Run notes, 15 November 2026, 15:34 to 15:44 UTC Machine: Windows 11 Pro 25H2, build 26200, 64-bit. Anode 0.11.4 (the current release), set up and signed in before the runs. The seat was session 2 and had been up for about 16 hours; its viewer was open, view only, on a second monitor. Two other programs had windows in the seat, a VPN client and another tool's overlay, both started with the seat and not used. .NET SDK 10.0.401. Claude Code 2.1.325. How each run was started - A fresh copy of data/TipCalc in an empty folder. - parent_watch.py started on the PC's own desktop (session 1), 2 seconds before the job, and stopped after it. - Claude Code, headless, from that folder, through our own job wrapper. The wrapper adds a spending cap and a fresh, empty Claude Code config folder (no user settings, plugins, skills, memory or other MCP servers; Claude Code's own bundled skills and plugins still load), then runs: claude -p --settings --setting-sources user --strict-mcp-config --model sonnet|haiku --max-budget-usd 3 --no-session-persistence --output-format stream-json --verbose --mcp-config mcp-config.json --allowedTools "" --max-turns 60 < prompt.txt - Allowed tools: Read, Edit, Glob, Grep, Bash(dotnet build:*), and Anode's anode_guide, seat_status, seat_lease, seat_capabilities, seat_run, seat_windows, seat_observe, seat_element, seat_wait, seat_screenshot, seat_click, seat_type, seat_key, seat_processes and seat_kill_process. Runs 5 and 6 also allowed seat_window. Not allowed in any run: seat_exec, seat_show, seat_hide, seat_start, seat_stop and the Steam, Android, gamepad, audio, browser and display tools. - After each job we read the app's result label ourselves in the seat (seat_observe or seat_wait) and closed the app. runs/runN.agent.json is made from the job's log by summarize_run.py; the raw log stays with us because it holds this PC's paths. Anode's pointer guard counter ("your pointer" in `anode status`, pointerGuard.suppressed in seat_status), read before and after each run: run 1: 7 to 8; run 2: 8 to 8; run 3: 8 to 9; runs 4, 5 and 6: 9 to 9. Its "forwarded" count was 0 before run 1 and 0 after the last run and the control. Things that went wrong, and why - Run 2 waited for the seat: our own session had just used the seat to take the screenshot and close run 1's app, and still held its lease. Anode told the agent "Desktop in use ... You are number 1 in line"; the agent's second request waited and got the seat. Run 2's 140.8 s includes that wait (the model's own time was 41.6 s). - Run 4 (Haiku) did not finish. Our allow-list left out Anode's seat_window tool; the model asked to close the app with it, Claude Code's permission check refused, and the model chose not to close the window another way. The rebuild then failed because the running app locked its exe. The three Sonnet runs with the same list closed the app through the Close button in its window's title bar instead. We added seat_window for runs 5 and 6. - Claude Code refused one compound shell command in runs 1, 2, 3, 5 and 6: in runs 3, 5 and 6, which started with a relative `cd TipCalc`, parts of the command needed approval; in runs 1 and 2 a `find` on a relative path after a cd couldn't be resolved while the wrapper's Read() deny rules were set. In run 2 it also refused seat_exec, and in runs 4 and 5 seat_hide. - In run 4 the model mistyped one observation id; Anode refused it and the model sent the right one. The summary drops observation ids, so here they are from the raw log: seat_observe returned s_68f0bd5b92a54e33a9ff987b183fdecf, call 12 sent s_68f00bd5b92a54e33a9ff987b183fdecf (one extra 0), call 13 sent the right one. The figure: grab_window.ps1 saved the window's rectangle (316 x 239, from seat_windows) inside the seat after run 1; we then cropped it to 300 x 230 to cut the shadow margin, which showed a strip of the windows behind it. On 16 November we cut the window's four corners to transparent, slightly beyond its own rounded corners, so nothing behind it shows. Who was at the PC: we did not arrange it. The watcher's input clock shows the PC's own desktop was in use during runs 1 (137 samples with new input), 2 (7) and 4 (38), and untouched during runs 3, 5 and 6 (0).