refactor: remove retry_count from ProviderConfig and enhance error handling in tool execution

- Removed the `retry_count` field from `ProviderConfig` as it is no longer needed.
- Introduced `argument_error` field in `ToolCall` to capture errors related to tool arguments.
- Updated various components to handle argument errors more gracefully, including in the `ToolDispatcher` and `ConversationOutput`.
- Enhanced tests to validate the new error handling and ensure proper functionality of tool calls.

server/migrations/0009_add_tool_call_argument_error.sql

@@ -0,0 +1 @@
+ALTER TABLE tool_round_calls ADD COLUMN argument_error TEXT;

server/src/config.rs

@@ -46,7 +46,6 @@ pub struct ProviderConfig {
     pub custom_headers: reqwest::header::HeaderMap,
     pub max_output_tokens: Option<u64>,
     pub request_timeout: Duration,
-    pub retry_count: u32,
     pub allowed_body_fields: Option<std::collections::HashSet<String>>,
 }
 

server/src/cursor/checkpoint/messages/decode.rs

@@ -102,6 +102,12 @@ pub fn decode_pending(value: &str) -> Result<RecoveredToolRound> {
         .into_iter()
         .enumerate()
         .map(|(index, call)| {
+            let argument_error = wire
+                .pointer("/providerOptions/cursor/pendingToolExecutionContracts")
+                .and_then(|contracts| contracts.get(&call.call_id))
+                .and_then(|contract| contract.get("argumentError"))
+                .and_then(Value::as_str)
+                .map(str::to_string);
             Ok(ToolCall {
                 index,
                 call_id: call.call_id,
@@ -109,6 +115,7 @@ pub fn decode_pending(value: &str) -> Result<RecoveredToolRound> {
                 name: call.name,
                 arguments_text: serde_json::to_string(&call.arguments)?,
                 arguments: call.arguments,
+                argument_error,
             })
         })
         .collect::<Result<Vec<_>>>()?;

server/src/cursor/checkpoint/messages/encode.rs

@@ -71,6 +71,7 @@ pub fn staged_tool_round(
             allowed_tools,
             dynamic_tools,
             started_at_ms,
+            tool_calls: Some(calls),
         }),
     )?)?)
 }
@@ -96,6 +97,7 @@ pub fn staged_final(
             allowed_tools,
             dynamic_tools,
             started_at_ms,
+            tool_calls: None,
         }),
     )?)?)
 }
@@ -105,6 +107,7 @@ pub(super) struct PendingContext<'a> {
     allowed_tools: &'a [String],
     dynamic_tools: &'a HashSet<String>,
     started_at_ms: u64,
+    tool_calls: Option<&'a [ToolCall]>,
 }
 
 pub(super) fn wire_message(
@@ -132,16 +135,23 @@ pub(super) fn wire_message(
                     calls
                         .iter()
                         .map(|call| {
-                            (
-                                call.call_id.clone(),
-                                json!({
+                            let mut contract = json!({
                                 "toolCallId": call.call_id,
                                 "outerToolName": call.name,
                                 "toolIdentifier": tool_identifier(&call.name, pending.dynamic_tools),
                                 "isDynamic": pending.dynamic_tools.contains(&call.name),
                                 "allowedToolNames": pending.allowed_tools,
-                                }),
-                            )
+                            });
+                            if let Some(error) = pending
+                                .tool_calls
+                                .and_then(|calls| {
+                                    calls.iter().find(|candidate| candidate.call_id == call.call_id)
+                                })
+                                .and_then(|call| call.argument_error.as_deref())
+                            {
+                                contract["argumentError"] = Value::String(error.into());
+                            }
+                            (call.call_id.clone(), contract)
                         })
                         .collect(),
                 ),

server/src/cursor/checkpoint/messages/tests.rs

@@ -30,6 +30,7 @@ fn pending_tool_round_is_one_complete_assistant_message_and_round_trips() {
             name: "Read".into(),
             arguments_text: r#"{"path":"/a"}"#.into(),
             arguments: json!({"path":"/a"}),
+            argument_error: Some("Read arguments are not valid JSON".into()),
         },
         ToolCall {
             index: 1,
@@ -38,6 +39,7 @@ fn pending_tool_round_is_one_complete_assistant_message_and_round_trips() {
             name: "Grep".into(),
             arguments_text: r#"{"pattern":"x"}"#.into(),
             arguments: json!({"pattern":"x"}),
+            argument_error: None,
         },
     ];
     let pending = staged_tool_round(
@@ -55,6 +57,10 @@ fn pending_tool_round_is_one_complete_assistant_message_and_round_trips() {
         wire["providerOptions"]["cursor"]["pendingToolExecutionContracts"]["a"]["toolIdentifier"],
         "READ"
     );
+    assert_eq!(
+        wire["providerOptions"]["cursor"]["pendingToolExecutionContracts"]["a"]["argumentError"],
+        "Read arguments are not valid JSON"
+    );
     assert_eq!(wire["role"], "assistant");
     assert_eq!(
         wire["providerOptions"]["cursor"]["pendingToolExecutionContracts"]
@@ -77,6 +83,10 @@ fn pending_tool_round_is_one_complete_assistant_message_and_round_trips() {
     assert_eq!(recovered.assistant.replay_state, Some(replay_state));
     assert_eq!(recovered.calls.len(), 2);
     assert_eq!(recovered.calls[0].call_id, "a");
+    assert_eq!(
+        recovered.calls[0].argument_error.as_deref(),
+        Some("Read arguments are not valid JSON")
+    );
     assert_eq!(recovered.calls[1].call_id, "b");
 }
 

server/src/cursor/checkpoint/steps.rs

@@ -75,6 +75,11 @@ impl StepBuffer {
         });
     }
 
+    pub fn finish_model_attempt(&mut self) {
+        self.finish_text();
+        self.finish_thinking(Duration::ZERO);
+    }
+
     pub fn discard_model_output(&mut self) {
         self.text.clear();
         self.thinking.clear();
@@ -101,6 +106,28 @@ impl StepBuffer {
 mod tests {
     use super::*;
 
+    #[test]
+    fn failed_attempt_output_is_retained_for_the_next_checkpoint() {
+        let mut buffer = StepBuffer::default();
+        buffer.text_delta("partial answer");
+        buffer.thinking_delta("partial reasoning");
+
+        buffer.finish_model_attempt();
+
+        let steps = buffer.take().steps;
+        assert_eq!(steps.len(), 2);
+        assert!(matches!(
+            &steps[0].message,
+            Some(pb::conversation_step::Message::AssistantMessage(message))
+                if message.text == "partial answer"
+        ));
+        assert!(matches!(
+            &steps[1].message,
+            Some(pb::conversation_step::Message::ThinkingMessage(message))
+                if message.text == "partial reasoning"
+        ));
+    }
+
     #[test]
     fn interrupted_model_output_is_not_persisted_as_checkpoint_steps() {
         let mut buffer = StepBuffer::default();

server/src/cursor/conversation/output.rs

@@ -21,7 +21,7 @@ use crate::{
         protocol::proto::agent::v1 as pb,
         services::blob_sync::BlobSynchronizer,
         tools::{
-            codec,
+            codec, compat,
             runtime::CursorToolRuntime,
             stream::ToolCallStream,
             tool_call_result::{ToolCompletion, ToolResultReceiver},
@@ -264,6 +264,32 @@ impl ConversationOutput {
                         streams.clear();
                         presentation.discard_model_output();
                     }
+                    RunEvent::ModelAttemptFailed { attempt, message } => {
+                        tracing::warn!(
+                            run_id = %self.run.run_id(),
+                            attempt,
+                            %message,
+                            "retrying model call from current checkpoint"
+                        );
+                        presentation.finish_model_attempt();
+                        for call in calls.values_mut() {
+                            if call.arguments.is_null() {
+                                call.arguments = serde_json::from_str(&call.arguments_text)
+                                    .unwrap_or_else(|_| serde_json::json!({}));
+                            }
+                            let completion = compat::failure_with_message(
+                                call,
+                                format!("Model attempt failed before tool completion: {message}"),
+                            );
+                            self.handle
+                                .emit(&codec::tool_completed(call, &completion))?;
+                            presentation.tool_completed(&completion);
+                        }
+                        response_text.clear();
+                        response_thinking.clear();
+                        calls.clear();
+                        streams.clear();
+                    }
                     RunEvent::TextStart => {}
                     RunEvent::TextEnd => {
                         if !self.context.compacting {
@@ -312,6 +338,7 @@ impl ConversationOutput {
                             name: name.clone(),
                             arguments_text: String::new(),
                             arguments: serde_json::Value::Null,
+                            argument_error: None,
                         };
                         self.emit_model_event(
                             crate::provider::ModelEvent::ToolCallStart {
@@ -335,10 +362,29 @@ impl ConversationOutput {
                         let stream = streams.get_mut(&index).ok_or_else(|| {
                             Error::Protocol(format!("missing Cursor tool stream: {index}"))
                         })?;
-                        for message in stream.arguments_delta(call, &delta)? {
+                        match stream.arguments_delta(call, &delta) {
+                            Ok(messages) => {
+                                for message in messages {
                                     self.handle.emit(&message)?;
                                 }
                             }
+                            Err(Error::Protocol(message)) => {
+                                tracing::warn!(
+                                    call_id = %call.call_id,
+                                    %message,
+                                    "ignoring invalid streaming tool arguments until completion"
+                                );
+                            }
+                            Err(Error::Json(error)) => {
+                                tracing::warn!(
+                                    call_id = %call.call_id,
+                                    %error,
+                                    "ignoring invalid streaming tool arguments until completion"
+                                );
+                            }
+                            Err(error) => return Err(error),
+                        }
+                    }
                     RunEvent::ToolCallEnd { index } => {
                         let call = calls.get_mut(&index).ok_or_else(|| {
                             Error::Protocol(format!("unknown completed tool index: {index}"))
@@ -349,7 +395,8 @@ impl ConversationOutput {
                         call.arguments = if call.arguments_text.trim().is_empty() {
                             serde_json::json!({})
                         } else {
-                            serde_json::from_str(&call.arguments_text)?
+                            serde_json::from_str(&call.arguments_text)
+                                .unwrap_or_else(|_| serde_json::json!({}))
                         };
                     }
                     RunEvent::Usage(usage) => {
@@ -359,10 +406,7 @@ impl ConversationOutput {
                             }
                         }
                         if !self.context.compacting {
-                            context_tokens = usage
-                                .input_tokens
-                                .zip(usage.output_tokens)
-                                .and_then(|(input, output)| input.checked_add(output));
+                            context_tokens = usage.context_input_tokens;
                         }
                         match &mut turn_usage {
                             Some(total) => *total += usage,
@@ -523,6 +567,7 @@ impl ConversationOutput {
                                 .map_err(|_| Error::Protocol("checkpoint worker stopped".into()))?
                             {
                                 Ok(checkpoint) => {
+                                    context_tokens = checkpoint_context_tokens(&checkpoint);
                                     compaction_checkpoint = Some(checkpoint);
                                     state.barrier.complete(Ok(()));
                                 }
@@ -1029,10 +1074,34 @@ pub(crate) fn finish_cancelled(handle: &TransportHandle) -> Result<()> {
     Ok(())
 }
 
