fix: separate Ticket keys from assignment identity

# Conflicts:
#	crates/agen/README.md
#	crates/agen/examples/engine_cancel_demo.rs
#	crates/agen/examples/engine_cli.rs
#	crates/agen/src/engine.rs
#	crates/agen/tests/annotated_history_test.rs
#	crates/agen/tests/callback_test.rs
#	crates/agen/tests/engine_fixtures.rs
#	crates/agen/tests/engine_state_test.rs
#	crates/agen/tests/parallel_execution_test.rs
#	crates/agen/tests/reasoning_round_trip_test.rs
#	crates/session-store/tests/session_test.rs
#	crates/worker/src/worker.rs

crates/agen/README.md

@@ -21,20 +21,22 @@ agen = { version = "0.2.1", features = ["codex"] }
 
 ## Quick start
 
-Supply an implementation of [`LlmClient`](https://docs.rs/agen/latest/agen/llm_client/trait.LlmClient.html), then run a turn. The first call consumes the mutable engine and returns a cache-locked engine for later turns.
+Supply an implementation of [`LlmClient`](https://docs.rs/agen/latest/agen/llm_client/trait.LlmClient.html), keep conversation history in your application, then run a turn. The first call consumes the mutable engine and returns a cache-locked engine for later turns.
 
 ```no_run
-use agen::Engine;
+use agen::{Engine, EngineError, History};
 use agen::llm_client::LlmClient;
 
-async fn conversation<C: LlmClient>(client: C) {
+async fn conversation<C: LlmClient>(client: C) -> Result<(), EngineError> {
+    let mut history = History::new();
     let output = Engine::new(client)
         .system_prompt("You are a concise assistant.")
-        .run("Explain typed state in one sentence.")
+        .run(&mut history, "Explain typed state in one sentence.")
         .await;
 
     let mut engine = output.engine;
-    let _exit = engine.run("Give a Rust example.").await;
+    let _result = engine.run(&mut history, "Give a Rust example.").await;
+    Ok(())
 }
 ```
 

crates/agen/examples/engine_cancel_demo.rs

@@ -29,6 +29,7 @@ async fn main() -> Result<(), Box<dyn std::error::Error>> {
     let base_url = scheme.default_base_url().to_string();
     let client = HttpTransport::new(scheme, model, base_url, ResolvedAuth::ApiKey(api_key), cap);
     let engine = Engine::new(client);
+    let mut history = agen::History::new();
 
     println!("🚀 Starting Engine...");
     println!("💡 Will cancel after 2 seconds\n");
@@ -45,7 +46,7 @@ async fn main() -> Result<(), Box<dyn std::error::Error>> {
 
     println!("📡 Sending request to LLM...");
 
-    let output = engine.run("Tell me a very long story about a brave knight. Make it as detailed as possible with many paragraphs.").await;
+    let output = engine.run(&mut history, "Tell me a very long story about a brave knight. Make it as detailed as possible with many paragraphs.").await;
     match output.result {
         EngineRunExit::Finished => println!("✅ Task completed normally"),
         EngineRunExit::Paused => println!("⏸️  Task paused"),

crates/agen/examples/engine_cli.rs

@@ -451,6 +451,7 @@ async fn main() -> Result<(), Box<dyn std::error::Error>> {
 
     // Create Engine
     let mut engine = Engine::new(client);
+    let mut history = agen::History::new();
 
     let tool_call_names = Arc::new(Mutex::new(HashMap::new()));
 
@@ -476,7 +477,7 @@ async fn main() -> Result<(), Box<dyn std::error::Error>> {
 
     // One-shot mode
     if let Some(prompt) = args.prompt {
-        let output = engine.run(&prompt).await;
+        let output = engine.run(&mut history, &prompt).await;
         if let EngineRunExit::Interrupted(StopReason::Unexpected(error)) = output.result {
             eprintln!("\n❌ Error: {error}");
         }
@@ -497,7 +498,7 @@ async fn main() -> Result<(), Box<dyn std::error::Error>> {
         return Ok(());
     }
 
-    let output = engine.run(first_input).await;
+    let output = engine.run(&mut history, first_input).await;
     let mut locked = output.engine;
 
     loop {
@@ -517,7 +518,9 @@ async fn main() -> Result<(), Box<dyn std::error::Error>> {
             break;
         }
 
-        if let EngineRunExit::Interrupted(StopReason::Unexpected(error)) = locked.run(input).await {
+        if let EngineRunExit::Interrupted(StopReason::Unexpected(error)) =
+            locked.run(&mut history, input).await
+        {
             eprintln!("\n❌ Error: {error}");
         }
     }

crates/agen/src/history.rs

@@ -0,0 +1,199 @@
+//! Typed conversation history containers.
+//!
+//! Agen keeps provider-visible [`Item`](crate::Item) values separate from any
+//! host-domain provenance. The host chooses the annotation type `A`, while Agen
+//! preserves each item and annotation as one entry for clone/truncate/restore
+//! style history operations.
+
+use serde::{Deserialize, Serialize};
+
+use crate::Item;
+
+/// One conversation-history entry with host-owned annotation.
+#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
+pub struct HistoryEntry<A = ()> {
+    /// Provider/model-visible conversation item.
+    pub item: Item,
+    /// Host-domain metadata kept with the item and never projected to providers.
+    pub annotation: A,
+}
+
+impl<A> HistoryEntry<A> {
+    /// Build an entry from an item and its annotation.
+    pub fn new(item: Item, annotation: A) -> Self {
+        Self { item, annotation }
+    }
+
+    /// Split the entry into its item and annotation.
+    pub fn into_parts(self) -> (Item, A) {
+        (self.item, self.annotation)
+    }
+}
+
+impl HistoryEntry<()> {
+    /// Build a unit-annotated entry.
+    pub fn from_item(item: Item) -> Self {
+        Self {
+            item,
+            annotation: (),
+        }
+    }
+}
+
+/// Conversation history with one annotation per item.
+#[derive(Debug, Clone, Serialize, Deserialize, PartialEq, Default)]
+pub struct History<A = ()> {
+    entries: Vec<HistoryEntry<A>>,
+}
+
+impl<A> History<A> {
+    /// Create an empty history.
+    pub fn new() -> Self {
+        Self {
+            entries: Vec::new(),
+        }
+    }
+
+    /// Build history from already annotated entries, preserving order.
+    pub fn from_entries(entries: Vec<HistoryEntry<A>>) -> Self {
+        Self { entries }
+    }
+
+    /// Replace all entries as one restore/rebuild operation and return the old entries.
+    pub fn replace_entries(&mut self, entries: Vec<HistoryEntry<A>>) -> Vec<HistoryEntry<A>> {
+        std::mem::replace(&mut self.entries, entries)
+    }
+
+    /// Borrow annotated entries.
+    pub fn entries(&self) -> &[HistoryEntry<A>] {
+        &self.entries
+    }
+
+    /// Mutably borrow annotated entries for host-owned rebuild operations.
+    pub fn entries_mut(&mut self) -> &mut [HistoryEntry<A>] {
+        &mut self.entries
+    }
+
+    /// Consume the history into annotated entries.
+    pub fn into_entries(self) -> Vec<HistoryEntry<A>> {
+        self.entries
+    }
+
+    /// Number of entries.
+    pub fn len(&self) -> usize {
+        self.entries.len()
+    }
+
+    /// Whether the history is empty.
+    pub fn is_empty(&self) -> bool {
+        self.entries.is_empty()
+    }
+
+    /// Iterate over annotated entries.
+    pub fn iter(&self) -> impl ExactSizeIterator<Item = &HistoryEntry<A>> {
+        self.entries.iter()
+    }
+
+    /// Iterate over provider-visible items only.
+    pub fn items(&self) -> impl ExactSizeIterator<Item = &Item> {
+        self.entries.iter().map(|entry| &entry.item)
+    }
+
+    /// Clone provider-visible items into a request-local projection.
+    pub fn items_cloned(&self) -> Vec<Item> {
+        self.items().cloned().collect()
+    }
+
+    /// Append an already annotated entry.
+    pub fn push_entry(&mut self, entry: HistoryEntry<A>) {
+        self.entries.push(entry);
+    }
+
+    /// Append many already annotated entries.
+    pub fn extend_entries(&mut self, entries: impl IntoIterator<Item = HistoryEntry<A>>) {
+        self.entries.extend(entries);
+    }
+
+    /// Commit one item through a trusted annotation callback before it becomes live.
+    ///
+    /// The callback may durably persist the item and returns the annotation that
+    /// must be stored with it. If the callback fails, the history is left unchanged.
+    pub fn append_with(
+        &mut self,
+        item: Item,
+        annotate: &mut impl FnMut(&Item) -> Result<A, String>,
+    ) -> Result<(), String> {
+        let annotation = annotate(&item)?;
+        self.entries.push(HistoryEntry { item, annotation });
+        Ok(())
+    }
+
+    /// Commit items through a trusted annotation callback before they become live.
+    ///
+    /// Items before a failure remain appended; the failing item and later items do
+    /// not enter history. This mirrors append-only durable logs where each accepted
+    /// item is already committed before the next item is attempted.
+    pub fn extend_with(
+        &mut self,
+        items: impl IntoIterator<Item = Item>,
+        annotate: &mut impl FnMut(&Item) -> Result<A, String>,
+    ) -> Result<(), String> {
+        for item in items {
+            self.append_with(item, annotate)?;
+        }
+        Ok(())
+    }
+
+    /// Truncate entries, preserving item+annotation pairing for retained entries.
+    pub fn truncate(&mut self, len: usize) {
+        self.entries.truncate(len);
+    }
+
+    /// Clear all entries.
+    pub fn clear(&mut self) {
+        self.entries.clear();
+    }
+}
+
+impl History<()> {
+    /// Build unit-annotated history from provider-visible items.
+    pub fn from_items(items: Vec<Item>) -> Self {
+        Self {
+            entries: items.into_iter().map(HistoryEntry::from_item).collect(),
+        }
+    }
+
+    /// Replace history from provider-visible items using unit annotations.
+    pub fn replace_items(&mut self, items: Vec<Item>) -> Vec<HistoryEntry<()>> {
+        self.replace_entries(items.into_iter().map(HistoryEntry::from_item).collect())
+    }
+
+    /// Append one item with unit annotation.
+    pub fn push(&mut self, item: Item) {
+        self.entries.push(HistoryEntry::from_item(item));
+    }
+
+    /// Append items with unit annotations.
+    pub fn extend_items(&mut self, items: impl IntoIterator<Item = Item>) {
+        self.entries
+            .extend(items.into_iter().map(HistoryEntry::from_item));
+    }
+}
+
+impl<A> IntoIterator for History<A> {
+    type Item = HistoryEntry<A>;
+    type IntoIter = std::vec::IntoIter<HistoryEntry<A>>;
+
+    fn into_iter(self) -> Self::IntoIter {
+        self.entries.into_iter()
+    }
+}
+
+impl<'a, A> IntoIterator for &'a History<A> {
+    type Item = &'a HistoryEntry<A>;
+    type IntoIter = std::slice::Iter<'a, HistoryEntry<A>>;
+
+    fn into_iter(self) -> Self::IntoIter {
+        self.entries.iter()
+    }
+}

crates/agen/src/lib.rs

@@ -2,6 +2,7 @@
 
 mod engine;
 mod handler;
+mod history;
 mod message;
 
 pub(crate) mod callback;
@@ -24,9 +25,14 @@ pub use engine::{
     LlmRetryNotice, StopReason, ToolRegistryError,
 };
 pub use handler::ToolUseBlockStart;
+pub use history::{History, HistoryEntry};
 pub use interceptor::Interceptor;
 pub use message::{ContentPart, Item, Message, Role};
-pub use tool::{ToolCall, ToolExecutionContext, ToolOutputLimits, ToolResult};
+pub use tool::{
+    ToolCall, ToolExecutionContext, ToolExecutionHandle, ToolExecutionPolicy,
+    ToolExecutionTerminal, ToolExecutionTerminalFuture, ToolOutputLimits, ToolResult,
+    ToolResultDisposition,
+};
 pub use usage_record::UsageRecord;
 
 /// Implementation dependencies used by code generated from `agen` macros.

crates/agen/src/llm_client/types.rs

@@ -9,7 +9,7 @@
 
 use std::{fmt, sync::Arc};
 
-use crate::tool::Attachment;
+use crate::tool::{Attachment, ToolResultDisposition};
 use base64::Engine as _;
 use serde::{Deserialize, Serialize};
 
@@ -121,6 +121,9 @@ pub enum Item {
         /// Detailed output (removed by pruning when old enough)
         #[serde(default, skip_serializing_if = "Option::is_none")]
         content: Option<String>,
+        /// Typed terminal state used for replay and recovery.
+        #[serde(default, skip_serializing_if = "ToolResultDisposition::is_success")]
+        disposition: ToolResultDisposition,
         /// Whether the tool result represents an execution error.
         #[serde(default, skip_serializing_if = "is_false")]
         is_error: bool,
@@ -261,7 +264,17 @@ impl Item {
         content: Option<String>,
         is_error: bool,
     ) -> Self {
-        Self::tool_result_item_with_attachments(call_id, summary, content, is_error, Vec::new())
+        Self::tool_result_item_with_disposition_and_attachments(
+            call_id,
+            summary,
+            content,
+            if is_error {
+                ToolResultDisposition::Error
+            } else {
+                ToolResultDisposition::Success
+            },
+            Vec::new(),
+        )
     }
 
     /// Create a tool result item with durable, prunable structured attachments.
@@ -272,11 +285,33 @@ impl Item {
         is_error: bool,
         attachments: Vec<Attachment>,
     ) -> Self {
+        Self::tool_result_item_with_disposition_and_attachments(
+            call_id,
+            summary,
+            content,
+            if is_error {
+                ToolResultDisposition::Error
+            } else {
+                ToolResultDisposition::Success
+            },
+            attachments,
+        )
+    }
+
+    pub fn tool_result_item_with_disposition_and_attachments(
+        call_id: impl Into<String>,
+        summary: impl Into<String>,
+        content: Option<String>,
+        disposition: ToolResultDisposition,
+        attachments: Vec<Attachment>,
+    ) -> Self {
+        let is_error = !disposition.is_success();
         Self::ToolResult {
             id: None,
             call_id: call_id.into(),
             summary: summary.into(),
             content,
+            disposition,
             is_error,
             attachments,
         }

crates/agen/src/state.rs

@@ -19,7 +19,7 @@ mod private {
 /// - Editing message history (add, delete, clear)
 /// - Registering tools and hooks
 ///
-/// Can transition to [`Locked`] state via `Engine::lock()`.
+/// Can transition to [`Locked`] state via `Engine::lock(&history)`.
 ///
 /// # Examples
 ///

crates/agen/src/tool.rs

@@ -3,7 +3,14 @@
 //! Traits for defining tools callable by LLM.
 //! Usually auto-implemented using the `#[tool]` macro.
 
-use std::{collections::HashMap, fmt, sync::Arc};
+use std::{
+    collections::HashMap,
+    fmt,
+    future::Future,
+    pin::Pin,
+    sync::Arc,
+    task::{Context, Poll},
+};
 
 use async_trait::async_trait;
 use base64::{Engine as _, engine::general_purpose::STANDARD};
@@ -23,6 +30,12 @@ pub enum ToolError {
     /// Internal error
     #[error("Internal error: {0}")]
     Internal(String),
+    /// Cooperative cancellation completed with bounded terminal output.
+    #[error("Tool execution cancelled")]
+    Cancelled(ToolOutput),
+    /// Execution was interrupted with a confirmed bounded terminal output.
+    #[error("Tool execution interrupted")]
+    Interrupted(ToolOutput),
 }
 
 // =============================================================================
@@ -158,6 +171,28 @@ pub enum Attachment {
     Image(ImageAttachment),
 }
 
+/// Terminal disposition of one started tool call.
+///
+/// `Cancelled` means the tool confirmed cancellation. `OutcomeUnknown` means
+/// execution stopped without confirmation, so neither completion nor side
+/// effects may be inferred.
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize, Default)]
+#[serde(rename_all = "snake_case")]
+pub enum ToolResultDisposition {
+    #[default]
+    Success,
+    Error,
+    Interrupted,
+    Cancelled,
+    OutcomeUnknown,
+}
+
+impl ToolResultDisposition {
+    pub const fn is_success(&self) -> bool {
+        matches!(self, Self::Success)
+    }
+}
+
 /// Tool execution result.
 ///
 /// Every output has a mandatory `summary` (1-2 lines) that persists in
@@ -322,6 +357,12 @@ impl ToolExecutionContext {
         }
     }
 
+    /// Identifies one live execution attempt without making the batch id a durable
+    /// replay or idempotency authority.
+    pub fn execution_id(&self) -> String {
+        format!("{}:{}", self.batch_id, self.call_id)
+    }
+
     /// Context for direct, non-engine calls in unit tests and low-level callers.
     pub fn direct() -> Self {
         Self::new("direct", "direct", 0)
@@ -334,6 +375,142 @@ impl Default for ToolExecutionContext {
     }
 }
 
+/// The provider-confirmed terminal result of one started tool execution.
+///
+/// `OutcomeUnknown` is reserved for an execution task that had to be force-closed
+/// or failed before the provider could confirm its terminal result.
+#[derive(Debug)]
+pub enum ToolExecutionTerminal {
+    Confirmed(Result<ToolOutput, ToolError>),
+    OutcomeUnknown,
+}
+
+/// The completion future paired with a [`ToolExecutionHandle`]. Dropping this
+/// future does not drop the provider execution: the spawned execution remains
+/// owned by its handle until it completes or is explicitly force-closed.
+pub struct ToolExecutionTerminalFuture {
+    task: tokio::task::JoinHandle<Result<ToolOutput, ToolError>>,
+}
+
+impl Future for ToolExecutionTerminalFuture {
+    type Output = ToolExecutionTerminal;
+
+    fn poll(mut self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<Self::Output> {
+        match Pin::new(&mut self.task).poll(cx) {
+            Poll::Ready(Ok(result)) => Poll::Ready(ToolExecutionTerminal::Confirmed(result)),
+            Poll::Ready(Err(_)) => Poll::Ready(ToolExecutionTerminal::OutcomeUnknown),
+            Poll::Pending => Poll::Pending,
+        }
+    }
+}
+
+/// Live ownership and control for one started tool execution.
+///
+/// Execution, cancellation, and terminal confirmation remain provider-owned:
+/// this handle starts `Tool::execute`, delegates cooperative cancellation to
+/// `Tool::cancel_execution`, and treats execution-future completion as the
+/// provider's terminal confirmation. Agen may force-close only after its caller's
+/// deadline expires, at which point the outcome is necessarily unknown.
+#[derive(Clone)]
+pub struct ToolExecutionHandle {
+    inner: Arc<ToolExecutionHandleInner>,
+}
+
+struct ToolExecutionHandleInner {
+    tool: Arc<dyn Tool>,
+    context: ToolExecutionContext,
+    abort: tokio::task::AbortHandle,
+}
+
+impl Drop for ToolExecutionHandleInner {
+    fn drop(&mut self) {
+        // Losing the final live owner is an explicit forced close, never a
+        // best-effort detached provider future.
+        self.abort.abort();
+    }
+}
+
+impl fmt::Debug for ToolExecutionHandle {
+    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
+        f.debug_struct("ToolExecutionHandle")
+            .field("call_id", &self.inner.context.call_id)
+            .field("batch_id", &self.inner.context.batch_id)
+            .finish_non_exhaustive()
+    }
+}
+
+impl ToolExecutionHandle {
+    pub fn start(
+        tool: Arc<dyn Tool>,
+        input_json: String,
+        context: ToolExecutionContext,
+    ) -> (Self, ToolExecutionTerminalFuture) {
+        let execution_tool = Arc::clone(&tool);
+        let execution_context = context.clone();
+        let task =
+            tokio::spawn(
+                async move { execution_tool.execute(&input_json, execution_context).await },
+            );
+        let abort = task.abort_handle();
+        (
+            Self {
+                inner: Arc::new(ToolExecutionHandleInner {
+                    tool,
+                    context,
+                    abort,
+                }),
+            },
+            ToolExecutionTerminalFuture { task },
+        )
+    }
+
+    pub fn context(&self) -> &ToolExecutionContext {
+        &self.inner.context
+    }
+
+    pub async fn cancel_before(&self, deadline: tokio::time::Instant) -> Result<(), ToolError> {
+        match tokio::time::timeout_at(
+            deadline,
+            self.inner.tool.cancel_execution(&self.inner.context),
+        )
+        .await
+        {
+            Ok(result) => result,
+            Err(_) => Err(ToolError::Internal(format!(
+                "tool cancellation request exceeded its deadline for call {}",
+                self.inner.context.call_id
+            ))),
+        }
+    }
+
+    pub fn force_close(&self) {
+        self.inner.abort.abort();
+    }
+}
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub struct ToolExecutionPolicy {
+    /// Time a pause waits for already-started providers to reach a natural safe
+    /// boundary before escalating to explicit cooperative cancellation.
+    pub pause_safe_boundary_timeout: std::time::Duration,
+    /// Maximum time allowed for a provider to accept one cooperative
+    /// cancellation request.
+    pub cancellation_request_timeout: std::time::Duration,
+    /// Maximum time allowed for all providers to confirm terminal results after
+    /// cancellation has been requested.
+    pub terminal_confirmation_timeout: std::time::Duration,
+}
+
+impl Default for ToolExecutionPolicy {
+    fn default() -> Self {
+        Self {
+            pause_safe_boundary_timeout: std::time::Duration::from_millis(100),
+            cancellation_request_timeout: std::time::Duration::from_millis(100),
+            terminal_confirmation_timeout: std::time::Duration::from_millis(500),
+        }
+    }
+}
+
 // =============================================================================
 // Tool trait
 // =============================================================================
@@ -402,6 +579,26 @@ pub trait Tool: Send + Sync {
         input_json: &str,
         ctx: ToolExecutionContext,
     ) -> Result<ToolOutput, ToolError>;
+
+    /// Request cooperative cancellation for one started call.
+    ///
+    /// Implementations that own cancellable provider operations should signal
+    /// every live execution identified by `call_id`, then let `execute` return
+    /// the confirmed bounded terminal output. Direct callers may use this
+    /// compatibility surface; Agen uses [`Tool::cancel_execution`] so providers
+    /// can bind cancellation to one exact live attempt.
+    async fn cancel(&self, _call_id: &str) -> Result<(), ToolError> {
+        Ok(())
+    }
+
+    /// Request cooperative cancellation for one exact started execution.
+    ///
+    /// The default preserves existing tools by delegating to `cancel(call_id)`.
+    /// Providers with their own execution registry should override this method
+    /// and key cancellation by [`ToolExecutionContext::execution_id`].
+    async fn cancel_execution(&self, ctx: &ToolExecutionContext) -> Result<(), ToolError> {
+        self.cancel(&ctx.call_id).await
+    }
 }
 
 // =============================================================================
@@ -429,6 +626,9 @@ pub struct ToolCall {
 pub struct ToolResult {
     /// Corresponding tool call ID
     pub tool_use_id: String,
+    /// Typed terminal state.
+    #[serde(default, skip_serializing_if = "ToolResultDisposition::is_success")]
+    pub disposition: ToolResultDisposition,
     /// Short summary (always kept in history)
     pub summary: String,
     /// Detailed output (prunable)
@@ -445,11 +645,20 @@ pub struct ToolResult {
 impl ToolResult {
     /// Create a success result from a [`ToolOutput`].
     pub fn from_output(tool_use_id: impl Into<String>, output: ToolOutput) -> Self {
+        Self::from_output_with_disposition(tool_use_id, output, ToolResultDisposition::Success)
+    }
+
+    pub fn from_output_with_disposition(
+        tool_use_id: impl Into<String>,
+        output: ToolOutput,
+        disposition: ToolResultDisposition,
+    ) -> Self {
         Self {
             tool_use_id: tool_use_id.into(),
+            disposition,
             summary: output.summary,
             content: output.content,
-            is_error: false,
+            is_error: !disposition.is_success(),
             attachments: output.attachments,
         }
     }
@@ -458,12 +667,28 @@ impl ToolResult {
     pub fn error(tool_use_id: impl Into<String>, message: impl Into<String>) -> Self {
         Self {
             tool_use_id: tool_use_id.into(),
+            disposition: ToolResultDisposition::Error,
             summary: message.into(),
             content: None,
             is_error: true,
             attachments: Vec::new(),
         }
     }
+
+    /// Close an execution whose completion and side effects cannot be confirmed.
+    pub fn outcome_unknown(tool_use_id: impl Into<String>) -> Self {
+        Self {
+            tool_use_id: tool_use_id.into(),
+            disposition: ToolResultDisposition::OutcomeUnknown,
+            summary: "Tool execution outcome unknown".to_string(),
+            content: Some(
+                "Execution was interrupted before completion could be confirmed. Completion and side effects are unknown."
+                    .to_string(),
+            ),
+            is_error: true,
+            attachments: Vec::new(),
+        }
+    }
 }
 
 #[cfg(test)]

crates/agen/tests/annotated_history_test.rs

@@ -0,0 +1,84 @@
+mod common;
+
+use agen::llm_client::event::{Event, ResponseStatus, StatusEvent};
+use agen::{Engine, EngineError, History, HistoryEntry, Item, Role};
+use common::MockLlmClient;
+
+fn completed_text_events(text: &str) -> Vec<Event> {
+    vec![
+        Event::text_block_start(0),
+        Event::text_delta(0, text),
+        Event::text_block_stop(0, None),
+        Event::Status(StatusEvent {
+            status: ResponseStatus::Completed,
+        }),
+    ]
+}
+
+#[tokio::test]
+async fn run_preserves_item_annotations_without_projecting_them() {
+    let client = MockLlmClient::new(completed_text_events("assistant reply"));
+    let engine = Engine::<_, agen::state::Mutable, String>::new_annotated(client);
+    let mut history = History::<String>::new();
+    let mut next = 0usize;
+    let mut annotate = |item: &Item| {
+        next += 1;
+        let kind = match item {
+            Item::Message { role, .. } => match role {
+                Role::User => "user",
+                Role::Assistant => "assistant",
+                Role::System => "system",
+            },
+            Item::ToolCall { .. } => "tool_call",
+            Item::ToolResult { .. } => "tool_result",
+            Item::Reasoning { .. } => "reasoning",
+        };
+        Ok(format!("{next}:{kind}"))
+    };
+
+    let output = engine
+        .run_with_annotation(&mut history, "hello", &mut annotate)
+        .await;
+
+    assert!(matches!(output.result, agen::EngineRunExit::Finished));
+    assert_eq!(history.len(), 2);
+    assert_eq!(history.entries()[0].annotation, "1:user");
+    assert_eq!(history.entries()[1].annotation, "2:assistant");
+    assert_eq!(history.items_cloned().len(), 2);
+}
+
+#[test]
+fn append_failure_does_not_make_item_live() {
+    let client = MockLlmClient::new(vec![]);
+    let mut engine = Engine::<_, agen::state::Mutable, usize>::new_annotated(client);
+    let mut history = History::<usize>::new();
+    let mut fail = |_item: &Item| Err("commit failed".to_string());
+
+    let err = engine
+        .append_history_with(&mut history, [Item::user_message("uncommitted")], &mut fail)
+        .unwrap_err();
+
+    assert!(matches!(err, EngineError::HistoryAppend(message) if message == "commit failed"));
+    assert!(history.is_empty());
+}
+
+#[test]
+fn replacement_keeps_items_and_annotations_together() {
+    let mut history = History::from_entries(vec![
+        HistoryEntry::new(Item::user_message("old"), "old-ann".to_string()),
+        HistoryEntry::new(Item::user_message("second"), "second-ann".to_string()),
+    ]);
+
+    history.truncate(1);
+    assert_eq!(history.entries()[0].item.as_text(), Some("old"));
+    assert_eq!(history.entries()[0].annotation, "old-ann");
+
+    let previous = history.replace_entries(vec![HistoryEntry::new(
+        Item::user_message("restored"),
+        "restored-ann".to_string(),
+    )]);
+
+    assert_eq!(previous.len(), 1);
+    assert_eq!(history.entries()[0].item.as_text(), Some("restored"));
+    assert_eq!(history.entries()[0].annotation, "restored-ann");
+}

crates/agen/tests/callback_test.rs

@@ -58,6 +58,7 @@ async fn test_callback_llm_retry_event() {
         max_attempts: 2,
         total_timeout: Duration::from_secs(1),
     });
+    let mut history = agen::History::new();
 
     let notices = Arc::new(Mutex::new(Vec::new()));
     let sink = notices.clone();
@@ -65,7 +66,7 @@ async fn test_callback_llm_retry_event() {
         sink.lock().unwrap().push((llm_call, notice.clone()));
     });
 
-    let result = engine.run("retry once").await;
+    let result = engine.run(&mut history, "retry once").await;
     assert!(
         matches!(result.result, agen::EngineRunExit::Finished),
         "engine should succeed after one retry"
@@ -94,6 +95,7 @@ async fn test_callback_text_block_events() {
 
     let client = MockLlmClient::new(events);
     let mut engine = Engine::new(client);
+    let mut history = agen::History::new();
 
     let text_deltas = Arc::new(Mutex::new(Vec::new()));
     let text_completes = Arc::new(Mutex::new(Vec::new()));
@@ -112,7 +114,7 @@ async fn test_callback_text_block_events() {
     });
 
     // Mutable::run consumes self, returns (Locked, EngineRunExit)
-    let result = engine.run("Greet me").await;
+    let result = engine.run(&mut history, "Greet me").await;
     assert!(
         matches!(result.result, agen::EngineRunExit::Finished),
         "Engine should complete"
@@ -143,6 +145,7 @@ async fn test_callback_tool_call_complete() {
 
     let client = MockLlmClient::new(events);
     let mut engine = Engine::new(client);
+    let mut history = agen::History::new();
 
     let tool_starts = Arc::new(Mutex::new(Vec::<(String, String)>::new()));
     let tool_completes = Arc::new(Mutex::new(Vec::new()));
@@ -161,7 +164,7 @@ async fn test_callback_tool_call_complete() {
     });
 
     // Mutable::run consumes self, returns (Locked, EngineRunExit)
-    let _ = engine.run("Weather please").await;
+    let _ = engine.run(&mut history, "Weather please").await;
 
     let starts = tool_starts.lock().unwrap();
     assert_eq!(starts.len(), 1);
@@ -189,6 +192,7 @@ async fn test_callback_turn_events() {
 
     let client = MockLlmClient::new(events);
     let mut engine = Engine::new(client);
+    let mut history = agen::History::new();
 
     let turn_starts = Arc::new(Mutex::new(Vec::new()));
     let turn_ends = Arc::new(Mutex::new(Vec::new()));
@@ -204,7 +208,7 @@ async fn test_callback_turn_events() {
     });
 
     // Mutable::run consumes self, returns (Locked, EngineRunExit)
-    let result = engine.run("Do something").await;
+    let result = engine.run(&mut history, "Do something").await;
     assert!(matches!(result.result, agen::EngineRunExit::Finished));
 
     let starts = turn_starts.lock().unwrap();
@@ -260,6 +264,7 @@ async fn test_callback_tool_result_events() {
 
     let client = MockLlmClient::new(events);
     let mut engine = Engine::new(client);
+    let mut history = agen::History::new();
 
     engine.register_tool(fixed_tool(
         "fixed",
@@ -282,7 +287,7 @@ async fn test_callback_tool_result_events() {
         ));
     });
 
-    let _ = engine.run("call it").await;
+    let _ = engine.run(&mut history, "call it").await;
 
     let observed = captured.lock().unwrap();
     assert_eq!(observed.len(), 1);
@@ -336,6 +341,7 @@ async fn test_callback_tool_result_error_path() {
 
     let client = MockLlmClient::new(events);
     let mut engine = Engine::new(client);
+    let mut history = agen::History::new();
 
     engine.register_tool(erroring_tool("erroring", "boom"));
 
@@ -351,7 +357,7 @@ async fn test_callback_tool_result_error_path() {
         ));
     });
 
-    let _ = engine.run("fail it").await;
+    let _ = engine.run(&mut history, "fail it").await;
 
     let observed = captured.lock().unwrap();
     assert_eq!(observed.len(), 1);
@@ -380,6 +386,7 @@ async fn test_callback_usage_events() {
 
     let client = MockLlmClient::new(events);
     let mut engine = Engine::new(client);
+    let mut history = agen::History::new();
 
     let usage_events = Arc::new(Mutex::new(Vec::new()));
 
@@ -389,7 +396,7 @@ async fn test_callback_usage_events() {
     });
 
     // Mutable::run consumes self, returns (Locked, EngineRunExit)
-    let _ = engine.run("Hello").await;
+    let _ = engine.run(&mut history, "Hello").await;
 
     let usages = usage_events.lock().unwrap();
     assert_eq!(usages.len(), 1);

crates/agen/tests/common/mod.rs

@@ -19,6 +19,7 @@ use std::sync::atomic::{AtomicUsize, Ordering};
 pub struct MockLlmClient {
     responses: Arc<Vec<Vec<Event>>>,
     call_count: Arc<AtomicUsize>,
+    requests: Arc<Mutex<Vec<Request>>>,
 }
 
 impl MockLlmClient {
@@ -30,6 +31,7 @@ impl MockLlmClient {
         Self {
             responses: Arc::new(responses),
             call_count: Arc::new(AtomicUsize::new(0)),
+            requests: Arc::new(Mutex::new(Vec::new())),
         }
     }
 
@@ -41,6 +43,10 @@ impl MockLlmClient {
     pub fn event_count(&self) -> usize {
         self.responses.iter().map(|v| v.len()).sum()
     }
+
+    pub fn requests(&self) -> Vec<Request> {
+        self.requests.lock().unwrap().clone()
+    }
 }
 
 #[async_trait]
@@ -51,8 +57,9 @@ impl LlmClient for MockLlmClient {
 
     async fn stream(
         &self,
-        _request: Request,
+        request: Request,
     ) -> Result<Pin<Box<dyn Stream<Item = Result<Event, ClientError>> + Send>>, ClientError> {
+        self.requests.lock().unwrap().push(request);
         let count = self.call_count.fetch_add(1, Ordering::SeqCst);
         if count >= self.responses.len() {
             return Err(ClientError::Api {

crates/agen/tests/engine_fixtures.rs

@@ -134,9 +134,10 @@ async fn test_engine_simple_text_response() {
 
     let client = MockLlmClient::from_fixture(&fixture_path).unwrap();
     let engine = Engine::new(client);
+    let mut history = agen::History::new();
 
     // Send a simple message (Mutable::run consumes self, returns tuple)
-    let result = engine.run("Hello").await;
+    let result = engine.run(&mut history, "Hello").await;
 
     assert!(
         matches!(result.result, agen::EngineRunExit::Finished),
@@ -159,6 +160,7 @@ async fn test_engine_tool_call() {
 
     let client = MockLlmClient::from_fixture(&fixture_path).unwrap();
     let mut engine = Engine::new(client);
+    let mut history = agen::History::new();
 
     // Register tool
     let weather_tool = MockWeatherTool::new();
@@ -166,7 +168,9 @@ async fn test_engine_tool_call() {
     engine.register_tool(weather_tool.definition());
 
     // Send message (Mutable::run consumes self, returns tuple)
-    let _result = engine.run("What's the weather in Tokyo?").await;
+    let _result = engine
+        .run(&mut history, "What's the weather in Tokyo?")
+        .await;
 
     // Verify tool was called
     // Note: max_turns=1 so no request is sent after tool result
@@ -198,9 +202,10 @@ async fn test_engine_with_programmatic_events() {
 
     let client = MockLlmClient::new(events);
     let engine = Engine::new(client);
+    let mut history = agen::History::new();
 
     // Mutable::run consumes self, returns tuple
-    let result = engine.run("Greet me").await;
+    let result = engine.run(&mut history, "Greet me").await;
 
     assert!(
         matches!(result.result, agen::EngineRunExit::Finished),

crates/agen/tests/engine_state_test.rs

@@ -12,45 +12,12 @@ use agen::Item;
 use agen::interceptor::{
     Interceptor, PreRequestAction, PreToolAction, ToolCallInfo, TurnEndAction,
 };
-use agen::llm_client::ClientError;
 use agen::llm_client::event::{Event, ResponseStatus, StatusEvent};
 use agen::tool::{Tool, ToolDefinition, ToolError, ToolMeta, ToolOutput};
-use agen::{Engine, EngineError, EngineRunExit, StopReason};
+use agen::{Engine, EngineError, EngineRunExit, History, StopReason};
 use async_trait::async_trait;
 use common::MockLlmClient;
 
-#[test]
-fn engine_source_has_no_worker_owned_interruption_marker() {
-    let source = include_str!("../src/engine.rs");
-    assert!(!source.contains("last_run_interrupted"));
-    assert!(!source.contains("set_last_run_interrupted"));
-}
-
-#[test]
-fn run_exit_classifies_known_and_unexpected_stops_without_message_parsing() {
-    assert!(matches!(
-        EngineRunExit::from(Err(EngineError::Cancelled)),
-        EngineRunExit::Interrupted(StopReason::Cancelled)
-    ));
-    assert!(matches!(
-        EngineRunExit::from(Err(EngineError::Client(ClientError::ContextWindowExceeded))),
-        EngineRunExit::Interrupted(StopReason::ContextWindowExceeded)
-    ));
-
-    let message_only = EngineError::Client(ClientError::Api {
-        status: Some(400),
-        code: None,
-        message: "context_length_exceeded".to_string(),
-        retry_after: None,
-    });
-    assert!(matches!(
-        EngineRunExit::from(Err(message_only)),
-        EngineRunExit::Interrupted(StopReason::Unexpected(EngineError::Client(
-            ClientError::Api { .. }
-        )))
-    ));
-}
-
 // =============================================================================
 // Mutable State Tests
 // =============================================================================
@@ -75,36 +42,37 @@ fn test_mutable_set_system_prompt() {
 fn test_mutable_history_manipulation() {
     let client = MockLlmClient::new(vec![]);
     let mut engine = Engine::new(client);
+    let mut history: History = History::new();
 
     // Initial state is empty
-    assert!(engine.history().is_empty());
+    assert!(history.is_empty());
 
     // Add to history
     engine
-        .append_history(vec![Item::user_message("Hello")])
+        .append_history(&mut history, vec![Item::user_message("Hello")])
         .unwrap();
     engine
-        .append_history(vec![Item::assistant_message("Hi there!")])
+        .append_history(&mut history, vec![Item::assistant_message("Hi there!")])
         .unwrap();
-    assert_eq!(engine.history().len(), 2);
+    assert_eq!(history.len(), 2);
 
     // Append to history via the callback-aware API.
     engine
-        .append_history(vec![Item::user_message("How are you?")])
+        .append_history(&mut history, vec![Item::user_message("How are you?")])
         .unwrap();
-    assert_eq!(engine.history().len(), 3);
+    assert_eq!(history.len(), 3);
 
     // Clear history
-    engine.clear_history();
-    assert!(engine.history().is_empty());
+    engine.clear_history(&mut history);
+    assert!(history.is_empty());
 
     // Set history
     let items = vec![
         Item::user_message("Test"),
         Item::assistant_message("Response"),
     ];
-    engine.set_history(items);
-    assert_eq!(engine.history().len(), 2);
+    engine.set_history(&mut history, items);
+    assert_eq!(history.len(), 2);
 }
 
 /// Verify that Engine can be constructed using builder pattern
@@ -112,9 +80,10 @@ fn test_mutable_history_manipulation() {
 fn test_mutable_builder_pattern() {
     let client = MockLlmClient::new(vec![]);
     let engine = Engine::new(client).system_prompt("System prompt");
+    let history: History = History::new();
 
     assert_eq!(engine.get_system_prompt(), Some("System prompt"));
-    assert!(engine.history().is_empty());
+    assert!(history.is_empty());
 }
 
