deep-research

Deep Research

Drift-prone skill. Provider registrations, tool names, result shapes, and quotas change. Never promise provider coverage from this file alone. Obtain a fresh MCP tools/list inventory, run the provider data probes, then route only exact tools that are both configured and live. The policy choke-point is references/research_harness/routing.py; preflight.py catches inventory, quota, and freshness drift before dispatch.

The existing spine remains unchanged:

understand -> plan -> search -> deep-read -> synthesize -> deliver

Research Harness M-16 adds exact-tool preflight, role-scoped routing, caching, academic consolidation, claim/evidence validation, bounded reflection, and a cited Markdown renderer to that spine. It is a small pure Python package, not a new MCP server or daemon.

When to Activate

• In-depth research, investigation, or current-state questions
• Competitive analysis, due diligence, market/technology evaluation
• Literature reviews and evidence synthesis
• Questions that need multiple independent sources and explicit uncertainty

Runtime MCP Tool Contract

Routing stores logical providers plus exact tool names. The configured server prefix is bound at runtime, so moving tools behind search-hub, another lazy hub, or a direct provider registration does not break policy. A same-name tool may be exposed as server__tool by one client and mcp__server__tool by another; normalize that client wrapper into the inventory boundary, not in the

Workflow

Step 1: Scope — write a research brief

Step 2: Preflight — inventory, probe, then bind

Step 3: Plan and dispatch with provider roles

Step 4: Academic path — discover, resolve, read, consolidate

Step 5: FACT-style claim/evidence table

Step 5a: Native Anthropic Citations when the execution path exposes them

Step 6: Synthesize once, in one context

Step 7: Reflect — targeted gap fill, at most two waves

Step 8: Fresh external RACE-style verifier

Step 9: Deliver and persist