fix: improve compaction continuation to prevent agent from stopping

[guazi04]

Issue for this PR

Closes #13217

Type of change

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

What does this PR do?

After compaction, the agent frequently stops working instead of continuing. This affects all providers and has been confirmed by 8+ users.

Root cause: Two issues combine:

1. The synthetic continuation message after compaction says "Continue if you have next steps, or stop and ask for clarification" — this gives the model explicit permission to stop. After seeing a compressed summary, models frequently choose to stop.

2. The compaction summary prompt doesn't ask for remaining tasks or next actions. Without clear direction in the summary, the model has nothing to continue with.

Fix (single file, compaction.ts, +47/-2 lines):

1. Replaced the continuation message with an assertive 3-line directive that tells the model to read the summary and take the next concrete step. It still allows asking questions when genuinely blocked on missing info.

2. Enhanced the summary prompt with 4 new sections: current status, remaining tasks, next actions, and open questions — so the summary captures what to do next.

3. Inject session todos into the compaction prompt unconditionally (even when a plugin provides a custom prompt), so the summary agent faithfully captures remaining work.

4. Include incomplete todos in the continuation message with an explicit instruction to resume the in-progress item.

I identified this by studying how oh-my-opencode's plugins (compaction-context-injector, todo-continuation-enforcer) solve this same problem — they enrich the compaction context and aggressively prevent the model from stopping. This PR brings the core fix upstream without requiring plugins.

How did you verify your code works?

• bun typecheck passes cleanly
• Backward-compatible: when no todos exist, behavior is identical except for the improved continuation message
• The overflow-specific media attachment prefix is preserved
• The replay path (when a previous user message exists) is unchanged
• The plugin hook (experimental.session.compacting) continues to work as before

Screenshots / recordings

N/A — backend logic change, no UI impact.

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:

I found related PRs that might be duplicates or closely related:

Potential Duplicates:

1. PR fix: auto-resume agent loop after context compaction #12970 - "fix: auto-resume agent loop after context compaction"

   • fix: auto-resume agent loop after context compaction #12970
   • Directly addresses the same issue of agent resuming after compaction

2. PR fix: compaction bugs #13946 and #13980 #14245 - "fix: compaction bugs Bug: opencode run exits after compaction when compaction model's token usage exceeds overflow threshold #13946 and Bug: compaction.reserved config is ignored for models without explicit 'input' limit #13980"

   • fix: compaction bugs #13946 and #13980 #14245
   • Fixes compaction-related bugs that may overlap with continuation issues

3. PR fix(opencode): skip overflow check on compaction summaries #15129 - "fix(opencode): skip overflow check on compaction summaries"

   • fix(opencode): skip overflow check on compaction summaries #15129
   • Related to compaction summary generation and context management

Why they might be related:

• PR fix: auto-resume agent loop after context compaction #12970 directly targets auto-resuming the agent after compaction, which is the core issue being fixed in fix: improve compaction continuation to prevent agent from stopping #16073
• PR fix: compaction bugs #13946 and #13980 #14245 addresses other compaction bugs that may be related to the stop/continuation behavior
• PR fix(opencode): skip overflow check on compaction summaries #15129 improves compaction summary handling, which relates to the enhancement of summary prompts in fix: improve compaction continuation to prevent agent from stopping #16073

[github-actions]

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

[IamTheCarl]

I make a fork that essentially rebases this commit onto the current dev branch (current as of when I made it, it's probably already behind) to try this.
https://github.com/IamTheCarl/opencode/tree/agent-stops-after-compaction

It's certainly a significant improvement. I'm using qwen3.5:9b on a locally hosted Ollama instance.
The only real issue is that sometimes it stops thinking "I need more information" but doesn't actually ask a direct question. I am very new to all of this so I'm not sure what information is helpful or how to collect useful information, so my naive guess is that that this is more of a problem with the model than with the prompt in my instance.