Summary

Fixes #28046 — max_tokens configured under custom_providers[].models.<model>.max_tokens was silently ignored. Output token requests fell back to the hard-coded 4096 default in chat_completion_helpers.py:269 / conversation_loop.py:2913, capping responses even when the user configured a higher per-model limit.

Root cause

agent/agent_init.py already had a per-model context_length lookup against custom_providers, but no equivalent for max_tokens. So self.max_tokens stayed None from the constructor default and the self.max_tokens or 4096 fallback kicked in at request time.

Fix

Added a parallel max_tokens lookup right after the existing context_length block in agent/agent_init.py:

• Only runs when agent.max_tokens is None (don't override an explicit constructor / CLI value).
• Matches on base_url then model against the custom_providers list — same matching as the context_length branch above.
• Coerces via int(); positive values win, non-positive / non-numeric values log a warning via _ra().logger.warning and leave max_tokens unchanged (so the 4096 default still kicks in).

Tests

tests/run_agent/test_custom_provider_max_tokens.py — 6 cases:

1. valid integer max_tokens is applied
2. string-numeric ("16000") parses and is applied
3. non-numeric ("32K") is rejected with a warning, stays None
4. zero is rejected with a warning, stays None
5. missing max_tokens leaves it None
6. explicit constructor max_tokens is not overridden by the custom_providers lookup

All 6 pass; broader tests/run_agent/ smoke run shows no regression.

Repro (from issue)

custom_providers:
  - name: xfyun
    base_url: https://maas-coding-api.cn-huabei-1.xf-yun.com/v2
    api_key: ${API_KEY}
    api_mode: chat_completions
    model: astron-code-latest
    models:
      astron-code-latest:
        context_length: 200000
        max_tokens: 32000
        reasoning: true

Before: responses cap at ~4096 tokens with finish_reason='length'.
After: configured max_tokens=32000 is honored.

Risk

Narrow — additive block, gated on max_tokens is None and _custom_providers. Mirrors a well-trodden code path right above it. Default behavior (no custom_providers.models.<m>.max_tokens set) is unchanged.