 /// Verify that multiple items can be added with append_history and callbacks fire.
@@ -124,6 +93,7 @@ fn test_mutable_append_history() {
     let observed = Arc::new(Mutex::new(Vec::new()));
     let observed_for_callback = Arc::clone(&observed);
     let mut engine = Engine::new(client);
+    let mut history: History = History::new();
     engine.on_history_append(move |item| {
         if let Some(text) = item.as_text() {
             observed_for_callback.lock().unwrap().push(text.to_string());
@@ -132,18 +102,21 @@ fn test_mutable_append_history() {
     });
 
     engine
-        .append_history(vec![Item::user_message("First")])
+        .append_history(&mut history, vec![Item::user_message("First")])
         .unwrap();
 
     engine
-        .append_history(vec![
+        .append_history(
+            &mut history,
+            vec![
                 Item::assistant_message("Response 1"),
                 Item::user_message("Second"),
                 Item::assistant_message("Response 2"),
-        ])
+            ],
+        )
         .unwrap();
 
-    assert_eq!(engine.history().len(), 4);
+    assert_eq!(history.len(), 4);
     assert_eq!(
         observed.lock().unwrap().as_slice(),
         ["First", "Response 1", "Second", "Response 2"]
@@ -218,6 +191,7 @@ async fn history_append_failure_stops_before_tool_execution() {
     ]);
     let tool = CountingTool::new("count_tool");
     let mut engine = Engine::new(client);
+    let mut history: History = History::new();
     engine.register_tool(tool.definition());
     engine.on_history_append(|item| {
         if item.is_tool_call() {
@@ -227,17 +201,15 @@ async fn history_append_failure_stops_before_tool_execution() {
         }
     });
 
-    let mut engine = engine.lock();
-    let exit = engine.run("use the tool").await;
+    let mut engine = engine.lock(&history);
+    let exit = engine.run(&mut history, "use the tool").await;
 
-    assert!(matches!(
-        exit,
-        EngineRunExit::Interrupted(StopReason::Unexpected(EngineError::HistoryAppend(ref message)))
-            if message == "simulated ENOSPC"
-    ));
+    assert!(
+        matches!(exit, EngineRunExit::Interrupted(StopReason::Unexpected(EngineError::HistoryAppend(ref message))) if message == "simulated ENOSPC")
+    );
     assert_eq!(tool.call_count(), 0);
-    assert_eq!(engine.history().len(), 1);
-    assert_eq!(engine.history()[0].as_text(), Some("use the tool"));
+    assert_eq!(history.len(), 1);
+    assert_eq!(history.entries()[0].item.as_text(), Some("use the tool"));
 }
 
 // =============================================================================
@@ -249,21 +221,22 @@ async fn history_append_failure_stops_before_tool_execution() {
 fn test_lock_transition() {
     let client = MockLlmClient::new(vec![]);
     let mut engine = Engine::new(client);
+    let mut history: History = History::new();
 
     engine.set_system_prompt("System");
     engine
-        .append_history(vec![Item::user_message("Hello")])
+        .append_history(&mut history, vec![Item::user_message("Hello")])
         .unwrap();
     engine
-        .append_history(vec![Item::assistant_message("Hi")])
+        .append_history(&mut history, vec![Item::assistant_message("Hi")])
         .unwrap();
 
     // Lock
-    let locked_engine = engine.lock();
+    let locked_engine = engine.lock(&history);
 
     // History and system prompt are still accessible in Locked state
     assert_eq!(locked_engine.get_system_prompt(), Some("System"));
-    assert_eq!(locked_engine.history().len(), 2);
+    assert_eq!(history.len(), 2);
     assert_eq!(locked_engine.locked_prefix_len(), 2);
 }
 
@@ -272,21 +245,22 @@ fn test_lock_transition() {
 fn test_unlock_transition() {
     let client = MockLlmClient::new(vec![]);
     let mut engine = Engine::new(client);
+    let mut history: History = History::new();
 
     engine
-        .append_history(vec![Item::user_message("Hello")])
+        .append_history(&mut history, vec![Item::user_message("Hello")])
         .unwrap();
-    let locked_engine = engine.lock();
+    let locked_engine = engine.lock(&history);
 
     // Unlock
     let mut engine = locked_engine.unlock();
 
     // History operations are available again in Mutable state
     engine
-        .append_history(vec![Item::assistant_message("Hi")])
+        .append_history(&mut history, vec![Item::assistant_message("Hi")])
         .unwrap();
-    engine.clear_history();
-    assert!(engine.history().is_empty());
+    engine.clear_history(&mut history);
+    assert!(history.is_empty());
 }
 
 // =============================================================================
@@ -307,20 +281,20 @@ async fn test_mutable_run_updates_history() -> Result<(), EngineError> {
 
     let client = MockLlmClient::new(events);
     let engine = Engine::new(client);
+    let mut history: History = History::new();
 
     // Execute (Mutable::run consumes self, returns EngineRunOutput)
-    let out = engine.run("Hi there").await;
-    let engine = out.engine;
+    let _out = engine.run(&mut history, "Hi there").await;
 
     // History is updated
-    let history = engine.history();
+    let entries = history.entries();
     assert_eq!(history.len(), 2); // user + assistant
 
     // User message
-    assert_eq!(history[0].as_text(), Some("Hi there"));
+    assert_eq!(entries[0].item.as_text(), Some("Hi there"));
 
     // Assistant message
-    assert_eq!(history[1].as_text(), Some("Hello, I'm an assistant!"));
+    assert_eq!(entries[1].item.as_text(), Some("Hello, I'm an assistant!"));
 
     Ok(())
 }
@@ -351,35 +325,36 @@ async fn test_locked_multi_turn_history_accumulation() {
     ]);
 
     let engine = Engine::new(client).system_prompt("You are helpful.");
+    let mut history: History = History::new();
 
     // Lock (after setting system prompt)
-    let mut locked_engine = engine.lock();
+    let mut locked_engine = engine.lock(&history);
     assert_eq!(locked_engine.locked_prefix_len(), 0); // No items yet
 
     // Turn 1
-    let result1 = locked_engine.run("Hello!").await;
+    let result1 = locked_engine.run(&mut history, "Hello!").await;
     assert!(matches!(result1, EngineRunExit::Finished));
-    assert_eq!(locked_engine.history().len(), 2); // user + assistant
+    assert_eq!(history.len(), 2); // user + assistant
 
     // Turn 2
-    let result2 = locked_engine.run("Can you help me?").await;
+    let result2 = locked_engine.run(&mut history, "Can you help me?").await;
     assert!(matches!(result2, EngineRunExit::Finished));
-    assert_eq!(locked_engine.history().len(), 4); // 2 * (user + assistant)
+    assert_eq!(history.len(), 4); // 2 * (user + assistant)
 
     // Verify history contents
-    let history = locked_engine.history();
+    let entries = history.entries();
 
     // Turn 1 user message
-    assert_eq!(history[0].as_text(), Some("Hello!"));
+    assert_eq!(entries[0].item.as_text(), Some("Hello!"));
 
     // Turn 1 assistant message
-    assert_eq!(history[1].as_text(), Some("Nice to meet you!"));
+    assert_eq!(entries[1].item.as_text(), Some("Nice to meet you!"));
 
     // Turn 2 user message
-    assert_eq!(history[2].as_text(), Some("Can you help me?"));
+    assert_eq!(entries[2].item.as_text(), Some("Can you help me?"));
 
     // Turn 2 assistant message
-    assert_eq!(history[3].as_text(), Some("I can help with that."));
+    assert_eq!(entries[3].item.as_text(), Some("I can help with that."));
 }
 
 /// Verify that locked_prefix_len correctly records history length at lock time
@@ -405,26 +380,33 @@ async fn test_locked_prefix_len_tracking() {
     ]);
 
     let mut engine = Engine::new(client);
+    let mut history: History = History::new();
 
     // Add items beforehand
     engine
-        .append_history(vec![Item::user_message("Pre-existing message 1")])
+        .append_history(
+            &mut history,
+            vec![Item::user_message("Pre-existing message 1")],
+        )
         .unwrap();
     engine
-        .append_history(vec![Item::assistant_message("Pre-existing response 1")])
+        .append_history(
+            &mut history,
+            vec![Item::assistant_message("Pre-existing response 1")],
+        )
         .unwrap();
 
-    assert_eq!(engine.history().len(), 2);
+    assert_eq!(history.len(), 2);
 
     // Lock
-    let mut locked_engine = engine.lock();
+    let mut locked_engine = engine.lock(&history);
     assert_eq!(locked_engine.locked_prefix_len(), 2); // 2 items at lock time
 
     // Execute turn
-    locked_engine.run("New message").await;
+    locked_engine.run(&mut history, "New message").await;
 
     // History grows but locked_prefix_len remains unchanged
-    assert_eq!(locked_engine.history().len(), 4); // 2 + 2
+    assert_eq!(history.len(), 4); // 2 + 2
     assert_eq!(locked_engine.locked_prefix_len(), 2); // Unchanged
 }
 
@@ -451,18 +433,22 @@ async fn test_turn_count_increment() -> Result<(), EngineError> {
     ]);
 
     let engine = Engine::new(client);
+    let mut history: History = History::new();
 
     assert_eq!(engine.turn_count(), 0);
     assert_eq!(engine.llm_call_count(), 0);
 
     // First run consumes Mutable, returns EngineRunOutput
-    let mut engine = engine.run("First").await.engine;
+    let mut engine = engine.run(&mut history, "First").await.engine;
     assert_eq!(engine.turn_count(), 1);
     // Retry not yet implemented → AgentTurn:LlmCall is 1:1.
     assert_eq!(engine.llm_call_count(), 1);
 
     // Subsequent runs on Locked take &mut self
-    engine.run("Second").await;
+    assert!(matches!(
+        engine.run(&mut history, "Second").await,
+        EngineRunExit::Finished
+    ));
     assert_eq!(engine.turn_count(), 2);
     assert_eq!(engine.llm_call_count(), 2);
 
@@ -482,28 +468,29 @@ async fn test_unlock_edit_relock() {
     ]]);
 
     let mut engine = Engine::new(client);
+    let mut history: History = History::new();
     engine
-        .append_history(vec![
-            Item::user_message("Hello"),
-            Item::assistant_message("Hi"),
-        ])
+        .append_history(
+            &mut history,
+            vec![Item::user_message("Hello"), Item::assistant_message("Hi")],
+        )
         .unwrap();
 
     // Lock -> Unlock
-    let locked = engine.lock();
+    let locked = engine.lock(&history);
     assert_eq!(locked.locked_prefix_len(), 2);
 
     let mut unlocked = locked.unlock();
 
     // Edit history
-    unlocked.clear_history();
+    unlocked.clear_history(&mut history);
     unlocked
-        .append_history(vec![Item::user_message("Fresh start")])
+        .append_history(&mut history, vec![Item::user_message("Fresh start")])
         .unwrap();
 
     // Re-lock
-    let relocked = unlocked.lock();
-    assert_eq!(relocked.history().len(), 1);
+    let relocked = unlocked.lock(&history);
+    assert_eq!(history.len(), 1);
     assert_eq!(relocked.locked_prefix_len(), 1);
 }
 
@@ -546,19 +533,26 @@ async fn test_lock_unlock_relock_tools_remain_effective() {
     ]);
 
     let mut engine = Engine::new(client);
+    let mut history: History = History::new();
     let tool_a = CountingTool::new("tool_a");
     engine.register_tool(tool_a.definition());
 
-    let mut locked = engine.lock();
-    locked.run("first").await;
+    let mut locked = engine.lock(&history);
+    assert!(matches!(
+        locked.run(&mut history, "first").await,
+        EngineRunExit::Finished
+    ));
     assert_eq!(tool_a.call_count(), 1, "tool_a should be called once");
 
     let mut unlocked = locked.unlock();
     let tool_b = CountingTool::new("tool_b");
     unlocked.register_tool(tool_b.definition());
 
-    let mut relocked = unlocked.lock();
-    relocked.run("second").await;
+    let mut relocked = unlocked.lock(&history);
+    assert!(matches!(
+        relocked.run(&mut history, "second").await,
+        EngineRunExit::Finished
+    ));
 
     assert_eq!(tool_a.call_count(), 1, "tool_a should not be called again");
     assert_eq!(tool_b.call_count(), 1, "tool_b should be called once");
@@ -573,8 +567,9 @@ async fn test_lock_unlock_relock_tools_remain_effective() {
 fn test_system_prompt_preserved_in_locked_state() {
     let client = MockLlmClient::new(vec![]);
     let engine = Engine::new(client).system_prompt("Important system prompt");
+    let history: History = History::new();
 
-    let locked = engine.lock();
+    let locked = engine.lock(&history);
     assert_eq!(locked.get_system_prompt(), Some("Important system prompt"));
 
     let unlocked = locked.unlock();
@@ -589,14 +584,15 @@ fn test_system_prompt_preserved_in_locked_state() {
 fn test_system_prompt_change_after_unlock() {
     let client = MockLlmClient::new(vec![]);
     let engine = Engine::new(client).system_prompt("Original prompt");
+    let history: History = History::new();
 
-    let locked = engine.lock();
+    let locked = engine.lock(&history);
     let mut unlocked = locked.unlock();
 
     unlocked.set_system_prompt("New prompt");
     assert_eq!(unlocked.get_system_prompt(), Some("New prompt"));
 
-    let relocked = unlocked.lock();
+    let relocked = unlocked.lock(&history);
     assert_eq!(relocked.get_system_prompt(), Some("New prompt"));
 }
 
@@ -658,17 +654,21 @@ impl Interceptor for ContinueTurnOnce {
 
 #[tokio::test]
 async fn max_turns_is_scoped_to_each_fresh_run() {
+    let mut history: History = History::new();
     let responses = vec![completed_text_events(), completed_text_events()];
     let mut engine = Engine::new(MockLlmClient::with_responses(responses));
     engine.set_max_turns(Some(1));
-    let mut engine = engine.lock();
+    let mut engine = engine.lock(&history);
 
-    assert!(matches!(engine.run("first").await, EngineRunExit::Finished));
+    assert!(matches!(
+        engine.run(&mut history, "first").await,
+        EngineRunExit::Finished
+    ));
     assert_eq!(engine.turn_count(), 1);
     assert_eq!(engine.active_run_turn_count(), None);
 
     assert!(matches!(
-        engine.run("second").await,
+        engine.run(&mut history, "second").await,
         EngineRunExit::Finished
     ));
     assert_eq!(engine.turn_count(), 2);
@@ -677,24 +677,32 @@ async fn max_turns_is_scoped_to_each_fresh_run() {
 
 #[tokio::test]
 async fn yielded_resume_keeps_the_same_unspent_turn_budget() {
+    let mut history: History = History::new();
     let mut engine = Engine::new(MockLlmClient::new(completed_text_events()));
     engine.set_max_turns(Some(1));
     engine.set_interceptor(YieldOnce {
         calls: AtomicUsize::new(0),
     });
-    let mut engine = engine.lock();
+    let mut engine = engine.lock(&history);
 
-    assert!(matches!(engine.run("start").await, EngineRunExit::Yielded));
+    assert!(matches!(
+        engine.run(&mut history, "start").await,
+        EngineRunExit::Yielded
+    ));
     assert_eq!(engine.turn_count(), 0);
     assert_eq!(engine.active_run_turn_count(), Some(0));
 
-    assert!(matches!(engine.resume().await, EngineRunExit::Finished));
+    assert!(matches!(
+        engine.resume(&mut history).await,
+        EngineRunExit::Finished
+    ));
     assert_eq!(engine.turn_count(), 1);
     assert_eq!(engine.active_run_turn_count(), None);
 }
 
 #[tokio::test]
 async fn paused_tool_resume_does_not_reset_the_consumed_turn_budget() {
+    let mut history: History = History::new();
     let events = vec![
         Event::tool_use_start(0, "call_1", "count_tool"),
         Event::tool_input_delta(0, "{}"),
@@ -710,15 +718,18 @@ async fn paused_tool_resume_does_not_reset_the_consumed_turn_budget() {
     engine.set_interceptor(PauseToolOnce {
         calls: AtomicUsize::new(0),
     });
-    let mut engine = engine.lock();
+    let mut engine = engine.lock(&history);
 
-    assert!(matches!(engine.run("call it").await, EngineRunExit::Paused));
+    assert!(matches!(
+        engine.run(&mut history, "call it").await,
+        EngineRunExit::Paused
+    ));
     assert_eq!(engine.turn_count(), 1);
     assert_eq!(engine.active_run_turn_count(), Some(1));
     assert_eq!(tool.call_count(), 0);
 
     assert!(matches!(
-        engine.resume().await,
+        engine.resume(&mut history).await,
         EngineRunExit::Interrupted(StopReason::LimitReached)
     ));
     assert_eq!(engine.turn_count(), 1);
@@ -728,6 +739,7 @@ async fn paused_tool_resume_does_not_reset_the_consumed_turn_budget() {
 
 #[tokio::test]
 async fn fresh_input_abandons_a_paused_run_and_starts_a_new_budget() {
+    let mut history: History = History::new();
     let tool_events = vec![
         Event::tool_use_start(0, "call_1", "count_tool"),
         Event::tool_input_delta(0, "{}"),
@@ -744,13 +756,16 @@ async fn fresh_input_abandons_a_paused_run_and_starts_a_new_budget() {
     engine.set_interceptor(PauseToolOnce {
         calls: AtomicUsize::new(0),
     });
-    let mut engine = engine.lock();
+    let mut engine = engine.lock(&history);
 
-    assert!(matches!(engine.run("pause").await, EngineRunExit::Paused));
+    assert!(matches!(
+        engine.run(&mut history, "pause").await,
+        EngineRunExit::Paused
+    ));
     assert_eq!(engine.active_run_turn_count(), Some(1));
 
     assert!(matches!(
-        engine.run("replace").await,
+        engine.run(&mut history, "replace").await,
         EngineRunExit::Finished
     ));
     assert_eq!(engine.turn_count(), 2);
@@ -760,15 +775,16 @@ async fn fresh_input_abandons_a_paused_run_and_starts_a_new_budget() {
 
 #[tokio::test]
 async fn interceptor_continuation_consumes_the_logical_run_budget() {
+    let mut history: History = History::new();
     let mut engine = Engine::new(MockLlmClient::new(completed_text_events()));
     engine.set_max_turns(Some(1));
     engine.set_interceptor(ContinueTurnOnce {
         calls: AtomicUsize::new(0),
     });
-    let mut engine = engine.lock();
+    let mut engine = engine.lock(&history);
 
     assert!(matches!(
-        engine.run("start").await,
+        engine.run(&mut history, "start").await,
         EngineRunExit::Interrupted(StopReason::LimitReached)
     ));
     assert_eq!(engine.turn_count(), 1);
@@ -778,14 +794,15 @@ async fn interceptor_continuation_consumes_the_logical_run_budget() {
 
 #[tokio::test]
 async fn restored_active_run_budget_is_enforced_before_another_llm_call() {
+    let mut history: History = History::new();
     let mut engine = Engine::new(MockLlmClient::new(completed_text_events()));
     engine.set_max_turns(Some(1));
     engine.set_turn_count(7);
     engine.set_active_run_turn_count(Some(1));
-    let mut engine = engine.lock();
+    let mut engine = engine.lock(&history);
 
     assert!(matches!(
-        engine.resume().await,
+        engine.resume(&mut history).await,
         EngineRunExit::Interrupted(StopReason::LimitReached)
     ));
     assert_eq!(engine.turn_count(), 7);

crates/agen/tests/parallel_execution_test.rs

@@ -6,12 +6,13 @@ use std::sync::atomic::{AtomicUsize, Ordering};
 use std::sync::{Arc, Mutex};
 use std::time::{Duration, Instant};
 
-use agen::Engine;
 use agen::interceptor::{Interceptor, PostToolAction, PreToolAction, ToolCallInfo, ToolResultInfo};
 use agen::llm_client::event::{Event, ResponseStatus, StatusEvent};
 use agen::tool::{
     Tool, ToolDefinition, ToolError, ToolExecutionContext, ToolMeta, ToolOutput, ToolResult,
+    ToolResultDisposition,
 };
+use agen::{Engine, History, Item, ToolExecutionPolicy};
 use async_trait::async_trait;
 
 mod common;
@@ -70,6 +71,144 @@ impl Tool for SlowTool {
     }
 }
 
+#[derive(Clone)]
+struct FirstAttemptHangsTool {
+    calls: Arc<AtomicUsize>,
+}
+
+impl FirstAttemptHangsTool {
+    fn new() -> Self {
+        Self {
+            calls: Arc::new(AtomicUsize::new(0)),
+        }
+    }
+
+    fn definition(&self) -> ToolDefinition {
+        let tool = self.clone();
+        Arc::new(move || {
+            let meta = ToolMeta::new("hang_once")
+                .description("Hangs on the first execution attempt")
+                .input_schema(serde_json::json!({"type": "object"}));
+            (meta, Arc::new(tool.clone()) as Arc<dyn Tool>)
+        })
+    }
+
+    fn call_count(&self) -> usize {
+        self.calls.load(Ordering::SeqCst)
+    }
+}
+
+#[async_trait]
+impl Tool for FirstAttemptHangsTool {
+    async fn execute(
+        &self,
+        _input_json: &str,
+        _ctx: ToolExecutionContext,
+    ) -> Result<ToolOutput, ToolError> {
+        let attempt = self.calls.fetch_add(1, Ordering::SeqCst);
+        if attempt == 0 {
+            std::future::pending::<()>().await;
+        }
+        Ok("completed on retry".to_string().into())
+    }
+}
+
+#[derive(Clone)]
+struct CooperativeCancelTool {
+    calls: Arc<AtomicUsize>,
+    cancelled: Arc<tokio::sync::Notify>,
+}
+
+impl CooperativeCancelTool {
+    fn new() -> Self {
+        Self {
+            calls: Arc::new(AtomicUsize::new(0)),
+            cancelled: Arc::new(tokio::sync::Notify::new()),
+        }
+    }
+
+    fn definition(&self) -> ToolDefinition {
+        let tool = self.clone();
+        Arc::new(move || {
+            let meta = ToolMeta::new("cooperative")
+                .description("Returns bounded progress after cancellation")
+                .input_schema(serde_json::json!({"type": "object"}));
+            (meta, Arc::new(tool.clone()) as Arc<dyn Tool>)
+        })
+    }
+}
+
+#[async_trait]
+impl Tool for CooperativeCancelTool {
+    async fn execute(
+        &self,
+        _input_json: &str,
+        _ctx: ToolExecutionContext,
+    ) -> Result<ToolOutput, ToolError> {
+        self.calls.fetch_add(1, Ordering::SeqCst);
+        self.cancelled.notified().await;
+        Err(ToolError::Cancelled(ToolOutput {
+            summary: "cooperative command cancelled".to_string(),
+            content: Some("stdout before cancellation\nstderr before cancellation".to_string()),
+            attachments: Vec::new(),
+        }))
+    }
+
+    async fn cancel(&self, _call_id: &str) -> Result<(), ToolError> {
+        self.cancelled.notify_one();
+        Ok(())
+    }
+}
+
+#[derive(Clone)]
+struct SafePauseTool {
+    calls: Arc<AtomicUsize>,
+    cancellations: Arc<AtomicUsize>,
+    release: Arc<tokio::sync::Notify>,
+}
+
+impl SafePauseTool {
+    fn new() -> Self {
+        Self {
+            calls: Arc::new(AtomicUsize::new(0)),
+            cancellations: Arc::new(AtomicUsize::new(0)),
+            release: Arc::new(tokio::sync::Notify::new()),
+        }
+    }
+
+    fn definition(&self) -> ToolDefinition {
+        let tool = self.clone();
+        Arc::new(move || {
+            let meta = ToolMeta::new("safe_pause")
+                .description("Waits for a safe-boundary release")
+                .input_schema(serde_json::json!({"type": "object"}));
+            (meta, Arc::new(tool.clone()) as Arc<dyn Tool>)
+        })
+    }
+}
+
+#[async_trait]
+impl Tool for SafePauseTool {
+    async fn execute(
+        &self,
+        _input_json: &str,
+        _ctx: ToolExecutionContext,
+    ) -> Result<ToolOutput, ToolError> {
+        self.calls.fetch_add(1, Ordering::SeqCst);
+        self.release.notified().await;
+        Ok(ToolOutput {
+            summary: "safe-boundary complete".to_string(),
+            content: Some("safe-boundary complete".to_string()),
+            attachments: Vec::new(),
+        })
+    }
+
+    async fn cancel(&self, _call_id: &str) -> Result<(), ToolError> {
+        self.cancellations.fetch_add(1, Ordering::SeqCst);
+        Ok(())
+    }
+}
+
 #[derive(Clone)]
 struct ContextRecordingTool {
     name: String,
@@ -145,6 +284,7 @@ async fn test_parallel_tool_execution() {
         ],
     ]);
     let mut engine = Engine::new(client);
+    let mut history: History = History::new();
     let tool1 = SlowTool::new("slow_tool_1", 100);
     let tool2 = SlowTool::new("slow_tool_2", 100);
     let tool3 = SlowTool::new("slow_tool_3", 100);
@@ -158,8 +298,8 @@ async fn test_parallel_tool_execution() {
     engine.register_tool(tool3.definition());
 
     let start = Instant::now();
-    // Mutable::run consumes self, returns (Locked, EngineRunExit)
-    let _result = engine.run("Run all tools").await;
+    // Mutable::run consumes self, returns (Locked, EngineResult)
+    let _result = engine.run(&mut history, "Run all tools").await;
     let elapsed = start.elapsed();
 
     // Verify all tools were called
@@ -178,6 +318,450 @@ async fn test_parallel_tool_execution() {
     println!("Parallel execution completed in {:?}", elapsed);
 }
 
+#[tokio::test]
+async fn completed_results_commit_before_publish_without_waiting_for_siblings() {
+    let client = MockLlmClient::with_responses(vec![
+        vec![
+            Event::tool_use_start(0, "call_slow", "slow_first"),
+            Event::tool_input_delta(0, r#"{}"#),
+            Event::tool_use_stop(0),
+            Event::tool_use_start(1, "call_fast", "fast_second"),
+            Event::tool_input_delta(1, r#"{}"#),
+            Event::tool_use_stop(1),
+            Event::Status(StatusEvent {
+                status: ResponseStatus::Completed,
+            }),
+        ],
+        vec![
+            Event::text_block_start(0),
+            Event::text_delta(0, "Done"),
+            Event::text_block_stop(0, None),
+            Event::Status(StatusEvent {
+                status: ResponseStatus::Completed,
+            }),
+        ],
+    ]);
+    let client_probe = client.clone();
+    let mut engine = Engine::new(client);
+    engine.register_tool(SlowTool::new("slow_first", 100).definition());
+    engine.register_tool(SlowTool::new("fast_second", 5).definition());
+
+    let observed = Arc::new(Mutex::new(Vec::<String>::new()));
+    let published = observed.clone();
+    engine.on_tool_result(move |result| {
+        published
+            .lock()
+            .unwrap()
+            .push(format!("publish:{}", result.tool_use_id));
+    });
+
+    let committed = observed.clone();
+    let mut annotate = move |item: &Item| {
+        if let Item::ToolResult { call_id, .. } = item {
+            committed.lock().unwrap().push(format!("commit:{call_id}"));
+        }
+        Ok(())
+    };
+    let mut history = History::new();
+    let _ = engine
+        .run_with_annotation(&mut history, "run both", &mut annotate)
+        .await;
+    observed.lock().unwrap().push("run-returned".to_string());
+
+    assert_eq!(
+        observed.lock().unwrap().as_slice(),
+        [
+            "commit:call_fast",
+            "publish:call_fast",
+            "commit:call_slow",
+            "publish:call_slow",
+            "run-returned",
+        ]
+    );
+
+    let committed_order: Vec<_> = history
+        .iter()
+        .filter_map(|entry| match &entry.item {
+            Item::ToolResult { call_id, .. } => Some(call_id.as_str()),
+            _ => None,
+        })
+        .collect();
+    assert_eq!(committed_order, ["call_fast", "call_slow"]);
+
+    let requests = client_probe.requests();
+    let projected_order: Vec<_> = requests[1]
+        .items
+        .iter()
+        .filter_map(|item| match item {
+            Item::ToolResult { call_id, .. } => Some(call_id.as_str()),
+            _ => None,
+        })
+        .collect();
+    assert_eq!(projected_order, ["call_slow", "call_fast"]);
+}
+
+#[tokio::test]
+async fn cancellation_preserves_completed_results_and_resume_skips_them() {
+    let client = MockLlmClient::with_responses(vec![
+        vec![
+            Event::tool_use_start(0, "call_hang", "hang_once"),
+            Event::tool_input_delta(0, r#"{}"#),
+            Event::tool_use_stop(0),
+            Event::tool_use_start(1, "call_fast_a", "fast_a"),
+            Event::tool_input_delta(1, r#"{}"#),
+            Event::tool_use_stop(1),
+            Event::tool_use_start(2, "call_fast_b", "fast_b"),
+            Event::tool_input_delta(2, r#"{}"#),
+            Event::tool_use_stop(2),
+            Event::Status(StatusEvent {
+                status: ResponseStatus::Completed,
+            }),
+        ],
+        vec![
+            Event::text_block_start(0),
+            Event::text_delta(0, "Recovered"),
+            Event::text_block_stop(0, None),
+            Event::Status(StatusEvent {
+                status: ResponseStatus::Completed,
+            }),
+        ],
+    ]);
+    let mut engine = Engine::new(client);
+    let hanging = FirstAttemptHangsTool::new();
+    let fast_a = SlowTool::new("fast_a", 1);
+    let fast_b = SlowTool::new("fast_b", 2);
+    engine.register_tool(hanging.definition());
+    engine.register_tool(fast_a.definition());
+    engine.register_tool(fast_b.definition());
+
+    let cancel = engine.cancel_sender();
+    let cancel_task = tokio::spawn(async move {
+        tokio::time::sleep(Duration::from_millis(30)).await;
+        cancel.send(()).await.unwrap();
+    });
+    let mut history = History::new();
+    let output = engine.run(&mut history, "start").await;
+    let mut engine = output.engine;
+    cancel_task.await.unwrap();
+
+    let completed_before_resume = history
+        .iter()
+        .filter(|entry| {
+            matches!(
+                &entry.item,
+                Item::ToolResult { call_id, .. }
+                    if call_id == "call_fast_a" || call_id == "call_fast_b"
+            )
+        })
+        .count();
+    let unknown_before_resume = history
+        .iter()
+        .filter(|entry| {
+            matches!(
+                &entry.item,
+                Item::ToolResult {
+                    call_id,
+                    disposition: ToolResultDisposition::OutcomeUnknown,
+                    ..
+                } if call_id == "call_hang"
+            )
+        })
+        .count();
+    assert_eq!(completed_before_resume, 2);
+    assert_eq!(unknown_before_resume, 1);
+    assert_eq!(fast_a.call_count(), 1);
+    assert_eq!(fast_b.call_count(), 1);
+    assert_eq!(hanging.call_count(), 1);
+
+    let _ = engine.resume(&mut history).await;
+
+    assert_eq!(
+        fast_a.call_count(),
+        1,
+        "completed call must not be re-executed"
+    );
+    assert_eq!(
+        fast_b.call_count(),
+        1,
+        "completed call must not be re-executed"
+    );
+    assert_eq!(
+        hanging.call_count(),
+        1,
+        "OutcomeUnknown is terminal and must not be re-executed"
+    );
+    let completed_after_resume = history
+        .iter()
+        .filter(|entry| {
+            matches!(
+                &entry.item,
+                Item::ToolResult { call_id, .. }
+                    if call_id == "call_fast_a" || call_id == "call_fast_b"
+            )
+        })
+        .count();
+    assert_eq!(completed_after_resume, 2);
+    assert_eq!(
+        history
+            .iter()
+            .filter(|entry| {
+                matches!(
+                    &entry.item,
+                    Item::ToolResult {
+                        call_id,
+                        disposition: ToolResultDisposition::OutcomeUnknown,
+                        ..
+                    } if call_id == "call_hang"
+                )
+            })
+            .count(),
+        1
+    );
+}
+
+#[tokio::test]
+async fn cooperative_cancellation_commits_bounded_terminal_output() {
+    let client = MockLlmClient::with_responses(vec![vec![
+        Event::tool_use_start(0, "call_cooperative", "cooperative"),
+        Event::tool_input_delta(0, r#"{}"#),
+        Event::tool_use_stop(0),
+        Event::Status(StatusEvent {
+            status: ResponseStatus::Completed,
+        }),
+    ]]);
+    let mut engine = Engine::new(client);
+    let tool = CooperativeCancelTool::new();
+    engine.register_tool(tool.definition());
+    let observed = Arc::new(Mutex::new(Vec::<&'static str>::new()));
+    let published = observed.clone();
+    engine.on_tool_result(move |_| published.lock().unwrap().push("published"));
+    let committed = observed.clone();
+    let mut annotate = move |item: &Item| {
+        if matches!(item, Item::ToolResult { .. }) {
+            committed.lock().unwrap().push("committed");
+        }
+        Ok(())
+    };
+
+    let cancel = engine.cancel_sender();
+    let cancel_task = tokio::spawn(async move {
+        tokio::time::sleep(Duration::from_millis(30)).await;
+        cancel.send(()).await.unwrap();
+    });
+    let mut history = History::new();
+    let output = engine
+        .run_with_annotation(&mut history, "start", &mut annotate)
+        .await;
+    observed.lock().unwrap().push("run-returned");
+    cancel_task.await.unwrap();
+
+    assert_eq!(
+        observed.lock().unwrap().as_slice(),
+        ["committed", "published", "run-returned"]
+    );
+    assert_eq!(tool.calls.load(Ordering::SeqCst), 1);
+    let terminal: Vec<_> = history
+        .iter()
+        .filter_map(|entry| match &entry.item {
+            Item::ToolResult {
+                call_id,
+                disposition,
+                content,
+                ..
+            } if call_id == "call_cooperative" => Some((*disposition, content.as_deref())),
+            _ => None,
+        })
+        .collect();
+    assert_eq!(terminal.len(), 1);
+    assert_eq!(terminal[0].0, ToolResultDisposition::Cancelled);
+    assert_eq!(
+        terminal[0].1,
+        Some("stdout before cancellation\nstderr before cancellation")
+    );
+    assert!(matches!(
+        output.result,
+        agen::EngineRunExit::Interrupted(agen::StopReason::Cancelled)
+    ));
+}
+
+#[tokio::test]
+async fn pause_waits_for_started_tool_terminal_without_cancelling_provider() {
+    let client = MockLlmClient::with_responses(vec![vec![
+        Event::tool_use_start(0, "call_safe_pause", "safe_pause"),
+        Event::tool_input_delta(0, r#"{}"#),
+        Event::tool_use_stop(0),
+        Event::Status(StatusEvent {
+            status: ResponseStatus::Completed,
+        }),
+    ]]);
+    let mut engine = Engine::new(client);
+    let tool = SafePauseTool::new();
+    engine.register_tool(tool.definition());
+
+    let pause = engine.pause_sender();
+    let calls = Arc::clone(&tool.calls);
+    let release = Arc::clone(&tool.release);
+    let control = tokio::spawn(async move {
+        tokio::time::timeout(Duration::from_secs(1), async {
+            while calls.load(Ordering::SeqCst) == 0 {
+                tokio::task::yield_now().await;
+            }
+        })
+        .await
+        .expect("tool execution starts");
+        pause.send(()).await.unwrap();
+        tokio::time::sleep(Duration::from_millis(50)).await;
+        release.notify_one();
+    });
+
+    let started_at = std::time::Instant::now();
+    let mut history = History::new();
+    let output = engine.run(&mut history, "pause safely").await;
+    control.await.unwrap();
+
+    assert!(started_at.elapsed() >= Duration::from_millis(50));
+    assert_eq!(tool.calls.load(Ordering::SeqCst), 1);
+    assert_eq!(tool.cancellations.load(Ordering::SeqCst), 0);
+    assert!(matches!(output.result, agen::EngineRunExit::Paused));
+    assert!(history.iter().any(|entry| matches!(
+        &entry.item,
+        Item::ToolResult {
+            call_id,
+            disposition: ToolResultDisposition::Success,
+            ..
+        } if call_id == "call_safe_pause"
+    )));
+}
+
+#[tokio::test]
+async fn pause_escalates_to_explicit_cancel_and_confirm_after_safe_boundary_deadline() {
+    let client = MockLlmClient::with_responses(vec![vec![
+        Event::tool_use_start(0, "call_pause_cancel", "cooperative"),
+        Event::tool_input_delta(0, r#"{}"#),
+        Event::tool_use_stop(0),
+        Event::Status(StatusEvent {
+            status: ResponseStatus::Completed,
+        }),
+    ]]);
+    let mut engine = Engine::new(client);
+    engine.set_tool_execution_policy(ToolExecutionPolicy {
+        pause_safe_boundary_timeout: Duration::from_millis(20),
+        cancellation_request_timeout: Duration::from_millis(50),
+        terminal_confirmation_timeout: Duration::from_millis(100),
+    });
+    let tool = CooperativeCancelTool::new();
+    engine.register_tool(tool.definition());
+
+    let pause = engine.pause_sender();
+    let calls = Arc::clone(&tool.calls);
+    let control = tokio::spawn(async move {
+        tokio::time::timeout(Duration::from_secs(1), async {
+            while calls.load(Ordering::SeqCst) == 0 {
+                tokio::task::yield_now().await;
+            }
+        })
+        .await
+        .expect("tool execution starts");
+        pause.send(()).await.unwrap();
+    });
+
+    let mut history = History::new();
+    let output = engine.run(&mut history, "pause with escalation").await;
+    control.await.unwrap();
+
+    assert!(matches!(output.result, agen::EngineRunExit::Paused));
+    assert!(history.iter().any(|entry| matches!(
+        &entry.item,
+        Item::ToolResult {
+            call_id,
+            disposition: ToolResultDisposition::Cancelled,
+            ..
+        } if call_id == "call_pause_cancel"
+    )));
+}
+
+#[tokio::test]
+async fn cancellation_completion_race_commits_one_terminal_output() {
+    for iteration in 0..24u64 {
+        let client = MockLlmClient::with_responses(vec![
+            vec![
+                Event::tool_use_start(0, "call_racy", "racy"),
+                Event::tool_input_delta(0, r#"{}"#),
+                Event::tool_use_stop(0),
+                Event::Status(StatusEvent {
+                    status: ResponseStatus::Completed,
+                }),
+            ],
+            vec![Event::Status(StatusEvent {
+                status: ResponseStatus::Completed,
+            })],
+        ]);
+        let mut engine = Engine::new(client);
+        let delay = 2 + iteration % 3;
+        let tool = SlowTool::new("racy", delay);
+        engine.register_tool(tool.definition());
+        let cancel = engine.cancel_sender();
+        let cancel_task = tokio::spawn(async move {
+            tokio::time::sleep(Duration::from_millis(delay)).await;
+            let _ = cancel.send(()).await;
+        });
+
+        let mut history = History::new();
+        let _ = engine.run(&mut history, "race").await;
+        cancel_task.await.unwrap();
+        let terminal_count = history
+            .iter()
+            .filter(|entry| {
+                matches!(
+                    &entry.item,
+                    Item::ToolResult { call_id, .. } if call_id == "call_racy"
+                )
+            })
+            .count();
+        assert_eq!(terminal_count, 1, "iteration {iteration}");
+        assert_eq!(tool.call_count(), 1, "iteration {iteration}");
+    }
+}
+
+#[tokio::test]
+async fn tool_result_commit_failure_prevents_publication() {
+    let client = MockLlmClient::with_responses(vec![vec![
+        Event::tool_use_start(0, "call_fast", "fast"),
+        Event::tool_input_delta(0, r#"{}"#),
+        Event::tool_use_stop(0),
+        Event::Status(StatusEvent {
+            status: ResponseStatus::Completed,
+        }),
+    ]]);
+    let mut engine = Engine::new(client);
+    engine.register_tool(SlowTool::new("fast", 1).definition());
+
+    let published = Arc::new(AtomicUsize::new(0));
+    let published_probe = published.clone();
+    engine.on_tool_result(move |_| {
+        published_probe.fetch_add(1, Ordering::SeqCst);
+    });
+
+    let mut history = History::new();
+    let mut reject_tool_result = |item: &Item| {
+        if matches!(item, Item::ToolResult { .. }) {
+            Err("session log unavailable".to_string())
+        } else {
+            Ok(())
+        }
+    };
+    let _ = engine
+        .run_with_annotation(&mut history, "start", &mut reject_tool_result)
+        .await;
+
+    assert_eq!(published.load(Ordering::SeqCst), 0);
+    assert!(
+        history
+            .iter()
+            .all(|entry| !matches!(entry.item, Item::ToolResult { .. }))
+    );
+}
+
 #[tokio::test]
 async fn test_tool_execution_context_order_and_batch_id() {
     let client = MockLlmClient::with_responses(vec![
@@ -205,13 +789,14 @@ async fn test_tool_execution_context_order_and_batch_id() {
         ],
     ]);
     let mut engine = Engine::new(client);
+    let mut history: History = History::new();
     let contexts = Arc::new(Mutex::new(Vec::new()));
 
     engine.register_tool(ContextRecordingTool::new("record_a", contexts.clone()).definition());
     engine.register_tool(ContextRecordingTool::new("record_b", contexts.clone()).definition());
     engine.register_tool(ContextRecordingTool::new("record_c", contexts.clone()).definition());
 
