Strip XML tags, runtime model metadata, run detail UI

CHANGELOG.md

@@ -55,6 +55,15 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
   tests and blocks all package builds (deb, rpm, arch, AppImage, snap,
   Homebrew, Docker) if they fail.
 
+### Added
+
+- **XML tag stripping**: Strip internal XML tags from LLM responses to prevent
+  tag leakage in chat (thinking, reasoning, scratchpad, etc.)
+- **Runtime model metadata**: Fetch model metadata from provider APIs for
+  accurate context window detection during auto-compaction
+- **Run detail UI**: Panel showing tool calls and message flow for agent runs,
+  accessible via expandable button on assistant messages
+
 ### Fixed
 
 - **Docker TLS setup**: All Docker examples now expose port 13132 for CA

crates/agents/src/lib.rs

@@ -14,6 +14,7 @@ pub use {
     runner::AgentRunError,
 };
 pub mod provider_chain;
+pub mod response_sanitizer;
 pub mod silent_turn;
 pub mod skills;
 pub mod tool_registry;

crates/agents/src/model.rs

@@ -364,6 +364,18 @@ pub trait LlmProvider: Send + Sync {
     ) -> Pin<Box<dyn Stream<Item = StreamEvent> + Send + '_>> {
         self.stream(messages)
     }
+
+    /// Fetch runtime model metadata from the provider API.
+    ///
+    /// The default implementation returns a `ModelMetadata` derived from the
+    /// static `context_window()` value. Providers that support a `/models`
+    /// endpoint can override this to fetch the actual context length at runtime.
+    async fn model_metadata(&self) -> anyhow::Result<ModelMetadata> {
+        Ok(ModelMetadata {
+            id: self.id().to_string(),
+            context_length: self.context_window(),
+        })
+    }
 }
 
 /// Response from an LLM completion call.
@@ -389,6 +401,13 @@ pub struct Usage {
     pub cache_write_tokens: u32,
 }
 
+/// Runtime model metadata fetched from provider APIs.
+#[derive(Debug, Clone)]
+pub struct ModelMetadata {
+    pub id: String,
+    pub context_length: u32,
+}
+
 #[allow(clippy::unwrap_used, clippy::expect_used)]
 #[cfg(test)]
 mod tests {
@@ -473,14 +492,11 @@ mod tests {
 
     #[test]
     fn to_openai_assistant_with_tools() {
-        let msg = ChatMessage::assistant_with_tools(
-            Some("thinking".into()),
-            vec![ToolCall {
-                id: "call_1".into(),
-                name: "exec".into(),
-                arguments: serde_json::json!({"cmd": "ls"}),
-            }],
-        );
+        let msg = ChatMessage::assistant_with_tools(Some("thinking".into()), vec![ToolCall {
+            id: "call_1".into(),
+            name: "exec".into(),
+            arguments: serde_json::json!({"cmd": "ls"}),
+        }]);
         let val = msg.to_openai_value();
         assert_eq!(val["role"], "assistant");
         assert_eq!(val["content"], "thinking");
@@ -657,4 +673,47 @@ mod tests {
         assert!(matches!(&msgs[0], ChatMessage::User { .. }));
         assert!(matches!(&msgs[1], ChatMessage::Assistant { .. }));
     }
+
+    // ── ModelMetadata default trait impl ────────────────────────────
+
+    /// Minimal provider to test default `model_metadata()` behavior.
+    struct StubProvider;
+
+    #[async_trait::async_trait]
+    impl LlmProvider for StubProvider {
+        fn name(&self) -> &str {
+            "stub"
+        }
+
+        fn id(&self) -> &str {
+            "stub-model"
+        }
+
+        fn context_window(&self) -> u32 {
+            42_000
+        }
+
+        async fn complete(
+            &self,
+            _: &[ChatMessage],
+            _: &[serde_json::Value],
+        ) -> anyhow::Result<CompletionResponse> {
+            anyhow::bail!("not implemented")
+        }
+
+        fn stream(
+            &self,
+            _: Vec<ChatMessage>,
+        ) -> Pin<Box<dyn Stream<Item = StreamEvent> + Send + '_>> {
+            Box::pin(tokio_stream::empty())
+        }
+    }
+
+    #[tokio::test]
+    async fn default_model_metadata_returns_context_window() {
+        let provider = StubProvider;
+        let meta = provider.model_metadata().await.unwrap();
+        assert_eq!(meta.id, "stub-model");
+        assert_eq!(meta.context_length, 42_000);
+    }
 }

crates/agents/src/providers/github_copilot.rs

@@ -235,14 +235,11 @@ async fn fetch_valid_copilot_token(
     }
 
     let copilot_resp: CopilotTokenResponse = resp.json().await?;
-    let _ = token_store.save(
-        "github-copilot-api",
-        &OAuthTokens {
-            access_token: Secret::new(copilot_resp.token.clone()),
-            refresh_token: None,
-            expires_at: Some(copilot_resp.expires_at),
-        },
-    );
+    let _ = token_store.save("github-copilot-api", &OAuthTokens {
+        access_token: Secret::new(copilot_resp.token.clone()),
+        refresh_token: None,
+        expires_at: Some(copilot_resp.expires_at),
+    });
 
     Ok(copilot_resp.token)
 }