Parent: #717. Supersedes the narrow background-only reading of #1001/#755.
Problem
The final acceptance gate must prove the actual operator target: many agents using foreground-equivalent capabilities at the same time while the human keeps using the computer. Background-only non-interference is insufficient and can mask capability loss.
Required FSV scenario
Use real Synapse MCP tools and primary agent sessions. No subagents as the proof surface.
Minimum acceptance:
- Launch or connect 50 agent sessions, or the maximum supported count plus a filed capacity issue if a real local hard limit is discovered after reversible setup work.
- Give each session an
agent_logical_foreground / foreground lane or owned target.
- Run a mix of tasks requiring click, type, read, screenshot, browser navigation/eval, and shell work.
- Human actively uses the real OS foreground during the run.
- Read physical SoTs after each trigger: session registry, foreground-lane registry, target ownership, action log rows, target DOM/UIA/file/window state, and human OS foreground/cursor samples.
Edges:
- concurrent lane acquisition pressure;
- one crashed/disconnected agent;
- one attempted cross-target action;
- one task requiring real OS foreground lease;
- stale client/tool-profile refresh.
No GitHub Actions/CI gate. No automated FSV harness.
Parent: #717. Supersedes the narrow background-only reading of #1001/#755.
Problem
The final acceptance gate must prove the actual operator target: many agents using foreground-equivalent capabilities at the same time while the human keeps using the computer. Background-only non-interference is insufficient and can mask capability loss.
Required FSV scenario
Use real Synapse MCP tools and primary agent sessions. No subagents as the proof surface.
Minimum acceptance:
agent_logical_foreground/ foreground lane or owned target.Edges:
No GitHub Actions/CI gate. No automated FSV harness.