Windows: MCP child exits before initialize response without a recorded exit status #401

Closed
opened 2026-10-20 20:59:27 +02:00 by buildagent · 1 comment
Member

0.41.4 follow-up (2026-10-21): PR #402 removed an unsafe test-cleanup path that killed an unrelated live MCP process through a stale registration PID. A controlled real-child regression reproduced that mechanism before the fix and passes afterwards. Final native Windows qualification passed all 4,775 ordinary tests and the full workflow; 0.41.4 is published. The EOF exit-status diagnostic remains in place. This does not conclusively attribute the original run 6116 failure. This issue stays open pending a direct failure witness; a green rerun is not a root-cause proof.

Observed in PR #400, Windows run 6116 / job 43551: 4760/4761 ordinary tests passed; concise_search_text_keeps_the_snippet_origin_disclosure failed at support/mcp.rs:513 during initialize with mcp server closed stdout before responding. No product error appeared on inherited stderr. The snippet assertions were never reached. The native Windows job does not use the new Linux compiler cache, and product sources are unchanged from v0.41.3.

The harness currently drops the child exit status on EOF, so the log cannot distinguish a normal refusal from a native exception or external termination. Add that diagnostic and a real-child regression, exercise the affected MCP target locally, then require a complete green native run. Keep the original startup failure open unless its cause is identified; a subsequent pass alone is not a root-cause fix.

Run: https://git.h-dv.de/h-dv/code-index/actions/runs/1294

PR #400 now records try_wait() child status on stdout EOF, without retrying or relaxing timing limits. A real CLI refusal regression verifies the child's exit code 2 is retained. All 15 affected-target tests passed locally in snapshot and daemon modes, all 980 MCP tests passed under the CI profile, and Windows-GNU compilation passed.

The final PR head 3d285208187c97f84c4f7baed6234d5d6a6b641a completed native Windows qualification successfully: 4,762 ordinary tests, all 27 serial timing tests, doctests and plugin qualification. Run: https://git.h-dv.de/h-dv/code-index/actions/runs/1306 (39m29s). This establishes the diagnostic change passes qualification; the original startup failure's cause remains unknown and this issue stays open.

> **0.41.4 follow-up (2026-10-21):** [PR #402](https://git.h-dv.de/h-dv/code-index/pulls/402) removed an unsafe test-cleanup path that killed an unrelated live MCP process through a stale registration PID. A controlled real-child regression reproduced that mechanism before the fix and passes afterwards. [Final native Windows qualification](https://git.h-dv.de/h-dv/code-index/actions/runs/1380) passed all 4,775 ordinary tests and the full workflow; [0.41.4](https://git.h-dv.de/h-dv/code-index/releases/tag/v0.41.4) is published. The EOF exit-status diagnostic remains in place. **This does not conclusively attribute the original run 6116 failure. This issue stays open pending a direct failure witness; a green rerun is not a root-cause proof.** Observed in PR #400, Windows run 6116 / job 43551: 4760/4761 ordinary tests passed; concise_search_text_keeps_the_snippet_origin_disclosure failed at support/mcp.rs:513 during initialize with `mcp server closed stdout before responding`. No product error appeared on inherited stderr. The snippet assertions were never reached. The native Windows job does not use the new Linux compiler cache, and product sources are unchanged from v0.41.3. The harness currently drops the child exit status on EOF, so the log cannot distinguish a normal refusal from a native exception or external termination. Add that diagnostic and a real-child regression, exercise the affected MCP target locally, then require a complete green native run. Keep the original startup failure open unless its cause is identified; a subsequent pass alone is not a root-cause fix. Run: https://git.h-dv.de/h-dv/code-index/actions/runs/1294 PR #400 now records `try_wait()` child status on stdout EOF, without retrying or relaxing timing limits. A real CLI refusal regression verifies the child's exit code 2 is retained. All 15 affected-target tests passed locally in snapshot and daemon modes, all 980 MCP tests passed under the CI profile, and Windows-GNU compilation passed. The final PR head `3d285208187c97f84c4f7baed6234d5d6a6b641a` completed native Windows qualification successfully: 4,762 ordinary tests, all 27 serial timing tests, doctests and plugin qualification. Run: https://git.h-dv.de/h-dv/code-index/actions/runs/1306 (39m29s). This establishes the diagnostic change passes qualification; the original startup failure's cause remains unknown and this issue stays open.
Author
Member

Triage 2026-11-09: Closing as not reproduced. The exit-status diagnostic is now in place (tests/support/mcp.rs try_wait on EOF), and PR #402 removed the stale-PID kill path, a plausible cause. Reopen on the next occurrence: the diagnostic will print the status.

Triage 2026-11-09: Closing as not reproduced. The exit-status diagnostic is now in place (`tests/support/mcp.rs` `try_wait` on EOF), and PR #402 removed the stale-PID kill path, a plausible cause. Reopen on the next occurrence: the diagnostic will print the status.
Sign in to join this conversation.
No milestone
No project
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference
h-dv/code-index#401
No description provided.