[myagizmaktav]

Named profiles retain each person’s role, exact provider/model/effort and source holds. The office shows the whole roster before work is assigned; passive avatars never start a CLI. Full display names become safe branch identifiers. Projects can also increase their issue-board limits without changing defaults.

An opt-in finite supervisor waits for current work and independent QA, then asks the named coordinator to select a suitable issue. It excludes held, assigned, reserved, previously inspected and completed native stages. Unready work is contract preparation only. Dispatch is journaled before sending; uncertain outcomes block replay, and manifests tie decisions to a unique cycle. It reuses an idle matching coordinator and surfaces its status in the team window. The deployment supplies its own resource/lease guard and a single-instance OS lock.

With AGENT_OFFICE_RESTORE_WORKERS=interrupted, page visits and server restarts retain live work while leaving completed and failed terminals asleep. This prevents passive observers from opening a CLI for every old desk. A launcher explicitly waiting for capacity stays in starting status instead of being reported as a human permission prompt. Worker resume can carry a new prompt, allowing a preserved failed task to resume without a duplicate desk.

Validation: typecheck, 629 tests and production build pass. Coverage includes source holds and exact routes, profile reuse refusing busy/changed/held workers, Git-valid full-name branches, passive roster identity, finite scheduler gates, uncertain dispatch, exact-cycle manifests, status-file symlink/privacy boundaries, interrupted-only restoration and resource-wait status. A live bounded installation verified a real Codex recovery on its original model/effort, supervisor waiting for active work, and the team/status window screenshot. No private profiles, credentials or deployment config are included.