ADR 0157: The desk owns its launched child sessions

Status: accepted; extends ADR 0119 and ADR 0151

The desk blocks while a Jump in shell, coding-agent command, or Project Script owns its terminal. When that child exits, the desk surveys the fleet again. Nothing identified the child as part of that suspended interaction, however. A user could invoke discern desk from the child, follow the existing linked-worktree redirect to the main checkout, and open a second desk while the first still waited underneath it.

Shell nesting is not the right signal. SHLVL varies by shell and starting environment, coding agents and Project Scripts are processes rather than necessarily shells, and walking the parent-process tree would be platform-specific. Worktree location is also insufficient because a child can change directory to the main checkout.

Every arbitrary-code child launched by the desk inherits DISCERN_DESK_SESSION=1, and the interactive desk refuses to start anywhere beneath that marker.

• The two launch boundaries cover the whole current class: inherited-terminal commands serve Jump in and every provider-owned coding-agent action, while the Project Script runner serves Run script. Descendants inherit the marker normally.
• runDesk checks the marker before JSON-mode, terminal, project-root, fleet-survey, or checkout-location handling. The named JSON form returns error = "desk_already_active"; human output tells the operator to exit the shell, agent, or script and return to the owning desk. Bare discern reaches the same guard before its interaction policy.
• The marker records process context, not authority. A caller can add or remove it, so no lifecycle operation, project state, or health verdict depends on it. Every non-desk verb keeps its existing behavior.
• discern doctor includes desk_session = true in its structured environment block and adds a human environment line when the marker is active. It remains an informational debugging fact, never a check or warning.

This does not reverse ADR 0119's rejection of agent or vendor detection. The marker identifies an action discern itself performed; it does not guess the caller from third-party environment conventions.

• A desk can have one active interaction stack. Changing directories inside its child cannot accidentally create a second supervisory loop.
• Shells, all configured coding agents, and Project Scripts share one ownership rule without provider-specific branches.
• Commands and tests run beneath a desk child also inherit the marker. Because only desk entry and doctor context consume it, ordinary project automation and discern lifecycle behavior remain deterministic.
• A deliberately detached descendant may retain the marker after the owning desk resumes or exits. It can unset the variable when detachment is intentional; the marker is an accident guard, not confinement.

• Read SHLVL. Rejected because it neither identifies the owning desk nor covers non-shell children.
• Inspect the parent-process tree. Rejected as platform-specific, brittle across launchers, and still unable to express discern's ownership as directly as a marker set at launch.
• Rely on the linked-worktree refusal. Rejected because its suggested cd to the main checkout removes the only condition it detects.
• Mark only Jump in shells. Rejected because coding agents and Project Scripts can invoke the same interactive command while the desk is blocked.
• Make every verb desk-session-aware. Rejected because an inherited, user-controlled marker is debugging context rather than lifecycle authority. Broader behavior would also leak ambient terminal context into project automation.