fix(routing): retry ACP transport disconnects instead of failing outright (#2680)

Co-authored-by: nova28 <17953305+nova28@users.noreply.github.com>
Co-authored-by: Carlos Florencio <carlosmflorencio@gmail.com>
This commit is contained in:
nova28
2026-08-15 23:00:55 +08:00
committed by GitHub
parent 1ae8f4cf9f
commit de4c7fbf9c
24 changed files with 645 additions and 12 deletions
@@ -70,6 +70,7 @@ The initial policies are:
| Semantic cause | Kanban interactive policy | Office unattended policy |
| --- | --- | --- |
| `network_unavailable`, `provider_overloaded`, `provider_unavailable`, `model_capacity` | Short same-profile retry when high-confidence and replay-safe | Retry the same route for a few seconds; only degrade and fall through after the short budget is exhausted |
| `agent_transport_lost` | Short same-profile retry when high-confidence (the ACP wire-level transport-death signature) | Launch path: short same-profile retry (3 attempts), then routing gives up on the candidate without degrading the provider, since the resume token is provider-specific and fallback is not allowed. Post-start path: the routing handler declines to act (fallback disallowed short-circuits before the retry check), so a mid-turn transport loss is not retried or parked by Office routing |
| `rate_limited` | Retry only a confirmed short throttle; honor a bounded retry hint | Retry the same route when a validated wait is at most 60 seconds; otherwise treat it as a long-horizon route block and fall through |
| `quota_limited` | Show reset/remediation information; no automatic prompt replay | Fall through to the next configured provider; if all routes are exhausted, park until a validated reset or the Office quota backoff |
| `auth_required`, `missing_credentials`, `subscription_required`, `model_unavailable`, `provider_not_configured` | User action; no automatic retry or route change | Mark the failed route user-action-required and try another configured route; if none works, block and surface inbox/settings remediation without a timed retry for that route |
@@ -48,8 +48,8 @@ failure consistently without sharing the same retry policy.
### Kanban interactive recovery
- `network_unavailable`, `provider_overloaded`, `provider_unavailable`,
`model_capacity`, and confirmed short `rate_limited` failures are eligible for
automatic retry.
`model_capacity`, `agent_transport_lost`, and confirmed short `rate_limited`
failures are eligible for automatic retry.
- Before replaying, Kandev verifies that the prompt has produced no assistant
content or tool activity, unless the adapter provides a resumable retry
guarantee. A transient failure after potentially effectful progress remains a
@@ -78,9 +78,13 @@ failure consistently without sharing the same retry policy.
- Office consumes the shared code through its own routing policy. It never uses
Kanban's short prompt timer or retry-attempt budget.
- The short same-route phase applies to every Office run, even when provider
fallback is disabled. Fallback after exhaustion still requires an enabled,
explicitly configured alternative.
- The short same-route phase applies to supported Office retry codes, even when
provider fallback is disabled. Fallback after exhaustion still requires an
enabled, explicitly configured alternative.
- `agent_transport_lost` is launch-only in Office. The launch path retries the
same execution profile three times, then stops routing that candidate without
degrading the provider or selecting a fallback. The post-start path declines
to handle this code, so the existing failure path handles the run.
- Network interruption, temporary provider unavailability/overload, model
capacity, and a validated rate-limit wait of at most 60 seconds first retry the
same execution profile after nominal 5, 10, and 20 second delays with jitter.