You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
feat(harness): Retry a dead attachment, so a typed HTTP key stops waiting for a relaunch #500
The third option from #469, named in ADR-0008's consequences when #482 shipped the second.
Where things stand
A Harness handle stays the configured Completer once it exists, alive or not. #482 made the interface honest about the consequence: when the child never spawned or has died, the HTTP rows stay editable and say plainly that a key typed there is read at the next launch, and the state line says static weights are answering meanwhile.
That is truthful, and it still asks the user to restart the app to recover.
What would actually fix it
Make the attachment re-openable, so a dead Harness is retried and the question stops arising. Then a user whose Harness died gets it back without a relaunch, and nobody has to be told that a typed key is waiting.
Two things make this larger than it looks, which is why #482 did not attempt it:
The attached session is a process-global held for the app's lifetime, so re-attaching means a re-initialisable global plus shutdown ordering inside the harness module.
ADR-0008's rule stands throughout: a session call that does not land falls to static weights. Recovery must not become a second mind answering behind one session's failure, which is what treating a dead handle as unconfigured would have done.
Worth deciding alongside it: whether the Completer source row can then switch without a relaunch at all, which #452 answered with "next launch" for the same reason.
Acceptance
A Harness that died is retried without a relaunch, and the buddy uses it again once it answers.
A failed session still falls to static weights, never to the HTTP Completer, unless the user chose that.
Activity