Performance degradation and long "Thinking" timeouts after multiple context compactions

Description

I've noticed a consistent performance issue during long sessions. After performing around 6–7 consecutive compaction operations, the following symptoms occur:

Stuck in "Thinking" state: The model hangs in the thinking phase indefinitely (in some cases for up to an hour before timing out or disconnecting).

Dropped responses: The model either stops generating output entirely or requires repeating the same prompt multiple times to get a response.

Cumulative degradation: The more compaction cycles the context undergoes within a single session, the worse the stability and response times become.

Additional Context:

This behavior reproduces regardless of the selected model (tested across various LLMs via API / OpenRouter).

The issue is directly tied to the accumulation of history and the number of compaction cycles.

Questions:

Is this a known issue, and could it be related to how context compaction is implemented in OpenCode?

Could this be caused by the compacted context/system prompt becoming too noisy or degraded for the model's attention mechanism as the context window fills up, or is it an issue with sending a massive history payload from the client side?

Are there recommended configuration settings for clearing or limiting system history to prevent this degradation without losing critical session context?

Plugins

No response

OpenCode version

No response

Steps to reproduce

No response

Screenshot and/or share link

No response

Operating System

No response

Terminal

No response

No files, tests, version, or reproduction steps are provided. Start by reproducing the problem after 6–7 consecutive compactions across the reported models, then trace context-compaction and history handling; done means identifying the cause of the hangs or degradation and verifying stable responses after repeated compactions.