-    let _ = engine.run("record contexts").await;
+    let _ = engine.run(&mut history, "record contexts").await;
 
     let mut contexts = contexts.lock().unwrap().clone();
     contexts.sort_by_key(|ctx| ctx.call_index);
@@ -256,11 +841,12 @@ async fn test_tool_execution_context_batch_id_changes_between_batches() {
         ],
     ]);
     let mut engine = Engine::new(client);
+    let mut history: History = History::new();
     let contexts = Arc::new(Mutex::new(Vec::new()));
 
     engine.register_tool(ContextRecordingTool::new("record", contexts.clone()).definition());
 
-    let _ = engine.run("record batches").await;
+    let _ = engine.run(&mut history, "record batches").await;
 
     let contexts = contexts.lock().unwrap().clone();
     assert_eq!(contexts.len(), 2);
@@ -298,6 +884,7 @@ async fn test_tool_execution_context_for_skipped_and_synthetic_paths() {
         ],
     ]);
     let mut engine = Engine::new(client);
+    let mut history: History = History::new();
     let executed_contexts = Arc::new(Mutex::new(Vec::new()));
     let pre_contexts = Arc::new(Mutex::new(Vec::new()));
     let post_contexts = Arc::new(Mutex::new(Vec::new()));
@@ -344,7 +931,9 @@ async fn test_tool_execution_context_for_skipped_and_synthetic_paths() {
         post_contexts: post_contexts.clone(),
     });
 
-    let _ = engine.run("record skipped and synthetic contexts").await;
+    let _ = engine
+        .run(&mut history, "record skipped and synthetic contexts")
+        .await;
 
     let mut pre_contexts = pre_contexts.lock().unwrap().clone();
     pre_contexts.sort_by_key(|ctx| ctx.call_index);
@@ -389,6 +978,7 @@ async fn test_before_tool_call_skip() {
 
     let client = MockLlmClient::new(events);
     let mut engine = Engine::new(client);
+    let mut history: History = History::new();
 
     let allowed_tool = SlowTool::new("allowed_tool", 10);
     let blocked_tool = SlowTool::new("blocked_tool", 10);
@@ -415,8 +1005,8 @@ async fn test_before_tool_call_skip() {
 
     engine.set_interceptor(BlockingPolicy);
 
-    // Mutable::run consumes self, returns (Locked, EngineRunExit)
-    let _result = engine.run("Test hook").await;
+    // Mutable::run consumes self, returns (Locked, EngineResult)
+    let _result = engine.run(&mut history, "Test hook").await;
 
     // allowed_tool is called, but blocked_tool is not
     assert_eq!(
@@ -457,6 +1047,7 @@ async fn test_post_tool_call_modification() {
     ]);
 
     let mut engine = Engine::new(client);
+    let mut history: History = History::new();
 
     #[derive(Clone)]
     struct SimpleTool;
@@ -502,8 +1093,8 @@ async fn test_post_tool_call_modification() {
         modified_content: modified_content.clone(),
     });
 
-    // Mutable::run consumes self, returns (Locked, EngineRunExit)
-    let result = engine.run("Test modification").await;
+    // Mutable::run consumes self, returns (Locked, EngineResult)
+    let result = engine.run(&mut history, "Test modification").await;
 
     assert!(
         matches!(result.result, agen::EngineRunExit::Finished),
@@ -543,6 +1134,7 @@ async fn test_before_tool_call_synthetic_result_committed() {
         ],
     ]);
     let mut engine = Engine::new(client);
+    let mut history: History = History::new();
     let blocked_tool = SlowTool::new("blocked_tool", 10);
     let blocked_clone = blocked_tool.clone();
     engine.register_tool(blocked_tool.definition());
@@ -561,10 +1153,10 @@ async fn test_before_tool_call_synthetic_result_committed() {
 
     engine.set_interceptor(SyntheticPolicy);
 
-    let result = engine.run("Test synthetic result").await;
+    let _result = engine.run(&mut history, "Test synthetic result").await;
 
     assert_eq!(blocked_clone.call_count(), 0, "Blocked tool should not run");
-    assert!(result.engine.history().iter().any(|item| matches!(
+    assert!(history.items().any(|item| matches!(
         item,
         agen::Item::ToolResult {
             call_id,
@@ -574,3 +1166,76 @@ async fn test_before_tool_call_synthetic_result_committed() {
         } if call_id == "call_1" && summary == "permission denied"
     )));
 }
+
+#[tokio::test]
+async fn post_tool_abort_commits_confirmed_result_before_stopping_run() {
+    let client = MockLlmClient::new(vec![
+        Event::tool_use_start(0, "call_confirmed", "confirmed"),
+        Event::tool_input_delta(0, r#"{}"#),
+        Event::tool_use_stop(0),
+        Event::Status(StatusEvent {
+            status: ResponseStatus::Completed,
+        }),
+    ]);
+    let mut engine = Engine::new(client);
+    let tool = SlowTool::new("confirmed", 1);
+    engine.register_tool(tool.definition());
+
+    struct AbortAfterResult;
+    #[async_trait]
+    impl Interceptor for AbortAfterResult {
+        async fn post_tool_call(&self, _info: &mut ToolResultInfo) -> PostToolAction {
+            PostToolAction::Abort("policy stopped the run".to_string())
+        }
+    }
+    engine.set_interceptor(AbortAfterResult);
+
+    let observed = Arc::new(Mutex::new(Vec::<&'static str>::new()));
+    let published = observed.clone();
+    engine.on_tool_result(move |_| published.lock().unwrap().push("published"));
+    let committed = observed.clone();
+    let mut annotate = move |item: &Item| {
+        if matches!(item, Item::ToolResult { .. }) {
+            committed.lock().unwrap().push("committed");
+        }
+        Ok(())
+    };
+
+    let mut history = History::new();
+    let output = engine
+        .run_with_annotation(&mut history, "run confirmed tool", &mut annotate)
+        .await;
+    observed.lock().unwrap().push("run-returned");
+
+    assert_eq!(tool.call_count(), 1);
+    assert_eq!(
+        observed.lock().unwrap().as_slice(),
+        ["committed", "published", "run-returned"]
+    );
+    assert!(matches!(
+        output.result,
+        agen::EngineRunExit::Interrupted(agen::StopReason::Unexpected(
+            agen::EngineError::Aborted(ref reason)
+        )) if reason == "policy stopped the run"
+    ));
+    let terminal: Vec<_> = history
+        .iter()
+        .filter_map(|entry| match &entry.item {
+            Item::ToolResult {
+                call_id,
+                disposition,
+                ..
+            } if call_id == "call_confirmed" => Some(*disposition),
+            _ => None,
+        })
+        .collect();
+    assert_eq!(terminal, [ToolResultDisposition::Success]);
+    assert!(!history.iter().any(|entry| matches!(
+        &entry.item,
+        Item::ToolResult {
+            call_id,
+            disposition: ToolResultDisposition::OutcomeUnknown,
+            ..
+        } if call_id == "call_confirmed"
+    )));
+}

crates/agen/tests/reasoning_round_trip_test.rs

@@ -13,12 +13,12 @@
 
 mod common;
 
-use agen::Engine;
 use agen::Item;
 use agen::llm_client::event::{
     BlockMetadata, BlockStart, BlockStop, BlockType, Event, ReasoningBlockData, ResponseStatus,
     StatusEvent,
 };
+use agen::{Engine, History};
 use common::MockLlmClient;
 
 fn reasoning_block(text: impl Into<String>, data: ReasoningBlockData) -> Vec<Event> {
@@ -65,15 +65,15 @@ async fn anthropic_thinking_round_trips_signature_into_history() {
     ]);
     let client = MockLlmClient::new(events);
     let engine = Engine::new(client);
-    let out = engine.run("question?").await;
-    let engine = out.engine;
+    let mut history: History = History::new();
+    let _out = engine.run(&mut history, "question?").await;
 
-    let history = engine.history();
+    let entries = history.entries();
     // user / reasoning / assistant_message
     assert_eq!(history.len(), 3, "history: {history:?}");
 
-    assert!(matches!(history[0], Item::Message { .. }));
-    match &history[1] {
+    assert!(matches!(entries[0].item, Item::Message { .. }));
+    match &entries[1].item {
         Item::Reasoning {
             text, signature, ..
         } => {
@@ -82,7 +82,7 @@ async fn anthropic_thinking_round_trips_signature_into_history() {
         }
         other => panic!("expected Reasoning, got {other:?}"),
     }
-    assert_eq!(history[2].as_text(), Some("Here's the answer"));
+    assert_eq!(entries[2].item.as_text(), Some("Here's the answer"));
 }
 
 /// OpenAI Responses 風: encrypted_content + summary を持った reasoning が
@@ -109,11 +109,11 @@ async fn openai_reasoning_round_trips_encrypted_and_summary() {
     ]);
     let client = MockLlmClient::new(events);
     let engine = Engine::new(client);
-    let out = engine.run("q").await;
-    let engine = out.engine;
+    let mut history: History = History::new();
+    let _out = engine.run(&mut history, "q").await;
 
-    let history = engine.history();
-    match &history[1] {
+    let entries = history.entries();
+    match &entries[1].item {
         Item::Reasoning {
             text,
             summary,
@@ -155,13 +155,13 @@ async fn reasoning_precedes_text_in_assistant_burst() {
     }));
     let client = MockLlmClient::new(events);
     let engine = Engine::new(client);
-    let out = engine.run("q").await;
-    let engine = out.engine;
+    let mut history: History = History::new();
+    let _out = engine.run(&mut history, "q").await;
 
-    let history = engine.history();
+    let entries = history.entries();
     // user / reasoning(先頭) / assistant_message
-    assert!(matches!(history[1], Item::Reasoning { .. }));
-    assert_eq!(history[2].as_text(), Some("intermediate"));
+    assert!(matches!(entries[1].item, Item::Reasoning { .. }));
+    assert_eq!(entries[2].item.as_text(), Some("intermediate"));
 }
 
 /// resume シナリオ: history.json 由来の Item::Reasoning(signature) を Engine に
@@ -207,14 +207,18 @@ async fn injected_reasoning_survives_into_outgoing_request() {
     };
 
     let mut engine = Engine::new(client);
+    let mut history: History = History::new();
     // resume: 既存 history を流し込む
-    engine.set_history(vec![
+    engine.set_history(
+        &mut history,
+        vec![
             Item::user_message("prior question"),
             Item::reasoning("prior thinking").with_signature("SIG-PRIOR"),
             Item::assistant_message("prior answer"),
-    ]);
+        ],
+    );
 
-    let _ = engine.run("follow up").await;
+    let _ = engine.run(&mut history, "follow up").await;
 
     let req = captured
         .lock()

crates/agen/tests/ui/locked_register_tool.rs

@@ -1,4 +1,4 @@
-use agen::Engine;
+use agen::{Engine, History};
 use agen::llm_client::capability::{
     CacheStrategy, ModelCapability, StructuredOutput, ToolCallingSupport,
 };
@@ -22,7 +22,8 @@ fn main() {
         cap,
     );
     let engine = Engine::new(client);
-    let mut locked = engine.lock();
+    let history = History::new();
+    let mut locked = engine.lock(&history);
     let def: agen::tool::ToolDefinition = Arc::new(|| panic!("unused"));
     let _ = locked.register_tool(def);
 }

crates/agen/tests/ui/locked_register_tool.stderr

@@ -1,8 +1,8 @@
 error[E0599]: no method named `register_tool` found for struct `Engine<HttpTransport<AnthropicScheme>, Locked>` in the current scope
-  --> tests/ui/locked_register_tool.rs:27:20
+  --> tests/ui/locked_register_tool.rs:28:20
    |
-27 |     let _ = locked.register_tool(def);
+28 |     let _ = locked.register_tool(def);
    |                    ^^^^^^^^^^^^^ method not found in `Engine<HttpTransport<AnthropicScheme>, Locked>`
    |
    = note: the method was found for
-           - `Engine<C>`
+           - `Engine<C, Mutable, A>`

crates/fs-operation/src/lib.rs

@@ -279,4 +279,58 @@ mod tests {
         assert_eq!(grep.matched_files, 2);
         assert!(!grep.output.contains("c.txt"));
     }
+
+    #[test]
+    fn grep_content_groups_lines_by_file_and_marks_matches() {
+        let temp = tempfile::tempdir().unwrap();
+        std::fs::write(
+            temp.path().join("first.txt"),
+            "before\nneedle one\nafter\nomitted one\nomitted two\nbefore distant\nneedle distant\nafter distant\n",
+        )
+        .unwrap();
+        std::fs::write(temp.path().join("second.txt"), "needle two\n").unwrap();
+        let root = temp.path().canonicalize().unwrap();
+        let readable = RootAccess(root.clone());
+
+        let grep = run_grep(
+            &root,
+            root.clone(),
+            GrepRequest {
+                pattern: "needle".to_string(),
+                path: FsPath::root(),
+                glob: Some("*.txt".to_string()),
+                output_mode: GrepOutputMode::Content,
+                case_insensitive: false,
+                before_context: 1,
+                after_context: 1,
+                multiline: false,
+                file_type: None,
+                limit: 20,
+                offset: 0,
+            },
+            &readable,
+        )
+        .unwrap();
+
+        assert_eq!(grep.match_count, 3);
+        assert_eq!(grep.matched_files, 2);
+        assert_eq!(
+            grep.output,
+            concat!(
+                "first.txt\n",
+                "   1 │ before\n",
+                " > 2 │ needle one\n",
+                "   3 │ after\n",
+                "   …\n",
+                "   6 │ before distant\n",
+                " > 7 │ needle distant\n",
+                "   8 │ after distant\n",
+                "\n",
+                "second.txt\n",
+                " > 1 │ needle two\n",
+            )
+        );
+        assert_eq!(grep.output.matches("first.txt").count(), 1);
+        assert_eq!(grep.output.matches("second.txt").count(), 1);
+    }
 }

crates/fs-operation/src/search.rs

@@ -1,3 +1,5 @@
+use std::collections::BTreeMap;
+use std::fmt::Write as _;
 use std::path::{Path, PathBuf};
 
 use crate::FsAccessPolicy;
@@ -57,20 +59,11 @@ impl GrepReport {
                 }
             }
             GrepOutputMode::Content => {
-                for line in &self.lines {
-                    let separator = if line.is_match { ':' } else { '-' };
-                    let path = logical_display(root, &line.path);
-                    if self.show_line_numbers
-                        && let Some(number) = line.line_number
-                    {
-                        output.push_str(&format!(
-                            "{path}{separator}{number}{separator}{}\n",
-                            line.text
+                output.push_str(&render_content_lines(
+                    root,
+                    &self.lines,
+                    self.show_line_numbers,
                 ));
-                    } else {
-                        output.push_str(&format!("{path}{separator}{}\n", line.text));
-                    }
-                }
             }
         }
         GrepResult {
@@ -82,6 +75,48 @@ impl GrepReport {
     }
 }
 
+fn render_content_lines(root: &Path, lines: &[ContentLine], show_line_numbers: bool) -> String {
+    let mut grouped = BTreeMap::<&Path, Vec<&ContentLine>>::new();
+    for line in lines {
+        grouped.entry(&line.path).or_default().push(line);
+    }
+
+    let mut output = String::new();
+    for (file_index, (path, file_lines)) in grouped.into_iter().enumerate() {
+        if file_index > 0 {
+            output.push('\n');
+        }
+        let _ = writeln!(output, "{}", logical_display(root, path));
+
+        let number_width = file_lines
+            .iter()
+            .filter_map(|line| line.line_number)
+            .map(|number| number.to_string().len())
+            .max()
+            .unwrap_or(1);
+        let mut previous_line_end = None;
+        for line in file_lines {
+            if let (Some(previous_end), Some(number)) = (previous_line_end, line.line_number)
+                && number > previous_end
+            {
+                let _ = writeln!(output, "   …");
+            }
+
+            let marker = if line.is_match { '>' } else { ' ' };
+            if show_line_numbers && let Some(number) = line.line_number {
+                let _ = writeln!(output, " {marker} {number:>number_width$} │ {}", line.text);
+            } else {
+                let _ = writeln!(output, " {marker} │ {}", line.text);
+            }
+            previous_line_end = line
+                .line_number
+                .map(|number| number + line.text.split('\n').count() as u64);
+        }
+    }
+
+    output
+}
+
 fn logical_display(root: &Path, path: &Path) -> String {
     path.strip_prefix(root)
         .unwrap_or(path)

crates/manifest/src/profile.rs

@@ -946,7 +946,7 @@ fn apply_role_profile(
     value["feature"]["sub_worker"] = serde_json::json!({ "enabled": sub_worker });
     value["feature"]["flow"] = serde_json::json!({ "enabled": slug == "coder" });
     value["feature"]["worker"] = serde_json::json!({
-        "enabled": matches!(slug, "companion" | "orchestrator"),
+        "enabled": slug == "orchestrator",
         "direct_spawn": slug != "orchestrator"
     });
     value["feature"]["manage_workdir"] = serde_json::json!({
@@ -1408,7 +1408,22 @@ mod tests {
     }
 
     #[test]
-    fn builtin_companion_can_manage_workdirs() {
+    fn builtin_coder_uses_sub_worker_control_without_worker_control() {
+        let tmp = TempDir::new().unwrap();
+        let resolved = ProfileResolver::new()
+            .with_workspace_base(tmp.path())
+            .resolve(
+                &ProfileSelector::source_named(ProfileRegistrySource::Builtin, "coder"),
+                ProfileResolveOptions::with_worker_name("coder-worker"),
+            )
+            .unwrap();
+
+        assert!(resolved.manifest.feature.sub_worker.enabled);
+        assert!(!resolved.manifest.feature.worker.enabled);
+    }
+
+    #[test]
+    fn builtin_companion_uses_sub_worker_control_without_worker_control() {
         let tmp = TempDir::new().unwrap();
         let resolved = ProfileResolver::new()
             .with_workspace_base(tmp.path())
@@ -1419,6 +1434,8 @@ mod tests {
             .unwrap();
 
         assert!(resolved.manifest.feature.manage_workdir.enabled);
+        assert!(resolved.manifest.feature.sub_worker.enabled);
+        assert!(!resolved.manifest.feature.worker.enabled);
     }
 
     #[test]

crates/memory/src/extract/payload.rs

@@ -9,7 +9,7 @@
 use schemars::JsonSchema;
 use serde::{Deserialize, Serialize};
 
-use crate::schema::{EvidenceKind, SourceEvidenceRef, SourceRef};
+use crate::schema::{EvidenceKind, EvidenceOrigin, SourceEvidenceRef, SourceRef};
 
 /// Current flat staging schema version.
 pub const STAGING_SCHEMA_VERSION: u32 = 2;
@@ -80,6 +80,8 @@ pub struct StagingEvidence {
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub entry_range: Option<[u64; 2]>,
     #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub origin: Option<EvidenceOrigin>,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub excerpt: Option<String>,
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub summary: Option<String>,
@@ -159,6 +161,7 @@ mod tests {
             id: "E001".into(),
             kind: EvidenceKind::new(EvidenceKind::MESSAGE),
             entry_range: Some([10, 12]),
+            origin: None,
             excerpt: Some("extract candidate taxonomy".into()),
             summary: Some("User and assistant discussed staging kinds".into()),
         };

crates/memory/src/schema/common.rs

@@ -67,6 +67,40 @@ impl EvidenceKind {
     }
 }
 
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize, JsonSchema)]
+#[serde(rename_all = "snake_case")]
+pub enum EvidenceOriginKind {
+    HumanInput,
+    WorkerInput,
+    FlowInstruction,
+    BackendInstruction,
+    ModelOutput,
+    ToolOutput,
+    DerivedSummary,
+    LegacyUnknown,
+}
+
+/// Bounded origin snapshot attached to extraction evidence. This is audit
+/// metadata only and cannot authorize Workspace operations.
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize, JsonSchema)]
+pub struct EvidenceOrigin {
+    pub kind: EvidenceOriginKind,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub account_id: Option<String>,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub workspace_id: Option<String>,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub runtime_id: Option<String>,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub worker_id: Option<String>,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub flow_selector: Option<String>,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub flow_definition_id: Option<String>,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub flow_definition_revision: Option<u64>,
+}
+
 /// Host-resolved source/evidence metadata for an individual staging claim.
 ///
 /// This deliberately stores only bounded anchor metadata: stable ids, entry
@@ -86,6 +120,9 @@ pub struct SourceEvidenceRef {
     /// Host-assigned evidence id within the referenced evidence set.
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub evidence_id: Option<String>,
+    /// Trusted typed origin snapshot for this logical evidence entry.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub origin: Option<EvidenceOrigin>,
     /// Extensible evidence kind tag.
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub evidence_kind: Option<EvidenceKind>,

crates/memory/src/schema/mod.rs

@@ -10,7 +10,10 @@ mod decision;
 mod request;
 mod summary;
 
-pub use common::{EvidenceKind, Frontmatter, SourceEvidenceRef, SourceRef, split_frontmatter};
+pub use common::{
+    EvidenceKind, EvidenceOrigin, EvidenceOriginKind, Frontmatter, SourceEvidenceRef, SourceRef,
+    split_frontmatter,
+};
 pub use decision::{DecisionFrontmatter, DecisionStatus};
 pub use request::RequestFrontmatter;
 pub use summary::SummaryFrontmatter;

crates/protocol/src/lib.rs

@@ -352,6 +352,18 @@ pub struct InternalWorkerSnapshot {
     pub internal_workers: Vec<InternalWorkerSnapshot>,
 }
 
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize, Default)]
+#[cfg_attr(feature = "typescript", derive(ts_rs::TS))]
+#[serde(rename_all = "snake_case")]
+pub enum ToolResultDisposition {
+    #[default]
+    Success,
+    Error,
+    Interrupted,
+    Cancelled,
+    OutcomeUnknown,
+}
+
 #[derive(Debug, Clone, Serialize, Deserialize)]
 #[cfg_attr(feature = "typescript", derive(ts_rs::TS))]
 #[serde(tag = "event", content = "data", rename_all = "snake_case")]
@@ -501,6 +513,8 @@ pub enum Event {
         /// summary-only, or when the result was pruned.
         #[serde(default, skip_serializing_if = "Option::is_none")]
         output: Option<String>,
+        #[serde(default, skip_serializing_if = "Option::is_none")]
+        disposition: Option<ToolResultDisposition>,
         #[serde(default)]
         is_error: bool,
     },
@@ -1839,6 +1853,7 @@ mod tests {
             id: "call_1".into(),
             summary: "Read 128 bytes".into(),
             output: Some("hello world".into()),
+            disposition: Some(ToolResultDisposition::Success),
             is_error: false,
         };
         let json = serde_json::to_string(&event).unwrap();
@@ -1855,11 +1870,13 @@ mod tests {
                 id,
                 summary,
                 output,
+                disposition,
                 is_error,
             } => {
                 assert_eq!(id, "call_1");
                 assert_eq!(summary, "Read 128 bytes");
                 assert_eq!(output.as_deref(), Some("hello world"));
+                assert_eq!(disposition, Some(ToolResultDisposition::Success));
                 assert!(!is_error);
             }
             other => panic!("expected ToolResult, got {other:?}"),
@@ -1872,6 +1889,7 @@ mod tests {
             id: "call_2".into(),
             summary: "ok".into(),
             output: None,
+            disposition: Some(ToolResultDisposition::Success),
             is_error: false,
         };
         let json = serde_json::to_string(&event).unwrap();
@@ -1887,6 +1905,7 @@ mod tests {
             id: "call_3".into(),
             summary: "invalid argument".into(),
             output: None,
+            disposition: Some(ToolResultDisposition::Error),
             is_error: true,
         };
         let json = serde_json::to_string(&event).unwrap();

crates/protocol/src/typescript.rs

@@ -8,7 +8,7 @@ use crate::{
     CompletionKind, ErrorCode, Event, Greeting, InFlightBlock, InFlightSnapshot,
     InFlightToolCallState, InternalWorkerKind, InternalWorkerRef, InternalWorkerSnapshot,
     InvokeKind, MemoryWorkerEvent, Method, Permission, RewindSummary, RewindTarget, RewindTargetId,
-    RunResult, ScopeRule, Segment, TurnResult, WorkerEvent, WorkerStatus,
+    RunResult, ScopeRule, Segment, ToolResultDisposition, TurnResult, WorkerEvent, WorkerStatus,
     subscription::{
         EventSubscriptionSelector, SubscriptionEvent, SubscriptionEventPayload, SubscriptionFrame,
         SubscriptionFramePayload, SubscriptionId, SubscriptionRejectionCode, SubscriptionRequest,
@@ -45,6 +45,7 @@ pub fn generated_protocol_types() -> String {
     push_decl::<TurnResult>(&cfg, &mut output);
     push_decl::<InvokeKind>(&cfg, &mut output);
     push_decl::<RunResult>(&cfg, &mut output);
+    push_decl::<ToolResultDisposition>(&cfg, &mut output);
     push_decl::<ErrorCode>(&cfg, &mut output);
     push_decl::<Permission>(&cfg, &mut output);
     push_decl::<InFlightToolCallState>(&cfg, &mut output);

crates/session-store/src/history.rs

@@ -0,0 +1,180 @@
+//! Serializable history entries with restore-authoritative logical identity and origin.
+
+use serde::{Deserialize, Serialize};
+
+use crate::{LoggedItem, SessionId};
+
+/// Stable logical identity of one model-visible history entry.
+///
+/// This value is generated at the trusted Worker session boundary and copied
+/// unchanged across fork, rewind, compaction retention, restore, and reboot.
+#[derive(Clone, Debug, PartialEq, Eq, Hash, Serialize, Deserialize)]
+#[serde(transparent)]
+pub struct LoggedSessionHistoryEntryId(pub String);
+
+impl LoggedSessionHistoryEntryId {
+    pub fn new() -> Self {
+        Self(uuid::Uuid::now_v7().to_string())
+    }
+}
+
+impl Default for LoggedSessionHistoryEntryId {
+    fn default() -> Self {
+        Self::new()
+    }
+}
+
+/// Bounded subject snapshot. It is evidence, not a live authorization handle.
+#[derive(Clone, Debug, PartialEq, Eq, Serialize, Deserialize)]
+pub struct LoggedWorkerSubject {
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub workspace_id: Option<String>,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub runtime_id: Option<String>,
+    pub worker_id: String,
+}
+
+#[derive(Clone, Debug, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(tag = "kind", rename_all = "snake_case")]
+pub enum LoggedSessionHistoryOrigin {
+    HumanInput {
+        account_id: String,
+    },
+    WorkerInput {
+        actor: LoggedWorkerSubject,
+    },
+    FlowInstruction {
+        selector: String,
+        definition_id: String,
+        definition_revision: u64,
+        instance_id: String,
+        state_id: String,
+    },
+    BackendInstruction {
+        #[serde(default, skip_serializing_if = "Option::is_none")]
+        operation_id: Option<String>,
+    },
+    ModelOutput {
+        worker: LoggedWorkerSubject,
+    },
+    ToolOutput {
+        worker: LoggedWorkerSubject,
+    },
+    DerivedSummary,
+    LegacyUnknown,
+}
+
+#[derive(Clone, Debug, PartialEq, Eq, Serialize, Deserialize)]
+pub struct LoggedHistoryDerivation {
+    pub sources: Vec<LoggedSessionHistoryEntryId>,
+}
+
+#[derive(Clone, Debug, PartialEq, Eq, Serialize, Deserialize)]
+pub struct LoggedSessionHistoryMetadata {
+    pub entry_id: LoggedSessionHistoryEntryId,
+    pub origin: LoggedSessionHistoryOrigin,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub derivation: Option<LoggedHistoryDerivation>,
+}
+
+impl LoggedSessionHistoryMetadata {
+    pub fn legacy_unknown() -> Self {
+        Self {
+            entry_id: LoggedSessionHistoryEntryId::new(),
+            origin: LoggedSessionHistoryOrigin::LegacyUnknown,
+            derivation: None,
+        }
+    }
+}
+
+/// Persisted item and metadata are one value so transforms cannot reorder or
+/// truncate one without the other.
+#[derive(Clone, Debug, PartialEq, Serialize, Deserialize)]
+pub struct LoggedHistoryEntry {
+    pub item: LoggedItem,
+    pub metadata: LoggedSessionHistoryMetadata,
+}
+
+/// Typed system-item history record. The typed system event remains available
+/// to client replay while its model-visible projection carries the same stable
+/// metadata used by live history.
+#[derive(Clone, Debug, Serialize, Deserialize)]
+pub struct LoggedSystemHistoryEntry {
+    pub item: crate::SystemItem,
+    pub metadata: LoggedSessionHistoryMetadata,
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::LoggedRole;
+    use agen::llm_client::RequestConfig;
+
+    #[test]
+    fn logged_history_entry_round_trip_preserves_id_origin_and_derivation() {
+        let source_id = LoggedSessionHistoryEntryId::new();
+        let entry = LoggedHistoryEntry {
+            item: LoggedItem::Message {
+                role: LoggedRole::User,
+                content: vec![crate::LoggedContentPart::Text {
+                    text: "preference".into(),
+                }],
+            },
+            metadata: LoggedSessionHistoryMetadata {
+                entry_id: LoggedSessionHistoryEntryId::new(),
+                origin: LoggedSessionHistoryOrigin::HumanInput {
+                    account_id: "account-1".into(),
+                },
+                derivation: Some(LoggedHistoryDerivation {
+                    sources: vec![source_id.clone()],
+                }),
+            },
+        };
+        let encoded = serde_json::to_vec(&entry).unwrap();
+        let decoded: LoggedHistoryEntry = serde_json::from_slice(&encoded).unwrap();
+        assert_eq!(decoded, entry);
+        assert_eq!(
+            decoded.metadata.derivation.unwrap().sources,
+            vec![source_id]
+        );
+    }
+
+    #[test]
+    fn annotated_segment_start_is_restore_visible_without_projecting_metadata() {
+        let session_id = uuid::Uuid::now_v7();
+        let history_entry = legacy_logged_history(LoggedItem::Message {
+            role: LoggedRole::Assistant,
+            content: vec![crate::LoggedContentPart::Text {
+                text: "answer".into(),
+            }],
+        });
+        let state = crate::collect_state(&[crate::LogEntry::AnnotatedSegmentStart {
+            ts: 1,
+            session_id,
+            system_prompt: None,
+            config: RequestConfig::default(),
+            history: vec![history_entry],
+            forked_from: None,
+            compacted_from: None,
+        }]);
+        assert_eq!(state.history[0].as_text(), Some("answer"));
+    }
+}
+
+/// Legacy Session Logs did not persist annotations. Decode helpers explicitly
+/// create `LegacyUnknown`; they never infer Human/System authority from role or
+/// plaintext.
+pub fn legacy_logged_history(item: LoggedItem) -> LoggedHistoryEntry {
+    LoggedHistoryEntry {
+        item,
+        metadata: LoggedSessionHistoryMetadata::legacy_unknown(),
+    }
+}
+
+pub fn legacy_segment_history(
+    session_id: SessionId,
+    items: impl IntoIterator<Item = LoggedItem>,
+) -> Vec<LoggedHistoryEntry> {
+    let _ = session_id;
+    items.into_iter().map(legacy_logged_history).collect()
+}

crates/session-store/src/lib.rs

@@ -32,6 +32,7 @@
 
 pub mod event_trace;
 pub mod fs_store;
+pub mod history;
 pub mod logged_item;
 pub mod segment;
 pub mod segment_log;
@@ -44,6 +45,11 @@ pub use agen::UsageRecord;
 pub use agen::llm_client::types::{ContentPart, Item, Role};
 pub use event_trace::{TraceEntry, TracePayload};
 pub use fs_store::FsStore;
+pub use history::{
+    LoggedHistoryDerivation, LoggedHistoryEntry, LoggedSessionHistoryEntryId,
+    LoggedSessionHistoryMetadata, LoggedSessionHistoryOrigin, LoggedSystemHistoryEntry,
+    LoggedWorkerSubject, legacy_logged_history, legacy_segment_history,
+};
 pub use logged_item::{LoggedContentPart, LoggedItem, LoggedRole, from_logged, to_logged};
 pub use segment::{
     SegmentStartState, append_entry, append_system_item, classify_history_item,

crates/session-store/src/logged_item.rs

@@ -14,7 +14,7 @@
 
 use agen::{
     llm_client::types::{ContentPart, Item, Role},
-    tool::{Attachment, ImageAttachment},
+    tool::{Attachment, ImageAttachment, ToolResultDisposition},
 };
 use base64::{Engine as _, engine::general_purpose::STANDARD};
 use serde::{Deserialize, Deserializer, Serialize, Serializer, de::Error as _};
@@ -61,6 +61,8 @@ pub enum LoggedItem {
         content: Option<String>,
         #[serde(default, skip_serializing_if = "Vec::is_empty")]
         attachments: Vec<LoggedAttachment>,
+        #[serde(default, skip_serializing_if = "ToolResultDisposition::is_success")]
+        disposition: ToolResultDisposition,
         #[serde(default, skip_serializing_if = "is_false")]
         is_error: bool,
     },
@@ -128,6 +130,7 @@ impl From<&Item> for LoggedItem {
                 summary,
                 content,
                 attachments,
+                disposition,
                 is_error,
                 ..
             } => Self::ToolResult {
@@ -135,6 +138,7 @@ impl From<&Item> for LoggedItem {
                 summary: summary.clone(),
                 content: content.clone(),
                 attachments: attachments.iter().map(LoggedAttachment::from).collect(),
+                disposition: *disposition,
                 is_error: *is_error,
             },
             Item::Reasoning {
@@ -184,15 +188,24 @@ impl From<LoggedItem> for Item {
                 summary,
                 content,
                 attachments,
+                disposition,
                 is_error,
-            } => Item::ToolResult {
+            } => {
+                let disposition = if is_error && disposition.is_success() {
+                    ToolResultDisposition::Error
+                } else {
+                    disposition
+                };
+                Item::ToolResult {
                     id: None,
                     call_id,
                     summary,
                     content,
+                    disposition,
                     is_error,
                     attachments: attachments.into_iter().map(Attachment::from).collect(),
-            },
+                }
+            }
             LoggedItem::Reasoning {
                 text,
                 summary,
@@ -430,6 +443,42 @@ mod tests {
         }
     }
 
+    #[test]
+    fn outcome_unknown_tool_result_round_trips_as_terminal() {
+        let original = Item::tool_result_item_with_disposition_and_attachments(
+            "call_unknown",
+            "outcome unknown",
+            Some("bounded progress".to_string()),
+            ToolResultDisposition::OutcomeUnknown,
+            Vec::new(),
+        );
+        let logged: LoggedItem = (&original).into();
+        let json = serde_json::to_string(&logged).unwrap();
+        assert!(json.contains(r#""disposition":"outcome_unknown""#));
+        match Item::from(serde_json::from_str::<LoggedItem>(&json).unwrap()) {
+            Item::ToolResult {
+                disposition,
+                is_error,
+                ..
+            } => {
+                assert_eq!(disposition, ToolResultDisposition::OutcomeUnknown);
+                assert!(is_error);
+            }
+            other => panic!("unexpected variant: {other:?}"),
+        }
+    }
+
+    #[test]
+    fn legacy_error_tool_result_infers_error_disposition() {
+        let legacy = r#"{"kind":"tool_result","call_id":"call_old","summary":"failed","content":null,"is_error":true}"#;
+        match Item::from(serde_json::from_str::<LoggedItem>(legacy).unwrap()) {
+            Item::ToolResult { disposition, .. } => {
+                assert_eq!(disposition, ToolResultDisposition::Error)
+            }
+            other => panic!("unexpected variant: {other:?}"),
+        }
+    }
+
     #[test]
     fn tool_result_persistence_round_trips_binary_attachments() {
         let original = Item::tool_result_item_with_attachments(

crates/session-store/src/segment_log.rs

@@ -14,6 +14,7 @@ use agen::{EngineResult, UsageRecord};
 use protocol::{InvokeKind, Segment};
 use serde::{Deserialize, Serialize};
 
+use crate::history::{LoggedHistoryEntry, LoggedSystemHistoryEntry};
 use crate::logged_item::LoggedItem;
 use crate::system_item::SystemItem;
 
@@ -70,6 +71,20 @@ pub enum LogEntry {
         compacted_from: Option<SegmentOrigin>,
     },
 
+    /// Schema-v2 segment seed. Retained entries keep their stable logical
+    /// identity and origin across fork/compaction/restore.
+    AnnotatedSegmentStart {
+        ts: u64,
+        session_id: crate::SessionId,
+        system_prompt: Option<String>,
+        config: RequestConfig,
+        history: Vec<LoggedHistoryEntry>,
+        #[serde(default, skip_serializing_if = "Option::is_none")]
+        forked_from: Option<SegmentOrigin>,
+        #[serde(default, skip_serializing_if = "Option::is_none")]
+        compacted_from: Option<SegmentOrigin>,
+    },
+
     /// IDLE → active marker. Records the start of a new self-driving
     /// cycle (Invoke range). The range extends implicitly until the
     /// next `Invoke` entry; this entry carries the trigger only — the
@@ -105,14 +120,37 @@ pub enum LogEntry {
         extensions: Vec<SessionExtension>,
     },
 
+    /// Schema-v2 user submission with its exact model-visible entries. Typed
+    /// Flow instructions and caller-attributed input remain separate entries.
+    AnnotatedUserInput {
+        ts: u64,
+        segments: Vec<Segment>,
+        #[serde(default, skip_serializing_if = "Vec::is_empty")]
+        extensions: Vec<SessionExtension>,
+        history: Vec<LoggedHistoryEntry>,
+    },
+
+    /// Schema-v2 model output and metadata committed as one journal record.
+    AnnotatedAssistantItem { ts: u64, entry: LoggedHistoryEntry },
+
     /// One assistant-side item appended to history — assistant message,
     /// reasoning, or tool call. Singular: one entry per history item so
     /// the wire-side `Event::*` lane and on-disk LogEntry stay 1:1.
     AssistantItem { ts: u64, item: LoggedItem },
 
