autoresearch

name: autoresearch description: | Autonomous iterative research protocol with mechanical quality gates. Multi-round search loops with per-round verification -- the agent does NOT decide when to stop, only the gate does. Works standalone or under Mercury dispatch. Triggers: "autoresearch", "自动研究", "深度调研", "deep research", "comprehensive research", "多轮调研".

Autoresearch Protocol

Purpose

Autonomous iterative research for comprehensive investigations. The agent does NOT decide when research is complete -- only the mechanical quality gate does. The loop terminates when all gate metrics pass, but the agent may never self-declare completion or skip the gate.

When This Applies

• researchScope is deep (Mercury dispatch)
• Research questions >= 3
• Cross-verification across >= 3 independent sources
• Architectural decision analysis
• User invokes /autoresearch or says 自动研究 / 深度调研

For lighter research, use the web-research skill instead.

Iron Rules

1. No completion claims without fresh verification evidence.
2. Every factual claim requires a source URL or an explicit UNVERIFIED tag.
3. Agent self-reports are not evidence.
4. "Should work", "probably", and "I believe" are banned.

Rationalization Prevention

Invocation & Bootstrap

/autoresearch

Optional directives:

• MAX_ROUNDS: N -- hard cap on iterations (default: 10)
• QUESTIONS: Q1; Q2; Q3 -- explicit research questions

Environment Detection (do this FIRST)

1. If Mercury_KB/04-research/ exists -> Mercury mode (use KB paths, RESULTS_FILE: results-{ISSUE_NUM}.jsonl)
2. Otherwise -> Standalone mode (use .research/reports/ and .research/state/, RESULTS_FILE: results.jsonl)

Research Manifest

On Round 1, create research-manifest.json in the state directory with topic, questions, max_rounds, started_at, mode. If no QUESTIONS provided, decompose topic into 3-7 sub-questions.

Research Loop

You are in a loop. DO NOT declare completion. DO NOT summarize prematurely. Only the mechanical quality gate can end this loop.

Round N:

1. RESTORE -- Read manifest + {RESULTS_FILE} + report
2. PLAN -- Pick 1-3 unanswered or weakest questions
3. SEARCH -- WebSearch + WebFetch, min 3 searches per question
4. WRITE -- Update report with findings, cite every claim with [URL]
5. GATE -- Run mechanical quality gate
6. LOG -- Append round JSON to {RESULTS_FILE}
7. BRANCH -- ALL PASSED -> VERIFICATION; ANY FAILED -> Round N+1

Quality Gate -- Mechanical Counting

Count (not self-assess):

1. total_questions from manifest
2. answered questions (>= 2 sentences + >= 1 URL each)
3. total_claims (declarative factual statements)
4. cited_claims (with [URL])
5. UNVERIFIED markers

ALL FOUR must pass. If any fails, continue to next round.

Results JSONL

Each round append to {RESULTS_FILE}: round, timestamp, questions_targeted, sources_found, sources_verified, question_answer_rate, citation_density, unverified_rate, iteration_depth, gate_passed, notes.

Final round adds: termination_reason, verification_verdict, verification_score.

Context Recovery

1. Read research-manifest.json
2. Read {RESULTS_FILE} -- find last round metrics
3. Identify weakest dimensions
4. Focus current round on those gaps

Verification

Step A: Mechanical Checklist (MANDATORY)

For each question confirm:

• Has dedicated section
• Section has >= 2 unique source URLs
• No unjustified UNVERIFIED claims
• Contradictions documented

Write to verification-{TOPIC}.md in state directory.

Step B: Codex Delegation (OPTIONAL)

Use native Codex delegation if available to run read-only verification:

• question coverage, citation density, actionability, risk honesty (1-5 each)
• Pass threshold: weighted average >= 4.0
• If delegation unavailable, mechanical verification is sufficient.
• Log verification_mode: mechanical_only

Termination & Output

State Externalization

Files: research-manifest.json, {RESULTS_FILE}, RESEARCH-{TOPIC}-*.md, verification-{TOPIC}.md

Mercury Integration

Auto-detected via Mercury_KB/04-research/ existence:

• Uses Mercury KB paths instead of .research/
• TaskBundle fields read from dispatch prompt
• Results JSONL uses issue number
• Receipt JSON follows SoT workflow