[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M6GXQCS5V00HCYA63XT2PS1V":3,"skill-same:01M6GXQCS5V00HCYA63XT2PS1V":100},{"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":26,"publishedAt":46,"linkOnly":47,"markdownContent":48,"markdownBody":49,"frontMatter":50,"refName":60,"commitSha":61,"skillDirectory":62,"forksCount":63,"repositoryHtmlUrl":64,"repositoryHomepageUrl":65,"ownerHtmlUrl":66,"ownerBlogUrl":65,"ownerDescription":67,"updatedAt":68,"install":69,"resources":81},"01M6GXQCS5V00HCYA63XT2PS1V","research",null,"Focused research investigations. Converts questions into structured findings with confidence levels and source citations. Single agent by default; with --parallel (or when the question decomposes into 3+ independent angles) it spawns scout agents whose findings are compressed into a unified brief. Does not make decisions; produces information that informs the next step.","SethGammon\u002FCitadel","SethGammon","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F5523585?v=4","01M6GXJ90FD2CE9QF5GQERWN06","Seth Gammon","01M6GXJ94VXYGK5KAW14EYE901",929,"MIT","https:\u002F\u002Fgithub.com\u002FSethGammon\u002FCitadel\u002Fblob\u002Fmain\u002Fskills\u002Fresearch\u002FSKILL.md",[18,23],{"categoryNo":19,"categoryName":20,"taxonomyType":21,"taxonomyTypeDesc":22},"01JBSEEDCAT000000000000021","资料检索与研究","SCENARIO","使用场景",{"categoryNo":24,"categoryName":25,"taxonomyType":21,"taxonomyTypeDesc":22},"01JBSEEDCAT000000000000023","Skill与智能体开发",[5,27,28,29,30,31,32,33,34,35,36,37,38,39,40,41,42,43,44,45],"options","fleet","true","with","information","make","multi","compare","next step","false","next","scout","default","question","decisions","parallel","investigate","Citations","brief","2026-11-10 10:12:30",false,"---\nname: research\nlicense: MIT\ndescription: >-\n  Focused research investigations. Converts questions into structured findings\n  with confidence levels and source citations. Single agent by default; with\n  --parallel (or when the question decomposes into 3+ independent angles) it\n  spawns scout agents whose findings are compressed into a unified brief.\n  Does not make decisions; produces information that informs the next step.\nuser-invocable: true\nauto-trigger: false\ntrigger_keywords:\n  - research\n  - investigate\n  - look into\n  - find out\n  - research fleet\n  - parallel research\n  - multi-angle research\n  - compare options\nlast-updated: 2026-07-21\n---\n\n# \u002Fresearch — Focused Investigation\n\n## When to Use\n\n- Evaluating whether a dependency has a newer version or has been superseded\n- Finding community best practices for a specific technical problem\n- Reading official documentation for an API or library\n- Investigating how other projects solve a similar problem\n- Checking if a pattern used in the codebase has known issues\n- Any time you need external information before making a decision\n\n**Don't use when:** you need to act on findings immediately (use \u002Fmarshal which calls \u002Fresearch internally).\n\n## Modes\n\n**Single (default):** one agent, 2-4 queries, 3-6 sources. Steps 1-5 below.\n\n**Parallel (`\u002Fresearch --parallel`):** scout agents investigate independent\nangles of the same question using Fleet wave mechanics. See Parallel Mode\nbelow. Without the flag, prefer parallel mode (and say so in the plan) when\nthe question naturally decomposes into 3+ independent angles: evaluating\nmultiple competing technologies or approaches, distinct sub-questions that\ndon't depend on each other, or time-sensitive research where parallel\nexecution matters.\n\nIf the question is narrow and focused, stay single-agent. Don't parallelize\nwhat a single agent can answer in 5 minutes.\n\n## Protocol (Single Mode)\n\n### Step 1: FORMULATE\n\nConvert the research question into 2-4 specific search queries:\n- Official docs query (e.g., \"express.js middleware error handling docs\")\n- Community\u002FGitHub query (e.g., \"express error middleware best practices site:github.com\")\n- Technical blog\u002Fcomparison query (e.g., \"express vs fastify error handling 2025\")\n- Release notes query if version-specific (e.g., \"express 5.x changelog breaking changes\")\n\nState the question clearly in one sentence before searching.\n\n### Step 2: SEARCH\n\nExecute searches and read actual content (not just snippets):\n- Use WebSearch for discovery, WebFetch for reading actual pages\n- Evaluate source credibility: official docs > GitHub repos with stars > recent blog posts > forum answers\n- Stop at 3-6 credible sources (not exhaustive — focused)\n- If a source contradicts another, note the disagreement\n\nThe WebFetch Restrictions under Parallel Mode apply in single mode too: never\nWebFetch rendered GitHub pages.