What does this PR do?

Adds a config-driven max_tokens override for OpenAI-compatible custom providers, mirroring the existing model.context_length lookup.

Custom providers (vLLM, llama.cpp, ollama, ...) that do not advertise a max_tokens default via /models cause responses to truncate with finish_reason='length' because the agent never sends a max_tokens hint. The chat_completions transport leaves max_tokens unset, the server uses its own (often conservative) default, and long answers get cut mid-sentence.

This PR introduces:

1. A new get_max_tokens_from_config(model, base_url, config) helper in hermes_cli/config.py — alongside the existing get_custom_provider_context_length.
2. A wire-up in AIAgent.__init__ that calls the helper after the existing context_length resolution and applies its value to self.max_tokens only when the constructor didn't pass an explicit value.

Lookup order (first valid positive int wins):

• model.max_tokens (top-level — applies to whichever model is active)
• custom_providers.<>.models.<model>.max_tokens (per-model, scoped to the entry whose base_url matches)

No behaviour change for built-in providers (Anthropic / OpenAI / OpenRouter / Bedrock / Gemini) — they still resolve max_tokens through their own adapter logic.

Aligns with the schema proposed in issue #15037.

Related Issue

Fixes the user-side workaround needed for #15037.

Type of Change

• ✨ New feature (non-breaking change that adds functionality)

Changes Made

• hermes_cli/config.py — add get_max_tokens_from_config(model, base_url, config=None) -> Optional[int]:
  • Reads top-level model.max_tokens first; logs a warning and falls through if non-int / non-positive.
  • Falls back to custom_providers.<>.models.<>.max_tokens (URL match is trailing-slash insensitive, value must be a positive int).
  • Returns None when neither location holds a valid value.
• run_agent.py — AIAgent.__init__: after the existing context_length resolution, call the helper and apply the value to self.max_tokens if the constructor didn't pass one.
• tests/hermes_cli/test_max_tokens_config.py — 15 unit tests covering top-level / per-model / precedence / edge cases / providers-dict (v12+) form.

How to Test

1. Configure a custom provider in ~/.hermes/config.yaml:

   model:
     default: qwen3.6-27b-fp8
     provider: custom
     base_url: http://192.168.x.x:8080/v1
     max_tokens: 32768

2. Send a prompt likely to produce a long response (e.g. "summarise this 50K-word document").
3. Before this fix: the response gets cut off at the server's default cap (often 2K–4K) with finish_reason='length'.
4. After this fix: the agent sends max_tokens=32768, and the response runs to a natural stop.

Or run the unit tests:

pytest tests/hermes_cli/test_max_tokens_config.py -v

Checklist

Code

• I've read the Contributing Guide
• My commit messages follow Conventional Commits (feat(config): ...)
• I searched for existing PRs — feat(config): support per-model max_tokens in custom_providers config #15037 proposes this schema but no implementation has landed
• My PR contains only changes related to this feature
• I've run pytest tests/ -q and the changes don't introduce new failures (vs. main baseline)
• I've added tests for the change (15 cases)
• I've tested on my platform: macOS 15 (Apple Silicon), Qwen 3.6 27B FP8 served via vLLM

Documentation & Housekeeping

• No breaking change to the schema — model.max_tokens is opt-in
• Existing context_length semantics unchanged