research（研究）

Deep Research

General-purpose deep research with multi-source synthesis, confidence scoring, and anti-hallucination verification. Adopts SOTA patterns from OpenAI Deep Research (multi-agent triage pipeline), Google Gemini Deep Research (user-reviewable plans), STORM (perspective-guided conversations), Perplexity (source confidence ratings), and LangChain ODR (supervisor-researcher with reflection).

Vocabulary

Dispatch

Auto-Detection Heuristic

If no mode keyword matches:

1. Ends with ? or starts with question word (who/what/when/where/why/how/is/are/can/does/should/will) → Investigate
2. Contains vs, versus, compared to, or between noun phrases → Compare
3. Declarative statement with factual claim, no question syntax → Fact-check
4. Broad field name with no specific question → ask: "Investigate a specific question, or survey the entire field?"
5. Ambiguous → ask: "Would you like me to investigate this question, verify this claim, or survey this field?"

Gallery (Empty Arguments)

Present research examples spanning domains:

Pick a number, paste your own question, or type guide me.

Skill Awareness

Before starting research, check if another skill is a better fit:

If the user confirms they want general research, proceed.

Complexity Classification

Score the query on 5 dimensions (0-2 each, total 0-10):

Present the scoring to the user. User can override tier with --depth <tier>.

Scaling Strategy

Scale work by query complexity and available orchestration capabilities:

Capability resolution: Treat named tools and orchestration APIs as preferred capabilities. Claude Code may use Task/TeamCreate; Codex may use dynamic subagents or parallel tool calls; other agents may use their native delegation or run the wave pipeline serially. If no delegation equivalent exists, use degraded orchestration: preserve wave order, reduce breadth, and report the limitation in methodology. Apply confidence ceilings only when source or retrieval capabilities are unavailable, per references/source-selection.md.

Wave Pipeline

All non-Quick research follows this 5-wave pipeline. Quick merges Waves 0+1+4 inline.

Wave 0: Triage (always inline, never parallelized)

1. Run !uv run python skills/research/scripts/research-scanner.py "$ARGUMENTS" for deterministic pre-scan
2. Decompose query into 2-5 sub-questions
3. Score complexity on the 5-dimension rubric
4. Check tool availability — probe key MCP tools; set degraded mode flags and confidence ceilings per references/source-selection.md
5. Select tools per domain signals — read references/source-selection.md
6. Check for existing journals — if track or resume, load prior state
7. Present triage to user — show: complexity score, sub-questions, planned strategy, estimated tier. User may override.

Wave 1: Broad Sweep (parallel)

Scale by tier:

Quick (inline): 1-2 tool calls sequentially. No subagents.

Standard (subagent wave): Dispatch 3-5 parallel subagents with the platform's available delegation primitive:

Subagent A → brave-search + duckduckgo-search for sub-question 1
Subagent B → exa + g-search for sub-question 2
Subagent C → context7 / deepwiki / arxiv / semantic-scholar for technical specifics
Subagent D → wikipedia / wikidata for factual grounding
[Subagent E → PubMed / openalex if academic domain detected]

Deep (agent team): Create a research team with the platform's available team primitive:

Lead: triage (Wave 0), orchestrate, judge reconcile (Wave 3), synthesize (Wave 4)
  |-- web-researcher:       brave-search, duckduckgo-search, exa, g-search
  |-- tech-researcher:      context7, deepwiki, arxiv, semantic-scholar, package-version
  |-- content-extractor:    fetcher, trafilatura, docling, wikipedia, wayback
  |-- [academic-researcher: arxiv, semantic-scholar, openalex, crossref, PubMed]
  |-- [adversarial-reviewer: devil's advocate — counter-search all emerging findings]

Spawn academic-researcher if domain signals include academic/scientific. Spawn adversarial-reviewer for Exhaustive tier or if verification complexity >= 2.

Exhaustive: Deep team + each teammate runs nested subagent waves internally when supported; otherwise use serial batches and label the run "degraded orchestration."

