diff --git a/src/openhuman/inference/provider/compatible_tests.rs b/src/openhuman/inference/provider/compatible_tests.rs
index 1e123a3336..b6c7214214 100644
--- a/src/openhuman/inference/provider/compatible_tests.rs
+++ b/src/openhuman/inference/provider/compatible_tests.rs
@@ -2196,6 +2196,36 @@ fn reasoning_alias_captured_in_stream_delta() {
     );
 }
 
+/// Regression for Sentry TAURI-RUST-A5N: a provider that emits BOTH `reasoning`
+/// and `reasoning_content` in the same message object must not fail with
+/// `duplicate field \`reasoning_content\``. Both keys deserialize and fold into
+/// the canonical field, which wins when both are present.
+#[test]
+fn reasoning_and_reasoning_content_both_present_does_not_error() {
+    let json = r#"{"choices":[{"message":{"content":null,"reasoning":"alias cot","reasoning_content":"canonical cot"}}]}"#;
+    let resp: ApiChatResponse = serde_json::from_str(json)
+        .expect("both reasoning keys must parse without a duplicate-field error");
+    assert_eq!(
+        resp.choices[0].message.reasoning_content.as_deref(),
+        Some("canonical cot"),
+        "canonical reasoning_content wins when both keys are present"
+    );
+}
+
+/// Same regression on the streaming delta path (TAURI-RUST-A5N also hits the
+/// native stream parser at `compatible_stream_native.rs`).
+#[test]
+fn reasoning_and_reasoning_content_both_present_in_stream_delta_does_not_error() {
+    let json = r#"{"choices":[{"delta":{"reasoning":"alias cot","reasoning_content":"canonical cot"},"finish_reason":null}]}"#;
+    let chunk: StreamChunkResponse = serde_json::from_str(json)
+        .expect("both reasoning keys must parse without a duplicate-field error");
+    assert_eq!(
+        chunk.choices[0].delta.reasoning_content.as_deref(),
+        Some("canonical cot"),
+        "canonical reasoning_content wins when both keys are present"
+    );
+}
+
 /// End-to-end: a tool-call turn whose reasoning arrived under the `reasoning`
 /// alias must still be surfaced by `parse_native_response` so the agent loop
 /// can replay it on the follow-up request (the issue #3094 failure path).
diff --git a/src/openhuman/inference/provider/compatible_types.rs b/src/openhuman/inference/provider/compatible_types.rs
index b5cd83e037..ade2234a71 100644
--- a/src/openhuman/inference/provider/compatible_types.rs
+++ b/src/openhuman/inference/provider/compatible_types.rs
@@ -4,7 +4,7 @@
 //! as appropriate). External code only sees the public API on
 //! [`super::OpenAiCompatibleProvider`].
 
-use serde::{Deserialize, Serialize};
+use serde::{Deserialize, Deserializer, Serialize};
 
 // ── Request bodies ────────────────────────────────────────────────────────────
 
@@ -316,25 +316,58 @@ pub(crate) struct OpenHumanBilling {
     pub(crate) charged_amount_usd: f64,
 }
 
-#[derive(Debug, Deserialize, Serialize)]
+#[derive(Debug, Serialize)]
 pub(crate) struct ResponseMessage {
-    #[serde(default)]
     pub(crate) content: Option<String>,
     /// Reasoning/thinking models may return their chain-of-thought in a
     /// dedicated field instead of (or alongside) `content`. DeepSeek, Qwen3 and
     /// GLM-4 name it `reasoning_content`; OpenRouter and vLLM/SGLang-backed
-    /// OpenAI-compatible proxies emit it as `reasoning`. Accept both so the CoT
-    /// is captured regardless of the (third-party) provider's field name — it
+    /// OpenAI-compatible proxies emit it as `reasoning`. Both names fold into
+    /// this single field (see the manual `Deserialize` impl below) — the CoT
     /// must be echoed back verbatim on tool-call turns or thinking models reject
     /// the follow-up request with HTTP 400.
-    #[serde(default, alias = "reasoning")]
     pub(crate) reasoning_content: Option<String>,
-    #[serde(default)]
     pub(crate) tool_calls: Option<Vec<ToolCall>>,
-    #[serde(default)]
     pub(crate) function_call: Option<Function>,
 }
 