+fn checkpoint_context_tokens(checkpoint: &pb::ConversationStateStructure) -> Option<u64> {
+    checkpoint
+        .token_details
+        .as_ref()
+        .map(|details| u64::from(details.used_tokens))
+}
+
 #[cfg(test)]
 mod tests {
-    use super::accept_tool_completion;
-    use crate::{run::CommandResult, Error};
+    use super::{accept_tool_completion, checkpoint_context_tokens};
+    use crate::{cursor::protocol::proto::agent::v1 as pb, run::CommandResult, Error};
+
+    #[test]
+    fn compacted_checkpoint_replaces_the_in_memory_context_usage() {
+        let compacted = pb::ConversationStateStructure {
+            token_details: Some(pb::ConversationTokenDetails {
+                used_tokens: 20_000,
+                ..Default::default()
+            }),
+            ..Default::default()
+        };
+
+        assert_eq!(checkpoint_context_tokens(&compacted), Some(20_000));
+        assert_eq!(
+            checkpoint_context_tokens(&pb::ConversationStateStructure::default()),
+            None
+        );
+    }
 
     #[test]
     fn closing_and_ended_runs_ignore_known_tool_completions() {

server/src/cursor/conversation/runtime.rs

@@ -11,7 +11,7 @@ use crate::{
         protocol::proto::agent::v1 as pb,
         services::{blob_sync::BlobSynchronizer, context_sync::RequestContextSynchronizer},
         tools::{
-            codec,
+            codec, compat,
             runtime::CursorToolRuntime,
             tool_call_result::{tool_result_channel, ToolResultReceiver, ToolResultSender},
             ClientToolEvent, ToolDispatcher,
@@ -262,17 +262,18 @@ impl ConversationRuntime {
                                                     .take_exec(throw.id)
                                                     .await
                                                 {
-                                                    Some(pending) => generation.results.send_error(
-                                                        crate::Error::Protocol(format!(
+                                                    Some(pending) => generation.results.send(
+                                                        compat::failure_with_message(
+                                                            &pending.call,
+                                                            format!(
                                                                 "Exec {} failed: {}",
                                                                 pending.call.call_id, throw.error
-                                                        )),
                                                             ),
-                                                    None => generation.results.send_error(
-                                                        crate::Error::Protocol(format!(
-                                                            "unknown ExecClientThrow id: {}",
-                                                            throw.id
-                                                        )),
+                                                        ),
+                                                    ),
+                                                    None => tracing::warn!(
+                                                        id = throw.id,
+                                                        "ignoring failure for unknown tool execution"
                                                     ),
                                                 }
                                             }

server/src/cursor/tools/codec/request.rs

@@ -542,6 +542,7 @@ mod tests {
             name: "Bash".into(),
             arguments_text: String::new(),
             arguments: json!({ "command": "ls -la" }),
+            argument_error: None,
         };
         let message = request(1, &call, &ExecContext::default()).unwrap();
         let Some(pb::agent_server_message::Message::ExecServerMessage(exec)) = message.message

server/src/cursor/tools/codec/response.rs

@@ -3,7 +3,7 @@ use crate::{
     cursor::{
         protocol::{events, proto::agent::v1 as pb},
         tools::{
-            edit,
+            compat, edit,
             runtime::{CursorToolRuntime, ExecStage, PendingExec},
             tool_call_result::{self as result, ToolCompletion},
         },
@@ -34,16 +34,15 @@ pub async fn client_event(
     let call = match pending.exec_call(message.id).await {
         Some(call) => call,
         None if pending.completed_call(message.id).await.is_some() => {
-            return Err(Error::Protocol(format!(
-                "duplicate terminal ExecClientMessage id: {}",
-                message.id
-            )))
+            tracing::warn!(id = message.id, "ignoring duplicate terminal tool response");
+            return Ok(ClientExecEvent::Pending);
         }
         None => {
-            return Err(Error::Protocol(format!(
-                "unknown ExecClientMessage id: {}",
-                message.id
-            )))
+            tracing::warn!(
+                id = message.id,
+                "ignoring response for unknown tool execution"
+            );
+            return Ok(ClientExecEvent::Pending);
         }
     };
     let Some(wire_result) = &message.message else {
@@ -174,9 +173,22 @@ pub async fn stream_closed(id: u32, pending: &CursorToolRuntime) -> Result<Optio
     }
     let rendered = match &entry.stage {
         ExecStage::DynamicMcp(definition) => {
-            super::render_dynamic_mcp(&entry.call, definition, false)
+            Ok(super::render_dynamic_mcp(&entry.call, definition, false))
         }
-        _ => super::render_tool_call(&entry.call, false)?,
+        _ => super::render_tool_call(&entry.call, false),
+    };
+    let rendered = match rendered {
+        Ok(rendered) => rendered,
+        Err(Error::Protocol(message)) => {
+            return Ok(Some(compat::failure_with_message(&entry.call, message)));
+        }
+        Err(Error::Json(error)) => {
+            return Ok(Some(compat::failure_with_message(
+                &entry.call,
+                error.to_string(),
+            )));
+        }
+        Err(error) => return Err(error),
     };
     Ok(Some(ToolCompletion::from_rendered(
         &entry.call,
@@ -212,10 +224,10 @@ async fn advance_edit(
         pb::exec_client_message::Message::ReadResult(result)
         | pb::exec_client_message::Message::RedactedReadResult(result) => result,
         _ => {
-            return Err(Error::Protocol(format!(
-                "expected ReadResult for edit tool {}",
-                entry.call.name
-            )))
+            let message = format!("expected ReadResult for edit tool {}", entry.call.name);
+            return Ok(ClientExecEvent::Completed(Box::new(
+                compat::failure_with_message(&entry.call, message),
+            )));
         }
     };
     let write = match edit::after_read(&entry.call, read) {
@@ -260,9 +272,14 @@ fn completed(
     pending: PendingExec,
     result: pb::exec_client_message::Message,
 ) -> Result<ClientExecEvent> {
-    Ok(ClientExecEvent::Completed(Box::new(result::from_exec(
-        pending, &result,
-    )?)))
+    let call = pending.call.clone();
+    let completion = match result::from_exec(pending, &result) {
+        Ok(completion) => completion,
+        Err(Error::Protocol(message)) => compat::failure_with_message(&call, message),
+        Err(Error::Json(error)) => compat::failure_with_message(&call, error.to_string()),
+        Err(error) => return Err(error),
+    };
+    Ok(ClientExecEvent::Completed(Box::new(completion)))
 }
 
 fn shell_exit_result(

server/src/cursor/tools/compat.rs

@@ -49,7 +49,10 @@ pub(crate) fn render(call: &ToolCall, completed: bool) -> pb::ToolCall {
 }
 
 pub(crate) fn failure(call: &ToolCall) -> ToolCompletion {
-    let error = failure_message(&call.name);
+    failure_with_message(call, failure_message(&call.name))
+}
+
+pub(crate) fn failure_with_message(call: &ToolCall, error: String) -> ToolCompletion {
     let arguments = call
         .arguments
         .as_object()

server/src/cursor/tools/edit.rs

@@ -265,6 +265,7 @@ mod tests {
                 "old_string": old_string,
                 "new_string": "replacement",
             }),
+            argument_error: None,
         }
     }
 

