hermes - 💡(How to fix) Fix feat(config): support per-model max_tokens in custom_providers config [1 participants]

Problem

When using models with large output limits (e.g., DeepSeek V4 with 384K max_output_tokens), there is currently no way to configure max_tokens per-provider or per-model in config.yaml.

The custom_providers section already supports context_length per-model, but max_tokens is not read from config at all. The AIAgent.__init__ only accepts max_tokens as a direct parameter, and neither cli.py nor gateway/run.py pass any config-based max_tokens value through.

This forces users who want to take advantage of large output limits (DeepSeek V4 384K, Gemini 2M, etc.) to either:

1. Accept the API server's default (which may be far lower than the model's capability)
2. Hardcode a global max_tokens that doesn't work across providers

Technical Details

Resolution chain currently works for context_length:

• config.yaml -> model.<provider_name>.models.<model_name>.context_length
• run_agent.py reads from custom_providers config in _config_context_length()

But no equivalent exists for max_tokens:

• run_agent.py:804 -- self.max_tokens only comes from __init__ parameter
• run_agent.py:1366-1431 -- custom_providers loop only reads context_length, not max_tokens
• run_agent.py:6644-6645 -- API call only sends max_tokens if self.max_tokens is not None
• cli.py:2795-2922 -- no max_tokens passed to AIAgent
• gateway/run.py:960 -- same
• Only batch_runner.py:329 has config.get("max_tokens"), but reads root-level config, not model-level

Proposed Solution

Add max_tokens support to the custom_providers model config, similar to how context_length works:

custom_providers:
  deepseek-v4:
    base_url: https://api.deepseek.com/v1
    models:
      deepseek-chat:
        context_length: 1000000
        max_tokens: 384000

And/or add a top-level default_max_tokens config key for models without explicit config.

Additional Context

• DeepSeek V4 (released 2026-04-24) supports up to 384K output tokens
• Related issue: #9489 (retry-based max_tokens boost, different scope)
• This is purely a config-reader enhancement -- no API behavior changes needed