fix: custom provider with slash model name no longer rerouted to OpenRouter

[nesquena-hermes]

Summary

Fixes a routing bug where custom provider configurations with slash-containing model names (e.g. google/gemma-4-26b-a4b) were silently rerouted to OpenRouter instead of the user's custom endpoint.

Problem

resolve_model_provider() in api/config.py checks custom_providers list entries first, but if a user configures only model.default: google/gemma-4-26b-a4b with provider: custom and base_url: http://... (without a custom_providers list), the code falls through to:

if prefix in _PROVIDER_MODELS and prefix != config_provider:
    return model_id, 'openrouter', None  # BUG: 'google' is in _PROVIDER_MODELS

This routes to OpenRouter, which returns a 401. The CLI doesn't have this problem because it uses a different routing path.

Fix

One guard added before the _PROVIDER_MODELS heuristic in resolve_model_provider():

# If a custom endpoint base_url is configured, trust it entirely.
# Don't apply the OpenRouter slash heuristic — a slash in the model
# name is just part of the model identifier, not a provider routing hint.
if config_base_url:
    return model_id, config_provider, config_base_url

When base_url is explicitly configured, the function returns immediately with the configured provider and URL. The OpenRouter slash heuristic only fires when no base_url is set.

Tests

Added test_custom_endpoint_slash_model_routes_to_custom_not_openrouter to tests/test_model_resolver.py:

• Verifies provider=custom + base_url + google/gemma-4-26b-a4b → routes to custom, not openrouter
• Verifies provider=openrouter + same model still routes to openrouter (no regression)

625 passed, 0 failed, 0 skipped (up from 624 — +1 new test)

Fixes #230

[nesquena]

Full Review: PR #255 — custom provider slash model routing fix

Security Audit

Clean. One logic branch added to resolve_model_provider() — when config_base_url is set and the model has a slash, return as-is instead of rerouting to OpenRouter. No new inputs, no new endpoints.

Code Review

The fix is correct and minimal:

if config_base_url:
    return model_id, config_provider, config_base_url

This runs after the OpenRouter early-return (which preserves slash models for OpenRouter) and after the prefix-match check (which strips provider prefixes for direct API). The logic is: "if the user configured a base_url, trust their routing — don't interpret a slash in the model name as OpenRouter syntax."

The regression test is thorough — tests both the custom provider case (should NOT go to OpenRouter) and the OpenRouter case (should still work as before). Good boundary coverage.

This fixes the same issue as external PR #189 (@smurmann) via a different mechanism. #189 checks custom_providers config entries; this one checks config_base_url. This approach is simpler (5 lines vs 14) and covers the broader case — any provider with a base_url, not just explicitly declared custom providers.

Tests

574 passed, 0 failed. (1 pre-existing failure from unmerged #243.)

Stacked Branch Note

PRs #256 and #257 are stacked on top of this branch. Merge order: #255 first, then #256, then #257. After #255 merges, #256 and #257 will need rebasing to drop the already-merged commits.

Verdict

Approved. Clean, minimal, correct. Ready to merge.

[nesquena-hermes]

Agent review — APPROVED ✅ (merged to stage)

Reviewed full diff and verified the fix in resolve_model_provider(). The guard if config_base_url: return model_id, config_provider, config_base_url is correctly positioned after the openrouter early-return and before the slash-based OpenRouter heuristic — meaning it only fires when a custom endpoint is configured, and openrouter configurations are never affected.

The regression test covers both paths: custom+base_url with slash model → stays custom, openrouter with slash model → still routes to openrouter.

Tests: 625 passed on review branch. Stage total: 645 passed, 0 failed.
Merged to stage branch.

[nesquena-hermes]

Merged to master via stage branch in PR #258 (v0.47.0).