server/src/cursor/tools/mod.rs

@@ -103,10 +103,20 @@ impl ToolDispatcher {
             }
             let message_index = first_tool_index + position;
             let publish_started = !state.started.contains(&call.call_id);
+            if let Some(error) = &call.argument_error {
+                dispatched.push(validation_failure(call, error.clone()));
+                continue;
+            }
             let edit_path = if dynamic_mcp.contains_key(&call.name) {
                 None
             } else {
-                edit::execution_path(call)?
+                match edit::execution_path(call) {
+                    Ok(path) => path,
+                    Err(error) => {
+                        dispatched.push(recover_validation_failure(call, error)?);
+                        continue;
+                    }
+                }
             };
             if let Some(path) = edit_path {
                 let next = self.edit_schedule.lock().await.start_or_defer(
@@ -121,22 +131,28 @@ impl ToolDispatcher {
                 let Some(next) = next else {
                     continue;
                 };
-                dispatched.push(
-                    self.start(
+                let started = self
+                    .start(
                         &next.call,
                         next.message_index,
                         next.publish_started,
                         dynamic_mcp,
                         &next.context,
                     )
-                    .await?,
-                );
+                    .await;
+                dispatched.push(match started {
+                    Ok(started) => started,
+                    Err(error) => recover_validation_failure(&next.call, error)?,
+                });
                 continue;
             }
-            dispatched.push(
-                self.start(call, message_index, publish_started, dynamic_mcp, context)
-                    .await?,
-            );
+            let started = self
+                .start(call, message_index, publish_started, dynamic_mcp, context)
+                .await;
+            dispatched.push(match started {
+                Ok(started) => started,
+                Err(error) => recover_validation_failure(call, error)?,
+            });
         }
         Ok(dispatched)
     }
@@ -146,7 +162,8 @@ impl ToolDispatcher {
         let Some(next) = next else {
             return Ok(None);
         };
-        self.start(
+        match self
+            .start(
                 &next.call,
                 next.message_index,
                 next.publish_started,
@@ -154,7 +171,10 @@ impl ToolDispatcher {
                 &next.context,
             )
             .await
-        .map(Some)
+        {
+            Ok(started) => Ok(Some(started)),
+            Err(error) => recover_validation_failure(&next.call, error).map(Some),
+        }
     }
 
     pub async fn interrupt_for_message(&self) -> Vec<u32> {
@@ -203,34 +223,64 @@ impl ToolDispatcher {
         let pending = match self.runtime.take_interaction(response.id).await {
             Some(pending) => pending,
             None if self.runtime.completed_call(response.id).await.is_some() => {
-                return Err(Error::Protocol(format!(
-                    "duplicate terminal InteractionResponse id: {}",
-                    response.id
-                )));
+                tracing::warn!(
+                    id = response.id,
+                    "ignoring duplicate terminal interaction response"
+                );
+                return Ok(ClientToolEvent::Pending);
             }
             None => {
-                return Err(Error::Protocol(format!(
-                    "unknown InteractionResponse id: {}",
-                    response.id
-                )));
+                tracing::warn!(
+                    id = response.id,
+                    "ignoring response for unknown interaction"
+                );
+                return Ok(ClientToolEvent::Pending);
             }
         };
-        Ok(
-            match tool_call_dispatch::resume_interaction(
+        let call = pending.call.clone();
+        let continuation = match tool_call_dispatch::resume_interaction(
             &self.results,
             &self.search,
             &self.fetch,
             pending,
             response,
         )
-            .await?
+        .await
         {
+            Ok(continuation) => continuation,
+            Err(Error::Protocol(message)) => {
+                return Ok(ClientToolEvent::Completed(Box::new(
+                    compat::failure_with_message(&call, message),
+                )));
+            }
+            Err(Error::Json(error)) => {
+                return Ok(ClientToolEvent::Completed(Box::new(
+                    compat::failure_with_message(&call, error.to_string()),
+                )));
+            }
+            Err(error) => return Err(error),
+        };
+        Ok(match continuation {
             tool_call_dispatch::InteractionContinuation::Completed(completion) => {
                 ClientToolEvent::Completed(completion)
             }
             tool_call_dispatch::InteractionContinuation::Pending => ClientToolEvent::Pending,
-            },
-        )
+        })
+    }
+}
+
+fn validation_failure(call: &ToolCall, message: String) -> DispatchedTool {
+    DispatchedTool {
+        messages: Vec::new(),
+        completion: Some(compat::failure_with_message(call, message)),
+    }
+}
+
+fn recover_validation_failure(call: &ToolCall, error: Error) -> Result<DispatchedTool> {
+    match error {
+        Error::Protocol(message) => Ok(validation_failure(call, message)),
+        Error::Json(error) => Ok(validation_failure(call, error.to_string())),
+        error => Err(error),
     }
 }
 

server/src/cursor/tools/stream.rs

@@ -248,6 +248,7 @@ mod tests {
             name: "Task".into(),
             arguments_text: arguments_text.into(),
             arguments: serde_json::Value::Null,
+            argument_error: None,
         }
     }
 

server/src/cursor/tools/tool_call_result/interaction.rs

@@ -406,6 +406,7 @@ mod tests {
                 name: "WebFetch".into(),
                 arguments_text: r#"{"url":"https://example.com"}"#.into(),
                 arguments: json!({"url": "https://example.com"}),
+                argument_error: None,
             },
             started_at_ms: 1,
         }

server/src/model/observability.rs

@@ -9,6 +9,7 @@ mod usage {
     #[derive(Clone, Copy, Debug, Default, Serialize, Deserialize, PartialEq, Eq)]
     pub struct Usage {
         pub input_tokens: Option<u64>,
+        pub context_input_tokens: Option<u64>,
         pub output_tokens: Option<u64>,
         pub total_tokens: Option<u64>,
         pub cache_read_tokens: Option<u64>,
@@ -19,6 +20,7 @@ mod usage {
     impl AddAssign for Usage {
         fn add_assign(&mut self, rhs: Self) {
             self.input_tokens = sum(self.input_tokens, rhs.input_tokens);
+            self.context_input_tokens = sum(self.context_input_tokens, rhs.context_input_tokens);
             self.output_tokens = sum(self.output_tokens, rhs.output_tokens);
             self.total_tokens = sum(self.total_tokens, rhs.total_tokens);
             self.cache_read_tokens = sum(self.cache_read_tokens, rhs.cache_read_tokens);

server/src/model/tool.rs

@@ -19,6 +19,7 @@ pub struct ToolCall {
     pub name: String,
     pub arguments_text: String,
     pub arguments: Value,
+    pub argument_error: Option<String>,
 }
 
 #[derive(Clone, Debug, Serialize, Deserialize, PartialEq, Eq)]

server/src/plugin/wire.rs

@@ -147,8 +147,10 @@ pub fn model_event(value: &serde_json::Value) -> Result<ModelEvent> {
                 .get("usage")
                 .ok_or_else(|| Error::Protocol("plugin usage event requires usage".into()))?;
             let tokens = |name: &str| usage.get(name).and_then(serde_json::Value::as_u64);
+            let input_tokens = tokens("inputTokens");
             ModelEvent::Usage(Usage {
-                input_tokens: tokens("inputTokens"),
+                input_tokens,
+                context_input_tokens: input_tokens,
                 output_tokens: tokens("outputTokens"),
                 total_tokens: tokens("totalTokens"),
                 cache_read_tokens: tokens("cacheReadTokens"),
@@ -285,6 +287,7 @@ mod tests {
             usage,
             ModelEvent::Usage(Usage {
                 input_tokens: Some(10),
+                context_input_tokens: Some(10),
                 output_tokens: Some(2),
                 total_tokens: None,
                 cache_read_tokens: Some(4),

server/src/provider/anthropic.rs

@@ -12,9 +12,9 @@ use crate::{
 };
 
 use super::{
+    attempt::{send_once, Attempt},
     map_sse_error, merge_extra_params, provider_event_error,
     recorder::recorded_headers,
-    retry::{send_with_retry, Attempt, RetryPolicy},
     CallRecorder, FinishReason, ModelEvent, Provider, ProviderStream,
 };
 
@@ -90,17 +90,14 @@ impl Provider for AnthropicProvider {
             if let Some(recorder) = &recorder {
                 recorder.request(request_headers.clone(), &body).await?;
             }
-            let attempt = send_with_retry(
+            let attempt = send_once(
                 "Anthropic",
                 || client.post(&config.request_url)
                     .header("x-api-key", &config.api_key).header("anthropic-version", "2023-06-01")
                     .headers(config.custom_headers.clone())
                     .json(&body),
-                RetryPolicy { retries: config.retry_count, ..RetryPolicy::default() },
                 &cancellation,
                 recorder.as_ref(),
-                request_headers,
-                &body,
             ).await?;
             let Attempt::Response(response) = attempt else { return };
             yield ModelEvent::Start { model_call_id: call_id };
@@ -321,6 +318,7 @@ fn apply_model(body: &mut Value, model: &crate::model::ModelSpec) -> Result<()>
 
 fn merge_usage(total: &mut Usage, update: Usage) {
     merge_usage_field(&mut total.input_tokens, update.input_tokens);
+    merge_usage_field(&mut total.context_input_tokens, update.context_input_tokens);
     merge_usage_field(&mut total.output_tokens, update.output_tokens);
     merge_usage_field(&mut total.cache_read_tokens, update.cache_read_tokens);
     merge_usage_field(&mut total.cache_write_tokens, update.cache_write_tokens);
@@ -475,14 +473,42 @@ fn required_u64(value: &Value, name: &str) -> Result<u64> {
 }
 
 fn anthropic_usage(value: &Value) -> Usage {
+    let input_tokens = value.get("input_tokens").and_then(Value::as_u64);
+    let cache_read_tokens = value.get("cache_read_input_tokens").and_then(Value::as_u64);
+    let cache_write_tokens = value
+        .get("cache_creation_input_tokens")
+        .and_then(Value::as_u64);
+    let context_input_tokens = input_tokens.map(|input| {
+        input
+            .saturating_add(cache_read_tokens.unwrap_or_default())
+            .saturating_add(cache_write_tokens.unwrap_or_default())
+    });
     Usage {
-        input_tokens: value.get("input_tokens").and_then(Value::as_u64),
+        input_tokens,
+        context_input_tokens,
         output_tokens: value.get("output_tokens").and_then(Value::as_u64),
         total_tokens: value.get("total_tokens").and_then(Value::as_u64),
-        cache_read_tokens: value.get("cache_read_input_tokens").and_then(Value::as_u64),
-        cache_write_tokens: value
-            .get("cache_creation_input_tokens")
-            .and_then(Value::as_u64),
+        cache_read_tokens,
+        cache_write_tokens,
         reasoning_tokens: None,
     }
 }
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn cached_tokens_are_included_once_in_anthropic_context_input() {
+        let usage = anthropic_usage(&serde_json::json!({
+            "input_tokens": 10,
+            "output_tokens": 5,
+            "cache_read_input_tokens": 20,
+            "cache_creation_input_tokens": 30
+        }));
+
+        assert_eq!(usage.input_tokens, Some(10));
+        assert_eq!(usage.context_input_tokens, Some(60));
+        assert_eq!(usage.output_tokens, Some(5));
+    }
+}

