question into a broad search target.\n\nIdentify:\n\n- **Problem:** what behavior, architecture, API, workflow, or convention are we researching?\n- **Use:** what decision, implementation, or review claim will this evidence inform?\n- **Search domain:** likely languages, ecosystems, repo types, or organizations\n- **Comparison axes:** 3-6 dimensions that will make examples comparable\n- **Exclusions:** things that would be irrelevant or misleading\n- **Mode:** prior-art answer or landscape report\n\nWhen composed inside a local or PR review, take the concrete design and claims from the caller. Research external comparators for those claims, then return the evidence to the caller; do not expand into a second full PR review or issue a standalone merge verdict.\n\nIf the user gives a vague question, choose a reasonable starting frame and state it. Ask only when the missing detail changes the search domain or the output mode.\n\nGood broad-search questions:\n\n- \"How do other agent frameworks organize skills and commands?\"\n- \"Find examples of projects using ADRs with AI coding instructions.\"\n- \"How do TypeScript CLIs manage plugin manifests?\"\n- \"What patterns do repos use for eval fixtures next to skills?\"\n- \"Compare how AI-agent frameworks organize skills, prompts, and commands.\"\n\nNot this skill:\n\n- \"Where is `validateToken` defined in this repo?\"\n- \"What breaks if I rename this function?\"\n- \"Review this PR.\"\n\n## Phase 2: Search Plan\n\nBefore calling research tools, turn the question into a small falsifiable campaign. The plan exists to prevent broad searches from drifting, so it should name what evidence could change the answer rather than repeat generic workflow steps.\n\nKeep at least two hypotheses alive for each unsettled claim. Include the cheapest disconfirming check so the research can reject an attractive example instead of only collecting supporting snippets.\n\nWrite a short plan and proceed immediately:\n\n```markdown\n## Octocode Research Plan\n**Question:** \u003Cuser question>\n**Mode:** \u003Cprior-art answer | landscape report>\n**Corpus:** \u003Corganizations, ecosystems, repo families, languages>\n**Active \u002F skipped surfaces:** \u003Ccode, tests, docs, PRs\u002Fhistory, packages; why skipped>\n**Axes:** \u003C3-6 fixed comparison axes>\n**Budget:** \u003Cdiscovery queries, finalist count, deep-read or time limit>\n**Stop test:** \u003Cevidence that answers the question and rejects the main alternative>\n\n| Claim \u002F hypotheses | Evidence needed | Cheapest first probe | Disconfirming signal |\n|---|---|---|---|\n| \u003Cclaim; H1 vs H2> | \u003Csource\u002Ftest\u002Fhistory anchor> | \u003Cquery + tool route> | \u003Cwhat would weaken or reject H1> |\n```\n\nKeep every receipt line. Shorten values for a small question rather than silently dropping surfaces, budget, or the stop test.\n\nDefault to 2-4 claims, 2-3 discovery queries, 3-5 finalists, and 3-5 decisive iterations. Landscape mode may expand to 5-8 finalists. Adjust the budget when the question is clearly smaller or larger.\n\nNever ask for approval of the research plan. Present it as a concise progress update and start the first probes in the same turn. Approval is only relevant for a separate side effect such as cloning many repositories, running untrusted code, or materially expanding the requested scope.\n\nDuring execution, keep a tiny private ledger: `claim → evidence → confidence → next check`. Update it after every decisive result. Stop when the answer is grounded and the main alternative is rejected, no cheap check can change the conclusion, the budget is reached, or two iterations change no claim or confidence.\n\n## Phase 3: Discover Candidate Projects\n\nUse external search first. Derive literal, alias, adjacent-implementation, and package terms from the plan. Prefer several smaller searches over one giant query, and batch 1-3 independent probes when the tool supports it.