fix(inference): replay deepseek reasoning_content on tool-call turns (Sentry TAURI-RUST-4KB) (#2876)

Co-authored-by: Ghost Scripter <ghostscripter@zerolend.xyz>

src/openhuman/agent/harness/parse.rs

@@ -597,7 +597,17 @@ pub(crate) fn parse_structured_tool_calls(tool_calls: &[ToolCall]) -> Vec<Parsed
 /// Build assistant history entry in JSON format for native tool-call APIs.
 /// `convert_messages` in the OpenRouter provider parses this JSON to reconstruct
 /// the proper `NativeMessage` with structured `tool_calls`.
-pub(crate) fn build_native_assistant_history(text: &str, tool_calls: &[ToolCall]) -> String {
+///
+/// `reasoning_content` carries the model's thinking output (when the provider
+/// surfaced it). It is persisted so the next request can replay it: DeepSeek's
+/// thinking mode rejects an `assistant` turn that carries `tool_calls` if its
+/// `reasoning_content` is not passed back (Sentry TAURI-RUST-4KB). Omitted from
+/// the JSON when empty, so non-reasoning models are unaffected.
+pub(crate) fn build_native_assistant_history(
+    text: &str,
+    reasoning_content: Option<&str>,
+    tool_calls: &[ToolCall],
+) -> String {
     let calls_json: Vec<serde_json::Value> = tool_calls
         .iter()
         .map(|tc| {
@@ -615,11 +625,16 @@ pub(crate) fn build_native_assistant_history(text: &str, tool_calls: &[ToolCall]
         serde_json::Value::String(text.trim().to_string())
     };
 
-    serde_json::json!({
+    let mut entry = serde_json::json!({
         "content": content,
         "tool_calls": calls_json,
-    })
-    .to_string()
+    });
+
+    if let Some(reasoning) = reasoning_content.map(str::trim).filter(|r| !r.is_empty()) {
+        entry["reasoning_content"] = serde_json::Value::String(reasoning.to_string());
+    }
+
+    entry.to_string()
 }
 
 pub(crate) fn build_assistant_history_with_tool_calls(

src/openhuman/agent/harness/parse_tests.rs

@@ -267,10 +267,25 @@ fn structured_tool_call_and_history_helpers_round_trip_expected_shapes() {
     assert_eq!(parsed.len(), 1);
     assert_eq!(parsed[0].arguments, serde_json::json!({ "value": "hello" }));
 
-    let native = build_native_assistant_history("done", &tool_calls);
+    let native = build_native_assistant_history("done", None, &tool_calls);
     let native_json: serde_json::Value = serde_json::from_str(&native).expect("valid json");
     assert_eq!(native_json["content"], "done");
     assert_eq!(native_json["tool_calls"][0]["id"], "call-1");
+    // No reasoning supplied -> field omitted entirely (non-reasoning models
+    // must not gain a spurious `reasoning_content` key).
+    assert!(native_json.get("reasoning_content").is_none());
+
+    // DeepSeek thinking mode: reasoning must round-trip onto the tool-call
+    // turn (Sentry TAURI-RUST-4KB).
+    let native_reasoning =
+        build_native_assistant_history("done", Some("  step-by-step thoughts  "), &tool_calls);
+    let reasoning_json: serde_json::Value =
+        serde_json::from_str(&native_reasoning).expect("valid json");
+    assert_eq!(reasoning_json["reasoning_content"], "step-by-step thoughts");
+    // Whitespace-only reasoning is treated as absent.
+    let native_blank = build_native_assistant_history("done", Some("   "), &tool_calls);
+    let blank_json: serde_json::Value = serde_json::from_str(&native_blank).expect("valid json");
+    assert!(blank_json.get("reasoning_content").is_none());
 
     let xml_history = build_assistant_history_with_tool_calls("", &tool_calls);
     assert!(xml_history.contains("<tool_call>"));

src/openhuman/agent/harness/subagent_runner/ops.rs

@@ -1526,8 +1526,11 @@ async fn run_inner_loop(
         if force_text_mode {
             history.push(ChatMessage::assistant(response_text.clone()));
         } else {
-            let assistant_history_content =
-                super::super::parse::build_native_assistant_history(&response_text, &native_calls);
+            let assistant_history_content = super::super::parse::build_native_assistant_history(
+                &response_text,
+                resp.reasoning_content.as_deref(),
+                &native_calls,
+            );
             history.push(ChatMessage::assistant(assistant_history_content));
         }
 

