Summary

Add Step 0c to get_model_context_length() that reads per-provider context_length overrides from providers.<name>.context_length in config.yaml.

Problem

Custom endpoints (e.g. llama-server) rarely expose context_length via /models probing, causing Hermes to fall back to artificially low values (e.g. 32K). This blocks cron jobs and any scenario where the model's true context window (e.g. 64K) can't be auto-detected.

The existing Step 0b (custom_providers per-model override) requires base_url matching, which is unavailable for cron jobs — their base_url is resolved later in the AIAgent init sequence.

Solution

Step 0c reads providers.<name>.context_length directly from config.yaml, independently of the custom_providers list. It also handles the custom provider alias — when the full provider name (e.g. custom:qwj-local) gets normalized to bare custom during AIAgent init, it falls back to prefix-matching providers.custom:* keys.

Test Plan

• Direct function test: get_model_context_length() with provider=custom:qwj-local and config_context_length=None returns 65536 (from providers.custom:qwj-local.context_length: 65536)
• Cron job test: previously failing with 32K error, now runs successfully with Qwen 27B (64K context)
• Non-custom providers unaffected: DeepSeek continues to auto-detect 1M context
• Fallback intact: if no provider-level config, falls through to existing probing chain

Configuration Example

providers:
  custom:qwj-local:
    base_url: http://192.168.0.68:18081/v1
    api_key: dummy
    context_length: 65536