mrbeandev/dsh-reconnect

Retries recoverable model requests in DSH with exponential backoff, provider Retry-After support and a settings card.

Retries recoverable model requests so a temporary network, gateway or provider failure does not end an agent turn or a long-running task. It emits the standard llm/retry and llm/retry-started session events, shows the retry count and countdown in the conversation UI, and exposes the delay cap, quota retries and unknown-error behaviour in a settings card whose changes take effect immediately.

Agent Capabilities ★ 0
View on GitHub ↗

Install

dsh plugin --profile web add dsh-reconnect

The package declares a dsh.bundle patch, so dsh plugin also adds it to the profile's active bundle list with no manual cordis.yml edit. Restart the running web process afterwards, then look for Settings → Plugins → Plugin configuration → ReConnect automatic retry to confirm the card is active. Update with dsh plugin --profile web update dsh-reconnect and remove with dsh plugin --profile web remove dsh-reconnect. A GitHub install can pin an exact tag or commit for reproducible installs. Needs Node ^22.19.0 or >=24.0.0 and the host settings service for the card.

Compatibility

It listens to the agent/request-error waterfall and takes over only after the built-in normal provider retry policy is exhausted, or when an always policy delegates recovery downstream. EMPTY_RESPONSE, RATE_LIMIT, SERVER, STREAM_CLOSED, TIMEOUT and TRANSPORT retry indefinitely; a missing or unconfigured model retries while configuration recovers; QUOTA is not retried unless retryQuota is on; unknown errors retry indefinitely by default and fall back to a three-attempt limit when that is disabled; tool, argument, authentication, credential, context-overflow and aborted errors are never retried. Tool execution failures in tool/result are outside its retry boundary — it uses the machine-readable failure code rather than inferring failures from tool output. Backoff is 1s, 2s, 4s and so on, capped at 60s by default, and a positive provider Retry-After is honoured as an uncapped minimum, so maxDelayMs limits only the local delay. It stops immediately when the turn aborts or the plugin unloads.

Details

Alternatives

humblebanana/dsh-record-replay · JuneLearn/dsh-reasoning-settings · dpskh/dsh-checkpoint

More plugins in Agent Capabilities

Browse more in Agent Capabilities

Guides for Agent Capabilities plugins