STREAMING, THINKING & TOOLS · ANTIGRAVITY TOOLS
Antigravity Manager: Fix Streaming & Tool Errors
Find where the request stopped: sending the input, waiting for text, calling a tool or continuing after its result. Save the conversation and inspect any partial file changes before retrying. Then compare one short request with the failing task.
For version 5.0.0 · Updated
Find the stage that stopped progressing
Use the client’s error and the corresponding Manager request log to identify the last completed stage. A large upload, a slow first response and a pause during a tool call need different checks.
| Last stage reached | Next check |
|---|---|
| Connection fails | Check the address, port, TLS and client environment with the connection guide. |
| Input is still being sent | Compare a small text request with the large body. Record body size, elapsed time and repeated upload attempts. Follow the media guide for file-specific problems. |
| Upload completed; waiting for headers | Record the header-wait error and whether Manager attempted another endpoint or account. Check the timeout settings for that build. |
| Headers arrived; waiting for useful content | Compare a short request on the same model. Note whether thinking events, text or only heartbeats arrive. |
| Text or thinking appeared, then silence | Record the last event and silence duration. Check whether the model is preparing a large tool call or the connection was interrupted. |
| Tool call appeared | Check execution, permissions and the matching result returned by the client; then inspect the following model turn. |
Record the Manager build, client version, protocol and resolved model before changing settings. If the client hides stream events, use Manager’s request logs to compare the failing request with a small one. Keep payloads containing private prompts or files local.
“Stream interrupted before completion” or long silence
Try a fresh short text request on the same model, then repeat the smallest version of the failing task. This shows whether the failure follows the connection, the conversation history or a particular tool response.
- Save the partial result. Inspect any files the agent touched. Check which commands or tool calls already ran before resuming the task.
- Compare streaming and non-streaming if the client supports both. Keep the prompt, model and protocol fixed. Manager can collect an upstream stream into a non-streaming JSON response, so capture the complete error from either mode.
- Check the route carrying the stream. If a reverse proxy or tunnel is involved, compare with direct access through your trusted local or private network. Check that intermediary’s buffering and idle timeout when direct access works.
- Keep a small Unicode example when relevant. If Cyrillic text, another language or an emoji triggers the error, retain that text in the reproduction. Use the harmless file test below to check the tool round trip.
If the model writes a brief introduction and then pauses before a large Write or Edit call, wait for the actual call or a timeout message. The arguments may arrive as a large block. If the connection closes at that point, save the last event and duration, then compare a request to write a much smaller file in a disposable folder.
Which stream timeout should I check?
In Manager 5.0.0, open API Proxy → Model Config → Stream Timeout Settings. Match the observed pause to the control:
- Initial Response Grace (TTFT): waiting for the first useful content, including a slow start or large input.
- Transition Grace (Epoch Reset): the gap after thinking and before text or a tool call.
- Streaming Sliding Timeout: silence while a response is already progressing.
Save the existing values before a comparison. Adjust only the relevant control, use Save Timeout Settings, and repeat the same small task. Check your client and reverse proxy timeouts as well: either can close a connection earlier.
Use Reset to Defaults to return to Manager’s default values when needed, or restore your saved custom values. For large-body upload and header-wait changes, see the version-specific guidance.
Manager sends downstream stream heartbeats while waiting. Judge progress by content and tool events. Leave upstream HTTP/2 keepalive behavior to the installed build; old HTTP/2 PING workarounds from 4.8.9 were removed in 4.9.3.
Thinking budget out of range or missing signature
Use the exact field named in the error to choose the check. A rejected reasoning value needs a settings comparison; a missing or invalid signature needs a conversation-history comparison.
Thinking budget or level rejected
Open API Proxy → Model Config → Thinking Settings → Thinking Budget Control and check Authority Source. Confirm whether Gateway Authoritative Control or Client Direct Control is supplying the reasoning setting.
