Skip to content
Chapter 7 · Decisions 7-20

Always allow parallel sessions in one workspace

Different conversations are independent units of work, even when their working directory is identical. Requiring users to find and enable shared coordination made switching between running sessions appear unsupported.

Workspace coordination is always shared. Admission remains atomic per session: idempotency replay precedes the active-session check, and a second active run in the same session is still rejected. Different sessions no longer contend for an exclusive workspace admission lock, including compatibility hosts that do not compose the coordinator store.

The existing membership scheduler bounds execution, with 32 workspace slots by default and a global limit of drive workers (four when accepted; sixteen since 2026-10-22, because a drive worker is held through a whole terminal command and four builds stalled every other session and subagent). Model calls keep their separate provider-step bound. The concurrency API retains its route and fields but accepts only shared mode and capacities from 2 to 32. Requests for exclusive mode or one slot fail with the existing named policy-validation error. Schema v19 upgrades legacy exclusive/single-slot coordinators; existing shared capacity choices above one remain intact.

Workspace trust, permission presets, per-session grants, approvals, provider pins and tool execution authority retain their existing owners. A busy or parked session does not block an unrelated session in the same directory. Cancellation and client navigation continue to target a run/session, not the workspace.

All sessions see the same real filesystem. Existing hash guards and external change detection remain in force; this decision does not introduce worktrees, per-path locks, automatic merge or OS confinement. Those need separate designs.

An opt-in GUI toggle was rejected: parallel sessions are normal product behavior. Unlimited execution was also rejected because the scheduler must retain bounded provider, process and memory usage.

Amendment (2026-11-16): no execution ceiling

At the owner's request the "unlimited execution was rejected" decision above is reversed. The runtime loop no longer has a drive-worker limit (formerly sixteen), the provider-step bound has no default (formerly four; RU_MAX_CONCURRENT_MODEL_STEPS remains as an opt-in cap, for example 1 for a single-slot local runtime), and a workspace has no membership cap by default (formerly 32). The concurrency API accepts shared mode with max_concurrent_runs of 0 (no cap, the default) or any cap of two or more; schema v37 rebuilds workspace_coordinators for that and moves rows at the old 32 ceiling to no cap while keeping smaller user caps. Exclusive mode and single-slot caps stay rejected.

The trade-off the original text named is accepted: provider, process and memory use are no longer bounded by the scheduler. Provider rate limits, the machine's resources, the per-run step and terminal bounds, and an optional cap are what remain. Session admission, trust, permissions and execution authority are unchanged.