\n\n### Step 3: EXTRACT\n\nFor each finding, record:\n- **What**: The specific fact, recommendation, or pattern discovered\n- **Source**: URL or reference\n- **Relevance**: How this applies to the original question (one sentence)\n- **Confidence**: high (official docs, verified), medium (community consensus), low (single source, opinion)\n- **Action**: What the codebase should do with this information (or \"informational only\")\n\n### Step 4: WRITE\n\nWrite findings to `.planning\u002Fresearch\u002F{topic-slug}.md`:\n\n````markdown\n# Research: {Topic}\n\n> Question: {The original question}\n> Date: {ISO date}\n> Confidence: {overall: high\u002Fmedium\u002Flow}\n\n## Findings\n\n### 1. {Finding title}\n**What:** {description}\n**Source:** {URL}\n**Confidence:** {high\u002Fmedium\u002Flow}\n**Action:** {recommendation or \"informational\"}\n\n### 2. {Finding title}\n...\n\n## Summary\n{2-3 sentences: what was learned, what the recommendation is}\n\n## Open Questions\n{Anything that couldn't be resolved — needs human judgment or deeper investigation}\n````\n\n### Step 5: RETURN\n\nReturn the summary and recommendation to the caller (user, Marshal, or Archon).\nThe research document persists for future reference.\n\n## Parallel Mode (--parallel)\n\nThe former \u002Fresearch-fleet behavior now lives in this parallel mode. Spawns\nmultiple scout agents, each investigating a different angle of the same\nquestion. Findings are compressed between waves. Produces a unified research\nbrief from multiple independent perspectives.\n\n**Inputs:** the question, plus optional **angles** (specific sub-questions to\ninvestigate). If not provided, decompose the question into 3-5 angles\nautomatically.\n\n### Step P1: DECOMPOSE\n\nBreak the research question into 3-5 independent angles:\n\nExample: \"Should we migrate from Express to Fastify?\"\n- Scout 1: Performance benchmarks (Express vs Fastify vs Hono, latest data)\n- Scout 2: Migration effort (breaking changes, middleware compatibility, ecosystem)\n- Scout 3: Community health (GitHub stars trend, npm downloads, maintainer activity)\n- Scout 4: Production war stories (who migrated, what broke, was it worth it)\n\nEach angle must be:\n- Independent (scout doesn't need another scout's findings to do its work)\n- Specific (one clear question per scout)\n- Answerable (3-6 sources should be sufficient)\n\n#### Step P1b: INITIALIZE GRAPH (Explicit Opt-In)\n\nWith explicit `--operation-graph`, assign 3-5 opaque angle IDs and initialize a graph plus bound operation:\n`node .citadel\u002Fscripts\u002Foperation-graph-runner.js research-init --project-root . --graph .planning\u002Fresearch\u002Ffleet-{slug}\u002Foperation-graph.json --operation .planning\u002Fresearch\u002Ffleet-{slug}\u002Foperation-spec.json --journal .planning\u002Fresearch\u002Ffleet-{slug}\u002Fgraph-journal --run-id research-{slug} --angles {id1,id2,id3}`\nFor each P2-P5 node, call `node .citadel\u002Fscripts\u002Foperation-graph-effects.js start` with the graph journal, `--effects ...\u002Feffect-journal`, node ID, and payload digest before work; call `node .citadel\u002Fscripts\u002Foperation-graph-effects.js complete` with its evidence digest after verification. Keep prompts and findings outside graph state.\nOn resume run `node .citadel\u002Fscripts\u002Foperation-graph-effects.js status`; resolve blocked nonrepeatable work with reviewed evidence. After the arbiter passes, run `node .citadel\u002Fscripts\u002Foperation-graph-effects.js receipt` with the bound operation. Without the flag, create no graph state.\n\n### Step P2: DEPLOY WAVE 1\n\nSpawn one scout agent per angle using Fleet wave mechanics:\n\nFor each scout:\n1. Create an isolated worktree\n2. Inject the scout's specific angle as the research question\n3. Each scout follows the single-mode protocol (formulate, search, extract, write)\n4. Each scout writes its findings to `.planning\u002Fresearch\u002Ffleet-{slug}\u002F{angle-slug}.md`\n\nAll scouts run in parallel. Wait for all to complete.\n\n### Step P3: COMPRESS\n\nAfter Wave 1 completes:\n\n1. Read all scout findings\n2. Identify:\n   - **Consensus**: findings that multiple scouts independently confirmed\n   - **Conflicts**: findings that contradict each other (flag these prominently)\n   - **Gaps**: angles that didn't produce strong results (consider a Wave 2)\n   - **Surprises**: unexpected findings that change the framing of the question\n3. Compress into a unified brief (~500 tokens)\n\n### Step P4: WAVE 2 (Optional)\n\nIf gaps or conflicts exist:\n\n1. Spawn targeted scouts to resolve specific conflicts or fill gaps\n2. Each Wave 2 scout receives the compressed brief from Wave 1 as context\n3. Wave 2 scouts don't re-research what Wave 1 already covered\n\nSkip Wave 2 if Wave 1 produced clear, consistent findings.