\n\nResolve named repositories, PRs, packages, and refs from supplied URLs, thread context, caller evidence, or current repository metadata before keyword discovery. PR numbers are repository-local; do not search an organization for a bare number when the owning repository can be resolved first.\n\nTypical octocode tool choices:\n\n| Need | Tool |\n|---|---|\n| Find candidate repos | `ghSearchRepos` |\n| Find implementation examples | `ghSearchCode` |\n| Inspect project layout | `ghViewRepoStructure` |\n| Read source\u002Fdocs | `ghGetFileContent` |\n| Understand adoption\u002Fhistory | `ghSearchPullRequests` |\n| Find package ecosystem signals | `npmSearch` |\n\nFor CLI calls, pass only fields from the tool's current `--scheme`; do not add\nlegacy planning fields that the schema does not expose. Keep the call's purpose\nin the research plan and claim ledger.\n\nFollow hints in tool responses. If a tool suggests narrowing, pagination, related files, or a better next call, use that guidance unless you have a clear reason not to.\n\nTreat `empty` as evidence only for the searched scope and `error` as a failed call, not absence. Change one filter, synonym, or search surface; after two unproductive refinements, pivot or mark the gap instead of looping.\n\nRank candidates cheaply before deep reads:\n\n- **Fit:** directly implements the capability, adjacent, or keyword-only\n- **Evidence:** exact code\u002Ftests, committed docs\u002Fexamples, or search snippet only\n- **Activity:** current releases\u002Fcommits\u002Fissues or visibly stale\u002Farchived\n- **Reuse:** clear API\u002Flicense\u002Fdependencies or substantial integration drag\n- **Risk:** known caveats, important unknowns, or blockers\n\nStars and downloads are tiebreakers, not proof of quality.\n\n## Phase 4: Inspect Representative Examples\n\nSelect a small, useful sample rather than collecting endless links.\n\nDefault sample:\n\n- 3-5 candidate projects for normal questions\n- 5-8 candidate projects for ecosystem landscape questions\n- 2-3 deep reads when examples are complex\n\nFor each representative example, capture:\n\n- Repo\u002Fproject name\n- What problem it solves\n- The concrete file, config, API, or workflow pattern\n- Why it is relevant to the user's question\n- Tradeoff or caveat\n- Full GitHub URL with line numbers when available\n\nFavor implementation files and committed docs over README claims. README-only evidence is allowed, but label it as documentation evidence.\n\nStart from the strongest available handle: use a known path for an exact read,\nan identifier for semantic references, a code shape for structural search, and\na why-question for PR or commit history. Do not force every question through\nthe same tool sequence.\n\nFor every nontrivial claim, inspect at least two evidence dimensions:\n\n- **Structure:** repository layout and where the mechanism lives\n- **Stream:** exact source, test, config, or committed documentation\n- **Connections\u002Fhistory:** callers, integrations, PRs, issues, releases, or commit intent\n\nGrade proof by what it establishes: **semantic** evidence proves identity or\nreachability, **structural** evidence proves code shape, **lexical** evidence\nproves text coverage, and **provider** evidence is limited by the remote index.\nTreat provider snippets as leads, and do not infer unused, safe, or absent from\nlexical or provider evidence alone.\n\nUse targeted `matchString` or line-range reads after discovery. If the same remote area needs three or more deep reads, or the conclusion needs semantic navigation or strong absence proof, materialize that bounded area when authorized instead of repeatedly reading it through the provider.\n\n## Phase 5A: Synthesize Prior-Art Answers\n\nUse this phase when the mode is **prior-art answer**. Group findings by pattern, not by search order.\n\nUse this output shape:\n\n```markdown\n## TL;DR\n\u003C2-4 sentences answering the question directly.