Token Economy — Efficiency Layer

"21 个 skill 不必全跑。L1 改个变量名不需要双脑审查。"

Token Budget by Task Tier

Tier	Budget	Skill Activation	Model
L1: Trivial	500 tok	None. Direct execution.	Haiku subagent
L2: Simple	2000 tok	reasoning-lora only (1-pass)	Haiku subagent
L3: Complex	5000 tok	thinking-dimensions → reasoning-lora	Full model
L4: Creative	10000 tok	divergent → thinking → reasoning	Full model
L5: High-Risk	20000 tok	Full chain including failure-predictor + second-brain	Full model

Six Savings Mechanisms

1. Model Routing (biggest win — 50-70% savings)

L1-L2 tasks → Haiku subagent instead of full model
Haiku costs 70% less than Opus, 50% less than Sonnet
Quality impact: zero for L1-L2 tasks (simple, well-understood)

2. Skill Chain Compression

L2 task "add error handling to this function":
  Full chain: divergent → thinking-dims → reasoning → execute (5000 tok)
  Compressed: reasoning-lora 1-pass → execute (1500 tok, 70% saved)

3. Context Pruning

When context > 60% of window:
  → Trim conversation history older than 10 turns
  → Compress old tool outputs to 1-line summaries
  → Drop files no longer relevant to current task

4. Response Compression

L1: "Done. Renamed X to Y in file.ts."
L2: "Fixed. Added try/catch. Tests pass."
L3+: Full recap per compact.md #11

5. Verification Tiering

L1: No verification (trust the simple change)
L2: precommit-pipeline only (skip second-brain)
L3: precommit-pipeline + second-brain (skip skill-judge)
L5: Full verification chain

6. Cache Awareness

Same file read 3+ times in one session → cache in context, don't re-read
Same search query → use cached results
Same skill chain for same task type → reuse without re-routing

Token Tracking

Log every task's token consumption:

token_log(task_type, tier, tokens_used, skills_activated, model_used, timestamp)

Weekly report:

=== Token Economy Report ===
Total saved: 45,000 tokens (38% of baseline)
L1-L2 Haiku routing: saved 28,000 tokens
Chain compression: saved 10,000 tokens
Context pruning: saved 7,000 tokens
Most expensive task type: architecture (avg 8,200 tok)
Most efficient: typo fix (avg 120 tok)

Override

User can say "full analysis" or "deep dive" → bypass token economy, run full chain on any tier. Token economy is the default, not a gatekeeper.

Integration

• neural-orchestrator: Token economy sits BETWEEN orchestrator routing and skill execution
• brain-intuition: Intuition + token economy = fastest path (L3 instinct skips everything)
• brain-self-model: Self-model feeds "I'm slow at X" → token economy pre-loads context for X
• delegation-lora: Token economy's model routing aligns with delegation (Haiku for simple)