Skip to content

fix(triggers): wait for a slow executor up to its own timeout, not 30 s (FD-5333) - #544

Merged
sarah-chitty-1 merged 1 commit into
mainfrom
sarahchitty/fd-5333-executor-read-timeout
Nov 11, 2026
Merged

sarah-chitty-1 merged 1 commit into
mainfrom
sarahchitty/fd-5333-executor-read-timeout

Conversation

@sarah-chitty-1

@sarah-chitty-1 sarah-chitty-1 commented Nov 8, 2026 •

Copy link
Copy Markdown
Collaborator

What

A trigger's nl action calls its executor (an AI-played contact) over A2A. The executor holds message/send open until it has replied, but the gateway gave that call a fixed 30 s timeout. Any contact slower than 30 s failed the fire: the trigger was marked failed and its later actions (like a notify) never ran, even though the reply usually still landed. The executor's own timeout (executor_timeout_seconds, 120 by default) never came into play.

In ML's prod runs this cut off 22% of contact replies (40% on Holiday Rush, 25% on Tree Removal). Linear: FD-5333.

Change

  • The send now waits up to the executor's timeout_seconds (never less than 30 s, so shorter settings behave as before).
  • Connecting and the follow-up task checks keep the 30 s limit.
  • While the call is open the trigger stays "firing", as before.
  • A contact slower than its executor timeout still fails the same way as today (ReadTimeout, trigger re-armed).

Tests

Two new unit tests against a local fake executor that answers late:

  • one slower than the HTTP timeout: the fire succeeds and the later action runs;
  • one slower than the executor timeout: the fire fails as before.

Both fail without the fix. The gateway and trigger unit tests pass: 941 with the fix, 939 on main, no failures either side.

Worth a look

  • A fire now holds its HTTP connection open for up to the executor timeout instead of 30 s. That's intended, but slow contacts keep a trigger "firing" longer.
  • Asking the executor for a non-blocking send ("configuration": {"blocking": false}) and relying on the existing polling would avoid the long-held call. It's left out here because it depends on every executor server supporting it. It would matter if a proxy in front of an executor drops idle connections before the executor timeout.
  • The 30 s barrier default (TRIGGER_BARRIER_TIMEOUT_S) is unchanged: a barriered call is still answered after 30 s while the fire carries on.

🤖 Generated with Claude Code

RetriggerView in GreptileConfidence Score: 5/5

The PR appears safe to merge.

Summary

The executor’s message/send now uses a read timeout of max(30, timeout_seconds). Connecting and task polls keep their 30-second limits.

  • Trigger actions wait for slow executor replies up to the executor's timeout.

Reviews (1) · Last reviewed commit: "fix(triggers): wait for a slow executor ..." · Reviewed by Greptile

… s (FD-5333)

An executor holds message/send open until its task ends, so the 30 s read
timeout failed every fire whose contact took longer, even though the reply
usually still landed. The send now reads for the executor's timeout_seconds
(never less than 30 s); connect and task polls keep 30 s.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@sarah-chitty-1
sarah-chitty-1 merged commit 7f44e73 into main Nov 11, 2026
34 checks passed
@sarah-chitty-1
sarah-chitty-1 deleted the sarahchitty/fd-5333-executor-read-timeout branch November 11, 2026 20:50
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants