Skip to content
5,000+ phones repaired • FREE tracked Royal Mail postage • Warranty on every repair • Need help?
5,000+ phones repaired FREE tracked Royal Mail postage Warranty on every repair (terms) Harlow workshop • UK-wide postal repairs Need help? Contact us

OpenClaw streamidletimeout: Diagnose a Stalled Run

5/11/2026 by James Waterston

OpenClaw stream_idle_timeout: Diagnose a Stalled Run

If OpenClaw shows stream_idle_timeout, start with the gateway logs before changing the idle timeout settings. Reported stalls include silent stops, explicit run timeouts and network failures, so the same recovery step won’t suit every case.

We recommend identifying the first recorded failure, then matching it to the affected run. Increasing a timeout won’t address every interrupted connection, and sending “continue” can restart activity without explaining why it stopped.

Begin by separating a session that stopped normally but prematurely from one that logged an error.

Key takeaways

  • A stuck session can show stopReason: 'stop' rather than 'toolUse', without an explicit timeout at that point.
  • One report configured both timeout values to 500 seconds, yet logged an embedded-run limit of 15,000 milliseconds.
  • Successful tool output doesn’t prove the next model request succeeded. Look for aborts, connection errors and context overflow after toolResult.

Our starting rule is straightforward: follow one run through its logs. Otherwise, a later fallback error can distract you from the event that interrupted the original request.

Recognise the two different kinds of stalled run

A flat vector workflow diagram with connected nodes, a checkmark, a gear, a clock, and a pause symbol.

A silent stop before the expected tool call

One OpenClaw issue describes a session stuck after a user message, with the agent apparently still thinking when a tool call was expected. Instead of returning stopReason: 'toolUse', the model returned stopReason: 'stop'.

The session then appeared stuck until another user message, such as “continue”, triggered activity. Recorded gaps reached 20 minutes or more before the user re-triggered the agent.

That report involved GPT-5.4, OpenClaw 2026.3.8 and Telegram. We would preserve the stop reason alongside the message timestamps because those details distinguish this symptom from a logged timeout.

An explicit failure after activity has started

Other reports describe a different sequence: a tool completes, but the following assistant response never arrives. The run then logs an abort, timeout or connection error.

In the Ollama glm-5.1:cloud report, a successful toolResult preceded This operation was aborted, with stopReason set to error.

Reports cover OpenClaw versions 2026.3.8, 2026.4.10 and 2026.5.7. However, those reports contain different symptoms. Keep the exact version and failure sequence together rather than treating every unresponsive session as an identical fault.

Trace stream_idle_timeout through the gateway logs

Work backwards from the failed response, but reconstruct events in timestamp order. We recommend this sequence:

  1. Find the affected user message and its run identifier. Keep unrelated sessions out of the comparison.
  2. Locate the last completed event, such as assistant output, a tool call or toolResult.
  3. Read forward to the first timeout, abort or connection error. Record its exact wording and timestamp.
  4. Follow any fallback decision into the replacement model request. Keep that outcome separate from the original failure.

These log entries point to different events:

Log entryWhat it records
embedded run timeout: runId=... timeoutMs=15000An embedded-run timeout reporting a 15,000 ms limit
decision=fallback_model reason=timeoutA decision to use a fallback model because of a timeout
FailoverError: LLM request timed out.A request timeout reported during failover handling
This operation was abortedAn aborted operation

The useful distinction is what happened first. A model fallback line records a recovery decision; it doesn’t explain every earlier event.

Isometric diagnostic stack diagram with translucent layers showing a brain, circuit icons, and a highlighted fault path.

Keep a compact log excerpt containing the last successful event, the first failure and the fallback decision. That gives you a clearer diagnostic record than a screenshot of the final error alone.

Compare the configured timeout with the logged limit

What the 500-second report actually shows

Two configuration keys appear in the timeout report: agents.defaults.timeoutSeconds and agents.defaults.llm.idleTimeoutSeconds.

The reporter set both to 500 seconds. Nevertheless, the gateway logged embedded run timeout: runId=... timeoutMs=15000.

The arithmetic matters: 15,000 milliseconds is 15 seconds, while 500 seconds is 8 minutes and 20 seconds. Those are different limits. Don’t describe that logged event as a request reaching the configured 500 seconds.

We would retain both configuration values beside the log excerpt. Together, they document the mismatch without assuming why it occurred.

A useful comparison needs matching run details

For your diagnostic record, pair the configured values with the OpenClaw version, provider, model identifier and affected run ID. Also retain whether streaming was enabled and whether the failure happened before or after a tool result.

