Fix remote compaction estimator/payload instruction small mismatch (openai#10692)

## Summary
This PR fixes a deterministic mismatch in remote compaction where
pre-trim estimation and the `/v1/responses/compact` payload could use
different base instructions.

Before this change:
- pre-trim estimation used model-derived instructions
(`model_info.get_model_instructions(...)`)
- compact payload used session base instructions
(`sess.get_base_instructions()`)

After this change:
- remote pre-trim estimation and compact payload both use the same
`BaseInstructions` instance from session state.

## Changes
- Added a shared estimator entry point in `ContextManager`:
- `estimate_token_count_with_base_instructions(&self, base_instructions:
&BaseInstructions) -> Option<i64>`
- Kept `estimate_token_count(&TurnContext)` as a thin wrapper that
resolves model/personality instructions and delegates to the new helper.
- Updated remote compaction flow to fetch base instructions once and
reuse it for both:
  - trim preflight estimation
  - compact request payload construction
- Added regression coverage for parity and behavior:
  - unit test verifying explicit-base estimator behavior
- integration test proving remote compaction uses session override
instructions and trims accordingly

## Why this matters
This removes a deterministic divergence source where pre-trim could
think the request fits while the actual compact request exceeded context
because its instructions were longer/different.

## Scope
In scope:
- estimator/payload base-instructions parity in remote compaction

Out of scope:
- retry-on-`context_length_exceeded`
- compaction threshold/headroom policy changes
- broader trimming policy changes

## Codex author:
`codex fork 019c2b24-c2df-7b31-a482-fb8cf7a28559`

Author: charley-oai

codex-rs/core/src/codex.rs

@@ -2117,10 +2117,10 @@ impl Session {
     }
 
     pub(crate) async fn recompute_token_usage(&self, turn_context: &TurnContext) {
-        let Some(estimated_total_tokens) = self
-            .clone_history()
-            .await
-            .estimate_token_count(turn_context)
+        let history = self.clone_history().await;
+        let base_instructions = self.get_base_instructions().await;
+        let Some(estimated_total_tokens) =
+            history.estimate_token_count_with_base_instructions(&base_instructions)
         else {
             return;
         };
@@ -4782,6 +4782,7 @@ mod tests {
     use crate::turn_diff_tracker::TurnDiffTracker;
     use codex_app_server_protocol::AppInfo;
     use codex_app_server_protocol::AuthMode;
+    use codex_protocol::models::BaseInstructions;
     use codex_protocol::models::ContentItem;
     use codex_protocol::models::ResponseItem;
     use std::path::Path;
@@ -5061,6 +5062,46 @@ mod tests {
         assert_eq!(actual, Some(info2));
     }
 
+    #[tokio::test]
+    async fn recompute_token_usage_uses_session_base_instructions() {
+        let (session, turn_context) = make_session_and_context().await;
+
+        let override_instructions = "SESSION_OVERRIDE_INSTRUCTIONS_ONLY".repeat(120);
+        {
+            let mut state = session.state.lock().await;
+            state.session_configuration.base_instructions = override_instructions.clone();
+        }
+
+        let item = user_message("hello");
+        session
+            .record_into_history(std::slice::from_ref(&item), &turn_context)
+            .await;
+
+        let history = session.clone_history().await;
+        let session_base_instructions = BaseInstructions {
+            text: override_instructions,
+        };
+        let expected_tokens = history
+            .estimate_token_count_with_base_instructions(&session_base_instructions)
+            .expect("estimate with session base instructions");
+        let model_estimated_tokens = history
+            .estimate_token_count(&turn_context)
+            .expect("estimate with model instructions");
+        assert_ne!(expected_tokens, model_estimated_tokens);
+
+        session.recompute_token_usage(&turn_context).await;
+
+        let actual_tokens = session
+            .state
+            .lock()
+            .await
+            .token_info()
+            .expect("token info")
+            .last_token_usage
+            .total_tokens;
+        assert_eq!(actual_tokens, expected_tokens.max(0));
+    }
+
     #[tokio::test]
     async fn record_initial_history_reconstructs_forked_transcript() {
         let (session, turn_context) = make_session_and_context().await;

codex-rs/core/src/compact_remote.rs

@@ -12,6 +12,7 @@ use crate::protocol::RolloutItem;
 use crate::protocol::TurnStartedEvent;
 use codex_protocol::items::ContextCompactionItem;
 use codex_protocol::items::TurnItem;
+use codex_protocol::models::BaseInstructions;
 use codex_protocol::models::ResponseItem;
 use tracing::info;
 
