hermes - 💡(How to fix) Fix [Feature Request] Smart reasoning_effort routing based on task complexity [1 participants]

Problem

reasoning_effort is currently a static configuration value applied uniformly to all tasks. However, task complexity varies dramatically:

• Low complexity (single-file patch, formatting, translation): medium reasoning depth is overkill, wasting time and tokens
• High complexity (multi-file refactors, architecture design, research): medium is insufficient, causing excessive iteration

Real-world example: a user reported that changing one HTML image width (a 2-second task) took nearly 2 hours because reasoning_effort: medium and max_turns: 60 were the defaults.

See: https://mp.weixin.qq.com/s/JZ0oCCPmeJp27EdEMTid0Q

Proposed Solutions (3 tiers, increasing complexity)

Option A — Zero code: Documentation (lowest cost)

Explicitly document the recommended reasoning_effort settings per task type, and promote the existing /reasoning CLI command for dynamic adjustment.

Option B — Medium cost: Smart Routing extension (recommended)

Extend the existing smart_model_routing mechanism to also handle reasoning_effort:

smart_reasoning_routing:
  enabled: true
  rules:
    - match: ["patch", "edit", "single file"]
      reasoning_effort: low
    - match: ["search", "research", "analyze"]
      reasoning_effort: high
  default: medium

Option C — Full solution: LLM-driven auto-classification

Add a lightweight classification step at task intake (sub-100ms) using a smaller model to determine complexity and inject the appropriate reasoning_effort hint into the system prompt automatically.

Current workaround (already exists but underutilized)

/reasoning low   # for simple tasks
/reasoning high  # for complex tasks
hermes chat -t terminal,file   # restrict tools for file editing tasks