src/openhuman/agent/harness/tool_loop.rs

@@ -583,7 +583,11 @@ pub(crate) async fn run_tool_call_loop(
                     let assistant_history_content = if resp.tool_calls.is_empty() {
                         response_text.clone()
                     } else {
-                        build_native_assistant_history(&response_text, &resp.tool_calls)
+                        build_native_assistant_history(
+                            &response_text,
+                            resp.reasoning_content.as_deref(),
+                            &resp.tool_calls,
+                        )
                     };
 
                     let native_calls = resp.tool_calls;

src/openhuman/inference/provider/compatible.rs

@@ -593,6 +593,22 @@ impl OpenAiCompatibleProvider {
                                         .and_then(serde_json::Value::as_str)
                                         .map(ToString::to_string);
 
+                                    // Replay the assistant's reasoning so
+                                    // DeepSeek thinking mode accepts the
+                                    // tool-call turn on the follow-up request
+                                    // (Sentry TAURI-RUST-4KB). Prefer the value
+                                    // embedded in the JSON content (written by
+                                    // `build_native_assistant_history` in the
+                                    // tool-loop path); fall back to the value
+                                    // stored in `extra_metadata` (written by the
+                                    // main session-turn path).
+                                    let reasoning_content = value
+                                        .get("reasoning_content")
+                                        .and_then(serde_json::Value::as_str)
+                                        .filter(|s| !s.trim().is_empty())
+                                        .map(ToString::to_string)
+                                        .or_else(|| reasoning_content.clone());
+
                                     return NativeMessage {
                                         role: "assistant".to_string(),
                                         content,
@@ -792,7 +808,12 @@ impl OpenAiCompatibleProvider {
         // the next turn for thinking models (e.g. DeepSeek-R1, Qwen3) whose APIs
         // return HTTP 400 ("reasoning_content in thinking mode must be passed back")
         // when the field is omitted from subsequent assistant messages.
-        let reasoning_content = message.reasoning_content.clone();
+        let reasoning_content = message
+            .reasoning_content
+            .as_deref()
+            .map(str::trim)
+            .filter(|s| !s.is_empty())
+            .map(str::to_owned);
         let mut tool_calls = message
             .tool_calls
             .unwrap_or_default()
@@ -1739,6 +1760,15 @@ impl Provider for OpenAiCompatibleProvider {
             .ok_or_else(|| anyhow::anyhow!("No response from {}", self.name))?;
 
         let text = choice.message.effective_content_optional();
+        // See `parse_native_response`: replay reasoning on the follow-up
+        // request so DeepSeek thinking mode accepts the tool-call turn.
+        let reasoning_content = choice
+            .message
+            .reasoning_content
+            .as_deref()
+            .map(str::trim)
+            .filter(|s| !s.is_empty())
+            .map(ToString::to_string);
         let tool_calls = choice
             .message
             .tool_calls
@@ -1760,7 +1790,7 @@ impl Provider for OpenAiCompatibleProvider {
             text,
             tool_calls,
             usage,
-            reasoning_content: None,
+            reasoning_content,
         })
     }
 

src/openhuman/inference/provider/compatible_tests.rs

@@ -658,6 +658,49 @@ fn parse_native_response_preserves_tool_call_id() {
     assert_eq!(parsed.tool_calls[0].name, "shell");
 }
 
+/// DeepSeek thinking mode emits the chain-of-thought in `reasoning_content`
+/// alongside the tool call. `parse_native_response` must surface it so the
+/// agent loop can replay it on the follow-up request (Sentry TAURI-RUST-4KB).
+#[test]
+fn parse_native_response_captures_reasoning_content() {
+    let message = ResponseMessage {
+        content: None,
+        tool_calls: Some(vec![ToolCall {
+            id: Some("call_r".to_string()),
+            kind: Some("function".to_string()),
+            function: Some(Function {
+                name: Some("shell".to_string()),
+                arguments: Some(serde_json::Value::String("{}".to_string())),
+            }),
+        }]),
+        function_call: None,
+        reasoning_content: Some("  weighing the options  ".to_string()),
+    };
+
+    let parsed =
+        OpenAiCompatibleProvider::parse_native_response(wrap_message(message), "deepseek").unwrap();
+    assert_eq!(
+        parsed.reasoning_content.as_deref(),
+        Some("weighing the options")
+    );
+}
+
+/// Whitespace-only / empty reasoning is normalised to `None` so it never
+/// produces a spurious `reasoning_content` key on the wire.
+#[test]
+fn parse_native_response_blank_reasoning_is_none() {
+    let message = ResponseMessage {
+        content: Some("hello".to_string()),
+        tool_calls: None,
+        function_call: None,
+        reasoning_content: Some("   ".to_string()),
+    };
+
+    let parsed =
+        OpenAiCompatibleProvider::parse_native_response(wrap_message(message), "deepseek").unwrap();
+    assert!(parsed.reasoning_content.is_none());
+}
+
 #[test]
 fn convert_messages_for_native_maps_tool_result_payload() {
     // A `tool` result must be opened by a preceding `assistant(tool_calls)`,
@@ -898,6 +941,53 @@ fn tool_invariants_drop_orphan_but_keep_following_cycle() {
     assert_eq!(converted[1].tool_call_id.as_deref(), Some("call_b"));
 }
 
+/// DeepSeek thinking mode (Sentry TAURI-RUST-4KB): an `assistant` turn that
+/// carries `tool_calls` must replay its `reasoning_content` on the follow-up
+/// request, otherwise DeepSeek returns
+/// `400 The reasoning_content in the thinking mode must be passed back to the
+/// API.` The history JSON written by `build_native_assistant_history` carries
+/// `reasoning_content`; `convert_messages_for_native` must lift it back onto
+/// the wire message.
+#[test]
+fn convert_preserves_reasoning_content_on_tool_call_turn() {
+    let input = vec![ChatMessage::assistant(
+        r#"{"content":null,"reasoning_content":"let me think about this","tool_calls":[{"id":"call_x","name":"shell","arguments":"{}"}]}"#,
+    )];
+
+    let converted = OpenAiCompatibleProvider::convert_messages_for_native(&input);
+
+    assert_eq!(converted.len(), 1);
+    assert_eq!(
+        converted[0].reasoning_content.as_deref(),
+        Some("let me think about this")
+    );
+
+    // The wire payload must actually carry the field for DeepSeek to accept it.
+    let wire = serde_json::to_value(&converted[0]).unwrap();
+    assert_eq!(wire["reasoning_content"], "let me think about this");
+}
+
+/// Assistant tool-call turns from non-reasoning models carry no
+/// `reasoning_content`; it must never appear on the wire for them (most
+/// OpenAI-compatible providers don't recognise the field).
+#[test]
+fn convert_omits_reasoning_content_when_absent() {
+    let input = vec![ChatMessage::assistant(
+        r#"{"content":"sure","tool_calls":[{"id":"call_y","name":"shell","arguments":"{}"}]}"#,
+    )];
+
+    let converted = OpenAiCompatibleProvider::convert_messages_for_native(&input);
+
+    assert_eq!(converted.len(), 1);
+    assert!(converted[0].reasoning_content.is_none());
+
+    let wire = serde_json::to_value(&converted[0]).unwrap();
+    assert!(
+        wire.get("reasoning_content").is_none(),
+        "reasoning_content must be omitted from the wire when absent"
+    );
+}
+
 #[test]
 fn chat_message_identity_metadata_is_not_provider_wire_payload() {
     let message = ChatMessage {
@@ -1561,9 +1651,9 @@ fn enrich_404_message_adds_hint_when_no_fallback() {
 // ── reasoning_content round-trip tests (issue #2800 / Sentry TAURI-RUST-4WC) ─
 
 /// `parse_native_response` must capture `reasoning_content` from a non-streaming
-/// response and surface it on `ChatResponse`.
+/// `ApiChatResponse` and surface it on `ChatResponse`.
 #[test]
-fn parse_native_response_captures_reasoning_content() {
+fn parse_native_response_captures_reasoning_content_from_api_response() {
     let api_resp = ApiChatResponse {
         choices: vec![Choice {
             message: ResponseMessage {