\n\n### Step P5: REPORT\n\nWrite the unified report to `.planning\u002Fresearch\u002Ffleet-{slug}\u002FREPORT.md`:\n\n```markdown\n# Research Fleet: {Topic}\n\n> Question: {The original question}\n> Date: {ISO date}\n> Scouts: {N} across {waves} wave(s)\n> Confidence: {overall: high\u002Fmedium\u002Flow}\n\n## Consensus Findings\n{Findings confirmed by 2+ scouts}\n\n## Conflicts\n{Findings where scouts disagreed — present both sides}\n\n## Key Findings by Angle\n\n### {Angle 1}: {title}\n{Summary from scout 1}\nSource: {scout report path}\n\n### {Angle 2}: {title}\n{Summary from scout 2}\nSource: {scout report path}\n\n...\n\n## Recommendation\n{2-3 sentences: what the evidence says, what the recommendation is}\n\n## Open Questions\n{What couldn't be resolved — needs human judgment}\n```\n\nLog the completed wave through Citadel's project-local telemetry helper:\n```bash\nnode .citadel\u002Fscripts\u002Ftelemetry-log.cjs --event wave-complete --agent research-fleet --session fleet-{slug} --status success\n```\nIf the helper is unavailable, skip telemetry without blocking the research result.\nKeep the scout count and total wave count authoritative in `REPORT.md` and the\nparallel-mode HANDOFF rather than duplicating them in telemetry metadata.\n\n### Safety Rules (Parallel Mode)\n\n- Maximum 5 scouts per wave (don't burn tokens on diminishing angles)\n- Maximum 2 waves (if Wave 2 doesn't resolve it, the question needs human judgment)\n- Each scout follows the single-mode quality gates (sources, confidence, evidence)\n- Scout findings are independent. No scout reads another scout's output during the same wave.\n- **Scout timeout: 15 minutes** (configurable via `harness.json` `agentTimeouts.research`).\n  If a scout exceeds its timeout, skip it and proceed with other scouts' results.\n  A timed-out scout's angle becomes a \"Gap\" in the final report.\n\n### WebFetch Restrictions\n\nEvery scout prompt MUST include this instruction:\n\n> **Do NOT use WebFetch on GitHub repository pages** (github.com\u002F{user}\u002F{repo}).\n> These pages are massive HTML documents (500KB+) that hang the fetcher indefinitely.\n> Instead:\n> - Use **WebSearch** to find information about repos (search snippets contain what you need)\n> - If you need a repo's README content, fetch the **raw** URL:\n>   `https:\u002F\u002Fraw.githubusercontent.com\u002F{user}\u002F{repo}\u002F{branch}\u002FREADME.md`\n> - Never fetch rendered GitHub pages: issues, pull requests, repo root, or file views\n\nThis restriction exists because a real research-fleet run hung for 38+ minutes\non `WebFetch(https:\u002F\u002Fgithub.com\u002Fjehna\u002Freadme-best-practices)` with zero output.\nThe circuit breaker didn't catch it because the tool didn't *fail* — it just\nnever completed.\n\n## What \u002Fresearch Does NOT Do\n\n- Make architectural decisions (that's the caller's job)\n- Install packages or modify code\n- Search exhaustively (2-4 queries, 3-6 sources per agent, done)\n- Evaluate subjective opinions as facts\n- Recommend without evidence\n\n## Fringe Cases\n\n- **No web access available**: Fall back to local-only research. Search the codebase, read docs files, check `package.json`, and produce findings from local sources. In parallel mode, run all scouts in local-only mode. Note the limitation in the findings document's confidence level.\n- **Search returns nothing relevant**: Broaden the query (remove version-specific terms, try synonyms), try one more angle. If still empty, report uncertainty explicitly: \"No strong evidence found. Recommend human review.\" A parallel-mode scout that comes up empty reports \"No strong evidence found for this angle\" rather than fabricating findings; the gap becomes an Open Question in the final report.\n- **`.planning\u002Fresearch\u002F` does not exist**: Create it (in parallel mode, including the `fleet-{slug}\u002F` subdirectory) before writing any findings. Never error on a missing output directory.\n- **Conflicting sources**: Surface the conflict explicitly in the findings rather than silently picking one. Both sides belong in the document.\n- **Question is too broad for 3-6 sources**: Escalate to parallel mode, or narrow to the single most important sub-question, answer it well, and note what was scoped out.\n- **Question decomposes into fewer than 3 independent angles**: Stay in (or fall back to) single mode rather than forcing artificial parallelism, even if --parallel was passed.