>\n\n## Patterns\n- **\u003CPattern name>:** \u003Cwhat multiple projects do and why it matters>\n  - Evidence: \u003Crepo\u002Ffile citation>\n  - Tradeoff: \u003Ccost or limitation>\n\n## Examples\n| Project | Pattern | Why it matters | Evidence |\n|---|---|---|---|\n| \u003Crepo> | \u003Cshort pattern> | \u003Crelevance> | \u003CURL> |\n\n## Takeaways\n1. \u003Cactionable synthesis>\n2. \u003Cactionable synthesis>\n3. \u003Cactionable synthesis>\n\n## Gaps\n\u003CWhat was not proven, weakly evidenced, or worth a follow-up search.>\n```\n\nFor small answers, keep it shorter, but always include citations.\n\n## Phase 5B: Write Landscape Reports\n\nUse this phase when the mode is **landscape report**.\n\nBefore characterizing any project, lock 4-7 structural axes. Default axes for an agent\u002Fskills framework:\n\n- **Runtime \u002F host** — Claude Code? Codex? custom CLI? language runtime?\n- **Invocation model** — slash commands? model auto-invocation? both? how is ambiguity resolved?\n- **Taxonomy** — how are skills\u002Fprompts\u002Fagents organized? explicit levels? flat?\n- **Artifact output** — in-repo? per-user? session-only? how is it cleaned up?\n- **Evaluation** — LLM-as-judge? scripted? regression over fixtures? none?\n- **Distribution** — package? symlink? pointer files? copy-paste from docs?\n- **Composition** — can skills call other skills? how? explicit vs implicit?\n\nWrite the axes down before searching. Revising axes after you've characterized a project biases the comparison.\n\nAim for 5-10 projects, not 30. Prefer projects that are active, show real usage, and are structural rather than one-off prompt dumps.\n\nFor each project, fill in each axis in one sentence. Cite a file\u002Fline when a sentence describes a concrete mechanism. Uncited axis values get flagged as `[inference]`.\n\nThe multi-project matrix and cross-linked evidence justify a visual artifact.\nResolve `$PROJECT_ID` (see\n[..\u002Fvs-internal-shared\u002FSKILL.md](..\u002Fvs-internal-shared\u002FSKILL.md)) and write one\nHTMDX artifact to\n`~\u002F.vs\u002F$PROJECT_ID\u002Fvs-octocode\u002FYYYY-MM-DD-landscape.html`. Follow the\n[shared rich-artifact contract](..\u002Fvs-internal-shared\u002Freferences\u002Frich-artifacts.md):\npin its exact runtime, keep all editable content in the HTMDX source block, and\ncreate no Markdown twin.\n\n```mdx\n---\ntitle: GitHub landscape\nupdated: YYYY-MM-DD\n---\n\n# GitHub landscape — YYYY-MM-DD\n\n\u003CExecutiveSummary>\nTODO: Write two to four sentences describing the main clusters and where this\nproject sits.\n\u003C\u002FExecutiveSummary>\n\n## Axes\n- **Runtime \u002F host** — ...\n- **Invocation model** — ...\n- **Taxonomy** — ...\n\n## Projects\n\n### owner\u002Frepo — one-line positioning\n- Runtime \u002F host: ...\n- Invocation model: ...\n- Taxonomy: ...\n- Artifact output: ...\n- Evaluation: ...\n- Distribution: ...\n- Composition: ...\n- Notable choices: ...\n- Stars \u002F activity: ... (as of YYYY-MM-DD)\n\n## Matrix\n\n\u003CDataTable>\n| Project | Runtime | Invocation | Taxonomy | Artifacts | Evaluation | Distribution | Composition |\n|---------|---------|------------|----------|-----------|------------|--------------|-------------|\n| owner\u002Frepo | ... | ... | ... | ... | ... | ... | ... |\n| **this project: vs** | ... | ... | ... | ... | ... | ... | ... |\n\u003C\u002FDataTable>\n\n## Clusters and outliers\n- Cluster 1: projects that chose X for axis Y (list)\n- Outlier: ... (why they're different)\n- Where this project sits: ... (one or two sentences)\n\n## Where this project is differentiated\n- ... (axes where this project makes a non-obvious choice — evidence-backed)\n\n## Where this project is lagging\n- ... (axes where a clear better choice exists in the ecosystem — honest)\n\n## Open questions\n- [ ] ...\n```\n\nPrint the report path plus a short clusters summary.\n\n## Evidence Rules\n\n- Use full GitHub URLs with line numbers for external code.\n- Separate direct evidence from inference.\n- Do not claim popularity or best practice from one repo.\n- Do not recommend copying code verbatim.\n- If evidence is thin, say so.\n- If examples disagree, explain the split instead of forcing consensus.