User-mode systemd units and launchd agents now prepend the loading user's
~/.local/bin to PATH, so user-installed agent CLIs (claude, codex, etc.,
installed via npm user prefix, pipx, fnm/nvm) are discoverable by the
kandev daemon. Without this they appeared as `not_installed` even though
they were on the user's interactive PATH.
System-mode units keep the original system-only PATH.
* fix(cli): prepend node bin dir to service unit PATH so npx-based agents resolve
systemd/launchd units hard-coded a system-only PATH. Under per-user node
managers (fnm, nvm, asdf, volta, mise), node lives in a versioned
subdirectory like ~/.local/share/fnm/node-versions/v24.14.0/installation/bin,
so the daemon could exec node (ExecStart resolves the absolute path) but
could not exec npx — every ACP agent built with `npx -y <bridge-pkg>`
(claude, codex, opencode, …) failed at spawn time with "peer disconnected
before response".
`captureLauncher()` already records the node binary path. Its parent dir
is exactly where the matching npm/npx live, so prepend it to the unit's
PATH (with dedup so /usr/local/bin etc. don't appear twice).
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
* fix(cli): guard pathWithNodeBinDir against non-POSIX nodePath
`path.dirname` returns `'.'` on POSIX when given a path with no `/`
separator (e.g. a Windows-style `C:\…\node.exe`). Prepending `.` to a
service unit's PATH would put CWD ahead of system dirs in the daemon
context — a latent privilege-escalation footgun if any future caller
ever passed a relative or bare-filename nodePath. Production is safe
because `captureLauncher()` uses `process.execPath` (always absolute on
Linux/macOS), so the guard is defensive, not a behavior change.
Adds regression tests for both systemd and launchd, plus a previously
missing launchd system-mode + versioned-bin-dir test.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
---------
Co-authored-by: Kandev Agent <agent@kandev.dev>
Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>