deep-explore

Deep Explore Skill

Trigger

• Keywords: deep explore, large-scale research, multi-agent explore, codebase survey, parallel investigation, comprehensive research, deep dive multiple areas

When NOT to Use

Scenario	Alternative
Quick single-area lookup	/code-explore
Dual Claude+Codex confirmation	/code-investigate
Git history tracking	/git-investigate
Code review	/codex-review-fast

Workflow Overview

flowchart TD
    A[Phase 0: Intent Analysis] --> B{Files > 25?}
    B -->|No| C[Redirect to /code-explore]
    B -->|Yes| D[Wave 1: Breadth]
    D --> E[Gather + Claim Registry]
    E --> F[Wave 2: Depth]
    F --> G{Completeness Gate}
    G -->|score >= 80 + no critical Qs| H[Report]
    G -->|score < 80 OR critical Qs| I[Wave 3: Cross-cutting]
    I --> H

Phase 0: Intent Analysis

1. Parse user's research question
2. Estimate scope: Grep for related keywords → count unique files
3. Routing guard: if estimated files <= 25, redirect to /code-explore
4. Plan shards: identify 2-3 non-overlapping exploration areas

Shard Planning Strategy

Priority	Method	When
1	User-specified areas (--areas)	User knows what to split
2	Domain-based clustering	Auto-detect service/provider/hooks/rules boundaries
3	Directory-based fallback	When domain boundaries unclear

Output: ownership matrix (shard → files).

Wave 1: Breadth (Mandatory)

Fan-out 2-3 Explore agents in parallel. Each agent explores one shard.

Agent({
  description: "Wave 1 Shard A: <area description>",
  subagent_type: "Explore",
  run_in_background: true,
  prompt: <agent-prompt template from references/agent-prompt.md>
});

All agents dispatched in a single message (parallel). Wait for all to complete.

Agent Contract (80/20)

Allocation	Scope	Limit
80%	Primary assigned shard	Unlimited findings
20%	Peripheral vision (security, edge cases, cross-cutting)	Max 2 peripheral findings

Peripheral findings must be evidence-backed (file:line) and tagged: cross-cutting | security | reliability | operability.

Inter-Wave: Gather + Plan

After each wave, the orchestrator:

1. Collect all agent results
2. Build claim registry (see references/synthesis.md)
3. Rank open questions by impact × uncertainty
4. Build context packet for next wave

Context Packet (what to pass forward)

Pass	Don't Pass
Evidence-backed facts (file:line)	Prior wave conclusions as truth
Open Qs ranked by impact × uncertainty	Narrative interpretations
Do-not-repeat ledger (explored files, executed queries)	Full raw findings dump
Contradiction list	Agent opinions without evidence

Wave 2: Depth (Mandatory)

Focus on top hotspots from Wave 1, ranked by impact × uncertainty × blast_radius.

• 2-3 agents, each assigned one hotspot for deep dive
• Context packet from Wave 1 provided (facts + questions, not conclusions)
• Agents verify Wave 1 hypotheses independently

Completeness Gate

After Wave 2 (and Wave 3 if run), compute completeness score.

2-Signal Completeness Score

score = round(100 × (0.7 × (1 - novelty_rate) + 0.3 × is_zero(critical_open)))

Where:
  novelty_rate = unique_new_findings / max(1, total_valid_findings)
  critical_open = count(questions where impact=high AND uncertainty=high)
  is_zero(x) = 1 if x == 0, else 0

Zero findings → novelty_rate = 0 → score = 70 (below threshold → continue).

Stop Conditions

Condition	Action
score >= 80 AND critical_open == 0	Stop, output report
score >= 80 AND critical_open > 0	Wave 3 (if not yet run)
Wave 3 done AND score < 80	Stop with Inconclusive + next actions

Hard-Fail Overrides (force continue)

• Unanswered critical user question
• High-severity contradiction unresolved
• Evidence missing for high-impact claim

Precedence (--waves is hard ceiling)

User --waves	Hard-fail active	Behavior
--waves 2	No	Stop after Wave 2
--waves 2	Yes	Stop after Wave 2 with Inconclusive
--waves 3	No	Adaptive (stop early if score met)
--waves 3	Yes	Force Wave 3

User --waves never exceeded. Hard-fail emits Inconclusive with explanation.

