[BUG] Agent teams crash on spawn with --model opus teammates (2.1.119, Bun stack trace)

What's Wrong?

Agent teams crash at spawn when teammates are assigned --model opus. The tmux panes get created and the claude launch command is typed in, but every spawned claude process exits immediately inside the Bun-bundled CLI. The team appears spawned (TeamCreate returns success, panes exist, swarm view shows N teammates), but no teammate ever runs a tool.

Symptoms (5/5 teammates affected in my repro):

• 0 tokens used per teammate
• 0 tool calls
• TaskList empty
• watchdog-status.md never written
• Inbox messages queued unread in ~/.claude/teams/<name>/inboxes/*.json
• Pane scrollback shows a Bun stack trace from /$bunfs/root/src/entrypoints/cli.js

Workaround that works: spawn teammates with --model sonnet (or --model haiku for cheap monitors). Orchestrator stays on Opus. Mixed-model teams spawn and run to completion.

Ruled out:

• Solo --model opus sessions work fine — this is /team-specific
• Sonnet/Haiku teammates work fine
• brew upgrade claude-code confirms 2.1.119 is the latest available

Suspected trigger: --model opus + CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 on 2.1.119.

What Should Happen?

Spawned teammates with --model opus should boot the same way Sonnet/Haiku teammates do: process stays alive, registers with the orchestrator, begins consuming inbox messages, emits tool calls. TaskList should report the teammate as running, and tokens/tool-use counters should advance within the normal 30–90s startup window.

Error Messages / Logs

Bun-bundled CLI stack trace visible in tmux pane scrollback (minified frame names):

createInst → PL → LYH → rZH → RD6 → iZH → jM_ → aH → LH → FH
  at /$bunfs/root/src/entrypoints/cli.js

Captured via:

tmux -L claude-swarm-<pid> capture-pane -t %<n> -p -S -2000

Happy to attach the full unminified scrollback if it helps narrow it down.

Steps to Reproduce

1. Open a fresh Claude Code session: claude
2. Switch orchestrator to Opus: /model opus
3. Trigger an agent team with 4–5 teammates, at least one on Opus:

/team "investigate <any task>"
(Or directly: a create an agent team with N teammates: directive where one or more teammate lines specify Opus.)

1. Wait 60–90 seconds.

Expected: teammates start emitting tool calls, TaskList populates, tokens advance.
Actual: TaskList stays empty, every teammate at 0 tokens / 0 tool uses indefinitely.

To confirm crash (not just slow startup):

tmux -L claude-swarm-<pid> list-panes -F '#{pane_id} #{pane_current_command}'

The panes show zsh (process exited) instead of claude (still running).

Recovery: tmux -L claude-swarm-<pid> kill-server, then respawn with Sonnet/Haiku teammates only.

Environment

• Claude Code Version: 2.1.119 (current latest per brew upgrade claude-code)
• Model: Opus 4.7 (1M context) orchestrator + Opus teammates triggers crash
• OS: macOS Darwin 24.6.0
• Shell: zsh
• Setup: agent-teams enabled (CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1), custom /team slash command spawning 3–5 teammates from personas
• Regression: unknown — first time using /team with Opus teammates on 2.1.119

Additional Information

Crash reproduces consistently when any teammate line specifies Opus, regardless of persona/role/prompt content. Happy to provide minimal repro spawn config, full pane scrollback, or run additional diagnostics.