heavy-thinking-manual

name	heavy-thinking-manual
description	Heavy Thinking Manual — K=8 parallel reasoning chains for pre-implementation analysis. Triggered by: fix bug, build feature, ideation, create spec. Auto-loads context from Hermes Memory, Session History, and Project Files. Outputs structured analysis to .task-context/{task-id}/.

---

Heavy Thinking Manual

Mission

Apply Heavy Thinking (K=8 parallel reasoning chains) to analyze tasks BEFORE implementation. Ensure context sufficiency, identify core problems, and prepare enriched input for execution.

Trigger: Khi user mention keywords như fix bug, build feature, lên ý tưởng, tạo spec.

Goal: Không nhảy vào implement cho đến khi analysis hoàn thành.

---

Boot Sequence

1. Đọc knowledge/k-chains-lens.md — 8 lens taxonomy
2. Đọc data/trigger-keywords.yaml — trigger keywords config
3. Xác định task type từ trigger keyword
4. Proceed to Phase 1

---

Workflow Phases

Phase 1: Context Loading

Mục tiêu: Load tất cả context sources và audit gaps.

Thực hiện:

1. Load Hermes Memory

   memory(action='list')
   

   • User preferences
   • Past session summaries
   • Tool configurations

2. Load Session History

   session_search(query="", limit=20)
   

   • Current conversation
   • Recent commands
   • Pending tasks

3. Load Project Files

   • AGENTS.md / CLAUDE.md
   • Relevant source files
   • Configurations

4. Context Audit

   • Report loaded sources
   • Flag missing sources
   • Ask user confirmation if critical sources missing

Output: context-sources.json

---

Phase 2: Chain Execution (K=8)

Mục tiêu: Spawn 8 parallel reasoning chains, each analyzing from different lens.

Thực hiện:

1. Determine Chain Configuration

   • Default: K=8 chains (1-8)
   • Based on task type, may add extra chains

2. Spawn Chains (Primary: opencode-go/deepseek-v4-flash, Fallback: delegate_task)

   Using opencode-go:

   opencode --model opencode-go/deepseek-v4-flash run "Analyze task from [lens] perspective. Context: [enriched_context]. Output format: [chain_format]"
   

   Using delegate_task (fallback):

   delegate_task(
     goal="Analyze task from [lens] perspective",
     context="[enriched_context]",
     toolsets=["terminal", "file"]
   )
   

3. Chain Isolation (CRITICAL)

   • Each chain is INDEPENDENT
   • No cross-chain visibility
   • Only Deliberator sees all chains
   • See knowledge/chain-isolation-rules.md

8 Chain Lenses:

Chain	Lens	Focus
1	Context & State	Input sufficiency, cached vs fresh
2	Handoff & Contract	Session/agent/skill boundaries
3	Error Handling	Hallucination, silent failures
4	Propagation	Codebase impact, side effects
5	Quality Assurance	Verification, diff, metrics
6	Risk Assessment	Failure modes, risk mitigation
7	Alternative Paths	Other valid approaches
8	Dependency Analysis	External dependencies

Task-Type Augmentation:

Task Type	Extra Chain	Focus
fix bug	Root Cause	Bug reproduction
build feature	Impact	Scope analysis
ideation	Alternatives	Idea validation
spec	Dependencies	Complete spec

Output: analysis-chains.md (combined findings)

---

Phase 3: Deliberation

Mục tiêu: Synthesize chain findings into CORE problems and CASE recommendations.

Thực hiện:

1. Findings Aggregation

   • Collect all findings
   • Deduplicate
   • Categorize by theme

2. Consensus Identification

   • Strong (≥7/8 agree)
   • Moderate (5-6/8 agree)
   • Weak (3-4/8 agree)
   • None (<3 agree)

3. Conflict Resolution

   • Apply logic, not voting
   • Document minority positions
   • Flag unresolved for human review

4. Synthesis

   • Identify 3-5 CORE problems
   • Apply CASE System:
     • PREVENT: Context enrichment, acceptance criteria
     • DETECT: Verification gates, quality metrics
     • RECOVER: Rollback, escalation

Output: deliberation.md + prepared-context.json

---

Phase 4: Output Generation

Mục tiêu: Write all output files to .task-context/{task-id}/.

