Docs

Two steps inside any Claude Code session:

1. Step 1. Adds this repo as a plugin marketplace in your Claude Code config.
2. Step 2. Installs the team-room plugin from the marketplace you just added. Format is plugin@marketplace, which is why the name appears twice.

Requires claude CLI and codex CLI on PATH (the plugin shells out to both). Repo: constantinexanthos/team-room.

The plugin bundles three things:

• An MCP stdio server that exposes four tools (team_room_ask, team_room_status, team_room_recent, team_room_cancel).
• A skillthat loads into Claude's context when the user asks a strategic question, telling Claude when to reach for the room.
• Slash commands (/team-room, /team-room-status, /team-room-rooms) for direct invocation.

Ask Claude in natural language. The skill picks it up and tells Claude to invoke the tool.

# you:
"Ask team-room: should we ship the auth refactor as one PR or split it?"

# claude:
team_room_ask({
  question: "Should we ship the auth refactor as one PR or split it?",
  mode: "dialogue",
  wait: true,
})

Claude and ChatGPT deliberate over up to 8 short turns, then return a structured final_brief.

Every session ends in exactly one terminal state. The MCP tool surfaces outcome and final_brief as the primary artifact, with the full transcript in messages for expansion.

Dialogue mode is a function-labeled working session. Each turn opens with a tag the orchestrator captures as metadata and uses for terminal-state detection.

[frame]

Turn 1: decision, criteria, uncertainty, and lens.

[frame-clear]

Turn 1 fast-path. The frame is obvious, skip straight to evidence.

[reshape]

Turn 2: improve the frame before adding substance.

[evidence / build / refine]

Mid-dialogue moves.

[push-back]

Substantive disagreement, mapped not graded.

[converge]

Terminal. Joint read for the user.

[fork]

Terminal. Explicit unresolved disagreement.

• Substantive uptake. Every non-first turn opens by naming what it's taking from the prior turn. Generic agreement is called out in prompt as collaboration theater.
• Map the fork. Disagreement uses the “condition under which the other's view is right” framing, not scoring rubrics.
• Asymmetry as lens. Claude and ChatGPT surface their training differences explicitly (“my UX lens flags...”, “my code-base prior says...”).

team_room_ask

Open a session. Required: question. Optional: mode (dialogue or rounds), wait, timeout_s, topic, project_id. Returns outcome, final_brief, and messages.

team_room_status

Get state for a topic. While in-flight: live status. After completion: idle plus final_brief.

team_room_recent

List recent topics by last-modified time.

team_room_cancel

SIGTERM an in-flight orchestrator on a topic.

• dialogue (default). Collaborative micro-turn working session. Use 99% of the time.
• rounds. Opt-in adversarial review. R1 = parallel independent answers. R2 = each critiques the other's R1. Only when you explicitly want stress-testing.

You don't see the dialogue turns in your terminal. Claude calls team_room_ask, the plugin runs the deliberation in the background, and the response Claude quotes to you is the joint_read field. The full transcript is available in the tool-call output if you expand it.

Sessions take 30 to 90 seconds typically (3 to 5 turns). While it's running, Claude shows the tool call as pending.

State lives at $TEAM_ROOM_DIR (default ~/.team-room/). Per topic, three files: <topic>.jsonl (transcript), <topic>.state.json (live state), <topic>.brief.json (structured envelope).

Tunables (all optional):

• TEAM_ROOM_DIR=~/.team-room
• TEAM_ROOM_MAX_TURNS=8
• TEAM_ROOM_TURN_WORDS=150
• TEAM_ROOM_AGENT_TIMEOUT=480
• CODEX_REASONING_EFFORT=high