@@ -49,8 +50,12 @@ async fn run_remote_compact_task_inner_impl(
     sess.emit_turn_item_started(turn_context, &compaction_item)
         .await;
     let mut history = sess.clone_history().await;
-    let deleted_items =
-        trim_function_call_history_to_fit_context_window(&mut history, turn_context.as_ref());
+    let base_instructions = sess.get_base_instructions().await;
+    let deleted_items = trim_function_call_history_to_fit_context_window(
+        &mut history,
+        turn_context.as_ref(),
+        &base_instructions,
+    );
     if deleted_items > 0 {
         info!(
             turn_id = %turn_context.sub_id,
@@ -71,7 +76,7 @@ async fn run_remote_compact_task_inner_impl(
         input: history.for_prompt(),
         tools: vec![],
         parallel_tool_calls: false,
-        base_instructions: sess.get_base_instructions().await,
+        base_instructions,
         personality: turn_context.personality,
         output_schema: None,
     };
@@ -107,14 +112,15 @@ async fn run_remote_compact_task_inner_impl(
 fn trim_function_call_history_to_fit_context_window(
     history: &mut ContextManager,
     turn_context: &TurnContext,
+    base_instructions: &BaseInstructions,
 ) -> usize {
     let mut deleted_items = 0usize;
     let Some(context_window) = turn_context.model_context_window() else {
         return deleted_items;
     };
 
     while history
-        .estimate_token_count(turn_context)
+        .estimate_token_count_with_base_instructions(base_instructions)
         .is_some_and(|estimated_tokens| estimated_tokens > context_window)
     {
         let Some(last_item) = history.raw_items().last() else {

codex-rs/core/src/context_manager/history.rs

@@ -9,6 +9,7 @@ use crate::truncate::approx_tokens_from_byte_count;
 use crate::truncate::truncate_function_output_items_with_policy;
 use crate::truncate::truncate_text;
 use crate::user_shell_command::is_user_shell_command_text;
+use codex_protocol::models::BaseInstructions;
 use codex_protocol::models::ContentItem;
 use codex_protocol::models::FunctionCallOutputBody;
 use codex_protocol::models::FunctionCallOutputContentItem;
@@ -88,8 +89,18 @@ impl ContextManager {
     pub(crate) fn estimate_token_count(&self, turn_context: &TurnContext) -> Option<i64> {
         let model_info = &turn_context.model_info;
         let personality = turn_context.personality.or(turn_context.config.personality);
-        let base_instructions = model_info.get_model_instructions(personality);
-        let base_tokens = i64::try_from(approx_token_count(&base_instructions)).unwrap_or(i64::MAX);
+        let base_instructions = BaseInstructions {
+            text: model_info.get_model_instructions(personality),
+        };
+        self.estimate_token_count_with_base_instructions(&base_instructions)
+    }
+
+    pub(crate) fn estimate_token_count_with_base_instructions(
+        &self,
+        base_instructions: &BaseInstructions,
+    ) -> Option<i64> {
+        let base_tokens =
+            i64::try_from(approx_token_count(&base_instructions.text)).unwrap_or(i64::MAX);
 
         let items_tokens = self.items.iter().fold(0i64, |acc, item| {
             acc.saturating_add(estimate_item_token_count(item))

codex-rs/core/src/context_manager/history_tests.rs

@@ -2,6 +2,7 @@ use super::*;
 use crate::truncate;
 use crate::truncate::TruncationPolicy;
 use codex_git::GhostCommit;
+use codex_protocol::models::BaseInstructions;
 use codex_protocol::models::ContentItem;
 use codex_protocol::models::FunctionCallOutputBody;
 use codex_protocol::models::FunctionCallOutputContentItem;
@@ -103,6 +104,10 @@ fn truncate_exec_output(content: &str) -> String {
     truncate::truncate_text(content, TruncationPolicy::Tokens(EXEC_FORMAT_MAX_TOKENS))
 }
 
+fn approx_token_count_for_text(text: &str) -> i64 {
+    i64::try_from(text.len().saturating_add(3) / 4).unwrap_or(i64::MAX)
+}
+
 #[test]
 fn filters_non_api_messages() {
     let mut h = ContextManager::default();
@@ -250,6 +255,28 @@ fn get_history_for_prompt_drops_ghost_commits() {
     assert_eq!(filtered, vec![]);
 }
 
+#[test]
+fn estimate_token_count_with_base_instructions_uses_provided_text() {
+    let history = create_history_with_items(vec![assistant_msg("hello from history")]);
+    let short_base = BaseInstructions {
+        text: "short".to_string(),
+    };
+    let long_base = BaseInstructions {
+        text: "x".repeat(1_000),
+    };
+
+    let short_estimate = history
+        .estimate_token_count_with_base_instructions(&short_base)
+        .expect("token estimate");
+    let long_estimate = history
+        .estimate_token_count_with_base_instructions(&long_base)
+        .expect("token estimate");
+
+    let expected_delta = approx_token_count_for_text(&long_base.text)
+        - approx_token_count_for_text(&short_base.text);
+    assert_eq!(long_estimate - short_estimate, expected_delta);
+}
+
 #[test]
 fn remove_first_item_removes_matching_output_for_function_call() {
     let items = vec![