Architecture

┌─────────────┐
│  CLI tool   │  (Claude Code, Codex, OpenClaw, Cursor, Cline...)
└──────┬──────┘
       │ http://localhost:20128/v1
       ↓
┌─────────────────────────────────────────────┐
│         9RouterX (Smart Router)           │
│  • RTK Token Saver (cut tool_result tokens) │
│  • Format translation (OpenAI ↔ Claude)     │
│  • Quota tracking                           │
│  • Auto token refresh                       │
└──────┬──────────────────────────────────────┘
       │
       ├─→ [Tier 1: SUBSCRIPTION] Claude Code, Codex, GitHub Copilot
       │   ↓ quota exhausted
       ├─→ [Tier 2: CHEAP] GLM ($0.6/1M), MiniMax ($0.2/1M)
       │   ↓ budget limit
       └─→ [Tier 3: FREE] Kiro, OpenCode Free, Vertex ($300 credits)

RTK Token Saver

RTK compresses tool_result content before it reaches the model, saving 20-40% of tokens. The guards it keeps:

• A failed command's output is never compressed — the agent comes back to read it again.
• Tool names are matched case-insensitively (read = Read), the same rule tokensaver.protectedTools uses.
• A JSON document is left intact.
• The filter is chosen by the command that produced the output, not by its content.
• Truncation follows upstream's caps (RAW_CAP = 10MB, smart truncation only engages once output spans at least 250 lines).

Combo strategies

Each combo carries its own strategy:

Design decisions that matter

• Concurrency slots are per attempt. In a combo/fallback, one request can try several providers. A slot must be released when that attempt ends (success, failure, or failover), not once at the end of the request — otherwise slots leak permanently and the node locks itself out.
• A congestion violation writes no cooldown. One burst over a limit must not globally park a healthy node. Only a debug trace is written.
• One source of truth for policy. Rate limits, budgets, and the model gate all read the same resolver — no limit logic scattered around.

Further reading

Full internals: ARCHITECTURE.md.