model-routing（模型路由）

Model Routing Skill (Compact)

Automatically route each request to the best model.

Routing rules:

• Use gpt-5-codex for coding, debugging, build errors, stack traces, logs, repo edits, and implementation tasks.
• Use gpt-5 for deep reasoning, architecture, large-context analysis, complex planning, and hard diagnosis.
• Use gpt-4.1-nano for short trivial requests with no code and no deep reasoning.
• Use gpt-5-mini for everything else.

Priority:

1. code_heavy
2. heavy_reasoning
3. simple
4. normal

Fallbacks:

• gpt-5-codex -> gpt-5 -> gpt-5-mini -> gpt-4.1-nano
• gpt-5 -> gpt-5-mini -> gpt-4.1-nano
• gpt-5-mini -> gpt-4.1-nano

Retry on:

• rate limit
• TPM exceeded
• timeout
• temporary unavailable
• capacity errors

Concurrency:

• max 2 concurrent requests
• reduce to 1 for large prompts or repeated TPM failures
• queue excess work

Respect explicit user overrides and any org/channel/session model locks first. Do not override a forced model. Only auto-route when routing is enabled and no explicit model is set.

Safety caps (defaults, may be tuned by admin):

• Max input tokens per class: simple 4k, general 8k, code 12k, heavy 20k
• Max output tokens per class: simple 400, general 800, code 1200, heavy 1500
• Per-turn budget ceiling: $0.02 (downgrade on breach)

Telemetry (lightweight): record {taskType, chosenModel, capsApplied} to internal trace; do not expose to users unless they ask.

Fallbacks: never upgrade to a more expensive model on retry; only downgrade tiers.