\n- **A scout times out (parallel mode)**: Treat its angle as a gap. Record it in the final report's Open Questions section. Do not block the rest of the wave.\n\n## Contextual Gates\n\n**Reversibility:** Green. Writes files under `.planning\u002Fresearch\u002F` only; delete to undo.\n**Cost:** Single mode spawns no agents and needs no confirmation. Parallel mode spawns up to 5 scouts per wave (max 2 waves); state the scout count before deploying.\n**Trust:** No gates. Read-only investigation plus report files, safe at all trust levels.\n\n## Quality Gates\n\n- Every finding must have a source URL\n- Confidence levels must be justified (not guessed)\n- Summary must answer the original question or state why it can't be answered\n- Research document must be written before returning findings\n\nParallel mode additionally:\n- Every scout must produce a findings document\n- Conflicts must be explicitly flagged, not silently resolved\n- The compressed brief must be written before spawning Wave 2\n\n## Exit Protocol\n\nOutput findings summary, then the HANDOFF for the mode that ran.\n\nSingle mode:\n\n```\n---HANDOFF---\n- Research: {topic}\n- Findings: {count} sources analyzed\n- Recommendation: {one-line summary}\n- Document: .planning\u002Fresearch\u002F{slug}.md\n- Reversibility: green — one file written to .planning\u002Fresearch\u002F; delete to undo\n---\n```\n\nParallel mode:\n\n```\n---HANDOFF---\n- Research (parallel): {topic}\n- Scouts: {N} across {waves} wave(s)\n- Consensus: {one-line summary of agreed findings}\n- Conflicts: {any unresolved disagreements}\n- Recommendation: {one-line}\n- Report: .planning\u002Fresearch\u002Ffleet-{slug}\u002FREPORT.md\n- Reversibility: green — delete `.planning\u002Fresearch\u002Ffleet-{slug}\u002F` to undo\n---\n```\n","# \u002Fresearch — Focused Investigation\n\n## When to Use\n\n- Evaluating whether a dependency has a newer version or has been superseded\n- Finding community best practices for a specific technical problem\n- Reading official documentation for an API or library\n- Investigating how other projects solve a similar problem\n- Checking if a pattern used in the codebase has known issues\n- Any time you need external information before making a decision\n\n**Don't use when:** you need to act on findings immediately (use \u002Fmarshal which calls \u002Fresearch internally).\n\n## Modes\n\n**Single (default):** one agent, 2-4 queries, 3-6 sources. Steps 1-5 below.\n\n**Parallel (`\u002Fresearch --parallel`):** scout agents investigate independent\nangles of the same question using Fleet wave mechanics. See Parallel Mode\nbelow. Without the flag, prefer parallel mode (and say so in the plan) when\nthe question naturally decomposes into 3+ independent angles: evaluating\nmultiple competing technologies or approaches, distinct sub-questions that\ndon't depend on each other, or time-sensitive research where parallel\nexecution matters.\n\nIf the question is narrow and focused, stay single-agent. Don't parallelize\nwhat a single agent can answer in 5 minutes.\n\n## Protocol (Single Mode)\n\n### Step 1: FORMULATE\n\nConvert the research question into 2-4 specific search queries:\n- Official docs query (e.g., \"express.js middleware error handling docs\")\n- Community\u002FGitHub query (e.g., \"express error middleware best practices site:github.com\")\n- Technical blog\u002Fcomparison query (e.g., \"express vs fastify error handling 2025\")\n- Release notes query if version-specific (e.g., \"express 5.x changelog breaking changes\")\n\nState the question clearly in one sentence before searching.\n\n### Step 2: SEARCH\n\nExecute searches and read actual content (not just snippets):\n- Use WebSearch for discovery, WebFetch for reading actual pages\n- Evaluate source credibility: official docs > GitHub repos with stars > recent blog posts > forum answers\n- Stop at 3-6 credible sources (not exhaustive — focused)\n- If a source contradicts another, note the disagreement\n\nThe WebFetch Restrictions under Parallel Mode apply in single mode too: never\nWebFetch rendered GitHub pages.\n\n### Step 3: EXTRACT\n\nFor each finding, record:\n- **What**: The specific fact, recommendation, or pattern discovered\n- **Source**: URL or reference\n- **Relevance**: How this applies to the original question (one sentence)\n- **Confidence**: high (official docs, verified), medium (community consensus), low (single source, opinion)\n- **Action**: What the codebase should do with this information (or \"informational only\")\n\n### Step 4: WRITE\n\nWrite findings to `.planning\u002Fresearch\u002F{topic-slug}.md`:\n\n````markdown\n# Research: {Topic}\n\n> Question: {The original question}\n> Date: {ISO date}\n> Confidence: {overall: high\u002Fmedium\u002Flow}\n\n## Findings\n\n### 1. {Finding title}\n**What:** {description}\n**Source:** {URL}\n**Confidence:** {high\u002Fmedium\u002Flow}\n**Action:** {recommendation or \"informational\"}\n\n### 2. {Finding title}\n...