orchestration leakage during context compaction #28355

Activity

1. github-actions commented on May 19, 2026

   

   github-actionsBot

   on May 19, 2026 – with GitHub Actions

   Contributor

   More actions

   This issue doesn't fully meet our contributing guidelines.

   What needs to be fixed:

   • Steps to reproduce is empty — please describe how to reproduce the behavior you observed
   • OpenCode version is empty — please include the version you are using
   • Operating System is empty — please include your OS
   • The description consists largely of a list of technical terms and incomplete example fragments (e.g. Ex: prompt: Expected output: left unfilled), rather than a clear account of what happened and how to reproduce it

   Please edit this issue to address the above within 2 hours, or it will be automatically closed.

   ---

   This issue might be a duplicate of existing issues. Please check:

   • Compaction and message export can leak reasoning across models #25190: Compaction and message export can leak reasoning across models — describes a similar scenario where reasoning/internal content leaks out during compaction or export

   If you believe this was flagged incorrectly, please let a maintainer know.
2. github-actions

   added

   needs:complianceThis means the issue will auto-close after 2 hours.This means the issue will auto-close after 2 hours.

   on May 19, 2026
3. github-actions

   assigned

   rekram1-node

   on May 19, 2026
4. jiten-core commented on May 19, 2026

   

   jiten-core

   on May 19, 2026

   Author

   More actions

   OS:
   Windows 11

   OpenCode version:
   v1.15.5

   Model:
   MiniMax M2.5

   Steps to reproduce:

   1. Open OpenCode Desktop v1.15.5

   2. Start a new session in Build mode

   3. Use MiniMax M2.5 as the model

   4. Send very small prompts such as:

      • “hi”
      • “banana”
      • “reply only with banana”

   5. Continue chatting for a few messages

   Observed behavior:
   The assistant first responds normally, but then the session sometimes becomes compacted and the chat shows structured task-summary sections such as:

   • Goal
   • Constraints & Preferences
   • Progress
   • Next Steps
   • Critical Context

   Sometimes “Session compacted” also appears in the timeline.

   After compaction, MiniMax may continue replying in this structured task-summary style even for casual prompts.

   Expected behavior:
   Casual prompts should continue producing normal conversational responses after compaction instead of switching into persistent task-summary formatting.

   Notes:
   This behavior appears much more frequently with MiniMax M2.5 than with Claude/GPT models in the same environment.
5. github-actions

   removed

   needs:complianceThis means the issue will auto-close after 2 hours.This means the issue will auto-close after 2 hours.

   on May 19, 2026
6. github-actions commented on May 19, 2026

   

   github-actionsBot

   on May 19, 2026 – with GitHub Actions

   Contributor

   More actions

   Thanks for updating the issue! Everything looks good now — the steps to reproduce, OS, version, and observed/expected behavior are all clearly described. We appreciate you taking the time to improve it.
7. JARVIS-coding-Agent commented on Jun 8, 2026

   

   JARVIS-coding-Agent

   on Jun 8, 2026

   More actions

   +1 — This also affects ACP downstream platforms (e.g. Discord, Slack) when opencode is used via opencode acp.

   Reproduction via ACP

   When opencode runs as an ACP agent behind a platform bridge (like OpenAB), the compacted planning block (Goal / Constraints & Preferences / Progress / Blocked / Next Steps / Critical Context / Relevant Files) is included in the result.content[].text returned via JSON-RPC.

   The downstream platform has no way to distinguish this internal orchestration state from the actual answer — it renders everything to the end user.

   Observed behavior (ACP mode)

   • Each tool call iteration appends another full planning dump to the response
   • A single user question produces 6-7 repetitions of the planning block before the actual answer
   • Internal state (model name, base URL, file paths) is leaked to end users
   • Model: Gemma 4 31B-IT via vLLM (non-Claude/GPT models are more susceptible, consistent with the original report)

   Impact

   For ACP consumers this is not just a UX annoyance — it is:

   1. Information leakage (internal infra details exposed to users)
   2. Token waste (repeated planning blocks consume output tokens)
   3. Unusable output (users see walls of planning text before getting the actual answer)

   Suggested fix scope

   Whatever fix is applied for the Desktop UI should also ensure the compaction summary format does not leak into the text parts returned via ACP session/prompt responses. Ideally the planning state should only be emitted via SessionUpdate::Plan (ref #30659), not as raw text in the assistant message.
8. github-actions commented on Aug 7, 2026

   

   github-actionsBot

   on Aug 7, 2026 – with GitHub Actions

   Contributor

   More actions

   To stay organized issues are automatically closed after 60 days of no activity. If the issue is still relevant please open a new one.
9. github-actions

   closed this as not plannedon Aug 7, 2026