Wave 3: Cross-cutting (Optional)

Triggered only when:

• Unresolved critical open questions are cross-cutting, OR
• Findings >70% concentrated in 1 subsystem, OR
• High-risk domain flags (auth/security/migration)

If no trigger conditions met, skip Wave 3 even if score < 80.

Agents in Wave 3 follow 80/20 contract. If trigger is cross-cutting, one agent may operate as conditional scout with broader mandate.

Agent Dispatch Contract

// Primary: Agent tool with subagent_type=Explore
Agent({
  description: "Wave N Shard X: <description>",
  subagent_type: "Explore",
  run_in_background: true,
  prompt: `<from references/agent-prompt.md>`
});

Fallback (if Explore dispatch fails):

1. Retry with subagent_type: "general-purpose"
2. If still fails → degrade to single-agent inline exploration
3. Mark coverage gap in report

Claim Registry (Synthesis)

See references/synthesis.md for full algorithm.

Step	Action
1. Normalize	{claim, evidence(file:line), shard, wave, confidence}
2. Dedup	Key = canonical_file_path + canonical_claim_text (±5 line tolerance)
3. Consensus	Same claim from 2+ shards → [consensus]
4. Conflict	Contradicting claims → evidence-weight resolution
5. Divergence	Unresolvable → explicit divergence section

Evidence redaction: follow @rules/logging.md — no secrets, file:line refs only.

Arguments

Argument	Description	Default
<query>	Research topic/question	Required
--agents N	Agents per wave (1-3)	3
--waves N	Max waves (2-3)	3 (adaptive)
--areas "a, b, c"	Manual shard specification	Auto-detect
--quick	Redirect to /code-explore	Off

Output

See references/synthesis.md for report template.

## Deep Exploration Report: <query>

### Completeness
- Score: <N>/100
- Waves executed: <N>
- Agents dispatched: <N>

### Executive Summary
<2-3 sentence answer to user's question>

### Per-Wave Findings
| Wave | Focus | Key Findings | Open Qs |
|------|-------|-------------|---------|

### Claim Registry
| # | Claim | Evidence | Source | Confidence | Consensus |
|---|-------|----------|--------|------------|-----------|

### Coverage Matrix
| Shard | Files Explored | Ownership |
|-------|---------------|-----------|

### Proactive Discoveries (from 20% peripheral)
| # | Finding | Tag | Evidence |
|---|---------|-----|----------|

### Divergence (if any)
| # | Claim A | Claim B | Source Agents | Status |
|---|---------|---------|--------------|--------|

### Residual Risks
- <remaining unknowns and suggested follow-up commands>

Verification Checklist

• Phase 0 correctly estimated scope and planned shards
• Wave 1 dispatched agents in parallel (single message)
• Wave 2 focused on highest-impact hotspots
• Completeness score computed and displayed
• Claim registry built with dedup and conflict resolution
• Inter-wave context did not pass conclusions as truth
• Report includes coverage matrix and residual risks
• No git add/commit/push executed

References

• Agent prompt template: references/agent-prompt.md
• Synthesis algorithm + report template: references/synthesis.md
• Standards: @rules/docs-writing.md

Examples

Input: /deep-explore "How does the review pipeline work end-to-end?"
Action: Phase 0 → Wave 1 (hooks, skills/review, rules/auto-loop) → Wave 2 (state machine, gate logic) → Report

Input: /deep-explore --areas "hooks, skills/codex-code-review, rules" "Review system architecture"
Action: Phase 0 (user shards) → Wave 1 (3 agents) → Wave 2 (hotspots) → Report

Input: /deep-explore --quick "How does review-state.js work?"
Action: Redirect to /code-explore (small scope)

Input: /deep-explore --agents 2 --waves 2 "Plugin install flow"
Action: Phase 0 → Wave 1 (2 agents) → Wave 2 (2 agents) → Report (max 2 waves)