The mechanics behind a session
Choose who participates, what the room is working on, how the conversation moves, and when edits need your approval.
Seats and authority
Each seat pairs a model with a role. The Drafter is the default writer; other roles review and contribute without changing files unless you explicitly grant them write authority under a Project policy that allows it. Your edit permissions still apply to authorized writers.
Tasks and modes
A task describes the work: Plan, Code, Debug, or Project. Start one with /plan, /code, /debug, or /project. A mode controls how the room takes turns:
- Turn-Based — you control the pace of the conversation.
- Brainstorm — the AIs take turns autonomously. Disk writes stay read-only until a quarantine branch is open; then changes stay on that branch until you choose to adopt or abandon them.
- Raise-Hand — an active speaker works while other participants review in the background and raise issues for you to resolve.
Switch with /mode <name>. Tasks, modes, and edit permissions are separate controls. During a Plan task, /freeze validates PLAN.md, emits an EXECUTION.md assignment manifest, and moves the task to Code. It does not start remote execution.
Hand-raises
In Raise-Hand mode, background participants raise issues for your attention. Each raise carries a priority:
- BLOCKER — pauses the room until you admit or dismiss it. Also blocks
/freeze. - CONCERN — queued for you to admit when convenient. The room keeps moving.
- FYI — an informational raise queued for review.
Permission modes and WriteGate
PermissionMode (/permission) is the autonomy knob — how often you click. WriteGate is what actually lands on disk. It is derived from task × interaction mode × permission (plus Cloud / pen-holder limits). Naming only auto-edit is misleading when the gate is still read-only.
- ask — when WriteGate is
ask-before-edit, each disk write waits for your approval. This is the default PermissionMode. - auto-edit — when WriteGate is
full-write, the pen-holder skips per-file prompts; the room still stops between turns. It does not unlock writes in Brainstorm-without-sandbox or a Cloud Project. - auto-chat-turn — turns advance automatically; you still admit hand-raises. WriteGate still governs disk.
- auto-chat-role — the room runs autonomously until a hand-raise resolves. WriteGate still governs disk.
Brainstorm stays read-only until a brainstorm/<timestamp> quarantine branch is confirmed open; only then are sandbox writes full-write, and adoption still needs an explicit ADOPT (nothing reaches your working branch silently). Cloud / Vault sessions never write the local workspace — model file blocks become reviewable drafts that only an explicit Save Cloud Draft (or Vault freeze-draft save) uploads. BLOCKER raises and the session cost cap still pause the room. Use /permission ask to tighten autonomy; /freeze is a Plan command, not a one-click return to ask.
Cost and context
In VS Code’s native Chat, address @orchestra and attach a Language Model tool to let the seated team use that exact tool for the current request. /capabilities shows the tools the live Extension Host actually exposes. A seated BYOK model prepares the schema input, VS Code owns validation and confirmation, and the result enters the room as untrusted evidence. Private extension agents and commands are not automatically callable, and the full AI Orchestra panel does not receive native Chat tool grants.
The cost meter in the header is live and per-session. /cost breaks it down by seat; /cost cap 2.00 sets a ceiling for new calls. The context pill estimates input tokens, window use, and input price before each provider dispatch. Use /context to separate Project instructions, repeated history, current evidence, and reserved output. /context save refreshes a compact CONTEXT.md in a local Project or opens an explicit Vault draft. The visible Session Memory remains, while older raw turns stop repeating in later provider calls.