Refocus compaction summaries on the active task

[ReyNeill]

Problem

When auto-compaction triggers mid-implementation, the summary plus preserved user messages can cause the model to re-check the entire session instead of continuing the current task.

Solution

Refocus the compaction prompt and summary prefix so the summary targets only the active task and explicitly avoids re-opening completed work unless the user asks.

Notes / Future options

• Keep only the most recent N user messages after compaction
• Store both an active-task summary and a full-session summary, reintroducing the latter only if needed
• Use retrieval for prior context instead of always re-injecting full history

Changes

• codex-rs/core/templates/compact/prompt.md
• codex-rs/core/templates/compact/summary_prefix.md

Tests

Not run (content-only changes).

[github-actions]

All contributors have signed the CLA ✍️ ✅
_{Posted by the CLA Assistant Lite bot.}

[ReyNeill]

I have read the CLA Document and I hereby sign the CLA

[etraut-openai]

Is there a bug report filed for this issue? In our contributor's guidance we ask that all bug fix PRs start with a bug report. This helps us prioritize and track issues, especially in cases where we decide to reject a PR.

[etraut-openai]

We do significant testing with our system prompts. Changes to these prompts can easily introduce regressions, so we have a high bar for accepting PRs that modify prompts.

[ReyNeill]

Is there a bug report filed for this issue? In our contributor's guidance we ask that all bug fix PRs start with a bug report. This helps us prioritize and track issues, especially in cases where we decide to reject a PR.

Understood, would that mean delete this PR and start a new one quoting the big report, or I can just simply add it here?

[etraut-openai]

@ReyNeill, no need to create a new PR. Just add a link to a bug report.

[ReyNeill]

I happen to run with the issue right now and submitted the bug: https://github.com/openai/codex/issues/new?template=2-bug-report.yml&steps=Uploaded%20thread:%20019b704d-04f3-7031-b1bb-fb5af42be238

thread ID 019b704d-04f3-7031-b1bb-fb5af42be238

The model wastes a lot of tokens and time verifying all of the session's tasks instead of continuing the task it was doing when the session history auto-compacted

[etraut-openai]

How much testing have you done with your modified prompt? I'm trying to get a sense for whether the solution is speculative or whether you have strong evidence that it improves the behavior, at least for your usage.

[ReyNeill]

Not enough testing, but I'll do it and comeback. Thanks for the fast feedback speed.