Each subagent/teammate returns structured findings:

{
  "sub_question": "...",
  "findings": [{
    "claim": "...",
    "confidence": 0.6,
    "evidence": [{"tool": "brave-search", "url": "https://...", "timestamp": "2026-04-24T12:00:00Z", "excerpt": "..."}],
    "cross_validation": "unknown",
    "bias_markers": [],
    "gaps": []
  }],
  "leads": ["url1", "url2"],
  "gaps": ["could not find data on X"]
}

Wave 1.5: Perspective Expansion (Deep/Exhaustive only)

STORM-style perspective-guided conversation. Spawn 2-4 perspective subagents:

Each perspective agent reviews Wave 1 findings and generates 2-3 additional sub-questions from their viewpoint. These sub-questions feed into Wave 2.

Wave 2: Deep Dive (parallel, targeted)

1. Rank leads from Wave 1 by potential value (citation frequency, source authority, relevance)
2. Dispatch deep-read subagents — use fetcher/trafilatura/docling to extract full content from top leads
3. Follow citation chains — if a source cites another, fetch the original
4. Fill gaps — for each gap identified in Wave 1, dispatch targeted searches
5. Use thinking MCPs:

- cascade-thinking for multi-perspective analysis of complex findings - structured-thinking for tracking evidence chains and contradictions - think-strategies for complex question decomposition (Standard+ only)

Wave 3: Cross-Validation (parallel)

The anti-hallucination wave. Read references/confidence-rubric.md and references/self-verification.md.

For every claim surviving Waves 1-2:

1. Independence check — are supporting sources truly independent? Sources citing each other are NOT independent.
2. Counter-search — explicitly search for evidence AGAINST each major claim using a different search engine
3. Freshness check — verify sources are current (flag if >1 year old for time-sensitive topics)
4. Contradiction scan — read references/contradiction-protocol.md, identify and classify disagreements
5. Confidence scoring — assign 0.0-1.0 per references/confidence-rubric.md
6. Bias sweep — check each finding against 10 bias categories (7 core + 3 LLM-specific) per references/bias-detection.md

Self-Verification (3+ findings survive): Spawn devil's advocate subagent per references/self-verification.md:

For each finding, attempt to disprove it. Search for counterarguments. Check if evidence is outdated. Verify claims actually follow from cited evidence. Flag LLM confabulations.

Adjust confidence: Survives +0.05, Weakened -0.10, Disproven set to 0.0. Adjustments are subject to hard caps — single-source claims remain capped at 0.60 even after survival adjustment.

Wave 4: Synthesis (always inline, lead only)

Produce the final research product. Read references/output-formats.md for templates.

The synthesis is NOT a summary. It must:

1. Answer directly — answer the user's question clearly
2. Map evidence — all verified findings with confidence and citations
3. Surface contradictions — where sources disagree, with analysis of why
4. Show confidence landscape — what is known confidently, what is uncertain, what is unknown
5. Audit biases — biases detected during research
6. Identify gaps — what evidence is missing, what further research would help
7. Distill takeaways — 3-7 numbered key findings
8. Cite sources — full bibliography with provenance

Output format adapts to mode:

• Investigate → Research Brief (Standard) or Deep Report (Deep/Exhaustive)
• Fact-check → Quick Answer with verdict + evidence
• Compare → Decision Matrix
• Survey → Annotated Bibliography
• User can override with --format brief|deep|bib|matrix

Confidence Scoring

Hard rules:

• No claim reported at >= 0.7 unless supported by 2+ independent sources
• Single-source claims cap at 0.6 regardless of source authority
• Degraded mode (all research tools unavailable): max confidence 0.4, all findings labeled "unverified"

Merged confidence (for claims supported by multiple sources): c_merged = 1 - (1-c1)(1-c2)...(1-cN) capped at 0.99

Evidence Chain Structure

Every finding carries this structure:

FINDING RR-{seq:03d}: [claim statement]
  CONFIDENCE: [0.0-1.0]
  EVIDENCE:
    1. [source_tool] [url] [access_timestamp] — [relevant excerpt, max 100 words]
    2. [source_tool] [url] [access_timestamp] — [relevant excerpt, max 100 words]
  CROSS-VALIDATION: [agrees|contradicts|partial] across [N] independent sources
  BIAS MARKERS: [none | list of detected biases with category]
  GAPS: [none | what additional evidence would strengthen this finding]

Use !uv run python skills/research/scripts/finding-formatter.py --format markdown to normalize.

Source Selection

Read references/source-selection.md during Wave 0 for the full tool-to-domain mapping. Summary:

Multi-engine protocol: For any claim requiring verification, use minimum 2 different search engines. Different engines have different indices and biases. Agreement across engines increases confidence.

Progressive Disclosure

Load only the next required reference:

1. Start with this file for routing, classification, and wave order.
2. Load references/source-selection.md during Wave 0 only.
3. Load validation references during Wave 3 only: confidence, contradiction, self-verification, and bias files as needed.
4. Load references/output-formats.md or references/dashboard-schema.md only when producing final output or exports.
5. Never preload all references; summarize tool limitations instead of filling context with unused mappings.

Bias Detection

Check every finding against 10 bias categories. Read references/bias-detection.md for full detection signals and mitigation strategies.

State Management

• Journal path: ~/.{gemini|copilot|codex|claude}/research/
• Archive path: ~/.{gemini|copilot|codex|claude}/research/archive/
• Filename convention: {YYYY-MM-DD}-{domain}-{slug}.md

- {domain}: tech, academic, market, policy, factcheck, compare, survey, track, general - {slug}: 3-5 word semantic summary, kebab-case - Collision: append -v2, -v3

• Format: YAML frontmatter + markdown body + <!-- STATE --> blocks

Save protocol:

• Quick: save once at end with status: Complete
• Standard/Deep/Exhaustive: save after Wave 1 with status: In Progress, update after each wave, finalize after synthesis

Resume protocol:

1. resume (no args): find status: In Progress journals. One → auto-resume. Multiple → show list.
2. resume N: Nth journal from list output (reverse chronological).
3. resume keyword: search frontmatter query and domain_tags for match.

Use !uv run python skills/research/scripts/journal-store.py for all journal operations.

State snapshot (appended after each wave save):

<!-- STATE
wave_completed: 2
findings_count: 12
leads_pending: ["url1", "url2"]
gaps: ["topic X needs more sources"]
contradictions: 1
next_action: "Wave 3: cross-validate top 8 findings"
-->

In-Session Commands (Deep/Exhaustive)

Available during active research sessions:

Read references/session-commands.md for full protocols.

Reference File Index

Loading rule: Load ONE reference at a time per the "Read When" column. Do not preload.

Critical Rules

1. No claim >= 0.7 unless supported by 2+ independent sources — single-source claims cap at 0.6
2. Never fabricate citations — if URL, author, title, or date cannot be verified, use vague attribution ("a study in this tradition") rather than inventing specifics
3. Always surface contradictions explicitly — never silently resolve disagreements; present both sides with evidence
4. Always present triage scoring before executing research — user must see and can override complexity tier
5. Save journal after every wave in Deep/Exhaustive mode — enables resume after interruption
6. Never skip Wave 3 (cross-validation) for Standard/Deep/Exhaustive tiers — this is the anti-hallucination mechanism
7. Multi-engine search is mandatory for fact-checking — use minimum 3 different search tools (e.g., brave-search + duckduckgo-search + exa)
8. Apply the Accounting Rule after every parallel dispatch — N dispatched = N accounted for before proceeding to next wave
9. Distinguish facts from interpretations in all output — factual claims carry evidence; interpretive claims are explicitly labeled as analysis
10. Flag all LLM-prior findings — claims matching common training data but lacking fresh evidence must be flagged with bias marker
11. Max confidence 0.4 in degraded mode — when all research tools are unavailable, report all findings as "unverified — based on training knowledge"
12. Load ONE reference file at a time — do not preload all references into context
13. Track mode must load prior journal before searching — avoid re-researching what is already known
14. The synthesis is not a summary — it must integrate findings into novel analysis, identify patterns across sources, and surface emergent insights not present in any single source
15. PreToolUse write guard is non-negotiable — the research skill never modifies source files; it only creates/updates journals in ~/.{gemini|copilot|codex|claude}/research/
16. Stop hook must pass — verify.py stop confirms the skill did not leave tracked research-source files dirty
17. Normalize legacy findings before synthesis — top-level source_url, source_tool, and confidence_raw must be converted into the canonical evidence[] + confidence shape