Route extension hints into context-window metadata (#40533)

## What changed

- Add a `ContextWindow` prompt slot for thread-scoped extension contributions.
- Include fragments in that slot in the token-budget context-window message.
- Keep the MCP `notes/thread_hint` bridge when the native history-notes extension is disabled, without falling back to it after a native request failure.

GitOrigin-RevId: befcb7cc8570b846094b2b542a112b186b39a30a

codex-rs/core/src/context/token_budget_context.rs

@@ -15,7 +15,7 @@ pub(crate) struct TokenBudgetContext {
     first_window_id: Uuid,
     previous_window_id: Option<Uuid>,
     window_id: Uuid,
-    mcp_result: Option<String>,
+    thread_hint: Option<String>,
 }
 
 impl TokenBudgetContext {
@@ -24,14 +24,14 @@ impl TokenBudgetContext {
         first_window_id: Uuid,
         previous_window_id: Option<Uuid>,
         window_id: Uuid,
-        mcp_result: Option<String>,
+        thread_hint: Option<String>,
     ) -> Self {
         Self {
             agent_path,
             first_window_id,
             previous_window_id,
             window_id,
-            mcp_result,
+            thread_hint,
         }
     }
 }
@@ -68,8 +68,8 @@ impl ContextualUserFragment for TokenBudgetContext {
         if let Some(previous_window_id) = self.previous_window_id {
             lines.push(format!("Previous context window id: {previous_window_id}"));
         }
-        if let Some(mcp_result) = &self.mcp_result {
-            lines.push(mcp_result.clone());
+        if let Some(thread_hint) = &self.thread_hint {
+            lines.push(thread_hint.clone());
         }
         format!("\n{}\n", lines.join("\n"))
     }

codex-rs/core/src/session/mod.rs

@@ -65,6 +65,7 @@ use codex_execpolicy::prefix_rule_migration;
 use codex_extension_api::ConversationHistorySnapshot;
 use codex_extension_api::ExtensionDataInit;
 use codex_extension_api::LoadedUserInstructions;
+use codex_extension_api::PromptSlot;
 use codex_extension_api::TurnContextContributionInput;
 use codex_features::FEATURES;
 use codex_features::Feature;
@@ -3590,6 +3591,7 @@ impl Session {
         let mut developer_sections = Vec::<RenderedFragment>::with_capacity(8);
         let mut contextual_user_sections = Vec::<RenderedFragment>::with_capacity(2);
         let mut separate_developer_sections = Vec::<RenderedFragment>::new();
+        let mut context_window_hints = Vec::new();
         let (session_source, auto_compact_window_ids) = {
             let state = self.state.lock().await;
             (
@@ -3649,7 +3651,14 @@ impl Session {
                 )
                 .await
             {
-                developer_sections.push(fragment.into());
+                match fragment.slot() {
+                    PromptSlot::ContextWindow => {
+                        context_window_hints.push(fragment.text().to_string());
+                    }
+                    PromptSlot::DeveloperPolicy | PromptSlot::DeveloperCapabilities => {
+                        developer_sections.push(fragment.into());
+                    }
+                }
             }
         }
         for contributor in &context_contributors {
@@ -3671,34 +3680,44 @@ impl Session {
         if turn_context.config.features.enabled(Feature::TokenBudget)
             && turn_context.model_context_window().is_some()
         {
-            let mcp_result = self
-                .services
-                .mcp_runtime
-                .latest_call_tool(
-                    "notes",
-                    "thread_hint",
-                    /*environment_id*/ None,
-                    /*arguments*/ None,
-                    Some(serde_json::json!({
-                        "threadId": self.thread_id().to_string(),
-                    })),
-                    /*requested_timeout*/ None,
-                    /*wait_for_server*/ true,
-                )
-                .await
-                .ok()
-                .and_then(|result| {
-                    let text = result
-                        .content
-                        .iter()
-                        .filter_map(|content| {
-                            content.get("text").and_then(serde_json::Value::as_str)
-                        })
-                        .filter(|text| !text.is_empty())
-                        .collect::<Vec<_>>()
-                        .join("\n");
-                    (!text.is_empty()).then_some(text)
-                });
+            // Keep the legacy bridge hint when native Notes is disabled. A failed
+            // native request must not fall back to the bridge.
+            if !turn_context
+                .config
+                .token_budget
+                .as_ref()
+                .is_some_and(|config| config.use_history_notes_extension)
+                && let Some(mcp_result) = self
+                    .services
+                    .mcp_runtime
+                    .latest_call_tool(
+                        "notes",
+                        "thread_hint",
+                        /*environment_id*/ None,
+                        /*arguments*/ None,
+                        Some(serde_json::json!({
+                            "threadId": self.thread_id().to_string(),
+                        })),
+                        /*requested_timeout*/ None,
+                        /*wait_for_server*/ true,
+                    )
+                    .await
+                    .ok()
+                    .and_then(|result| {
+                        let text = result
+                            .content
+                            .iter()
+                            .filter_map(|content| {
+                                content.get("text").and_then(serde_json::Value::as_str)
+                            })
+                            .filter(|text| !text.is_empty())
+                            .collect::<Vec<_>>()
+                            .join("\n");
+                        (!text.is_empty()).then_some(text)
+                    })
+            {
+                context_window_hints.push(mcp_result);
+            }
             separate_developer_sections.push(
                 crate::context::TokenBudgetContext::new(
                     session_source
@@ -3707,7 +3726,7 @@ impl Session {
                     auto_compact_window_ids.first_window_id,
                     auto_compact_window_ids.previous_window_id,
                     auto_compact_window_ids.window_id,
-                    mcp_result,
+                    (!context_window_hints.is_empty()).then(|| context_window_hints.join("\n")),
                 )
                 .render_fragment(),
             );

codex-rs/core/src/session/world_state.rs

@@ -116,7 +116,7 @@ impl Session {
                 window_ids.first_window_id,
                 window_ids.previous_window_id,
                 window_ids.window_id,
-                /*mcp_result*/ None,
+                /*thread_hint*/ None,
             ));
             if let Some(guidance) = turn_context
                 .config

codex-rs/ext/extension-api/src/contributors/prompt.rs

@@ -8,6 +8,8 @@ use codex_protocol::models::ContentItemKind;
 pub enum PromptSlot {
     DeveloperPolicy,
     DeveloperCapabilities,
+    /// Text inside the context-window message, supplied by `contribute_thread_context`.
+    ContextWindow,
 }
 
 #[derive(Clone, Debug, PartialEq, Eq)]