+    /// Schema-v2 tool output and metadata committed as one journal record.
+    AnnotatedToolResult { ts: u64, entry: LoggedHistoryEntry },
+
     /// One tool-execution result appended to history.
     ToolResult { ts: u64, item: LoggedItem },
 
+    /// Schema-v2 typed system event and model-visible metadata committed
+    /// together.
+    AnnotatedSystemItem {
+        ts: u64,
+        entry: LoggedSystemHistoryEntry,
+    },
+
     /// One typed agent-injected system item: notification, child-Worker
     /// lifecycle event, `@<path>` / `/<slug>` resolution payload. Each
     /// `SystemItem` carries kind metadata that the LLM
@@ -278,6 +316,22 @@ pub fn collect_state(entries: &[LogEntry]) -> RestoredState {
                 state.config = config.clone();
                 state.history = history.iter().cloned().map(Item::from).collect();
             }
+            LogEntry::AnnotatedSegmentStart {
+                session_id,
+                system_prompt,
+                config,
+                history,
+                ..
+            } => {
+                state.session_id = Some(*session_id);
+                state.system_prompt = system_prompt.clone();
+                state.config = config.clone();
+                state.history = history
+                    .iter()
+                    .cloned()
+                    .map(|entry| Item::from(entry.item))
+                    .collect();
+            }
             LogEntry::Invoke { .. } => {
                 // A terminal run record below clears or refines this. If the
                 // log ends first, restore must treat the turn as interrupted.
@@ -298,6 +352,29 @@ pub fn collect_state(entries: &[LogEntry]) -> RestoredState {
                         .map(|extension| (extension.domain.clone(), extension.payload.clone())),
                 );
             }
+            LogEntry::AnnotatedUserInput {
+                segments,
+                extensions,
+                history,
+                ..
+            } => {
+                state
+                    .history
+                    .extend(history.iter().cloned().map(|entry| Item::from(entry.item)));
+                state.user_segments.push(segments.clone());
+                state.extensions.extend(
+                    extensions
+                        .iter()
+                        .map(|extension| (extension.domain.clone(), extension.payload.clone())),
+                );
+            }
+            LogEntry::AnnotatedAssistantItem { entry, .. }
+            | LogEntry::AnnotatedToolResult { entry, .. } => {
+                state.history.push(Item::from(entry.item.clone()));
+            }
+            LogEntry::AnnotatedSystemItem { entry, .. } => {
+                state.history.push(entry.item.to_history_item());
+            }
             LogEntry::AssistantItem { item, .. } => {
                 state.history.push(Item::from(item.clone()));
             }

crates/session-store/src/worker_session_store.rs

@@ -20,7 +20,8 @@ use std::path::{Path, PathBuf};
 use std::sync::{Arc, Mutex};
 use std::time::SystemTime;
 
-const SESSION_SCHEMA_VERSION: u32 = 1;
+const SESSION_SCHEMA_VERSION: u32 = 2;
+const LEGACY_SESSION_SCHEMA_VERSION: u32 = 1;
 const SESSION_FILE: &str = "session.json";
 const SEGMENTS_DIR: &str = "segments";
 
@@ -44,16 +45,23 @@ impl WorkerSessionStore {
         fs::create_dir_all(root.join(SEGMENTS_DIR))?;
         let session_id = match fs::read(root.join(SESSION_FILE)) {
             Ok(bytes) => {
-                let manifest: SessionManifest = serde_json::from_slice(&bytes)?;
-                if manifest.schema_version != SESSION_SCHEMA_VERSION {
+                let mut manifest: SessionManifest = serde_json::from_slice(&bytes)?;
+                match manifest.schema_version {
+                    SESSION_SCHEMA_VERSION => {}
+                    LEGACY_SESSION_SCHEMA_VERSION => {
+                        validate_legacy_segment_logs(&root)?;
+                        manifest.schema_version = SESSION_SCHEMA_VERSION;
+                        atomic_write_json(&root.join(SESSION_FILE), &manifest)?;
+                    }
+                    version => {
                         return Err(StoreError::Corrupt {
                             line: 0,
                             message: format!(
-                            "unsupported Worker Session schema version {}, expected {}",
-                            manifest.schema_version, SESSION_SCHEMA_VERSION
+                                "unsupported Worker Session schema version {version}, expected {SESSION_SCHEMA_VERSION}"
                             ),
                         });
                     }
+                }
                 Some(manifest.session_id)
             }
             Err(error) if error.kind() == std::io::ErrorKind::NotFound => None,
@@ -278,6 +286,37 @@ impl Store for WorkerSessionStore {
     }
 }
 
+fn validate_legacy_segment_logs(root: &Path) -> Result<(), StoreError> {
+    let segments = root.join(SEGMENTS_DIR);
+    if !segments.exists() {
+        return Ok(());
+    }
+    for entry in fs::read_dir(&segments)? {
+        let entry = entry?;
+        let path = entry.path();
+        let Some(name) = path.file_name().and_then(|name| name.to_str()) else {
+            continue;
+        };
+        if !name.ends_with(".jsonl") || name.ends_with(".trace.jsonl") {
+            continue;
+        }
+        let contents = fs::read_to_string(&path)?;
+        for (line_index, line) in contents.lines().enumerate() {
+            if line.trim().is_empty() {
+                continue;
+            }
+            serde_json::from_str::<LogEntry>(line).map_err(|error| StoreError::Corrupt {
+                line: line_index + 1,
+                message: format!(
+                    "cannot migrate legacy Worker Session log {}: {error}",
+                    path.display()
+                ),
+            })?;
+        }
+    }
+    Ok(())
+}
+
 fn atomic_write_json<T: Serialize>(path: &Path, value: &T) -> Result<(), StoreError> {
     let mut bytes = serde_json::to_vec_pretty(value)?;
     bytes.push(b'\n');
@@ -405,6 +444,54 @@ mod tests {
         assert_eq!(store.list_sessions().unwrap(), vec![session_id]);
     }
 
+    #[test]
+    fn schema_v1_logs_are_validated_and_promoted_to_v2() {
+        let root = tempfile::tempdir().unwrap();
+        let session_id = new_session_id();
+        let segment_id = new_segment_id();
+        WorkerSessionStore::new(root.path())
+            .unwrap()
+            .create_segment(session_id, segment_id, &[])
+            .unwrap();
+        let manifest_path = root.path().join(SESSION_FILE);
+        let mut manifest: SessionManifest =
+            serde_json::from_slice(&fs::read(&manifest_path).unwrap()).unwrap();
+        manifest.schema_version = LEGACY_SESSION_SCHEMA_VERSION;
+        atomic_write_json(&manifest_path, &manifest).unwrap();
+
+        let reopened = WorkerSessionStore::new(root.path()).unwrap();
+        assert_eq!(reopened.session_id().unwrap(), Some(session_id));
+        let migrated: SessionManifest =
+            serde_json::from_slice(&fs::read(&manifest_path).unwrap()).unwrap();
+        assert_eq!(migrated.schema_version, SESSION_SCHEMA_VERSION);
+    }
+
+    #[test]
+    fn schema_v1_migration_rejects_corrupt_log_before_manifest_update() {
+        let root = tempfile::tempdir().unwrap();
+        let session_id = new_session_id();
+        let manifest = SessionManifest {
+            schema_version: LEGACY_SESSION_SCHEMA_VERSION,
+            session_id,
+        };
+        atomic_write_json(&root.path().join(SESSION_FILE), &manifest).unwrap();
+        fs::create_dir_all(root.path().join(SEGMENTS_DIR)).unwrap();
+        fs::write(
+            root.path().join(SEGMENTS_DIR).join("broken.jsonl"),
+            "{not-json}\n",
+        )
+        .unwrap();
+
+        let error = match WorkerSessionStore::new(root.path()) {
+            Ok(_) => panic!("corrupt legacy Session log must reject migration"),
+            Err(error) => error,
+        };
+        assert!(matches!(error, StoreError::Corrupt { .. }));
+        let persisted: SessionManifest =
+            serde_json::from_slice(&fs::read(root.path().join(SESSION_FILE)).unwrap()).unwrap();
+        assert_eq!(persisted.schema_version, LEGACY_SESSION_SCHEMA_VERSION);
+    }
+
     #[test]
     fn reopen_preserves_session_and_segment_ids() {
         let root = tempfile::tempdir().unwrap();

crates/session-store/tests/session_test.rs

@@ -1,12 +1,13 @@
 mod common;
 
+use std::ops::{Deref, DerefMut};
 use std::sync::Arc;
 
-use agen::Engine;
 use agen::interceptor::{Interceptor, TurnEndAction};
 use agen::llm_client::event::{Event, ResponseStatus, StatusEvent};
 use agen::llm_client::types::{Item, RequestConfig};
 use agen::tool::{Tool, ToolDefinition, ToolError, ToolMeta, ToolOutput};
+use agen::{Engine, History};
 use async_trait::async_trait;
 use common::MockLlmClient;
 use session_store::{FsStore, LogEntry, SegmentStartState, Store, collect_state};
@@ -94,15 +95,47 @@ fn make_store() -> (tempfile::TempDir, FsStore) {
     (dir, store)
 }
 
+struct TestWorker {
+    engine: Engine<MockLlmClient>,
+    history: History,
+}
+
+impl TestWorker {
+    fn new(engine: Engine<MockLlmClient>) -> Self {
+        Self {
+            engine,
+            history: History::new(),
+        }
+    }
+
+    fn history(&self) -> Vec<Item> {
+        self.history.items_cloned()
+    }
+}
+
+impl Deref for TestWorker {
+    type Target = Engine<MockLlmClient>;
+
+    fn deref(&self) -> &Self::Target {
+        &self.engine
+    }
+}
+
+impl DerefMut for TestWorker {
+    fn deref_mut(&mut self) -> &mut Self::Target {
+        &mut self.engine
+    }
+}
+
 /// Run a worker turn and persist via session-store functions.
 /// Takes ownership of the worker (needed for lock/unlock) and returns it.
 async fn run_and_persist(
-    worker: Engine<MockLlmClient>,
+    mut worker: TestWorker,
     store: &FsStore,
     session_id: session_store::SessionId,
     segment_id: session_store::SegmentId,
     input: &str,
-) -> (Engine<MockLlmClient>, agen::EngineRunExit) {
+) -> (TestWorker, agen::EngineRunExit) {
     // Mirror Worker's run-entry contract: log the user input as segments
     // before the worker pushes its flattened user_message; save_delta
     // skips the resulting user_message item to avoid double-write.
@@ -114,13 +147,14 @@ async fn run_and_persist(
     )
     .unwrap();
 
-    let history_before = worker.history().len();
+    let history_before = worker.history.len();
 
-    let mut locked = worker.lock();
-    let result = locked.run(input).await;
-    let worker = locked.unlock();
+    let mut locked = worker.engine.lock(&worker.history);
+    let result = locked.run(&mut worker.history, input).await;
+    worker.engine = locked.unlock();
 
-    let new_items = &worker.history()[history_before..];
+    let projected = worker.history();
+    let new_items = &projected[history_before..];
     session_store::save_delta(store, session_id, segment_id, new_items).unwrap();
     session_store::save_turn_end(store, session_id, segment_id, worker.turn_count()).unwrap();
 
@@ -178,14 +212,14 @@ async fn run_and_persist(
 async fn session_run_logs_entries() {
     let (_dir, store) = make_store();
     let client = MockLlmClient::new(simple_text_events());
-    let worker = Engine::new(client);
+    let worker = TestWorker::new(Engine::new(client));
 
     let (sid, segid) = session_store::create_segment(
         &store,
         SegmentStartState {
             system_prompt: worker.get_system_prompt(),
             config: worker.request_config(),
-            history: worker.history(),
+            history: &worker.history(),
         },
     )
     .unwrap();
@@ -222,7 +256,7 @@ async fn session_run_logs_entries() {
 async fn session_restore_round_trip() {
     let (_dir, store) = make_store();
     let client = MockLlmClient::new(simple_text_events());
-    let mut worker = Engine::new(client);
+    let mut worker = TestWorker::new(Engine::new(client));
     worker.set_system_prompt("You are helpful.");
 
     let (sid, segid) = session_store::create_segment(
@@ -230,7 +264,7 @@ async fn session_restore_round_trip() {
         SegmentStartState {
             system_prompt: worker.get_system_prompt(),
             config: worker.request_config(),
-            history: worker.history(),
+            history: &worker.history(),
         },
     )
     .unwrap();
@@ -261,7 +295,7 @@ async fn session_restore_round_trip() {
 async fn session_run_with_tool_call() {
     let (_dir, store) = make_store();
     let client = MockLlmClient::with_responses(tool_call_events());
-    let mut worker = Engine::new(client);
+    let mut worker = TestWorker::new(Engine::new(client));
     worker.register_tool(weather_tool_definition());
 
     let (sid, segid) = session_store::create_segment(
@@ -269,7 +303,7 @@ async fn session_run_with_tool_call() {
         SegmentStartState {
             system_prompt: worker.get_system_prompt(),
             config: worker.request_config(),
-            history: worker.history(),
+            history: &worker.history(),
         },
     )
     .unwrap();
@@ -295,7 +329,7 @@ async fn session_resume_after_pause() {
 
     // First run: tool call with pause policy → Paused
     let client = MockLlmClient::with_responses(tool_call_events());
-    let mut worker = Engine::new(client);
+    let mut worker = TestWorker::new(Engine::new(client));
     worker.register_tool(weather_tool_definition());
     worker.set_interceptor(PausePolicy);
 
@@ -304,7 +338,7 @@ async fn session_resume_after_pause() {
         SegmentStartState {
             system_prompt: worker.get_system_prompt(),
             config: worker.request_config(),
-            history: worker.history(),
+            history: &worker.history(),
         },
     )
     .unwrap();
@@ -335,7 +369,7 @@ async fn session_resume_after_pause() {
 async fn session_fork_creates_new_session() {
     let (_dir, store) = make_store();
     let client = MockLlmClient::new(simple_text_events());
-    let mut worker = Engine::new(client);
+    let mut worker = TestWorker::new(Engine::new(client));
     worker.set_system_prompt("System prompt");
 
     let (sid, segid) = session_store::create_segment(
@@ -343,7 +377,7 @@ async fn session_fork_creates_new_session() {
         SegmentStartState {
             system_prompt: worker.get_system_prompt(),
             config: worker.request_config(),
-            history: worker.history(),
+            history: &worker.history(),
         },
     )
     .unwrap();
@@ -356,7 +390,7 @@ async fn session_fork_creates_new_session() {
         SegmentStartState {
             system_prompt: worker.get_system_prompt(),
             config: worker.request_config(),
-            history: worker.history(),
+            history: &worker.history(),
         },
     )
     .unwrap();
@@ -377,14 +411,14 @@ async fn session_fork_creates_new_session() {
 async fn session_fork_at_truncates_within_session() {
     let (_dir, store) = make_store();
     let client = MockLlmClient::new(simple_text_events());
-    let worker = Engine::new(client);
+    let worker = TestWorker::new(Engine::new(client));
 
     let (sid, segid) = session_store::create_segment(
         &store,
         SegmentStartState {
             system_prompt: worker.get_system_prompt(),
             config: worker.request_config(),
-            history: worker.history(),
+            history: &worker.history(),
         },
     )
     .unwrap();
@@ -422,14 +456,14 @@ async fn session_fork_at_truncates_within_session() {
 async fn session_config_changed_logged() {
     let (_dir, store) = make_store();
     let client = MockLlmClient::new(vec![]);
-    let mut worker = Engine::new(client);
+    let mut worker = TestWorker::new(Engine::new(client));
 
     let (sid, segid) = session_store::create_segment(
         &store,
         SegmentStartState {
             system_prompt: worker.get_system_prompt(),
             config: worker.request_config(),
-            history: worker.history(),
+            history: &worker.history(),
         },
     )
     .unwrap();
@@ -455,14 +489,14 @@ async fn session_auto_forks_on_conflict() {
 
     // Create a segment
     let client_a = MockLlmClient::new(simple_text_events());
-    let worker_a = Engine::new(client_a);
+    let worker_a = TestWorker::new(Engine::new(client_a));
 
     let (sid, original_segid) = session_store::create_segment(
         &store,
         SegmentStartState {
             system_prompt: worker_a.get_system_prompt(),
             config: worker_a.request_config(),
-            history: worker_a.history(),
+            history: &worker_a.history(),
         },
     )
     .unwrap();
@@ -488,7 +522,7 @@ async fn session_auto_forks_on_conflict() {
         SegmentStartState {
             system_prompt: worker_a.get_system_prompt(),
             config: worker_a.request_config(),
-            history: worker_a.history(),
+            history: &worker_a.history(),
         },
     )
     .unwrap();
@@ -540,14 +574,14 @@ async fn session_auto_forks_on_conflict() {
 async fn nested_past_fork_leaves_ancestors_immutable() {
     let (_dir, store) = make_store();
     let client = MockLlmClient::new(simple_text_events());
-    let worker = Engine::new(client);
+    let worker = TestWorker::new(Engine::new(client));
 
     let (sid, root_segid) = session_store::create_segment(
         &store,
         SegmentStartState {
             system_prompt: worker.get_system_prompt(),
             config: worker.request_config(),
-            history: worker.history(),
+            history: &worker.history(),
         },
     )
     .unwrap();

crates/ticket/src/tool.rs

@@ -1546,7 +1546,7 @@ fn model_ticket_reference(
     match ticket.meta.resource_key {
         Some(resource_key) if is_canonical_ticket_resource_key(&resource_key) => Ok(resource_key),
         Some(_) => Err(ToolError::ExecutionFailed(format!(
-            "{tool_name} failed: required Ticket human key is unavailable"
+            "{tool_name} failed: required Ticket key is unavailable"
         ))),
         None => Ok(ticket.meta.id),
     }

crates/tools/src/bash.rs

@@ -1,5 +1,6 @@
+use std::collections::{HashMap, HashSet};
 use std::path::PathBuf;
-use std::sync::Arc;
+use std::sync::{Arc, Mutex};
 
 use agen::tool::{Tool, ToolDefinition, ToolError, ToolMeta, ToolOutput};
 use async_trait::async_trait;
@@ -20,21 +21,65 @@ struct BashParams {
 
 pub(crate) struct BashTool {
     session: WorkdirSessionHandle,
+    state: Arc<Mutex<BashExecutionState>>,
+}
+
+#[derive(Clone)]
+struct ActiveCommand {
+    call_id: String,
+    execution_nonce: u64,
+    handle: CommandHandle,
+}
+
+#[derive(Default)]
+struct BashExecutionState {
+    active: HashMap<String, ActiveCommand>,
+    cancellation_requested: HashSet<String>,
+    legacy_cancellation_requested: HashSet<String>,
+    next_execution_nonce: u64,
 }
 
 struct CommandGuard {
     session: WorkdirSessionHandle,
+    state: Arc<Mutex<BashExecutionState>>,
+    execution_id: String,
+    execution_nonce: u64,
     handle: Option<CommandHandle>,
 }
 
 impl Drop for CommandGuard {
     fn drop(&mut self) {
-        if let Some(handle) = self.handle.take() {
+        let Some(handle) = self.handle.take() else {
+            return;
+        };
         let workdir = self.session.clone();
+        let state = Arc::clone(&self.state);
+        let execution_id = self.execution_id.clone();
+        let execution_nonce = self.execution_nonce;
+        // A dropped provider future is not terminal confirmation. Keep the live
+        // execution registered until cleanup has both requested cancellation and
+        // observed terminal command output, so cancellation/session teardown
+        // cannot race with an apparently empty registry.
         tokio::spawn(async move {
-                let _ = workdir.cancel_command(handle).await;
-            });
+            let _ = workdir.cancel_command(handle.clone()).await;
+            let _ = workdir
+                .command_output(CommandOutputRequest {
+                    handle,
+                    cursor: 0,
+                    limit: INLINE_BYTE_BUDGET,
+                    wait: true,
+                })
+                .await;
+            let mut state = state.lock().unwrap();
+            if state
+                .active
+                .get(&execution_id)
+                .is_some_and(|active| active.execution_nonce == execution_nonce)
+            {
+                state.active.remove(&execution_id);
+                state.cancellation_requested.remove(&execution_id);
             }
+        });
     }
 }
 
@@ -52,20 +97,50 @@ impl Tool for BashTool {
             .unwrap_or(DEFAULT_TIMEOUT_SECS)
             .clamp(1, MAX_TIMEOUT_SECS);
         let cmd_summary = truncate_for_summary(&params.command);
+        let execution_id = ctx.execution_id();
+        let call_id = ctx.call_id;
+        let execution_nonce = {
+            let mut state = self.state.lock().unwrap();
+            state.next_execution_nonce = state.next_execution_nonce.wrapping_add(1);
+            state.next_execution_nonce
+        };
+        let mut guard = CommandGuard {
+            session: self.session.clone(),
+            state: self.state.clone(),
+            execution_id: execution_id.clone(),
+            execution_nonce,
+            handle: None,
+        };
         let handle = self
             .session
             .start_command(CommandRequest {
                 command: params.command,
                 timeout_secs,
                 output_limit: INLINE_BYTE_BUDGET,
-                tool_call_id: Some(ctx.call_id),
+                tool_call_id: Some(call_id.clone()),
             })
             .await
             .map_err(crate::ToolsError::from)?;
-        let mut guard = CommandGuard {
-            session: self.session.clone(),
-            handle: Some(handle.clone()),
+        let cancel_after_start = {
+            let mut state = self.state.lock().unwrap();
+            state.active.insert(
+                execution_id.clone(),
+                ActiveCommand {
+                    call_id: call_id.clone(),
+                    execution_nonce,
+                    handle: handle.clone(),
+                },
+            );
+            state.cancellation_requested.contains(&execution_id)
+                || state.legacy_cancellation_requested.contains(&call_id)
         };
+        guard.handle = Some(handle.clone());
+        if cancel_after_start {
+            self.session
+                .cancel_command(handle.clone())
+                .await
+                .map_err(crate::ToolsError::from)?;
+        }
         let output = self
             .session
             .command_output(CommandOutputRequest {
@@ -76,9 +151,27 @@ impl Tool for BashTool {
             })
             .await
             .map_err(crate::ToolsError::from)?;
+        let cancellation_requested = {
+            let mut state = self.state.lock().unwrap();
+            let owns_registration = state
+                .active
+                .get(&execution_id)
+                .is_some_and(|active| active.execution_nonce == execution_nonce);
+            let exact = if owns_registration {
+                state.active.remove(&execution_id);
+                state.cancellation_requested.remove(&execution_id)
+            } else {
+                false
+            };
+            let legacy = state.legacy_cancellation_requested.remove(&call_id);
+            exact || legacy
+        };
         guard.handle = None;
 
-        let summary = if output.timed_out {
+        let timed_out = output.timed_out;
+        let summary = if cancellation_requested {
+            format!("$ {cmd_summary} (cancelled)")
+        } else if output.timed_out {
             format!("$ {cmd_summary} (timed out after {timeout_secs}s)")
         } else {
             match output.exit_code {
@@ -97,11 +190,62 @@ impl Tool for BashTool {
         } else {
             Some(output.content)
         };
-        Ok(ToolOutput {
+        let output = ToolOutput {
             summary,
             content,
             attachments: Vec::new(),
-        })
+        };
+        if cancellation_requested {
+            Err(ToolError::Cancelled(output))
+        } else if timed_out {
+            Err(ToolError::Interrupted(output))
+        } else {
+            Ok(output)
+        }
+    }
+
+    async fn cancel(&self, call_id: &str) -> Result<(), ToolError> {
+        let handles = {
+            let mut state = self.state.lock().unwrap();
+            state
+                .legacy_cancellation_requested
+                .insert(call_id.to_string());
+            state
+                .active
+                .values()
+                .filter(|active| active.call_id == call_id)
+                .map(|active| active.handle.clone())
+                .collect::<Vec<_>>()
+        };
+        for handle in handles {
+            self.session
+                .cancel_command(handle)
+                .await
+                .map_err(crate::ToolsError::from)?;
+        }
+        Ok(())
+    }
+
+    async fn cancel_execution(
+        &self,
+        ctx: &agen::tool::ToolExecutionContext,
+    ) -> Result<(), ToolError> {
+        let execution_id = ctx.execution_id();
+        let handle = {
+            let mut state = self.state.lock().unwrap();
+            state.cancellation_requested.insert(execution_id.clone());
+            state
+                .active
+                .get(&execution_id)
+                .map(|active| active.handle.clone())
+        };
+        if let Some(handle) = handle {
+            self.session
+                .cancel_command(handle)
+                .await
+                .map_err(crate::ToolsError::from)?;
+        }
+        Ok(())
     }
 }
 
@@ -123,6 +267,7 @@ pub fn bash_tool(session: WorkdirSessionHandle, _output_dir: PathBuf) -> ToolDef
             .input_schema(serde_json::to_value(schema).expect("Bash schema serialization"));
         let tool: Arc<dyn Tool> = Arc::new(BashTool {
             session: session.clone(),
+            state: Arc::new(Mutex::new(BashExecutionState::default())),
         });
         (meta, tool)
     })

crates/tools/src/grep.rs

@@ -129,7 +129,7 @@ pub fn grep_tool(session: WorkdirSessionHandle) -> ToolDefinition {
     Arc::new(move || {
         let schema = schemars::schema_for!(GrepParams);
         let meta = ToolMeta::new("Grep")
-            .description("Search Workdir file contents with a regex. Glob/Grep traversal executes inside the WorkdirSession provider. Results are bounded and Workdir-relative.")
+            .description("Search Workdir file contents with a regex. Content results group lines by file; `>` marks matching lines and unmarked lines are context. Glob/Grep traversal executes inside the WorkdirSession provider. Results are bounded and Workdir-relative.")
             .input_schema(serde_json::to_value(schema).expect("Grep schema serialization"));
         let tool: Arc<dyn Tool> = Arc::new(GrepTool {
             session: session.clone(),

crates/tools/tests/integration.rs

@@ -7,7 +7,10 @@
 use std::path::Path;
 use std::sync::Arc;
 
-use agen::tool::{Tool, ToolDefinition, ToolMeta};
+use agen::tool::{
+    Tool, ToolDefinition, ToolError, ToolExecutionContext, ToolExecutionHandle,
+    ToolExecutionTerminal, ToolMeta,
+};
 use manifest::{Permission, Scope, ScopeConfig, ScopeRule};
 use serde_json::json;
 use tempfile::TempDir;
@@ -401,5 +404,84 @@ async fn bash_provider_output_does_not_expose_internal_paths() {
     assert_eq!(std::fs::read_dir(spill.path()).unwrap().count(), 0);
 }
 
+#[tokio::test]
+async fn bash_cancellation_returns_bounded_progress_as_terminal_output() {
+    let (dir, _spill, reg) = setup();
+    let marker = dir.path().join("must-not-run-after-cancel");
+    let command = format!(
+        "printf 'before\\n'; printf 'err-before\\n' >&2; sleep 1; touch {}; printf 'after\\n'",
+        marker.display()
+    );
+    let input = serde_json::to_string(&json!({ "command": command })).unwrap();
+    let context = ToolExecutionContext::new("call-heavy", "attempt-heavy", 0);
+    let bash = reg.get("Bash");
+    let executing = bash.clone();
+    let execution_context = context.clone();
+    let execution = tokio::spawn(async move { executing.execute(&input, execution_context).await });
+
+    tokio::time::sleep(std::time::Duration::from_millis(100)).await;
+    bash.cancel_execution(&context)
+        .await
+        .expect("signal exact execution cancellation");
+    let error = tokio::time::timeout(std::time::Duration::from_secs(2), execution)
+        .await
+        .expect("cancelled Bash should terminate inside the Engine grace budget")
+        .expect("Bash task join");
+
+    let ToolError::Cancelled(output) = error.expect_err("cancelled command is non-success") else {
+        panic!("expected typed cancellation result");
+    };
+    let content = output.content.expect("bounded progress output");
+    assert!(
+        content.contains("before"),
+        "missing pre-cancel stdout: {content}"
+    );
+    assert!(
+        content.contains("err-before"),
+        "missing pre-cancel stderr: {content}"
+    );
+    assert!(
+        !content.contains("after"),
+        "post-cancel output leaked: {content}"
+    );
+    assert!(content.len() <= 16 * 1024, "output must remain bounded");
+
+    tokio::time::sleep(std::time::Duration::from_millis(1_100)).await;
+    assert!(
+        !marker.exists(),
+        "the cancelled command continued executing after terminal confirmation"
+    );
+}
+
+#[tokio::test]
+async fn bash_force_close_cleanup_stops_command_and_keeps_session_reusable() {
+    let (dir, _spill, reg) = setup();
+    let marker = dir.path().join("must-not-survive-force-close");
+    let command = format!("sleep 1; touch {}", marker.display());
+    let input = serde_json::to_string(&json!({ "command": command })).unwrap();
+    let bash = reg.get("Bash");
+    let context = ToolExecutionContext::new("call-force", "attempt-force", 0);
+    let (handle, terminal) = ToolExecutionHandle::start(bash.clone(), input, context);
+
+    tokio::time::sleep(std::time::Duration::from_millis(100)).await;
+    handle.force_close();
+    assert!(matches!(
+        terminal.await,
+        ToolExecutionTerminal::OutcomeUnknown
+    ));
+
+    tokio::time::sleep(std::time::Duration::from_millis(1_100)).await;
+    assert!(
+        !marker.exists(),
+        "CommandGuard cleanup allowed a force-closed command to continue"
+    );
+
+    let output = bash
+        .execute(r#"{"command":"printf 'reused'"}"#, Default::default())
+        .await
+        .expect("workdir session remains reusable after cleanup");
+    assert_eq!(output.content.as_deref(), Some("reused"));
+}
+
 // Sanity: unused Path import guard
 const _: fn() -> &'static Path = || Path::new("/");

crates/tui/src/app.rs

@@ -1244,6 +1244,7 @@ impl App {
                 id,
                 summary,
                 output,
+                disposition: _,
                 is_error,
             } => {
                 self.latest_llm_wait_event = None;

crates/worker-runtime/src/worker_backend.rs

@@ -38,9 +38,7 @@ use crate::working_directory::{
 };
 use async_trait::async_trait;
 use protocol::{Event, Method, Segment, WorkerStatus};
-use session_store::{
-    CombinedStore, LogEntry, WorkerAggregateStore, WorkerSessionStore, collect_state,
-};
+use session_store::{CombinedStore, LogEntry, WorkerAggregateStore, WorkerSessionStore};
 #[cfg(test)]
 use session_store::{FsStore, FsWorkerStore};
 use tokio::runtime::Runtime;
@@ -68,8 +66,10 @@ const RUNTIME_TASK_TIMEOUT: Duration = Duration::from_secs(10);
 const USER_INPUT_COMMIT_TIMEOUT: Duration = Duration::from_secs(9);
 
 fn user_input_has_submission(entry: &LogEntry, submission_id: &str) -> bool {
-    let LogEntry::UserInput { extensions, .. } = entry else {
-        return false;
+    let extensions = match entry {
+        LogEntry::UserInput { extensions, .. }
+        | LogEntry::AnnotatedUserInput { extensions, .. } => extensions,
+        _ => return false,
     };
     extensions.iter().any(|extension| {
         extension.domain == WORKER_INPUT_SUBMISSION_EXTENSION_DOMAIN
@@ -212,11 +212,11 @@ impl WorkerObservationProvider for RuntimeGrantedWorkerObservationProvider {
             return Err(WorkerObservationError::NotFound);
         }
         let entries = sink.subscribe_with_snapshot().0;
-        let state = collect_state(&entries);
-        Ok(WorkerSessionCapture {
-            segment_id: format!("runtime:{runtime_id}:worker:{worker_id}"),
-            items: state.history,
-        })
+        WorkerSessionCapture::from_log_entries(
+            format!("runtime:{runtime_id}:worker:{worker_id}"),
+            &entries,
+        )
+        .map_err(WorkerObservationError::Unavailable)
     }
 }
 
@@ -2507,7 +2507,9 @@ mod tests {
             let scope = Scope::writable(&scope_root).map_err(|err| err.to_string())?;
             let worker = Worker::new(
                 manifest,
-                Engine::new(self.client.clone()),
+                Engine::<_, agen::state::Mutable, worker::SessionHistoryMetadata>::new_annotated(
+                    self.client.clone(),
+                ),
                 store,
                 workspace_context,
                 filesystem_authority,
@@ -3243,14 +3245,17 @@ mod tests {
             matches!(
                 entry,
                 LogEntry::UserInput { segments, .. }
+                | LogEntry::AnnotatedUserInput { segments, .. }
                     if segments == &vec![Segment::text("start the ticket")]
             )
         }));
         let submission_id = entries
             .iter()
             .find_map(|entry| {
-                let LogEntry::UserInput { extensions, .. } = entry else {
-                    return None;
+                let extensions = match entry {
+                    LogEntry::UserInput { extensions, .. }
+                    | LogEntry::AnnotatedUserInput { extensions, .. } => extensions,
+                    _ => return None,
                 };
                 extensions
                     .iter()

crates/worker/examples/worker_cli.rs

@@ -71,7 +71,7 @@ async fn main() -> Result<(), Box<dyn std::error::Error>> {
     }
 
     // 5. Extract the assistant's reply from history
-    let history = worker.engine().history();
+    let history = worker.history();
     if let Some(text) = history
         .iter()
         .rev()

crates/worker/src/compact/prune.rs

@@ -22,7 +22,7 @@ use crate::compact::token_counter::{
     EstimateSource, savings_for_prune_impl, token_estimates_for_prune_impl,
 };
 
-impl<C: LlmClient, St: Store> Worker<C, St> {
+impl<C: LlmClient + 'static, St: Store> Worker<C, St> {
     /// Enable prune projection on the underlying Engine.
     ///
     /// Registers the config and token/savings-estimator closures on the Engine.

crates/worker/src/compact/token_counter.rs

@@ -242,13 +242,13 @@ pub(crate) fn savings_for_prune_impl(
 
 // ── Worker に生やす公開 API ───────────────────────────────────────────────
 
-impl<C: LlmClient, St: Store> Worker<C, St> {
+impl<C: LlmClient + 'static, St: Store> Worker<C, St> {
     /// 現在の history 全体の推定トークン数。
     ///
     /// 最後の measurement と、その後に追加された未測定分の byte/4 外挿。
     pub fn total_tokens(&self) -> TokenEstimate {
         let usage = self.usage_history();
-        agen::token_counter::total_tokens(self.history(), &usage)
+        agen::token_counter::total_tokens(&self.history(), &usage)
     }
 
     /// 任意の history index 時点でのプロンプト全長推定。
@@ -259,7 +259,7 @@ impl<C: LlmClient, St: Store> Worker<C, St> {
     /// pointer 以降に増えたプロンプト長を測るのに使う。
     pub fn total_tokens_at(&self, history_len: usize) -> TokenEstimate {
         let usage = self.usage_history();
-        agen::token_counter::total_tokens_at(self.history(), &usage, history_len)
+        agen::token_counter::total_tokens_at(&self.history(), &usage, history_len)
     }
 
