[codex] Harden overflow auto-compaction recovery by aibrahim-oai · Pull Request #22141 · openai/codex

aibrahim-oai · 2026-05-11T12:05:22Z

Why

Dogfooder feedback exposed two correctness gaps in normal-loop overflow recovery:

a sampling request that hit ContextWindowExceeded could keep re-entering auto-compaction indefinitely if the compacted retry still did not fit, and
compacted user-history rebuilds flattened user messages down to text, so an overflowing multimodal turn such as [image, "what is this?"] could be retried without the image after compaction.

That means recovery could either fail to terminate cleanly or continue with a materially weakened version of the user request.

What changed

Move normal-loop ContextWindowExceeded handling into the sampling retry loop, so successful rescue compaction consumes the provider retry budget instead of creating an unbounded outer-turn loop.
Keep compacted user history structured: collect_user_messages now returns user UserInput content rather than flattened strings, and build_compacted_history reconstructs full user messages from that structured representation.
Preserve image inputs through compacted-history rebuilds while retaining the existing text-budget truncation behavior for user text.
Reuse the shared user_message_text projection where text-only analysis is still needed, such as connector mention extraction.
Preserve existing compaction-task failure handling and client-session reset behavior while bounding repeated overflow retries.

Coverage

context_window_error_sets_total_tokens_to_model_window_before_auto_compact_recovery
normal_loop_context_window_error_auto_compacts_and_resumes_turn
normal_loop_context_window_error_auto_remote_compacts_and_resumes_turn
normal_loop_context_window_error_auto_remote_v2_compacts_and_preserves_user_turn
normal_loop_context_window_error_stops_after_sample_retry_budget
normal_loop_context_window_error_stops_after_remote_compaction_failure
collect_user_messages_preserves_user_content

Verification

Not run locally; relying on PR CI for this update.

Recover regular turns that overflow the model context by reusing the existing mid-turn auto-compaction flow and retrying the turn. Co-authored-by: Codex <noreply@openai.com>

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: a175208c81

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

chatgpt-codex-connector · 2026-05-11T12:17:22Z

+                    client_session.reset_websocket_session();
+                }
+                can_drain_pending_input = false;
+                continue;


Guard context-window recovery retries

This branch retries unconditionally after every successful auto-compaction. If the compacted history still exceeds the model window (for example an oversized current prompt/user input or a remote compact result that is still too large), the next sampling request raises ContextWindowExceeded again and the loop re-enters compaction forever instead of surfacing the error. Add a per-turn retry/fit guard.

Useful? React with 👍 / 👎.

jif-oai

lgtm after those 2

jif-oai · 2026-05-11T12:25:15Z

+                    client_session.reset_websocket_session();
+                }
+                can_drain_pending_input = false;
+                continue;


jif-oai · 2026-05-11T12:25:45Z

                // Aborted turn is reported via a different event.
                break;
            }
+            Err(CodexErr::ContextWindowExceeded) => {


I think this can mutate the current turn before the model has ever seen it. On the local path, compaction rebuilds user history from flattened text only and can truncate it, so an overflowing turn with a long prompt or an image can be retried as a different prompt. We probably need to preserve/re-append the exact in-flight input here, not recover it from compacted history

Count overflow rescue compactions against the existing sampling retry budget and preserve the exact current user turn across compaction-backed retries. Co-authored-by: Codex <noreply@openai.com>

This reverts commit 1482455.

aibrahim-oai · 2026-05-11T13:55:38Z

@codex review this

chatgpt-codex-connector

💡 Codex Review

codex/codex-rs/core/src/compact.rs

Line 265 in b20b795

    
           let mut new_history = build_compacted_history(Vec::new(), &user_messages, &summary_text);

Do not rely on compacted text for the current turn

The recovery path installs build_compacted_history(...) directly after overflow. That helper applies the 20k-token compacted-history cap to all user messages, including the just-submitted overflowing turn, so a large current prompt can be truncated before the retry. The commit says the exact current turn is preserved, but this code never exempts it from compaction.

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

chatgpt-codex-connector · 2026-05-11T14:08:21Z

+                if reset_client_session {
+                    client_session.reset_websocket_session();
+                }
+                continue;


Restore the overflowing user turn before retrying

After this continue, the retry is built from compacted history, but the original input was already consumed and is never reinstalled. Local/remote compaction rebuilds history from text summaries, so an overflowing turn with images or text beyond the compaction limit can be retried with content dropped or truncated before the model ever sees it.

Useful? React with 👍 / 👎.

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: c3a51fb7c2

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

chatgpt-codex-connector · 2026-05-11T14:50:11Z

                .await
-                .for_prompt(&turn_context.model_info.input_modalities)
+                .for_prompt(&turn_context.model_info.input_modalities);
+            restore_current_turn_for_sampling_retry(prompt_input, current_turn_input.as_ref())


Preserve the restored turn for follow-up sampling

The restored user input is only spliced into this one retry prompt; sess history remains the compacted/truncated version. If the recovered request returns a tool call or otherwise needs a follow-up, the outer loop rebuilds the next prompt from session history, so images or oversized text from the overflowing turn disappear before the final model call.

