feat: reconnect with backoff and session re-establishment #2

Merged

w33ble merged 1 commit from feat/reconnect into main

2026-09-29 07:27:26 -06:00

hijynx commented

2026-09-28 14:07:59 -06:00

Collaborator

Summary

Reconnect loop around probe/Connect and the message pumps: initial-connect failures and post-connect drops retry with backoff (1x/2x/4x/8x of --initial-retry-interval, capped) until --max-disconnected-time elapses (unset = retry forever).
Session re-establishment on each reconnect: replays the last-seen initialize on the fresh transport, discards the response, sends notifications/initialized, then flushes stdin lines buffered while down (64-line cap, drop-oldest).
Per-call ErrRejected write failures no longer kill the connection; stdin EOF mid-backoff still exits 0.
New flags: --max-disconnected-time (default infinity) and --initial-retry-interval (default 5s), both with MCP_BRIDGE_* env fallbacks.
Fixes parseTrailingFlags no-op so flags after the positional URL are honored.
Tests: backoff calc, flag/env parsing (incl. invalid values), rejection detection; existing smoke test untouched. Build, vet, lint, and tests all green.

Why

Previously any network hiccup or remote restart killed the bridge process and the MCP client saw a dead server until the client session was restarted manually. The proxy now rides out transient outages transparently.

Summary - Reconnect loop around probe/Connect and the message pumps: initial-connect failures and post-connect drops retry with backoff (1x/2x/4x/8x of --initial-retry-interval, capped) until --max-disconnected-time elapses (unset = retry forever). - Session re-establishment on each reconnect: replays the last-seen initialize on the fresh transport, discards the response, sends notifications/initialized, then flushes stdin lines buffered while down (64-line cap, drop-oldest). - Per-call ErrRejected write failures no longer kill the connection; stdin EOF mid-backoff still exits 0. - New flags: --max-disconnected-time (default infinity) and --initial-retry-interval (default 5s), both with MCP_BRIDGE_* env fallbacks. - Fixes parseTrailingFlags no-op so flags after the positional URL are honored. - Tests: backoff calc, flag/env parsing (incl. invalid values), rejection detection; existing smoke test untouched. Build, vet, lint, and tests all green. Why - Previously any network hiccup or remote restart killed the bridge process and the MCP client saw a dead server until the client session was restarted manually. The proxy now rides out transient outages transparently.

hijynx added 1 commit

2026-09-28 14:07:59 -06:00

feat: reconnect with backoff and session re-establishment

CI / Build, vet, test, lint (pull_request) Successful in 52s

Details

CI / Build, vet, test, lint (push) Successful in 2m29s

Details

76c5a8bf1d

Add --max-disconnected-time (default infinity) and
--initial-retry-interval (default 5s), with MCP_BRIDGE_* env
fallbacks. Retry initial connect and post-connect drops with
1x/2x/4x/8x backoff; replay stored initialize, send
notifications/initialized, then flush stdin buffered while down.
Per-call ErrRejected failures no longer kill the connection.

Also fix parseTrailingFlags no-op so flags after the positional
URL are honored.