Then compare the elapsed time with the logged timeoutMs value. This separates an observed timing limit from an expectation based on configuration.

If another attempt still reports timeoutMs=15000, record that unchanged result. Repeatedly increasing the two settings adds little diagnostic value when the same 15-second limit remains in the log.

Separate slow CPU inference from network failure

Local Ollama can spend substantial time starting

In the CPU-based Ollama report, qwen2.5:7b took 2 minutes and 32 seconds to complete a warm-up request. The runner needed 11.87 seconds to start.

Those timings matter when assessing a short timeout. Runner startup alone occupied much of the 15-second limit reported in the gateway log.

For local Ollama diagnosis, measure startup separately from a subsequent Ollama request after the model is loaded. Keep the same model and a comparable prompt so the comparison remains useful.

A successful warm-up establishes that the model answered that request. It doesn’t establish that the next OpenClaw run will complete within its logged limit.

Cloud connection errors need their own evidence

The glm-5.1:cloud report included fetch failed | read ECONNRESET. It also recorded Client network socket disconnected before secure TLS connection was established.

Preserve those messages exactly. Together, they show a failure in the network connection, a detail missing from a generic “LLM request timed out” summary.

Also record the provider used before and after failover. The reported fallback chain could route from ollama/glm-5.1:cloud to nvidia/z-ai/glm-5.1 after a timeout.

That distinction matters because a response from the fallback model doesn’t demonstrate that the original provider connection recovered. Treat the two requests as separate outcomes, even when they belong to the same user task.

Follow tool results through context compaction

A completed tool call is an important checkpoint, but the agent still needs to produce its next response. When a run stalls there, inspect everything between toolResult and the next assistant message.

The cloud Ollama report also contained Context overflow: estimated context size exceeds safe threshold during tool loop. Auto-compaction followed on attempt 1/3.

Keep that sequence intact. Context overflow, compaction and a later abort are separate recorded events, even when they occur close together.

For a controlled comparison, we recommend a shorter task with less accumulated tool output. Preserve the original session logs first, then change one condition at a time. Otherwise, a fresh session, different model and changed timeout leave you unable to identify which change affected the outcome.

Also retain the tool’s completion status. A successful tool result followed by an aborted model request is different from a tool that never returned.

We wouldn’t repeat a completed tool action merely because the assistant response is missing. First establish whether that action changed anything outside the conversation.

Recover without repeating completed actions

Sending “continue” re-triggered activity in the silent-stop report. That is useful recovery evidence, but it doesn’t establish a lasting remedy.

Before retrying, inspect the last tool result. This matters most when a tool can send messages, write files or change an external system.

An absent assistant response doesn’t mean the preceding tool action failed. Retrying the entire task can repeat an action that already completed.

Proposed remedies include auto-retry on partial output, a session health check, timeout-based recovery prompts and an external monitor service. These remain proposals rather than confirmed fixes.

We favour a recovery approach that records what completed before requesting another attempt. Any automated retry should also have a bounded attempt count and retain the original failure details.

For monitoring, distinguish useful progress from connection activity. Record completed messages and tool results, then flag a run that remains inactive beyond your chosen operational threshold. An alert should preserve enough context to support a safe decision about resuming or stopping the task.

Questions about stuck OpenClaw sessions

Does sending “continue” solve the problem?

It can re-trigger a session, as the GPT-5.4 report describes. However, that observation doesn’t show why the agent stopped or whether the same interruption will recur.

We would use the new message as a recovery action after reviewing completed tool work. Keep the original timestamps and stop reason rather than replacing the diagnostic record with “worked after retry”.

What should a useful issue report contain?

Include the OpenClaw version, model identifier, provider, streaming setting and both timeout values. Add the affected run ID and a timestamped sequence covering the last successful event, first failure and fallback outcome.

For local Ollama, include measured startup and warm-request timings. For cloud requests, retain connection errors verbatim. Remove credentials and private content, but preserve the order and timing of events.

Does a successful fallback explain the original stall?

No. It shows that a replacement request produced an outcome. In the reported chain, the request could move between Ollama and NVIDIA, so the provider path changed.

Keep the original failure and fallback response in separate parts of your record. That makes it easier to distinguish recovery success from evidence about the first request.

Start with the first failure, then choose recovery

The strongest diagnosis comes from a clear sequence: the last completed event, the first failure and the recovery decision. Exact log entries are more useful than treating every stalled session as one timeout problem.

We recommend preserving that sequence before changing settings or retrying. It gives you a firmer basis for recovery and reduces the risk of repeating work the agent already completed.