\n\n## Summary\n{2-3 sentences: what was learned, what the recommendation is}\n\n## Open Questions\n{Anything that couldn't be resolved — needs human judgment or deeper investigation}\n````\n\n### Step 5: RETURN\n\nReturn the summary and recommendation to the caller (user, Marshal, or Archon).\nThe research document persists for future reference.\n\n## Parallel Mode (--parallel)\n\nThe former \u002Fresearch-fleet behavior now lives in this parallel mode. Spawns\nmultiple scout agents, each investigating a different angle of the same\nquestion. Findings are compressed between waves. Produces a unified research\nbrief from multiple independent perspectives.\n\n**Inputs:** the question, plus optional **angles** (specific sub-questions to\ninvestigate). If not provided, decompose the question into 3-5 angles\nautomatically.\n\n### Step P1: DECOMPOSE\n\nBreak the research question into 3-5 independent angles:\n\nExample: \"Should we migrate from Express to Fastify?\"\n- Scout 1: Performance benchmarks (Express vs Fastify vs Hono, latest data)\n- Scout 2: Migration effort (breaking changes, middleware compatibility, ecosystem)\n- Scout 3: Community health (GitHub stars trend, npm downloads, maintainer activity)\n- Scout 4: Production war stories (who migrated, what broke, was it worth it)\n\nEach angle must be:\n- Independent (scout doesn't need another scout's findings to do its work)\n- Specific (one clear question per scout)\n- Answerable (3-6 sources should be sufficient)\n\n#### Step P1b: INITIALIZE GRAPH (Explicit Opt-In)\n\nWith explicit `--operation-graph`, assign 3-5 opaque angle IDs and initialize a graph plus bound operation:\n`node .citadel\u002Fscripts\u002Foperation-graph-runner.js research-init --project-root . --graph .planning\u002Fresearch\u002Ffleet-{slug}\u002Foperation-graph.json --operation .planning\u002Fresearch\u002Ffleet-{slug}\u002Foperation-spec.json --journal .planning\u002Fresearch\u002Ffleet-{slug}\u002Fgraph-journal --run-id research-{slug} --angles {id1,id2,id3}`\nFor each P2-P5 node, call `node .citadel\u002Fscripts\u002Foperation-graph-effects.js start` with the graph journal, `--effects ...\u002Feffect-journal`, node ID, and payload digest before work; call `node .citadel\u002Fscripts\u002Foperation-graph-effects.js complete` with its evidence digest after verification. Keep prompts and findings outside graph state.\nOn resume run `node .citadel\u002Fscripts\u002Foperation-graph-effects.js status`; resolve blocked nonrepeatable work with reviewed evidence. After the arbiter passes, run `node .citadel\u002Fscripts\u002Foperation-graph-effects.js receipt` with the bound operation. Without the flag, create no graph state.\n\n### Step P2: DEPLOY WAVE 1\n\nSpawn one scout agent per angle using Fleet wave mechanics:\n\nFor each scout:\n1. Create an isolated worktree\n2. Inject the scout's specific angle as the research question\n3. Each scout follows the single-mode protocol (formulate, search, extract, write)\n4. Each scout writes its findings to `.planning\u002Fresearch\u002Ffleet-{slug}\u002F{angle-slug}.md`\n\nAll scouts run in parallel. Wait for all to complete.\n\n### Step P3: COMPRESS\n\nAfter Wave 1 completes:\n\n1. Read all scout findings\n2. Identify:\n   - **Consensus**: findings that multiple scouts independently confirmed\n   - **Conflicts**: findings that contradict each other (flag these prominently)\n   - **Gaps**: angles that didn't produce strong results (consider a Wave 2)\n   - **Surprises**: unexpected findings that change the framing of the question\n3. Compress into a unified brief (~500 tokens)\n\n### Step P4: WAVE 2 (Optional)\n\nIf gaps or conflicts exist:\n\n1. Spawn targeted scouts to resolve specific conflicts or fill gaps\n2. Each Wave 2 scout receives the compressed brief from Wave 1 as context\n3. Wave 2 scouts don't re-research what Wave 1 already covered\n\nSkip Wave 2 if Wave 1 produced clear, consistent findings.