Merge pull request #4173 from nesquena/stage-titlegen

Release NV (v0.51.409): provider-gate title-gen reasoning extra_body (#4161/#2083, consolidates #4162+#3944)

CHANGELOG.md

@@ -3,6 +3,12 @@
 
 ## [Unreleased]
 
+## [v0.51.409] — 2026-06-14 — Release NV (provider-gate title-gen reasoning extra_body, #4161/#2083)
+
+### Fixed
+
+- **Auto-title generation no longer sends an unsupported `reasoning` parameter to providers that reject it, restoring LLM-quality titles (#4161).** To suppress thinking on reasoning models (#2083), title generation injects `extra_body={"reasoning": {"enabled": false}}` — but OpenAI Chat Completions and Azure OpenAI reject unknown top-level params with a 400, so every new session silently fell back to a low-quality heuristic title for users on those providers. The reasoning-disable is now gated: the auxiliary title path skips it for reject-listed routes (OpenAI / Azure) while keeping it for local endpoints, OpenRouter, and other reasoning-aware providers; and the agent title path gates it behind the agent's canonical `_supports_reasoning_extra_body()` route check, which also excludes OpenRouter Anthropic mandatory-reasoning models (Claude Sonnet 4.6 / Opus 4.8) that are reasoning-capable but reject a disable. MiniMax keeps its existing `reasoning_split` handling. (#4161, #2083)
+
 ## [v0.51.408] — 2026-06-14 — Release NU (reasoning selector for nested Gemini custom-provider routes, #3431)
 
 ### Fixed

api/streaming.py

@@ -2413,6 +2413,48 @@ def _is_minimax_route(provider: str = '', model: str = '', base_url: str = '') -
     return 'minimax' in text or 'minimaxi.com' in text
 
 
+def _route_rejects_reasoning_extra(provider: str = '', model: str = '', base_url: str = '') -> bool:
+    """Routes known to reject an ``extra_body`` ``reasoning`` parameter with HTTP 400.
+
+    Title generation injects ``extra_body={"reasoning": {"enabled": False}}`` to
+    suppress thinking on reasoning-capable models (#2083). But OpenAI Chat
+    Completions (and Azure OpenAI) reject unknown top-level params with a 400, so
+    that inject silently fails the title call and falls back to a low-quality
+    heuristic title (#4161). Skip the inject for those routes.
+
+    OpenRouter Anthropic mandatory-reasoning models (Claude Sonnet 4.6 / Opus 4.8)
+    are reasoning-capable but reject a reasoning *disable* — title gen only needs
+    reasoning off, so skip the inject for them too rather than risk the same 400.
+    """
+    provider_lower = str(provider or '').strip().lower()
+    model_lower = str(model or '').strip().lower()
+    # Hostname-based match (not substring) so a proxy URL that merely *contains*
+    # one of these strings in a path segment isn't mis-classified.
+    host = ''
+    try:
+        from urllib.parse import urlsplit
+        host = (urlsplit(str(base_url or '').strip()).hostname or '').lower()
+    except Exception:
+        host = ''
+    if host == 'api.openai.com' or host.endswith('.openai.azure.com'):
+        return True
+    # Azure AI Foundry chat-completions hosts (also reject the reasoning param).
+    if host.endswith('.services.ai.azure.com') or host.endswith('.cognitiveservices.azure.com'):
+        return True
+    if provider_lower in ('openai', 'openai-api', 'openai-codex'):
+        return True
+    if (
+        provider_lower in ('azure', 'azure-foundry', 'azure-ai-foundry', 'azure-ai')
+        or provider_lower.startswith('azure/')
+        or provider_lower.startswith('azure-')
+    ):
+        return True
+    if (host == 'openrouter.ai' or host.endswith('.openrouter.ai')) and model_lower.startswith('anthropic/'):
+        # Anthropic on OpenRouter: mandatory-reasoning families reject a disable.
+        return True
+    return False
+
+
 def _get_aux_title_config() -> dict:
     """Return title_generation auxiliary config, or an empty dict on errors."""
     try:
@@ -2594,7 +2636,9 @@ def generate_title_raw_via_aux(
     if not caller_supplied_route:
         api_key = str(configured.get('api_key', '') or '').strip()
     base_max_tokens = _title_completion_budget(provider, model, base_url)
-    reasoning_extra = {"reasoning": {"enabled": False}}
+    reasoning_extra = {}
+    if not _route_rejects_reasoning_extra(provider, model, base_url):
+        reasoning_extra["reasoning"] = {"enabled": False}
     if _is_minimax_route(provider, model, base_url):
         reasoning_extra["reasoning_split"] = True
     try:
@@ -2619,7 +2663,7 @@ def generate_title_raw_via_aux(
                         max_tokens=max_tokens,
                         temperature=0.2,
                         timeout=_timeout,
-                        extra_body=reasoning_extra,
+                        extra_body=reasoning_extra or None,
                     )
                     raw, empty_status = _extract_title_response(resp, aux=True)
                     if raw:
@@ -2707,10 +2751,22 @@ def generate_title_raw_via_agent(agent, user_text: str, assistant_text: str) ->
                         api_kwargs.pop('tools', None)
                         api_kwargs['temperature'] = 0.1
                         api_kwargs['timeout'] = 15.0
+                        # Reasoning suppression for title gen is already handled
+                        # route-correctly by `_build_api_kwargs()` from the
+                        # `agent.reasoning_config = {"enabled": False}` set above —
+                        # each provider profile applies (or deliberately omits) the
+                        # disable in the form its endpoint accepts (OpenAI/Nous omit
+                        # the field; LM Studio uses top-level reasoning_effort;
+                        # OpenRouter Anthropic mandatory-reasoning is omitted). Do NOT
+                        # re-inject a generic `reasoning:{enabled:False}` here — that
+                        # re-adds a 400-rejected param on top of the profile output
+                        # (#4161). MiniMax still needs reasoning_split, which the
+                        # profile path does not add.
+                        _tg_extra = dict(api_kwargs.get('extra_body') or {})
                         if _is_minimax_route(getattr(agent, 'provider', ''), getattr(agent, 'model', ''), getattr(agent, 'base_url', '')):
-                            extra_body = dict(api_kwargs.get('extra_body') or {})
-                            extra_body['reasoning_split'] = True
-                            api_kwargs['extra_body'] = extra_body
+                            _tg_extra['reasoning_split'] = True
+                        if _tg_extra:
+                            api_kwargs['extra_body'] = _tg_extra
                         if 'max_completion_tokens' in api_kwargs:
                             api_kwargs['max_completion_tokens'] = max_tokens
                         else:

