fix: enable reasoning for third-party OpenAI-compatible proxies

[Sun-sunshine06]

Summary

Fixes #188.

The issue was that when a user configured:

• OpenAI-compatible protocol
• A third-party proxy (e.g. univibe)
• A reasoning-capable model ID (e.g. )

would return because it only enabled reasoning for official OpenAI API. This caused the synthesized PiModel to have , and the gateway would reject the request or return empty content, resulting in and the UI showing completed but with an empty preview.

The fix

Extend the heuristic: for third-party OpenAI-compatible gateways, still check if the model ID matches a known reasoning model pattern (same patterns already used for OpenRouter). This enables for Claude 4, o1/o3/gpt-5, qwq, deepseek-r, etc when proxied through OpenAI-compatible endpoints.

Testing

• Added test cases covering the issue scenario
• Existing tests still pass
• Fix verified by the issue author's log: the model on will now correctly get

Closes #188.

🤖 Generated with Claude Code

[github-actions]

Findings

• [Major] Third-party openai-chat misses canonical gpt-5 model IDs — the new pattern only matches gpt-5-... or gpt-5...., so plain gpt-5 on OpenAI-compatible gateways will be classified as non-reasoning and lose reasoning behavior. Evidence packages/providers/src/index.ts:202.
  Suggested fix:

  const REASONING_MODEL_ID_PATTERN = new RegExp(
    [
      ':thinking$',
      '(^|/)claude-(?:opus|sonnet)-4',
      '^o1[-.]', '^o3[-.]', '^o4[-.]', '^gpt-5(?:[-.]|$)',
      '^minimax/minimax-m\\d',
      '^deepseek/deepseek-r\\d',
      '^qwen/qwq',
    ].join('|'),
    'i',
  );

Summary

• Review mode: initial
• 1 issue found on changed lines (model-ID classification regression for proxied gpt-5).

Testing

• Not run (automation)

open-codesign Bot

[github-actions]

[Major] ^gpt-5[-.] misses plain gpt-5 IDs on third-party OpenAI-compatible gateways, so inferReasoning('openai-chat', 'gpt-5', '<custom base>') becomes false.

Suggested fix:

'^gpt-5(?:[-.]|$)'

Please also add a regression test in index.test.ts for proxied plain gpt-5.

[github-actions]

Findings

• [Minor] Missing direct regression test for bare proxied gpt-5 ID — this follow-up changes the matcher to support plain gpt-5, but the new tests only cover openai/gpt-5.1; without the exact gpt-5 case, the prior bug can regress unnoticed. Evidence packages/providers/src/index.test.ts:575
  Suggested fix:

  it('returns true for third-party openai-chat bare gpt-5 model ID', () => {
    expect(inferReasoning('openai-chat', 'gpt-5', 'https://my-proxy.example/v1')).toBe(true);
  });

Summary

• Review mode: follow-up after new commits
• 1 issue found on changed lines (regression-test coverage gap for the exact prior bug case).

Testing

• Not run (automation)

open-codesign Bot

[github-actions]

[Minor] The matcher fix now permits bare gpt-5, but this test block only adds openai/gpt-5.1; please add an explicit proxied gpt-5 regression case to lock in the exact prior bug.

Suggested fix:

expect(inferReasoning('openai-chat', 'gpt-5', 'https://my-proxy.example/v1')).toBe(true);