| name | real-claude-teammate-tmux |
| description | Launch, verify, and regression-protect real Claude Code teammates running in tmux with the `--teammate-mode tmux` primitive. Use when the user wants a real parallel Sonnet Claude Code team (not pseudo-agent narration), when fixing slash commands that drifted into `Agent(...)` / `TaskCreate` / `SendMessage` pseudo-primitives, when wiring `/team-claude` or `/sidekick` style commands to actually start teammates, when the user says "use real claude team" / "use tmux" / "sonnet only" / "actually start the team" / "use the real claude team", or when a regression contract test is needed so future docs cannot regress into Agent/TaskList pseudo-primitive language. Covers the launch recipe, the verification recipe, the Sonnet-only model policy, the regression-contract test pattern, and the /rg + /advice sign-off workflow. |
| version | 1.0.0 |
| author | Hermes Agent |
| license | MIT |
| platforms | ["macos","linux"] |
| metadata | {"hermes":{"tags":["claude-code","teammates","tmux","sonnet","--teammate-mode","regression-contract","rg","advice","sidekick","team-claude"],"related_skills":["claude-code","agento","advice","finish-the-job","system-wide-review","test-tui-claude-feature-via-cmux"]}} |
| changelog | ["1.0.0 (2026-07-10): Initial authoring. Captures the real-vs-pseudo Claude Code teammate lesson from jleechanorg/claude-commands PR #321 and `tests/test_real_claude_team_contract.py`. Pin the Sonnet-only model policy from user feedback (\"just use sonnet models with claude code\") and the /rg + /advice sign-off workflow pattern."] |
real-claude-teammate-tmux
When a slash command's docs (or a session's plan) need to "start a Claude team",
they MUST spawn real Claude Code processes in tmux, NOT narrate a pseudo team
from an in-memory Agent(...) / TaskCreate / TaskUpdate / SendMessage
surface. The real primitive is:
claude --model sonnet --teammate-mode tmux
This skill encodes the launch recipe, the verification recipe, the
Sonnet-only model policy, the regression-contract test pattern, and the
/rg + /advice sign-off workflow that proved the pattern end-to-end
(jleechanorg/claude-commands PR #321, 2026-07-10).
When to use this skill
- A project slash command (
/team-claude, /sidekick, /swarm, etc.) is
supposed to start parallel Claude Code teammates but currently describes
pseudo-primitive Agent(...) / TaskCreate / TaskUpdate / SendMessage
/ team_name=... calls that don't actually spawn teammates.
- The user explicitly says "use real claude team", "use tmux", "sonnet only",
"actually start the team", "use the real claude team", or complains that
/team-claude / /sidekick "don't always start the real claude team".
- A teammate fan-out is needed and the orchestrator must verify each
teammate is a real process, not a narrative.
- A regression contract test is needed so a future edit cannot regress the
docs back into pseudo-primitive language.
- The user wants the
/rg (RED → CODE → GREEN) loop with /advice Reviewer A
sign-off on a fix.
Hard contracts (non-negotiable)
- Real primitive only. Every teammate is a real Claude Code process
launched via
tmux new-session. The launch command prefix is always:
claude --model sonnet --teammate-mode tmux
- Sonnet-only by default. Do not spawn haiku/fable/opus scout lanes even
if a slash command's docs nominally support them. The user has the final
word on the model. No
--fallback-model without explicit user approval.
- No pseudo-primitives. No
Agent(subagent_type=...),
TaskCreate / TaskList / / / ,
/ strings in the production
docs. They are documentation drift, not real.