server/src/provider/attempt.rs

@@ -0,0 +1,100 @@
+//! Sends one Provider HTTP attempt without applying retry policy.
+
+use tokio_util::sync::CancellationToken;
+
+use crate::{Error, Result};
+
+use super::CallRecorder;
+
+#[derive(Debug)]
+pub(crate) enum Attempt {
+    Response(reqwest::Response),
+    Cancelled,
+}
+
+pub(crate) async fn send_once<F>(
+    label: &str,
+    build: F,
+    cancellation: &CancellationToken,
+    recorder: Option<&CallRecorder>,
+) -> Result<Attempt>
+where
+    F: FnOnce() -> reqwest::RequestBuilder,
+{
+    let response = tokio::select! {
+        _ = cancellation.cancelled() => return Ok(Attempt::Cancelled),
+        response = build().send() => response,
+    }?;
+    if let Some(recorder) = recorder {
+        recorder
+            .response_headers(response.status().as_u16())
+            .await?;
+    }
+    if response.status().is_success() {
+        return Ok(Attempt::Response(response));
+    }
+    let status = response.status();
+    let bytes = tokio::select! {
+        _ = cancellation.cancelled() => return Ok(Attempt::Cancelled),
+        bytes = response.bytes() => bytes,
+    }?;
+    Err(Error::Provider(format!(
+        "{label} {status}: {}",
+        String::from_utf8_lossy(&bytes)
+    )))
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use tokio::io::{AsyncReadExt, AsyncWriteExt};
+
+    async fn server(response: &'static [u8]) -> String {
+        let listener = tokio::net::TcpListener::bind("127.0.0.1:0").await.unwrap();
+        let address = listener.local_addr().unwrap();
+        tokio::spawn(async move {
+            let (mut socket, _) = listener.accept().await.unwrap();
+            let mut request = [0_u8; 1024];
+            let _ = socket.read(&mut request).await;
+            socket.write_all(response).await.unwrap();
+        });
+        format!("http://{address}")
+    }
+
+    #[tokio::test]
+    async fn non_success_status_is_one_failed_attempt() {
+        let url =
+            server(b"HTTP/1.1 503 Service Unavailable\r\nContent-Length: 4\r\n\r\ndown").await;
+        let client = reqwest::Client::new();
+        let error = send_once("test", || client.get(&url), &CancellationToken::new(), None)
+            .await
+            .unwrap_err();
+        assert!(
+            matches!(error, Error::Provider(message) if message.contains("503") && message.contains("down"))
+        );
+    }
+
+    #[tokio::test]
+    async fn response_body_transport_failure_is_one_failed_attempt() {
+        let url =
+            server(b"HTTP/1.1 500 Internal Server Error\r\nContent-Length: 100\r\n\r\nshort").await;
+        let client = reqwest::Client::new();
+        let error = send_once("test", || client.get(&url), &CancellationToken::new(), None)
+            .await
+            .unwrap_err();
+        assert!(matches!(error, Error::Http(_)));
+    }
+
+    #[tokio::test]
+    async fn request_transport_failure_is_one_failed_attempt() {
+        let listener = tokio::net::TcpListener::bind("127.0.0.1:0").await.unwrap();
+        let address = listener.local_addr().unwrap();
+        drop(listener);
+        let url = format!("http://{address}");
+        let client = reqwest::Client::new();
+        let error = send_once("test", || client.get(&url), &CancellationToken::new(), None)
+            .await
+            .unwrap_err();
+        assert!(matches!(error, Error::Http(_)));
+    }
+}

server/src/provider/mod.rs

@@ -1,11 +1,11 @@
 //! Defines the provider interface and exports provider implementations.
 mod anthropic;
+mod attempt;
 mod event;
 mod normalize;
 mod openai_chat;
 mod openai_responses;
 mod recorder;
-mod retry;
 mod router;
 
 use std::pin::Pin;

server/src/provider/openai_chat.rs

@@ -17,10 +17,10 @@ use crate::{
 };
 
 use super::{
-    apply_body_allowlist, apply_openai_prompt_cache_key, map_sse_error, merge_extra_params,
-    provider_event_error,
+    apply_body_allowlist, apply_openai_prompt_cache_key,
+    attempt::{send_once, Attempt},
+    map_sse_error, merge_extra_params, provider_event_error,
     recorder::recorded_headers,
-    retry::{send_with_retry, Attempt, RetryPolicy},
     CallRecorder, FinishReason, ModelEvent, Provider, ProviderStream,
 };
 
@@ -92,15 +92,12 @@ impl Provider for OpenAiChatProvider {
             if let Some(recorder) = &recorder {
                 recorder.request(request_headers.clone(), &body).await?;
             }
-            let attempt = send_with_retry(
+            let attempt = send_once(
                 "OpenAI Chat",
                 || client.post(&config.request_url)
                     .bearer_auth(&config.api_key).headers(config.custom_headers.clone()).json(&body),
-                RetryPolicy { retries: config.retry_count, ..RetryPolicy::default() },
                 &cancellation,
                 recorder.as_ref(),
-                request_headers,
-                &body,
             ).await?;
             let Attempt::Response(response) = attempt else { return };
             yield ModelEvent::Start { model_call_id: call_id };
@@ -421,8 +418,10 @@ fn merge_chat_fragment(target: &mut String, fragment: &str) {
 }
 
 pub(crate) fn openai_usage(value: &Value) -> Usage {
+    let input_tokens = value.get("prompt_tokens").and_then(Value::as_u64);
     Usage {
-        input_tokens: value.get("prompt_tokens").and_then(Value::as_u64),
+        input_tokens,
+        context_input_tokens: input_tokens,
         output_tokens: value.get("completion_tokens").and_then(Value::as_u64),
         total_tokens: value.get("total_tokens").and_then(Value::as_u64),
         cache_read_tokens: value

server/src/provider/openai_responses.rs

@@ -14,10 +14,10 @@ use crate::{
 };
 
 use super::{
-    apply_body_allowlist, apply_openai_prompt_cache_key, map_sse_error, merge_extra_params,
-    provider_event_error,
+    apply_body_allowlist, apply_openai_prompt_cache_key,
+    attempt::{send_once, Attempt},
+    map_sse_error, merge_extra_params, provider_event_error,
     recorder::recorded_headers,
-    retry::{send_with_retry, Attempt, RetryPolicy},
     CallRecorder, FinishReason, ModelEvent, Provider, ProviderStream,
 };
 
@@ -89,15 +89,12 @@ impl Provider for OpenAiResponsesProvider {
             if let Some(recorder) = &recorder {
                 recorder.request(request_headers.clone(), &body).await?;
             }
-            let attempt = send_with_retry(
+            let attempt = send_once(
                 "OpenAI Responses",
                 || client.post(&config.request_url)
                     .bearer_auth(&config.api_key).headers(config.custom_headers.clone()).json(&body),
-                RetryPolicy { retries: config.retry_count, ..RetryPolicy::default() },
                 &cancellation,
                 recorder.as_ref(),
-                request_headers,
-                &body,
             ).await?;
             let Attempt::Response(response) = attempt else { return };
             yield ModelEvent::Start { model_call_id: call_id };
@@ -505,8 +502,10 @@ fn required_u64(value: &Value, name: &str) -> Result<u64> {
 }
 
 fn responses_usage(value: &Value) -> Usage {
+    let input_tokens = value.get("input_tokens").and_then(Value::as_u64);
     Usage {
-        input_tokens: value.get("input_tokens").and_then(Value::as_u64),
+        input_tokens,
+        context_input_tokens: input_tokens,
         output_tokens: value.get("output_tokens").and_then(Value::as_u64),
         total_tokens: value.get("total_tokens").and_then(Value::as_u64),
         cache_read_tokens: value

server/src/provider/recorder.rs

@@ -1,7 +1,7 @@
 //! Records provider requests, responses, usage, and timing.
 use std::{
     sync::{
-        atomic::{AtomicBool, AtomicI64, AtomicU32, AtomicU64, Ordering},
+        atomic::{AtomicBool, AtomicI64, AtomicU64, Ordering},
         Arc,
     },
     time::Instant,
@@ -71,7 +71,6 @@ struct Inner {
     base_call: NewLlmCall,
     detailed: bool,
     attempt: Mutex<AttemptState>,
-    next_attempt: AtomicU32,
     next_generation: AtomicU64,
     finished: AtomicBool,
 }
@@ -120,7 +119,6 @@ impl CallRecorder {
                 base_call: call.clone(),
                 detailed: call.detailed,
                 attempt: Mutex::new(AttemptState::new(call.call_id.clone())),
-                next_attempt: AtomicU32::new(0),
                 next_generation: AtomicU64::new(0),
                 finished: AtomicBool::new(false),
             }),
@@ -284,34 +282,6 @@ impl CallRecorder {
         self.finish("cancelled", None, None, None).await
     }
 