     /// 末尾から `retained` トークン以上を残すための分割位置。
@@ -267,7 +267,7 @@ impl<C: LlmClient, St: Store> Worker<C, St> {
     /// `history[..cut.index]` が要約／破棄される側、`history[cut.index..]` が残る側。
     pub fn split_for_retained(&self, retained: u64) -> SplitPoint {
         let usage = self.usage_history();
-        split_for_retained_impl(self.history(), &usage, retained)
+        split_for_retained_impl(&self.history(), &usage, retained)
     }
 }
 

crates/worker/src/controller.rs

@@ -485,6 +485,7 @@ impl WorkerController {
         // into the controller task so the in-flight turn can be reached
         // via these handles while worker itself is borrowed by drive_turn.
         let cancel_tx = worker.engine_mut().cancel_sender();
+        let pause_tx = worker.engine_mut().pause_sender();
         let notify_buffer = worker.notify_buffer_handle();
 
         tokio::spawn(controller_loop(
@@ -494,6 +495,7 @@ impl WorkerController {
             shared_state,
             runtime_dir,
             cancel_tx,
+            pause_tx,
             notify_buffer,
             self_parent_socket,
             spawner_name,
@@ -763,6 +765,19 @@ pub(crate) fn wire_event_bridges_on_engine<C, St>(
             id: result.tool_use_id.clone(),
             summary: result.summary.clone(),
             output: result.content.clone(),
+            disposition: Some(match result.disposition {
+                agen::ToolResultDisposition::Success => protocol::ToolResultDisposition::Success,
+                agen::ToolResultDisposition::Error => protocol::ToolResultDisposition::Error,
+                agen::ToolResultDisposition::Interrupted => {
+                    protocol::ToolResultDisposition::Interrupted
+                }
+                agen::ToolResultDisposition::Cancelled => {
+                    protocol::ToolResultDisposition::Cancelled
+                }
+                agen::ToolResultDisposition::OutcomeUnknown => {
+                    protocol::ToolResultDisposition::OutcomeUnknown
+                }
+            }),
             is_error: result.is_error,
         });
     });
@@ -983,14 +998,6 @@ where
 
     if feature_config.sub_worker.enabled {
         worker.register_worker_orchestration_instruction();
-        if !feature_config.worker.enabled {
-            feature_registry.add_module(
-                crate::feature::builtin::manage_worker::sub_worker_control_feature(
-                    worker.workspace_client_handle(),
-                    spawned_registry.clone(),
-                ),
-            );
-        }
     }
 
     let host_worker_observation_provider = worker.worker_observation_provider();
@@ -1123,6 +1130,7 @@ async fn controller_loop<C, St>(
     shared_state: Arc<WorkerSharedState>,
     runtime_dir: Arc<RuntimeDir>,
     cancel_tx: mpsc::Sender<()>,
+    pause_tx: mpsc::Sender<()>,
     notify_buffer: NotifyBuffer,
     self_parent_socket: Option<PathBuf>,
     spawner_name: String,
@@ -1169,6 +1177,9 @@ async fn controller_loop<C, St>(
             // clear at run start prevents stale partial output left by an older
             // interrupted/error turn from being carried into the next snapshot.
             worker.clear_in_flight_events();
+            let parent_originated = run.is_parent_originated();
+            let user_input_run = matches!(&run, PendingRun::Run(_) | PendingRun::RunTracked { .. });
+            if !user_input_run {
                 set_controller_status(
                     &shared_state,
                     &runtime_dir,
@@ -1176,15 +1187,25 @@ async fn controller_loop<C, St>(
                     WorkerStatus::Running,
                 )
                 .await;
-            let parent_originated = run.is_parent_originated();
+            }
             let (mut new_status, shutdown) = match run {
                 PendingRun::Run(input) => {
+                    let (input_commit_tx, input_commit_rx) = oneshot::channel();
                     drive_turn(
-                        worker.run(input),
+                        worker.run_with_input_extensions_and_commit_hook(
+                            input,
+                            Vec::new(),
+                            move || {
+                                let _ = input_commit_tx.send(());
+                            },
+                        ),
                         &mut method_rx,
                         &event_tx,
                         &cancel_tx,
+                        &pause_tx,
                         &shared_state,
+                        &runtime_dir,
+                        Some(input_commit_rx),
                         &notify_buffer,
                         self_parent_socket.as_ref(),
                         &spawner_name,
@@ -1194,12 +1215,22 @@ async fn controller_loop<C, St>(
                     .await
                 }
                 PendingRun::RunTracked { input, extension } => {
+                    let (input_commit_tx, input_commit_rx) = oneshot::channel();
                     drive_turn(
-                        worker.run_with_input_extensions(input, vec![extension]),
+                        worker.run_with_input_extensions_and_commit_hook(
+                            input,
+                            vec![extension],
+                            move || {
+                                let _ = input_commit_tx.send(());
+                            },
+                        ),
                         &mut method_rx,
                         &event_tx,
                         &cancel_tx,
+                        &pause_tx,
                         &shared_state,
+                        &runtime_dir,
+                        Some(input_commit_rx),
                         &notify_buffer,
                         self_parent_socket.as_ref(),
                         &spawner_name,
@@ -1214,7 +1245,10 @@ async fn controller_loop<C, St>(
                         &mut method_rx,
                         &event_tx,
                         &cancel_tx,
+                        &pause_tx,
                         &shared_state,
+                        &runtime_dir,
+                        None,
                         &notify_buffer,
                         self_parent_socket.as_ref(),
                         &spawner_name,
@@ -1229,7 +1263,10 @@ async fn controller_loop<C, St>(
                         &mut method_rx,
                         &event_tx,
                         &cancel_tx,
+                        &pause_tx,
                         &shared_state,
+                        &runtime_dir,
+                        None,
                         &notify_buffer,
                         self_parent_socket.as_ref(),
                         &spawner_name,
@@ -1626,7 +1663,10 @@ async fn drive_turn<F>(
     method_rx: &mut mpsc::Receiver<Method>,
     event_tx: &broadcast::Sender<Event>,
     cancel_tx: &mpsc::Sender<()>,
+    pause_tx: &mpsc::Sender<()>,
     shared_state: &Arc<WorkerSharedState>,
+    runtime_dir: &RuntimeDir,
+    mut input_commit_rx: Option<oneshot::Receiver<()>>,
     notify_buffer: &NotifyBuffer,
     parent_socket: Option<&PathBuf>,
     self_name: &str,
@@ -1642,10 +1682,34 @@ where
 
     loop {
         tokio::select! {
+            // If input commit and provider completion become ready together, expose
+            // Running only after processing the commit fence. This makes the
+            // Running snapshot contract deterministic even for immediate clients.
+            biased;
+            committed = async {
+                input_commit_rx
+                    .as_mut()
+                    .expect("input commit receiver guarded by select condition")
+                    .await
+            }, if input_commit_rx.is_some() => {
+                input_commit_rx = None;
+                if committed.is_ok() {
+                    set_controller_status(
+                        shared_state,
+                        runtime_dir,
+                        event_tx,
+                        WorkerStatus::Running,
+                    )
+                    .await;
+                }
+            }
             result = &mut worker_future => {
                 return match result {
                     Ok(r) => {
                         let (status, run_result) = match r {
+                            WorkerRunResult::Finished if pause_requested => {
+                                (WorkerStatus::Paused, RunResult::Paused)
+                            }
                             WorkerRunResult::Finished => (WorkerStatus::Idle, RunResult::Finished),
                             WorkerRunResult::Paused => (WorkerStatus::Paused, RunResult::Paused),
                             WorkerRunResult::LimitReached => (WorkerStatus::Idle, RunResult::LimitReached),
@@ -1718,7 +1782,7 @@ where
                     }
                     Some(Method::Pause) => {
                         pause_requested = true;
-                        let _ = cancel_tx.try_send(());
+                        let _ = pause_tx.try_send(());
                     }
                     Some(Method::Shutdown) => {
                         shutdown_requested = true;
@@ -1780,7 +1844,7 @@ where
 
 fn emit_rewind_targets<C, St>(worker: &Worker<C, St>, event_tx: &broadcast::Sender<Event>)
 where
-    C: LlmClient,
+    C: LlmClient + 'static,
     St: Store,
 {
     match worker.list_rewind_targets() {
@@ -1806,7 +1870,7 @@ fn apply_rewind<C, St>(
     expected_head_entries: usize,
 ) -> bool
 where
-    C: LlmClient,
+    C: LlmClient + 'static,
     St: Store,
 {
     match worker.rewind_to(target, expected_head_entries) {
@@ -1854,7 +1918,7 @@ fn model_supports_image_attachments(model: &manifest::ModelManifest) -> bool {
 
 fn build_greeting<C, St>(worker: &Worker<C, St>) -> protocol::Greeting
 where
-    C: LlmClient,
+    C: LlmClient + 'static,
     St: Store,
 {
     let manifest = worker.manifest();
@@ -1970,11 +2034,13 @@ mod tests {
         event_tx: broadcast::Sender<Event>,
         cancel_tx: mpsc::Sender<()>,
         _cancel_rx: mpsc::Receiver<()>,
+        pause_tx: mpsc::Sender<()>,
+        _pause_rx: mpsc::Receiver<()>,
         shared_state: Arc<WorkerSharedState>,
         notify_buffer: NotifyBuffer,
         spawned_registry: Arc<SpawnedWorkerRegistry>,
         parent_socket_path: PathBuf,
-        _runtime_dir: Arc<RuntimeDir>,
+        runtime_dir: Arc<RuntimeDir>,
         _temp: TempDir,
     }
 
@@ -1988,6 +2054,7 @@ mod tests {
         let (method_tx, method_rx) = mpsc::channel::<Method>(16);
         let (event_tx, _) = broadcast::channel::<Event>(16);
         let (cancel_tx, cancel_rx) = mpsc::channel::<()>(1);
+        let (pause_tx, pause_rx) = mpsc::channel::<()>(1);
         let shared_state = Arc::new(WorkerSharedState::new(
             "child-worker".to_string(),
             session_store::new_segment_id(),
@@ -2013,11 +2080,13 @@ mod tests {
             event_tx,
             cancel_tx,
             _cancel_rx: cancel_rx,
+            pause_tx,
+            _pause_rx: pause_rx,
             shared_state,
             notify_buffer,
             spawned_registry,
             parent_socket_path,
-            _runtime_dir: runtime_dir,
+            runtime_dir,
             _temp: temp,
         }
     }
@@ -2070,7 +2139,10 @@ mod tests {
             &mut env.method_rx,
             &env.event_tx,
             &env.cancel_tx,
+            &env.pause_tx,
             &env.shared_state,
+            &env.runtime_dir,
+            None,
             &env.notify_buffer,
             Some(&env.parent_socket_path),
             "child-worker",
@@ -2091,6 +2163,44 @@ mod tests {
         }
     }
 
+    #[tokio::test]
+    async fn pause_waits_for_run_boundary_and_uses_safe_pause_channel() {
+        let mut env = make_env().await;
+        let method_tx = env._method_tx.clone();
+        tokio::spawn(async move {
+            tokio::time::sleep(Duration::from_millis(10)).await;
+            method_tx.send(Method::Pause).await.expect("send pause");
+        });
+
+        let worker_future = async {
+            tokio::time::sleep(Duration::from_millis(100)).await;
+            Ok::<_, WorkerError>(WorkerRunResult::Finished)
+        };
+        let started_at = std::time::Instant::now();
+        let (status, shutdown) = drive_turn(
+            worker_future,
+            &mut env.method_rx,
+            &env.event_tx,
+            &env.cancel_tx,
+            &env.pause_tx,
+            &env.shared_state,
+            &env.runtime_dir,
+            None,
+            &env.notify_buffer,
+            None,
+            "child-worker",
+            &env.spawned_registry,
+            true,
+        )
+        .await;
+
+        assert_eq!(status, WorkerStatus::Paused);
+        assert!(!shutdown);
+        assert!(started_at.elapsed() >= Duration::from_millis(100));
+        assert!(env._pause_rx.try_recv().is_ok());
+        assert!(env._cancel_rx.try_recv().is_err());
+    }
+
     #[tokio::test]
     async fn non_parent_originated_finished_stays_silent() {
         let mut env = make_env().await;
@@ -2102,7 +2212,10 @@ mod tests {
             &mut env.method_rx,
             &env.event_tx,
             &env.cancel_tx,
+            &env.pause_tx,
             &env.shared_state,
+            &env.runtime_dir,
+            None,
             &env.notify_buffer,
             Some(&env.parent_socket_path),
             "child-worker",
@@ -2137,7 +2250,10 @@ mod tests {
             &mut env.method_rx,
             &env.event_tx,
             &env.cancel_tx,
+            &env.pause_tx,
             &env.shared_state,
+            &env.runtime_dir,
+            None,
             &env.notify_buffer,
             Some(&env.parent_socket_path),
             "child-worker",
@@ -2178,7 +2294,10 @@ mod tests {
             &mut env.method_rx,
             &env.event_tx,
             &env.cancel_tx,
+            &env.pause_tx,
             &env.shared_state,
+            &env.runtime_dir,
+            None,
             &env.notify_buffer,
             Some(&env.parent_socket_path),
             "child-worker",
@@ -2217,7 +2336,10 @@ mod tests {
             &mut env.method_rx,
             &env.event_tx,
             &env.cancel_tx,
+            &env.pause_tx,
             &env.shared_state,
+            &env.runtime_dir,
+            None,
             &env.notify_buffer,
             Some(&env.parent_socket_path),
             "parent",
@@ -2253,7 +2375,10 @@ mod tests {
             &mut env.method_rx,
             &env.event_tx,
             &env.cancel_tx,
+            &env.pause_tx,
             &env.shared_state,
+            &env.runtime_dir,
+            None,
             &env.notify_buffer,
             Some(&env.parent_socket_path),
             "parent",
@@ -2287,7 +2412,10 @@ mod tests {
             &mut env.method_rx,
             &env.event_tx,
             &env.cancel_tx,
+            &env.pause_tx,
             &env.shared_state,
+            &env.runtime_dir,
+            None,
             &env.notify_buffer,
             Some(&env.parent_socket_path),
             "parent",
@@ -2320,7 +2448,10 @@ mod tests {
             &mut env.method_rx,
             &env.event_tx,
             &env.cancel_tx,
+            &env.pause_tx,
             &env.shared_state,
+            &env.runtime_dir,
+            None,
             &env.notify_buffer,
             Some(&env.parent_socket_path),
             "child-worker",

crates/worker/src/feature.rs

@@ -1795,9 +1795,9 @@ impl FeatureRegistryBuilder {
     }
 
     /// Install modules into the existing Engine tool path and hook builder.
-    pub(crate) fn install_into_engine<C: LlmClient>(
+    pub(crate) fn install_into_engine<C: LlmClient, A>(
         self,
-        worker: &mut Engine<C, Mutable>,
+        worker: &mut Engine<C, Mutable, A>,
         hook_builder: &mut HookRegistryBuilder,
     ) -> FeatureRegistryInstallReport {
         let mut pending_tools = Vec::new();

crates/worker/src/feature/builtin/memory_extract.rs

@@ -6,7 +6,9 @@ use memory::backend::{
     MemoryBackendOperation, MemoryBackendOperationResult, MemoryStageCandidateOperation,
 };
 use memory::extract::{CandidateKind, ExtractedCandidate, StagingEvidence};
-use memory::schema::{EvidenceKind, SourceEvidenceRef, SourceRef};
+use memory::schema::{
+    EvidenceKind, EvidenceOrigin, EvidenceOriginKind, SourceEvidenceRef, SourceRef,
+};
 use schemars::JsonSchema;
 use serde::Deserialize;
 
@@ -174,17 +176,29 @@ impl Tool for StageMemoryCandidateTool {
                 "StageMemoryCandidate requires at least one entry_ref".to_string(),
             ));
         }
-        let mut evidence = Vec::with_capacity(params.entry_refs.len());
-        let mut source_refs = Vec::with_capacity(params.entry_refs.len());
+        let mut entries = Vec::with_capacity(params.entry_refs.len());
         for entry_ref in &params.entry_refs {
-            let projection = self.state.view.evidence_for(entry_ref).ok_or_else(|| {
+            entries.push(self.state.view.evidence_for(entry_ref).ok_or_else(|| {
                 ToolError::InvalidArgument(format!(
                     "unknown SessionEntryRef {entry_ref:?} for this extraction capture"
                 ))
-            })?;
-            evidence.push(staging_evidence(&projection));
-            source_refs.push(source_evidence_ref(&projection));
+            })?);
         }
+        if matches!(params.kind, CandidateKind::Preference)
+            && entries.iter().any(|entry| {
+                !matches!(
+                    entry.origin,
+                    crate::WorkerHistoryProvenance::HumanInput { .. }
+                )
+            })
+        {
+            return Err(ToolError::InvalidArgument(
+                "preference candidates require exclusively HumanInput evidence; model, Worker, Flow, backend, derived, and legacy-unknown origins are not preference authority"
+                    .to_string(),
+            ));
+        }
+        let evidence = entries.iter().map(staging_evidence).collect();
+        let source_refs = entries.iter().map(source_evidence_ref).collect();
         let candidate = ExtractedCandidate {
             kind: params.kind,
             claim: params.claim,
@@ -310,11 +324,65 @@ fn evidence_kind(entry: &SessionEntryEvidence) -> EvidenceKind {
     }
 }
 
+fn evidence_origin(origin: &crate::WorkerHistoryProvenance) -> EvidenceOrigin {
+    use crate::WorkerHistoryProvenance as Origin;
+    let mut evidence = EvidenceOrigin {
+        kind: EvidenceOriginKind::LegacyUnknown,
+        account_id: None,
+        workspace_id: None,
+        runtime_id: None,
+        worker_id: None,
+        flow_selector: None,
+        flow_definition_id: None,
+        flow_definition_revision: None,
+    };
+    match origin {
+        Origin::HumanInput { account_id } => {
+            evidence.kind = EvidenceOriginKind::HumanInput;
+            evidence.account_id = Some(account_id.clone());
+        }
+        Origin::WorkerInput { actor } => {
+            evidence.kind = EvidenceOriginKind::WorkerInput;
+            evidence.workspace_id = actor.workspace_id.clone();
+            evidence.runtime_id = actor.runtime_id.clone();
+            evidence.worker_id = Some(actor.worker_id.clone());
+        }
+        Origin::FlowInstruction {
+            selector,
+            definition_id,
+            definition_revision,
+            ..
+        } => {
+            evidence.kind = EvidenceOriginKind::FlowInstruction;
+            evidence.flow_selector = Some(selector.clone());
+            evidence.flow_definition_id = Some(definition_id.clone());
+            evidence.flow_definition_revision = Some(*definition_revision);
+        }
+        Origin::BackendInstruction { .. } => evidence.kind = EvidenceOriginKind::BackendInstruction,
+        Origin::ModelOutput { worker } => {
+            evidence.kind = EvidenceOriginKind::ModelOutput;
+            evidence.workspace_id = worker.workspace_id.clone();
+            evidence.runtime_id = worker.runtime_id.clone();
+            evidence.worker_id = Some(worker.worker_id.clone());
+        }
+        Origin::ToolOutput { worker } => {
+            evidence.kind = EvidenceOriginKind::ToolOutput;
+            evidence.workspace_id = worker.workspace_id.clone();
+            evidence.runtime_id = worker.runtime_id.clone();
+            evidence.worker_id = Some(worker.worker_id.clone());
+        }
+        Origin::DerivedSummary => evidence.kind = EvidenceOriginKind::DerivedSummary,
+        Origin::LegacyUnknown => evidence.kind = EvidenceOriginKind::LegacyUnknown,
+    }
+    evidence
+}
+
 fn staging_evidence(entry: &SessionEntryEvidence) -> StagingEvidence {
     StagingEvidence {
         id: entry.entry_ref.to_string(),
         kind: evidence_kind(entry),
         entry_range: Some(entry.entry_range),
+        origin: Some(evidence_origin(&entry.origin)),
         excerpt: Some(entry.excerpt.clone()),
         summary: Some(entry.summary.clone()),
     }
@@ -325,6 +393,7 @@ fn source_evidence_ref(entry: &SessionEntryEvidence) -> SourceEvidenceRef {
         segment_id: Some(entry.segment_id.clone()),
         entry_range: Some(entry.entry_range),
         evidence_id: Some(entry.entry_ref.to_string()),
+        origin: Some(evidence_origin(&entry.origin)),
         evidence_kind: Some(evidence_kind(entry)),
         label: Some(entry.label.clone()),
         summary: Some(entry.summary.clone()),
@@ -432,6 +501,15 @@ mod tests {
         assert!(input.contains("StageMemoryCandidate.entry_refs"));
     }
 
+    #[test]
+    fn human_origin_projects_account_authority_into_evidence() {
+        let origin = evidence_origin(&crate::WorkerHistoryProvenance::HumanInput {
+            account_id: "account-1".into(),
+        });
+        assert_eq!(origin.kind, EvidenceOriginKind::HumanInput);
+        assert_eq!(origin.account_id.as_deref(), Some("account-1"));
+    }
+
     #[test]
     fn backend_input_failures_remain_invalid_argument_tool_errors() {
         let backend = map_memory_stage_error(WorkspaceMemoryBackendError::Backend(
@@ -445,6 +523,19 @@ mod tests {
         assert!(matches!(http, ToolError::InvalidArgument(_)));
     }
 
+    #[tokio::test]
+    async fn preference_rejects_legacy_unknown_before_backend_mutation() {
+        let tool = StageMemoryCandidateTool { state: state() };
+        let error = tool
+            .execute(
+                r#"{"kind":"preference","claim":"claim","why_useful":"useful","entry_refs":["E00000000"]}"#,
+                agen::tool::ToolExecutionContext::direct(),
+            )
+            .await
+            .unwrap_err();
+        assert!(format!("{error:?}").contains("exclusively HumanInput evidence"));
+    }
+
     #[tokio::test]
     async fn stage_rejects_entry_ref_outside_capture_before_backend_mutation() {
         let tool = StageMemoryCandidateTool { state: state() };

crates/worker/src/feature/builtin/objective.rs

@@ -209,9 +209,7 @@ impl WorkspaceHttpObjectiveBackend {
             .and_then(serde_json::Value::as_str)
             .filter(|key| is_canonical_resource_key(key, "T-"))
             .map(ToOwned::to_owned)
-            .ok_or_else(|| {
-                ToolError::ExecutionFailed("required T- human key is unavailable".to_string())
-            })
+            .ok_or_else(|| ToolError::ExecutionFailed("required T- key is unavailable".to_string()))
     }
 
     fn objective_url(&self, id: &str) -> String {
@@ -295,7 +293,7 @@ fn is_canonical_resource_key(resource_key: &str, prefix: &str) -> bool {
 fn objective_output(summary: String, response: ObjectiveDetail) -> Result<ToolOutput, ToolError> {
     if !is_canonical_resource_key(&response.resource_key, "O-") {
         return Err(ToolError::ExecutionFailed(
-            "required O- human key is unavailable".to_string(),
+            "required O- key is unavailable".to_string(),
         ));
     }
     let projected = serde_json::json!({
@@ -712,7 +710,7 @@ mod tests {
     }
 
     #[tokio::test(flavor = "multi_thread")]
-    async fn objective_show_summary_uses_projected_human_key() {
+    async fn objective_show_summary_uses_projected_key() {
         let listener = TcpListener::bind("127.0.0.1:0").unwrap();
         let base_url = format!("http://{}", listener.local_addr().unwrap());
         let server = thread::spawn(move || {
@@ -764,7 +762,7 @@ mod tests {
     }
 
     #[tokio::test(flavor = "multi_thread")]
-    async fn objective_link_summaries_resolve_internal_ticket_ids_to_human_keys() {
+    async fn objective_link_summaries_resolve_internal_ticket_ids_to_keys() {
         let listener = TcpListener::bind("127.0.0.1:0").unwrap();
         let base_url = format!("http://{}", listener.local_addr().unwrap());
         let server = thread::spawn(move || {
@@ -833,7 +831,7 @@ mod tests {
     }
 
     #[test]
-    fn objective_output_rejects_noncanonical_human_keys() {
+    fn objective_output_rejects_noncanonical_keys() {
         let response = ObjectiveDetail {
             resource_key: "O-internal".to_string(),
             title: "Objective".to_string(),

crates/worker/src/feature/builtin/resource_projection.rs

@@ -228,7 +228,7 @@ pub(super) fn project_ticket_query(value: Value) -> Result<ModelTicketQueryRespo
 fn project_ticket_query_item(value: &Value) -> Result<ModelTicketQueryItem, String> {
     let item = object(value, "Ticket query item")?;
     Ok(ModelTicketQueryItem {
-        ticket: human_ref(item, "resource_key", "T-")?,
+        ticket: resource_ref(item, "resource_key", "T-")?,
         title: string_field(item, "title")?,
         state: string_field(item, "state")?,
         readiness: optional_string(item, "readiness")?,
@@ -245,7 +245,7 @@ fn project_ticket_query_item(value: &Value) -> Result<ModelTicketQueryItem, Stri
             .transpose()?,
         linked_objectives: string_array(item, "linked_objective_keys")?
             .into_iter()
-            .map(|key| validate_human_ref(key, "O-"))
+            .map(|key| validate_resource_ref(key, "O-"))
             .collect::<Result<Vec<_>, _>>()?,
         relation_count: usize_field(item, "relation_count")?,
         blocker_count: usize_field(item, "blocker_count")?,
@@ -273,7 +273,7 @@ pub(super) fn project_ticket_detail(value: Value) -> Result<ModelTicketDetail, S
         .collect::<Result<Vec<_>, _>>()?;
 
     Ok(ModelTicketDetail {
-        ticket: human_ref(root, "resource_key", "T-")?,
+        ticket: resource_ref(root, "resource_key", "T-")?,
         title: string_field(root, "title")?,
         body: string_field(root, "body")?,
         state: string_field(root, "state")?,
@@ -332,10 +332,10 @@ fn project_objective_query_item(value: &Value) -> Result<ModelObjectiveQueryItem
     let item = object(value, "Objective query item")?;
     let linked_tickets = string_array(item, "linked_ticket_keys")?
         .into_iter()
-        .map(|key| validate_human_ref(key, "T-"))
+        .map(|key| validate_resource_ref(key, "T-"))
         .collect::<Result<Vec<_>, _>>()?;
     Ok(ModelObjectiveQueryItem {
-        objective: human_ref(item, "resource_key", "O-")?,
+        objective: resource_ref(item, "resource_key", "O-")?,
         title: string_field(item, "title")?,
         summary: optional_string(item, "snippet")?,
         state: string_field(item, "state")?,
@@ -349,7 +349,7 @@ fn project_objective_query_item(value: &Value) -> Result<ModelObjectiveQueryItem
 pub(super) fn project_objective_detail(value: Value) -> Result<ModelObjectiveDetail, String> {
     let root = object(&value, "Objective detail response")?;
     Ok(ModelObjectiveDetail {
-        objective: human_ref(root, "resource_key", "O-")?,
+        objective: resource_ref(root, "resource_key", "O-")?,
         title: string_field(root, "title")?,
         body: string_field(root, "body")?,
         state: string_field(root, "state")?,
@@ -373,7 +373,7 @@ pub(super) fn project_objective_detail(value: Value) -> Result<ModelObjectiveDet
 fn project_worker(value: &Value) -> Result<ModelWorkerSummary, String> {
     let worker = object(value, "Worker summary")?;
     Ok(ModelWorkerSummary {
-        worker: human_ref(worker, "worker_resource_key", "W-")?,
+        worker: resource_ref(worker, "worker_resource_key", "W-")?,
     })
 }
 
@@ -439,7 +439,7 @@ fn project_relation(
         None => optional_string(relation, "at")?,
     };
     Ok(ModelRelation {
-        ticket: human_ref(relation, ticket_key, "T-")?,
+        ticket: resource_ref(relation, ticket_key, "T-")?,
         kind,
         note,
         created_at,
@@ -449,7 +449,7 @@ fn project_relation(
 fn project_blocker(value: &Value) -> Result<ModelBlocker, String> {
     let blocker = object(value, "Ticket blocker")?;
     Ok(ModelBlocker {
-        ticket: human_ref(blocker, "blocking_resource_key", "T-")?,
+        ticket: resource_ref(blocker, "blocking_resource_key", "T-")?,
         kind: string_field(blocker, "relation_kind")?,
         state: optional_string(blocker, "blocking_state")?,
         resolved: bool_field(blocker, "resolved")?,
@@ -466,7 +466,7 @@ fn project_notice(value: &Value) -> Result<ModelNotice, String> {
 fn project_objective_summary(value: &Value) -> Result<ModelObjectiveSummary, String> {
     let summary = object(value, "Objective summary")?;
     Ok(ModelObjectiveSummary {
-        objective: human_ref(summary, "resource_key", "O-")?,
+        objective: resource_ref(summary, "resource_key", "O-")?,
         title: string_field(summary, "title")?,
         state: string_field(summary, "state")?,
     })
@@ -475,7 +475,7 @@ fn project_objective_summary(value: &Value) -> Result<ModelObjectiveSummary, Str
 fn project_ticket_summary(value: &Value) -> Result<ModelTicketSummary, String> {
     let summary = object(value, "Ticket summary")?;
     Ok(ModelTicketSummary {
-        ticket: human_ref(summary, "resource_key", "T-")?,
+        ticket: resource_ref(summary, "resource_key", "T-")?,
         title: string_field(summary, "title")?,
         state: string_field(summary, "state")?,
     })
@@ -491,9 +491,7 @@ fn project_assignment(
     let principal = match kind.as_str() {
         "worker" => current_coder
             .map(|coder| coder.worker.clone())
-            .ok_or_else(|| {
-                "Worker assignment is missing a Workspace human key projection".to_string()
-            })?,
+            .ok_or_else(|| "Worker assignment is missing a Workspace key projection".to_string())?,
         "workspace_agent" => format!("workspace-agent:{}", string_field(principal, "agent_key")?),
         "user" => "user".to_string(),
         other => format!("source:{other}"),
@@ -645,23 +643,23 @@ fn string_array(object: &Map<String, Value>, key: &str) -> Result<Vec<String>, S
         .collect()
 }
 
-fn human_ref(object: &Map<String, Value>, key: &str, prefix: &str) -> Result<String, String> {
+fn resource_ref(object: &Map<String, Value>, key: &str, prefix: &str) -> Result<String, String> {
     let value = object
         .get(key)
         .and_then(Value::as_str)
         .map(ToOwned::to_owned)
-        .ok_or_else(|| format!("required {prefix} human key is unavailable"))?;
-    validate_human_ref(value, prefix)
+        .ok_or_else(|| format!("required {prefix} key is unavailable"))?;
+    validate_resource_ref(value, prefix)
 }
 
-fn validate_human_ref(value: String, prefix: &str) -> Result<String, String> {
+fn validate_resource_ref(value: String, prefix: &str) -> Result<String, String> {
     let valid = value.strip_prefix(prefix).is_some_and(|sequence| {
         !sequence.is_empty() && sequence.bytes().all(|byte| byte.is_ascii_digit())
     });
     if valid {
         Ok(value)
     } else {
-        Err(format!("required {prefix} human key is unavailable"))
+        Err(format!("required {prefix} key is unavailable"))
     }
 }
 
@@ -671,7 +669,7 @@ mod tests {
     use serde_json::json;
 
     #[test]
-    fn objective_projection_exposes_only_human_resource_references() {
+    fn objective_projection_exposes_only_resource_references() {
         let projected = project_objective_detail(json!({
             "id": "00001M10HW6BV",
             "resource_key": "O-543",
@@ -779,9 +777,9 @@ mod tests {
     }
 
     #[test]
-    fn human_resource_projection_rejects_noncanonical_keys() {
+    fn resource_projection_rejects_noncanonical_keys() {
         for (key, prefix) in [("T-key", "T-"), ("O-", "O-"), ("W-1x", "W-")] {
-            assert!(validate_human_ref(key.to_string(), prefix).is_err());
+            assert!(validate_resource_ref(key.to_string(), prefix).is_err());
         }
     }
 

crates/worker/src/feature/builtin/session_explore.rs

@@ -193,6 +193,7 @@ impl Tool for ShowOverviewTool {
             .map(|entry| {
                 serde_json::json!({
                     "entry_ref": entry.id,
+                    "origin": entry.origin,
                     "entry_range": entry.entry_range,
                     "kind": entry.kind.as_str(),
                     "label": entry.label,
@@ -234,15 +235,16 @@ impl Tool for SearchEntriesTool {
             .transpose()?;
         let from = params.from.as_deref().map(parse_entry_ref).transpose()?;
         let through = params.through.as_deref().map(parse_entry_ref).transpose()?;
+        let view = self.state.view();
         if let (Some(from), Some(through)) = (&from, &through) {
-            if from.source_index() > through.source_index() {
+            if view.source_index_for_ref(from) > view.source_index_for_ref(through) {
                 return Err(ToolError::InvalidArgument(
                     "SearchEntries from must not be after through".to_string(),
                 ));
             }
         }
         let limit = bounded_limit(params.limit, DEFAULT_PAGE_LIMIT, MAX_PAGE_LIMIT);
-        let hits = self.state.view().search(&SearchOptions {
+        let hits = view.search(&SearchOptions {
             query: params.query,
             kind,
             tool_part,
@@ -318,6 +320,7 @@ impl Tool for ReadEntryTool {
             .map(|entry| {
                 serde_json::json!({
                     "entry_ref": entry.id,
+                    "origin": entry.origin,
                     "entry_range": entry.entry_range,
                     "kind": entry.kind.as_str(),
                     "tool_part": entry.tool_part.map(|part| format!("{part:?}").to_lowercase()),

crates/worker/src/feature/builtin/ticket.rs

@@ -306,9 +306,11 @@ struct BackendTicketService {
     backend: TicketToolBackend,
 }
 
-impl TicketService for BackendTicketService {
-    fn ticket_handoff(&self, ticket_ref: &str) -> Result<TicketHandoff, TicketError> {
-        let ticket = self.backend.show(ticket_ref.into())?;
+struct WorkspaceTicketService {
+    backend: WorkspaceHttpTicketBackend,
+}
+
+fn ticket_handoff_from_record(ticket: Ticket) -> Result<TicketHandoff, TicketError> {
     let resource_key = ticket
         .meta
         .resource_key
@@ -320,6 +322,17 @@ impl TicketService for BackendTicketService {
         workflow_state: ticket.meta.workflow_state,
     })
 }
+
+impl TicketService for BackendTicketService {
+    fn ticket_handoff(&self, ticket_ref: &str) -> Result<TicketHandoff, TicketError> {
+        ticket_handoff_from_record(self.backend.show(ticket_ref.into())?)
+    }
+}
+
+impl TicketService for WorkspaceTicketService {
+    fn ticket_handoff(&self, ticket_ref: &str) -> Result<TicketHandoff, TicketError> {
+        ticket_handoff_from_record(self.backend.show_unprojected(ticket_ref)?)
+    }
 }
 
 fn ticket_workflow_instruction() -> FeatureInstructionDeclaration {
@@ -640,9 +653,14 @@ impl FeatureModule for TicketFeature {
         let Some(backend) = self.tool_backend(context) else {
             return Ok(());
         };
-        let ticket_service: Arc<dyn TicketService> = Arc::new(BackendTicketService {
+        let ticket_service: Arc<dyn TicketService> = match &self.backend {
+            TicketFeatureBackend::WorkspaceClient(client) => Arc::new(WorkspaceTicketService {
+                backend: WorkspaceHttpTicketBackend::new(client.clone()),
+            }),
+            TicketFeatureBackend::Local { .. } => Arc::new(BackendTicketService {
                 backend: backend.clone(),
-        });
+            }),
+        };
         context.services().provide(
             ServiceDeclaration::new(
                 ServiceId::builtin(TICKET_SERVICE_ID),
@@ -714,6 +732,26 @@ impl WorkspaceHttpTicketBackend {
         Self::invoke_client(client, workspace_id, operation)
     }
 
+    fn show_unprojected(&self, ticket_ref: &str) -> TicketResult<Ticket> {
+        let client = self.client.clone();
+        let workspace_id = self.client.workspace_id().unwrap_or_default().to_string();
+        let ticket_path = Self::ticket_path(&TicketIdOrSlug::from(ticket_ref));
+        let request = move || {
+            Self::request_unprojected(
+                client,
+                WorkspaceRequestMethod::Get,
+                format!("/api/w/{workspace_id}/tickets/{ticket_path}/record"),
+                None,
+            )
+        };
+        if tokio::runtime::Handle::try_current().is_ok() {
+            return std::thread::spawn(request).join().map_err(|_| {
+                TicketError::Conflict("ticket REST request thread panicked".to_string())
+            })?;
+        }
+        request()
+    }
+
     fn ticket_path(id: &TicketIdOrSlug) -> String {
         let value = match id {
             TicketIdOrSlug::Id(value)
@@ -738,6 +776,29 @@ impl WorkspaceHttpTicketBackend {
         endpoint: String,
         body: Option<serde_json::Value>,
     ) -> TicketResult<T> {
+        let mut value = Self::request_value(client, method, endpoint, body)?;
+        Self::canonicalize_ticket_references(&mut value);
+        serde_json::from_value(value)
+            .map_err(|error| TicketError::Conflict(format!("decode ticket REST response: {error}")))
+    }
+
+    fn request_unprojected<T: serde::de::DeserializeOwned>(
+        client: Arc<dyn WorkspaceClient>,
+        method: WorkspaceRequestMethod,
+        endpoint: String,
+        body: Option<serde_json::Value>,
+    ) -> TicketResult<T> {
+        let value = Self::request_value(client, method, endpoint, body)?;
+        serde_json::from_value(value)
+            .map_err(|error| TicketError::Conflict(format!("decode ticket REST response: {error}")))
+    }
+
+    fn request_value(
+        client: Arc<dyn WorkspaceClient>,
+        method: WorkspaceRequestMethod,
+        endpoint: String,
+        body: Option<serde_json::Value>,
+    ) -> TicketResult<Value> {
         let request = match body {
             Some(body) => WorkspaceRequest::json(method, endpoint, body.to_string()),
             None if method == WorkspaceRequestMethod::Get => WorkspaceRequest::get(endpoint),
@@ -756,11 +817,7 @@ impl WorkspaceHttpTicketBackend {
                 response.status
             )));
         }
-        let mut value: Value = serde_json::from_str(&response.body).map_err(|error| {
-            TicketError::Conflict(format!("decode ticket REST response: {error}"))
-        })?;
-        Self::canonicalize_ticket_references(&mut value);
-        serde_json::from_value(value)
+        serde_json::from_str(&response.body)
             .map_err(|error| TicketError::Conflict(format!("decode ticket REST response: {error}")))
     }
 
@@ -810,9 +867,7 @@ impl WorkspaceHttpTicketBackend {
             .and_then(Value::as_str)
             .filter(|key| is_canonical_ticket_resource_key(key))
             .map(ToOwned::to_owned)
-            .ok_or_else(|| {
-                TicketError::Conflict("required Ticket human key is unavailable".to_string())
-            })
+            .ok_or_else(|| TicketError::Conflict("required Ticket key is unavailable".to_string()))
     }
 
     fn request_unit(
@@ -889,7 +944,7 @@ impl WorkspaceHttpTicketBackend {
                     .is_some_and(is_canonical_ticket_resource_key)
                 {
                     return Err(TicketError::Conflict(
-                        "required Ticket human key is unavailable".to_string(),
+                        "required Ticket key is unavailable".to_string(),
                     ));
                 }
                 Ok(TicketBackendOperationResult::Ticket(ticket))
@@ -1861,7 +1916,7 @@ provider = "github"
     }
 
     #[test]
-    fn workspace_http_backend_records_relation_with_authoritative_human_keys() {
+    fn workspace_http_backend_records_relation_with_authoritative_keys() {
         let listener = TcpListener::bind("127.0.0.1:0").unwrap();
         let addr = listener.local_addr().unwrap();
         let server = thread::spawn(move || {
@@ -2000,6 +2055,46 @@ provider = "github"
         assert_eq!(removed.target, "T-2");
     }
 
+    #[test]
+    fn workspace_ticket_service_preserves_internal_identity_for_handoff() {
+        let temp = TempDir::new().unwrap();
+        let local = LocalTicketBackend::new(temp.path().join("tickets"));
+        let created = local.create(NewTicket::new("Ticket handoff")).unwrap();
+        let mut ticket = local.show(TicketIdOrSlug::Id(created.id.clone())).unwrap();
+        ticket.meta.resource_key = Some("T-548".to_string());
+        ticket.meta.workflow_state = TicketWorkflowState::Queued;
+
+        let listener = TcpListener::bind("127.0.0.1:0").unwrap();
+        let base_url = format!("http://{}", listener.local_addr().unwrap());
+        let response_body = serde_json::to_string(&ticket).unwrap();
+        let server = thread::spawn(move || {
+            let (mut stream, _) = listener.accept().unwrap();
+            let mut buffer = [0_u8; 8192];
+            let len = stream.read(&mut buffer).unwrap();
+            let request = String::from_utf8_lossy(&buffer[..len]);
+            assert!(request.starts_with("GET /api/w/workspace-a/tickets/T-548/record HTTP/1.1"));
+            write!(
+                stream,
+                "HTTP/1.1 200 OK\r\nContent-Type: application/json\r\nContent-Length: {}\r\n\r\n{}",
+                response_body.len(),
+                response_body
+            )
+            .unwrap();
+        });
+
+        let service = WorkspaceTicketService {
+            backend: WorkspaceHttpTicketBackend::new(Arc::new(
+                crate::worker::TestWorkspaceHttpClient::new("workspace-a", base_url),
+            )),
+        };
+        let handoff = service.ticket_handoff("T-548").unwrap();
+
+        server.join().unwrap();
+        assert_eq!(handoff.id, created.id);
+        assert_eq!(handoff.resource_key, "T-548");
+        assert_eq!(handoff.workflow_state, TicketWorkflowState::Queued);
+    }
+
     #[test]
     fn ticket_handoff_accepts_only_canonical_ticket_resource_keys() {
         assert!(is_canonical_ticket_resource_key("T-482"));

crates/worker/src/feature/builtin/worker_observation.rs

@@ -1,11 +1,12 @@
 use std::sync::Arc;
 
+#[cfg(test)]
 use agen::Item;
 use agen::tool::{Tool, ToolDefinition, ToolError, ToolMeta, ToolOutput};
 use async_trait::async_trait;
 use schemars::JsonSchema;
 use serde::{Deserialize, Serialize};
-use session_store::collect_state;
+use session_store::{LogEntry, collect_state};
 
 use super::manage_worker::{WORKER_CONTROL_SERVICE_ID, WorkerControlService};
 use crate::feature::{
@@ -60,7 +61,27 @@ pub struct WorkerObservationSubject {
 #[derive(Debug, Clone)]
 pub struct WorkerSessionCapture {
     pub segment_id: String,
-    pub items: Vec<Item>,
+    pub entries: Vec<agen::HistoryEntry<crate::SessionHistoryMetadata>>,
+}
+
+impl WorkerSessionCapture {
+    pub fn from_log_entries(
+        segment_id: impl Into<String>,
+        log_entries: &[LogEntry],
+    ) -> Result<Self, String> {
+        let segment_id = segment_id.into();
+        let state = collect_state(log_entries);
+        let parsed_segment_id = segment_id.parse().unwrap_or_default();
+        let entries = crate::session_history::restore_history_entries(
+            state.session_id.unwrap_or_default(),
+            parsed_segment_id,
+            log_entries,
+        )?;
+        Ok(Self {
+            segment_id,
+            entries,
+        })
+    }
 }
 
