fix(session): prevent double compaction after auto-compact

[sgfvamll]

Issue for this PR

Resolves #26230
Resolves #20621

Type of change

• Bug fix
• New feature
• Refactor / code improvement
• Documentation

What does this PR do?

After auto-compaction, filterCompacted retains tail assistant messages that carry stale cumulative token counts (~100K). The backward scan in the prompt loop picks up these retained tails as lastFinished, and the overflow pre-check immediately fires a second compaction — every single time. This wastes tokens and doubles Copilot request consumption.

Root cause: The backward scan at prompt.ts:1418-1426 searches from the end of filterCompacted's output and breaks as soon as it finds lastUser + lastFinished. After compaction, the output is:

[0]  compaction_user        (COMPACTION part with tail_start_id)
[1]  compaction_summary     summary=true, finish=stop
[2-20] retained tail assistants  tokens=97K-103K, finish=tool-calls (stale)
[21] auto-continue user     tokens=0

The scan finds lastUser=[21] and lastFinished=[20] (a retained tail), then breaks — never reaching the compaction part at [0]. Since tasks is empty, the compaction task handler is skipped. The pre-check then sees lastFinished with stale 103K tokens, isOverflow returns true, and a second compaction fires.

Fix: Add a guard that only triggers the overflow pre-check when lastFinished is newer than the last compaction user message. Retained tails have IDs older than the compaction user, so they fail this check.

4-line production change in packages/opencode/src/session/prompt.ts.

How did you verify your code works?

• Reproduced the bug empirically against a real session (587 messages, 19 compaction pairs — all doubled)
• Imported the actual filterCompacted function and traced the exact output ordering and backward scan behavior
• bun typecheck — clean
• bun test test/session/prompt.test.ts --test-name-pattern "double-compact" — 1 pass
• Regression test seeds a post-compaction state (compaction pair + auto-continue user) and asserts no second compaction is created and the model is called exactly once

Screenshots / recordings

N/A — no UI changes.

Checklist

• I have tested my changes locally
• I have not included unrelated changes in this PR

[github-actions]

The following comment was made by an LLM, it may be inaccurate:

Potential Related PR Found:

• PR fix(session): prevent double compaction when task already pending #26235: "fix(session): prevent double compaction when task already pending" - fix(session): prevent double compaction when task already pending #26235

This PR appears to be related as it also addresses double compaction issues in sessions. You should verify if PR #26235 overlaps with the fix in PR #26484 or if they address different scenarios of the same problem.

[github-actions]

Thanks for updating your PR! It now meets our contributing guidelines. 👍

[sgfvamll]

I'm not sure if the root cause analysis is right. But it fixed the issue I encountered. May serve as a workround.
I'm not sure if this fix might cause other issues in some corner cases.

[github-actions]

Automated PR Cleanup

Thank you for contributing to opencode.

Due to the high volume of PRs from users and AI agents, we periodically close older PRs using automated criteria so maintainers can focus review time on the most active and community-supported contributions.

This PR was closed because it matched the following cleanup criteria:

• The PR was created more than 1 month ago
• The PR had fewer than 2 positive reactions
• Positive reactions are counted as thumbs-up, heart, celebration, or rocket reactions on the PR

PRs created within the last month are not affected by this cleanup.

If you believe this PR was closed incorrectly, or if you are still actively working on it, please leave a comment explaining why it should be reopened. A maintainer can review and reopen it if appropriate.

Thanks again for taking the time to contribute.