compaction failed: context summary errors and recovery

Compaction fails in two ways: no useful summary, or a summary that hits the output limit. Both signal context pressure, but one calls for threshold tuning and the other for a larger-output model.

Two compaction errors

Two messages, two dimensions: quality and length.

The first is "cannot compress usefully." Long, scattered history or a weak model produces a poor summary that the system rejects.

The second is "compressed but too long." The summary itself hits the output limit, usually with very large history or high detail requirements (source).

Both may come with continuation failures. Seeing a still-truncated message means repeated attempts also failed; do not keep retrying.

Why compaction fails

Four factors dominate.

1. History too large. Too much content at once, summary either does not fit or loses quality.
2. Output limit too small. The model cannot finish the summary; it is cut midway.
3. Topics too scattered. Mixed tasks make trade-offs impossible.
4. Model capability insufficient. A small model loses the thread and is judged unusable.

All four are input-output mismatches. Too much input and too little output guarantee failure; tuning only delays it.

Check history size before output limit. Reversing the order leads to model swapping when the issue is history.

Automatic or manual

Automatic and manual compaction fail differently.

1. Automatic: triggered at a threshold; failures may repeat every turn.
2. Manual: triggered by you; retry with a different model or scope.
3. Diagnosis: check whether the error appears at fixed turns or after your action.
4. Automatic failure: stop at a stable point and recover manually.
5. Manual failure: change model or narrow scope.

Automatic failure is easily missed. It retries silently until the session stalls; fixed-turn errors suggest it.

Manual compression is controllable. Save key conclusions first; a failure does not break ongoing work.

Recovery order: light to heavy

Preserve data first.

1. Save key conclusions. Write decisions, paths, and todos into a new message or external note.
2. Open a new session. Start with the summary or key conclusions, bypassing oversized history.
3. Adjust the threshold. Trigger compaction earlier; too low increases calls.
4. Switch to a larger-output model. When summaries need more room, a model change beats parameter tuning.
5. Split history. Divide by task to reduce per-compaction pressure.
6. Remove irrelevant content. Delete long logs, duplicate attachments, and dead dialogue.

A new session is the most reliable stopgap. It does not depend on compaction success and restores work immediately; the cost is lost context.

Threshold and model must be considered together. Threshold alone leaves summaries cut; model alone leaves frequency unchanged.

After two failures

Two consecutive failures mean the current path is wrong.

1. Stop automatic retries to avoid repeated failure records.
2. Start a new session with key conclusions.
3. Keep only necessary background, not the whole history.
4. Retrieve old details with search or logs instead of recompressing.
5. Record history size, model, and threshold for the next adjustment.

Two failures are a signal, not bad luck. Same input and model will likely produce the same result.

New sessions also need size control. Moving the entire history copies the problem; carry only conclusions and references.

Prevention: keep history bounded

Compaction failure is preventable by controlling history size.

1. Close stages. At the end of a task, write conclusions and open a new session.
2. Avoid pasting large blocks. Use paths or summaries for long logs and files.
3. Clean periodically. Remove irrelevant attachments and duplicate dialogue.
4. Watch threshold warnings. Start converging when compaction appears, not after failure.
5. Choose an adequate model. Long sessions need stronger output capacity.

Prevention costs less than recovery. Two lines of conclusions per task beats rebuilding context afterward.

Next steps

Read the error, then recover by cost. For the mechanism, read What context compaction is. For window overflow, read Context window overflow. For output limits, read Output token limit and truncated replies.

Sources: DeepSeek Harness docs, dsh CLI README, deepseek-harness repository