 #[derive(Debug, thiserror::Error)]
@@ -161,9 +182,17 @@ impl WorkerObservationProvider for WorkspaceClientWorkerObservationProvider {
             })
             .collect::<Result<Vec<session_store::LogEntry>, _>>()?;
         let state = collect_state(&entries);
+        let segment_id = response.segment_id;
+        let parsed_segment_id = segment_id.parse().unwrap_or_default();
+        let typed_entries = crate::session_history::restore_history_entries(
+            state.session_id.unwrap_or_default(),
+            parsed_segment_id,
+            &entries,
+        )
+        .map_err(WorkerObservationError::Unavailable)?;
         Ok(WorkerSessionCapture {
-            segment_id: response.segment_id,
-            items: state.history,
+            segment_id,
+            entries: typed_entries,
         })
     }
 }
@@ -392,9 +421,15 @@ impl WorkerObservationProvider for SpawnedSubWorkerObservationProvider {
             .ok_or(WorkerObservationError::NotFound)?;
         let entries = record.session.entries();
         let state = collect_state(&entries);
+        let typed_entries = crate::session_history::restore_history_entries(
+            state.session_id.unwrap_or_default(),
+            Default::default(),
+            &entries,
+        )
+        .map_err(WorkerObservationError::Unavailable)?;
         Ok(WorkerSessionCapture {
             segment_id: format!("subworker:{name}"),
-            items: state.history,
+            entries: typed_entries,
         })
     }
 }
@@ -508,6 +543,7 @@ impl Tool for ViewSessionOverviewTool {
             .map(|entry| {
                 serde_json::json!({
                     "entry_ref": entry.id,
+                    "origin": entry.origin,
                     "entry_range": entry.entry_range,
                     "kind": entry.kind.as_str(),
                     "label": entry.label,
@@ -547,7 +583,7 @@ impl Tool for SearchSessionEntriesTool {
         let from = params.from.as_deref().map(parse_entry_ref).transpose()?;
         let through = params.through.as_deref().map(parse_entry_ref).transpose()?;
         if let (Some(from), Some(through)) = (&from, &through) {
-            if from.source_index() > through.source_index() {
+            if view.source_index_for_ref(from) > view.source_index_for_ref(through) {
                 return Err(ToolError::InvalidArgument(
                     "SearchSessionEntries from must not be after through".to_string(),
                 ));
@@ -573,6 +609,7 @@ impl Tool for SearchSessionEntriesTool {
             .map(|entry| {
                 serde_json::json!({
                     "entry_ref": entry.id,
+                    "origin": entry.origin,
                     "entry_range": entry.entry_range,
                     "kind": entry.kind.as_str(),
                     "tool_part": entry.tool_part.map(|part| format!("{part:?}").to_lowercase()),
@@ -628,6 +665,7 @@ impl Tool for ReadSessionEntryTool {
             .map(|entry| {
                 serde_json::json!({
                     "entry_ref": entry.id,
+                    "origin": entry.origin,
                     "entry_range": entry.entry_range,
                     "kind": entry.kind.as_str(),
                     "tool_part": entry.tool_part.map(|part| format!("{part:?}").to_lowercase()),
@@ -661,7 +699,10 @@ async fn latest_view(
         .capture_worker_session(subject)
         .await
         .map_err(tool_error)?;
-    Ok(SessionCapture::new(capture.segment_id, capture.items))
+    Ok(SessionCapture::from_history_entries(
+        capture.segment_id,
+        capture.entries,
+    ))
 }
 
 fn parse_input<T: serde::de::DeserializeOwned>(
@@ -751,9 +792,23 @@ mod tests {
             if subject != &granted_subject() {
                 return Err(WorkerObservationError::NotFound);
             }
+            let entries = self
+                .captures
+                .lock()
+                .unwrap()
+                .clone()
+                .into_iter()
+                .enumerate()
+                .map(|(index, item)| {
+                    let mut metadata = crate::SessionHistoryMetadata::legacy_unknown();
+                    metadata.entry_id =
+                        session_store::LoggedSessionHistoryEntryId(format!("fake-{index:08}"));
+                    agen::HistoryEntry::new(item, metadata)
+                })
+                .collect();
             Ok(WorkerSessionCapture {
                 segment_id: "segment".to_string(),
-                items: self.captures.lock().unwrap().clone(),
+                entries,
             })
         }
     }
@@ -796,7 +851,7 @@ mod tests {
         let read = read_definition(provider.clone())().1;
         let hidden = read
             .execute(
-                r#"{"subject":{"kind":"runtime_worker","runtime_id":"runtime-1","worker_id":"unauthorized"},"entry_ref":"E00000000"}"#,
+                r#"{"subject":{"kind":"runtime_worker","runtime_id":"runtime-1","worker_id":"unauthorized"},"entry_ref":"Efake-00000000"}"#,
                 agen::tool::ToolExecutionContext::direct(),
             )
             .await
@@ -810,7 +865,7 @@ mod tests {
             .push(message("a1", Role::Assistant, "second"));
         let output = read
             .execute(
-                r#"{"subject":{"kind":"runtime_worker","runtime_id":"runtime-1","worker_id":"granted"},"entry_ref":"E00000000"}"#,
+                r#"{"subject":{"kind":"runtime_worker","runtime_id":"runtime-1","worker_id":"granted"},"entry_ref":"Efake-00000000"}"#,
                 agen::tool::ToolExecutionContext::direct(),
             )
             .await
@@ -819,7 +874,7 @@ mod tests {
 
         let output = read
             .execute(
-                r#"{"subject":{"kind":"runtime_worker","runtime_id":"runtime-1","worker_id":"granted"},"entry_ref":"E00000001"}"#,
+                r#"{"subject":{"kind":"runtime_worker","runtime_id":"runtime-1","worker_id":"granted"},"entry_ref":"Efake-00000001"}"#,
                 agen::tool::ToolExecutionContext::direct(),
             )
             .await

crates/worker/src/internal_worker.rs

@@ -55,7 +55,17 @@ pub(crate) struct InternalWorkerSpec {
     pub input: String,
     pub cache_key: Option<String>,
     pub max_turns: Option<u32>,
-    pub engine_configurator: Option<Box<dyn FnOnce(&mut Engine<Box<dyn LlmClient>>) + Send>>,
+    pub engine_configurator: Option<
+        Box<
+            dyn FnOnce(
+                    &mut Engine<
+                        Box<dyn LlmClient>,
+                        agen::state::Mutable,
+                        crate::SessionHistoryMetadata,
+                    >,
+                ) + Send,
+        >,
+    >,
     pub features: FeatureRegistryBuilder,
     pub required_tools: &'static [&'static str],
     pub authority: InternalWorkerAuthority,
@@ -124,7 +134,9 @@ where
 
     let last_usage = Arc::new(Mutex::new(None::<UsageEvent>));
     let usage_slot = last_usage.clone();
-    let mut engine = Engine::new(client).system_prompt(system_prompt);
+    let mut engine =
+        Engine::<_, agen::state::Mutable, crate::SessionHistoryMetadata>::new_annotated(client)
+            .system_prompt(system_prompt);
     engine.on_usage(move |usage| {
         if let Ok(mut slot) = usage_slot.lock() {
             *slot = Some(usage.clone());
@@ -536,7 +548,9 @@ pub(crate) async fn spawn_internal_worker_session(
 
     let last_usage = Arc::new(Mutex::new(None::<UsageEvent>));
     let usage_slot = last_usage.clone();
-    let mut engine = Engine::new(client).system_prompt(system_prompt);
+    let mut engine =
+        Engine::<_, agen::state::Mutable, crate::SessionHistoryMetadata>::new_annotated(client)
+            .system_prompt(system_prompt);
     engine.on_usage(move |usage| {
         if let Ok(mut slot) = usage_slot.lock() {
             *slot = Some(usage.clone());
@@ -633,7 +647,9 @@ pub(crate) fn prepare_internal_worker_from_spec(
         manifest.compaction = None;
         manifest.memory = None;
 
-        let mut engine = Engine::new(client).system_prompt(system_prompt);
+        let mut engine =
+            Engine::<_, agen::state::Mutable, crate::SessionHistoryMetadata>::new_annotated(client)
+                .system_prompt(system_prompt);
         engine.set_cache_key(cache_key);
         engine.set_max_turns(max_turns);
         if let Some(configure) = engine_configurator {

crates/worker/src/interrupt_prep.rs

@@ -13,7 +13,7 @@
 
 #[cfg(test)]
 use crate::prompt::catalog::PromptCatalog;
-use agen::Item;
+use agen::{Item, ToolResultDisposition};
 
 /// Build synthetic `Item::ToolResult` items for every unanswered
 /// `Item::ToolCall` in `history`, preserving order.
@@ -28,7 +28,16 @@ pub(crate) fn orphan_tool_result_closures(history: &[Item], summary: &str) -> Ve
     for item in history {
         if let Item::ToolCall { call_id, .. } = item {
             if !answered.contains(call_id.as_str()) {
-                out.push(Item::tool_result(call_id.clone(), summary));
+                out.push(Item::tool_result_item_with_disposition_and_attachments(
+                    call_id.clone(),
+                    summary,
+                    Some(
+                        "Execution ended before completion could be confirmed. Completion and side effects are unknown."
+                            .to_string(),
+                    ),
+                    ToolResultDisposition::OutcomeUnknown,
+                    Vec::new(),
+                ));
             }
         }
     }
@@ -77,10 +86,12 @@ mod tests {
             Item::ToolResult {
                 call_id,
                 summary: got,
+                disposition,
                 ..
             } => {
                 assert_eq!(call_id, "c1");
                 assert_eq!(got, &summary);
+                assert_eq!(*disposition, ToolResultDisposition::OutcomeUnknown);
             }
             other => panic!("expected ToolResult, got {other:?}"),
         }

crates/worker/src/ipc/interceptor.rs

@@ -8,6 +8,7 @@
 //! decisions (continue / skip / abort / pause).
 
 use std::borrow::Cow;
+use std::collections::VecDeque;
 use std::sync::atomic::{AtomicUsize, Ordering};
 use std::sync::{Arc, Mutex};
 
@@ -33,7 +34,9 @@ use crate::hook::{
 };
 use crate::ipc::notify_buffer::{NotifyBuffer, build_system_item_with_provenance};
 use crate::prompt::catalog::PromptCatalog;
+use crate::session_history::SessionHistoryMetadata;
 use crate::worker::SystemItemCommitter;
+use agen::HistoryEntry;
 use agen::token_counter::total_tokens;
 
 /// Maximum number of bytes copied into `TurnEndInfo::final_text_preview`.
@@ -73,6 +76,7 @@ pub(crate) struct WorkerInterceptor {
     /// worker. `None` in tests / `Worker::new` paths where no writer is
     /// attached.
     log_writer: Option<Arc<dyn SystemItemCommitter>>,
+    pending_committed_history: Arc<Mutex<VecDeque<HistoryEntry<SessionHistoryMetadata>>>>,
     /// Next turn index assigned by `on_prompt_submit`.
     next_turn_index: AtomicUsize,
     /// Tool calls observed in the current turn (reset on each new prompt).
@@ -80,6 +84,7 @@ pub(crate) struct WorkerInterceptor {
 }
 
 impl WorkerInterceptor {
+    #[cfg(test)]
     pub(crate) fn new(
         registry: Arc<HookRegistry>,
         compact_state: Option<Arc<CompactState>>,
@@ -88,6 +93,28 @@ impl WorkerInterceptor {
         pending_attachments: Arc<Mutex<Vec<SystemItem>>>,
         prompts: Arc<ArcSwap<PromptCatalog>>,
         log_writer: Option<Arc<dyn SystemItemCommitter>>,
+    ) -> Self {
+        Self::new_with_history_queue(
+            registry,
+            compact_state,
+            usage_history,
+            pending_notifies,
+            pending_attachments,
+            prompts,
+            log_writer,
+            Arc::new(Mutex::new(VecDeque::new())),
+        )
+    }
+
+    pub(crate) fn new_with_history_queue(
+        registry: Arc<HookRegistry>,
+        compact_state: Option<Arc<CompactState>>,
+        usage_history: Option<Arc<Mutex<Vec<UsageRecord>>>>,
+        pending_notifies: NotifyBuffer,
+        pending_attachments: Arc<Mutex<Vec<SystemItem>>>,
+        prompts: Arc<ArcSwap<PromptCatalog>>,
+        log_writer: Option<Arc<dyn SystemItemCommitter>>,
+        pending_committed_history: Arc<Mutex<VecDeque<HistoryEntry<SessionHistoryMetadata>>>>,
     ) -> Self {
         Self {
             registry,
@@ -99,6 +126,7 @@ impl WorkerInterceptor {
             prompts,
             prompt_workspace_id: None,
             log_writer,
+            pending_committed_history,
             next_turn_index: AtomicUsize::new(0),
             tool_calls_this_turn: AtomicUsize::new(0),
         }
@@ -125,7 +153,11 @@ impl WorkerInterceptor {
             return Ok(());
         };
         for item in items {
-            writer.commit_system_item(item.clone())?;
+            let entry = writer.commit_system_item(item.clone())?;
+            self.pending_committed_history
+                .lock()
+                .expect("pending committed history poisoned")
+                .push_back(entry);
         }
         Ok(())
     }
@@ -507,7 +539,12 @@ mod tests {
             &self,
             entry: session_store::LogEntry,
         ) -> Result<(), session_store::StoreError> {
-            if let session_store::LogEntry::SystemItem { item, .. } = entry {
+            let item = match entry {
+                session_store::LogEntry::SystemItem { item, .. } => Some(item),
+                session_store::LogEntry::AnnotatedSystemItem { entry, .. } => Some(entry.item),
+                _ => None,
+            };
+            if let Some(item) = item {
                 self.committed
                     .lock()
                     .expect("committed system-item list poisoned")

crates/worker/src/ipc/protocol_session.rs

@@ -29,15 +29,21 @@ pub fn subscribe_worker_protocol_session(handle: &WorkerHandle) -> WorkerProtoco
 
 pub fn live_log_entry_event(entry: LogEntry) -> Option<Event> {
     match entry {
-        LogEntry::SegmentStart { .. } => {
+        entry @ (LogEntry::SegmentStart { .. } | LogEntry::AnnotatedSegmentStart { .. }) => {
             let value = serde_json::to_value(&entry).expect("LogEntry is Serialize");
             Some(Event::SegmentRotated { entry: value })
         }
-        LogEntry::UserInput { segments, .. } => Some(Event::UserMessage { segments }),
+        LogEntry::UserInput { segments, .. } | LogEntry::AnnotatedUserInput { segments, .. } => {
+            Some(Event::UserMessage { segments })
+        }
         LogEntry::SystemItem { item, .. } => {
             let value = serde_json::to_value(&item).expect("SystemItem is Serialize");
             Some(Event::SystemItem { item: value })
         }
+        LogEntry::AnnotatedSystemItem { entry, .. } => {
+            let value = serde_json::to_value(&entry.item).expect("SystemItem is Serialize");
+            Some(Event::SystemItem { item: value })
+        }
         LogEntry::Invoke { trigger, .. } => Some(Event::InvokeStart { kind: trigger }),
         other => {
             // `SegmentLogSink::is_live_relevant` keeps non-live-relevant

crates/worker/src/lib.rs

@@ -12,6 +12,7 @@ pub mod prompt;
 pub mod runtime;
 pub mod segment_log_sink;
 mod session_capture;
+mod session_history;
 pub mod shared_state;
 mod shutdown_after_idle;
 pub mod skill;
@@ -42,6 +43,10 @@ pub use prompt::system::{SystemPromptContext, SystemPromptError, SystemPromptTem
 pub use protocol::{ErrorCode, Event, Method, TurnResult, WorkerStatus};
 pub use runtime::dir::RuntimeDir;
 pub use segment_log_sink::SegmentLogSink;
+pub use session_history::{
+    SessionHistoryDerivation, SessionHistoryEntryId, SessionHistoryMetadata,
+    WorkerHistoryProvenance, WorkerSubjectSnapshot,
+};
 pub use shared_state::WorkerSharedState;
 pub use worker::{
     LocalWorkingDirectory, WORKER_INPUT_SUBMISSION_EXTENSION_DOMAIN, Worker, WorkerError,

crates/worker/src/permission.rs

@@ -34,7 +34,7 @@ impl PermissionHook {
     }
 }
 
-impl<C: LlmClient, St: Store> Worker<C, St> {
+impl<C: LlmClient + 'static, St: Store> Worker<C, St> {
     pub(crate) fn apply_permissions_from_manifest(&mut self) {
         let Some(permissions) = self.manifest().permissions.clone() else {
             return;

crates/worker/src/segment_log_sink.rs

@@ -121,8 +121,11 @@ impl SegmentLogSink {
         matches!(
             entry,
             LogEntry::SegmentStart { .. }
+                | LogEntry::AnnotatedSegmentStart { .. }
                 | LogEntry::UserInput { .. }
+                | LogEntry::AnnotatedUserInput { .. }
                 | LogEntry::SystemItem { .. }
+                | LogEntry::AnnotatedSystemItem { .. }
                 | LogEntry::Invoke { .. }
         )
     }

crates/worker/src/session_capture.rs

@@ -6,7 +6,8 @@
 
 use std::sync::Arc;
 
-use agen::{Item, Role};
+use crate::session_history::{SessionHistoryMetadata, WorkerHistoryProvenance};
+use agen::{HistoryEntry, Item, Role};
 use serde::{Deserialize, Serialize};
 
 const DEFAULT_SEARCH_LIMIT: usize = 20;
@@ -21,14 +22,21 @@ const OVERVIEW_ANCHOR_STRIDE: usize = 8;
 pub(crate) struct SessionEntryRef(String);
 
 impl SessionEntryRef {
-    pub(crate) fn new(source_index: usize) -> Self {
-        Self(format!("E{source_index:08}"))
+    pub(crate) fn from_history_entry_id(entry_id: &crate::SessionHistoryEntryId) -> Self {
+        Self(format!("E{}", entry_id.0))
     }
 
     pub(crate) fn parse(value: &str) -> Option<Self> {
-        let reference = Self(value.to_string());
-        reference.source_index()?;
-        Some(reference)
+        let suffix = value.strip_prefix('E')?;
+        if suffix.is_empty()
+            || suffix.len() > 64
+            || !suffix
+                .bytes()
+                .all(|byte| byte.is_ascii_alphanumeric() || byte == b'-')
+        {
+            return None;
+        }
+        Some(Self(value.to_string()))
     }
 
     pub(crate) fn as_str(&self) -> &str {
@@ -97,6 +105,7 @@ impl ToolPart {
 #[derive(Debug, Clone)]
 pub(crate) struct OverviewItem {
     pub id: SessionEntryRef,
+    pub origin: WorkerHistoryProvenance,
     pub entry_range: [u64; 2],
     pub kind: ReferenceKind,
     pub label: String,
@@ -107,6 +116,7 @@ pub(crate) struct OverviewItem {
 #[derive(Debug, Clone)]
 pub(crate) struct ReferenceEntry {
     pub id: SessionEntryRef,
+    pub origin: WorkerHistoryProvenance,
     pub entry_range: [u64; 2],
     pub kind: ReferenceKind,
     pub tool_part: Option<ToolPart>,
@@ -132,6 +142,7 @@ pub(crate) struct SearchOptions {
 #[derive(Debug, Clone)]
 pub(crate) struct SearchHit {
     pub id: SessionEntryRef,
+    pub origin: WorkerHistoryProvenance,
     pub kind: ReferenceKind,
     pub tool_part: Option<ToolPart>,
     pub tool_name: Option<String>,
@@ -177,6 +188,7 @@ impl Default for ReadOptions {
 #[derive(Debug, Clone)]
 pub(crate) struct ReadEntry {
     pub id: SessionEntryRef,
+    pub origin: WorkerHistoryProvenance,
     pub kind: ReferenceKind,
     pub tool_part: Option<ToolPart>,
     pub tool_name: Option<String>,
@@ -195,6 +207,7 @@ pub(crate) struct ReadResult {
 pub(crate) struct SessionEntryEvidence {
     pub segment_id: String,
     pub entry_ref: SessionEntryRef,
+    pub origin: WorkerHistoryProvenance,
     pub entry_range: [u64; 2],
     pub kind: ReferenceKind,
     pub tool_part: Option<ToolPart>,
@@ -206,26 +219,42 @@ pub(crate) struct SessionEntryEvidence {
 #[derive(Debug, Clone)]
 pub(crate) struct SessionCapture {
     segment_id: String,
-    items: Arc<Vec<Item>>,
+    entries: Arc<Vec<HistoryEntry<SessionHistoryMetadata>>>,
     overview: Vec<OverviewItem>,
     index: Vec<ReferenceEntry>,
 }
 
 impl SessionCapture {
     pub(crate) fn new(segment_id: impl Into<String>, items: Vec<Item>) -> Self {
+        let entries = items
+            .into_iter()
+            .enumerate()
+            .map(|(index, item)| {
+                let mut metadata = SessionHistoryMetadata::legacy_unknown();
+                metadata.entry_id =
+                    session_store::LoggedSessionHistoryEntryId(format!("{index:08}"));
+                HistoryEntry::new(item, metadata)
+            })
+            .collect();
+        Self::from_history_entries(segment_id, entries)
+    }
+
+    pub(crate) fn from_history_entries(
+        segment_id: impl Into<String>,
+        entries: Vec<HistoryEntry<SessionHistoryMetadata>>,
+    ) -> Self {
         let segment_id = segment_id.into();
-        let items = Arc::new(items);
+        let entries = Arc::new(entries);
         let mut overview = Vec::new();
         let mut index = Vec::new();
 
-        for (idx, item) in items.iter().enumerate() {
+        for (idx, entry) in entries.iter().enumerate() {
+            let item = &entry.item;
             let entry_range = [idx as u64, idx as u64];
             match item {
                 Item::Message { role, content, .. } => {
-                    let kind = match role {
-                        Role::User => ReferenceKind::User,
-                        Role::Assistant => ReferenceKind::Assistant,
-                        Role::System => continue,
+                    let Some(kind) = message_reference_kind(&entry.annotation.origin, role) else {
+                        continue;
                     };
                     let text = content
                         .iter()
@@ -234,9 +263,10 @@ impl SessionCapture {
                         .join("");
                     let label = format!("{} message", kind.as_str());
                     let summary = truncate_chars(&text, 240);
-                    let id = SessionEntryRef::new(idx);
+                    let id = SessionEntryRef::from_history_entry_id(&entry.annotation.entry_id);
                     index.push(ReferenceEntry {
                         id: id.clone(),
+                        origin: entry.annotation.origin.clone(),
                         entry_range,
                         kind,
                         tool_part: None,
@@ -248,6 +278,7 @@ impl SessionCapture {
                     if matches!(kind, ReferenceKind::User | ReferenceKind::Assistant) {
                         overview.push(OverviewItem {
                             id: id.clone(),
+                            origin: entry.annotation.origin.clone(),
                             entry_range,
                             kind,
                             label,
@@ -261,7 +292,8 @@ impl SessionCapture {
                 } => {
                     let text = format!("{name}\n{arguments}");
                     index.push(ReferenceEntry {
-                        id: SessionEntryRef::new(idx),
+                        id: SessionEntryRef::from_history_entry_id(&entry.annotation.entry_id),
+                        origin: entry.annotation.origin.clone(),
                         entry_range,
                         kind: ReferenceKind::Tool,
                         tool_part: Some(ToolPart::Input),
@@ -287,7 +319,8 @@ impl SessionCapture {
                         content.as_deref().unwrap_or_default(),
                     );
                     index.push(ReferenceEntry {
-                        id: SessionEntryRef::new(idx),
+                        id: SessionEntryRef::from_history_entry_id(&entry.annotation.entry_id),
+                        origin: entry.annotation.origin.clone(),
                         entry_range,
                         kind: ReferenceKind::Tool,
                         tool_part: Some(ToolPart::Output),
@@ -327,7 +360,7 @@ impl SessionCapture {
 
         Self {
             segment_id,
-            items,
+            entries,
             overview,
             index,
         }
@@ -337,6 +370,14 @@ impl SessionCapture {
         &self.overview
     }
 
+    pub(crate) fn source_index_for_ref(&self, reference: &SessionEntryRef) -> Option<u64> {
+        self.index
+            .iter()
+            .find(|entry| entry.id == *reference)
+            .map(|entry| entry.entry_range[0])
+            .or_else(|| reference.source_index())
+    }
+
     pub(crate) fn search(&self, options: &SearchOptions) -> Vec<SearchHit> {
         let query = options.query.trim().to_lowercase();
         let limit = options
@@ -347,12 +388,12 @@ impl SessionCapture {
         let min_entry_index = options
             .from
             .as_ref()
-            .and_then(SessionEntryRef::source_index)
+            .and_then(|reference| self.source_index_for_ref(reference))
             .unwrap_or_else(|| options.min_entry_index.unwrap_or(0));
         let max_entry_index = options
             .through
             .as_ref()
-            .and_then(SessionEntryRef::source_index)
+            .and_then(|reference| self.source_index_for_ref(reference))
             .unwrap_or(u64::MAX);
         let mut skipped = 0usize;
         let mut hits = Vec::new();
@@ -391,6 +432,7 @@ impl SessionCapture {
             }
             hits.push(SearchHit {
                 id: entry.id.clone(),
+                origin: entry.origin.clone(),
                 kind: entry.kind,
                 tool_part: entry.tool_part,
                 tool_name: entry.tool_name.clone(),
@@ -442,13 +484,18 @@ impl SessionCapture {
                     }
                 }
             }
-            let Some(item) = self.items.get(entry.entry_range[0] as usize) else {
+            let Some(item) = self
+                .entries
+                .get(entry.entry_range[0] as usize)
+                .map(|entry| &entry.item)
+            else {
                 continue;
             };
             let text = render_item(item, entry, options.detail, max_bytes.saturating_sub(bytes));
             bytes = bytes.saturating_add(text.len());
             entries.push(ReadEntry {
                 id: entry.id.clone(),
+                origin: entry.origin.clone(),
                 kind: entry.kind,
                 tool_part: entry.tool_part,
                 tool_name: entry.tool_name.clone(),
@@ -485,6 +532,7 @@ impl SessionCapture {
         Some(SessionEntryEvidence {
             segment_id: self.segment_id.clone(),
             entry_ref: entry.id.clone(),
+            origin: entry.origin.clone(),
             entry_range: entry.entry_range,
             kind: entry.kind,
             tool_part: entry.tool_part,
@@ -495,6 +543,28 @@ impl SessionCapture {
     }
 }
 
+fn message_reference_kind(
+    origin: &WorkerHistoryProvenance,
+    provider_role: &Role,
+) -> Option<ReferenceKind> {
+    match origin {
+        WorkerHistoryProvenance::HumanInput { .. }
+        | WorkerHistoryProvenance::WorkerInput { .. } => Some(ReferenceKind::User),
+        WorkerHistoryProvenance::ModelOutput { .. } => Some(ReferenceKind::Assistant),
+        WorkerHistoryProvenance::ToolOutput { .. } => Some(ReferenceKind::Tool),
+        WorkerHistoryProvenance::LegacyUnknown => match provider_role {
+            Role::User => Some(ReferenceKind::User),
+            Role::Assistant => Some(ReferenceKind::Assistant),
+            Role::System => None,
+        },
+        // Flow/backend/system content remains out of the observation surface
+        // even when represented with a provider user/system role.
+        WorkerHistoryProvenance::FlowInstruction { .. }
+        | WorkerHistoryProvenance::BackendInstruction { .. }
+        | WorkerHistoryProvenance::DerivedSummary => None,
+    }
+}
+
 fn render_item(
     item: &Item,
     entry: &ReferenceEntry,
@@ -563,6 +633,60 @@ fn truncate_chars(text: &str, max_chars: usize) -> String {
 mod tests {
     use super::*;
 
+    #[test]
+    fn flow_user_role_is_excluded_while_explicit_human_origin_remains_evidence() {
+        let entries = vec![
+            crate::session_history::history_entry(
+                Item::user_message("trusted flow instruction"),
+                WorkerHistoryProvenance::FlowInstruction {
+                    selector: "builtin:coder-review".into(),
+                    definition_id: "coder-review".into(),
+                    definition_revision: 3,
+                    instance_id: "instance".into(),
+                    state_id: "implement".into(),
+                },
+            ),
+            crate::session_history::history_entry(
+                Item::user_message("remember my preference"),
+                WorkerHistoryProvenance::HumanInput {
+                    account_id: "account-1".into(),
+                },
+            ),
+        ];
+        let capture = SessionCapture::from_history_entries("segment", entries);
+        let overview = capture.overview();
+        assert_eq!(overview.len(), 1);
+        assert!(matches!(
+            overview[0].origin,
+            WorkerHistoryProvenance::HumanInput { .. }
+        ));
+        let evidence = capture.evidence_for(overview[0].id.as_str()).unwrap();
+        assert!(evidence.excerpt.ends_with("remember my preference"));
+        assert!(matches!(
+            evidence.origin,
+            WorkerHistoryProvenance::HumanInput { .. }
+        ));
+    }
+
+    #[test]
+    fn stable_logical_ref_survives_retention_and_restore_projection() {
+        let retained = crate::session_history::history_entry(
+            Item::assistant_message("retained"),
+            WorkerHistoryProvenance::ModelOutput {
+                worker: crate::session_history::worker_subject(Default::default()),
+            },
+        );
+        let expected_ref = SessionEntryRef::from_history_entry_id(&retained.annotation.entry_id);
+        let before = SessionCapture::from_history_entries("old", vec![retained.clone()]);
+        let after = SessionCapture::from_history_entries("new", vec![retained]);
+        assert_eq!(before.overview()[0].id, expected_ref);
+        assert_eq!(after.overview()[0].id, expected_ref);
+        assert_eq!(
+            after.evidence_for(expected_ref.as_str()).unwrap().entry_ref,
+            expected_ref
+        );
+    }
+
     #[test]
     fn overview_contains_user_and_assistant_only() {
         let view = SessionCapture::new(

crates/worker/src/session_history.rs

@@ -0,0 +1,219 @@
+//! Restore-authoritative metadata for model-visible Worker history.
+//!
+//! Agen transports this annotation without interpreting it. Session Log v2
+//! stores each item and metadata in one typed record; legacy records are
+//! retained only as explicit `LegacyUnknown` entries.
+
+use agen::{HistoryEntry, Item};
+use protocol::Segment;
+use session_store::{
+    LogEntry, LoggedHistoryDerivation, LoggedHistoryEntry, LoggedSessionHistoryEntryId,
+    LoggedSessionHistoryMetadata, LoggedSessionHistoryOrigin, LoggedWorkerSubject, SegmentId,
+    SessionId,
+};
+
+pub type SessionHistoryEntryId = LoggedSessionHistoryEntryId;
+pub type SessionHistoryMetadata = LoggedSessionHistoryMetadata;
+pub type WorkerHistoryProvenance = LoggedSessionHistoryOrigin;
+pub type SessionHistoryDerivation = LoggedHistoryDerivation;
+pub type WorkerSubjectSnapshot = LoggedWorkerSubject;
+
+pub(crate) fn worker_subject(session_id: SessionId) -> WorkerSubjectSnapshot {
+    WorkerSubjectSnapshot {
+        workspace_id: None,
+        runtime_id: None,
+        worker_id: session_id.to_string(),
+    }
+}
+
+pub(crate) fn metadata(
+    origin: WorkerHistoryProvenance,
+    derivation: Option<SessionHistoryDerivation>,
+) -> SessionHistoryMetadata {
+    SessionHistoryMetadata {
+        entry_id: SessionHistoryEntryId::new(),
+        origin,
+        derivation,
+    }
+}
+
+pub(crate) fn history_entry(
+    item: Item,
+    origin: WorkerHistoryProvenance,
+) -> HistoryEntry<SessionHistoryMetadata> {
+    HistoryEntry::new(item, metadata(origin, None))
+}
+
+pub(crate) fn to_logged_history_entry(
+    entry: &HistoryEntry<SessionHistoryMetadata>,
+) -> LoggedHistoryEntry {
+    LoggedHistoryEntry {
+        item: entry.item.clone().into(),
+        metadata: entry.annotation.clone(),
+    }
+}
+
+fn legacy_entry(item: Item) -> HistoryEntry<SessionHistoryMetadata> {
+    HistoryEntry::new(item, SessionHistoryMetadata::legacy_unknown())
+}
+
+fn from_logged(entry: &LoggedHistoryEntry) -> HistoryEntry<SessionHistoryMetadata> {
+    HistoryEntry::new(Item::from(entry.item.clone()), entry.metadata.clone())
+}
+
+/// Rebuild typed Worker history directly from the append-only Session Log.
+/// Missing legacy metadata is never inferred from role or plaintext.
+pub(crate) fn restore_history_entries(
+    _session_id: SessionId,
+    _segment_id: SegmentId,
+    entries: &[LogEntry],
+) -> Result<Vec<HistoryEntry<SessionHistoryMetadata>>, String> {
+    let mut history = Vec::new();
+    for entry in entries {
+        match entry {
+            LogEntry::AnnotatedSegmentStart { history: seed, .. } => {
+                history = seed.iter().map(from_logged).collect();
+            }
+            LogEntry::SegmentStart { history: seed, .. } => {
+                history = seed
+                    .iter()
+                    .cloned()
+                    .map(Item::from)
+                    .map(legacy_entry)
+                    .collect();
+            }
+            LogEntry::AnnotatedUserInput { history: input, .. } => {
+                history.extend(input.iter().map(from_logged))
+            }
+            LogEntry::UserInput { segments, .. } => history.push(legacy_entry(Item::user_message(
+                Segment::flatten_to_text(segments),
+            ))),
+            LogEntry::AnnotatedAssistantItem { entry, .. }
+            | LogEntry::AnnotatedToolResult { entry, .. } => history.push(from_logged(entry)),
+            LogEntry::AssistantItem { item, .. } | LogEntry::ToolResult { item, .. } => {
+                history.push(legacy_entry(Item::from(item.clone())));
+            }
+            LogEntry::AnnotatedSystemItem { entry, .. } => history.push(HistoryEntry::new(
+                entry.item.to_history_item(),
+                entry.metadata.clone(),
+            )),
+            LogEntry::SystemItem { item, .. } => {
+                history.push(legacy_entry(item.to_history_item()));
+            }
+            _ => {}
+        }
+    }
+    Ok(history)
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use agen::llm_client::RequestConfig;
+    use session_store::LogEntry;
+
+    #[test]
+    fn legacy_user_role_is_not_inferred_as_human_authority() {
+        let entries = vec![LogEntry::UserInput {
+            ts: 1,
+            segments: vec![Segment::text("legacy")],
+            extensions: Vec::new(),
+        }];
+        let restored =
+            restore_history_entries(SessionId::now_v7(), SegmentId::now_v7(), &entries).unwrap();
+        assert!(matches!(
+            restored[0].annotation.origin,
+            WorkerHistoryProvenance::LegacyUnknown
+        ));
+    }
+
+    #[test]
+    fn typed_flow_and_unknown_caller_input_round_trip_without_role_inference() {
+        let session_id = SessionId::now_v7();
+        let projected = vec![
+            history_entry(
+                Item::user_message("flow instructions"),
+                WorkerHistoryProvenance::FlowInstruction {
+                    selector: "builtin:coder-review".to_string(),
+                    definition_id: "coder-review".to_string(),
+                    definition_revision: 7,
+                    instance_id: "flow-instance".to_string(),
+                    state_id: "implement".to_string(),
+                },
+            ),
+            history_entry(
+                Item::user_message("implement"),
+                WorkerHistoryProvenance::LegacyUnknown,
+            ),
+        ];
+        let entries = vec![
+            LogEntry::AnnotatedSegmentStart {
+                ts: 0,
+                session_id,
+                system_prompt: None,
+                config: RequestConfig::default(),
+                history: Vec::new(),
+                forked_from: None,
+                compacted_from: None,
+            },
+            LogEntry::AnnotatedUserInput {
+                ts: 1,
+                segments: vec![
+                    Segment::Flow {
+                        selector: "builtin:coder-review".to_string(),
+                    },
+                    Segment::text("implement"),
+                ],
+                extensions: Vec::new(),
+                history: projected.iter().map(to_logged_history_entry).collect(),
+            },
+        ];
+        let restored = restore_history_entries(session_id, SegmentId::now_v7(), &entries).unwrap();
+        assert_eq!(restored, projected);
+    }
+
+    #[test]
+    fn annotated_restore_preserves_logical_ids_across_reboot() {
+        let session_id = SessionId::now_v7();
+        let entry = history_entry(
+            Item::assistant_message("persisted"),
+            WorkerHistoryProvenance::ModelOutput {
+                worker: worker_subject(session_id),
+            },
+        );
+        let log = vec![LogEntry::AnnotatedSegmentStart {
+            ts: 0,
+            session_id,
+            system_prompt: None,
+            config: RequestConfig::default(),
+            history: vec![to_logged_history_entry(&entry)],
+            forked_from: None,
+            compacted_from: None,
+        }];
+        let first = restore_history_entries(session_id, SegmentId::now_v7(), &log).unwrap();
+        let second = restore_history_entries(session_id, SegmentId::now_v7(), &log).unwrap();
+        assert_eq!(first[0].annotation.entry_id, entry.annotation.entry_id);
+        assert_eq!(second[0].annotation.entry_id, entry.annotation.entry_id);
+    }
+
+    #[test]
+    fn compacted_derivation_uses_stable_logical_entry_ids() {
+        let source = history_entry(
+            Item::user_message("source"),
+            WorkerHistoryProvenance::LegacyUnknown,
+        );
+        let summary = HistoryEntry::new(
+            Item::system_message("summary"),
+            metadata(
+                WorkerHistoryProvenance::DerivedSummary,
+                Some(SessionHistoryDerivation {
+                    sources: vec![source.annotation.entry_id.clone()],
+                }),
+            ),
+        );
+        assert_eq!(
+            summary.annotation.derivation.unwrap().sources,
+            vec![source.annotation.entry_id]
+        );
+    }
+}

crates/worker/src/spawn/tool.rs

@@ -1252,7 +1252,7 @@ extract_threshold = 4000
         )
         .await
         .unwrap();
-        assert!(first_capture.items.iter().any(|item| {
+        assert!(first_capture.entries.iter().map(|entry| &entry.item).any(|item| {
             matches!(item, Item::Message { role: Role::Assistant, content, .. } if content.iter().any(|part| matches!(part, ContentPart::Text { text } if text.contains("reviewed"))))
         }));
 
@@ -1274,7 +1274,7 @@ extract_threshold = 4000
         )
         .await
         .unwrap();
-        assert!(latest_capture.items.len() > first_capture.items.len());
+        assert!(latest_capture.entries.len() > first_capture.entries.len());
 
         fail_requests.store(true, Ordering::SeqCst);
         send.execute(

crates/worker/tests/compact_events_test.rs

@@ -163,7 +163,8 @@ async fn make_worker_with_manifest(
     let scope = worker::Scope::writable(&pwd).unwrap();
     std::mem::forget(pwd_tmp);
 
-    let worker = Engine::new(client);
+    let worker =
+        Engine::<_, agen::state::Mutable, worker::SessionHistoryMetadata>::new_annotated(client);
     let mut worker = Worker::new(
         manifest,
         worker,
@@ -204,7 +205,14 @@ fn system_texts_in_sink_session_start(
 ) -> Vec<String> {
     let (entries, _rx) = worker.sink().subscribe_with_snapshot();
     for entry in entries.into_iter().rev() {
-        if let session_store::LogEntry::SegmentStart { history, .. } = entry {
+        let history = match entry {
+            session_store::LogEntry::AnnotatedSegmentStart { history, .. } => history
+                .into_iter()
+                .map(|entry| entry.item)
+                .collect::<Vec<_>>(),
+            session_store::LogEntry::SegmentStart { history, .. } => history,
+            _ => continue,
+        };
         return history
             .into_iter()
             .filter_map(|logged| {
@@ -226,7 +234,6 @@ fn system_texts_in_sink_session_start(
             })
             .collect();
     }
-    }
     Vec::new()
 }
 
@@ -337,7 +344,12 @@ permission = "write"
     // New segment records forked_from pointing at the source.
     let new_entries = store.read_all(session_id, new_segment_id).unwrap();
     match &new_entries[0] {
-        LogEntry::SegmentStart {
+        LogEntry::AnnotatedSegmentStart {
+            session_id: seg_session,
+            forked_from: Some(origin),
+            ..
+        }
+        | LogEntry::SegmentStart {
             session_id: seg_session,
             forked_from: Some(origin),
             ..

