deep-research-investigate

Deep Research: Investigate Thread (Phase 2)

Execute iterative web research for a SINGLE research thread using the Triplet Engine: [Scope -> Search -> Extract] repeated until convergence. All output goes to files; no user interaction.

Invariant Principles

1. Citation-Mandatory: Every fact MUST have a URL. Facts without URLs are UNVERIFIED. No exceptions.
2. No Fabrication: If search does not find something, say so. Never invent facts, URLs, or sources.
3. Micro-Report Atomic: One micro-report per round. The micro-report is the unit of evidence. No rounds without a micro-report.
4. Convergence-Driven: Stop when sub-questions are answered or progress has plateaued, not when a round count is reached.
5. Drift-Resistant: Every result must pass relevance checks before extraction. Off-topic results are discarded, not forced to fit.

Parameters

Parameter	Required	Purpose
thread	Yes	Thread definition (name, sub-questions, subjects, source_strategy)
brief	Yes	Research Brief with disambiguation keys and known facts
round_budget	Yes	Maximum number of search rounds
artifact_dir	Yes	Directory for micro-report output files

Execution States

[Initialize] -> [Scope] -> [Search] -> [Extract] -> [Converged?]
                  ^                                      |
                  |          NO                           |
                  +--------------------------------------+
                                    |
                                   YES
                                    |
                                    v
                              [Completion Report]

Phase 1: Initialize Thread State

Parse context and initialize tracking structures.

thread_state = {
  name: thread.name,
  sub_questions: thread.sub_questions,     # list of { id, question, status: OPEN }
  subjects: thread.subjects,               # list of { name, search_rounds: 0, status: UNCOVERED }
  source_strategy: thread.source_strategy, # SURVEY -> EXTRACT -> DIVERSIFY -> VERIFY
  strategy_phase: "SURVEY",
  round_history: [],
  known_facts: brief.known_facts,
  disambiguation: brief.disambiguation,
  current_round: 0,
  plateau_level: 0,
  converged: false
}