Deep Think v2

Adaptive multi-phase reasoning with evidence-based depth, targeted critique→reflect cycles, convergence shortcuts, and 3-pass synthesis. Agent count scales with problem complexity (2-5).

Prerequisites

# Track A (Agent Teams) only — Track B works without this
export CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1

Execution Track Detection

Before starting, detect which track to use:

System prompt contains "Plan mode is active"?
  → Yes: Track B (Task Orchestration) — output to plan file
  → No:  Track A (Agent Teams) — output to .deep-think/06-answer/answer.md

Personas

Core Personas (always available)

Pragmatist Focus: What works in production at 3am when things break. Hard Constraint: Must include a concrete implementation sequence with time estimates. Consider: maintenance burden, onboarding, debugging at scale. Favor battle-tested over novel. Ask: "Will this still make sense in 2 years?"

First-Principles Focus: Derive from fundamentals. Question every assumption. Hard Constraint: Must question at least one widely-accepted assumption. Don't accept "best practices" — ask WHY they're best. Maybe they're not. Focus on correctness and logical soundness above all else.

Adversarial Focus: Everything will fail. Find out how. Hard Constraint: Must provide at least one specific failure scenario (no abstract concerns). Consider: malicious input, race conditions, resource exhaustion, cascading failures. Your job is to BREAK every approach. Be paranoid. Be concrete.

Optimizer Focus: Think in O(n), cache lines, memory bandwidth, network round-trips. Hard Constraint: Must include a quantitative cost model (even if estimated). Quantify EVERYTHING. Don't say "faster" — say "3x faster because..." Profile before you optimize. Know your bottlenecks.

Dynamic Persona

Domain Specialist (replaces fixed Innovator) Assigned by team lead in Phase 0 based on problem type:

• Performance → "Performance Engineer"
• Architecture → "Systems Architect"
• UX decisions → "UX Researcher"
• DB design → "Data Engineer"
• Security → "Security Engineer" Hard Constraint: Must reference at least 2 established best practices or anti-patterns from the domain.

Tier-based Persona Assignment

Architecture

You (Team Lead)
├── Phase 0: Triage & Frame (you — classify, assign tier)
├── Phase 1: Parallel Paths (2-5 agents, evidence-based depth)
│   └── Personas assigned per tier
├── Phase 2: Convergence Check (you — shortcut if aligned)
├── Phase 3a: Targeted Critique (max-disagreement pairs)
├── Phase 3b: Author Reflection (Tier 3 only)
├── Phase 4: 3-Pass Synthesis (Extract → Reconcile → Compose)
└── Phase 5: Final Answer (structured confidence)

Workflow

Phase 0: Triage & Frame (Team Lead)

Evaluate the problem on 4 dimensions (0-2 each):

Tier assignment:

Phase 0 Output:

## Problem Frame
- Question: [one sentence]
- Type: [architecture / optimization / design / debugging / knowledge]
- Dimensions: Space=[X] Tension=[X] Uncertainty=[X] Scope=[X]
- Tier: [1/2/3] (score: X/8)
- Key Constraints: [list]
- Success Criteria: [what a good answer looks like]
- Persona Assignments: [who gets which focus question]
- Domain Specialist Role: [specific role name, if Tier 1 or 3]

Pre-Mortem (Tier 3 only):

After framing, add:

## Pre-Mortem
Imagine 3 months from now, our chosen approach has failed badly.
- What went wrong? (3 scenarios)
- What early warning signs should we watch for?
- Which assumptions, if wrong, would invalidate the entire approach?

Pre-mortem scenarios become mandatory evaluation criteria for all paths.

Initialize workspace (Track A only):

python scripts/deep_think.py init "your question" -c tier2 -w .deep-think

Phase 1: Parallel Paths

Track A (Agent Teams):

Create agent team "deep-think".

Each assigned persona:
- Read .deep-think/00-triage/frame.md
- Write solution to .deep-think/03-paths/path-{name}.md
- Follow the structured path template (see phase-templates.md)
- When done, message team-lead with Core Thesis + key differentiator

Track B (Task Orchestration):

For each assigned persona, launch a Task agent:
- Task(name="{persona}", prompt="[problem frame] + [persona instructions] + [path template]")
- Collect returned text from each Task

Path Structure (replaces word count minimum):

Each path MUST include these sections:

1. Core Thesis (1-2 sentences)
2. Evidence Chain (each claim tagged: [CODE] [BENCH] [PATTERN] [REASON] [ASSUME])
3. Implementation Sequence (dependency-ordered steps)
4. Risk Register (table: risk / likelihood / impact / mitigation)
5. What This Path Uniquely Offers (1-2 things other paths won't surface)

[ASSUME]-tagged claims are treated as hypotheses in synthesis. They appear as unverified assumptions in the final answer.

Phase 2: Convergence Check (Team Lead)

After all paths complete, read each path's Core Thesis.

If (N-1) or more of N paths converge on the same core recommendation:
  → Skip challenge round
  → Proceed to synthesis with note: "Strong convergence detected — high confidence in common recommendation"
  → Record minority opinion in "Dissenting Views"
Otherwise:
  → Proceed to Phase 3a

Expected effect: 40-50% time savings on strong models (Opus 4.6) for convergent problems.

Phase 3a: Targeted Critique

Assignment: max-disagreement pairs (NOT round-robin). Team lead assigns each critic to the path they most disagree with.

Critique quality rules: All critiques MUST include specific scenarios. Critiques without concrete scenarios are ignored in synthesis.

