deep-research-workflow

name	deep-research-workflow
description	Run contract-governed deep research with parallel investigators, an adversarial verifier, and a cited synthesis. Use for high-stakes or broad research requests that need source independence, conflict detection, and explicit evidence quality. Triggers on deep research, comprehensive research, investigate thoroughly, 深度调研, 深入研究, 全面调查.
license	Proprietary. LICENSE.txt has complete terms
compatibility	Uses deep_research, subagents_orchestrate, and sessions_spawn. Works on macOS, Linux, and Windows.

Contract-Governed Deep Research

Use this workflow when a normal research pass is not enough. The workflow has four explicit stages and may only claim completion after the verifier contract passes.

1. Frame and decompose

Write down:

• the decision or question;
• the time boundary and relevant geography or product scope;
• 3–6 non-overlapping sub-questions;
• claims that would materially change the answer;
• acceptable evidence and exclusions.

Do not delegate vague copies of the full question. Each work package must own a distinct sub-question.

2. Gather in parallel

Call subagents_orchestrate with one researcher work package per sub-question. Set wait: true. Every work package must:

1. use deep_research with multiple focused queries;
2. prefer primary, current sources;
3. return a claim ledger with claim, source URL, source date, authority, and whether the source is primary;
4. identify disagreement, missing evidence, and circular sourcing;
5. include these exit criteria:

[
  {"type":"content_contains","text":"Sources"},
  {"type":"content_contains","text":"Caveats"}
]

Ask each investigator to cite stable URLs, not search-result pages. A failed worker contract is evidence of incomplete research, not a reason to silently drop that sub-question.

3. Verify adversarially

After fan-in, call sessions_spawn with specialist researcher, include the complete investigator outputs as context, and ask it to disprove the draft conclusions. The verifier must:

• map every material claim to at least one cited source;
• require two independent domains for claims marked high confidence, unless a single authoritative primary source is controlling;
• flag duplicated domains and sources that merely repeat the same origin;
• identify conflicts and select the controlling source using authority, directness, and recency;
• report unsupported claims and uncovered sub-questions;
• calculate:
  • citation coverage = supported material claims / material claims;
  • source independence = distinct controlling domains / controlling sources;
  • conflict disposition = resolved conflicts / detected conflicts;
• finish with VERDICT: PASS or VERDICT: FAIL.

Use this verifier exit contract:

[
  {"type":"content_contains","text":"VERDICT:"},
  {"type":"content_contains","text":"citation coverage"},
  {"type":"content_contains","text":"unsupported claims"}
]

Do not synthesize a confident final answer after VERDICT: FAIL. Repair the named gaps with one targeted gather-and-verify retry. If it still fails, return the partial result and say exactly what remains unverified.

4. Synthesize

The final response must contain:

1. the bottom line;
2. findings organized around the original decision, not around worker names;
3. inline citations for every material factual claim;
4. explicit disagreement and uncertainty;
5. a compact evidence-quality footer:

Evidence quality
- Citation coverage: <supported>/<material>
- Independent controlling domains: <count>
- Conflicts: <resolved>/<detected>
- Verifier: PASS|FAIL

Never present the number of search hits or agent agreement as corroboration. Independence is measured at the controlling source/domain level.