hermes - 💡(How to fix) Fix get_model_context_length() unconditionally queries OpenRouter metadata even for non-OpenRouter users, causing false "below minimum 64K" rejections

get_model_context_length() unconditionally queries OpenRouter metadata even for non-OpenRouter users, causing false "below minimum 64K" rejections

Problem

agent/model_metadata.py::get_model_context_length() calls fetch_model_metadata() (OpenRouter /models API) as step 6 in the resolution chain, regardless of which provider the user actually configured.

This means users on Kimi Coding, Moonshot, Ollama Cloud, or any other non-OpenRouter provider can receive incorrect context_length values from OpenRouter's stale metadata — specifically, OpenRouter reports 32768 for moonshotai/kimi-k2.6 when the model actually supports 262144.

The 64K minimum-context guard in run_agent.py then rejects the model with:

ValueError: Model kimi-k2.6 has a context window of 32,768 tokens, which is below the minimum 64,000 required by Hermes Agent.

Impact

• Cron jobs fail silently (logged but not surfaced to user)
• Gateway DM sessions fail to initialize agent
• Any code path that constructs AIAgent without explicit config_context_length override is affected

Reproduction

1. Configure Hermes with provider: kimi-coding, model: kimi-k2.6, no model.context_length override
2. Run any cron job or gateway session
3. AIAgent.__init__ → ContextCompressor.__init__ → get_model_context_length() → fetch_model_metadata() returns 32768 from OpenRouter cache → ValueError

Root Cause

fetch_model_metadata() unconditionally hits https://openrouter.ai/api/v1/models even when:

• OPENROUTER_API_KEY is not set
• The active provider is not OpenRouter
• The endpoint base_url is not OpenRouter

OpenRouter's data is treated as provider-unaware fallback authority, but it can be stale or wrong for models it doesn't directly host.

Current Workaround

Commit 91eef6255 added a narrow guard: reject exactly 32768 for Kimi-family models from OpenRouter metadata and fall through to hardcoded defaults. This fixes the symptom for kimi-k2.6 but leaves the architecture issue intact.

Suggested Fix

1. Skip OpenRouter fallback when OpenRouter is not configured — if OPENROUTER_API_KEY is absent and the active provider is known non-OpenRouter, don't query OpenRouter metadata.
2. Or: Move OpenRouter fallback to the absolute end of the chain (step 9), after hardcoded defaults and local queries.
3. Or: Add a model.context_length auto-population during hermes setup so users don't silently depend on runtime probing.

Environment

• Hermes version: post-91eef6255
• Provider: kimi-coding
• Model: kimi-k2.6
• macOS 26.4.1