Skip to content

codex-busy-detection.md

Latest commit

 

History

History
80 lines (53 loc) · 4.98 KB

codex-busy-detection.md

File metadata and controls

80 lines (53 loc) · 4.98 KB
Raw

Codex 0.145.0 busy-detection evidence

This doc records the empirical basis for the codex agent-liveness backstop in bin/fm-watch.sh (codex_static_pane_upgrade) and the related facts in .agents/skills/harness-adapters/SKILL.md. It is evidence, not narrative: every claim below carries the date, version, exact command, and observed output.

Why the backstop exists

Codex 0.145.0 renders its esc to interrupt busy row only during the pre-answer phase of a turn. It drops the row while an answer streams and, depending on the tool, during a mid-turn tool call, so a healthy codex worker can sit on a pane that renders no busy text at all. Firstmate's ordinary pane evidence still covers the phases where the pane carries a visible signal:

  • streaming token bursts change pane content, so the pane hash moves and the watcher does not treat those changed samples as stale;
  • the pre-answer phase renders the row, so window_is_busy reads busy.

The gap is a worker on a STATIC pane between visible updates, or in a static tool-call phase, with the row absent for longer than the two polls the non-terminal-stale path needs. That worker reads idle to window_is_busy and would surface as a possible wedge. The backstop adds a signal that does not read interface text at all: the codex agent PROCESS itself, via fm_backend_agent_alive.

Environment

  • Date: 2026-08-02
  • codex --version: codex-cli 0.145.0
  • Model in the running TUI: gpt-5.6-sol (the bundled default)
  • Backend: tmux 3.4, on a scratch tmux socket (tmux -L codexverify) isolated from the fleet's server
  • codex launched with --dangerously-bypass-approvals-and-sandbox in a scratch cwd for the streaming samples

Fact 1 - the codex process stays the pane's foreground command for the whole turn

This is the load-bearing assumption of the backstop: fm_backend_tmux_agent_alive maps *codex* to alive, so if the foreground command drops to a bare shell mid-turn the backstop would misfire.

Method: sample tmux display-message -p '#{pane_current_command}' every 0.4s across a full turn that included a sleep 18 shell tool call followed by a streamed answer.

Result: codex in 100/100 samples across a ~40s turn (and 90/90 in an earlier ~36s turn). A bare idle shell reports bash, which the classifier maps to dead; codex exiting on Ctrl+C flipped the foreground command from codex to bash immediately (see Fact 3).

Fact 2 - the busy row is absent for a long contiguous stretch, and the pane is STATIC within it

Method: sample every 0.4s, recording (a) whether esc to interrupt appears anywhere in the full pane capture and (b) whether the whole-pane content hash changed since the previous sample.

Turn 1 (sleep 18 tool call then a 3-sentence answer), busy-row timeline relative to turn start:

  • +0.9s .. +25s: esc to interrupt present (pre-answer plus this tool call)
  • +26.4s .. end (38 consecutive samples, ~15s): row absent

Turn 2 (an 8-sentence explanation, more streaming), cross-tabulating pane-state against the busy row over 120 samples:

     10 pane=CHANGED busy=no
      6 pane=CHANGED busy=ROW
    102 pane=STATIC   busy=no
      2 pane=STATIC   busy=ROW

Every one of the 102 STATIC busy=no samples had cmd=codex. The transitions show the exposure directly: from +4.7s to the end of the turn the row was absent, and the pane alternated STATIC/CHANGED as text streamed in bursts, e.g. STATIC/no at +5.1s, +5.9s, +6.8s, +7.7s, +8.9s, +9.8s, +10.6s, +12.3s. Two consecutive watcher polls landing in such a stretch is exactly the false-wedge the backstop absorbs. The row string itself is unchanged and still matches whenever it renders, which is why the busy-signature matching in fm-watch.sh/fm-tmux-lib.sh is left as-is.

Fact 3 - exit and resume mechanics (re-confirmed)

Single Ctrl+C from an idle empty composer quit codex outright: the foreground command went from codex to bash after one press, with no confirmation.

On that clean quit codex printed:

To continue this session, run codex resume 019fc488-b6af-7353-a380-2d9b32b88db0

This confirms the To continue this session, run <command> form and that the runnable command for an unnamed session is codex resume <uuid>. These match the facts already recorded in the harness-adapters codex section; they were re-observed here as a side effect of the busy-row runs and are noted for completeness.

Carried-forward facts not re-verified here

The trust-dialog Escape-selects-QUIT fact and the $-popup settle were not re-exercised in this session (a non-git scratch cwd did not raise the trust dialog on 0.145.0). They remain as recorded in the harness-adapters codex section from the 2026-07-26 re-verification.

Maintaining this file

Keep this file as evidence for the codex busy-detection backstop. If codex changes when it renders the busy row, or the pane's foreground command stops being codex mid-turn, re-run the two sampling methods above and update both the numbers here and codex_static_pane_upgrade in bin/fm-watch.sh.