Skip to content

Token Management

Ankit Jha edited this page May 29, 2026 · 1 revision

Token Management

Claude Code is the orchestrator. If its context fills up, everything stops. Hydra enforces a strict tiered response to rising token usage.


Claude Token Thresholds

claude_pct Mode Actions
0–49% normal Full orchestration, all features active
50–64% compact Run /compact — recommended now
65–69% caution Run /compact URGENTLY. Stop self-reviewing agy output
70–74% warning Downgrade all tasks 1 tier. Freeze escalations
75–79% critical Routing only. Only CORE if truly necessary
80%+ emergency STOP all generation. Route to Qwen. Start new session

Rule: Run /compact proactively at 50%. Too late after 75%.


Global Model Token Rules

The 70/75/80% rule applies to all models. route.sh enforces via fallback chains:

model_token_thresholds:
  warn_at_percent:      70
  fallback_at_percent:  75
  hard_limit_percent:   80

Checking Current Usage

dispatch/route.sh --status
# Claude token usage : 52%  [mode: compact]
# Exhausted pools    : none

Token Pool Exhaustion

When a pool runs out:

agy_claude exhausted → skip T2, T3 → try T4 (Flash Thinking) → T5 (Pro) → T10 (Qwen)
agy_flash exhausted  → skip T4,T7,T8,T9 → T10 (Qwen) only

Qwen (T10) is always available — it runs locally regardless of API state.

Clone this wiki locally