[kh0pper]

What was failing

tests/workspace-calendar.test.js "T10-I3: after_etag comes from the PUT's ETag…" failed in CI (seen on #535, and once in a local full suite) with Cannot read properties of undefined (reading 'uid'): ws_cal_create_event answered workspace_unreachable ("connection failed").

Root cause

It depends on load, not on the date or on test order.

• The test right before it (T10-I2 "a rule too dense to expand…") runs a synchronous repeat expansion. On the CI runner it took 6.6 s; locally it takes about 2 s.
• During that time the event loop is blocked. The fake Nextcloud's keep-alive timeout is Node's default 5 s, so the server closes the idle pooled socket. fetch never sees the close in time and sends T10-I3's first request (the calendar PROPFIND) on that socket. The request fails with ECONNRESET, and ncFetch maps that to workspace_unreachable.
• Reproduced: blocking the loop for 3 to 7 s at the start of T10-I3 fails it every time (the logged cause is fetch failed / ECONNRESET on the PROPFIND). Blocking for 2.5 s or less passes.

The product has the same hazard. A real Nextcloud (Apache) also drops idle keep-alive sockets after a few seconds, and the same dense expansion blocks the real Workspace server too, so the next tool call could fail with "Crow Workspace did not answer".

Fix

ncFetch (bundles/workspace/server/nc/http.js) resends a request once when it fails with a stale-socket reset (ECONNRESET / EPIPE / UND_ERR_SOCKET). The resend goes on a fresh connection, and only for methods that are safe to send twice: GET, HEAD, OPTIONS, PROPFIND, REPORT, SEARCH, PUT and DELETE. A conditional PUT whose first attempt did land answers 412, which becomes changed_concurrently, never a double write. Some failures are never resent:

• POST, MOVE, MKCOL, PROPPATCH and LOCK.
• Timeouts.
• Refused connections.

Workspace manifest bumped to 0.2.3 (registry regenerated).

Tests

• New tests/workspace-nc-http.test.js uses a server that resets the second request on a kept-alive socket. It checks that each idempotent method is resent once on a new connection, that POST and MOVE go out exactly once and fail as workspace_unreachable, and that a refused connection is not retried. The resend cases fail without the fix and pass with it.
• tests/workspace-calendar.test.js passes, including with a 6 to 7 s forced event-loop block before T10-I3 (it failed 100% that way before the fix).
• Full suite: 10426 tests, 10391 pass, 0 fail, 35 skipped (taskset -c 0-11,16-27 npm test).

https://claude.ai/code/session_01N9fXzvNy8rCEArj1y5p6fp