- Save the current authority, mode and family values. Keep the model fixed for the test.
- For gateway control, compare with Default Mode (Upstream Adaptive). For client control, use the client’s supported default/adaptive setting and remove custom reasoning overrides through its configuration.
- Send one short request, then a small tool task. Add your custom setting back only after the default case works.
thinkingBudget is a numeric budget; thinkingLevel and client effort express levels. Supported controls and ranges vary by model. Use the thinking-settings guide and the limits reported for the selected model. Preserve the rejected field and value in a report if the default case still fails.
Missing or invalid thought signature
Keep the failing history and start a fresh conversation on the same model. Run a short message, one tool call and a follow-up. Note whether the failure began after a restart, a client update or a switch between Claude and Gemini.
If the fresh conversation works, preserve the old transcript and continue in a new session with a short summary of the work you need. Keep the model family consistent during that comparison. Manager’s signature restoration depends on the history and session context; use the Claude Code or OpenCode setup for the provider you selected.
Signature and tool-configuration fields vary by protocol, model family and client. If an error names toolConfig, compare a fresh request using that provider’s default tool settings. Keep the protocol fixed and inspect the field the adapter sends before adding an override.
Leave stored signatures intact. Report the failing block type and error while keeping signature values private. Avoid copying signatures between turns or removing every signature from a saved conversation.
“Field required”, repeated tools or no tool result
Follow one tool call from the model to the client and back. Check the call name, arguments and ID, whether the client executed it, and whether a result with the matching ID was sent in the next request.
For messages.…content.…text: Field required, save the complete field path. Inspect the failing block for a missing text value, an empty or whitespace-only string, or a different content type. For an error naming a tool result, check the call/result pairing instead. If you maintain the client adapter, fix the malformed block where it is created and retest a small conversation.
Run a harmless file-read round trip
Create stream-check.txt in a disposable folder with this exact content. Open that folder in the agent client and allow access to the test file.
STREAM-CHECK-742
Привет мир. Hello world.- In a fresh conversation, ask: “Use your file-reading tool to read stream-check.txt. Quote both lines exactly.”
- Confirm the tool call appeared, the client read that file, and the returned result contains both lines. Approve the read if your client requests permission.
- Check that the model quotes the contents correctly. Then ask: “What identifier was on the first line?” The expected answer is
STREAM-CHECK-742.
If the read never starts, check the client’s tool permissions and execution error. If it runs but the model repeats it, inspect the matching tool result and the next request. If the follow-up fails, capture that second turn as part of the reproduction.
For Write or Edit failures, repeat with a tiny test file and inspect its actual contents and diff. Before retrying an interrupted production task, check for changes already applied so the next attempt starts from the current files.
“safety_settings category invalid”
Find the rejected array entry in the error, then check the active client provider, endpoint and custom safety configuration. Compare a fresh request using the provider’s supported default settings.
If you configured a category that this model or endpoint rejects, restore its supported value or remove that custom entry through the client settings. Keep the previous configuration for comparison. Leave the remaining safety controls at their intended settings.
Retest the same short prompt and model. If the invalid field returns, check the request the adapter actually sends, including field spelling and enum value. Save a redacted structure containing the rejected entry and the client version for the report.
Zero tokens, EOF or “done” with an unfinished task
Inspect the actual answer and tool effects first. Then compare the usage fields and final stream events to find whether the failure is in generation, execution, stream handling or the statistics display.
- Useful output exists, usage shows zero: record the response’s usage fields and the corresponding Monitor or statistics entry. Continue with usage and logging checks.
- Only an introduction or thinking arrived: capture the last complete event, silence duration and error. Return to the stream checks.
- The tool ran, but the task is incomplete: check its returned result and the following model request. Inspect the real file or command result before resuming.
- No useful content arrived: preserve the full response body, even with HTTP 200. Look for an error, a refusal or an empty content block.
What do EOF, [DONE], end_turn and finish_reason mean?
