On this page
Opus 5.5 streaming progress can disappear even while a tool loop keeps working. The first checks are the thinking display setting and the stream parser: progress can arrive in thinking blocks, and the default display omits their text. Listening only for ordinary text deltas can make an active task look idle.
Set thinking.display to summarized when you want readable summaries, route events by type, and keep tool status separate from model commentary. This guide uses Anthropic's direct Claude API documentation checked on October 15, 2026. Its stream fixtures were executed locally; they are synthetic events, not a recorded Opus API run.
Diagnose silence before changing timeouts
Anthropic's Opus 5.5 what's-new page describes the response change: notes between tool calls can arrive in thinking blocks whose text is empty at the default display. The request can succeed without producing the text your old interface expected.
Different kinds of silence need different fixes:
| Observed state | Evidence to inspect | Next action |
|---|---|---|
| Events arrive, thinking text is empty | Display setting and typed blocks | Request readable summaries if needed |
| Nonempty thinking deltas arrive, UI stays blank | Parser routes and rendering | Handle thinking deltas separately |
| A tool has started and returns no result yet | Tool request ID and execution log | Show actual tool state and investigate the tool |
| Stream stops with an error | Error event and provider request ID | Handle the reported failure |
| Final answer arrives but work is incomplete | Task acceptance evidence | Keep the task incomplete |
Inspect the event stream before treating silence as a slow model. Redact private data when saving a diagnostic sample, but retain event types, timing, block indices and stop state. These fields show whether the server, parser, tool executor or renderer stopped supplying useful information.
Avoid resetting a task simply because its commentary is quiet. If a tool has already started, a new run can duplicate work. A status check should identify the original task and its running operation. Whether to cancel or resume depends on that operation's actual state.
Request summaries and preserve the provider blocks
The migration guide gives this setting for readable thinking summaries:
This is a documentary payload, not a submitted request. Use your account's actual endpoint and authorization. The same guide says to select content blocks by type and pass thinking blocks back unmodified in tool loops. Treat the readable summary as display content, not as a replacement for the original block used by the conversation.
Opus 5.5 keeps adaptive thinking on. Trying to restore the old interface by setting thinking to disabled introduces a different request error. If you need to alter latency or cost, use the documented effort control and evaluate the task again; do not use a display setting as a performance benchmark. Model changes
The migration guide also names display: "updates" as a beta option. Start with the stable summarized setting when that meets the product need, then verify the beta requirements before adopting updates. A screenshot of a settings field does not prove your SDK serialized it or the provider accepted it. Save the sanitized actual payload for the request you test.
Route stream events by meaning
A text-only listener typically looks for text_delta and appends its text to the answer. That listener can correctly display the final answer while discarding every progress summary. Keep separate channels for final answer text, model summaries and real tool execution state.
Here is a small event router with synthetic fixtures:
The assertions were executed. They check that readable summaries reach the progress channel and empty thinking does not become an empty answer. They do not implement a complete SSE client, assemble tool arguments or preserve the conversation. Keep those responsibilities in your actual streaming implementation and test them against provider events.
Do not append summaries to the final answer simply to make something visible. A summary can describe a planned check that has not finished. In the interface, label it as progress and keep the answer's acceptance state tied to actual results.
Show tool progress from the tool itself
For a task that searches documents, runs tests and prepares a patch, model commentary is useful orientation. The tool executor is the source for whether a search, test or file operation started, returned or failed.
Represent those events with a stable operation ID and an explicit state. A search might be queued, running, returned or failed. A test can have a command, exit code and result artifact. A patch can have a diff and verification state. The interface can then stay informative even when the model produces no commentary.
A useful progress row says “Running the checkout regression command” and points to the associated command or task. Avoid labels such as “Almost done” unless the system has a real completion estimate. For a tool that exposes no progress, show its start time and running state; invented percentages make the silence harder to diagnose.
Keep tool failure visible if the model continues with other work. A later message does not erase an earlier failed test. The final result should state which checks passed and which remain unresolved, with the corresponding artifacts. This makes the progress UI part of task review rather than a decorative activity indicator.
Verify the fix at three boundaries
First, verify request serialization. Capture the outgoing model ID and thinking object from the actual SDK or gateway. A unit fixture for your application configuration does not prove a gateway preserved the field.
Second, verify stream handling. On an authorized live test, retain a sanitized event trace and check that nonempty thinking deltas become progress, text deltas become answer content, and errors or stream completion reach their own handlers. If no summary is produced, record that outcome; the display setting permits readable content but does not supply a promised update cadence.
Third, verify user-visible completion. Run a representative tool task with a real result artifact. Check that the interface shows the tool's actual state, the final answer, and the acceptance evidence. Reload or reconnect to the task if your product supports that flow and confirm that the original operation is still identified correctly.
Use the same task and tools before and after the parser change. Compare visibility and correctness separately from elapsed time. A fix that restores progress text may leave the task duration unchanged, which is still a useful improvement when users can understand the work.
Build a stream review workbench with Atoms
An Atoms web app can help a team inspect sanitized event traces and their rendered output. Start with fixture upload or manual input. No native Opus connection is established by this workflow.
Ask for an event timeline with separate lanes for answer text, model progress, tool operations and errors. Preserve task ID, event order, block index and timestamp. Add a review panel showing expected and observed states. Never mark business completion from a stream-stop event alone.
In the preview, load the fixtures above and verify their lane assignments. Add a failed tool fixture followed by a final answer and check that the failure stays visible. Save a trace, reload it and compare the same event order. If you need shared persistence, use the Atoms Cloud guide and verify access to stored traces before adding real customer material.
Conclusion
Restore readable summaries with the documented display setting, route blocks by type, and report real tool state independently. A stream ending proves that the response ended; the task's artifacts prove what was completed. Build the stream review workbench in Atoms to make both parts reviewable.
Frequently asked questions
01Q1: Does a quiet Opus 5.5 stream mean the task stopped?
It can mean the default display omitted thinking text or the parser ignored it. Inspect events and tool state before deciding that the task stopped.
02Q2: Should I disable thinking to get progress back?
Opus 5.5 does not support disabling thinking. Use its documented display setting and typed event handling.
03Q3: Are summaries the full internal reasoning?
The migration guide describes readable summaries as display text. Preserve the original provider blocks for conversation replay and keep displayed summaries separate from the final answer.
04Q4: Did the fixtures reproduce a live provider stream?
They exercise the local event router with synthetic data. Provider acceptance, event timing and task completion still need a live authorized test.