-    pub async fn retry(
-        &self,
-        error: &crate::Error,
-        headers: serde_json::Value,
-        body: &serde_json::Value,
-    ) -> Result<()> {
-        self.failed(error).await?;
-
-        let attempt_number = self.inner.next_attempt.fetch_add(1, Ordering::Relaxed) + 1;
-        let mut call = self.inner.base_call.clone();
-        call.call_id = format!("{}:retry-{attempt_number}", self.inner.base_call.call_id);
-        {
-            let mut attempt = self.inner.attempt.lock().await;
-            *attempt = AttemptState::new(call.call_id.clone());
-            self.inner.finished.store(false, Ordering::Release);
-            if let Err(error) = self.inner.store.start_llm_call(&call).await {
-                self.inner.finished.store(true, Ordering::Release);
-                return Err(error);
-            }
-        }
-
-        if let Err(error) = self.request(headers, body).await {
-            self.failed(&error).await?;
-            return Err(error);
-        }
-        Ok(())
-    }
-
     async fn finish(
         &self,
         status: &str,

server/src/provider/retry.rs

@@ -1,84 +0,0 @@
-//! Applies provider retry and backoff behavior.
-use std::time::Duration;
-
-use tokio_util::sync::CancellationToken;
-
-use crate::{Error, Result};
-
-use super::CallRecorder;
-
-#[derive(Clone, Copy, Debug)]
-pub(crate) struct RetryPolicy {
-    pub retries: u32,
-    pub delay: Duration,
-}
-
-impl Default for RetryPolicy {
-    fn default() -> Self {
-        Self {
-            retries: 5,
-            delay: Duration::from_secs(5),
-        }
-    }
-}
-
-#[derive(Debug)]
-pub(crate) enum Attempt {
-    Response(reqwest::Response),
-    Cancelled,
-}
-
-pub(crate) async fn send_with_retry<F>(
-    label: &str,
-    build: F,
-    policy: RetryPolicy,
-    cancellation: &CancellationToken,
-    recorder: Option<&CallRecorder>,
-    request_headers: serde_json::Value,
-    request_body: &serde_json::Value,
-) -> Result<Attempt>
-where
-    F: Fn() -> reqwest::RequestBuilder,
-{
-    for attempt in 0..=policy.retries {
-        let response = tokio::select! {
-            _ = cancellation.cancelled() => return Ok(Attempt::Cancelled),
-            response = build().send() => response,
-        }?;
-        if let Some(recorder) = recorder {
-            recorder
-                .response_headers(response.status().as_u16())
-                .await?;
-        }
-        if response.status().is_success() {
-            return Ok(Attempt::Response(response));
-        }
-        let status = response.status();
-        let bytes = response.bytes().await?;
-        let error = Error::Provider(format!(
-            "{label} {status}: {}",
-            String::from_utf8_lossy(&bytes)
-        ));
-        if attempt == policy.retries {
-            return Err(error);
-        }
-        tracing::warn!(
-            provider = label,
-            status = status.as_u16(),
-            attempt = attempt + 1,
-            retries = policy.retries,
-            delay_ms = policy.delay.as_millis(),
-            "provider returned a non-success status, retrying"
-        );
-        if let Some(recorder) = recorder {
-            recorder
-                .retry(&error, request_headers.clone(), request_body)
-                .await?;
-        }
-        tokio::select! {
-            _ = cancellation.cancelled() => return Ok(Attempt::Cancelled),
-            _ = tokio::time::sleep(policy.delay) => {}
-        }
-    }
-    unreachable!("the retry loop returns on the final attempt")
-}

server/src/provider/router.rs

@@ -18,8 +18,6 @@ use super::{
     OpenAiChatProvider, OpenAiResponsesProvider, Provider, ProviderStream,
 };
 
-const BUILTIN_PROVIDER_RETRIES: u32 = 5;
-
 pub struct ProviderRouter {
     store: Store,
     plugins: PluginRegistry,
@@ -91,7 +89,6 @@ impl Provider for ProviderRouter {
                         custom_headers: if model.custom_headers_enabled { custom_headers(&model.custom_headers)? } else { reqwest::header::HeaderMap::new() },
                         max_output_tokens: model.max_output_tokens(),
                         request_timeout,
-                        retry_count: BUILTIN_PROVIDER_RETRIES,
                         allowed_body_fields: None,
                     };
                     let client = crate::network::client_builder(&store).await?.timeout(request_timeout).build()?;

server/src/run/engine.rs

@@ -14,8 +14,10 @@ use crate::{
 };
 
 use super::{
-    consume_model_cycle, CommitBarrier, CommitCause, MessagesCommitted, ModelCycleFailure,
-    RunCommand, RunEvent, RunFailure, RunOutcome, RunPort,
+    consume_model_cycle,
+    model_retry::{should_retry, MODEL_RETRY_DELAY},
+    CommitBarrier, CommitCause, MessagesCommitted, RunCommand, RunEvent, RunFailure, RunOutcome,
+    RunPort,
 };
 
 pub struct RunEngine {
@@ -208,12 +210,20 @@ impl RunEngine {
                 model: prepared.model.clone(),
                 history,
             };
+            let mut retries = 0_u32;
+            let mut pending_insertions = Vec::new();
+            let cycle = 'attempt: loop {
+                let call_id = if retries == 0 {
+                    format!("{}:{provider_call_index}", prepared.run_id)
+                } else {
+                    format!("{}:{provider_call_index}:retry-{retries}", prepared.run_id)
+                };
                 let invocation = crate::model::ModelInvocation {
-                call_id: format!("{}:{provider_call_index}", prepared.run_id),
+                    call_id,
                     run_id: prepared.run_id.to_string(),
                     conversation_id: prepared.conversation_id.to_string(),
                     provider_call_index,
-                request,
+                    request: request.clone(),
                 };
                 let cycle_cancellation = cancellation.child_token();
                 let cycle_events = client.events.clone();
@@ -223,7 +233,6 @@ impl RunEngine {
                     &cycle_cancellation,
                 );
                 tokio::pin!(cycle);
-            let mut pending_insertions = Vec::new();
                 let cycle = loop {
                     tokio::select! {
                         biased;
@@ -306,21 +315,108 @@ impl RunEngine {
                         result = &mut cycle => break result,
                     }
                 };
-            let cycle = match cycle {
-                Ok(cycle) => cycle,
-                Err(ModelCycleFailure {
-                    failure,
-                    usage: cycle_usage,
-                    ..
-                }) => {
-                    if let Some(cycle_usage) = cycle_usage {
+                match cycle {
+                    Ok(cycle) => break 'attempt cycle,
+                    Err(cycle_failure) => {
+                        if let Some(cycle_usage) = cycle_failure.usage {
                             accumulate_usage(&mut usage, cycle_usage);
                         }
                         if cancellation.is_cancelled() {
                             let _ = emit(client, RunEvent::CycleInterrupted).await;
                             return (RunOutcome::Cancelled, usage);
                         }
-                    return (RunOutcome::Failed(failure), usage);
+                        if !should_retry(&cycle_failure, retries) {
+                            return (RunOutcome::Failed(cycle_failure.failure), usage);
+                        }
+                        retries += 1;
+                        let message = failure_message(&cycle_failure.failure);
+                        tracing::warn!(
+                            provider_call_index,
+                            retries,
+                            max_retries = super::model_retry::MAX_MODEL_RETRIES,
+                            delay_ms = MODEL_RETRY_DELAY.as_millis() as u64,
+                            %message,
+                            checkpoint_id = checkpoint.0,
+                            "model attempt failed; retrying from current checkpoint"
+                        );
+                        if emit(
+                            client,
+                            RunEvent::ModelAttemptFailed {
+                                attempt: retries,
+                                message,
+                            },
+                        )
+                        .await
+                        .is_err()
+                        {
+                            return (client_failure(), usage);
+                        }
+
+                        let delay = tokio::time::sleep(MODEL_RETRY_DELAY);
+                        tokio::pin!(delay);
+                        loop {
+                            tokio::select! {
+                                biased;
+                                command = client.commands.recv() => {
+                                    let interruption = match command {
+                                        Some(RunCommand::InsertMessages(insertion)) => {
+                                            pending_insertions.push(insertion);
+                                            continue;
+                                        }
+                                        Some(RunCommand::BreakMessages(messages)) => messages,
+                                        Some(RunCommand::Cancel) => {
+                                            let _ = emit(client, RunEvent::CycleInterrupted).await;
+                                            return (RunOutcome::Cancelled, usage);
+                                        }
+                                        Some(RunCommand::ToolResult(_)) => {
+                                            return (
+                                                RunOutcome::Failed(RunFailure::Protocol(
+                                                    "received a tool result while waiting to retry the model".into(),
+                                                )),
+                                                usage,
+                                            );
+                                        }
+                                        None => return (client_failure(), usage),
+                                    };
+                                    if emit(client, RunEvent::CycleInterrupted).await.is_err() {
+                                        return (client_failure(), usage);
+                                    }
+                                    checkpoint = match super::messages::append_batches(
+                                        &self.store,
+                                        prepared,
+                                        client,
+                                        cancellation,
+                                        checkpoint,
+                                        std::mem::take(&mut pending_insertions),
+                                    )
+                                    .await
+                                    {
+                                        Ok((checkpoint, _)) => checkpoint,
+                                        Err(outcome) => return (outcome, usage),
+                                    };
+                                    checkpoint = match super::messages::append_batches(
+                                        &self.store,
+                                        prepared,
+                                        client,
+                                        cancellation,
+                                        checkpoint,
+                                        vec![interruption],
+                                    )
+                                    .await
+                                    {
+                                        Ok((checkpoint, _)) => checkpoint,
+                                        Err(outcome) => return (outcome, usage),
+                                    };
+                                    continue 'model;
+                                }
+                                _ = cancellation.cancelled() => {
+                                    let _ = emit(client, RunEvent::CycleInterrupted).await;
+                                    return (RunOutcome::Cancelled, usage);
+                                }
+                                _ = &mut delay => break,
+                            }
+                        }
+                    }
                 }
             };
             if let Some(cycle_usage) = cycle.usage {

