OpenClaw "Auto-compaction timed out before the next model request. Try again or run /compact."

Why it happens

Compaction is a model call. OpenClaw sends the older part of the transcript to a model and asks for a summary. If that model is slow, overloaded or rate limited, the summary times out or fails and the turn cannot continue.

1. A slow summary model. By default compaction uses the session's primary model, so a local model on modest hardware has to summarize a huge transcript. A 905k-token session on 2026.9.7 timed out every turn after about 3 minutes (issue #164220).
2. Thinking enabled for summaries. Reasoning tokens eat the time window; OpenClaw already prefers thinking off for native local Ollama.
3. Provider errors during summarization: a 429 quota, overload, or auth failure. Before the fix for issue #166438, these surfaced only as a generic agent-runner-failure event.
4. Compaction cancelled: a safeguard hook or a stop cancelled the compaction. Usually the log line next to it gives the reason.

The fix

1. 1 Find the real reason: openclaw logs --follow, then look for auto-compaction failed for, preflight compaction failed, or [compaction-diag] lines.
2. 2 Use a faster summary model: openclaw config set agents.defaults.compaction.model "<provider/model>".
3. 3 Give slow models more time: openclaw config set agents.defaults.compaction.timeoutSeconds 600. The window resets on every streamed token, and a whole compaction is capped at 10 windows.
4. 4 Turn summary thinking down: set agents.defaults.compaction.thinkingLevel to off or low.
5. 5 Restart the Gateway, then run /compact in the chat. If it still fails, /new starts a clean session; the old transcript stays on disk.

Why the log only said agent-runner-failure

Issue #166438 reported that during a multi-hour compaction failure loop the only signal was a generic agent-runner-failure event and the heartbeat consecutiveErrors counter. Nothing said that compaction was the failing stage or why. The fix (PR #167016, merged 8 October 2026) logs the actual reason and compaction model for both /compact and auto-compaction. On builds without it, match the time of the failure against the gateway log's model fallback decision lines and your heartbeat run history to find the provider error.

A related change (PR #164246, merged 4 October 2026) stops a summary timeout from dead-ending every later turn: OpenClaw commits the compaction without a summary and logs reason=timeout summary=deterministic. Check your version's release notes before relying on either fix.

Still failing?

• If the reason is a 429 or overload, the compaction model's provider is the problem; switch compaction.model to another provider.
• If compaction never triggers before the context fills, set agents.defaults.compaction.maxActiveTranscriptBytes to compact earlier.
• If the agent forgets details after compaction succeeds, see the silent context compression decoder for keepRecentTokens and postCompactionSections.

Related errors

• OpenClaw silently compresses context and loses work
• OpenClaw "Unknown model"