Thực hiện:

1. Create Task Directory

   task_id = "{date}-{trigger-keyword}-{uuid-short}"
   # e.g., 2026-05-11-fix-bug-a1b2c3
   

2. Write Output Files

   • task-meta.yaml — Task metadata
   • context-sources.json — Context audit
   • analysis-chains.md — Chain findings
   • deliberation.md — Synthesis
   • prepared-context.json — Enriched input
   • diagrams.mmd — Mermaid diagrams
   • checklist.yaml — Verification checklist

3. Quality Gate Review

   • Run through loop/quality-gate.md
   • Verify all schemas validate
   • Flag any issues

---

Phase 5: Delivery

Mục tiêu: Present results to user and wait for directive.

Thực hiện:

1. Present Summary

   • Task ID and trigger
   • Context sources loaded
   • CORE problems identified
   • Prepared context ready

2. Interaction Points:

   • IP1: Context loaded → "Found X sources. Missing: Y. Continue?"
   • IP2: Chains complete → Show deliberation summary
   • IP3: Quality gate fail → Report failed items
   • IP4: Task too broad → Request scope clarification
   • IP5: Output complete → "Proceed to implementation?"

3. Wait for User Directive

   • Proceed to implementation
   • Request changes
   • Cancel task

---

Task ID Pattern

{date}-{trigger-keyword}-{uuid-short}

Examples:

• 2026-05-11-fix-bug-a1b2c3
• 2026-05-11-build-feature-d4e5f6
• 2026-05-11-ideation-g7h8i9

---

Output Directory Structure

.task-context/{task-id}/
├── task-meta.yaml          # Task metadata
├── context-sources.json    # Context audit (loaded + missing)
├── analysis-chains.md      # K=8 chain findings
├── deliberation.md         # Synthesis + CORE problems
├── prepared-context.json   # Enriched input for implementation
├── diagrams.mmd            # Mermaid diagrams
└── checklist.yaml          # Verification checklist

---

Guardrails

ID	Rule	Description
G1	No Premature Implementation	Must complete all 5 phases before any code/write
G2	Context Audit Required	Must report loaded + missing sources
G3	Chain Isolation	Chains must not see each other
G4	Schema Validation	Output must validate against schemas
G5	User Confirmation	Must wait for user at each interaction point

---

Anti-Pitfalls

1. Jumping to Solution: Analyze first, implement second
2. Context Assumption: Don't assume context is sufficient
3. Chain Contamination: Keep chains independent
4. Incomplete Synthesis: Deliberate over ALL chains
5. Format Drift: Follow schemas exactly

---

Related Knowledge

• knowledge/k-chains-lens.md — 8 lens definitions
• knowledge/context-sources.md — Context loading guide
• knowledge/deliberation-process.md — Synthesis methodology
• knowledge/chain-isolation-rules.md — Anti-contamination rules
• loop/quality-gate.md — Verification checklist
• loop/chain-isolation-enforcer.md — Isolation enforcement

---

Execution Example

# Trigger detected: user says "fix bug in auth"
task_id = "2026-05-11-fix-bug-a1b2c3"

# Phase 1: Load Context
context = load_all_context()
# → memory + session + project files

# Phase 2: Spawn K=8 Chains
chains = spawn_k_chains(task_id, context)
# → 8 parallel analyses

# Phase 3: Deliberation
deliberation = deliberate(chains)
# → CORE problems + CASE recommendations

# Phase 4: Write Output
write_output_files(task_id, context, chains, deliberation)
# → .task-context/{task-id}/

# Phase 5: Present & Wait
present_to_user(task_id)
# → User approves or requests changes