server/src/run/event.rs

@@ -104,6 +104,10 @@ pub enum RunEvent {
     AutoCompactionStarted,
     AutoCompactionCompleted,
     CycleInterrupted,
+    ModelAttemptFailed {
+        attempt: u32,
+        message: String,
+    },
     TextStart,
     TextDelta(String),
     TextEnd,

server/src/run/mod.rs

@@ -7,6 +7,7 @@ mod event;
 mod handle;
 mod messages;
 mod model_cycle;
+mod model_retry;
 mod port;
 mod tool_round;
 

server/src/run/model_cycle.rs

@@ -29,6 +29,7 @@ pub struct ModelCycleFailure {
     pub partial_text: String,
     pub partial_reasoning: String,
     pub usage: Option<Usage>,
+    pub retryable: bool,
 }
 
 struct OpenTool {
@@ -186,6 +187,7 @@ pub async fn consume_model_cycle(
                                 name: name.clone(),
                                 arguments_text: String::new(),
                                 arguments: serde_json::Value::Null,
+                                argument_error: None,
                             },
                             ended: false,
                         });
@@ -218,14 +220,27 @@ pub async fn consume_model_cycle(
                         serde_json::from_str(&tool.call.arguments_text)
                     };
                     match arguments {
-                        Ok(arguments) => {
+                        Ok(arguments) if arguments.is_object() => {
                             tool.call.arguments = arguments;
+                        }
+                        Ok(_) => {
+                            tool.call.arguments = serde_json::json!({});
+                            tool.call.argument_error = Some(format!(
+                                "{} arguments must be a JSON object",
+                                tool.call.name
+                            ));
+                        }
+                        Err(error) => {
+                            tool.call.arguments = serde_json::json!({});
+                            tool.call.argument_error = Some(format!(
+                                "{} arguments are not valid JSON: {error}",
+                                tool.call.name
+                            ));
+                        }
+                    }
                     tool.ended = true;
                     send(client, RunEvent::ToolCallEnd { index }).await
                 }
-                        Err(_) => Err("provider ended a tool call with invalid JSON arguments"),
-                    }
-                }
                 Some(_) => Err("provider emitted duplicate ToolCallEnd"),
                 None => Err("provider ended an unknown tool index"),
             },
@@ -239,6 +254,13 @@ pub async fn consume_model_cycle(
             ModelEvent::Usage(value) => {
                 if usage.replace(value).is_some() {
                     Err("provider emitted duplicate Usage")
+                } else if send(client, RunEvent::Usage(value)).await.is_err() {
+                    return Err(failure(
+                        RunFailure::Client("client event channel closed".into()),
+                        text,
+                        reasoning,
+                        usage,
+                    ));
                 } else {
                     Ok(())
                 }
@@ -280,7 +302,7 @@ pub async fn consume_model_cycle(
         .map(|tool| tool.call)
         .collect::<Vec<_>>();
     if finish_reason == FinishReason::Length {
-        return Err(failure(
+        return Err(terminal_failure(
             RunFailure::Provider("model stopped before completing the response".into()),
             text,
             reasoning,
@@ -296,16 +318,6 @@ pub async fn consume_model_cycle(
             usage,
         ));
     }
-    if let Some(usage) = usage {
-        if send(client, RunEvent::Usage(usage)).await.is_err() {
-            return Err(failure(
-                RunFailure::Client("client event channel closed".into()),
-                text,
-                reasoning,
-                Some(usage),
-            ));
-        }
-    }
     let model_call_id = model_call_id.ok_or_else(|| {
         failure(
             RunFailure::Protocol("provider completed without Start".into()),
@@ -360,10 +372,83 @@ fn failure(
     partial_reasoning: String,
     usage: Option<Usage>,
 ) -> ModelCycleFailure {
+    let retryable = matches!(failure, RunFailure::Protocol(_) | RunFailure::Provider(_));
     ModelCycleFailure {
         failure,
         partial_text,
         partial_reasoning,
         usage,
+        retryable,
+    }
+}
+
+fn terminal_failure(
+    failure: RunFailure,
+    partial_text: String,
+    partial_reasoning: String,
+    usage: Option<Usage>,
+) -> ModelCycleFailure {
+    ModelCycleFailure {
+        failure,
+        partial_text,
+        partial_reasoning,
+        usage,
+        retryable: false,
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::{
+        model::Usage,
+        provider::{FinishReason, ModelEvent},
+    };
+    use tokio_stream::wrappers::ReceiverStream;
+
+    #[tokio::test]
+    async fn usage_is_forwarded_before_the_provider_call_finishes() {
+        let (provider_tx, provider_rx) = tokio::sync::mpsc::channel(4);
+        let stream = Box::pin(ReceiverStream::new(provider_rx));
+        let (event_tx, mut event_rx) = tokio::sync::mpsc::channel(4);
+        let cancellation = CancellationToken::new();
+        let cycle_cancellation = cancellation.clone();
+        let cycle = tokio::spawn(async move {
+            consume_model_cycle(stream, &event_tx, &cycle_cancellation).await
+        });
+        let usage = Usage {
+            input_tokens: Some(100),
+            context_input_tokens: Some(100),
+            output_tokens: Some(20),
+            total_tokens: Some(120),
+            ..Default::default()
+        };
+
+        provider_tx
+            .send(Ok(ModelEvent::Start {
+                model_call_id: "call".into(),
+            }))
+            .await
+            .unwrap();
+        provider_tx
+            .send(Ok(ModelEvent::Usage(usage)))
+            .await
+            .unwrap();
+
+        let event = tokio::time::timeout(std::time::Duration::from_secs(1), event_rx.recv())
+            .await
+            .unwrap()
+            .unwrap();
+        assert!(matches!(event, RunEvent::Usage(value) if value == usage));
+        assert!(!cycle.is_finished(), "usage must arrive before Done");
+
+        provider_tx
+            .send(Ok(ModelEvent::Done(FinishReason::Stop)))
+            .await
+            .unwrap();
+        drop(provider_tx);
+        let result = cycle.await.unwrap().unwrap();
+        assert_eq!(result.usage, Some(usage));
+        assert!(event_rx.try_recv().is_err(), "usage must be forwarded once");
     }
 }

server/src/run/model_retry.rs

@@ -0,0 +1,42 @@
+//! Defines retry policy for one logical model call.
+
+use std::time::Duration;
+
+use super::ModelCycleFailure;
+
+pub(super) const MAX_MODEL_RETRIES: u32 = 8;
+pub(super) const MODEL_RETRY_DELAY: Duration = Duration::from_secs(5);
+
+pub(super) fn should_retry(failure: &ModelCycleFailure, retries: u32) -> bool {
+    failure.retryable && retries < MAX_MODEL_RETRIES
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::run::{ModelCycleFailure, RunFailure};
+
+    fn failure(retryable: bool) -> ModelCycleFailure {
+        ModelCycleFailure {
+            failure: RunFailure::Provider("failed".into()),
+            partial_text: String::new(),
+            partial_reasoning: String::new(),
+            usage: None,
+            retryable,
+        }
+    }
+
+    #[test]
+    fn permits_eight_retries_after_the_initial_attempt() {
+        let retryable = failure(true);
+        for retries in 0..MAX_MODEL_RETRIES {
+            assert!(should_retry(&retryable, retries));
+        }
+        assert!(!should_retry(&retryable, MAX_MODEL_RETRIES));
+    }
+
+    #[test]
+    fn terminal_failures_never_retry() {
+        assert!(!should_retry(&failure(false), 0));
+    }
+}

server/src/store/migrations.rs

@@ -445,9 +445,20 @@ mod tests {
             .await
             .unwrap();
 
+            let argument_error_column_exists: i64 = sqlx::query_scalar(
+                "SELECT EXISTS(
+                    SELECT 1 FROM pragma_table_info('tool_round_calls')
+                    WHERE name = 'argument_error'
+                 )",
+            )
+            .fetch_one(&pool)
+            .await
+            .unwrap();
+
             assert_eq!(checksum_after, checksum_before);
-            assert_eq!(versions, vec![1, 2, 3, 4, 5, 6, 7, 8]);
+            assert_eq!(versions, vec![1, 2, 3, 4, 5, 6, 7, 8, 9]);
             assert_eq!(checkpoint_table_exists, 1);
+            assert_eq!(argument_error_column_exists, 1);
         }
     }
 

server/src/store/tool_rounds.rs

@@ -93,22 +93,16 @@ impl Store {
         for call in calls {
             sqlx::query(
                 "INSERT INTO tool_round_calls
-                 (round_id, call_index, call_id, model_call_id, name, arguments_json, status)
-                 VALUES (?, ?, ?, ?, ?, ?, 'pending')",
+                 (round_id, call_index, call_id, model_call_id, name, arguments_json, argument_error, status)
+                 VALUES (?, ?, ?, ?, ?, ?, ?, 'pending')",
             )
             .bind(round_id.as_str())
             .bind(call.index as i64)
             .bind(&call.call_id)
             .bind(&call.model_call_id)
             .bind(&call.name)
-            // A no-argument tool call streams no argument text; persist it as an
-            // empty object so the `arguments_json` column always holds valid JSON
-            // and can be re-parsed on load.
-            .bind(if call.arguments_text.trim().is_empty() {
-                "{}"
-            } else {
-                call.arguments_text.as_str()
-            })
+            .bind(serde_json::to_string(&call.arguments)?)
+            .bind(call.argument_error.as_deref())
             .execute(&mut *tx)
             .await?;
         }
@@ -276,7 +270,7 @@ impl Store {
             return Ok(None);
         };
         let rows = sqlx::query(
-            "SELECT call_index, call_id, model_call_id, name, arguments_json, status
+            "SELECT call_index, call_id, model_call_id, name, arguments_json, argument_error, status
              FROM tool_round_calls WHERE round_id = ? ORDER BY call_index",
         )
         .bind(round_id.as_str())
@@ -287,7 +281,7 @@ impl Store {
         for row in rows {
             let arguments_text: String = row.get(4);
             let call_id: String = row.get(1);
-            if row.get::<&str, _>(5) == "completed" {
+            if row.get::<&str, _>(6) == "completed" {
                 completed.push(call_id.clone());
             }
             calls.push(ToolCall {
@@ -297,6 +291,7 @@ impl Store {
                 name: row.get(3),
                 arguments: serde_json::from_str(&arguments_text)?,
                 arguments_text,
+                argument_error: row.get(5),
             });
         }
         Ok(Some(ToolRoundSnapshot {