\n\n### Step P5: REPORT\n\nWrite the unified report to `.planning\u002Fresearch\u002Ffleet-{slug}\u002FREPORT.md`:\n\n```markdown\n# Research Fleet: {Topic}\n\n> Question: {The original question}\n> Date: {ISO date}\n> Scouts: {N} across {waves} wave(s)\n> Confidence: {overall: high\u002Fmedium\u002Flow}\n\n## Consensus Findings\n{Findings confirmed by 2+ scouts}\n\n## Conflicts\n{Findings where scouts disagreed — present both sides}\n\n## Key Findings by Angle\n\n### {Angle 1}: {title}\n{Summary from scout 1}\nSource: {scout report path}\n\n### {Angle 2}: {title}\n{Summary from scout 2}\nSource: {scout report path}\n\n...\n\n## Recommendation\n{2-3 sentences: what the evidence says, what the recommendation is}\n\n## Open Questions\n{What couldn't be resolved — needs human judgment}\n```\n\nLog the completed wave through Citadel's project-local telemetry helper:\n```bash\nnode .citadel\u002Fscripts\u002Ftelemetry-log.cjs --event wave-complete --agent research-fleet --session fleet-{slug} --status success\n```\nIf the helper is unavailable, skip telemetry without blocking the research result.\nKeep the scout count and total wave count authoritative in `REPORT.md` and the\nparallel-mode HANDOFF rather than duplicating them in telemetry metadata.\n\n### Safety Rules (Parallel Mode)\n\n- Maximum 5 scouts per wave (don't burn tokens on diminishing angles)\n- Maximum 2 waves (if Wave 2 doesn't resolve it, the question needs human judgment)\n- Each scout follows the single-mode quality gates (sources, confidence, evidence)\n- Scout findings are independent. No scout reads another scout's output during the same wave.\n- **Scout timeout: 15 minutes** (configurable via `harness.json` `agentTimeouts.research`).\n  If a scout exceeds its timeout, skip it and proceed with other scouts' results.\n  A timed-out scout's angle becomes a \"Gap\" in the final report.\n\n### WebFetch Restrictions\n\nEvery scout prompt MUST include this instruction:\n\n> **Do NOT use WebFetch on GitHub repository pages** (github.com\u002F{user}\u002F{repo}).\n> These pages are massive HTML documents (500KB+) that hang the fetcher indefinitely.\n> Instead:\n> - Use **WebSearch** to find information about repos (search snippets contain what you need)\n> - If you need a repo's README content, fetch the **raw** URL:\n>   `https:\u002F\u002Fraw.githubusercontent.com\u002F{user}\u002F{repo}\u002F{branch}\u002FREADME.md`\n> - Never fetch rendered GitHub pages: issues, pull requests, repo root, or file views\n\nThis restriction exists because a real research-fleet run hung for 38+ minutes\non `WebFetch(https:\u002F\u002Fgithub.com\u002Fjehna\u002Freadme-best-practices)` with zero output.\nThe circuit breaker didn't catch it because the tool didn't *fail* — it just\nnever completed.\n\n## What \u002Fresearch Does NOT Do\n\n- Make architectural decisions (that's the caller's job)\n- Install packages or modify code\n- Search exhaustively (2-4 queries, 3-6 sources per agent, done)\n- Evaluate subjective opinions as facts\n- Recommend without evidence\n\n## Fringe Cases\n\n- **No web access available**: Fall back to local-only research. Search the codebase, read docs files, check `package.json`, and produce findings from local sources. In parallel mode, run all scouts in local-only mode. Note the limitation in the findings document's confidence level.\n- **Search returns nothing relevant**: Broaden the query (remove version-specific terms, try synonyms), try one more angle. If still empty, report uncertainty explicitly: \"No strong evidence found. Recommend human review.\" A parallel-mode scout that comes up empty reports \"No strong evidence found for this angle\" rather than fabricating findings; the gap becomes an Open Question in the final report.\n- **`.planning\u002Fresearch\u002F` does not exist**: Create it (in parallel mode, including the `fleet-{slug}\u002F` subdirectory) before writing any findings. Never error on a missing output directory.\n- **Conflicting sources**: Surface the conflict explicitly in the findings rather than silently picking one. Both sides belong in the document.\n- **Question is too broad for 3-6 sources**: Escalate to parallel mode, or narrow to the single most important sub-question, answer it well, and note what was scoped out.\n- **Question decomposes into fewer than 3 independent angles**: Stay in (or fall back to) single mode rather than forcing artificial parallelism, even if --parallel was passed.\n- **A scout times out (parallel mode)**: Treat its angle as a gap. Record it in the final report's Open Questions section. Do not block the rest of the wave.\n\n## Contextual Gates\n\n**Reversibility:** Green. Writes files under `.planning\u002Fresearch\u002F` only; delete to undo.\n**Cost:** Single mode spawns no agents and needs no confirmation. Parallel mode spawns up to 5 scouts per wave (max 2 waves); state the scout count before deploying.