OpenClaw merged a high-priority agent-runtime fix in PR #125357, restoring failure causes when subagent errors are reported through the pending-error timeout path.
The bug was easy to miss because the status looked legitimate. A child agent could fail inside the lifecycle error retry grace window, and the parent would receive a timeout result. But that wait result already carried both the failure text and a pendingError flag. The announce path rebuilt it as a plain timeout anyway, dropping the useful cause.
For a parent agent, that distinction is not cosmetic. A genuine budget timeout suggests retrying with a narrower task or reporting that the child ran out of time. A runtime fault, provider failure, or malformed tool-call path points somewhere else entirely. Without the cause, the parent can neither recover well nor explain the real blocker to the user.
What Changed
The fix is narrow. In subagent-announce-output.ts, the timeout branch now preserves terminalOutcome.error or the wait error only when wait.pendingError === true. Genuine budget timeouts keep the existing plain timed out label.
The user-visible internal status label now has room for the recovered cause:
- budget timeout:
timed out; - pending-error timeout:
timed out: <cause>; - failure:
failed: <cause>; - success:
completed; ready for parent review.
The PR also caps the status label before it enters parent-agent prompt context. That matters because failure text can come from providers or lifecycle paths, and previously the child result was bounded while the status line itself could grow without a shared cap. The implementation reuses OpenClaw's UTF-16-safe truncation helper, so it does not split emoji or other astral characters at the boundary.
Why It Matters
Subagents are most useful when the parent can treat their result as reliable evidence. A vague timeout is almost the opposite: it blurs whether the task was too large, the runtime failed, or the provider returned an actionable error.
PR #125357 improves that contract without changing ordinary timeout behavior. It simply keeps the cause when the runtime already knows the timeout is standing in for a pending error.
That should make multi-agent work easier to diagnose in long-running tasks, especially where child sessions run out of process and the parent learns about them through agent.wait.
Validation
The PR includes unit coverage for three projection cases: pending-error timeouts keep the cause, budget timeouts do not gain one, and timeout error text without pendingError is ignored. It also tests that the formatted announce message includes the new timed out: <cause> label.
Runtime evidence was produced against a real gateway and real HTTP model endpoints. The PR also reports clean production and test type lanes, lint, formatting, dead-code, architecture, max-lines, Plugin SDK declaration, docs, and protocol checks.
Bottom Line
OpenClaw parent agents now get a more faithful subagent completion signal. When a child actually fails, the parent can see the cause instead of being left with a misleading generic timeout.