server/tests/compaction.rs

@@ -62,6 +62,7 @@ async fn summarize_replaces_model_history_and_preserves_cursor_history() {
         ModelEvent::TextEnd,
         ModelEvent::Usage(Usage {
             input_tokens: Some(4_012),
+            context_input_tokens: Some(4_012),
             output_tokens: Some(9),
             total_tokens: Some(4_021),
             ..Default::default()
@@ -442,6 +443,7 @@ fn text_response(text: &str, input: u64, output: u64) -> Vec<ModelEvent> {
         ModelEvent::TextEnd,
         ModelEvent::Usage(Usage {
             input_tokens: Some(input),
+            context_input_tokens: Some(input),
             output_tokens: Some(output),
             total_tokens: Some(input + output),
             ..Default::default()

server/tests/error_lifecycle.rs

@@ -51,10 +51,35 @@ async fn abort_command_cancels_the_run_and_closes_output() {
 }
 
 #[tokio::test]
-async fn provider_failure_keeps_the_initial_checkpoint_then_returns_structured_error() {
+async fn provider_failure_retries_from_the_current_checkpoint_without_hiding_partial_output() {
     let (_directory, store) = fixtures::temp_store().await;
     let provider = fake_provider::FakeProvider::default();
-    provider.push_error(Error::Provider("provider failed".into()));
+    provider.push_results(vec![
+        Ok(ModelEvent::Start {
+            model_call_id: "attempt-0".into(),
+        }),
+        Ok(ModelEvent::TextStart),
+        Ok(ModelEvent::TextDelta("partial ".into())),
+        Ok(ModelEvent::ToolCallStart {
+            index: 0,
+            call_id: "failed-tool".into(),
+            name: "Read".into(),
+        }),
+        Ok(ModelEvent::ToolCallArgumentsDelta {
+            index: 0,
+            delta: "{\"path\":".into(),
+        }),
+        Err(Error::Provider("stream disconnected".into())),
+    ]);
+    provider.push(vec![
+        ModelEvent::Start {
+            model_call_id: "attempt-1".into(),
+        },
+        ModelEvent::TextStart,
+        ModelEvent::TextDelta("completed".into()),
+        ModelEvent::TextEnd,
+        ModelEvent::Done(FinishReason::Stop),
+    ]);
     let assets = PromptAssets::load(
         std::path::Path::new(env!("CARGO_MANIFEST_DIR"))
             .join("prompt/cursor")
@@ -63,7 +88,106 @@ async fn provider_failure_keeps_the_initial_checkpoint_then_returns_structured_e
     .unwrap();
     let registry = TransportRegistry::new(
         store.clone(),
-        Arc::new(provider),
+        Arc::new(provider.clone()),
+        PromptCompiler::new(assets),
+    );
+    let handle = registry.get_or_create("retry-request").await.unwrap();
+    let mut output = handle.subscribe();
+    handle
+        .command(TransportCommand::Append {
+            seqno: 0,
+            message: Box::new(protocol_client_run("retry", "retry-user")),
+        })
+        .await
+        .unwrap();
+
+    let mut seqno = 1;
+    let mut text = String::new();
+    let mut failed_tool_completed = false;
+    loop {
+        let frame = tokio::time::timeout(std::time::Duration::from_secs(60), output.recv())
+            .await
+            .unwrap()
+            .unwrap();
+        let (flags, payload) = connect::decode_frames(&frame).unwrap().pop().unwrap();
+        if flags & connect::END_STREAM_FLAG != 0 {
+            assert_eq!(
+                serde_json::from_slice::<serde_json::Value>(&payload).unwrap(),
+                serde_json::json!({})
+            );
+            break;
+        }
+        let server = pb::AgentServerMessage::decode(payload).unwrap();
+        match server.message {
+            Some(pb::agent_server_message::Message::KvServerMessage(kv)) => {
+                handle
+                    .command(TransportCommand::Append {
+                        seqno,
+                        message: Box::new(kv_ack(kv.id)),
+                    })
+                    .await
+                    .unwrap();
+                seqno += 1;
+            }
+            Some(pb::agent_server_message::Message::InteractionUpdate(update)) => {
+                match update.message {
+                    Some(pb::interaction_update::Message::TextDelta(delta)) => {
+                        text.push_str(&delta.text);
+                    }
+                    Some(pb::interaction_update::Message::ToolCallCompleted(completed))
+                        if completed.call_id == "failed-tool" =>
+                    {
+                        let tool = completed.tool_call.expect("failed tool completion");
+                        failed_tool_completed = tool.completed_at_ms.is_some();
+                    }
+                    _ => {}
+                }
+            }
+            _ => {}
+        }
+    }
+
+    assert_eq!(text, "partial completed");
+    assert!(
+        failed_tool_completed,
+        "failed attempt must terminate its partial tool card"
+    );
+    let requests = provider.requests();
+    assert_eq!(requests.len(), 2);
+    assert_eq!(requests[0], requests[1]);
+    let messages = store
+        .load_current_messages(&cursor_server::model::ConversationId::new(
+            "protocol-failed-conversation",
+        ))
+        .await
+        .unwrap();
+    assert!(messages.iter().any(|message| {
+        matches!(&message.content, MessageContent::Assistant { text, .. } if text == "completed")
+    }));
+    assert!(!messages.iter().any(|message| {
+        matches!(&message.content, MessageContent::Assistant { text, .. } if text.contains("partial"))
+    }));
+}
+
+#[tokio::test]
+async fn provider_failure_keeps_the_initial_checkpoint_then_returns_structured_error() {
+    let (_directory, store) = fixtures::temp_store().await;
+    let provider = fake_provider::FakeProvider::default();
+    provider.push(vec![
+        ModelEvent::Start {
+            model_call_id: "length-limited".into(),
+        },
+        ModelEvent::Done(FinishReason::Length),
+    ]);
+    let assets = PromptAssets::load(
+        std::path::Path::new(env!("CARGO_MANIFEST_DIR"))
+            .join("prompt/cursor")
+            .as_path(),
+    )
+    .unwrap();
+    let registry = TransportRegistry::new(
+        store.clone(),
+        Arc::new(provider.clone()),
         PromptCompiler::new(assets),
     );
     let handle = registry.get_or_create("failed-request").await.unwrap();
@@ -148,6 +272,7 @@ async fn provider_failure_keeps_the_initial_checkpoint_then_returns_structured_e
         None
     );
 
+    assert_eq!(provider.requests().len(), 1, "Length must not retry");
     let messages = store
         .load_current_messages(&cursor_server::model::ConversationId::new(
             "failed-conversation",
@@ -164,7 +289,7 @@ async fn provider_failure_keeps_the_initial_checkpoint_then_returns_structured_e
 }
 
 #[tokio::test]
-async fn runtime_protocol_failure_returns_connect_error_end_stream_and_closes() {
+async fn unknown_tool_response_id_is_ignored_and_the_run_continues() {
     let (_directory, store) = fixtures::temp_store().await;
     let provider = fake_provider::FakeProvider::default();
     provider.push(vec![
@@ -183,6 +308,15 @@ async fn runtime_protocol_failure_returns_connect_error_end_stream_and_closes()
         ModelEvent::ToolCallEnd { index: 0 },
         ModelEvent::Done(FinishReason::ToolUse),
     ]);
+    provider.push(vec![
+        ModelEvent::Start {
+            model_call_id: "model-call-2".into(),
+        },
+        ModelEvent::TextStart,
+        ModelEvent::TextDelta("done".into()),
+        ModelEvent::TextEnd,
+        ModelEvent::Done(FinishReason::Stop),
+    ]);
     let assets = PromptAssets::load(
         std::path::Path::new(env!("CARGO_MANIFEST_DIR"))
             .join("prompt/cursor")
@@ -209,7 +343,7 @@ async fn runtime_protocol_failure_returns_connect_error_end_stream_and_closes()
 
     let mut append_seqno = 1;
     let mut saw_turn_ended = false;
-    let error_json = loop {
+    let end_stream = loop {
         let frame = tokio::time::timeout(std::time::Duration::from_secs(5), output.recv())
             .await
             .unwrap()
@@ -231,25 +365,43 @@ async fn runtime_protocol_failure_returns_connect_error_end_stream_and_closes()
                 append_seqno += 1;
             }
             Some(pb::agent_server_message::Message::ExecServerMessage(exec)) => {
-                // An unknown numeric bridge id is a runtime protocol error.
-                handle
-                    .command(TransportCommand::Append {
-                        seqno: append_seqno,
-                        message: Box::new(pb::AgentClientMessage {
-                            message: Some(pb::agent_client_message::Message::ExecClientMessage(
+                // Unknown bridge ids are ignored; the valid response still completes the tool.
+                for message in [
                     pb::ExecClientMessage {
                         id: exec.id + 1_000,
                         exec_id: String::new(),
                         message: None,
                         ..Default::default()
                     },
+                    pb::ExecClientMessage {
+                        id: exec.id,
+                        exec_id: String::new(),
+                        message: Some(pb::exec_client_message::Message::ReadResult(
+                            pb::ReadResult {
+                                result: Some(pb::read_result::Result::Success(pb::ReadSuccess {
+                                    path: "/tmp/a".into(),
+                                    output: Some(pb::read_success::Output::Content("value".into())),
+                                    ..Default::default()
+                                })),
+                            },
                         )),
+                        ..Default::default()
+                    },
+                ] {
+                    handle
+                        .command(TransportCommand::Append {
+                            seqno: append_seqno,
+                            message: Box::new(pb::AgentClientMessage {
+                                message: Some(
+                                    pb::agent_client_message::Message::ExecClientMessage(message),
+                                ),
                             }),
                         })
                         .await
                         .unwrap();
                     append_seqno += 1;
                 }
+            }
             Some(pb::agent_server_message::Message::InteractionUpdate(update)) => {
                 if matches!(
                     update.message,
@@ -266,12 +418,8 @@ async fn runtime_protocol_failure_returns_connect_error_end_stream_and_closes()
         }
     };
 
