Deep Research

Announce: "Using deep-research skill for multi-threaded investigation with verification."

Invariant Principles

1. Tag Every Claim: No finding enters the report without a confidence level and source URL
2. Surface Every Conflict: When sources disagree, document both positions
3. Respect the User's Frame: When research contradicts user-provided facts, STOP and present conflict via AskUserQuestion. Never silently override.
4. Verify Before Synthesizing: All findings pass through fact-checking and dehallucination

Inputs/Outputs

Artifacts at ~/.local/spellbook/docs/<project-encoded>/research-<topic-slug>/: research-brief.md, research-plan.md, micro-reports/, verified-claims.md, research-report.md

Registries

Subject Registry: Track all named entities from request. Each must get >= 1 investigation round. If any subject has 0 rounds after 50% of budget, FORCE a dedicated round.

Conflict Register: Log when sources disagree {claim, source_a, source_b, status: OPEN|RESOLVED|FLAGGED}. All must be RESOLVED or FLAGGED before Phase 4. Choosing one side without citation is FORBIDDEN.

Confidence Tags: VERIFIED (primary source URL) | CORROBORATED (2+ independent) | PLAUSIBLE (consistent, unconfirmed) | INFERRED (derived logically) | UNVERIFIED (no source) | CONTESTED (sources disagree)

Plateau Breaker: URL overlap >= 60% or 0 new facts for 2 rounds triggers: L1 query reformulation, L2 source type change, L3 STOP and report gaps. Hard limit: 3 stale rounds = mandatory L3.

Phases

Phase 0: Interview

What is the user actually asking? What named entities appear? What do they already know?

Execute: /deep-research-interview with user's request and constraints. Output: research-brief.md with refined question, subject registry, success criteria, depth. Gate: All subjects registered, research type classified, brief written.

Phase 1: Plan

Execute: /deep-research-plan with research brief. Output: research-plan.md with thread definitions, source strategies, round budgets. Gate: Threads independent, all subjects assigned, convergence criteria set.

Phase 2: Investigate (Parallel)

Threads independent? Each subagent has complete context? CURRENT_AGENT_TYPE set?

Dispatch one subagent per thread:

Task(description="Investigate: <thread>", subagent_type=CURRENT_AGENT_TYPE,
  prompt="Execute /deep-research-investigate. Thread: <def>. Budget: <N>.
  Brief: <summary>. Write micro-reports to <path>. Apply confidence tags,
  conflict register, plateau breaker.")

Gate: All threads returned, every subject has >= 1 round, conflicts consolidated.

Phase 3: Verify

Dispatch fact-checking subagent on micro-reports/*.md with research-adapted extraction (SourceCredibility, CrossReference, DateValidity agents). Then dispatch dehallucination on verified-claims.md focusing on precision fabrication and source conflation.

Gate: All claims have verdicts, no REFUTED presented as fact, dehallucination passed.

Phase 4: Synthesize

Assembly: reorder to reader-logical order, apply confidence tags inline, build bibliography, insert FLAGGED conflicts with both positions.

Completeness check against research-brief.success_criteria. If gaps: dispatch targeted Phase 2 (max 1 loop) or acknowledge gaps.

Gate: Success criteria addressed, all subjects in report, bibliography complete.

Circuit Breakers

<FINAL_EMPHASIS> Research is only as valuable as its honesty. Tag uncertainty. Surface conflicts. Acknowledge gaps. Fabrication is unrecoverable. Honest incompleteness is always preferable. </FINAL_EMPHASIS>