crates/worker/tests/controller_test.rs

@@ -32,16 +32,29 @@ fn history_from_sink(handle: &WorkerHandle) -> Vec<Item> {
     let mut items = Vec::new();
     for entry in entries {
         match entry {
+            LogEntry::AnnotatedSegmentStart { history, .. } => {
+                items.extend(history.into_iter().map(|entry| Item::from(entry.item)));
+            }
             LogEntry::SegmentStart { history, .. } => {
                 items.extend(history.into_iter().map(Item::from));
             }
+            LogEntry::AnnotatedUserInput { history, .. } => {
+                items.extend(history.into_iter().map(|entry| Item::from(entry.item)));
+            }
             LogEntry::UserInput { segments, .. } => {
                 let text = protocol::Segment::flatten_to_text(&segments);
                 items.push(Item::user_message(text));
             }
+            LogEntry::AnnotatedAssistantItem { entry, .. }
+            | LogEntry::AnnotatedToolResult { entry, .. } => {
+                items.push(Item::from(entry.item));
+            }
             LogEntry::AssistantItem { item, .. } | LogEntry::ToolResult { item, .. } => {
                 items.push(Item::from(item));
             }
+            LogEntry::AnnotatedSystemItem { entry, .. } => {
+                items.push(entry.item.to_history_item());
+            }
             LogEntry::SystemItem { item, .. } => {
                 items.push(item.to_history_item());
             }
@@ -51,6 +64,14 @@ fn history_from_sink(handle: &WorkerHandle) -> Vec<Item> {
     items
 }
 
+fn system_item(entry: &LogEntry) -> Option<&session_store::SystemItem> {
+    match entry {
+        LogEntry::AnnotatedSystemItem { entry, .. } => Some(&entry.item),
+        LogEntry::SystemItem { item, .. } => Some(item),
+        _ => None,
+    }
+}
+
 // ---------------------------------------------------------------------------
 // Mock LLM Client
 // ---------------------------------------------------------------------------
@@ -192,7 +213,8 @@ async fn make_worker_with_pwd_and_manifest(
     let scope = manifest::Scope::writable(&pwd).unwrap();
     std::mem::forget(pwd_tmp);
 
-    let worker = Engine::new(client);
+    let worker =
+        Engine::<_, agen::state::Mutable, worker::SessionHistoryMetadata>::new_annotated(client);
     let authority = WorkerFilesystemAuthority::local(pwd.clone(), pwd.clone());
     let worker = Worker::new(
         manifest,
@@ -784,13 +806,30 @@ async fn snapshot_includes_user_input_for_in_flight_turn() {
     let client = MockClient::sequential(vec![MockResponse::Hang(simple_text_events())]);
     let worker = make_worker(client).await;
     let handle = spawn_controller(worker).await;
+    let mut events = handle.subscribe();
 
     handle
         .send(Method::run_text("hello in-flight"))
         .await
         .unwrap();
-    wait_for_status(&handle, WorkerStatus::Running).await;
+    tokio::time::timeout(std::time::Duration::from_secs(2), async {
+        loop {
+            if matches!(
+                events.recv().await,
+                Ok(Event::Status {
+                    status: WorkerStatus::Running,
+                })
+            ) {
+                break;
+            }
+        }
+    })
+    .await
+    .expect("running status event");
 
+    // The Running event is the in-flight visibility fence: the committed
+    // annotated input must already be available to an immediately attaching
+    // subscriber rather than racing behind this status transition.
     let stream = tokio::net::UnixStream::connect(handle.runtime_dir.socket_path())
         .await
         .unwrap();
@@ -804,10 +843,12 @@ async fn snapshot_includes_user_input_for_in_flight_turn() {
                 // Walk the entries, find a `LogEntry::UserInput` and
                 // confirm its segments flatten to our submitted text.
                 let mut found = false;
-                for value in entries {
+                for value in &entries {
                     let entry: session_store::LogEntry =
-                        serde_json::from_value(value).expect("LogEntry deserialise");
-                    if let session_store::LogEntry::UserInput { segments, .. } = entry {
+                        serde_json::from_value(value.clone()).expect("LogEntry deserialise");
+                    if let session_store::LogEntry::UserInput { segments, .. }
+                    | session_store::LogEntry::AnnotatedUserInput { segments, .. } = entry
+                    {
                         let text = protocol::Segment::flatten_to_text(&segments);
                         if text == "hello in-flight" {
                             found = true;
@@ -815,7 +856,10 @@ async fn snapshot_includes_user_input_for_in_flight_turn() {
                         }
                     }
                 }
-                assert!(found, "snapshot must carry the in-flight UserInput entry");
+                assert!(
+                    found,
+                    "snapshot must carry the in-flight UserInput entry: {entries:?}"
+                );
                 return;
             }
             Event::Alert(_) => continue,
@@ -1086,7 +1130,7 @@ async fn run_with_paste_segment_inlines_content_and_emits_typed_user_message() {
                 _ => {}
             },
             entry = entry_rx.recv() => match entry {
-                Ok(session_store::LogEntry::UserInput { segments, .. }) => {
+                Ok(session_store::LogEntry::UserInput { segments, .. } | session_store::LogEntry::AnnotatedUserInput { segments, .. }) => {
                     user_input_segments = Some(segments);
                     if saw_turn_end {
                         break;
@@ -1317,11 +1361,8 @@ async fn notify_while_idle_auto_starts_turn_and_injects_system_message() {
     let (entries, _) = handle.sink.subscribe_with_snapshot();
     let saw_notify_in_mirror = entries.iter().any(|e| {
         matches!(
-            e,
-            session_store::LogEntry::SystemItem {
-                item: session_store::SystemItem::Notification { message, .. },
-                ..
-            } if message == "turn finished"
+            system_item(e),
+            Some(session_store::SystemItem::Notification { message, .. }) if message == "turn finished"
         )
     });
     assert!(
@@ -1463,14 +1504,11 @@ async fn worker_event_turn_ended_while_idle_auto_starts_turn_and_injects_system_
     let (entries, _) = handle.sink.subscribe_with_snapshot();
     let saw_worker_event_in_mirror = entries.iter().any(|e| {
         matches!(
-            e,
-            session_store::LogEntry::SystemItem {
-                item: session_store::SystemItem::WorkerEvent {
+            system_item(e),
+            Some(session_store::SystemItem::WorkerEvent {
                 event: protocol::WorkerEvent::TurnEnded { worker_name },
                 ..
-                },
-                ..
-            } if worker_name == "child"
+            }) if worker_name == "child"
         )
     });
     assert!(
@@ -1552,14 +1590,11 @@ async fn worker_event_scope_sub_delegated_while_idle_stays_control_plane_only()
     let (entries, _) = handle.sink.subscribe_with_snapshot();
     let saw_scope_event_in_mirror = entries.iter().any(|entry| {
         matches!(
-            entry,
-            session_store::LogEntry::SystemItem {
-                item: session_store::SystemItem::WorkerEvent {
+            system_item(entry),
+            Some(session_store::SystemItem::WorkerEvent {
                 event: protocol::WorkerEvent::ScopeSubDelegated { .. },
                 ..
-                },
-                ..
-            }
+            })
         )
     });
     assert!(
@@ -2134,9 +2169,13 @@ async fn paused_then_run_closes_orphan_tool_use_for_next_request() {
     for item in items {
         match item {
             agen::Item::ToolResult {
-                call_id, summary, ..
+                call_id,
+                summary,
+                disposition,
+                ..
             } if call_id == "call_orphan" => {
-                assert_eq!(summary, "[Interrupted by user]");
+                assert_eq!(summary, "Tool execution outcome unknown");
+                assert_eq!(*disposition, agen::ToolResultDisposition::OutcomeUnknown);
                 saw_synthetic_tool_result = true;
             }
             agen::Item::Message { role, content, .. } if *role == agen::Role::System => {
@@ -2327,8 +2366,11 @@ async fn paused_cancel_abandons_resume_and_next_input_is_fresh_run() {
     assert!(
         items.iter().any(|item| matches!(
             item,
-            agen::Item::ToolResult { call_id, summary, .. }
-                if call_id == "call_cancelled" && summary == "[Interrupted by user]"
+            agen::Item::ToolResult {
+                call_id,
+                disposition: agen::ToolResultDisposition::OutcomeUnknown,
+                ..
+            } if call_id == "call_cancelled"
         )),
         "paused cancel should close orphan tool_use before future requests: {items:?}"
     );
@@ -2373,7 +2415,8 @@ async fn snapshot_contains_user_input(handle: &WorkerHandle, needle: &str) -> bo
                     let entry: session_store::LogEntry =
                         serde_json::from_value(value).expect("LogEntry deserialise");
                     match entry {
-                        session_store::LogEntry::UserInput { segments, .. } => {
+                        session_store::LogEntry::UserInput { segments, .. }
+                        | session_store::LogEntry::AnnotatedUserInput { segments, .. } => {
                             protocol::Segment::flatten_to_text(&segments).contains(needle)
                         }
                         _ => false,

crates/worker/tests/session_metrics_test.rs

@@ -188,7 +188,8 @@ async fn make_worker(
     let pwd = pwd_tmp.path().to_path_buf();
     let scope = worker::Scope::writable(&pwd).unwrap();
 
-    let mut worker = Engine::new(client);
+    let mut worker =
+        Engine::<_, agen::state::Mutable, worker::SessionHistoryMetadata>::new_annotated(client);
     worker.register_tool(big_content_tool_definition(tool_name));
 
     let worker = Worker::new(
@@ -460,7 +461,8 @@ async fn metric_write_failure_emits_warn_alert_and_does_not_abort_run() {
     // protected token budget covers the only user message). That is enough to drive
     // the failure path: at least one metric attempts to write.
     let client = MockClient::new(vec![text_response_with_cache("hi", 0, 0)]);
-    let worker = Engine::new(client);
+    let worker =
+        Engine::<_, agen::state::Mutable, worker::SessionHistoryMetadata>::new_annotated(client);
     let mut worker = Worker::new(
         manifest,
         worker,
@@ -536,7 +538,8 @@ permission = "write"
     let pwd_tmp = tempfile::tempdir().unwrap();
     let pwd = pwd_tmp.path().to_path_buf();
     let scope = worker::Scope::writable(&pwd).unwrap();
-    let worker = Engine::new(client);
+    let worker =
+        Engine::<_, agen::state::Mutable, worker::SessionHistoryMetadata>::new_annotated(client);
     let mut worker = Worker::new(
         manifest,
         worker,

crates/worker/tests/system_prompt_template_test.rs

@@ -130,7 +130,8 @@ async fn make_worker_with_body(
         EffectivePromptCatalog::new(templates, 1, "test-schema", "test-toolchain").unwrap();
     let loader = PromptCatalogSource::builtins_only().with_effective_catalog(projection);
 
-    let worker = Engine::new(client);
+    let worker =
+        Engine::<_, agen::state::Mutable, worker::SessionHistoryMetadata>::new_annotated(client);
     let mut worker = Worker::new(
         manifest,
         worker,

crates/workspace-server/src/hosts.rs

@@ -4449,6 +4449,13 @@ mod tests {
             .resolve_profile("builtin:companion", root.path(), "embedded-test-companion")
             .unwrap();
         assert!(companion.feature.manage_workdir.enabled);
+        assert!(companion.feature.sub_worker.enabled);
+        assert!(!companion.feature.worker.enabled);
+        let coder = archive
+            .resolve_profile("builtin:coder", root.path(), "embedded-test-coder")
+            .unwrap();
+        assert!(coder.feature.sub_worker.enabled);
+        assert!(!coder.feature.worker.enabled);
     }
 
     #[test]
@@ -4475,6 +4482,8 @@ mod tests {
             .resolve_profile("builtin:coder", root.path(), "remote-test-worker")
             .unwrap();
         assert_eq!(manifest.worker.name, "remote-test-worker");
+        assert!(manifest.feature.sub_worker.enabled);
+        assert!(!manifest.feature.worker.enabled);
     }
 
     #[test]

crates/workspace-server/src/server.rs

@@ -5458,6 +5458,36 @@ fn resolve_workspace_ticket_reference(
         .ok_or_else(|| Error::Ticket(ticket::TicketError::NotFound(reference.to_string())).into())
 }
 
+fn resolve_workspace_ticket_identity(
+    api: &WorkspaceApi,
+    workspace_id: &str,
+    reference: &str,
+) -> ApiResult<String> {
+    let ticket_id = resolve_workspace_ticket_reference(api, workspace_id, reference)?;
+    let ticket = browser_ticket_backend(api)?
+        .show(ticket_id.clone().into())
+        .map_err(Error::from)?;
+    if ticket.meta.id != ticket_id {
+        return Err(Error::InvalidInput(
+            "resolved Ticket identity does not match Ticket authority".to_string(),
+        )
+        .into());
+    }
+    Ok(ticket_id)
+}
+
+fn resolve_workspace_worker_ticket_assignment(
+    api: &WorkspaceApi,
+    workspace_id: &str,
+    assignment: &mut Option<CreateWorkspaceWorkerTicketAssignmentRequest>,
+) -> ApiResult<()> {
+    if let Some(assignment) = assignment {
+        assignment.ticket_id =
+            resolve_workspace_ticket_identity(api, workspace_id, &assignment.ticket_id)?;
+    }
+    Ok(())
+}
+
 #[derive(Debug, serde::Deserialize)]
 struct MergeRequestListHttpQuery {
     state: Option<String>,
@@ -8217,6 +8247,11 @@ async fn spawn_known_worker(
 ) -> ApiResult<Json<BrowserCreateWorkerResponse>> {
     validate_workspace_scope(&api, &path.workspace_id)?;
     let source = authenticate_worker_mutation_source(&api, &path.workspace_id, &headers)?;
+    resolve_workspace_worker_ticket_assignment(
+        &api,
+        &path.workspace_id,
+        &mut request.ticket_assignment,
+    )?;
     let relation = if request.ticket_assignment.is_some() {
         "assigned"
     } else {
@@ -9261,9 +9296,8 @@ fn cleanup_working_directory_for_runtime(
             result.diagnostics,
         ));
     };
-    let record = workdir_record_from_summary(&api, runtime_id, &working_directory.summary);
-    api.store.upsert_workdir_registry(&record)?;
     let mut summary = working_directory.summary;
+    persist_workdir_cleanup_observation(&api, runtime_id, &summary)?;
     apply_workdir_occupancy_projection(&api, &mut summary)?;
     Ok(Json(BrowserWorkingDirectoryDetailResponse {
         workspace_id: api.config.workspace_id.clone(),
@@ -14154,11 +14188,7 @@ fn sync_runtime_workdir_observations(
                         api.store.upsert_workdir_registry(&updated)?;
                     }
                 } else {
-                    record.materialization_status =
-                        workdir_status_from_runtime_miss(result.diagnostics.as_slice()).to_string();
-                    record.cleanliness = "unknown".to_string();
-                    record.updated_at = now_registry_timestamp();
-                    api.store.upsert_workdir_registry(&record)?;
+                    persist_workdir_runtime_miss(api, record, result.diagnostics.as_slice())?;
                 }
             }
             Err(_) => {
@@ -14172,17 +14202,54 @@ fn sync_runtime_workdir_observations(
     Ok(response.diagnostics)
 }
 
-fn workdir_status_from_runtime_miss(diagnostics: &[RuntimeDiagnostic]) -> &'static str {
-    if diagnostics
+fn persist_workdir_cleanup_observation(
+    api: &WorkspaceApi,
+    runtime_id: &str,
+    summary: &WorkingDirectorySummary,
+) -> ApiResult<()> {
+    if summary.status == WorkingDirectoryStatusKind::NotFound {
+        api.store.delete_workdir_registry(
+            &api.config.workspace_id,
+            summary.working_directory_id.as_str(),
+        )?;
+    } else {
+        let record = workdir_record_from_summary(api, runtime_id, summary);
+        api.store.upsert_workdir_registry(&record)?;
+    }
+    Ok(())
+}
+
+fn workdir_runtime_miss_is_not_found(diagnostics: &[RuntimeDiagnostic]) -> bool {
+    diagnostics
         .iter()
         .any(|diagnostic| diagnostic.code == "working_directory_not_found")
-    {
+}
+
+fn workdir_status_from_runtime_miss(diagnostics: &[RuntimeDiagnostic]) -> &'static str {
+    if workdir_runtime_miss_is_not_found(diagnostics) {
         "not_found"
     } else {
         "unknown"
     }
 }
 
+fn persist_workdir_runtime_miss(
+    api: &WorkspaceApi,
+    mut record: WorkdirRegistryRecord,
+    diagnostics: &[RuntimeDiagnostic],
+) -> ApiResult<()> {
+    if workdir_runtime_miss_is_not_found(diagnostics) {
+        api.store
+            .delete_workdir_registry(&api.config.workspace_id, record.workdir_id.as_str())?;
+    } else {
+        record.materialization_status = "unknown".to_string();
+        record.cleanliness = "unknown".to_string();
+        record.updated_at = now_registry_timestamp();
+        api.store.upsert_workdir_registry(&record)?;
+    }
+    Ok(())
+}
+
 fn sync_all_runtime_workdir_observations(api: &WorkspaceApi) -> Vec<RuntimeDiagnostic> {
     let mut diagnostics = Vec::new();
     let runtimes = api.runtime.list_runtimes(api.config.max_records.min(200));
@@ -16980,22 +17047,24 @@ mod tests {
 
     #[test]
     fn workdir_runtime_miss_uses_exact_typed_code() {
-        assert_eq!(
-            workdir_status_from_runtime_miss(&[RuntimeDiagnostic {
+        let typed_not_found = [RuntimeDiagnostic {
             code: "working_directory_not_found".to_string(),
             severity: DiagnosticSeverity::Warning,
             message: "missing".to_string(),
-            }]),
+        }];
+        assert_eq!(
+            workdir_status_from_runtime_miss(&typed_not_found),
             "not_found"
         );
-        assert_eq!(
-            workdir_status_from_runtime_miss(&[RuntimeDiagnostic {
+        assert!(workdir_runtime_miss_is_not_found(&typed_not_found));
+
+        let unrelated = [RuntimeDiagnostic {
             code: "some_other_not_found".to_string(),
             severity: DiagnosticSeverity::Warning,
             message: "not a typed workdir miss".to_string(),
-            }]),
-            "unknown"
-        );
+        }];
+        assert_eq!(workdir_status_from_runtime_miss(&unrelated), "unknown");
+        assert!(!workdir_runtime_miss_is_not_found(&unrelated));
     }
 
     struct DeterministicExecutionBackend {
@@ -20186,6 +20255,25 @@ mod tests {
                 .unwrap(),
             ticket_id
         );
+        assert_eq!(
+            resolve_workspace_ticket_identity(&api, TEST_WORKSPACE_ID, &ticket_resource_key)
+                .unwrap(),
+            ticket_id
+        );
+        assert_eq!(
+            resolve_workspace_ticket_identity(&api, TEST_WORKSPACE_ID, &ticket_id).unwrap(),
+            ticket_id
+        );
+        let mut assignment = Some(CreateWorkspaceWorkerTicketAssignmentRequest {
+            ticket_id: ticket_resource_key.clone(),
+            operation_id: "ticket-key-assignment".to_string(),
+        });
+        resolve_workspace_worker_ticket_assignment(&api, TEST_WORKSPACE_ID, &mut assignment)
+            .unwrap();
+        assert_eq!(assignment.unwrap().ticket_id, ticket_id);
+        let missing =
+            resolve_workspace_ticket_identity(&api, TEST_WORKSPACE_ID, "T-999999").unwrap_err();
+        assert_eq!(missing.into_response().status(), StatusCode::NOT_FOUND);
         let path = || ScopedRecordPath {
             workspace_id: TEST_WORKSPACE_ID.to_string(),
             id: ticket_id.clone(),
@@ -21308,6 +21396,87 @@ mod tests {
             .unwrap();
     }
 
+    #[tokio::test]
+    async fn confirmed_runtime_miss_removes_registry_record_but_unknown_is_retained() {
+        let workspace = tempfile::tempdir().unwrap();
+        init_clean_git_workspace(workspace.path());
+        let api = test_api(workspace.path()).await;
+        seed_cleanup_workdir(&api, "deleted-workdir", "present", "clean");
+        let deleted = api
+            .store
+            .get_workdir_registry(TEST_WORKSPACE_ID, "deleted-workdir")
+            .unwrap()
+            .unwrap();
+
+        persist_workdir_runtime_miss(
+            &api,
+            deleted,
+            &[RuntimeDiagnostic {
+                code: "working_directory_not_found".to_string(),
+                severity: DiagnosticSeverity::Warning,
+                message: "missing".to_string(),
+            }],
+        )
+        .unwrap();
+
+        assert!(
+            api.store
+                .get_workdir_registry(TEST_WORKSPACE_ID, "deleted-workdir")
+                .unwrap()
+                .is_none()
+        );
+
+        seed_cleanup_workdir(&api, "unknown-workdir", "present", "clean");
+        let unknown = api
+            .store
+            .get_workdir_registry(TEST_WORKSPACE_ID, "unknown-workdir")
+            .unwrap()
+            .unwrap();
+        persist_workdir_runtime_miss(
+            &api,
+            unknown,
+            &[RuntimeDiagnostic {
+                code: "runtime_unavailable".to_string(),
+                severity: DiagnosticSeverity::Warning,
+                message: "temporarily unavailable".to_string(),
+            }],
+        )
+        .unwrap();
+        assert_eq!(
+            api.store
+                .get_workdir_registry(TEST_WORKSPACE_ID, "unknown-workdir")
+                .unwrap()
+                .unwrap()
+                .materialization_status,
+            "unknown"
+        );
+    }
+
+    #[tokio::test]
+    async fn cleanup_not_found_observation_removes_registry_record() {
+        let workspace = tempfile::tempdir().unwrap();
+        init_clean_git_workspace(workspace.path());
+        let api = test_api(workspace.path()).await;
+        let working_directory_id = "cleanup-existing";
+        seed_cleanup_workdir(&api, working_directory_id, "present", "clean");
+        let record = api
+            .store
+            .get_workdir_registry(TEST_WORKSPACE_ID, working_directory_id)
+            .unwrap()
+            .unwrap();
+        let mut summary = workdir_summary_from_record(&record);
+        summary.status = WorkingDirectoryStatusKind::NotFound;
+
+        persist_workdir_cleanup_observation(&api, "runtime-test", &summary).unwrap();
+
+        assert!(
+            api.store
+                .get_workdir_registry(TEST_WORKSPACE_ID, working_directory_id)
+                .unwrap()
+                .is_none()
+        );
+    }
+
     fn seed_cleanup_link(api: &WorkspaceApi, runtime_worker_id: &str, workdir_id: &str) {
         let runtime_worker_id = runtime_worker_id.parse::<u64>().unwrap();
         api.store

docs/development/workspace-schema-migrations.md

@@ -19,7 +19,7 @@ The Workspace Server owns one control-plane SQLite database. Schema changes are
 
 Start exactly one instance of the new Server binary against the database. Startup applies migration 39 in one SQLite transaction after the Ticket and Merge Request component schemas are available. The migration:
 
-- rebuilds Ticket, Objective, assignment, Artifact, and human-key tables with Workspace-scoped composite identity;
+- rebuilds Ticket, Objective, assignment, Artifact, and resource-key tables with Workspace-scoped composite identity;
 - adds composite foreign keys for repository, Ticket, Objective, Worker, relation-target, and current-assignment references;
 - materializes assignment-specific Worker tombstones for pre-v39 historical assignments whose valid Worker UUID no longer has a matching live registry row (including Workers deleted by the legacy cleanup path and Workers moved between Runtimes); a Worker ID that resolves only in another Workspace remains a preflight error;
 - validates new historical assignment/event references with SQLite triggers while allowing those audit rows to survive later Ticket or Worker retention deletion; parent delete/Runtime-move triggers record exact Workspace-scoped tombstones, and startup accepts a missing live parent only when that tombstone exists, so an unrelated same ID in another Workspace cannot change the result; reservation operation ids remain intentionally unconstrained until their resources exist;

resources/profiles/coder.dcdl

@@ -10,7 +10,6 @@ import "./base.dcdl" // {
         web = { enabled = true; };
         sub_worker = { enabled = true; };
         flow = { enabled = true; };
-        worker = { enabled = true; };
         ticket = { enabled = true; thread = true; };
         merge_request = {
             show = true;

resources/profiles/companion.dcdl

@@ -8,7 +8,6 @@ import "./base.dcdl" // {
         memory = { enabled = true; };
         web = { enabled = true; };
         sub_worker = { enabled = true; };
-        worker = { enabled = true; };
         manage_workdir = { enabled = true; };
         ticket = { enabled = true; authoring = true; thread = true; };
     };

web/workspace/deno.json

@@ -6,7 +6,7 @@
     "dev": "deno run -A npm:vite@7.2.7 dev",
     "dev:backend": "cd ../.. && cargo run -p yoi-workspace-server --bin yoi-server -- serve --listen 127.0.0.1:8787",
     "check": "deno run -A npm:@sveltejs/kit@2.49.4 sync && deno run -A npm:svelte-check@4.3.4 --tsconfig ./tsconfig.json",
-    "test": "deno test --allow-read=src,test --allow-env=LOG,VSCODE_TEXTMATE_DEBUG src/lib/workspace/auth/model.test.ts src/lib/workspace/api/http.test.ts src/lib/workspace/header/breadcrumb-model.test.ts src/lib/workspace/console/chat-submit.test.ts src/lib/workspace/console/composer-command.test.ts src/lib/workspace/console/composer-completion.test.ts src/lib/workspace/console/markdown.test.ts test/console/ansi.test.ts src/lib/workspace/console/model.test.ts src/lib/workspace/console/tasks.test.ts src/lib/workspace/console/worker-console.ui.test.ts src/lib/workspace/settings/model.test.ts src/lib/workspace/sidebar/workers.test.ts src/lib/workspace/sidebar/workspace-switcher.test.ts src/lib/workspace/sidebar/worker-subscription.test.ts src/lib/workspace/sidebar/worker-launch.test.ts src/lib/workspace/tickets/merge-request-resources.test.ts src/lib/workspace/tickets/ticket-panel.test.ts test/merge-request-status.test.ts test/config-source/decodal-grammar.test.ts test/config-source/editor-state.test.ts test/config-source/fixed-schema-wrapper.test.ts test/config-source/toolchain.test.ts test/config-source/wasm-parity.test.ts",
+    "test": "deno test --allow-read=src,test --allow-env=LOG,VSCODE_TEXTMATE_DEBUG src/lib/workspace/auth/model.test.ts src/lib/workspace/api/http.test.ts src/lib/workspace/header/breadcrumb-model.test.ts src/lib/workspace/console/chat-submit.test.ts src/lib/workspace/console/composer-command.test.ts src/lib/workspace/console/composer-completion.test.ts src/lib/workspace/console/markdown.test.ts test/console/ansi.test.ts src/lib/workspace/console/model.test.ts src/lib/workspace/console/tasks.test.ts test/ticket-detail-route-reuse.test.ts src/lib/workspace/console/worker-console.ui.test.ts src/lib/workspace/settings/model.test.ts src/lib/workspace/sidebar/workers.test.ts src/lib/workspace/sidebar/workspace-switcher.test.ts src/lib/workspace/sidebar/worker-subscription.test.ts src/lib/workspace/sidebar/worker-launch.test.ts src/lib/workspace/tickets/merge-request-resources.test.ts src/lib/workspace/tickets/ticket-panel.test.ts test/merge-request-status.test.ts test/config-source/decodal-grammar.test.ts test/config-source/editor-state.test.ts test/config-source/fixed-schema-wrapper.test.ts test/config-source/toolchain.test.ts test/config-source/wasm-parity.test.ts",
     "build": "deno run -A npm:vite@7.2.7 build",
     "preview": "deno run -A npm:vite@7.2.7 preview"
   },

web/workspace/deno.lock

@@ -9,9 +9,9 @@
     "npm:@codemirror/view@6.43.8": "6.43.8",
     "npm:@fontsource/ibm-plex-mono@5.3.0": "5.3.0",
     "npm:@lezer/highlight@1.2.3": "1.2.3",
-    "npm:@sveltejs/adapter-static@3.0.9": "3.0.9_@sveltejs+kit@2.49.4__@sveltejs+vite-plugin-svelte@6.2.1___svelte@5.45.6___vite@7.2.7__svelte@5.45.6__typescript@5.9.3__vite@7.2.7_@sveltejs+vite-plugin-svelte@6.2.1__svelte@5.45.6__vite@7.2.7_svelte@5.45.6_typescript@5.9.3_vite@7.2.7",
-    "npm:@sveltejs/kit@2.49.4": "2.49.4_@sveltejs+vite-plugin-svelte@6.2.1__svelte@5.45.6__vite@7.2.7_svelte@5.45.6_typescript@5.9.3_vite@7.2.7",
-    "npm:@sveltejs/vite-plugin-svelte@6.2.1": "6.2.1_svelte@5.45.6_vite@7.2.7",
+    "npm:@sveltejs/adapter-static@3.0.9": "3.0.9_@sveltejs+kit@2.49.4__@sveltejs+vite-plugin-svelte@6.2.1___svelte@5.45.6___vite@7.2.7____yaml@2.9.0___yaml@2.9.0__svelte@5.45.6__typescript@5.9.3__vite@7.2.7___yaml@2.9.0__yaml@2.9.0_yaml@2.9.0",
+    "npm:@sveltejs/kit@2.49.4": "2.49.4_@sveltejs+vite-plugin-svelte@6.2.1__svelte@5.45.6__vite@7.2.7___yaml@2.9.0__yaml@2.9.0_svelte@5.45.6_typescript@5.9.3_vite@7.2.7__yaml@2.9.0_yaml@2.9.0",
+    "npm:@sveltejs/vite-plugin-svelte@6.2.1": "6.2.1_svelte@5.45.6_vite@7.2.7__yaml@2.9.0_yaml@2.9.0",
     "npm:clsx@2.1.1": "2.1.1",
     "npm:cookie@0.6.0": "0.6.0",
     "npm:decodal-codemirror@0.3.0": "0.3.0_@codemirror+language@6.12.4_@codemirror+view@6.43.8_@lezer+highlight@1.2.3_@lezer+lr@1.4.10",
@@ -23,7 +23,8 @@
     "npm:svelte-check@4.3.4": "4.3.4_svelte@5.45.6_typescript@5.9.3",
     "npm:svelte@5.45.6": "5.45.6",
     "npm:typescript@5.9.3": "5.9.3",
-    "npm:vite@7.2.7": "7.2.7"
+    "npm:vite@7.2.7": "7.2.7_yaml@2.9.0",
+    "npm:yaml@2.9.0": "2.9.0"
   },
   "jsr": {
     "@std/assert@1.0.19": {
@@ -433,13 +434,13 @@
         "acorn"
       ]
     },
-    "@sveltejs/adapter-static@3.0.9_@sveltejs+kit@2.49.4__@sveltejs+vite-plugin-svelte@6.2.1___svelte@5.45.6___vite@7.2.7__svelte@5.45.6__typescript@5.9.3__vite@7.2.7_@sveltejs+vite-plugin-svelte@6.2.1__svelte@5.45.6__vite@7.2.7_svelte@5.45.6_typescript@5.9.3_vite@7.2.7": {
+    "@sveltejs/adapter-static@3.0.9_@sveltejs+kit@2.49.4__@sveltejs+vite-plugin-svelte@6.2.1___svelte@5.45.6___vite@7.2.7____yaml@2.9.0___yaml@2.9.0__svelte@5.45.6__typescript@5.9.3__vite@7.2.7___yaml@2.9.0__yaml@2.9.0_yaml@2.9.0": {
       "integrity": "sha512-aytHXcMi7lb9ljsWUzXYQ0p5X1z9oWud2olu/EpmH7aCu4m84h7QLvb5Wp+CFirKcwoNnYvYWhyP/L8Vh1ztdw==",
       "dependencies": [
         "@sveltejs/kit"
       ]
     },
-    "@sveltejs/kit@2.49.4_@sveltejs+vite-plugin-svelte@6.2.1__svelte@5.45.6__vite@7.2.7_svelte@5.45.6_typescript@5.9.3_vite@7.2.7": {
+    "@sveltejs/kit@2.49.4_@sveltejs+vite-plugin-svelte@6.2.1__svelte@5.45.6__vite@7.2.7___yaml@2.9.0__yaml@2.9.0_svelte@5.45.6_typescript@5.9.3_vite@7.2.7__yaml@2.9.0_yaml@2.9.0": {
       "integrity": "sha512-JFtOqDoU0DI/+QSG8qnq5bKcehVb3tCHhOG4amsSYth5/KgO4EkJvi42xSAiyKmXAAULW1/Zdb6lkgGEgSxdZg==",
       "dependencies": [
         "@standard-schema/spec",
@@ -465,7 +466,7 @@
       ],
       "bin": true
     },
-    "@sveltejs/vite-plugin-svelte-inspector@5.0.2_@sveltejs+vite-plugin-svelte@6.2.1__svelte@5.45.6__vite@7.2.7_svelte@5.45.6_vite@7.2.7": {
+    "@sveltejs/vite-plugin-svelte-inspector@5.0.2_@sveltejs+vite-plugin-svelte@6.2.1__svelte@5.45.6__vite@7.2.7___yaml@2.9.0__yaml@2.9.0_svelte@5.45.6_vite@7.2.7__yaml@2.9.0_yaml@2.9.0": {
       "integrity": "sha512-TZzRTcEtZffICSAoZGkPSl6Etsj2torOVrx6Uw0KpXxrec9Gg6jFWQ60Q3+LmNGfZSxHRCZL7vXVZIWmuV50Ig==",
       "dependencies": [
         "@sveltejs/vite-plugin-svelte",
@@ -474,7 +475,7 @@
         "vite"
       ]
     },
-    "@sveltejs/vite-plugin-svelte@6.2.1_svelte@5.45.6_vite@7.2.7": {
+    "@sveltejs/vite-plugin-svelte@6.2.1_svelte@5.45.6_vite@7.2.7__yaml@2.9.0_yaml@2.9.0": {
       "integrity": "sha512-YZs/OSKOQAQCnJvM/P+F1URotNnYNeU3P2s4oIpzm1uFaqUEqRxUB0g5ejMjEb5Gjb9/PiBI5Ktrq4rUUF8UVQ==",
       "dependencies": [
         "@sveltejs/vite-plugin-svelte-inspector",
@@ -966,7 +967,7 @@
         "vfile-message"
       ]
     },
-    "vite@7.2.7": {
+    "vite@7.2.7_yaml@2.9.0": {
       "integrity": "sha512-ITcnkFeR3+fI8P1wMgItjGrR10170d8auB4EpMLPqmx6uxElH3a/hHGQabSHKdqd4FXWO1nFIp9rRn7JQ34ACQ==",
       "dependencies": [
         "esbuild",
@@ -974,14 +975,18 @@
         "picomatch",
         "postcss",
         "rollup",
-        "tinyglobby"
+        "tinyglobby",
+        "yaml"
       ],
       "optionalDependencies": [
         "fsevents"
       ],
+      "optionalPeers": [
+        "yaml"
+      ],
       "bin": true
     },
-    "vitefu@1.1.2_vite@7.2.7": {
+    "vitefu@1.1.2_vite@7.2.7__yaml@2.9.0_yaml@2.9.0": {
       "integrity": "sha512-zpKATdUbzbsycPFBN71nS2uzBUQiVnFoOrr2rvqv34S1lcAgMKKkjWleLGeiJlZ8lwCXvtWaRn7R3ZC16SYRuw==",
       "dependencies": [
         "vite"
@@ -993,6 +998,10 @@
     "w3c-keyname@2.2.8": {
       "integrity": "sha512-dpojBhNsCNN7T82Tm7k26A6G9ML3NkhDsnw9n/eoxSRlVBB4CEtIQ/KTCLI2Fwf3ataSXRhYFkQi3SlnFwPvPQ=="
     },
+    "yaml@2.9.0": {
+      "integrity": "sha512-2AvhNX3mb8zd6Zy7INTtSpl1F15HW6Wnqj0srWlkKLcpYl/gMIMJiyuGq2KeI2YFxUPjdlB+3Lc10seMLtL4cA==",
+      "bin": true
+    },
     "zimmerframe@1.1.4": {
       "integrity": "sha512-B58NGBEoc8Y9MWWCQGl/gq9xBCe4IiKM0a2x7GZdQKOW5Exr8S1W24J6OgM1njK8xCRGvAJIL/MxXHf6SkmQKQ=="
     },
@@ -1024,7 +1033,8 @@
     "packageJson": {
       "dependencies": [
         "npm:@fontsource/ibm-plex-mono@5.3.0",
-        "npm:gen-interface-jp@0.8.0"
+        "npm:gen-interface-jp@0.8.0",
+        "npm:yaml@2.9.0"
       ]
     }
   }

web/workspace/package.json

@@ -5,6 +5,7 @@
   "type": "module",
   "dependencies": {
     "@fontsource/ibm-plex-mono": "5.3.0",
-    "gen-interface-jp": "0.8.0"
+    "gen-interface-jp": "0.8.0",
+    "yaml": "2.9.0"
   }
 }

web/workspace/src/lib/generated/protocol.ts

@@ -16,6 +16,8 @@ export type InvokeKind = "user_send" | "notify" | "worker_event" | "system_remin
 
 export type RunResult = "finished" | "paused" | "limit_reached" | "rolled_back";
 
+export type ToolResultDisposition = "success" | "error" | "interrupted" | "cancelled" | "outcome_unknown";
+
 export type ErrorCode = "already_running" | "not_running" | "not_paused" | "provider_error" | "tool_error" | "invalid_request" | "internal";
 
 export type Permission = "read" | "write";
@@ -191,7 +193,7 @@ summary: string,
  * Full tool output. Absent when the tool chose to return
  * summary-only, or when the result was pruned.
  */
-output?: string | null, is_error: boolean, } } | { "event": "usage", "data": { input_tokens: number | null, output_tokens: number | null, cache_read_input_tokens?: number | null, } } | { "event": "run_end", "data": { result: RunResult, } } | { "event": "error", "data": { code: ErrorCode, message: string, } } | { "event": "snapshot", "data": { entries: Array<unknown>, greeting: Greeting, status: WorkerStatus,
+output?: string | null, disposition?: ToolResultDisposition | null, is_error: boolean, } } | { "event": "usage", "data": { input_tokens: number | null, output_tokens: number | null, cache_read_input_tokens?: number | null, } } | { "event": "run_end", "data": { result: RunResult, } } | { "event": "error", "data": { code: ErrorCode, message: string, } } | { "event": "snapshot", "data": { entries: Array<unknown>, greeting: Greeting, status: WorkerStatus,
 /**
  * Unfinished model output that has already streamed in the current
  * run but is not yet represented by committed snapshot entries.

web/workspace/src/lib/workspace/console/ConsoleLineItem.svelte

@@ -8,6 +8,7 @@
   };
 
   let { item }: Props = $props();
+  let detailOpen = $state(false);
   let nowMs = $state(Date.now());
 
   $effect(() => {
@@ -55,22 +56,20 @@
       line.kind !== 'activity' && line.kind !== 'task_reminder' && line.kind !== 'run_stats';
   }
 