EOF means the connection ended. [DONE] is a closing marker used by some streaming protocols. end_turn and finish_reason describe protocol-level stopping conditions.
Manager can add closing events after partial output to finish the downstream response cleanly. Check the answer, the requested tool effect and a follow-up turn before resuming work that depends on completion. Keep the last content block and closing event together when reporting a premature finish.
In Manager 5.0.0, collecting a Claude stream into a non-streaming response can leave output_tokens at zero when the final usage update contains only that field. If you see this with useful output, compare the original stream’s usage event with the collected response and include both in the report.
If Enable Usage Scaling is active in Experimental Settings, account for that when interpreting the displayed usage. Keep billing and remaining quota checks tied to their respective account data.
A long conversation fails, but a fresh one works
Save or export the conversation, then use the client’s supported compact or summarization flow. Review which decisions, file paths and unfinished steps the summary retains before continuing.
For Claude Code, use its native compact flow within the conversation. If compaction fails, preserve the error and a short task summary, then start a fresh session on the same model. Reintroduce the essential context and check the next tool round trip.
Choose the compression mechanism deliberately
- Client-native compact: the client summarizes its conversation and continues with a smaller active context. Review the retained facts before the next task.
- Gateway-generated summary: if your build forwards a condensed history, review that summary and the retained tool results before continuing. Record Experimental Settings → Context Compression Level when comparing builds. In the 5.0.0 Claude path, the client controls which tool-call turns are retained; begin with its native compact flow.
- Reactive 400: a configured compatibility flow can return a deliberate context-pressure error to trigger client compaction. If it repeats after compaction, record the build and sequence, then review that feature’s configuration.
A generic 400 naming a missing field, signature or invalid category should follow that field’s diagnostic branch. For context errors, inspect the summary and forwarded context locally: a conversation can remain visible in the UI while the model receives a condensed history.
Should I use Deep Compact or a client patch?
Deep Compact is an optional macOS/Cowork feature that combines a native compaction flow with more aggressive history pruning. The patch changes client application files. Keep a restorable copy and confirm support for the exact client version before opting in.
For routine recovery, begin with the supported native compact flow or a fresh session with a reviewed summary. Check patch compatibility after every client update. Backup and checksum checks verify particular files; also test application launch, conversation access, compaction and continued tool use on a disposable case.
The feature’s ./compact control intent belongs to its Cowork flow. Use your client’s documented compact control. The specification’s 35K active-history target is an engineering setting; use the model’s published context limit and actual usage data for capacity planning.
After compaction, ask the client to restate one key decision and continue a small pending step. If information is missing, supply it explicitly from your saved notes before returning to the full task.
When should I check for a version-specific fix?
Compare your exact Manager and client versions when a small, repeatable case keeps failing. Use the release notes for that pair and repeat the same case after an update.
- Large inputs repeatedly upload before failing: 5.0.1-beta.4 separates connect, upload, header-wait and stream budgets, and limits repeated large-body endpoint/account attempts by default. Check the matching build’s timeout configuration.
- Only the current user text or system/identity tags change: 5.0.1-beta.5 revises these sanitization boundaries. Compare a harmless incoming example with the forwarded text in your installed build.
- Stream interruption persists across versions: include the client version, model, last event and small reproduction. A recurrence after an update needs the new result and error.
Save work and back up the relevant data before changing builds. The release guide covers version selection; the reference covers configuration details. Match timeout overrides to that build and failure stage.
Report the smallest failing round trip
Include the first error, the last completed stage and the result of the short comparison. For a tool failure, add the call, execution result and following turn using harmless sample content.
Record Manager and client versions, OS, protocol, requested and resolved model, streaming setting, time, and whether a restart or model switch preceded the failure. Keep signature values, keys, account identifiers, private prompts and file contents out of shared logs.
For rate or capacity errors, use limit and pool troubleshooting. For normal provider configuration, use client setup.