You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Commit 05ec1d6
Browse filesBrowse the repository at this point in the historyBrowse files
fix(server): Claude probes no longer spawn in the packaged server's home cwd
Packaged desktop builds deliberately start the Node server with
cwd = $HOME (DesktopEnvironment's backendCwd). The Claude driver's
instance-level auxiliary spawns — the SDK capability probe and the
'claude --version' health check — inherited that directory: the driver
handed its startup cwd to the probe, and runClaudeCommand passed no cwd
to ChildProcess.make at all. Third-party 'claude' shims that treat
their cwd as the project boundary (container runtimes, mount-scoped
sandboxes, governance shims) then fail the probe, which runs first and
marks the provider Unavailable before any bound project is reached.
The driver now passes the server's stateDir — the app's own user-data
directory, present in dev, packaged, and WSL alike, and never a project
— as the explicit spawn cwd, and runClaudeCommand threads that cwd into
ChildProcess.make instead of silently inheriting. This is the narrow
option-B fix from the issue: backendCwd (fix A) and the cross-driver
env/cwd consolidation are untouched, and the per-workspace probe path
(snapshotForCwd) keeps receiving the real workspace cwd.
Model: GLM-5.3-Flash (Z.ai) · Harness: ZCode CLI
0 commit comments