OmniscioGuidesAuto-compaction failed
Claude Code: auto-compaction failed (summarization failed)
Last checked against Anthropic's documentation on November 14, 2026. By the Omniscio team. Updated .
What does the message look like?
The short wording people search for reads like this:
Auto-compaction failed: Summarization failedThat wording is not in Anthropic's documentation. Other coding agents print it for the same event, which is why it appears in search.
In Claude Code, the same failure prints the message below. The words after the colon are the real error:
Prompt is too long · automatic compaction failed: <the underlying error>If your screen shows that second line, this page is still the right one. Only the first line differs, and a -p-run transcript uses the bare Prompt is too long line instead.
What is compaction, and what failed?
Compaction is Claude Code's way of making room in a conversation that has grown too long. It asks a model to summarize the older parts, then carries on with the summary in place of them.
This message means that summary step failed. So there was no summary, and no room was made.
The cause is underneath. A failed summary is a symptom, and the real error is named after the colon.
Why did the summary fail?
There are four usual reasons.
- An error the summary request hit. A bad credential, a model that is unavailable, or the network. This is the one Anthropic documents, and the message names it after the colon.
- The conversation is too big to send. Compacting is itself a request. When the conversation is too large to send, it is too large to summarize.
- There was nothing to summarize. A single large paste can fill the window with one exchange. Claude Code does not attempt compaction on a single exchange now: it skips the attempt and explains what fills the request instead.
- It was refused. On versions before v2.1.282, a refused summarization request ended here. From v2.1.282 it retries on a fallback model instead.
How do I fix it?
Commands go into Claude Code's own chat box, typed with a slash in front, then Enter.
- Read the words after the colon, and fix that error. It is the real cause, and
/compactfails the same way until it is resolved. If you are not sure what it means, the list after these steps covers the three error kinds it usually names. - Then free space. Type
/compact, then words telling it what to keep, for example/compact keep only the plan and the diff. Use your own topic in place of "the plan and the diff". - If it fails again, press Esc twice. That takes you back past the turn that made the conversation huge. Then send a smaller message.
- Summarize only part of it. Type
/rewindand choose Summarize from here or Summarize up to here. That compacts a smaller chunk. - Run
/clear. Start fresh if the earlier conversation is no longer needed. This is the step that works when nothing can be summarized. Before you clear, ask Claude to write what you still need into a file.
If it names one of these, this is what it usually means:
- A sign-in or permission message. Run
/login, then try again. - A network or connection message. See our connection errors guide.
- A model that is unavailable. Try again in a few minutes, or switch to another model with
/model.
Running /context shows what is taking the space: messages, the system prompt, tools, or memory files.
Will Claude Code retry on its own?
Sometimes, and then it stops. Two changes are worth knowing.
- From v2.1.282, a summarization request that is refused is retried on a fallback model.
- From v2.1.76, auto-compaction stops after three consecutive failures. It does not retry forever.
So a session that keeps failing needs a manual step. Update Claude Code when you can, then work through the list above from the top.
Common questions
What does auto-compaction failed (summarization failed) mean?
Automatic compaction ran, and the step that summarizes your conversation failed. Something went wrong during that summary request, so Claude Code could not make room in the window. The cause is whatever error sits underneath.
Is that the exact wording Claude Code uses?
Not quite. Anthropic's error reference prints this event as 'Prompt is too long · automatic compaction failed:' followed by the underlying error. The shorter 'auto-compaction failed: summarization failed' wording is what other coding agents print for the same event. The fix is the same: find the error underneath.
Should I run /compact?
Only once the underlying error is fixed. Anthropic's rule is that /compact fails on the same error until you resolve it. If the failure is the conversation being too large to summarize, /compact will not help either; press Esc twice, or run /clear.
Will Claude Code retry on its own?
Since v2.1.282, a summarization request that is refused is retried on a fallback model. Since v2.1.76, auto-compaction stops after three consecutive failures rather than retrying forever, so a session that keeps failing needs a manual step.
Where these answers come from
- Claude Code error reference, for
Prompt is too long · automatic compaction failed:, the single-exchange case, the-pwording and what fills a window. Read November 14, 2026. - Claude Code troubleshooting, for the focused
/compactform and its example. Read November 14, 2026. - Checkpointing, for the
/rewindmenu and its Summarize from here and Summarize up to here entries. Read November 14, 2026. - Claude Code's changelog, entry for v2.1.282 (refused summarization retried on a fallback model) and v2.1.76 (stops after three consecutive failures). Read November 14, 2026.
- The wording
Auto-compaction failed: Summarization faileddoes not appear in Anthropic's Claude Code documentation or changelog. It comes from other coding agents that wrap a failed summary call, and this page does not attribute it to Claude Code. - Our compaction errors hub indexes the whole family, including the message that says the conversation could not be reduced and the one that says compaction is not making progress. The prompt is too long guide has the
Prompt is too longside of the same line.