Weak: "This approach may have performance issues at scale"
Strong: "With 1,650 cells + 30 columns, the O(n*m²) algorithm executes
        49,500 array traversals per render, causing ~110ms interaction delay"

Critique rating scale (3-level):

• SOUND: No significant gaps found
• HAS-GAPS: Specific issues identified but approach is viable with fixes
• FUNDAMENTALLY-FLAWED: Core thesis is broken

Targeted critique question examples:

• First-Principles → Pragmatist: "What practical constraints does this path underestimate?"
• Adversarial → Optimizer: "In what specific production scenario does this optimization fail?"
• Pragmatist → Adversarial: "Which of these risk assessments confuse actual probability with worst-case?"

Track A: Each critic reads the target path file and writes to .deep-think/03.5-challenges/. Track B: Team lead passes the target path's full text into the critic Task's prompt.

Phase 3b: Author Reflection (Tier 3 only)

After receiving critique, each path author:

• Accepted points → revised approach
• Rejected points → counter-argument with evidence
• Updated self-assessment: SOUND / HAS-GAPS

Track A: Send critique via message, author writes path-{name}-reflected.md. Track B: Launch new Task with original path + critique in prompt, collect reflected version.

Phase 4: 3-Pass Synthesis

Replaces weighted accumulation with distillation-based synthesis (PENCIL-inspired):

Pass 1: Extract From each path, extract ONLY:

• (a) Core thesis
• (b) Unique contributions
• (c) HIGH-confidence recommendations Discard everything else.

Pass 2: Reconcile Using the disagreement map:

• Factual disagreements → resolve by evidence
• Preference disagreements → present as explicit trade-offs (with recommended choice)
• Scope disagreements → "depends on context" + context guide for each

Pass 3: Compose Write the final answer fresh from reconciled materials. No text copying from original paths. This forces genuine integration.

Blind Spot Check (all tiers): Before finalizing, ask: "What question did NO path address that a domain expert would ask?"

Phase 5: Final Answer

Write the final answer with structured confidence (see phase-templates.md for full template).

Structured Confidence (replaces X/10):

## Confidence Assessment

### Convergence: [Y of N paths recommended this approach]
### Evidence Quality: HIGH (code-based) / MEDIUM (pattern-based) / LOW (reasoning only)
### Risk Level: [from adversarial analysis]

### High-Confidence Recommendations (all paths agree, evidence-backed)
- ...

### Medium-Confidence Recommendations (majority agree, some caveats)
- ...

### Low-Confidence Recommendations (exploratory, needs validation)
- ...

Shutdown

Track A:

Shutdown the deep-think team. Wait for all teammates to finish.

Then generate report:

python scripts/deep_think.py report -w .deep-think

Track B: Write final plan to plan file → ExitPlanMode for user approval.

Tier Workflows Summary

Tier 1 (집중, ~60% of use)

Phase 0: Triage & Frame → Tier 1
Phase 1: 2 agents parallel paths
Phase 2: Convergence check → if aligned, skip to synthesis
Phase 3a: Mutual critique ×1 (if not converged)
Phase 4: 3-pass synthesis
Phase 5: Final answer
Agents: 2 (+ team lead)

Tier 2 (표준, ~30% of use)

Phase 0: Triage & Frame → Tier 2, focus questions assigned
Phase 1: 3 agents parallel paths (structured requirements)
Phase 2: Convergence check
Phase 3a: Targeted critique (specific scenarios required)
Phase 4: 3-pass synthesis + blind spot check
Phase 5: Final answer (structured confidence)
Agents: 3 (+ team lead)

Tier 3 (복합, ~10% of use)

Phase 0: Triage & Frame + Pre-Mortem → Tier 3
Phase 1: 4-5 agents parallel paths (evidence anchoring)
Phase 2: Convergence check
Phase 3a: Targeted critique (specific scenarios required)
Phase 3b: Author reflection (accept/rebut critique)
Phase 4: 3-pass synthesis + blind spot check
Phase 5: Final answer (structured confidence + disagreement tracking)
Agents: 4-5 (+ team lead)

Output Structure

Track A (Agent Teams)

.deep-think/
├── 00-triage/frame.md                  # Problem frame + tier assignment
├── 01-analysis/analysis.md             # Detailed analysis
├── 02-decomposition/decomposition.md   # Sub-problems
├── 03-paths/
│   ├── path-pragmatist.md              # Structured evidence-based paths
│   ├── path-first-principles.md
│   ├── path-{persona}.md
│   └── path-{name}-reflected.md        # Tier 3: post-critique reflection
├── 03.5-challenges/
│   ├── critique-{from}-vs-{to}.md      # Targeted critiques
│   └── ...
├── 04-synthesis/
│   ├── pass1-extract.md                # Extracted elements
│   ├── pass2-reconcile.md              # Disagreement resolution
│   └── pass3-compose.md                # Fresh composition
├── 05-answer/answer.md                 # Final polished answer
├── session.json
└── REPORT.md

Track B (Plan Mode)

Output goes to plan file:

# [Topic] Implementation Plan

## Context
[Deep-think analysis summary: why this change is needed]
[How many personas participated, convergence/divergence status]

## Approach
[Recommended approach from synthesis]
[Which persona's perspective was adopted and why]

## Changes
| # | Change | Files |
[Concrete implementation steps]

## Confidence Assessment
[Structured confidence]

## Dissenting Views
[Minority opinions]

## Verification
[Test/validation methods]

→ Then ExitPlanMode for user approval.

Troubleshooting

Paths lacking evidence tags? → Message: "Tag each claim with [CODE], [BENCH], [PATTERN], [REASON], or [ASSUME]."

Critiques too vague? → Message: "Provide a specific scenario with numbers. Vague critiques are ignored in synthesis."

Convergence check unclear? → Compare Core Thesis sentences. If they recommend the same fundamental approach (even with different details), that's convergence.

Tier seems wrong? → Re-evaluate dimensions. When in doubt, tier up (Tier 2 → Tier 3) rather than down.