CLI Tools Integration Summary

How the context management CLI tools integrate with the Tachikoma orchestration system.

---

Architecture

User Request
    ↓
Tachikoma (Primary Agent)
    ↓
Intent Classification (intent-classifier skill)
    ↓
Context Loading ← USES: context-manager skill
    ↓
[IF research intent] ← USES: context7 skill
    ↓
Route to specialist (skill/subagent)
    ↓
Execute with context
    ↓
Synthesize results

---

Tools Overview

1. Context Manager (.opencode/skills/context-manager/)

Purpose: Zero-token context operations

When Tachikoma uses it:

• Before loading context → discover (verify files exist)
• Quick lookups → extract (0 tokens vs 300+ for LLM)
• System health → status
• Maintenance → cleanup

Example invocation:

# Before routing
tachikoma$ bun run .opencode/skills/context-manager/router.ts discover

# During Q&A
tachikoma$ bun run .opencode/skills/context-manager/router.ts extract coding-standards "naming"

---

2. Context7 (.opencode/skills/context7/)

Purpose: Live documentation for research intent

When Tachikoma uses it:

• Research intent detected (confidence > 0.8)
• Training data likely outdated (React, Next.js, etc.)
• User asks about library APIs

Example invocation:

# During research processing
tachikoma$ bun run .opencode/skills/context7/router.ts quick "React" "hooks"

# Result: Current React 19 docs (not React 18 from training)

---

Token Savings

Operation	Without CLI	With CLI	Savings
Context discovery	500 tokens	0 tokens	100%
Info extraction	300 tokens	0 tokens	100%
Live docs fetch	400 tokens	API call	Faster
Per session	~1200 tokens	~0 tokens	1200+ tokens

---

Routing Integration

intent-routes.yaml

routes:
  debug:
    confidence_threshold: 0.7
    context_modules:
      - core-contract
      - coding-standards
    skill: code-agent
    # Tachikoma runs: router.sh discover before loading

  research:
    confidence_threshold: 0.7
    context_modules:
      - core-contract
      - research-methods
    skill: research-agent
    # Tachikoma runs: router.sh quick "{lib}" "{topic}" for external docs
    tools:
      - Read
      - Bash # For CLI tools

---

Workflow Examples

Example 1: Debug Intent

User: "Fix the typo in README"

Tachikoma:
  1. Classify: intent=debug, confidence=0.95
  2. Context Manager: discover
     → bash router.sh discover
     → Verifies context files exist
  3. Load: core-contract, coding-standards
  4. Route: code-agent skill
  5. Execute: Fix typo

Token cost: 0 (discovery via CLI)

Example 2: Research Intent

User: "How do React Server Components work?"

Tachikoma:
  1. Classify: intent=research, confidence=0.91
  2. Context Manager: discover
     → Lists available context
  3. Context7: quick "React" "Server Components"
     → Fetches React 19 docs
     → Saves to .tmp/external-context/
  4. Load: core-contract, research-methods, + fetched docs
  5. Route: research-agent skill
  6. Execute: Explain with current examples

Token cost: 0 for discovery/fetch (CLI tools)

Example 3: Harvesting

User: "Save this analysis"

Tachikoma:
  1. Context Manager: harvest ANALYSIS.md
     → Creates .opencode/context-modules/40-analysis.md
  2. Updates intent-routes.yaml (optional)
  3. Confirms: Available for future use

---

File Structure

.opencode/
├── agents/
│   └── tachikoma.md          # Orchestrator (invokes CLI tools)
├── skills/
│   ├── context-manager/      # Context operations
│   │   ├── SKILL.md
│   │   └── router.sh
│   ├── context7/             # Live documentation
│   │   ├── SKILL.md
│   │   └── router.sh
│   └── intent-classifier/    # Routing decisions
├── context-modules/
│   └── *.md                  # Context modules (00-core-contract, etc.)
└── config/
    └── intent-routes.yaml    # Routing configuration

---

Agent-Only Design

No slash commands needed - Tachikoma invokes these automatically:

• User asks natural questions
• Tachikoma classifies intent
• Tachikoma runs CLI tools internally
• Results presented seamlessly

Why this works:

• Zero cognitive load for users
• Consistent with OpenCode patterns
• Maximum token efficiency
• No maintenance of command wrappers

---

Summary

The CLI tools are agent-facing infrastructure:

1. Tachikoma discovers context before loading (0 tokens)
2. Tachikoma extracts specific info when needed (0 tokens)
3. Tachikoma fetches live docs for research (current data)
4. Tachikoma harvests temporary files to permanent context

Users never run these directly - they just ask questions naturally, and Tachikoma uses the most efficient method automatically.