| name | deepresearch |
| description | Iterative web-current research with source ledger. Sonnet collects, Opus synthesizes. Parallel WebSearch + WebFetch across sub-queries. Output is a structured ResearchBrief with citations, persisted to memory.research_sources. The default first step of every umbrella when the goal is novel or time-sensitive. |
/deepresearch — sourced research with ledger
The research substrate for every umbrella. Anchors decisions on current-time evidence, not training-cutoff knowledge.
Process
- Memory-first ledger lookup. Before any web call, query
memory.research_sources for the topic within the last 7 days (default window). If a fresh prior brief exists, reuse it and skip steps 2–5. Outside the window, surface the prior brief as "possibly stale" alongside a fresh pass so the caller can compare. Disable the shortcut with memoryFirst: false (or --no-memory-first on the CLI) when the topic is known-volatile and must re-fetch every run.
- Decompose the topic into 3–8 sub-queries.
- Run WebSearch + WebFetch in parallel across the sub-queries.
- Read primary sources first (official docs, vendor announcements, RFCs, repos). De-prioritize SEO blogspam.
- Resolve conflicts by citing the most recent authoritative source.
- Synthesize via Opus into a ResearchBrief: topic, summary, keyFindings, sources (URL + relevance + excerpt + accessedAt), openQuestions.
- Persist sources to
memory.research_sources for future re-use.
When to invoke (manually)
- Before architectural decisions where SOTA matters (
/cmax and friends already invoke this automatically).
- When evaluating libraries, models, or services as of today.
- When the goal involves a regulation, pricing, or API that changes frequently.
Output
JSON ResearchBrief matching the schema in packages/core/src/types.ts. CLI: cmax research "<topic>" prints the brief and writes sources to memory.
Anti-pattern protection
- Citation linkage is schema-enforced. Every
keyFinding now carries ≥ 1 sourceUrl drawn from the brief's sources array — the JSON schema rejects findings without one. Uncitable findings are dropped at generation time rather than emitted naked; the dark-patterns no-fake-cite hook is the second line of defense, not the first.
- Every claim attaches ≥ 1 source URL with an
accessedAt field.
- Sources with relevance < 0.3 are dropped.
openQuestions is mandatory — research that returns "everything is clear" is suspicious.