+// Manual `Deserialize` so that `reasoning` and `reasoning_content` are accepted
+// as DISTINCT wire keys and then folded into the single canonical field.
+//
+// A serde `alias` maps both names onto one field slot, which makes a provider
+// that emits BOTH keys in the same object (some OpenRouter / vLLM-SGLang
+// proxies do) fail with `duplicate field \`reasoning_content\``, dropping the
+// entire response. Deserializing them as separate optional fields tolerates
+// any combination; the canonical `reasoning_content` wins when both are present.
+impl<'de> Deserialize<'de> for ResponseMessage {
+    fn deserialize<D>(deserializer: D) -> Result<Self, D::Error>
+    where
+        D: Deserializer<'de>,
+    {
+        #[derive(Deserialize)]
+        struct Shadow {
+            #[serde(default)]
+            content: Option<String>,
+            #[serde(default)]
+            reasoning_content: Option<String>,
+            #[serde(default)]
+            reasoning: Option<String>,
+            #[serde(default)]
+            tool_calls: Option<Vec<ToolCall>>,
+            #[serde(default)]
+            function_call: Option<Function>,
+        }
+
+        let shadow = Shadow::deserialize(deserializer)?;
+        Ok(ResponseMessage {
+            content: shadow.content,
+            reasoning_content: shadow.reasoning_content.or(shadow.reasoning),
+            tool_calls: shadow.tool_calls,
+            function_call: shadow.function_call,
+        })
+    }
+}
+
 impl ResponseMessage {
     /// Extract text content, falling back to `reasoning_content` when `content`
     /// is missing or empty. Reasoning/thinking models (Qwen3, GLM-4, etc.)
@@ -426,22 +459,50 @@ pub(crate) struct StreamChoice {
     pub(crate) finish_reason: Option<String>,
 }
 
-#[derive(Debug, Deserialize)]
+#[derive(Debug)]
 pub(crate) struct StreamDelta {
-    #[serde(default)]
     pub(crate) content: Option<String>,
     /// Reasoning/thinking models may stream their chain-of-thought via
     /// `reasoning_content` (DeepSeek/Qwen3/GLM-4) or `reasoning`
-    /// (OpenRouter, vLLM/SGLang proxies). Accept both delta field names.
-    #[serde(default, alias = "reasoning")]
+    /// (OpenRouter, vLLM/SGLang proxies). Both delta field names fold into
+    /// this single field (see the manual `Deserialize` impl below).
     pub(crate) reasoning_content: Option<String>,
     /// Native tool-call chunks. Each entry is keyed by `index`; the first
     /// chunk for a given index carries `id`/`type`/`function.name`, later
     /// chunks only carry fragments of `function.arguments`.
-    #[serde(default)]
     pub(crate) tool_calls: Option<Vec<StreamToolCallDelta>>,
 }
 
+// Manual `Deserialize` for the same reason as `ResponseMessage`: a streaming
+// delta that carries both `reasoning` and `reasoning_content` must not fail
+// with `duplicate field`. They deserialize as distinct keys and fold into the
+// canonical `reasoning_content` (canonical wins when both are present).
+impl<'de> Deserialize<'de> for StreamDelta {
+    fn deserialize<D>(deserializer: D) -> Result<Self, D::Error>
+    where
+        D: Deserializer<'de>,
+    {
+        #[derive(Deserialize)]
+        struct Shadow {
+            #[serde(default)]
+            content: Option<String>,
+            #[serde(default)]
+            reasoning_content: Option<String>,
+            #[serde(default)]
+            reasoning: Option<String>,
+            #[serde(default)]
+            tool_calls: Option<Vec<StreamToolCallDelta>>,
+        }
+
+        let shadow = Shadow::deserialize(deserializer)?;
+        Ok(StreamDelta {
+            content: shadow.content,
+            reasoning_content: shadow.reasoning_content.or(shadow.reasoning),
+            tool_calls: shadow.tool_calls,
+        })
+    }
+}
+
 #[derive(Debug, Deserialize)]
 pub(crate) struct StreamToolCallDelta {
     /// Index of this tool call within the assistant message. Multiple