/research

description: "Deep multi-agent research on a topic, question, or thesis. Launches parallel agents to search the web, ingests sources, and compiles them into active wiki articles. Thesis mode provides for/against evidence framing with a verdict."
argument-hint: "<topic|question> [--plan] [--mode thesis \"<claim>\"] [--new-topic] [--sources <N>] [--deep] [--retardmax] [--min-time <duration>] [--specialist <name>] [--no-specialists] [--wiki <name>] [--local] [--project <slug>] [--include-archived]"
allowed-tools: Read, Write, Edit, Glob, Grep, Bash(ls:*), Bash(wc:*), Bash(date:*), Bash(mkdir:*), Bash(python3:*), Bash(scripts/llm-wiki:*), Bash(${CLAUDE_PLUGIN_ROOT}/bin/llm-wiki:*), WebFetch, WebSearch, Agent

Your task

Conduct deep research on the topic in $ARGUMENTS. This is an automated pipeline: search → ingest → compile.

Parse $ARGUMENTS

• **topic/question**: The research subject — everything that is not a flag. Auto-detected as topic vs question (see Input Detection below).
• **--mode thesis "<claim>"**: Activate thesis mode. The claim is the scope constraint — sources that don't relate to it are skipped. Agents evaluate evidence for/against. A verdict is rendered at the end. See Thesis Research Mode below.
• **--new-topic**: Create a new topic wiki from the topic name and start researching into it. Works from any directory. Derives the wiki slug from the topic (e.g., "quantum error correction" → `quantum-error-correction`).
• **--sources <N>**: Target sources PER ROUND (default: 5, max: 20)
• **--deep**: 8 parallel agents with broader search angles
• **--retardmax**: 10 agents, skip planning, ingest aggressively
• **--min-time <duration>**: Minimum research time. Keep running research rounds until this duration is reached. Formats: `30m`, `1h`, `2h`, `4h`. Default: single round (no minimum).
• **--wiki <name>**: Target a specific existing topic wiki
• **--local**: Use project-local `.wiki/`
• **--plan**: Decompose the topic into 3-5 independent research paths, present the plan for confirmation, then execute all paths in parallel. Each path gets its own agent group (5 agents standard, 8 with `--deep`). Ingest runs in parallel across paths (each path writes unique raw files); compilation runs once after all paths complete (single pass sees all sources for better cross-referencing). Without `--plan`, research runs a single path as before. See Plan Mode below.
• **--project <slug>**: Tag all new outputs with this project. The research playbook/summary artifact is saved inside `output/projects/<slug>/` instead of flat `output/`. Compiled wiki articles get `project: <slug>` frontmatter. If the project doesn't exist, fail early with a helpful error. See `references/projects.md` for the projects architecture.
• **--include-archived**: Explicitly allow researching into an archived target

wiki. Keep the target archived and label the session accordingly.

• **--specialist <name>**: Explicitly apply an enabled personal specialist;

repeat up to three times. For project-local `.wiki/`, explicit selection is required because v1 has no persisted local-project allowlist.

• **--no-specialists**: Run the baseline research workflow without reading

specialist package bodies. Mutually exclusive with `--specialist`.

`--project <slug>` flag

If `--project <slug>` is set, the research playbook/summary artifact is saved inside `output/projects/<slug>/` instead of flat `output/`, and compiled wiki articles get `project: <slug>` frontmatter. Verify the project exists at `<wiki-root>/output/projects/<slug>/WHY.md` before starting; if not, fail early with a suggestion to run `/wiki:project new <slug> "goal"` first.

There is no ambient project focus — pass `--project` explicitly when you want project scope. The focus-session mechanism was removed in the v0.2 projects simplification; see `references/projects.md` § "Focus" for the rationale.

Inventory awareness

At the start, read inventory indexes when cheap. Active `p0`/`p1` items, candidates, corpora, questions, and watch items can seed research paths or explain why a topic matters. Do not treat inventory records as evidence; they are operational state.

At the end, propose inventory records only for durable follow-ups: important unresolved gaps, item/part decisions, candidate sources/corpora to revisit, watch items, or blocked tasks. Do not create a backlog for every minor search tangent. For larger pivots, show a 1-3 row sample of proposed records and ask before writing.

Archive awareness

Research ignores archived topic wikis by default. If `--wiki <name>` resolves to an archived registry entry or a path under `topics/.archive/`, stop and ask the user to restore it or rerun with `--include-archived`. When no target wiki is specified, archived topics may be mentioned as overlap only: read their master `_index.md` files, warn that a related archived topic exists, and ask whether to restore it or continue fresh. Do not ingest or compile into archived topics through auto-classification.

Specialist selection

After resolving the target and before decomposing the question, read `skills/wiki-manager/references/specialists.md`. If `--no-specialists` is set, record an empty specialist list and continue without reading `HUB/.skills/`.

For a hub topic, read `HUB/.skills/_index.md` when present, then run:

"${CLAUDE_PLUGIN_ROOT}/bin/llm-wiki" specialist list \
  --wiki <topic-slug> --json

In a source checkout use `scripts/llm-wiki`. With explicit `--specialist` flags, require every name in the returned allowlist and validate each package; stop rather than silently substituting. Without explicit flags, classify the task by domain, decision type, risk, jurisdiction, freshness, and evidence needs, then choose zero to three allowlisted specialists (normally one). Run `specialist validate <name>` for every selected method and stop on any finding. Do not load unselected package bodies.

For project-local `.wiki/`, make n