[zudsniper]

fixes the team_spawn crash + silent worktree failure from #40 on opencode 2.0.19+, plus the stuff that still stops teammates from working after that.

vs #41 / #43

both guard ctx.permission.rules, #41 also validates depends_on. this does both, plus:

▸ worktree root cause: #41 adds a warning but couldn't find why worktree.create made nothing. 2.0.19+ worktree endpoints need a projectID. ensemble never sent one → creation failed → everyone shared the base tree. now passes ctx.location.project.id to create/list/remove/refresh

▸ detached worktrees (2.0.32+): opencode now does git worktree add --detach (packages/core/src/git.ts). teammate commits land on a detached HEAD → team_merge: "not something we can merge". if the new worktree has no branch, ensemble creates it

▸ edits stuck on approval: v2 checks in-location files as relative paths, so edit <worktree>/** allow never matched → ask nobody answers. worktree teammates also get edit * allow. outside-the-worktree stays behind external_directory

▸ read-only detection: was name-only (plan, explore). custom reviewers got a worktree + shell * allow. now reads the agent's own rules via agent.get, names as fallback

not touched: the 4th #40 thing (notifications redelivered after purge).

on permission.rules: since 2.0.19 core merges session rules after agent rules (merge(agent, session), last match wins), so the session.create permissions already win. skipping rules() when it's gone loses nothing.

verification

✓ bun run typecheck clean
✓ bun test 857 pass / 0 fail (new: projectID pass-through, detached naming, read-only from rules, edit *, unknown depends_on)
✓ bun run lint only pre-existing warnings
✓ live on 2.0.24 + 2.0.32:

team_create → team_spawn(build, worktree) → own worktree + branch → commit
           → team_shutdown → team_merge ✓ → team_cleanup

✗ not verified: subagent-mode teammates (explore etc.) on opencode's free big-pickle. they get "free tier can only be used from within OpenCode", built-ins too, so it looks like a host/provider thing, not ensemble.

🤖 Generated with Claude Code

https://claude.ai/code/session_01GUtz6gBn7rhXaufK97d8cm

[hueyexe]

Merged — thank you for the comprehensive 2.x compat work. Post-merge review applied four follow-up fixes on top (commit 024967a): (1) team_spawn now binds the session to the worktree directory when workspace binding is unavailable, so the edit * allow rule is genuinely scoped to the worktree instead of the whole shared tree — the premise your comment relied on wasn't established without it; (2) gitCreateBranch failures are now logged and fall back to git switch for the existing-branch re-spawn race, and worktree.create records only the effective branch — recording the requested name unconditionally could point team_merge at a stale same-named branch and silently lose the teammate's new work; (3) the guard now uses ctx.permission?.rules to survive a missing permission domain entirely; (4) PR #41's isolationInfo spawn-result warning was cherry-picked so the lead is told plainly when isolation did not complete (issue #40 bug 2's reporting ask). One minor behavior to note: with agent.readOnly returning false, the ?? only falls through on undefined, so a host that reports its plan agent as writable gets plan teammates with worktrees and bash * allow — called out in the release notes.