tests/test_title_gen_reasoning_extra_gate.py

@@ -0,0 +1,56 @@
+"""Consolidated regression tests for title-gen reasoning extra_body gating.
+
+Consolidates #2083 (suppress thinking on reasoning models so titles aren't
+polluted) with #4161 (OpenAI/Azure Chat Completions reject the `reasoning`
+extra_body param with a 400 -> silent fallback to heuristic titles).
+
+Aux path (`generate_title_raw_via_aux`, no agent object): inject the
+reasoning-disable EXCEPT on reject-listed routes (OpenAI/Azure).
+Agent path (`generate_title_raw_via_agent`): inject only when the agent's
+canonical `_supports_reasoning_extra_body()` says the route is reasoning-
+tolerant AND the route is not reject-listed — which additionally excludes
+OpenRouter Anthropic mandatory-reasoning models (Claude Sonnet 4.6 / Opus 4.8)
+that are reasoning-capable but 400 on a disable.
+"""
+from __future__ import annotations
+
+from api.streaming import _route_rejects_reasoning_extra
+
+
+class TestAuxRejectList:
+    def test_openai_direct_is_reject_listed(self):
+        assert _route_rejects_reasoning_extra("openai", "gpt-5.5", "https://api.openai.com/v1") is True
+
+    def test_azure_is_reject_listed(self):
+        assert _route_rejects_reasoning_extra("azure", "gpt-4", "https://x.openai.azure.com/") is True
+        assert _route_rejects_reasoning_extra("azure/foo", "gpt-4", "") is True
+
+    def test_azure_foundry_aliases_reject_listed(self):
+        assert _route_rejects_reasoning_extra("azure-foundry", "gpt-5", "") is True
+        assert _route_rejects_reasoning_extra("azure-ai-foundry", "gpt-5", "") is True
+        assert _route_rejects_reasoning_extra("azure-ai", "gpt-5", "") is True
+        # Foundry host-based detection (services.ai.azure.com / cognitiveservices)
+        assert _route_rejects_reasoning_extra("custom", "gpt-5", "https://x.services.ai.azure.com/v1") is True
+
+    def test_hostname_match_not_substring(self):
+        # A proxy whose PATH merely contains api.openai.com must NOT be reject-listed.
+        assert _route_rejects_reasoning_extra("custom", "qwen3", "https://proxy.example.test/api.openai.com/v1") is False
+        # but the real OpenAI host is.
+        assert _route_rejects_reasoning_extra("custom", "gpt-5", "https://api.openai.com/v1") is True
+
+    def test_openai_codex_alias_is_reject_listed(self):
+        assert _route_rejects_reasoning_extra("openai-codex", "gpt-5", "") is True
+
+    def test_openrouter_non_anthropic_is_not_reject_listed(self):
+        assert _route_rejects_reasoning_extra("openrouter", "deepseek/deepseek-r1", "https://openrouter.ai/api/v1") is False
+
+    def test_openrouter_anthropic_mandatory_is_reject_listed(self):
+        # Promoted into the shared helper so BOTH aux and agent paths skip it.
+        assert _route_rejects_reasoning_extra("openrouter", "anthropic/claude-sonnet-4.6", "https://openrouter.ai/api/v1") is True
+        assert _route_rejects_reasoning_extra("openrouter", "anthropic/claude-opus-4.8", "https://openrouter.ai/api/v1") is True
+
+    def test_local_lmstudio_is_not_reject_listed(self):
+        assert _route_rejects_reasoning_extra("lmstudio", "qwen3-8b", "http://localhost:1234/v1") is False
+
+    def test_minimax_is_not_reject_listed(self):
+        assert _route_rejects_reasoning_extra("", "minimax-m2", "https://api.minimaxi.com/v1") is False