\n\n## Relationship To Nearby Skills\n\n- `\u002Fvs-octocode` answers broad GitHub-backed questions and can create a landscape map across multiple projects.\n- `\u002Fvs-steal` deeply inspects one named repo for portable ideas.\n- `\u002Fvs-rfc-research` turns evidence into a formal proposal or decision document.\n\n## Output style\n\nApply the [shared output style](..\u002Fvs-internal-shared\u002Freferences\u002Foutput-style.md)\nto every user-facing message.\n\n## Workflow\n\nDirect: emit **Next** only. Composed: return to caller.\n\n**Prev:** research question | `\u002Fvs-shape-it` | `\u002Fvs-pushback`\n**Next:** `\u002Fvs-rfc-research`\n**Relevant:** `\u002Fvs-steal` | `\u002Fvs-write`\n","# Octocode Research\n\nAnswer the user's question by looking outward at real GitHub projects. Find how external projects solve the same kind of problem, then synthesize the patterns with citations.\n\nThis skill is for GitHub-backed prior art, examples, and landscape comparison. It is not a local call-graph tracer, rename blast-radius checker, or PR reviewer. If the user asks \"where is this defined in this repo?\", use normal local code search or a dedicated local research skill instead.\n\n## Output Goal\n\nChoose one output mode from the user's wording:\n\n- **Prior-art answer:** compact synthesis of patterns, examples, tradeoffs, and takeaways.\n- **Landscape report:** structured comparison across multiple parallel projects on fixed axes.\n\nDo not turn the result into an RFC unless the user asks for one. Hand off to `\u002Fvs-rfc-research` when they want a decision document.\n\nFor a prior-art answer, stay in compact Markdown and give the user:\n\n1. What patterns exist across external projects\n2. Which examples are most relevant\n3. What tradeoffs those examples reveal\n4. What is worth borrowing, avoiding, or researching next\n\nFor a landscape report, produce a matrix-centered map of projects in the same space. Do not recommend a winner or propose copying everything the \"best\" project does; porting specific ideas belongs in `\u002Fvs-steal`, and architectural pivots belong in `\u002Fvs-shape-it`.\n\n## Tool Source\n\nThe vs plugin includes `octocode` in its plugin `.mcp.json`. Use its typed MCP tools directly when available.\n\nIf they are not active in this session, fall back to the Octocode CLI\n(`npx -y octocode@latest tools \u003Cname> --queries '\u003Cjson>' --compact`) and note in\nthe report that the plugin MCP server was not loaded. Check the interface once\nwith `npx -y octocode@latest context --minimal`; before the first raw CLI call\nto a tool, read its exact schema with\n`npx -y octocode@latest tools \u003Cname> --scheme --json --compact`. See the\n[shared Octocode access ladder](..\u002Fvs-internal-shared\u002Freferences\u002Foctocode-access.md).\n\nDo not load or invoke Octocode's prompt, skill, or orchestration workflows (for example, `reviewPR`). This skill owns the research plan and execution loop; wrapping it in another workflow can add an approval gate or route the task into local code review.\n\nIf host or repository policy requires a research subagent, delegate bounded probes to that subagent and have it call the typed tools directly. The policy changes who executes the probes, not who owns the plan; do not nest another research orchestration workflow.\n\nIf neither MCP nor the CLI can reach Octocode, say so. Stop rather than pretending web snippets or local files are equivalent evidence.\n\n## Phase 1: Frame The Question\n\nTranslate the user's question into a broad search target.\n\nIdentify:\n\n- **Problem:** what behavior, architecture, API, workflow, or convention are we researching?\n- **Use:** what decision, implementation, or review claim will this evidence inform?\n- **Search domain:** likely languages, ecosystems, repo types, or organizations\n- **Comparison axes:** 3-6 dimensions that will make examples comparable\n- **Exclusions:** things that would be irrelevant or misleading\n- **Mode:** prior-art answer or landscape report\n\nWhen composed inside a local or PR review, take the concrete design and claims from the caller. Research external comparators for those claims, then return the evidence to the caller; do not expand into a second full PR review or issue a standalone merge verdict.