mirror of
https://github.com/openclaw/openclaw.git
synced 2026-07-06 21:03:38 +00:00
e845a26fd6
* fix(agents): fail fast with attributable reason after MCP stdio session dies mid-run Wires MCP Client onclose/onerror during bundle-mcp session creation so a crashed/exited server flips session.connected instead of staying stale. Next tool/resource/prompt call throws a domain-specific 'is disconnected' error immediately instead of surfacing the SDK's generic 'Not connected'. A disconnected reused session is retired and rebuilt fresh on the next catalog pass rather than reused, since the SDK chains onclose/onerror cumulatively on repeat connect() and the stdio transport never clears its read buffer on an unexpected exit. * fix(agents): retire a reused MCP session that dies mid-refresh, not just pre-refresh codex review found: the catch-path retirement in getCatalog()'s per-server task only covered two cases (fresh session that never connected this pass, and a non-reused session that failed for any reason) - a reused session that was healthy when this pass started but disconnects mid-refresh (child process dies between ensureSessionConnected() returning and listAllToolsBestEffort() finishing) hit neither branch, so onclose flipped connected=false but the dead session object stayed in the map until the next catalog rebuild happened to notice it. Added the missing branch plus a regression test that kills the child process mid tools/list on a reused session and asserts the session is purged from the map within the same refresh (not just fail-fast on tool calls, which already worked). * fix(agents): retire a dead reused MCP session even across overlapping catalog generations codex round-2 review found: the mid-refresh retirement branch still skipped when sharedWithNewerGeneration was true, which protects a still-alive session another generation is actively using - but once onclose flips session.connected=false, the transport is dead for every generation sharing that object, so that guard no longer applies. Simplified to a single !session.connected branch that always retires (retireSessionIfCurrent already no-ops safely if a newer generation replaced the map entry), and dropped the now-write-only connectedForCatalog local it replaced. * test(scripts): allow MCP SDK callback suppressions * fix(agents): retire closed MCP sessions --------- Co-authored-by: Vincent Koc <vincentkoc@ieee.org>