Summary

Fixes #28046 — max_tokens configured under custom_providers[].models.<model>.max_tokens was silently ignored, always defaulting to 4096.

The codebase already had get_custom_provider_context_length() for reading per-model context_length from custom_providers, but no equivalent for max_tokens.

Changes

1. Extract get_custom_provider_model_field() — generic lookup helper that searches custom_providers entries for any per-model field. Replaces the inline logic in get_custom_provider_context_length().

2. Add get_custom_provider_max_tokens() — thin wrapper over the generic helper, symmetric to get_custom_provider_context_length().

3. Read max_tokens in agent_init.py — after the existing context_length custom_providers lookup, adds a symmetric block that calls get_custom_provider_max_tokens() when agent.max_tokens is None.

4. 10 regression tests — full coverage for the new lookup: matching, trailing-slash insensitivity, zero/negative rejection, string coercion, coexistence with context_length, first-match-wins, None inputs.

Backward Compatibility

• get_custom_provider_context_length() signature unchanged — custom_providers remains a positional arg (3 existing tests pass without modification)
• New code only runs when agent.max_tokens is None (no override from top-level config or constructor) — pure additive fallback
• 22/22 tests pass (12 existing context_length + 10 new max_tokens)

Root Cause Pattern

This is the same pattern as #28961 (pre_tool_call missing session_id) and #28984 (Typed Plugin Hook Protocol FR): configuration/state flows through call chains without schema enforcement, so adding a new config field requires manually updating every bridge — and omissions are silent.