Status dots flicker: session-active detection is nondeterministic #20

New issue

Closed

opened 2026-08-03 07:54:30 +01:00 by pedromcaraujo · 0 comments

pedromcaraujo commented

2026-08-03 07:54:30 +01:00

Owner

Symptom

The sidebar status dot flips between green (busy) and dim (idle) across
refreshes with nothing actually changing in the channel.

Cause

GET /status in web/tmux_ctl.py derives state from two signals, OR'd
together in _classify():

A — tmux capture-pane -p -t <name> grepped for esc to interrupt.
B — newest ~/.claude/projects/<slugged-cwd>/*.jsonl mtime within
BUSY_WINDOW = 45s.

Neither is authoritative, and the OR flips:

A is a snapshot of a repainting TUI. The spinner line is rewritten
continuously. A poll landing during a redraw, or while a long tool output
has scrolled the hint out of the visible region, simply misses it — next
poll it's back. This alone produces green -> dim -> green on refresh.
A only sees the session's active window/pane. A turn running in a
split or background window is invisible to it.
B doesn't cover a long turn. Claude appends at message boundaries, so
a multi-minute thinking/tool stretch writes nothing and age > 45,
leaving A as the sole signal — see (1). Symmetrically, the dot stays green
for 45s after a turn ends.
B is per-directory, not per-session. Two channels rooted at the same
cwd light together; a claude -p from any shell in that cwd lights both.

So A is precise but unreliably sampled; B is reliably sampled but imprecise.

Fix

Take the state from Claude Code itself via hooks rather than inferring it
from the screen. A hook script brackets the turn (UserPromptSubmit -> busy,
Stop -> idle, PreToolUse/PostToolUse as heartbeat, Notification ->
idle since a permission prompt is "waiting on you") and writes a per-channel
state file the sidecar reads.

Caveats that keep this a hybrid rather than a pure replacement:

Stop does not fire on an Esc interrupt, nor when the session is
killed. So the sidecar still needs liveness guards: pid check for the kill
case, and pane-mark hysteresis (idle only after N consecutive misses) for
the Esc case.
Sessions predating the hooks, or a deploy where hooks aren't wired, must
fall back to today's logic.

Hysteresis on the pane signal is worth having regardless — it kills the
redraw-race flicker without changing anything else.

## Symptom The sidebar status dot flips between green (busy) and dim (idle) across refreshes with nothing actually changing in the channel. ## Cause `GET /status` in `web/tmux_ctl.py` derives state from two signals, OR'd together in `_classify()`: - **A** — `tmux capture-pane -p -t <name>` grepped for `esc to interrupt`. - **B** — newest `~/.claude/projects/<slugged-cwd>/*.jsonl` mtime within `BUSY_WINDOW = 45s`. Neither is authoritative, and the OR flips: 1. **A is a snapshot of a repainting TUI.** The spinner line is rewritten continuously. A poll landing during a redraw, or while a long tool output has scrolled the hint out of the visible region, simply misses it — next poll it's back. This alone produces green -> dim -> green on refresh. 2. **A only sees the session's active window/pane.** A turn running in a split or background window is invisible to it. 3. **B doesn't cover a long turn.** Claude appends at message boundaries, so a multi-minute thinking/tool stretch writes nothing and `age > 45`, leaving A as the sole signal — see (1). Symmetrically, the dot stays green for 45s after a turn ends. 4. **B is per-directory, not per-session.** Two channels rooted at the same cwd light together; a `claude -p` from any shell in that cwd lights both. So A is precise but unreliably sampled; B is reliably sampled but imprecise. ## Fix Take the state from Claude Code itself via hooks rather than inferring it from the screen. A hook script brackets the turn (`UserPromptSubmit` -> busy, `Stop` -> idle, `PreToolUse`/`PostToolUse` as heartbeat, `Notification` -> idle since a permission prompt is "waiting on you") and writes a per-channel state file the sidecar reads. Caveats that keep this a hybrid rather than a pure replacement: - `Stop` does **not** fire on an Esc interrupt, nor when the session is killed. So the sidecar still needs liveness guards: pid check for the kill case, and pane-mark hysteresis (idle only after N consecutive misses) for the Esc case. - Sessions predating the hooks, or a deploy where hooks aren't wired, must fall back to today's logic. Hysteresis on the pane signal is worth having regardless — it kills the redraw-race flicker without changing anything else.

pedromcaraujo referenced this issue from a commit

2026-08-03 08:21:41 +01:00

forge: let Claude Code report the status dots instead of guessing

pedromcaraujo referenced this issue from a pull request that will close it,

2026-08-03 08:22:07 +01:00

Take the status dots from Claude Code's hooks instead of guessing #21

pedromcaraujo referenced this issue from a commit

2026-08-03 08:40:29 +01:00

forge: make the hook state authoritative, stop cross-checking the pane

pedromcaraujo referenced this issue

2026-08-03 08:40:44 +01:00

Take the status dots from Claude Code's hooks instead of guessing #21