[JDProfresh]

Problem

POST /api/sessions/:id/paste-image writes the upload under session.workingDir (through ensureUploadDir) with no bounded check first. If that workspace sits on a network mount that has stopped answering, the lstat/mkdir inside ensureUploadDir is an unbounded async call that holds one of libuv's few threadpool workers (4 by default, shared with every other fs, dns.lookup and crypto call) until the mount answers. A handful of uploads onto a dead mount can drain the pool and stall the whole server.

Cause and precedent

This is the #516 rule: a user-chosen path is probed BOUNDED, never with a sync or unbounded async fs call. #574 round one bounded the hourly upload sweep and the per-session delete with probePathKind() plus pastCap, but the upload REQUEST path was left on a raw ensureUploadDir. POST /api/sessions already does the right thing one function over: a single probePathKind(workingDir, { pastCap: true }) at entry that refuses an unknown mount. The paste route just never got the same guard.

Fix

• Probe session.workingDir with probePathKind(..., { pastCap: true }) right after the remote-session 400 and before anything reads the body or touches the workspace. Refuse only unknown (the mount is not answering) through describeUnknownPath('Workspace', ...) with OPERATION_FAILED, exactly as the create route does. absent and not-a-directory keep ensureUploadDir's own errors, since a live session's workspace should be neither.
• pastCap: true because this is one path at the user's explicit request, so unrelated dead mounts must not refuse it; it still stops at the ceiling that keeps one worker free.
• The probe is a preflight, not a wrapper around the write. A mount that dies between the probe and ensureUploadDir is the same residual race POST /api/sessions accepts: a pre-check cannot bound the write that follows it, and a second probe would only narrow, never close, that window. I kept one probe at entry and stated the race in the code comment and the invariants rather than chase it with more checks.

Tests

• test/routes/session-routes.test.ts (paste-image block): one test pins probePathKind(session.workingDir, { pastCap: true }) at the call site, with the mock cleared right before the request so only that upload's call can satisfy the pin. Another drives an unknown result and asserts the exact describeUnknownPath message, the 422 / OPERATION_FAILED envelope, and that nothing is written. Each guard was shown to fail with it removed: drop the probe and both go red; flip pastCap to false and the pin goes red; change the 'Workspace' label and the message assertion goes red.
• test/routes/session-create-unreachable-path.test.ts: added a paste-image case to the existing dead-mount harness (a hung stat on a simulated unreachable mount). It asserts the route answers well inside the block window and does no synchronous fs on the dead mount, the same shape as the create and quick-start cases already there.
• The full suite passes apart from one failure on my run, session-input-wait's wall-clock "blocks for its full timeout", which passed 3 of 3 on its own; it is a timing flake on my host, unrelated to this diff.

Verified

The dead-mount test reproduces the stall directly: with the mount's stat hung, the route returns the "not responding" refusal promptly instead of parking a worker, and records no synchronous fs touch. The probe pattern this copies has shipped on POST /api/sessions in released builds. It also runs in my dev install's 1.41.0 build together with the other open PRs of this batch, where image, video and document uploads behave as before on a healthy workspace.

Note

The Attach button and video uploads PR I'm sending right after this one adds a second caller of the same ensureUploadDir, so the entry probe covers it too. Both add to the Prompt uploads paragraph in CLAUDE.md and the paste-image test block; I'll rebase whichever lands second.

[Ark0N]

Thanks @JDProfresh, shipped in 1.41.1, merged just before #606 so the entry probe covers the video path too. No changes at merge.