Opens in a new tab
News

Claude compaction: keep the reason the worker is waiting

In brief

“Restart after the queue drains” is not the same as “restart approved.” Follow that small difference through Claude’s conversation summaries and Opus 5.5 replay checks.

6 min read

Is this the right starting point?
Before you read
Practical agent integration experience with conversation state, API messages and execution permissions.
What you will take away
Evaluate conversation compaction for preserved conditional authority, safe history replacement and model-migration compatibility.

Follow-up · September 22, 2026: Opus 5.5 adds conversation-replay checks

Sources
A red ribbon passes through an accordion booklet and three dark beads to a black crank, while a hand presses the booklet.
On this page4 sections

A worker is still emptying its queue. The instruction is clear: “Restart the worker after its queue drains.” Now imagine a summary that remembers only “restart approved.” It has kept the subject and lost the reason to wait. This hypothetical case captures the difficulty of conversation compaction: making an agent's working record smaller without changing the job it is allowed to do.

Anthropic introduced on-demand compaction on September 14, 2026. With the Claude API beta header compact-2026-09-04 and the top-level compaction parameter, an application can request a signed summary block. On later requests, that block goes before the new messages in place of the history it summarizes.

The smaller record can make a long conversation easier to carry forward. For our waiting worker, though, the important question is what happens next. Does the assistant still wait for the queue to drain? Do options rejected earlier stay rejected? Counting fewer tokens, the units of text the model processes, does not settle either question.

Nor does this mean keeping every word forever. Retain the audit records your data policy requires, and test the meaning needed for the next decision. The useful saving is in detail the workflow can afford to shorten, not in the condition that made a restart acceptable.

Choose when the conversation gets shorter

The compaction overview distinguishes two ways to reach that smaller record. An application can ask for a summary at a chosen point, or compaction can happen inside a request when a token threshold is reached. On-demand work can run in the background while selected recent exchanges stay word for word. Whether retained reasoning records, called thinking blocks, remain valid depends on the chosen model, platform and compatibility conditions.

For an incident assistant, a natural proposed stopping point might be the end of one bounded investigation. Its findings and open questions could be reviewed together before older conversation is replaced. That is an application design choice, not a built-in incident-management feature. It asks the team to do some integration work in return for choosing when detailed history becomes a shorter account.

Keep history, summary and current state distinct: Older conversation, Retain an original record; Signed summary, Replace summarized messages; Recent turns, Keep selected exchanges intact.
Older history, its signed summary and retained recent turns play different roles. Application controls still enforce permissions and check external state.

The word “after” is doing real work

Return to “restart approved.” If the queue-drain condition disappears, the assistant may propose restarting immediately. A separate execution check against the current queue should refuse that proposal while work remains. But if the conversation held the only copy of the condition, the application has lost the very comparison that would have caught the mistake.

A useful test keeps both the condition and its link to the restart. Try another case in which an early diagnosis was later disproved. After compaction, the rejected explanation should not return as an established fact merely because the summary remembered it more clearly than the correction. These are tests of meaning where meaning changes a decision.

Both “restart” and “queue” could appear in a bad summary. Read the proposal that follows it: does one still have to wait for the other? That small relationship is the difference between the original permission and a broader one nobody gave.

The summary helps the model propose work; it does not become the authority to execute it. As the operational agent skills guide explains, application controls still check the permitted operation and target. Keeping the condition in the conversation and in execution controls is not pointless duplication. One helps the model ask for the right action; the other can refuse when it asks for the wrong one.

Keep the correction that arrived during the summary

While a summary is being prepared, someone may add a correction. Record where its source history ends and keep the later exchanges when installing it, alongside the original evidence required by policy. Otherwise the application can successfully summarize yesterday's discussion and accidentally throw away today's correction.

Run a proposed pilot with the same investigation cases, once with compaction and once without. Compare completed investigations, missed constraints and total model usage, including the work of summarizing. A smaller active conversation does not necessarily mean less computation over the whole task. The AI token usage guide follows that broader accounting.

If no usable summary arrives, keep the existing history as the basis for a defined pause or retry. If a critical condition vanishes, stop expanding the pilot while the summarization strategy is repaired. Restoring old messages can restore context. It cannot undo a restart, or another external action, that already happened because the condition was lost.

The goal is a shorter working record whose next decision still makes sense, with a clear account of what was replaced. Changing the model means checking that handoff again, as the following dated update shows.

September 22 follow-up: carry the condition into Opus 5.5

The September 22 release notes introduced Claude Opus 5.5, and its feature documentation confirms on-demand compaction support. It is another candidate for the worker test, with the intended platform and history-handling settings recorded rather than assumed.

Suppose the model has returned a thinking block, a piece of its reasoning carried with the conversation. The client sends that block back unchanged and appends the next question about the queue. That is different from editing an earlier instruction or tool definition and then replaying the old block. The material before it, its prefix, has changed. Opus 5.5 checks that prefix.

With mismatch behavior unset, an ordinary request from an account created on or after August 31, 2026 can be rejected with a 400 error when the prefix does not match. Older accounts can pass failing blocks through. Anthropic documents append-only updates and an explicit option to drop the affected blocks. Dropping can let the request continue, but without that block's reasoning. A reply arriving successfully is not enough to show that the worker's earlier context survived.

There is a separate instruction question in the on-demand guide. Imagine introducing the queue condition in a system message partway through the conversation, then replacing that message with a summary. Instructions in a summarized system message cease to apply; those still needed must be stated again through the documented message flow. The words “queue drains” appearing in a readable summary are not the same as the instruction remaining in force. This is a specific migration check, not a claim that every earlier application was broken.

After the migration, replay the worker case. The condition must still govern the proposal, and execution controls must check the current queue. If either part is missing, the new workflow has lost something the old one needed. This documentation-based follow-up claims no hands-on Opus 5.5 test. It gives the team a way to tell an accepted API request from a continuation that still knows why the worker is waiting.

Back to the original report

Sources & context

Sources linked in this article. Read alongside the author’s analysis; a citation does not independently verify a publisher’s claims.

Report an error or outdated detail

More on Anthropic

All Anthropic coverage

Keep reading

Where to go next