Greptile Summary

This PR fixes a roundtrip-safety bug where memory_search followed by memory_get would fail when QMD session exports lived under the workspace's qmd/ subdirectory. Because qmd/ is a reserved virtual-path prefix, memory_search was returning the raw workspace-relative path (e.g. qmd/sessions/session-1.md) which memory_get could not resolve — it expected a collection-scoped form like qmd/sessions-<agentId>/session-1.md.

The fix is minimal and targeted:

• In buildSearchPath, a new guard detects when the workspace-relative path equals "qmd" or starts with "qmd/" and redirects to the collection-scoped virtual form qmd/<collection>/<collectionRelativePath>, matching what resolveReadPath already expects.
• The sanitized variable was moved above the insideWorkspace branch so both code paths share the same leading-slash-stripped value without duplication.
• A regression test verifies the end-to-end roundtrip: search returns the collection-scoped path, and readFile succeeds with that exact path.

No logic errors or security concerns were found. The change is well-scoped, correct, and the accompanying test directly covers the reported failure mode.

Confidence Score: 5/5

• This PR is safe to merge — it is a minimal, well-tested bug fix with no regressions or side effects.
• The change is small (two files, ~20 lines of production code), correctly addresses the described root cause, does not weaken any security boundary in resolveReadPath (path-escape checks are unaffected), and is backed by a focused regression test that validates both the search path and the read-back roundtrip.
• No files require special attention.

_{Last reviewed commit: 8543f21}

Follow-up on the earlier red CI:

• the previously failing jobs were rerun by pushing an empty retrigger commit
• the failures did not reproduce
• all relevant PR checks are now green, including:
  • check
  • build-artifacts
  • checks (node, test, ...)
  • checks (bun, test, ...)
  • checks (node, extensions, ...)
  • checks (node, protocol, ...)
  • all checks-windows shards
  • install-smoke

This looks like transient CI / setup-install flakiness rather than something caused by the QMD roundtrip fix itself.

Thanks for the original fix attempt here, @holgergruenhagen.

I reworked this on top of current main in #57560. The underlying bug is still real, but the current code path now lives in the QMD manager's workspace-path translation logic, so the narrow current fix needed a fresh port rather than landing this branch directly.

Closing this as superseded by #57560 to keep the queue clean while preserving attribution.

Thanks for the rework in #57560, @vincentkoc! Makes total sense to port it fresh given the code path has moved since. Appreciate you confirming the root cause and keeping the fix alive — looking forward to seeing it land. 🙌