[chriscoveries]

Revised per review. Terminal output does not establish whether a process is
working: a model can compute silently for long periods, and a spinner or a
prompt waiting for input can keep printing. This version separates
observation from inference.

What changes

• cccc-runtime records when each terminal session last wrote output, as
  a wall-clock time for display and a monotonic instant for the quiet
  duration. cccc_runtime::last_output(group_id, actor_id) returns it for
  the live session only. A new session starts without an observation.
• Actor fields add two display-only values for a running terminal Actor:
  pty_last_output_at (RFC 3339) and pty_quiet_seconds. They are null
  before the first output, for stopped Actors and for non-terminal runners.

What does not change

• A terminal Actor without authoritative turn information still reports
  waiting / pty_running_state_unknown. idle_seconds and
  effective_working_updated_at keep their current values. The new fields
  never feed the working state.
• Managed sessions, event-driven runtimes and headless runners keep
  precedence and behave as before.
• There is no timing window and no new setting.

Tests (unix, real PTY processes)

• quiet ongoing work (prints once, then computes silently): the state stays unknown;
• noisy input-waiting (a prompt redrawn every 0.2 s): the state stays unknown while the observation advances;
• print-then-sleep: observations only, and the last-output time is stable;
• a new PTY session resets the observations;
• a managed session keeps its state while its terminal prints.

Verification: GitHub Actions CI on the fork at exact head 337a6b2a421e6a57bf8b56cc6c33d8f0b1a2d361 (run 39806606736) passed all six jobs, including rust-linux. The test-only red reference 7413e80c22e598e5a8072c5d4b12bd50fb60fc08 on the #146 head c4c9e7a02048aa36320dc3a807ea5e8e03ce65f6 (run 39806615317) fails exactly the four output-inference cases and the managed-session case, where the observation field is absent.

[chriscoveries]

Live end-to-end run (claw: Debian 13, Skylake 8vCPU AVX2/AVX-512, 15G RAM), 2026-10-21, built cccc binaries in a throwaway CCCC_HOME, real daemon and real PTY. A custom-runtime Actor prints one line a second for 8 seconds, then stays alive and silent; cccc actor list is read every 0.5 s for 35 s.

• upstream main 3d3d10cb: waiting / pty_running_state_unknown the whole time, idle_seconds null.
• this branch (merged locally on main with the rest of this batch, 1664c16f): working / pty_output_recent from the first line (0.35 s after start) while it prints; idle / pty_output_quiet 10 s after the last line (17.3 s), with idle_seconds counting up from there (28 at the end).

[ChesterRa]

The missing PTY observability is real, but I am holding this version because output activity does not establish working or idle. A model can compute silently for longer than ten seconds; a spinner or a process waiting for input can keep printing. Changing the window cannot resolve those counterexamples.

Please separate observation from inference: expose last-output time / quiet duration for display, while retaining an unknown working state when the runtime supplies no authoritative turn information. Preserve managed-session precedence and reset observations on a new PTY session. Add cases for quiet ongoing work and noisy input-waiting, rather than only print-then-sleep.

That narrower observability change would be useful without changing the meaning of the shared working-state fields or introducing a global timing knob.

[chriscoveries]

CI note: on current main, rust-linux fails the five real-Chrome fixtures added in 1e5651b2 on every pull request. #175 moves them to the nightly job.