\n\nIf the user gives a vague question, choose a reasonable starting frame and state it. Ask only when the missing detail changes the search domain or the output mode.\n\nGood broad-search questions:\n\n- \"How do other agent frameworks organize skills and commands?\"\n- \"Find examples of projects using ADRs with AI coding instructions.\"\n- \"How do TypeScript CLIs manage plugin manifests?\"\n- \"What patterns do repos use for eval fixtures next to skills?\"\n- \"Compare how AI-agent frameworks organize skills, prompts, and commands.\"\n\nNot this skill:\n\n- \"Where is `validateToken` defined in this repo?\"\n- \"What breaks if I rename this function?\"\n- \"Review this PR.\"\n\n## Phase 2: Search Plan\n\nBefore calling research tools, turn the question into a small falsifiable campaign. The plan exists to prevent broad searches from drifting, so it should name what evidence could change the answer rather than repeat generic workflow steps.\n\nKeep at least two hypotheses alive for each unsettled claim. Include the cheapest disconfirming check so the research can reject an attractive example instead of only collecting supporting snippets.\n\nWrite a short plan and proceed immediately:\n\n```markdown\n## Octocode Research Plan\n**Question:** \u003Cuser question>\n**Mode:** \u003Cprior-art answer | landscape report>\n**Corpus:** \u003Corganizations, ecosystems, repo families, languages>\n**Active \u002F skipped surfaces:** \u003Ccode, tests, docs, PRs\u002Fhistory, packages; why skipped>\n**Axes:** \u003C3-6 fixed comparison axes>\n**Budget:** \u003Cdiscovery queries, finalist count, deep-read or time limit>\n**Stop test:** \u003Cevidence that answers the question and rejects the main alternative>\n\n| Claim \u002F hypotheses | Evidence needed | Cheapest first probe | Disconfirming signal |\n|---|---|---|---|\n| \u003Cclaim; H1 vs H2> | \u003Csource\u002Ftest\u002Fhistory anchor> | \u003Cquery + tool route> | \u003Cwhat would weaken or reject H1> |\n```\n\nKeep every receipt line. Shorten values for a small question rather than silently dropping surfaces, budget, or the stop test.\n\nDefault to 2-4 claims, 2-3 discovery queries, 3-5 finalists, and 3-5 decisive iterations. Landscape mode may expand to 5-8 finalists. Adjust the budget when the question is clearly smaller or larger.\n\nNever ask for approval of the research plan. Present it as a concise progress update and start the first probes in the same turn. Approval is only relevant for a separate side effect such as cloning many repositories, running untrusted code, or materially expanding the requested scope.\n\nDuring execution, keep a tiny private ledger: `claim → evidence → confidence → next check`. Update it after every decisive result. Stop when the answer is grounded and the main alternative is rejected, no cheap check can change the conclusion, the budget is reached, or two iterations change no claim or confidence.\n\n## Phase 3: Discover Candidate Projects\n\nUse external search first. Derive literal, alias, adjacent-implementation, and package terms from the plan. Prefer several smaller searches over one giant query, and batch 1-3 independent probes when the tool supports it.\n\nResolve named repositories, PRs, packages, and refs from supplied URLs, thread context, caller evidence, or current repository metadata before keyword discovery. PR numbers are repository-local; do not search an organization for a bare number when the owning repository can be resolved first.\n\nTypical octocode tool choices:\n\n| Need | Tool |\n|---|---|\n| Find candidate repos | `ghSearchRepos` |\n| Find implementation examples | `ghSearchCode` |\n| Inspect project layout | `ghViewRepoStructure` |\n| Read source\u002Fdocs | `ghGetFileContent` |\n| Understand adoption\u002Fhistory | `ghSearchPullRequests` |\n| Find package ecosystem signals | `npmSearch` |\n\nFor CLI calls, pass only fields from the tool's current `--scheme`; do not add\nlegacy planning fields that the schema does not expose. Keep the call's purpose\nin the research plan and claim ledger.