fix(agent): comprehensive DeepSeek V4 support — context windows, thinking mode, reasoning replay

Unifies approaches from PRs #14952, #14958, #15325, #15228, #15354
into a single cohesive implementation:

- Add 1M context window entries for V4 models (deepseek-v4-pro,
  deepseek-v4-flash, deepseek-chat, deepseek-reasoner)
- Plumb thinking.type toggle and reasoning_effort mapping for
  native DeepSeek API (only "high" and "max" are valid)
- Strip incompatible sampling params when thinking is enabled
- Inject reasoning_content="" on all assistant messages for
  DeepSeek replay (scoped to api.deepseek.com and OpenRouter)
- Fix _extract_reasoning isinstance checks for empty strings
- Preserve empty-string reasoning_content in normalize_response
- Add _copy_reasoning_content_for_api call in _handle_max_iterations

Fixes #15353. Supersedes #14952, #14958, #15325, #15228, #15354.

agent/model_metadata.py

@@ -162,8 +162,12 @@ DEFAULT_CONTEXT_LENGTHS = {
     "gemma-4-31b": 256000,
     "gemma-3": 131072,
     "gemma": 8192,  # fallback for older gemma models
-    # DeepSeek
-    "deepseek": 128000,
+    # DeepSeek — V4 family supports 1M context (api.deepseek.com docs)
+    "deepseek-v4-pro": 1000000,
+    "deepseek-v4-flash": 1000000,
+    "deepseek-chat": 1000000,
+    "deepseek-reasoner": 1000000,
+    "deepseek": 128000,  # fallback for older/unrecognised DeepSeek models
     # Meta
     "llama": 131072,
     # Qwen — specific model families before the catch-all.

agent/transports/chat_completions.py

@@ -239,6 +239,30 @@ class ChatCompletionsTransport(ProviderTransport):
                 "type": "enabled" if _kimi_thinking_enabled else "disabled",
             }
 
+        # DeepSeek: thinking mode toggle and effort mapping
+        is_deepseek = params.get("is_deepseek", False)
+        if is_deepseek:
+            _ds_thinking_enabled = True
+            if reasoning_config and isinstance(reasoning_config, dict):
+                if reasoning_config.get("enabled") is False:
+                    _ds_thinking_enabled = False
+            if _ds_thinking_enabled:
+                # DeepSeek only supports "high" and "max" effort values.
+                # Map low/medium/high → "high", xhigh/max → "max".
+                _ds_effort = "high"
+                if reasoning_config and isinstance(reasoning_config, dict):
+                    _e = (reasoning_config.get("effort") or "").strip().lower()
+                    if _e in ("xhigh", "max"):
+                        _ds_effort = "max"
+                extra_body["thinking"] = {"type": "enabled", "budget_tokens": 8192}
+                api_kwargs["reasoning_effort"] = _ds_effort
+                # DeepSeek rejects temperature/top_p/presence_penalty/
+                # frequency_penalty when thinking is enabled.
+                for _k in ("temperature", "top_p", "presence_penalty", "frequency_penalty"):
+                    api_kwargs.pop(_k, None)
+            else:
+                extra_body["thinking"] = {"type": "disabled"}
+
         # Reasoning
         if params.get("supports_reasoning", False):
             if is_github_models:
@@ -347,7 +371,7 @@ class ChatCompletionsTransport(ProviderTransport):
         reasoning_content = getattr(msg, "reasoning_content", None)
 
         provider_data: Dict[str, Any] = {}
-        if reasoning_content:
+        if reasoning_content is not None:
             provider_data["reasoning_content"] = reasoning_content
         rd = getattr(msg, "reasoning_details", None)
         if rd: