claude: fix Opus 5 wire incompatibilities

Drop temperature for models that removed sampling params (400 otherwise).
Preserve signed thinking blocks across tool rounds — Anthropic rejects a
tool_use whose preceding thinking block was dropped. Raise default
max_tokens to 32k for thinking-by-default models and request summarized
display so the reasoning pane is fed.

src/bridge/bifrost.rs

@@ -114,6 +114,23 @@ pub struct Message {
     /// Function name for `role: "tool"` messages (some providers require it).
     #[serde(skip_serializing_if = "Option::is_none", default)]
     pub name: Option<String>,
+    /// Signed thinking block from the assistant turn, carried so it can be
+    /// replayed. Anthropic requires the thinking block that preceded a
+    /// `tool_use` to come back with it, byte-identical and with its signature,
+    /// or the follow-up request is rejected. Only providers that return a
+    /// signed block populate this; it is skipped on the wire when absent, so
+    /// OpenAI-shaped gateways never see it.
+    #[serde(skip_serializing_if = "Option::is_none", default)]
+    pub thinking: Option<ThinkingBlock>,
+}
+
+/// An Anthropic thinking block preserved verbatim for replay. Both halves are
+/// required — text without its signature cannot be replayed, so this is only
+/// ever constructed when the provider returned both.
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct ThinkingBlock {
+    pub text: String,
+    pub signature: String,
 }
 
 impl Message {
@@ -125,6 +142,7 @@ impl Message {
             tool_calls: None,
             tool_call_id: None,
             name: None,
+            thinking: None,
         }
     }
 
@@ -136,6 +154,7 @@ impl Message {
             tool_calls: None,
             tool_call_id: None,
             name: None,
+            thinking: None,
         }
     }
 
@@ -147,9 +166,26 @@ impl Message {
             tool_calls: Some(calls),
             tool_call_id: None,
             name: None,
+            thinking: None,
         }
     }
 
+    /// Attach a signed thinking block for replay. No-op unless both the text
+    /// and the signature are present.
+    pub fn with_thinking(
+        mut self,
+        text: Option<impl Into<String>>,
+        signature: Option<impl Into<String>>,
+    ) -> Self {
+        if let (Some(text), Some(signature)) = (text, signature) {
+            self.thinking = Some(ThinkingBlock {
+                text: text.into(),
+                signature: signature.into(),
+            });
+        }
+        self
+    }
+
     /// Tool-result message bound to a prior assistant tool_call by id.
     pub fn tool_result(
         tool_call_id: impl Into<String>,
@@ -162,6 +198,7 @@ impl Message {
             tool_calls: None,
             tool_call_id: Some(tool_call_id.into()),
             name: Some(name.into()),
+            thinking: None,
         }
     }
 }
@@ -324,6 +361,10 @@ pub struct StreamToolCallFunction {
 pub struct CompletionResult {
     pub content: String,
     pub reasoning: Option<String>,
+    /// Opaque signature for the thinking block in `reasoning`. Anthropic
+    /// returns it and requires it back verbatim when the same assistant turn
+    /// carried a `tool_use`. `None` for providers that don't sign thinking.
+    pub reasoning_signature: Option<String>,
     pub tool_calls: Vec<ParsedToolCall>,
     pub finish_reason: Option<String>,
     pub usage: Option<Usage>,
@@ -695,6 +736,8 @@ impl BifrostClient {
         Ok(CompletionResult {
             content,
             reasoning,
+            // OpenAI-shaped gateways return reasoning text unsigned.
+            reasoning_signature: None,
             tool_calls,
             finish_reason: choice.finish_reason.clone(),
             usage: parsed.usage,

src/bridge/claude_subscription.rs

@@ -51,6 +51,37 @@ const CC_PLATFORM: &str = "claude_code_cli";
 const FALLBACK_CC_VERSION: &str = "2.1.196";
 const TOKEN_REFRESH_BUFFER: Duration = Duration::from_secs(5 * 60);
 const DEFAULT_MAX_TOKENS: u32 = 8192;
+/// Models that think by default cap thinking *plus* response text with
+/// `max_tokens`, so 8192 truncates mid-answer.
+const THINKING_MAX_TOKENS: u32 = 32000;
+
+/// Models that removed the sampling parameters. Sending `temperature`,
+/// `top_p` or `top_k` to one of these is a 400, so they are dropped rather
+/// than passed through.
+fn rejects_sampling_params(model: &str) -> bool {
+    const REMOVED: &[&str] = &[
+        "claude-opus-5",
+        "claude-opus-4-7",
+        "claude-opus-4-8",
+        "claude-sonnet-5",
+        "claude-fable-5",
+        "claude-mythos-5",
+        "claude-mythos-preview",
+    ];
+    REMOVED.iter().any(|m| model.starts_with(m))
+}
+
+/// Models where omitting `thinking` still runs adaptive thinking, which shares
+/// the `max_tokens` budget with the visible response.
+fn thinks_by_default(model: &str) -> bool {
+    const THINKS: &[&str] = &[
+        "claude-opus-5",
+        "claude-sonnet-5",
+        "claude-fable-5",
+        "claude-mythos-5",
+    ];
+    THINKS.iter().any(|m| model.starts_with(m))
+}
 
 /// Direct Claude Code subscription inference provider.
 pub struct ClaudeSubscriptionProvider {
@@ -160,14 +191,27 @@ impl ClaudeSubscriptionProvider {
     /// `/v1/messages` body.
     fn shape_body(&self, model: &str, request: &ChatCompletionRequest) -> Result<Vec<u8>> {
         let (messages, system_text, first_user_text) = translate_messages(&request.messages);
+        let default_max_tokens = if thinks_by_default(model) {
+            THINKING_MAX_TOKENS
+        } else {
+            DEFAULT_MAX_TOKENS
+        };
         let mut body = json!({
             "model": model,
             "messages": messages,
-            "max_tokens": request.max_tokens.unwrap_or(DEFAULT_MAX_TOKENS),
+            "max_tokens": request.max_tokens.unwrap_or(default_max_tokens),
             "stream": false,
         });
         if let Some(t) = request.temperature {
-            body["temperature"] = json!(t);
+            if !rejects_sampling_params(model) {
+                body["temperature"] = json!(t);
+            }
+        }
+        if thinks_by_default(model) {
+            // Thinking runs either way on these models; the default display is
+            // "omitted", which returns signed blocks whose text is empty and
+            // leaves the reasoning pane blank. Ask for the summary instead.
+            body["thinking"] = json!({ "type": "adaptive", "display": "summarized" });
         }
         if let Some(system) = system_text {
             body["system"] = json!(system);
@@ -598,6 +642,20 @@ fn translate_messages(openai: &[Message]) -> (Vec<Value>, Option<String>, String
                 push_merge(&mut merged, "user", block);
             }
             "assistant" => {
+                // The thinking block must lead the turn and must come back
+                // unmodified — Anthropic rejects a tool_use whose preceding
+                // thinking block was dropped or edited.
+                if let Some(thinking) = &msg.thinking {
+                    push_merge(
+                        &mut merged,
+                        "assistant",
+                        json!({
+                            "type": "thinking",
+                            "thinking": thinking.text,
+                            "signature": thinking.signature,
+                        }),
+                    );
+                }
                 if !msg.content.is_empty() {
                     push_merge(
                         &mut merged,
@@ -706,25 +764,24 @@ fn parse_completion(text: &str) -> Result<CompletionResult> {
         .unwrap_or_default();
     let content = content_parts.concat();
 