\n**Trust:** No gates. Read-only investigation plus report files, safe at all trust levels.\n\n## Quality Gates\n\n- Every finding must have a source URL\n- Confidence levels must be justified (not guessed)\n- Summary must answer the original question or state why it can't be answered\n- Research document must be written before returning findings\n\nParallel mode additionally:\n- Every scout must produce a findings document\n- Conflicts must be explicitly flagged, not silently resolved\n- The compressed brief must be written before spawning Wave 2\n\n## Exit Protocol\n\nOutput findings summary, then the HANDOFF for the mode that ran.\n\nSingle mode:\n\n```\n---HANDOFF---\n- Research: {topic}\n- Findings: {count} sources analyzed\n- Recommendation: {one-line summary}\n- Document: .planning\u002Fresearch\u002F{slug}.md\n- Reversibility: green — one file written to .planning\u002Fresearch\u002F; delete to undo\n---\n```\n\nParallel mode:\n\n```\n---HANDOFF---\n- Research (parallel): {topic}\n- Scouts: {N} across {waves} wave(s)\n- Consensus: {one-line summary of agreed findings}\n- Conflicts: {any unresolved disagreements}\n- Recommendation: {one-line}\n- Report: .planning\u002Fresearch\u002Ffleet-{slug}\u002FREPORT.md\n- Reversibility: green — delete `.planning\u002Fresearch\u002Ffleet-{slug}\u002F` to undo\n---\n```\n",{"name":5,"license":15,"description":7,"auto-trigger":47,"last-updated":51,"user-invocable":52,"trigger_keywords":53},1784592000000,true,[5,43,54,55,56,57,58,59],"look into","find out","research fleet","parallel research","multi-angle research","compare options","main","e1a3d45697a0f323aaedcbea0ab6a0387b5fe722","skills\u002Fresearch",84,"https:\u002F\u002Fgithub.com\u002FSethGammon\u002FCitadel","https:\u002F\u002Fsethgammon.github.io\u002FCitadel\u002F","https:\u002F\u002Fgithub.com\u002FSethGammon","Building Citadel, the repo-local operating layer for Claude Code and Codex. Persistent work, routing, verification, and coordinated agent fleets.","2026-11-06 02:04:44",{"source":8,"skillName":5,"commands":70,"prompt":80,"zipAvailable":52},[71,74,77],{"tool":72,"command":73},"npx","npx skills add SethGammon\u002FCitadel --skill research",{"tool":75,"command":76},"bunx","bunx skills add SethGammon\u002FCitadel --skill research",{"tool":78,"command":79},"pnpm","pnpm dlx skills add SethGammon\u002FCitadel --skill research","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002FSethGammon\u002FCitadel\u002Fblob\u002Fmain\u002Fskills\u002Fresearch\u002FSKILL.md",[82,88,92,96],{"path":83,"resourceType":84,"resourceTypeDesc":85,"sizeBytes":86,"sourceHtmlUrl":87},"__benchmarks__\u002Fexternal-library-research.md","REFERENCE","参考文档",1044,"https:\u002F\u002Fgithub.com\u002FSethGammon\u002FCitadel\u002Fblob\u002Fmain\u002Fskills\u002Fresearch\u002F__benchmarks__\u002Fexternal-library-research.md",{"path":89,"resourceType":84,"resourceTypeDesc":85,"sizeBytes":90,"sourceHtmlUrl":91},"__benchmarks__\u002Fno-web-results.md",1155,"https:\u002F\u002Fgithub.com\u002FSethGammon\u002FCitadel\u002Fblob\u002Fmain\u002Fskills\u002Fresearch\u002F__benchmarks__\u002Fno-web-results.md",{"path":93,"resourceType":84,"resourceTypeDesc":85,"sizeBytes":94,"sourceHtmlUrl":95},"__benchmarks__\u002Fparallel-decompose-angles.md",1279,"https:\u002F\u002Fgithub.com\u002FSethGammon\u002FCitadel\u002Fblob\u002Fmain\u002Fskills\u002Fresearch\u002F__benchmarks__\u002Fparallel-decompose-angles.md",{"path":97,"resourceType":84,"resourceTypeDesc":85,"sizeBytes":98,"sourceHtmlUrl":99},"__benchmarks__\u002Fparallel-no-research-dir.md",1147,"https:\u002F\u002Fgithub.com\u002FSethGammon\u002FCitadel\u002Fblob\u002Fmain\u002Fskills\u002Fresearch\u002F__benchmarks__\u002Fparallel-no-research-dir.md",[101,125,150,174,200,219],{"skillNo":102,"name":103,"nameZh":6,"description":104,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":105,"categories":106,"keywords":107,"publishedAt":46},"01M6GXSBKE9A2HQWF8QRAZ4WJ5","workspace","Multi-repo campaign coordinator. Same lifecycle as fleet -- scope claims, discovery relay, wave-based execution -- but the unit of work is a repo, not a file. Coordinates campaigns across repositories with shared context.","https:\u002F\u002Fgithub.com\u002FSethGammon\u002FCitadel\u002Fblob\u002Fmain\u002Fskills\u002Fworkspace\u002FSKILL.md",[],[103,108,109,110,111,112,113,114,115,116,117,118,119,36,120,121,122,123,28,124],"repos","snowflake","multi-repo","work","context","campaign","coordinate","but","across","campaigns","shared","as","scope","add","unit","coordinator","claims",{"skillNo":126,"name":127,"nameZh":6,"description":128,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":129,"categories":130,"keywords":132,"publishedAt":46},"01M6GXS873AD1Y6RTM6GV47F3E","wiki","Markdown-first