[BUG] Reaching context window 100% and not triggering auto-compaction

[JWIMaster]

Auto-compaction not triggering in 0.1.7-rc1

In DeepSeek Harness 0.1.7-rc1, auto-compaction does not appear to trigger when the context window reaches its limit.

Configuration:

• Context window: 80,000 tokens
• Max output tokens: 32,000
• The same context and max token settings are configured for the model in LM Studio.

With these settings, the context continues to grow without triggering automatic compaction.

However, using the latest NPM release, 0.1.5-rc3, with the same configuration, auto-compaction works as expected.

Expected behaviour

When the context approaches the configured context window limit, DeepSeek Harness should automatically trigger compaction, as it does in 0.1.5-rc3.

Actual behaviour

In 0.1.7-rc1, auto-compaction is not triggered, and the context continues to grow until the context limit is reached.

Reproduction

1. Run DeepSeek Harness 0.1.7-rc1.
2. Configure an LM Studio model with an 80k context window and 32k max output tokens.
3. Configure DeepSeek Harness with the same values.
4. Continue a conversation until the context approaches 80k tokens.
5. Observe that auto-compaction does not trigger.
6. Repeat with 0.1.5-rc3 using the same configuration. Auto-compaction triggers as expected.

[danielbitpro]

Bug is present in 0.1.5-rc3 as well, hit this bug twice already. The only way around it is to be sure to leave some context room for the LLM. Once you hit the bug, stop DSH, increase context window of the LLM by about 10%, then increase it in the settings of DSH as well and finally start DSH again. After that you can trigger a manual compaction, which will successfully finish. Once done, revert the context window setting and restart DSH so you can recover if you hit the bug again.
Note that I am using a local model, not sure if this is correct for cloud models.

[dsh-community-assistant[bot]]

Thanks for the report.

[apellaman]

It's broken in 0.1.7-rc.2 too

[JWIMaster]

Does anyone know if it's broken in 0.2.0 rc1/2?

[davidlav]

Yes, it's still bugged in 0.2.0-rc.2.

[yethee]

I can confirm that the same issue exists in version 0.2.0-rc2. Auto-compaction does not work.

[davidlav]

Explanations here:

#8079 (comment)

#8088 (comment)

[dsh-community-assistant[bot]]

To investigate further, please share: your operating system and app platform; a log with tokens and personal information removed.