🔄 fix: resume Anthropic context-window stops

[upman]

Summary

Resume once per agent/run after Anthropic's model_context_window_exceeded stop, preserving partial output and encrypted server-search results. Keep the raw streaming stop reason; pending client tools retain their normal route. A repeated stop sets the existing output_truncated halt reason.

context stop → retain history + resume prompt → model
                                               ├─ answer → finish
                                               └─ context stop → halt

Claude Code behavior

Local Claude Code 2.1.268 tests showed top-level context stops trigger a resume request with unchanged history, not compaction. Its nested WebSearch path instead returns links to the outer model. This follows the top-level continuation pattern, with a one-continuation safety bound and search history retained.

Validation

Check	Result
Regression tests	292 passed, including streaming and tracing
Local tarball in LibreChat	Normal response: 1 call; context stop: 2 calls, persisted answer and aggregate usage; repeated stop: bounded at 2
Gateway + Bedrock	Replayed a real context-stopped search response; LibreChat continued with encrypted history and received a coherent cited answer without another search

The Bedrock continuation used a full-capacity model context: this verifies replay, not compaction of a still-full window. The tested LibreChat build does not yet map output_truncated to its persisted unfinished flag.

[upman]

@codex review

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-09-16T21:19:04.213554Z	e9dc340	Manual request

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" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 976536a381

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".

[upman]

@codex re-review

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 1a99832eff

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".

[upman]

@codex re-review

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: e9dc340af1

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]

Persist context-stop provenance across durable resumes

When the continuation produces an approval-gated tool call and the host recreates Run before resuming its checkpoint, this in-memory map is empty because it is absent from RunStepResumeState. The post-tool assistant message therefore loses contextStopContinuationOf, causing the Langfuse root output to omit the pre-interrupt segments; another context-window stop can also receive a second billable retry despite the one-per-run contract. Persist and restore this map with the other process-safe resume state.

AGENTS.md reference: AGENTS.md:L144-L144

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

Scope the preemption guard to the current agent

In a parallel MultiAgentGraph, if one agent first seals without a boundary injection, it sets the graph-wide preemptIncomplete flag; a different agent that then returns model_context_window_exceeded fails this condition and never receives its continuation. The result depends on branch completion order, so this guard should check whether the current response/agent was preempted rather than the shared aggregate flag.

Useful? React with 👍 / 👎.