Useful? React with 👍 / 👎.

chatgpt-codex-connector · 2026-05-11T14:50:12Z

+fn latest_real_user_turn(input: &[ResponseItem]) -> Option<ResponseItem> {
+    input
+        .iter()
+        .rev()
+        .find_map(|item| match parse_turn_item(item) {
+            Some(TurnItem::UserMessage(user)) if !is_summary_message(&user.message()) => {
+                Some(item.clone())
+            }
+            _ => None,
+        })
+}


Restore every unseen user message before retrying

latest_real_user_turn preserves only one user ResponseItem. A sampling request can contain multiple not-yet-seen user messages, e.g. queued pending inputs accepted before the next model call; after overflow compaction, earlier pending messages are rebuilt from text summaries while only the last one gets its images/full text restored.

Useful? React with 👍 / 👎.

chatgpt-codex-connector · 2026-05-11T14:50:12Z

+fn latest_real_user_turn(input: &[ResponseItem]) -> Option<ResponseItem> {
+    input
+        .iter()
+        .rev()
+        .find_map(|item| match parse_turn_item(item) {
+            Some(TurnItem::UserMessage(user)) if !is_summary_message(&user.message()) => {
+                Some(item.clone())
+            }
+            _ => None,
+        })
+}


Preserve pending tool outputs across overflow recovery

When the overflow happens on a follow-up request after a tool call, the new content the model has not seen is the function/tool output. Compaction drops tool artifacts, but this preservation path only saves the last user message, so the immediate retry can ask the model to continue without the tool result it requested.

Useful? React with 👍 / 👎.

aibrahim-oai · 2026-05-11T15:08:52Z

@codex review this

chatgpt-codex-connector

💡 Codex Review

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

Lines 1060 to 1063 in 2aa791e

    
           let prompt = build_prompt( 
        
               prompt_input, 
        
               router.as_ref(), 
        
               turn_context.as_ref(),

Rebuild tool routing after recovery compaction

router is built once from the overflowing input before the loop, but after run_auto_compact the retry prompt is rebuilt from compacted history while build_prompt still uses the stale router. If compaction drops or truncates connector mentions, the recovered request can expose tools that are no longer present in the prompt (or miss changes from the compacted prompt).

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

Line 585 in 2aa791e

input_messages: sampling_request_input_messages,

Pass hooks the recovered request input

sampling_request_input_messages is captured before run_sampling_request; this change lets that call compact history and retry internally. On a successful recovered response, AfterAgent hooks still receive the pre-compaction messages here, so hooks can validate or log a prompt that is not the one the model actually answered.

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

chatgpt-codex-connector · 2026-05-11T15:17:18Z

+            UserInput::Text { .. } if !text_inserted => {
+                truncated.push(UserInput::Text {
+                    text: truncated_text.clone(),


Preserve text segment order when truncating user messages

When a user message has multiple text segments around images, user_message_text concatenates all text and this branch inserts the combined truncation into the first text item while dropping later text. A turn like [text A, image, text B] can be retried as [A+B prefix, image], moving captions/questions before their image and defeating structured-history preservation.

Useful? React with 👍 / 👎.

chatgpt-codex-connector · 2026-05-11T15:17:18Z

+            let message_text = user_message_text(message);
+            let tokens = approx_token_count(&message_text);
            if tokens <= remaining {
                selected_messages.push(message.clone());


Count preserved images against the compaction budget

The selection budget only counts text tokens, but the new structured path keeps every non-text item. Image-only messages have tokens == 0, so the reverse scan can preserve an unbounded number of historical images without reducing remaining, and mixed messages keep all images even when text is truncated. Image-heavy threads can still overflow after “successful” compaction.

Useful? React with 👍 / 👎.

aibrahim-oai · 2026-05-11T15:48:56Z

@codex review this

chatgpt-codex-connector · 2026-05-11T15:55:28Z

Codex Review: Didn't find any major issues. Bravo.

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

pakrym-oai · 2026-05-11T16:15:59Z

@@ -310,7 +311,7 @@ pub(crate) async fn run_turn(
        Vec::new()
    } else {
        let initial_input_for_turn: ResponseInputItem = ResponseInputItem::from(input.clone());
-        let response_item: ResponseItem = initial_input_for_turn.clone().into();
+        let response_item: ResponseItem = initial_input_for_turn.into();
        let user_prompt_submit_outcome = run_user_prompt_submit_hooks(
            &sess,
            &turn_context,
@@ -918,7 +919,11 @@ pub(super) fn filter_connectors_for_input(
        return Vec::new();
    }