\n\nFollow hints in tool responses. If a tool suggests narrowing, pagination, related files, or a better next call, use that guidance unless you have a clear reason not to.\n\nTreat `empty` as evidence only for the searched scope and `error` as a failed call, not absence. Change one filter, synonym, or search surface; after two unproductive refinements, pivot or mark the gap instead of looping.\n\nRank candidates cheaply before deep reads:\n\n- **Fit:** directly implements the capability, adjacent, or keyword-only\n- **Evidence:** exact code\u002Ftests, committed docs\u002Fexamples, or search snippet only\n- **Activity:** current releases\u002Fcommits\u002Fissues or visibly stale\u002Farchived\n- **Reuse:** clear API\u002Flicense\u002Fdependencies or substantial integration drag\n- **Risk:** known caveats, important unknowns, or blockers\n\nStars and downloads are tiebreakers, not proof of quality.\n\n## Phase 4: Inspect Representative Examples\n\nSelect a small, useful sample rather than collecting endless links.\n\nDefault sample:\n\n- 3-5 candidate projects for normal questions\n- 5-8 candidate projects for ecosystem landscape questions\n- 2-3 deep reads when examples are complex\n\nFor each representative example, capture:\n\n- Repo\u002Fproject name\n- What problem it solves\n- The concrete file, config, API, or workflow pattern\n- Why it is relevant to the user's question\n- Tradeoff or caveat\n- Full GitHub URL with line numbers when available\n\nFavor implementation files and committed docs over README claims. README-only evidence is allowed, but label it as documentation evidence.\n\nStart from the strongest available handle: use a known path for an exact read,\nan identifier for semantic references, a code shape for structural search, and\na why-question for PR or commit history. Do not force every question through\nthe same tool sequence.\n\nFor every nontrivial claim, inspect at least two evidence dimensions:\n\n- **Structure:** repository layout and where the mechanism lives\n- **Stream:** exact source, test, config, or committed documentation\n- **Connections\u002Fhistory:** callers, integrations, PRs, issues, releases, or commit intent\n\nGrade proof by what it establishes: **semantic** evidence proves identity or\nreachability, **structural** evidence proves code shape, **lexical** evidence\nproves text coverage, and **provider** evidence is limited by the remote index.\nTreat provider snippets as leads, and do not infer unused, safe, or absent from\nlexical or provider evidence alone.\n\nUse targeted `matchString` or line-range reads after discovery. If the same remote area needs three or more deep reads, or the conclusion needs semantic navigation or strong absence proof, materialize that bounded area when authorized instead of repeatedly reading it through the provider.\n\n## Phase 5A: Synthesize Prior-Art Answers\n\nUse this phase when the mode is **prior-art answer**. Group findings by pattern, not by search order.\n\nUse this output shape:\n\n```markdown\n## TL;DR\n\u003C2-4 sentences answering the question directly.>\n\n## Patterns\n- **\u003CPattern name>:** \u003Cwhat multiple projects do and why it matters>\n  - Evidence: \u003Crepo\u002Ffile citation>\n  - Tradeoff: \u003Ccost or limitation>\n\n## Examples\n| Project | Pattern | Why it matters | Evidence |\n|---|---|---|---|\n| \u003Crepo> | \u003Cshort pattern> | \u003Crelevance> | \u003CURL> |\n\n## Takeaways\n1. \u003Cactionable synthesis>\n2. \u003Cactionable synthesis>\n3. \u003Cactionable synthesis>\n\n## Gaps\n\u003CWhat was not proven, weakly evidenced, or worth a follow-up search.>\n```\n\nFor small answers, keep it shorter, but always include citations.\n\n## Phase 5B: Write Landscape Reports\n\nUse this phase when the mode is **landscape report**.\n\nBefore characterizing any project, lock 4-7 structural axes. Default axes for an agent\u002Fskills framework:\n\n- **Runtime \u002F host** — Claude Code? Codex? custom CLI? language runtime?