desktop-agent-activity

Desktop coding-agent activity

User asks whether a desktop-hosted coding agent (Orca + Claude Code, and similar PTY hosts) is still working, stuck, or idle.

Do not treat ps showing claude as "working". Alive sessions sit for hours with MCP children. Prove work with CPU + file/git mtimes + host telemetry.

Visual capture is optional confirmation, not the primary evidence.

Verdict (what to tell the user)

Start with the conclusion:working / idle but hanging / stopped。

Table lists each instance: TTY or PID · status · evidence (CPU is changing / source file just modified / how long the terminal has stopped writing). ≤5 bullet lines. Do not paste command output.

Check order (do all of 1–4 before claiming idle)

1. Process + cwd

   • ps for claude (and host app, e.g. Orca).
   • lsof -a -p <pid> -d cwd only — do not dump full lsof -p (large Claude processes time out).
   • Parent is often orca-tcc-login → login zsh on a ttysXXX.

2. CPU sample, 2–3 times over ~5s

   • Working: %CPU jumps (often 1–10%+).
   • Idle-alive: stays ~0, children are only MCP (chrome-devtools-mcp, server-github, kimi-cu mcp).
   • Work-ish children: caffeinate, compilers, git, test runners — not MCP.

3. TTY last-write + source mtimes

   • ls -lT /dev/ttysXXX — terminal still printing?
   • Walk the repo excluding node_modules / .next / .git for .ts/.tsx/.md mtime in the last hour.
   • git status -sb + recent commits. Uncommitted files with fresh mtimes = still editing.

4. Host telemetry (Orca)

   • ~/Library/Application Support/orca/agent-hooks/last-status.json — last prompt, state, worktree path.
   • orca-stats.json agent_start / agent_stop.
   • Paths and parse pitfalls: references/orca-claude-liveness.md.

Optional: screenshot

computer_use capture of the host window can confirm the TUI. If capture returns empty / 0×0, stop looping captures and ship the process+file verdict. Do not raise the window unless the user asked to look at the screen itself.

Pitfalls

• Alive ≠ working. MCP children spawn at session start and stay.
• ~/.claude/projects/**/*.jsonl mtime can lag days while the process is live. Never use stale jsonl as "not working".
• orca-stats unmatched agent_start can be months-old ghosts. Prefer last-status.json + CPU + files.
• last-status state=done can lag behind an already-resumed edit loop. state=working can stick on the last tool. Corroborate with file mtimes. Parse pitfalls: references/orca-claude-liveness.md.
• terminal-history/ entries are directories, not files.
• Iterating Application Support/orca can hit disappearing Electron cookies — skip FileNotFoundError, don't abort the check.

Out of scope

Starting, stopping, or steering the agent. This skill is read-only status.