-    let mentions = collect_tool_mentions_from_messages(&user_messages);
+    let user_message_texts = user_messages
+        .iter()
+        .map(|message| user_message_text(message))
+        .collect::<Vec<_>>();
+    let mentions = collect_tool_mentions_from_messages(&user_message_texts);
    let mention_names_lower = mentions
        .plain_names
        .iter()
@@ -1041,6 +1046,7 @@ async fn run_sampling_request(
            Arc::clone(&turn_diff_tracker),
        )
        .await;
+    let max_retries = turn_context.provider.info().stream_max_retries();
    let mut retries = 0;
    let mut initial_input = Some(input);
    loop {
@@ -1074,7 +1080,27 @@ async fn run_sampling_request(
            }
            Err(CodexErr::ContextWindowExceeded) => {
                sess.set_total_tokens_full(&turn_context).await;
-                return Err(CodexErr::ContextWindowExceeded);
+                if retries >= max_retries {
+                    return Err(CodexErr::ContextWindowExceeded);
+                }
+                retries += 1;
+                let reset_client_session = match run_auto_compact(
+                    &sess,
+                    &turn_context,
+                    client_session,
+                    InitialContextInjection::BeforeLastUserMessage,
+                    CompactionReason::ContextLimit,
+                    CompactionPhase::MidTurn,
+                )
+                .await
+                {
+                    Ok(reset_client_session) => reset_client_session,
+                    Err(_) => return Err(CodexErr::TurnAborted),
+                };
+                if reset_client_session {
+                    client_session.reset_websocket_session();


you don't need to do this.

core: auto-compact on sampling context overflow

a175208

Recover regular turns that overflow the model context by reusing the existing mid-turn auto-compaction flow and retrying the turn. Co-authored-by: Codex <noreply@openai.com>

aibrahim-oai marked this pull request as ready for review May 11, 2026 12:09

aibrahim-oai requested a review from a team as a code owner May 11, 2026 12:09

chatgpt-codex-connector Bot reviewed May 11, 2026

View reviewed changes

jif-oai approved these changes May 11, 2026

View reviewed changes

aibrahim-oai changed the title ~~[codex] Recover sampling overflow with auto-compaction~~ [codex] Harden overflow auto-compaction recovery May 11, 2026

aibrahim-oai and others added 4 commits May 11, 2026 13:07

core: harden overflow compaction retries

1482455

Count overflow rescue compactions against the existing sampling retry budget and preserve the exact current user turn across compaction-backed retries. Co-authored-by: Codex <noreply@openai.com>

Revert "core: harden overflow compaction retries"

c568bf1

This reverts commit 1482455.

core: bound overflow compaction retries in sampling

a568d99

core: keep sampling recovery self contained

b20b795

chatgpt-codex-connector Bot reviewed May 11, 2026

View reviewed changes

aibrahim-oai added 2 commits May 11, 2026 17:12

core: restore overflow input after compact retry

6c445f6

core: restore retry input in sampling

c3a51fb

chatgpt-codex-connector Bot reviewed May 11, 2026

View reviewed changes

aibrahim-oai added 3 commits May 11, 2026 17:58

core: preserve images in compacted user history

fb557f6

core: reuse compacted user text helper

4d17876

core: keep empty compacted user text

4b691eb

aibrahim-oai enabled auto-merge (squash) May 11, 2026 15:04

Merge branch 'main' into dev/auto-compact-context-overflow

2aa791e

chatgpt-codex-connector Bot reviewed May 11, 2026

View reviewed changes

aibrahim-oai added 2 commits May 11, 2026 18:25

core: preserve compacted text segment order

c2bfd5e

core: fix compact truncation test clone

77769ef

pakrym-oai reviewed May 11, 2026

View reviewed changes

aibrahim-oai merged commit 15e79f3 into main May 11, 2026
37 of 38 checks passed

aibrahim-oai deleted the dev/auto-compact-context-overflow branch May 11, 2026 16:16

github-actions Bot locked and limited conversation to collaborators May 11, 2026

	let prompt = build_prompt(
	prompt_input,
	router.as_ref(),
	turn_context.as_ref(),

Conversation

aibrahim-oai commented May 11, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Why

What changed

Coverage

Verification

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

chatgpt-codex-connector Bot May 11, 2026

Choose a reason for hiding this comment

Uh oh!

jif-oai May 11, 2026

Choose a reason for hiding this comment

Uh oh!

jif-oai left a comment

Choose a reason for hiding this comment

Uh oh!

jif-oai May 11, 2026

Choose a reason for hiding this comment

Uh oh!

jif-oai May 11, 2026

Choose a reason for hiding this comment

Uh oh!

aibrahim-oai commented May 11, 2026

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

chatgpt-codex-connector Bot May 11, 2026

Choose a reason for hiding this comment

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

chatgpt-codex-connector Bot May 11, 2026

Choose a reason for hiding this comment

Uh oh!

chatgpt-codex-connector Bot May 11, 2026

Choose a reason for hiding this comment

Uh oh!

chatgpt-codex-connector Bot May 11, 2026

Choose a reason for hiding this comment

Uh oh!

aibrahim-oai commented May 11, 2026

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

chatgpt-codex-connector Bot May 11, 2026

Choose a reason for hiding this comment

Uh oh!

chatgpt-codex-connector Bot May 11, 2026

Choose a reason for hiding this comment

Uh oh!

aibrahim-oai commented May 11, 2026

Uh oh!

chatgpt-codex-connector Bot commented May 11, 2026

Uh oh!

pakrym-oai May 11, 2026

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

aibrahim-oai commented May 11, 2026 •

edited

Loading