knowledge base where the LLM acts as librarian. Ingests raw sources, compiles and interlinks topic files, self-maintains an index. No vector DB or embeddings required -- uses LLM-native navigation over structured markdown up to ~400K words.","https:\u002F\u002Fgithub.com\u002FSethGammon\u002FCitadel\u002Fblob\u002Fmain\u002Fskills\u002Fwiki\u002FSKILL.md",[131],{"categoryNo":19,"categoryName":20,"taxonomyType":21,"taxonomyTypeDesc":22},[127,133,134,135,136,137,138,139,140,141,142,143,144,145,146,147,119,36,148,149],"vector","embeddings","librarian","markdown","knowledge","karpathy","knowledge management","karpathy wiki","up","files","maintain","self","knowledge base","project","db","sources","native",{"skillNo":151,"name":152,"nameZh":6,"description":153,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":154,"categories":155,"keywords":156,"publishedAt":46},"01M6GXS2840SJ61285SZ43EJC3","watch","File sentinel that monitors the working directory for changes and marker comments, then auto-triggers appropriate skills. Poll-based via git diff against the last scan commit. Writes intake items for batch processing and routes marker actions through \u002Fdo. Use for automatic reactions to file changes; do NOT use for one-off inspection or tasks needing human judgment per file.","https:\u002F\u002Fgithub.com\u002FSethGammon\u002FCitadel\u002Fblob\u002Fmain\u002Fskills\u002Fwatch\u002FSKILL.md",[],[152,157,158,159,160,161,162,163,164,165,166,167,142,168,169,170,171,172,36,173],"directory","commit","tasks","git","batch","actions","start","sentinel","diff","auto","stop","do","scan","human","judgment","reactions","processing",{"skillNo":175,"name":176,"nameZh":6,"description":177,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":178,"categories":179,"keywords":183,"publishedAt":46},"01M6GXRXHYJ5RD9RZP81S2VFSW","verify","Self-test the Citadel hook pipeline from within a live session. Exercises real tool calls (Write, Edit, Bash, Read) and checks that hooks fired, telemetry accumulated, and no errors occurred. Reports HOOK HEALTH: PASS or HOOK HEALTH: FAIL with a per-hook breakdown.","https:\u002F\u002Fgithub.com\u002FSethGammon\u002FCitadel\u002Fblob\u002Fmain\u002Fskills\u002Fverify\u002FSKILL.md",[180],{"categoryNo":181,"categoryName":182,"taxonomyType":21,"taxonomyTypeDesc":22},"01JBSEEDCAT000000000000008","测试与质量",[176,184,185,186,187,188,189,190,191,192,193,194,195,144,196,36,197,198,199,29],"citadel","telemetry","harness","pipeline","Bash","tool","read","write","edit","session","hooks","reports","live","hook","errors","calls",{"skillNo":201,"name":202,"nameZh":6,"description":203,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":204,"categories":205,"keywords":206,"publishedAt":46},"01M6GXRS9MCJFMQZ917H2XH3AC","unharness","Safely leave Citadel using the active adoption receipt. Produces a no-write, reviewable plan, preserves a portable archive, removes only exact owned material, and reports modified or externally registered surfaces as retained or unknown. Legacy installs must be imported before exact leave is claimed.","https:\u002F\u002Fgithub.com\u002FSethGammon\u002FCitadel\u002Fblob\u002Fmain\u002Fskills\u002Funharness\u002FSKILL.md",[],[202,207,184,186,208,209,210,211,212,141,195,213,214,215,216,217,119,36,218,29],"uninstall","adoption","plan","archive","remove","only","Legacy","surfaces","clean","imported","active","material",{"skillNo":220,"name":221,"nameZh":6,"description":222,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":223,"categories":224,"keywords":233,"publishedAt":46},"01M6GXRMEJ2VHD43R9FNPD4910","triage","GitHub issue and PR investigator. Pulls open issues\u002FPRs, classifies them, searches the codebase for root cause or reviews contributed code, proposes fixes with file:line references, and optionally implements fixes. Use for investigating GitHub issues and reviewing PRs; do NOT use for general code review unrelated to GitHub issues.","https:\u002F\u002Fgithub.com\u002FSethGammon\u002FCitadel\u002Fblob\u002Fmain\u002Fskills\u002Ftriage\u002FSKILL.md",[225,228],{"categoryNo":226,"categoryName":227,"taxonomyType":21,"taxonomyTypeDesc":22},"01JBSEEDCAT000000000000007","代码审查",{"categoryNo":229,"categoryName":230,"taxonomyType":231,"taxonomyTypeDesc":232},"01JBSEEDCAT000000000000030","开发工程师","PROFESSION","适用职业",[221,234,235,236,237,238,239,240,241,242,243,244,245,246,168,36,247,248,249,29],"code review","review","review pr","pr","code","github issues","issues","open","use","unrelated","issue","general","references","line","investigator","reviews"]