Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/jsagir/mindrian-os-plugin/llms.txt

Use this file to discover all available pages before exploring further.

When you need fresh evidence from the web cross-referenced with the Brain methodology graph, /mos:research runs a dual-source pull — public signal plus calibrated framework. It extracts your room context first, surfaces each finding with a candidate filing location and a decision card, then wires accepted findings as typed graph data other commands can consume. With Brain connected, every finding is semantically cross-referenced against 12,485 methodology vectors. Without Brain (Tier 0), research still runs and files typed evidence — the web-only path works fully.

Usage

/mos:research [topic | url]
/mos:research --broad [topic]
ArgumentDescription
topicA natural-language question or keyword string (e.g. "competitive landscape for B2B SaaS invoicing")
urlA bare http(s):// URL — routes to the URL ingest pipeline
--broadOverrides the computed lens set with a fixed three-lens preset: scholarly, industry, patent — all at equal weight
If you run /mos:research with no argument, Larry will ask: “What do you want me to research? Give me a specific question or topic related to your venture.”

What Happens

1

Room context extraction

Larry reads your room — the active section, current JTBD, persona role, and any existing evidence gaps tagged needs_evidence. This becomes a conversational context summary surfaced before any fetch: “Speaking to your investor lens, you are in the build-thesis workflow… I will research this topic against THIS context.”
2

Lens set computation

From the context summary, Larry computes an ordered, weighted lens set — the sources to query (e.g. scholarly, industry, patent, Brain). The lens names are surfaced so you can see which sources will run. With --broad, this is overridden by the fixed three-lens preset.
3

Execution via source-lens driver

The driver queries each lens cache-first (30-day TTL). On a miss it fetches live, writes back to cache, and deduplicates against prior research. Results are ranked by evidence tier and relevance. Up to the top 5 findings are returned.
4

Findings presented in venture context

Each finding is shown with: title, one-line summary, source URL, retrieval timestamp, evidence tier, and a pre-mapped candidate room section with a match score against that section’s existing claim graph. Raw search results are never dumped.
5

Filing decision per finding (F.1 gate)

For each finding, Larry renders a filing selector with five options: file to the primary section (recommended when confidence clears 0.7), file to a secondary section, split across both, defer to milestone audit, or reject with a captured reason. Nothing files without your explicit verb.
6

Wiring accepted findings to the graph

Approved findings are written as EvidenceClaim nodes with review_status: proposed. Typed edges are added: INFORMS to the target section, CONTRADICTS when a finding conflicts with an existing claim, SUPERSEDES when it replaces a weaker-tier source. Rejected findings are wired with a REJECTED_BECAUSE edge — rejection is data, not a dead end.
7

Next-move surface or chain return

If you ran research standalone, Larry surfaces an F.1 next-move selector naming the methodologies that can now consume the wired claims. If research was dispatched by another command (e.g. /mos:build-thesis), it returns only the accepted EvidenceClaim node IDs so the caller resumes with exactly the evidence it needed.

Output / Artifacts

EvidenceClaim nodes

Each approved finding becomes a typed EvidenceClaim node in your room’s SQLite graph at room/.mindrian/room.db, with provenance: URL, retrieved_at timestamp, and evidence tier.

Typed edges

INFORMS, CONTRADICTS, SUPERSEDES, and REJECTED_BECAUSE cascade edges are written to the local graph — never to the Brain. These edges are queryable via /mos:graph.

Research artifact

Filed findings land as cited artifacts under room/**/research/, matching the section the F.1 gate resolved.

Memory events

A research_filed or research_rejected memory event is written per finding, carrying URL, retrieval time, and evidence tier for future deduplication.
Because /mos:research declares emits_evidence_claims: true in its frontmatter, other commands that declare requires_evidence: can auto-dispatch it when room evidence is thin. Larry proposes the dispatch — you decide whether to run it.

Brain Behavior

ModeWhat changes
With Brain (Tier 1)The brain lens is included in the lens set, semantically cross-referencing findings against 12,485 methodology vectors. Connections between findings and methodology are surfaced.
Without Brain (Tier 0)The brain lens is skipped with a typed skipped provider status. All other lenses run normally. Research still files typed evidence — nothing is blocked.
Zero local room data ever reaches the Brain. The only Brain touch in this command is a read-only generic methodology query. All graph writes land in your local room/.mindrian/room.db via navigation.cjs.

Recovery Modes

When an engine breaks mid-run, the run does not dead-end and does not lie. The research_mode field on every result tells you exactly what kind of run occurred:
research_modeMeaning
normalEvery engine worked. Nothing to disclose.
web_degraded_local_fallbackA live web leg failed; cached or local data covered the gap. The outage stays visible in providers.
local_onlyThe run was deliberately offline — your choice, not a failure.
insufficient_evidenceZero items anywhere. A cold corpus is a typed verdict, never a silent success with empty arrays.
llm_engine_recoveryAn engine broke and a bounded, fenced LLM recovery did the work; everything it did is disclosed.
manual_intervention_requiredThe run needs you — the message names the exact missing thing (a credential, or a Claude spend cap).

Examples

# Research a URL — surfaces a readback gate (Ingest / Ingest+Explore / Skip) before fetching
/mos:research https://example.com/market-report

# Research a topic — context-aware, files findings as typed EvidenceClaim nodes
/mos:research "competitive landscape for B2B SaaS invoicing"

# Broad preset — scholarly + industry + patent lenses at equal weight
/mos:research --broad "network effects in two-sided marketplaces"

/mos:graph

Query the typed evidence nodes and edges wired by research using natural language.

/mos:find-connections

Use Brain semantic search to surface cross-domain patterns connected to your filed evidence.

/mos:opportunities

Scan for funding opportunities informed by the room context research builds.

/mos:qualify-opportunity

Qualify a surfaced opportunity at a structured decision card after research surfaces candidates.

Build docs developers (and LLMs) love