Skip to content

0019-completion-timeout-for-hanging-process.md

Latest commit

 

History

History
16 lines (9 loc) · 2.48 KB

0019-completion-timeout-for-hanging-process.md

File metadata and controls

16 lines (9 loc) · 2.48 KB
Raw

Completion timeout: force-complete a hanging process after the completion signal

When an agent emits the completion signal but its process does not exit — because a child it spawned (a subprocess like gh, or a long-lived MCP server) holds the sandbox exec's stdout open so EOF never arrives — Sandcastle previously hung until the full idle timeout (default 10 minutes) and then failed the iteration with AgentIdleTimeoutError, discarding the already-committed work from the result. We now scan the agent's accumulated output stream for the completion signal and, once it is seen, replace the idle timeout with a one-minute completion timeout; if the process is still silent when that expires, we complete the iteration successfully (collecting commits via the normal path) and warn that the process is hanging. A clean process exit within the window still completes normally, so healthy runs are unaffected and gain no added latency.

Considered alternatives

  • Short-circuit on the agent's terminal stream event rather than the signal string. Rejected: no reliable terminal event exists across providers. Claude Code emits a single terminal result event, but Sandcastle synthesizes a result event for every agent message for Codex (AgentProvider.ts:551) and OpenCode (715), so keying on it would terminate those agents after their first message. Building a real per-provider terminal event is larger work than this bug warrants.
  • Kill the process immediately on signal detection. Rejected: useful data trails the signal — Codex's token usage rides on turn.completed, Claude Code's canonical result text comes last, and a run using both completion signal and structured output may emit the <tag> payload after the marker. The silence-based window captures trailing output before giving up.

Scope

The completion timeout is gated strictly on the completion signal. A process that hangs before any signal is emitted is indistinguishable from an agent genuinely stuck mid-work, so it still rides the full idle timeout and fails.

Consequences

Force-completing abandons the hanging process. For container-based sandbox providers, container teardown (docker rm -f) kills it. For the no-sandbox provider, close() is a no-op and there is no proc.kill() anywhere in the codebase, so the abandoned agent process and its children leak on the host. This extends an existing timeout/abort-path leak to the common success case and may warrant an explicit process kill in follow-up.