-    let reasoning = value
+    // The first thinking block carries the signature that has to be replayed
+    // alongside any tool_use in this same turn. Keep its text and signature
+    // paired — a concatenation of several blocks could not be signed.
+    let thinking_block = value
         .get("content")
         .and_then(Value::as_array)
         .and_then(|arr| {
-            let r: Vec<String> = arr
-                .iter()
-                .filter(|b| b.get("type").and_then(Value::as_str) == Some("thinking"))
-                .filter_map(|b| {
-                    b.get("thinking")
-                        .and_then(Value::as_str)
-                        .map(str::to_string)
-                })
-                .collect();
-            if r.is_empty() {
-                None
-            } else {
-                Some(r.concat())
-            }
+            arr.iter()
+                .find(|b| b.get("type").and_then(Value::as_str) == Some("thinking"))
         });
+    let reasoning = thinking_block
+        .and_then(|b| b.get("thinking"))
+        .and_then(Value::as_str)
+        .map(str::to_string);
+    let reasoning_signature = thinking_block
+        .and_then(|b| b.get("signature"))
+        .and_then(Value::as_str)
+        .map(str::to_string);
 
     let tool_calls = value
         .get("content")
@@ -762,6 +819,7 @@ fn parse_completion(text: &str) -> Result<CompletionResult> {
     Ok(CompletionResult {
         content,
         reasoning,
+        reasoning_signature,
         tool_calls,
         finish_reason,
         usage,
@@ -899,6 +957,7 @@ mod tests {
                 }]),
                 tool_call_id: None,
                 name: None,
+                thinking: None,
             },
             Message::tool_result("call_1", "read", "example-host"),
             Message::text("user", "thanks"),

src/bridge/providers/responses.rs

@@ -186,6 +186,7 @@ pub fn accumulate_sse(body: &str) -> Result<CompletionResult> {
     Ok(CompletionResult {
         content,
         reasoning: (!reasoning.is_empty()).then_some(reasoning),
+        reasoning_signature: None,
         tool_calls,
         finish_reason: Some(finish_reason.to_string()),
         usage,

src/core/reflection/mod.rs

@@ -236,10 +236,12 @@ impl ReflectionEngine {
                     tc.arguments.to_string(),
                 ))
                 .collect();
-            chat_messages.push(Message::assistant_tool_calls(
-                response.content.clone(),
-                calls,
-            ));
+            chat_messages.push(
+                Message::assistant_tool_calls(response.content.clone(), calls).with_thinking(
+                    response.reasoning.clone(),
+                    response.reasoning_signature.clone(),
+                ),
+            );
 
             for tc in &response.tool_calls {
                 let input_str = tc.arguments.to_string();

src/server/consciousness_engine.rs

@@ -868,10 +868,12 @@ Resolve entries: `[YYYY-MM-DD HH:MM] RESOLVED — note`"#;
                     tc.arguments.to_string(),
                 ))
                 .collect();
-            messages.push(Message::assistant_tool_calls(
-                response.content.clone(),
-                calls,
-            ));
+            messages.push(
+                Message::assistant_tool_calls(response.content.clone(), calls).with_thinking(
+                    response.reasoning.clone(),
+                    response.reasoning_signature.clone(),
+                ),
+            );
 
             // Execute each tool call and bind the result by tool_call_id.
             // Halt and intrusive are signal tools — we read them out of the

src/server/subagent.rs

@@ -166,10 +166,13 @@ impl SubagentRunner for ServerSubagentRunner {
                     tc.arguments.to_string(),
                 ))
                 .collect();
-            messages.push(BifrostMessage::assistant_tool_calls(
-                response.content.clone(),
-                calls,
-            ));
+            messages.push(
+                BifrostMessage::assistant_tool_calls(response.content.clone(), calls)
+                    .with_thinking(
+                        response.reasoning.clone(),
+                        response.reasoning_signature.clone(),
+                    ),
+            );
 
             // Execute tools with context, bind each result by tool_call_id
             for tc in &response.tool_calls {

src/server/turn.rs

@@ -523,10 +523,12 @@ pub(crate) async fn run_turn(
                 tc.arguments.to_string(),
             ))
             .collect();
-        messages.push(BifrostMessage::assistant_tool_calls(
-            response.content.clone(),
-            calls,
-        ));
+        messages.push(
+            BifrostMessage::assistant_tool_calls(response.content.clone(), calls).with_thinking(
+                response.reasoning.clone(),
+                response.reasoning_signature.clone(),
+            ),
+        );
 
         // Stream any text the model produced alongside tool calls as italic
         // interstitial narration. Configurable via tui.show_interstitial.