feat: proj:<name> spawn prefix (spawn in a SPAWN_CWD subdirectory)

[kwliang1]

What

Adds a proj:<name> (alias dir:<name>) topic prefix to doSpawnSession that spawns the session directly inside an existing subdirectory of SPAWN_CWD — the real checkout, no worktree. Mirrors the existing worktree:/wt: prefix parsing.

proj:car-ship-prototype <your task>

Why

Per-project tooling that keys off the session cwd currently can't distinguish projects, because every plain spawn shares the single SPAWN_CWD. Concretely: claude-mem shards memory by git-repo-root / directory — but with all sessions launching at SPAWN_CWD it labels everything one project and pools it.

proj:<name> lets each project's sessions launch in their own directory, so such tooling shards per project, without the ephemeral-branch / cleanup overhead of worktree: (which is great for isolated build tasks but deletes its branch on exit — wrong fit for normal interactive work on the real checkout).

How

• Parse ^(?:proj|dir):(\S+)\s+ from the topic (same pattern as worktree:), strip it so it doesn't leak into thread names/prompts.
• If set (and not a worktree spawn), resolve effectiveCwd = resolve(SPAWN_CWD, name), erroring if the dir doesn't exist. existsSync is already imported.
• Worktree path is unchanged; proj: and worktree: are mutually exclusive (only one matches ^).

Testing

• bun build daemon.ts --target=bun ✅
• bun build bridge.ts --target=bun ✅

Notes

Draft for review. No change to default behavior — plain spawns still land at SPAWN_CWD.

🤖 Generated with Claude Code

[sf8193]

PR #60 Review — proj:<name> spawn prefix

Reviewed by two specialist agents (sp-reviewer: architecture, typescript-reviewer: correctness). Findings merged and deduplicated below.

Not in diff but noted: The spawn_session tool schema in daemon/bridge-dispatch.ts has no proj parameter — agent sessions can't discover or use this feature via the tool contract. The router also has no spawn-proj: shorthand (unlike spawn-wt: for worktrees). Both make proj: invisible to callers without reading source.

[sf8193]

Blocker — path traversal (flagged by both reviewers)

resolve(spawnCwd, projTarget) will happily resolve ../../etc or an absolute path. There is no guard that projDir stays inside spawnCwd. A user with Discord spawn access can set effectiveCwd to any filesystem path.

Fix:

if (!projDir.startsWith(spawnCwd + '/')) {
  throw new Error(`proj target "${projTarget}" escapes SPAWN_CWD`)
}

[sf8193]

Should-fix — no mutual exclusion with worktreeTarget (flagged by both reviewers)

worktree:foo proj:bar topic will set both targets. The projTarget block runs first and sets effectiveCwd, then the worktreeTarget block silently overwrites it. The proj directive is discarded without error.

Fix: add a guard after both parsers:

if (projTarget && worktreeTarget) {
  throw new Error('proj:/dir: and worktree:/wt: prefixes cannot be combined')
}

[sf8193]

Should-fix — resume/respawn loses proj CWD (flagged by both reviewers)

The proj: prefix is stripped from topic here. SessionInfo stores the stripped topic but never records projTarget or projDir. When tryResume/tryRespawn reconstruct a session from dead.topic, the proj-pinned CWD is lost — the session respawns under spawnCwd instead of projDir.

Fix: persist projDir (or projTarget) in the registry entry alongside worktreeRepo/worktreePath, and pass it through the respawn path.

[sf8193]

Should-fix — bare proj:name silently falls through

The regex /^(?:proj|dir):(\S+)\s+/ requires trailing whitespace. A bare proj:myapp (no trailing text) won't match — the prefix becomes part of the topic/prompt with no error. The worktree regex has the same pattern, but this is a new instance worth fixing.

Fix: /^(?:proj|dir):(\S+)(?:\s+|$)/ and handle the end-of-string case when slicing.

[sf8193]

Nit — log line missing session name

All other spawn-path log lines include tmuxName for grepability (e.g. daemon: spawn ${tmuxName}: ...). This line omits it. tmuxName is already assigned by this point (line 166).