Response protocol: digit positions + Enter confirms (digit alone is not
enough); arrows+Enter work everywhere. Registry detects waiting state
for both permission and AskUserQuestion dialogs (same waitingFor label
— fine-grained dialog typing needs screen reading). Sandboxed bash runs
no-prompt for most commands: real waits come from edits/network/
questions. Pre-trust via ~/.claude.json projects[dir]
.hasTrustDialogAccepted captured and plausible. Esc-deny and plan
approval deferred to P4 reliability campaign.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Demonstrated on CLI 2.1.173: resuming a live session interleaves both
TUIs into one transcript (no lock, no warning) — liveness detection via
registry pid+procStart is mandatory; --fork-session is safe on live
sessions; --resume must run in the session's original cwd ("No
conversation found" otherwise); registry files are cleaned on graceful
exit AND SIGTERM, may be GC'd later after SIGKILL — never reason on
file presence. ANSI-stripped TUI text loses spaces (cursor-positioned
painting) — confirms @xterm/headless for screen parsing.
Flood: 21 MB through node-pty with 10s pause => 0 bytes leaked, no
loss, 4.7 ms echo after flood.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
99.8% of 1318 real transcripts (449 MB) yield sessionId+cwd in 1.37s
(head+tail reads only). Munge cwd→dir validated on 100% of files.
Registry pid/procStart liveness check validated 3/3. 12 line types
inventoried, ai-title is the best title source.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Arboretum — self-hosted web dashboard for git worktrees and the Claude
Code sessions running on them. README states the positioning vs
GitKraken Agent Mode, Happy/CloudCLI and Anthropic Remote Control, the
security model (localhost-first + Tailscale), and the Claude usage note.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