-  function toolSummary(line: ConsoleLine): { label: string; suffix: string; rest: string } {
-    const [firstLine = '', ...rest] = line.body.split('\n');
-    const [label, suffix = ''] = firstLine.split(' — ', 2);
-    return {
-      label,
-      suffix,
-      rest: rest.join('\n')
-    };
+  function toolLabel(line: ConsoleLine): string {
+    return line.toolCallLabel ?? line.toolCall?.name ?? line.title;
+  }
+
+  function toolStatus(line: ConsoleLine): string {
+    return line.toolStatus ?? line.toolCall?.state ?? '';
   }
 
   function shouldRenderMarkdown(line: ConsoleLine): boolean {
     return line.kind === 'user' || line.kind === 'assistant' || line.kind === 'system';
   }
 
-  function bodyTextAfterToolSummary(line: ConsoleLine): string {
-    return toolSummary(line).rest;
+  function toolBodyText(line: ConsoleLine): string {
+    return detailOpen ? (line.expandedBody ?? line.body) : line.body;
   }
 </script>
 
@@ -107,20 +106,27 @@
     </div>
   {:else if item.kind === 'tool'}
     <div class="tool-summary">
-      <span class="tool-label">{toolSummary(item).label}</span>
-      <span class="tool-separator"> — </span>
-      <span class={`tool-suffix ${item.toolCall?.state ?? ''}`}>{toolSummary(item).suffix}</span>
+      <span class="tool-label">{toolLabel(item)}</span>
+      <span class={`tool-status ${item.toolCall?.state ?? ''}`}>{toolStatus(item)}</span>
+      {#if item.detail}
+        <button
+          type="button"
+          class="tool-detail-button"
+          aria-expanded={detailOpen}
+          onclick={() => (detailOpen = !detailOpen)}
+        >detail</button>
+      {/if}
     </div>
   {/if}
   {#if item.compaction}
     <!-- rendered as one lifecycle item above -->
   {:else if item.kind === 'tool'}
-    {#if bodyTextAfterToolSummary(item)}
+    {#if toolBodyText(item)}
       <p class="console-plain-text">
         {#if isBashTool(item)}
-          <AnsiText text={bodyTextAfterToolSummary(item)} />
+          <AnsiText text={toolBodyText(item)} />
         {:else}
-          {bodyTextAfterToolSummary(item)}
+          {toolBodyText(item)}
         {/if}
       </p>
     {/if}
@@ -147,11 +153,10 @@
       {/each}
     </div>
   {/if}
-  {#if item.detail}
-    <details class="message-detail">
-      <summary>detail</summary>
+  {#if item.detail && detailOpen}
+    <div class="message-detail" role="region" aria-label={`${toolLabel(item)} detail`}>
       <p>{item.detail}</p>
-    </details>
+    </div>
   {/if}
 </li>
 
@@ -306,48 +311,46 @@
   .tool-summary {
     display: flex;
     align-items: baseline;
-    gap: 0;
+    gap: 0.5rem;
     color: var(--text-muted);
     font-size: 0.88rem;
     font-weight: 750;
   }
 
   .tool-label {
-    flex: 0 0 auto;
-    color: var(--tui-cyan);
-    white-space: nowrap;
-  }
-
-  .tool-separator {
-    flex: 0 0 auto;
-    white-space: nowrap;
-  }
-
-  .tool-suffix {
     flex: 1 1 auto;
     min-width: 0;
-    overflow-wrap: anywhere;
+    overflow: hidden;
+    color: var(--tui-cyan);
+    text-overflow: ellipsis;
+    white-space: nowrap;
   }
 
-  .tool-separator,
-  .tool-suffix {
+  .tool-status {
+    flex: 0 0 auto;
     color: var(--tui-dark-gray);
+    font-size: 0.72rem;
+    white-space: nowrap;
   }
 
-  .tool-state-error .tool-suffix {
+  .tool-state-error .tool-status {
     color: var(--tui-red);
   }
 
-  .tool-state-running .tool-suffix,
-  .tool-state-streaming_args .tool-suffix,
-  .tool-state-pending .tool-suffix {
+  .tool-state-running .tool-status,
+  .tool-state-streaming_args .tool-status,
+  .tool-state-pending .tool-status {
     color: var(--tui-yellow);
   }
 
-  .tool-state-done .tool-suffix {
+  .tool-state-done .tool-status {
     color: var(--tui-dark-gray);
   }
 
+  .console-line.error-line .tool-status {
+    color: var(--tui-red);
+  }
+
   .message-heading {
     display: flex;
     align-items: center;
@@ -412,13 +415,47 @@
     color: var(--code);
   }
 
+  .tool-detail-button {
+    margin-inline-start: auto;
+    border: 1px solid var(--line);
+    border-radius: 0.35rem;
+    padding: 0.08rem 0.35rem;
+    background: var(--bg-raised);
+    color: var(--text-muted);
+    cursor: pointer;
+    font: inherit;
+    font-size: 0.68rem;
+    font-weight: 750;
+    opacity: 0;
+    pointer-events: none;
+    transition: opacity 120ms ease;
+  }
+
+  .console-line:hover .tool-detail-button,
+  .tool-detail-button:focus-visible,
+  .tool-detail-button[aria-expanded='true'] {
+    opacity: 1;
+    pointer-events: auto;
+  }
+
   .message-detail {
+    margin-top: 0.35rem;
+    border-left: 2px solid var(--line);
+    padding-left: 0.6rem;
     color: var(--text-muted);
     font-size: 0.84rem;
   }
 
-  .message-detail summary {
-    cursor: pointer;
-    font-weight: 800;
+  .message-detail p {
+    margin: 0;
+    overflow-wrap: anywhere;
+    white-space: pre-wrap;
+  }
+
+  @media (hover: none) {
+    .tool-detail-button {
+      opacity: 1;
+      pointer-events: auto;
+    }
   }
 </style>

web/workspace/src/lib/workspace/console/model.test.ts

@@ -313,26 +313,25 @@ Deno.test("projectConsole groups tool call lifecycle into one Call block", () =>
     !toolLines[0].streaming,
     "completed tool call should not remain streaming",
   );
-  assert(
-    toolLines[0].body.includes("$ pwd"),
-    "Bash command should be summarized",
-  );
+  assertEquals(toolLines[0].toolCallLabel, "Bash($ pwd)");
+  assertEquals(toolLines[0].toolStatus, "done");
   assert(
     toolLines[0].body.includes("/repo"),
     "tool result should be folded into the Call block",
   );
   assert(
     toolLines[0].body.includes("line9"),
-    "Bash result preview should include the ninth output line",
+    "Bash preview should include the ninth output line",
   );
   assert(
     !toolLines[0].body.includes("line10") &&
-      !toolLines[0].body.includes("line12"),
-    "Bash result preview should be capped at ten display lines",
+      toolLines[0].body.includes("… +3 more lines"),
+    "Bash preview should retain its line cap",
   );
   assert(
-    toolLines[0].body.includes("… +3 more lines"),
-    "Bash result preview should show omitted output count",
+    toolLines[0].expandedBody?.includes("line12") === true &&
+      !toolLines[0].expandedBody?.includes("more lines"),
+    "Bash detail should show every returned output line",
   );
   assert(
     toolLines[0].detail?.includes("id: call-1"),
@@ -421,7 +420,8 @@ Deno.test("projectConsole streams distinct Bash stdout and stderr through termin
   ]);
 
   const [line] = projection.lines.filter((line) => line.kind === "tool");
-  assert(line.body.includes("Bash — failed (exit 7)"), line.body);
+  assertEquals(line.toolCallLabel, "Bash($ long-command)");
+  assertEquals(line.toolStatus, "failed (exit 7)");
   assert(!line.body.includes("elapsed"), line.body);
   assert(!line.body.includes("stdout:"), line.body);
   assert(line.body.includes("ready\n"), line.body);
@@ -463,7 +463,8 @@ Deno.test("snapshot restores bounded in-flight Bash command output", () => {
 
   const projection = projectConsole([{ eventId: "snapshot-command", event: snapshot }]);
   const [line] = projection.lines.filter((line) => line.kind === "tool");
-  assert(line.body.includes("Bash — running…"), line.body);
+  assertEquals(line.toolCallLabel, "Bash($ slow)");
+  assertEquals(line.toolStatus, "running…");
   assert(!line.body.includes("elapsed"), line.body);
   assert(!line.body.includes("stdout:"), line.body);
   assert(line.body.includes("[… earlier stdout omitted]\ntail\n"), line.body);
@@ -474,7 +475,7 @@ Deno.test("snapshot restores bounded in-flight Bash command output", () => {
   assertEquals(line.streaming, true);
 });
 
-Deno.test("projectConsole caps default tool request and result previews", () => {
+Deno.test("projectConsole caps default preview but keeps complete detail body", () => {
   const projection = projectConsole([
     {
       eventId: "70",
@@ -508,19 +509,100 @@ Deno.test("projectConsole caps default tool request and result previews", () =>
 
   const [line] = projection.lines.filter((line) => line.kind === "tool");
   assertEquals(line.title, "Call · CustomTool");
-  assertEquals(line.body.split("\n").length, 7);
-  assert(line.body.includes("CustomTool — done"), "tool state should be shown");
+  assertEquals(line.toolCallLabel, 'CustomTool("first":"one","second":"two","third":"three","fourth":"four")');
+  assertEquals(line.toolStatus, "done");
+  assertEquals(line.body.split("\n").length, 3);
   assert(
-    line.body.includes('"first": "one"'),
-    "request preview should be shown",
+    line.body.includes("out1") && line.body.includes("… +3 more lines"),
+    "normal display should retain the capped response preview",
+  );
+  assert(!line.body.includes("first"), "request arguments should stay in the Call signature and detail");
+  assert(
+    line.detail?.includes("arguments:\nfirst: one") === true &&
+      line.detail?.includes("fourth: four") === true,
+    "detail metadata should render complete request arguments as YAML",
+  );
+  assert(
+    line.expandedBody?.includes("out5") === true &&
+      !line.expandedBody?.includes("more lines"),
+    "detail body should contain the complete result",
   );
-  assert(line.body.includes("out1"), "result preview should be shown");
-  assert(!line.body.includes("third"), "request preview should be capped");
-  assert(!line.body.includes("out3"), "result preview should be capped");
-  assert(line.body.includes("… +"), "overflow marker should be shown");
 });
 
-Deno.test("projectConsole shows Grep query and caps result preview to five entries", () => {
+Deno.test("projectConsole renders JSON tool responses as YAML", () => {
+  const projection = projectConsole([
+    {
+      eventId: "json-call",
+      event: {
+        event: "tool_call_done",
+        data: {
+          id: "json-tool",
+          name: "CustomTool",
+          arguments: "{}",
+        },
+      } satisfies Event,
+    },
+    {
+      eventId: "json-result",
+      event: {
+        event: "tool_result",
+        data: {
+          id: "json-tool",
+          summary: "json completed",
+          output: JSON.stringify({
+            status: "ok",
+            items: [{ id: 1 }, { id: 2 }],
+          }),
+          is_error: false,
+        },
+      } satisfies Event,
+    },
+    {
+      eventId: "invalid-json-call",
+      event: {
+        event: "tool_call_done",
+        data: {
+          id: "invalid-json-tool",
+          name: "CustomTool",
+          arguments: "{}",
+        },
+      } satisfies Event,
+    },
+    {
+      eventId: "invalid-json-result",
+      event: {
+        event: "tool_result",
+        data: {
+          id: "invalid-json-tool",
+          summary: "invalid json",
+          output: '{"status": broken}',
+          is_error: false,
+        },
+      } satisfies Event,
+    },
+  ]);
+
+  const toolLines = projection.lines.filter((line) => line.kind === "tool");
+  const jsonLine = toolLines.find((line) => line.id.includes("json-tool"));
+  const invalidLine = toolLines.find((line) => line.id.includes("invalid-json-tool"));
+  assert(jsonLine, "JSON tool line should be projected");
+  assert(invalidLine, "invalid JSON tool line should be projected");
+  assert(
+    jsonLine.expandedBody?.includes("status: ok") === true &&
+      jsonLine.expandedBody?.includes("  - id: 2") === true,
+    "detail body should serialize parsed JSON as YAML",
+  );
+  assert(
+    jsonLine.body.includes("more lines"),
+    "normal preview should cap the pretty-printed JSON",
+  );
+  assert(
+    invalidLine.expandedBody?.includes('{"status": broken}') === true,
+    "invalid JSON-looking output should remain unchanged",
+  );
+});
+
+Deno.test("projectConsole caps Grep preview but keeps complete detail body", () => {
   const projection = projectConsole([
     {
       eventId: "72",
@@ -549,17 +631,19 @@ Deno.test("projectConsole shows Grep query and caps result preview to five entri
 
   const [line] = projection.lines.filter((line) => line.kind === "tool");
   assertEquals(line.title, "Call · Grep");
-  assert(
-    line.body.includes("Grep — 6 matches"),
-    "Grep summary should be shown",
-  );
-  assert(line.body.includes("query: needle"), "Grep query should be shown");
+  assertEquals(line.toolCallLabel, "Grep(needle)");
+  assertEquals(line.toolStatus, "done");
   assert(line.body.includes("hit1"), "first result should be shown");
   assert(line.body.includes("hit5"), "fifth result should be shown");
-  assert(!line.body.includes("hit6"), "sixth result should be capped");
+  assert(!line.body.includes("hit6"), "normal preview should retain its result cap");
   assert(
     line.body.includes("… +1 more results"),
-    "overflow marker should be shown",
+    "preview should show the omitted result count",
+  );
+  assert(
+    line.expandedBody?.includes("hit6") === true &&
+      !line.expandedBody?.includes("more results"),
+    "detail body should show every Grep result",
   );
 });
 
@@ -594,17 +678,15 @@ Deno.test("projectConsole keeps Grep error detail in the body", () => {
 
   const [line] = projection.lines.filter((line) => line.kind === "tool");
   assertEquals(line.title, "Call · Grep");
-  assert(
-    line.body.includes("Grep — Failed"),
-    "error suffix should stay short",
-  );
+  assertEquals(line.toolCallLabel, "Grep(needle)");
+  assertEquals(line.toolStatus, "error");
   assert(
     line.body.includes(message),
     "error detail should remain visible in the body",
   );
   assert(
-    !line.body.includes(`Grep — ${message}`),
-    "error detail should not be repeated in the suffix",
+    !line.toolCallLabel?.includes(message),
+    "error detail should not be repeated in the Call signature",
   );
 });
 
@@ -879,9 +961,10 @@ Deno.test("projectConsole keeps streaming tool call updates in the same Call blo
   assertEquals(toolLines.length, 1);
   assertEquals(toolLines[0].title, "Call · Read");
   assert(toolLines[0].streaming, "streaming tool call should remain streaming");
+  assertEquals(toolLines[0].toolCallLabel, "Read(1 file)");
+  assertEquals(toolLines[0].toolStatus, "reading…");
   assert(
-    toolLines[0].body.includes("/tmp/a.md") &&
-      toolLines[0].body.includes("Read — reading"),
+    toolLines[0].body.includes("/tmp/a.md"),
     "Read call should render aggregate progress and path without content",
   );
 });
@@ -1018,10 +1101,8 @@ Deno.test("projectConsole aggregates Read calls without showing file content", (
   const toolLines = projection.lines.filter((line) => line.kind === "tool");
   assertEquals(toolLines.length, 1);
   assertEquals(toolLines[0].title, "Call · Read");
-  assert(
-    toolLines[0].body.includes("Read — 2 files read"),
-    "aggregate count should be shown",
-  );
+  assertEquals(toolLines[0].toolCallLabel, "Read(2 files)");
+  assertEquals(toolLines[0].toolStatus, "done");
   assert(
     toolLines[0].body.includes("/tmp/a.md"),
     "first path should be listed",
@@ -1073,7 +1154,9 @@ Deno.test("projectConsole renders Edit calls with structured diff lines", () =>
 
   const [line] = projection.lines.filter((line) => line.kind === "tool");
   assertEquals(line.title, "Call · Edit");
-  assert(line.body.includes("diff: -1 +2"), "diff summary should be shown");
+  assertEquals(line.toolCallLabel, "Edit(/tmp/a.md)");
+  assertEquals(line.toolStatus, "done");
+  assertEquals(line.body, "ok");
   assertEquals(line.diff?.map((row) => row.kind), [
     "context",
     "remove",
@@ -1242,13 +1325,13 @@ Deno.test("projectConsole renders snapshot entries and in-flight output", () =>
   assertEquals(projection.status, "running");
   assertEquals(
     projection.lines.map((line) =>
-      `${line.kind}:${line.body}:${line.streaming}`
+      `${line.kind}:${line.toolCallLabel ? `${line.toolCallLabel}\n${line.body}` : line.body}:${line.streaming}`
     ),
     [
       "user:seed user:false",
       "user:new user:false",
       "assistant:assistant reply:false",
-      "tool:Read — 1 file read\n  /tmp/a.md:false",
+      "tool:Read(1 file)\n  /tmp/a.md:false",
       "status:Compacting…:true",
       "in_flight:partial:true",
     ],
@@ -1476,25 +1559,26 @@ Deno.test("projectConsole relativizes known tool path displays from snapshot cwd
     },
   ]);
 
-  const bodies = projection.lines.filter((line) => line.kind === "tool").map((
-    line,
-  ) => line.body);
-  assertEquals(bodies[0], "Read — 1 file read\n  src/main.rs");
+  const toolLines = projection.lines.filter((line) => line.kind === "tool");
+  const bodies = toolLines.map((line) => line.body);
+  assertEquals(toolLines[0].toolCallLabel, "Read(1 file)");
+  assertEquals(bodies[0], "  src/main.rs");
   assert(
     projection.lines[0].detail?.includes("from src/main.rs"),
     "Read summary detail path should be relative",
   );
   assert(
-    bodies.some((body) =>
-      body.includes("Write — out.txt") && body.includes("Wrote out.txt")
+    toolLines.some((line) =>
+      line.toolCallLabel === "Write(out.txt)" && line.body.includes("Wrote out.txt")
     ),
-    "Write header and known result path should be relative",
+    "Write signature and known result path should be relative",
   );
   assert(
-    bodies.some((body) =>
-      body.includes("Edit — src/main.rs") && body.includes("Edited src/main.rs")
+    toolLines.some((line) =>
+      line.toolCallLabel === "Edit(src/main.rs)" &&
+      line.body.includes("Edited src/main.rs")
     ),
-    "Edit header and known result path should be relative",
+    "Edit signature and known result path should be relative",
   );
   assert(
     bodies.some((body) =>

web/workspace/src/lib/workspace/console/model.ts

@@ -11,6 +11,7 @@ import type {
   InternalWorkerSnapshot,
   Segment,
 } from "$lib/generated/protocol";
+import { stringify as stringifyYaml } from "yaml";
 import { workspaceRoute } from "$lib/workspace/api/http";
 import {
   applyRunActivityEvent,
@@ -86,6 +87,9 @@ export type ConsoleLine = {
   kind: ConsoleLineKind;
   title: string;
   body: string;
+  expandedBody?: string;
+  toolCallLabel?: string;
+  toolStatus?: string;
   detail?: string;
   compaction?: ConsoleCompaction;
   diff?: ConsoleDiffLine[];
@@ -1376,7 +1380,10 @@ function refreshedToolLine(item: ConsoleLine): ConsoleLine {
     title: item.title.startsWith("Call · Tool result")
       ? item.title
       : `Call · ${toolCall.name}`,
-    body: renderToolCall(toolCall),
+    body: renderToolResponse(toolCall),
+    expandedBody: renderToolResponse(toolCall, true),
+    toolCallLabel: toolCallSignature(toolCall),
+    toolStatus: toolCallStatus(toolCall),
     detail: toolCallDetail(toolCall),
     diff: toolCall.name === "Edit" ? editDiff(toolCall) : undefined,
     streaming: !["done", "error"].includes(toolCall.state) && !commandTerminal,
@@ -1384,7 +1391,7 @@ function refreshedToolLine(item: ConsoleLine): ConsoleLine {
   };
 }
 
-function renderToolCall(toolCall: ToolCallView): string {
+function renderToolResponse(toolCall: ToolCallView, expanded = false): string {
   switch (toolCall.name) {
     case "Read":
       return renderReadTool(toolCall);
@@ -1395,14 +1402,54 @@ function renderToolCall(toolCall: ToolCallView): string {
     case "Glob":
       return renderSearchTool(toolCall);
     case "Grep":
-      return renderGrepTool(toolCall);
+      return renderGrepTool(toolCall, expanded);
     case "Bash":
-      return renderBashTool(toolCall);
+      return renderBashTool(toolCall, expanded);
     default:
-      return renderDefaultTool(toolCall);
+      return renderDefaultTool(toolCall, expanded);
   }
 }
 
+function toolCallSignature(toolCall: ToolCallView): string {
+  const args = parsedArgs(toolCall);
+  switch (toolCall.name) {
+    case "Read":
+      return `Read(${readPath(toolCall)})`;
+    case "Write":
+    case "Edit": {
+      const path = displayPath(stringField(args, "file_path") ?? "?", toolCall.cwd);
+      return `${toolCall.name}(${path})`;
+    }
+    case "Glob":
+      return `Glob(${stringField(args, "pattern") ?? genericCallArguments(toolCall)})`;
+    case "Grep":
+      return `Grep(${stringField(args, "pattern") ?? genericCallArguments(toolCall)})`;
+    case "Bash": {
+      const command = stringField(args, "command");
+      return `Bash(${command ? `$ ${singleLine(command)}` : genericCallArguments(toolCall)})`;
+    }
+    default:
+      return `${toolCall.name}(${genericCallArguments(toolCall)})`;
+  }
+}
+
+function genericCallArguments(toolCall: ToolCallView): string {
+  const raw = toolCall.arguments ?? toolCall.argsStream;
+  if (!raw.trim()) return "";
+  const parsed = parseJson(raw);
+  if (parsed === undefined) return singleLine(raw);
+  const serialized = JSON.stringify(parsed) ?? "null";
+  return isRecord(parsed) ? serialized.slice(1, -1) : serialized;
+}
+
+function singleLine(value: string): string {
+  return value.replace(/\s+/g, " ").trim();
+}
+
+function toolCallStatus(toolCall: ToolCallView): string {
+  return toolCall.name === "Bash" ? commandStateSuffix(toolCall) : stateSuffix(toolCall.state);
+}
+
 function aggregateReadToolLines(lines: ConsoleLine[]): ConsoleLine[] {
   const result: ConsoleLine[] = [];
   let index = 0;
@@ -1434,9 +1481,6 @@ function readAggregateLine(group: ConsoleLine[]): ConsoleLine {
   const paths = calls.map(readPath);
   const visiblePaths = inProgress ? paths.slice(-3) : paths;
   const body = compactLines([
-    inProgress
-      ? `Read — reading (${count} file${plural(count)}…)`
-      : `Read — ${count} file${plural(count)} read`,
     visiblePaths.map((path) => `  ${path}`).join("\n"),
     inProgress && paths.length > visiblePaths.length
       ? `  … (${paths.length - visiblePaths.length} earlier)`
@@ -1447,6 +1491,8 @@ function readAggregateLine(group: ConsoleLine[]): ConsoleLine {
     kind: "tool",
     title: "Call · Read",
     body,
+    toolCallLabel: `Read(${count} file${plural(count)})`,
+    toolStatus: hasError ? "failed" : inProgress ? "reading…" : "done",
     detail: calls.map(readDetail).join("\n\n"),
     eventId: group.at(-1)?.eventId,
     source: "event",
@@ -1483,32 +1529,16 @@ function readDetail(toolCall: ToolCallView): string {
   ]);
 }
 
-function renderReadTool(toolCall: ToolCallView): string {
-  return `Read — ${readPath(toolCall)} (${stateSuffix(toolCall.state)})`;
+function renderReadTool(_toolCall: ToolCallView): string {
+  return "";
 }
 
 function renderWriteTool(toolCall: ToolCallView): string {
-  const args = parsedArgs(toolCall);
-  const path = displayPath(stringField(args, "file_path") ?? "?", toolCall.cwd);
-  const content = stringField(args, "content");
-  return compactLines([
-    `Write — ${path} (${stateSuffix(toolCall.state)})`,
-    cappedSection(content, 5),
-    knownToolResultText(toolCall),
-  ]);
+  return knownToolResultText(toolCall) ?? "";
 }
 
 function renderEditTool(toolCall: ToolCallView): string {
-  const args = parsedArgs(toolCall);
-  const path = displayPath(stringField(args, "file_path") ?? "?", toolCall.cwd);
-  const diff = editDiff(toolCall) ?? [];
-  const removes = diff.filter((line) => line.kind === "remove").length;
-  const adds = diff.filter((line) => line.kind === "add").length;
-  return compactLines([
-    `Edit — ${path} (${stateSuffix(toolCall.state)})`,
-    diff.length > 0 ? `diff: -${removes} +${adds}` : undefined,
-    knownToolResultText(toolCall),
-  ]);
+  return knownToolResultText(toolCall) ?? "";
 }
 
 function editDiff(toolCall: ToolCallView): ConsoleDiffLine[] | undefined {
@@ -1591,52 +1621,20 @@ function lcsTable(oldLines: string[], newLines: string[]): number[][] {
 }
 
 function renderSearchTool(toolCall: ToolCallView): string {
-  const summary = toolCall.summary?.trim();
-  return compactLines([
-    `${toolCall.name} — ${toolHeaderSuffix(toolCall, summary)}`,
-    knownToolResultText(toolCall),
-  ]);
+  return knownToolResultText(toolCall) ?? "";
 }
 
-function renderGrepTool(toolCall: ToolCallView): string {
-  const summary = toolCall.summary?.trim();
-  return compactLines([
-    `Grep — ${toolHeaderSuffix(toolCall, summary)}`,
-    grepQueryText(toolCall),
-    cappedResultSection(knownToolResultText(toolCall), 5),
-  ]);
+function renderGrepTool(toolCall: ToolCallView, expanded: boolean): string {
+  const result = knownToolResultText(toolCall);
+  return expanded ? result ?? "" : cappedResultSection(result, 5) ?? "";
 }
 
-function toolHeaderSuffix(
-  toolCall: ToolCallView,
-  summary?: string,
-): string {
-  if (toolCall.state === "error") {
-    return "Failed";
+function renderBashTool(toolCall: ToolCallView, expanded: boolean): string {
+  if (["done", "error"].includes(toolCall.state)) {
+    const result = resultText(toolCall);
+    return expanded ? result ?? "" : cappedDisplaySection(result, 10) ?? "";
   }
-  return summary ? firstLine(summary) : stateSuffix(toolCall.state);
-}
-
-function grepQueryText(toolCall: ToolCallView): string | undefined {
-  const args = parsedArgs(toolCall);
-  const pattern = stringField(args, "pattern");
-  if (pattern) {
-    return `query: ${pattern}`;
-  }
-  const renderedArgs = argsText(toolCall);
-  return renderedArgs ? `query:\n${renderedArgs}` : undefined;
-}
-
-function renderBashTool(toolCall: ToolCallView): string {
-  const args = parsedArgs(toolCall);
-  const command = stringField(args, "command");
-  return compactLines([
-    `Bash — ${commandStateSuffix(toolCall)}`,
-    command ? `$ ${command}` : argsText(toolCall),
-    ["done", "error"].includes(toolCall.state)
-      ? cappedDisplaySection(resultText(toolCall), 10)
-      : renderLiveCommandOutput(toolCall.command),
-  ]);
+  return renderLiveCommandOutput(toolCall.command) ?? "";
 }
 
 function commandStateSuffix(toolCall: ToolCallView): string {
@@ -1690,12 +1688,9 @@ function renderLiveCommandOutput(command?: CommandSnapshot): string | undefined
   ]);
 }
 
-function renderDefaultTool(toolCall: ToolCallView): string {
-  return compactLines([
-    `${toolCall.name} — ${stateSuffix(toolCall.state)}`,
-    cappedDisplaySection(argsText(toolCall), 3),
-    cappedDisplaySection(resultText(toolCall), 3),
-  ]);
+function renderDefaultTool(toolCall: ToolCallView, expanded: boolean): string {
+  const result = resultText(toolCall);
+  return expanded ? result ?? "" : cappedDisplaySection(result, 3) ?? "";
 }
 
 function toolCallDetail(toolCall: ToolCallView): string {
@@ -1713,10 +1708,30 @@ function toolCallDetail(toolCall: ToolCallView): string {
 }
 
 function resultText(toolCall: ToolCallView): string | undefined {
-  if (toolCall.output) {
-    return toolCall.output;
+  const text = toolCall.output || toolCall.summary;
+  return text ? formatJsonResponseAsYaml(text) : undefined;
+}
+
+function formatJsonResponseAsYaml(text: string): string {
+  const trimmed = text.trim();
+  if (
+    !(
+      (trimmed.startsWith("{") && trimmed.endsWith("}")) ||
+      (trimmed.startsWith("[") && trimmed.endsWith("]"))
+    )
+  ) {
+    return text;
+  }
+
+  try {
+    const parsed: unknown = JSON.parse(trimmed);
+    if (parsed === null || typeof parsed !== "object") {
+      return text;
+    }
+    return stringifyYaml(parsed).trimEnd();
+  } catch {
+    return text;
   }
-  return toolCall.summary;
 }
 
 function knownToolResultText(toolCall: ToolCallView): string | undefined {
@@ -1803,7 +1818,7 @@ function argsText(toolCall: ToolCallView): string {
     return "";
   }
   const parsed = parseJson(raw);
-  return parsed === undefined ? raw : jsonPreview(parsed);
+  return parsed === undefined ? raw : stringifyYaml(parsed).trimEnd();
 }
 
 function parsedArgs(
@@ -1830,21 +1845,6 @@ function compactLines(lines: Array<string | undefined | null | false>): string {
   return lines.filter((line): line is string => Boolean(line)).join("\n");
 }
 
-function cappedSection(
-  value: string | undefined,
-  cap: number,
-): string | undefined {
-  if (!value) {
-    return undefined;
-  }
-  const lines = value.split(/\r?\n/);
-  const shown = lines.slice(0, cap);
-  if (lines.length > cap) {
-    shown.push(`… +${lines.length - cap} more lines`);
-  }
-  return shown.join("\n");
-}
-
 function cappedDisplaySection(
   value: string | undefined,
   maxLines: number,

web/workspace/src/lib/workspace/console/worker-console.ui.test.ts

@@ -251,7 +251,8 @@ Deno.test("workspace Tickets surface provides Kanban and lifecycle controls", as
     ticketDetailLoad.includes("/repositories") &&
       ticketDetailPage.includes('mutate("state", "/state"') &&
       ticketDetailPage.includes("async function queueTicket") &&
-      ticketDetailPage.includes("`${ticketPath}/queue`") &&
+      ticketDetailPage.includes("const path = ticketPath") &&
+      ticketDetailPage.includes("`${path}/queue`") &&
       !ticketDetailPage.includes("/merge-request/merge") &&
       ticketDetailPage.includes("mergeRequest.selector_from") &&
       ticketDetailPage.includes("mergeRequest.review_status") &&
@@ -402,7 +403,7 @@ Deno.test("Worker Console renders markdown only for message rows", async () => {
       consoleLine.includes("item.kind === 'tool'") &&
       consoleLine.includes("{#if isBashTool(item)}") &&
       consoleLine.includes(
-        "<AnsiText text={bodyTextAfterToolSummary(item)} />",
+        "<AnsiText text={toolBodyText(item)} />",
       ) &&
       consoleLine.includes(
         ".console-line.tool-bash .console-plain-text",
@@ -419,6 +420,30 @@ Deno.test("Worker Console renders markdown only for message rows", async () => {
   );
 });
 
+Deno.test("Worker Console expands uncapped tool body from the hover detail action", async () => {
+  const consoleLine = await Deno.readTextFile(
+    new URL("./ConsoleLineItem.svelte", import.meta.url),
+  );
+
+  assert(
+    consoleLine.includes(
+      "return detailOpen ? (line.expandedBody ?? line.body) : line.body",
+    ) &&
+      consoleLine.includes("line.toolCallLabel ?? line.toolCall?.name") &&
+      consoleLine.includes('class={`tool-status') &&
+      consoleLine.includes('class="tool-detail-button"') &&
+      consoleLine.includes("aria-expanded={detailOpen}") &&
+      consoleLine.includes("detailOpen = !detailOpen") &&
+      consoleLine.includes("item.detail && detailOpen") &&
+      consoleLine.includes('role="region"') &&
+      consoleLine.includes(".console-line:hover .tool-detail-button") &&
+      consoleLine.includes(".tool-detail-button:focus-visible") &&
+      consoleLine.includes("@media (hover: none)") &&
+      !consoleLine.includes('<details class="message-detail">'),
+    "Normal tool display should keep its preview while detail reveals the uncapped body and existing metadata",
+  );
+});
+
 Deno.test("Worker Console renders Edit diffs without preformatted template gaps", async () => {
   const consoleLine = await Deno.readTextFile(
     new URL("./ConsoleLineItem.svelte", import.meta.url),

web/workspace/src/routes/w/[workspaceId]/tickets/[ticketId]/+page.svelte

@@ -36,7 +36,7 @@
   const initialData = untrack(() => data);
   const loadedTicket = initialData.ticket.data;
   if (!loadedTicket) throw new Error(initialData.ticket.error ?? "ticket load failed");
-  const loadedRepositories = initialData.repositories.data;
+  const loadedRepositories = $derived(data.repositories.data);
 
   type QueueOutcome = {
     requested_ticket: string;
@@ -62,6 +62,8 @@
   let manualRuntimeId = $state("");
   let manualWorkerId = $state("");
   let cancellationReason = $state("");
+  let routeTicketSnapshot = `${initialData.ticketId}:${loadedTicket.item_revision}`;
+  let routeGeneration = 0;
   const coderAssignment = $derived(
     ticket.assignments.find((assignment) => assignment.role === "coder") ?? null,
   );
@@ -95,13 +97,43 @@
 
   function applyTicket(updatedTicket: TicketDetail): void {
     ticket = updatedTicket;
-    editTitle = ticket.title;
-    editBody = ticket.body;
-    repositoryId = ticket.repository_id ?? "";
-    refSelector = ticket.ref_selector ?? "";
-    nextState = ticket.state;
+    editTitle = updatedTicket.title;
+    editBody = updatedTicket.body;
+    repositoryId = updatedTicket.repository_id ?? "";
+    refSelector = updatedTicket.ref_selector ?? "";
+    nextState = updatedTicket.state;
   }
 
+  function resetTicketView(updatedTicket: TicketDetail): void {
+    applyTicket(updatedTicket);
+    editing = false;
+    transitionReason = "";
+    threadRole = "comment";
+    threadBody = "";
+    resolution = "";
+    busy = null;
+    errorMessage = null;
+    queueMessage = null;
+    readyOperationKey = null;
+    manualRuntimeId = "";
+    manualWorkerId = "";
+    cancellationReason = "";
+  }
+
+  $effect(() => {
+    const incomingTicketId = data.ticketId;
+    const incomingTicket = data.ticket.data;
+    if (!incomingTicket) return;
+    const incomingSnapshot = `${incomingTicketId}:${incomingTicket.item_revision}`;
+
+    untrack(() => {
+      if (incomingSnapshot === routeTicketSnapshot) return;
+      routeTicketSnapshot = incomingSnapshot;
+      routeGeneration += 1;
+      resetTicketView(incomingTicket);
+    });
+  });
+
   async function mutate(
     action: string,
     suffix: string,
@@ -109,40 +141,51 @@
     method = "POST",
   ): Promise<boolean> {
     if (busy) return false;
+    const generation = routeGeneration;
+    const path = `${ticketPath}${suffix}`;
     busy = action;
     errorMessage = null;
     try {
-      const path = `${ticketPath}${suffix}`;
       const response = await workspaceApiJsonWithBody<TicketDetail>(path, {
         method,
         ...(body === undefined ? {} : { body: JSON.stringify(body) }),
       });
+      if (generation !== routeGeneration) return false;
       applyTicket(response);
       return true;
     } catch (error) {
+      if (generation === routeGeneration) {
         errorMessage = error instanceof Error ? error.message : String(error);
+      }
       return false;
     } finally {
-      busy = null;
+      if (generation === routeGeneration) busy = null;
     }
   }
 
   async function queueTicket(): Promise<void> {
     if (busy) return;
+    const generation = routeGeneration;
+    const path = ticketPath;
     busy = "queue";
     errorMessage = null;
     queueMessage = null;
     try {
       const outcome = await workspaceApiJsonWithBody<QueueOutcome>(
-        `${ticketPath}/queue`,
+        `${path}/queue`,
         { method: "POST", body: JSON.stringify({}) },
       );
+      if (generation !== routeGeneration) return;
+      const updatedTicket = await workspaceApiJson<TicketDetail>(path);
+      if (generation !== routeGeneration) return;
       queueMessage = `Queued ${outcome.queued_tickets.length} Ticket(s): ${outcome.queued_tickets.join(", ")}`;
-      applyTicket(await workspaceApiJson<TicketDetail>(ticketPath));
+      applyTicket(updatedTicket);
     } catch (error) {
+      if (generation === routeGeneration) {
         errorMessage = error instanceof Error ? error.message : String(error);
+      }
     } finally {
-      busy = null;
+      if (generation === routeGeneration) busy = null;
     }
   }
 
@@ -152,11 +195,13 @@
     principal: Record<string, string>,
   ): Promise<void> {
     if (busy) return;
+    const generation = routeGeneration;
+    const path = ticketPath;
     busy = action;
     errorMessage = null;
     try {
       await workspaceApiJsonWithBody(
-        `${ticketPath}/assignments/${role}`,
+        `${path}/assignments/${role}`,
         {
           method: "PUT",
           body: JSON.stringify({
@@ -166,11 +211,16 @@
           }),
         },
       );
-      applyTicket(await workspaceApiJson<TicketDetail>(ticketPath));
+      if (generation !== routeGeneration) return;
+      const updatedTicket = await workspaceApiJson<TicketDetail>(path);
+      if (generation !== routeGeneration) return;
+      applyTicket(updatedTicket);
     } catch (error) {
+      if (generation === routeGeneration) {
         errorMessage = error instanceof Error ? error.message : String(error);
+      }
     } finally {
-      busy = null;
+      if (generation === routeGeneration) busy = null;
     }
   }
 

web/workspace/test/ticket-detail-route-reuse.test.ts

@@ -0,0 +1,50 @@
+import { assert, assertStringIncludes } from "jsr:@std/assert";
+
+const pageSource = await Deno.readTextFile(
+  new URL(
+    "../src/routes/w/[workspaceId]/tickets/[ticketId]/+page.svelte",
+    import.meta.url,
+  ),
+);
+
+Deno.test("ticket detail synchronizes reused route data", () => {
+  const effectStart = pageSource.indexOf("$effect(() => {");
+  assert(effectStart >= 0, "ticket detail must react to reused route props");
+
+  const effectSource = pageSource.slice(effectStart);
+  for (
+    const token of [
+      "data.ticketId",
+      "data.ticket.data",
+      "incomingTicket.item_revision",
+      "routeGeneration += 1",
+      "resetTicketView(incomingTicket)",
+    ]
+  ) {
+    assertStringIncludes(effectSource, token);
+  }
+});
+
+Deno.test("ticket detail fences stale mutation responses", () => {
+  for (
+    const operation of [
+      "async function mutate(",
+      "async function queueTicket(",
+      "async function mutateAssignment(",
+    ]
+  ) {
+    const operationStart = pageSource.indexOf(operation);
+    assert(operationStart >= 0, `missing ${operation}`);
+    const nextOperation = pageSource.indexOf(
+      "\n  async function ",
+      operationStart + 1,
+    );
+    const operationSource = pageSource.slice(
+      operationStart,
+      nextOperation === -1 ? undefined : nextOperation,
+    );
+    assertStringIncludes(operationSource, "const generation = routeGeneration");
+    assertStringIncludes(operationSource, "generation !== routeGeneration");
+    assertStringIncludes(operationSource, "generation === routeGeneration");
+  }
+});