\n- **Invocation model** — slash commands? model auto-invocation? both? how is ambiguity resolved?\n- **Taxonomy** — how are skills\u002Fprompts\u002Fagents organized? explicit levels? flat?\n- **Artifact output** — in-repo? per-user? session-only? how is it cleaned up?\n- **Evaluation** — LLM-as-judge? scripted? regression over fixtures? none?\n- **Distribution** — package? symlink? pointer files? copy-paste from docs?\n- **Composition** — can skills call other skills? how? explicit vs implicit?\n\nWrite the axes down before searching. Revising axes after you've characterized a project biases the comparison.\n\nAim for 5-10 projects, not 30. Prefer projects that are active, show real usage, and are structural rather than one-off prompt dumps.\n\nFor each project, fill in each axis in one sentence. Cite a file\u002Fline when a sentence describes a concrete mechanism. Uncited axis values get flagged as `[inference]`.\n\nThe multi-project matrix and cross-linked evidence justify a visual artifact.\nResolve `$PROJECT_ID` (see\n[..\u002Fvs-internal-shared\u002FSKILL.md](..\u002Fvs-internal-shared\u002FSKILL.md)) and write one\nHTMDX artifact to\n`~\u002F.vs\u002F$PROJECT_ID\u002Fvs-octocode\u002FYYYY-MM-DD-landscape.html`. Follow the\n[shared rich-artifact contract](..\u002Fvs-internal-shared\u002Freferences\u002Frich-artifacts.md):\npin its exact runtime, keep all editable content in the HTMDX source block, and\ncreate no Markdown twin.\n\n```mdx\n---\ntitle: GitHub landscape\nupdated: YYYY-MM-DD\n---\n\n# GitHub landscape — YYYY-MM-DD\n\n\u003CExecutiveSummary>\nTODO: Write two to four sentences describing the main clusters and where this\nproject sits.\n\u003C\u002FExecutiveSummary>\n\n## Axes\n- **Runtime \u002F host** — ...\n- **Invocation model** — ...\n- **Taxonomy** — ...\n\n## Projects\n\n### owner\u002Frepo — one-line positioning\n- Runtime \u002F host: ...\n- Invocation model: ...\n- Taxonomy: ...\n- Artifact output: ...\n- Evaluation: ...\n- Distribution: ...\n- Composition: ...\n- Notable choices: ...\n- Stars \u002F activity: ... (as of YYYY-MM-DD)\n\n## Matrix\n\n\u003CDataTable>\n| Project | Runtime | Invocation | Taxonomy | Artifacts | Evaluation | Distribution | Composition |\n|---------|---------|------------|----------|-----------|------------|--------------|-------------|\n| owner\u002Frepo | ... | ... | ... | ... | ... | ... | ... |\n| **this project: vs** | ... | ... | ... | ... | ... | ... | ... |\n\u003C\u002FDataTable>\n\n## Clusters and outliers\n- Cluster 1: projects that chose X for axis Y (list)\n- Outlier: ... (why they're different)\n- Where this project sits: ... (one or two sentences)\n\n## Where this project is differentiated\n- ... (axes where this project makes a non-obvious choice — evidence-backed)\n\n## Where this project is lagging\n- ... (axes where a clear better choice exists in the ecosystem — honest)\n\n## Open questions\n- [ ] ...\n```\n\nPrint the report path plus a short clusters summary.\n\n## Evidence Rules\n\n- Use full GitHub URLs with line numbers for external code.\n- Separate direct evidence from inference.\n- Do not claim popularity or best practice from one repo.\n- Do not recommend copying code verbatim.\n- If evidence is thin, say so.\n- If examples disagree, explain the split instead of forcing consensus.\n\n## Relationship To Nearby Skills\n\n- `\u002Fvs-octocode` answers broad GitHub-backed questions and can create a landscape map across multiple projects.\n- `\u002Fvs-steal` deeply inspects one named repo for portable ideas.\n- `\u002Fvs-rfc-research` turns evidence into a formal proposal or decision document.\n\n## Output style\n\nApply the [shared output style](..\u002Fvs-internal-shared\u002Freferences\u002Foutput-style.md)\nto every user-facing message.\n\n## Workflow\n\nDirect: emit **Next** only. Composed: return to caller.