-    assert!(!saw_turn_ended);
-    assert_eq!(error_json["error"]["code"], "invalid_argument");
-    assert_eq!(
-        error_json["error"]["message"],
-        "unknown ExecClientMessage id: 1001"
-    );
+    assert!(saw_turn_ended);
+    assert_eq!(end_stream, serde_json::json!({}));
     assert_eq!(
         tokio::time::timeout(std::time::Duration::from_secs(1), output.recv())
             .await
@@ -279,28 +427,14 @@ async fn runtime_protocol_failure_returns_connect_error_end_stream_and_closes()
         None
     );
 
-    let deadline = tokio::time::Instant::now() + std::time::Duration::from_secs(1);
-    let (status, failure_summary) = loop {
-        let row: (String, Option<String>) =
+    let (status, failure_summary): (String, Option<String>) =
         sqlx::query_as("SELECT status, failure_summary FROM runs WHERE cursor_request_id = ?")
             .bind("protocol-failed-request")
             .fetch_one(store.pool())
             .await
             .unwrap();
-        if row.0 != "running" {
-            break row;
-        }
-        assert!(
-            tokio::time::Instant::now() < deadline,
-            "Run remained running after the Cursor session failed"
-        );
-        tokio::time::sleep(std::time::Duration::from_millis(10)).await;
-    };
-    assert_eq!(status, "failed");
-    assert_eq!(
-        failure_summary.as_deref(),
-        Some("unknown ExecClientMessage id: 1001")
-    );
+    assert_eq!(status, "completed");
+    assert_eq!(failure_summary, None);
 }
 
 #[tokio::test]

server/tests/interrupt.rs

@@ -1580,6 +1580,7 @@ fn text_response(text: &str) -> Vec<ModelEvent> {
         ModelEvent::TextEnd,
         ModelEvent::Usage(Usage {
             input_tokens: Some(1),
+            context_input_tokens: Some(1),
             output_tokens: Some(1),
             total_tokens: Some(2),
             ..Default::default()

server/tests/support/fake_provider.rs

@@ -31,10 +31,13 @@ pub struct FakeProvider {
 
 impl FakeProvider {
     pub fn push(&self, events: Vec<ModelEvent>) {
+        self.push_results(events.into_iter().map(Ok).collect());
+    }
+    pub fn push_results(&self, events: Vec<Result<ModelEvent, Error>>) {
         self.responses
             .lock()
             .unwrap()
-            .push_back(FakeResponse::Events(events.into_iter().map(Ok).collect()));
+            .push_back(FakeResponse::Events(events));
     }
     pub fn push_error(&self, error: Error) {
         self.responses

server/tests/tool_round.rs

@@ -25,9 +25,11 @@ use cursor_server::{
         OPENAI_CHAT_ENDPOINT,
     },
     provider::{FinishReason, ModelEvent},
+    run::consume_model_cycle,
 };
 use prost::Message;
 use serde_json::json;
+use tokio_util::sync::CancellationToken;
 
 fn call(id: &str, name: &str) -> ToolCall {
     ToolCall {
@@ -37,6 +39,7 @@ fn call(id: &str, name: &str) -> ToolCall {
         name: name.into(),
         arguments_text: "{}".into(),
         arguments: json!({}),
+        argument_error: None,
     }
 }
 
@@ -68,6 +71,37 @@ fn mcp_context(server: &str, provider: &str, tool: &str) -> ExecContext {
     context
 }
 
+#[tokio::test]
+async fn malformed_provider_tool_json_is_kept_as_a_tool_validation_error() {
+    let stream = Box::pin(futures_util::stream::iter(vec![
+        Ok(ModelEvent::Start {
+            model_call_id: "model-call".into(),
+        }),
+        Ok(ModelEvent::ToolCallStart {
+            index: 0,
+            call_id: "call-malformed".into(),
+            name: "Read".into(),
+        }),
+        Ok(ModelEvent::ToolCallArgumentsDelta {
+            index: 0,
+            delta: "{\"path\":".into(),
+        }),
+        Ok(ModelEvent::ToolCallEnd { index: 0 }),
+        Ok(ModelEvent::Done(FinishReason::ToolUse)),
+    ]));
+    let (events, _receiver) = tokio::sync::mpsc::channel(16);
+    let result = consume_model_cycle(stream, &events, &CancellationToken::new())
+        .await
+        .expect("malformed tool JSON must not fail the model cycle");
+
+    assert_eq!(result.calls.len(), 1);
+    assert!(result.calls[0]
+        .argument_error
+        .as_deref()
+        .is_some_and(|message| message.contains("not valid JSON")));
+    assert_eq!(result.calls[0].arguments, json!({}));
+}
+
 #[test]
 fn dynamic_mcp_call_routes_to_the_captured_exec_message() {
     let call = ToolCall {
@@ -77,6 +111,7 @@ fn dynamic_mcp_call_routes_to_the_captured_exec_message() {
         name: "mcp_repo_lookup".into(),
         arguments_text: "{\"query\":\"x\"}".into(),
         arguments: json!({"query": "x"}),
+        argument_error: None,
     };
     let definition = pb::McpToolDefinition {
         name: "mcp_repo_lookup".into(),
@@ -372,6 +407,52 @@ async fn unknown_mcp_descriptor_returns_a_tool_error_without_client_discovery()
     assert!(completion.result().content.contains("descriptor not found"));
 }
 
+#[tokio::test]
+async fn invalid_tool_arguments_complete_as_tool_errors() {
+    let dispatcher = ToolDispatcher::new(CursorToolRuntime::default());
+    let completed = HashSet::new();
+    let started = HashSet::new();
+    let state = || ToolBatchState {
+        completed: &completed,
+        started: &started,
+        response_text: "",
+        response_thinking: "",
+    };
+
+    let mut malformed = call("call-malformed", "Read");
+    malformed.argument_error = Some("Read arguments are not valid JSON".into());
+    let mut missing = call("call-missing", "Shell");
+    missing.arguments = json!({"description": "missing command"});
+    let mut wrong_type = call("call-type", "Shell");
+    wrong_type.arguments = json!({"command": 42});
+    let mut invalid_timeout = call("call-timeout", "Shell");
+    invalid_timeout.arguments = json!({"command": "pwd", "block_until_ms": -1});
+
+    for (invocation, expected) in [
+        (malformed, "not valid JSON"),
+        (missing, "missing command"),
+        (wrong_type, "missing command"),
+        (invalid_timeout, "out of range"),
+    ] {
+        let dispatched = dispatcher
+            .start_batch(
+                &[invocation],
+                state(),
+                &[],
+                &BTreeMap::new(),
+                &exec_context(),
+            )
+            .await
+            .unwrap();
+        let completion = dispatched[0]
+            .completion
+            .as_ref()
+            .expect("invalid arguments must complete as a tool error");
+        assert!(completion.result().is_error);
+        assert!(completion.result().content.contains(expected));
+    }
+}
+
 #[tokio::test]
 async fn shell_uses_background_timeout_and_preserves_stream_identity() {
     let mut shell = call("call-shell", "Shell");
@@ -701,12 +782,15 @@ async fn an_exec_result_must_match_the_reserved_tool() {
         },
         &pending,
     )
-    .await;
-    let Err(error) = result else {
-        panic!("mismatched result must fail")
+    .await
+    .unwrap();
+    let codec::ClientExecEvent::Completed(completion) = result else {
+        panic!("mismatched result must complete as a tool error")
     };
-    assert!(error
-        .to_string()
+    assert!(completion.result().is_error);
+    assert!(completion
+        .result()
+        .content
         .contains("unexpected Exec result for tool Read"));
     assert!(pending.exec_call(id).await.is_none());
     assert_eq!(pending.completed_call(id).await.as_deref(), Some("call-1"));
@@ -718,15 +802,13 @@ async fn an_exec_result_must_match_the_reserved_tool() {
         },
         &pending,
     )
-    .await;
-    let Err(duplicate) = duplicate else {
-        panic!("duplicate terminal result must fail")
-    };
-    assert!(duplicate.to_string().contains("duplicate terminal"));
+    .await
+    .unwrap();
+    assert!(matches!(duplicate, codec::ClientExecEvent::Pending));
 }
 
 #[tokio::test]
-async fn unknown_exec_id_is_a_protocol_error() {
+async fn unknown_exec_id_is_ignored() {
     let result = codec::client_event(
         &pb::ExecClientMessage {
             id: 999,
@@ -737,15 +819,9 @@ async fn unknown_exec_id_is_a_protocol_error() {
         },
         &CursorToolRuntime::default(),
     )
-    .await;
-    let Err(error) = result else {
-        panic!("unknown Exec id must fail")
-    };
-    assert!(matches!(
-        error,
-        cursor_server::Error::Protocol(message)
-            if message == "unknown ExecClientMessage id: 999"
-    ));
+    .await
+    .unwrap();
+    assert!(matches!(result, codec::ClientExecEvent::Pending));
 }
 
 #[tokio::test]