Files
hermes-agent/apps
Teknium 2b7f49673f fix(desktop): self-heal dropped SSH/HTTP registered remote connections
A dropped registered remote connection (SSH or HTTP) never recovered on
its own: the next boot attempt failed with a transient transport error
("Could not verify the existing SSH backend", ERR_CONNECTION_RESET,
mint timeout), the failure was correctly NOT latched, but nothing ever
re-attempted the boot — the renderer's reconnect machinery only arms
after a completed boot. The app parked on "Desktop boot failed" until
the user manually deleted and re-entered the same connection details,
which merely forced the fresh bootstrap an automatic retry would have
performed (issue 82679, feature ask 80430).

Root causes and fixes:

- electron/backend-start-failure.ts: new isRetryableRemoteBootFailure()
  predicate — a remote, non-reauth boot failure is transient and may be
  retried; local failures and confirmed 401/403 rejections are not
  (a missing capability differs from a transient failure).
- electron/main.ts: the boot-failure progress broadcast now carries
  `retryable` (rides with `error` through updateBootProgress), and a
  failed reuse probe against a cached SSH master tears the stale
  master/tunnel down so the next attempt bootstraps fresh — exactly
  what manual re-entry did.
- use-gateway-boot.ts: bounded self-heal loop for a failed boot whose
  progress is marked retryable — up to 5 re-attempts with the same
  full-jitter backoff as the socket reconnect loop (2s base, 15s cap).
  Exhausted retries end in the real boot-failure recovery overlay,
  never an infinite spinner. Reset on success and on soft switch;
  timer cleared on unmount.
- store/boot.ts: resumeDesktopBootForRetry() re-arms the overlay with a
  retry status while an automatic retry is in flight.

Secondaries already had full-jitter backoff (store/gateway.ts); this
closes the same class for the PRIMARY/registered-connection path.

Tests: predicate matrix (retryable vs reauth-latch mutually exclusive),
plus renderer hook tests proving a transient SSH failure self-heals on
the next attempt, retries are bounded (6 total dials then the recovery
overlay, no further attempts), and non-retryable failures never enter
the loop. Sabotage-verified (disabling either half fails 4 tests).

Fixes #82679
Fixes #80430
2026-08-16 19:58:37 -07:00
..