\n\n**Prev:** research question | `\u002Fvs-shape-it` | `\u002Fvs-pushback`\n**Next:** `\u002Fvs-rfc-research`\n**Relevant:** `\u002Fvs-steal` | `\u002Fvs-write`\n",{"name":5,"description":7},"master","72448991571c9e5f0e76cccf189b8522d442e6eb","skills\u002Fvs-octocode",1,"https:\u002F\u002Fgithub.com\u002Fvltansky\u002Fvs","https:\u002F\u002Fgithub.com\u002Fvltansky","2026-11-06 18:47:54",{"source":8,"skillName":5,"commands":32,"prompt":42,"zipAvailable":43},[33,36,39],{"tool":34,"command":35},"npx","npx skills add vltansky\u002Fvs --skill vs-octocode",{"tool":37,"command":38},"bunx","bunx skills add vltansky\u002Fvs --skill vs-octocode",{"tool":40,"command":41},"pnpm","pnpm dlx skills add vltansky\u002Fvs --skill vs-octocode","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Fvltansky\u002Fvs\u002Fblob\u002Fmaster\u002Fskills\u002Fvs-octocode\u002FSKILL.md",true,[45,51,57],{"path":46,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":49,"sourceHtmlUrl":50},"agents\u002Fopenai.yaml","REFERENCE","参考文档",290,"https:\u002F\u002Fgithub.com\u002Fvltansky\u002Fvs\u002Fblob\u002Fmaster\u002Fskills\u002Fvs-octocode\u002Fagents\u002Fopenai.yaml",{"path":52,"resourceType":53,"resourceTypeDesc":54,"sizeBytes":55,"sourceHtmlUrl":56},"test\u002Fcompetitors.eval.ts","SCRIPT","脚本",1416,"https:\u002F\u002Fgithub.com\u002Fvltansky\u002Fvs\u002Fblob\u002Fmaster\u002Fskills\u002Fvs-octocode\u002Ftest\u002Fcompetitors.eval.ts",{"path":58,"resourceType":53,"resourceTypeDesc":54,"sizeBytes":59,"sourceHtmlUrl":60},"test\u002Fresearch.eval.ts",3968,"https:\u002F\u002Fgithub.com\u002Fvltansky\u002Fvs\u002Fblob\u002Fmaster\u002Fskills\u002Fvs-octocode\u002Ftest\u002Fresearch.eval.ts",[62,69,75,80,87,92],{"skillNo":63,"name":64,"nameZh":6,"description":65,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":66,"categories":67,"keywords":68,"publishedAt":19},"01M77XSKWFMPKFXM0311QKEJHH","vs-write","Use when asked to write, rewrite, edit, tighten, simplify, or polish prose for clarity, or to put text in STE, ASD-STE100, or controlled language. Produces plain, direct, accessible copy without losing substance or precision.","https:\u002F\u002Fgithub.com\u002Fvltansky\u002Fvs\u002Fblob\u002Fmaster\u002Fskills\u002Fvs-write\u002FSKILL.md",[],[],{"skillNo":70,"name":71,"nameZh":6,"description":6,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":72,"categories":73,"keywords":74,"publishedAt":19},"01M77XSJNX2V560HSKQKQD67KP","vs-fake-verify","https:\u002F\u002Fgithub.com\u002Fvltansky\u002Fvs\u002Fblob\u002Fmaster\u002Fskills\u002Fvs-verify\u002Ftest\u002Ffixtures\u002Fpath-end-state\u002Fstub-rejector\u002FSKILL.md",[],[],{"skillNo":76,"name":71,"nameZh":6,"description":6,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":77,"categories":78,"keywords":79,"publishedAt":19},"01M77XSHHBEMMYZ6B4PD384NZH","https:\u002F\u002Fgithub.com\u002Fvltansky\u002Fvs\u002Fblob\u002Fmaster\u002Fskills\u002Fvs-verify\u002Ftest\u002Ffixtures\u002Fpath-end-state\u002Fstructure-paste\u002FSKILL.md",[],[],{"skillNo":81,"name":82,"nameZh":6,"description":83,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":84,"categories":85,"keywords":86,"publishedAt":19},"01M77XSE2Y7MW0KP3DTFNEE2E8","vs-verify","Use when asked to verify, prove it works, check completion, or before claiming done. Produces evidence-backed PASS\u002FWARN\u002FFAIL status.","https:\u002F\u002Fgithub.com\u002Fvltansky\u002Fvs\u002Fblob\u002Fmaster\u002Fskills\u002Fvs-verify\u002Ftest\u002Ffixtures\u002Fpath-end-state\u002Fpublished-pair\u002FSKILL.md",[],[],{"skillNo":88,"name":71,"nameZh":6,"description":6,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":89,"categories":90,"keywords":91,"publishedAt":19},"01M77XSAXFQXNE6VCRC7275Z0X","https:\u002F\u002Fgithub.com\u002Fvltansky\u002Fvs\u002Fblob\u002Fmaster\u002Fskills\u002Fvs-verify\u002Ftest\u002Ffixtures\u002Fpath-end-state\u002Fphrase-complete-beside\u002FSKILL.md",[],[],{"skillNo":93,"name":82,"nameZh":6,"description":83,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":94,"categories":95,"keywords":96,"publishedAt":19},"01M77XS6R0E9X1E5KKSS00WYSS","https:\u002F\u002Fgithub.com\u002Fvltansky\u002Fvs\u002Fblob\u002Fmaster\u002Fskills\u002Fvs-verify\u002FSKILL.md",[],[]]