Repository navigation
fix(base): wake idle agy/Gemini panes - #41
Conversation
The idle-composer probe recognised only Claude Code's U+276F and Codex's
U+203A, so an idle agy pane was never woken by a live push and the message
sat unread until the agent took a turn.
Recognise agy's bare ">" composer as a third idle glyph, and accept
"esc to cancel" alongside "esc to interrupt" as the active-turn marker so a
busy agy pane is skipped rather than woken.
A dirty agy composer ("> git status") still fails the bare-prompt check.
Review, Sol 6.1, 2026-10-02Reviewed dfa2cbe.
51 focused predicate tests passed using PR-head sources in a write-free, in-memory harness. An additional Gemini |
Review, Sonnet, 2026-10-03PR 41, fix(base): wake idle agy/Gemini panes SummaryAdds CriticalNone. MajorM1. Bash PS2 continuation prompt (and any
|
Review (Sol + Sonnet) found two Majors in the original dfa2cbe fix: M1 - adding bare > to the composer glyphs also matched Bash's PS2 continuation prompt and a node REPL, so paneTextIsIdle / paneAcceptsSend read an unfinished shell as a clean composer and the live push could type inbox text into the shell. A bare > now counts only when agy's frame border sits directly above it (isAgyPromptLine); otherwise it is refused. M2 - the glyph list is consumed by findPromptLineIndex with line.includes(glyph). That is harmless for the two near-unique TUI glyphs but not for >, which appears in statuslines, diffs and redirects. A statusline line containing > below a Claude composer would hijack the last-glyph-line scan and silently skip the wake. lineHasPromptGlyph now anchors > to the start of the line and keeps includes for U+276F/U+203A. Adds regression tests that fail on the pre-fix head: PS2 continuation, node REPL, a > in a statusline below a Claude composer, and the positive agy-frame wake. Revert-to-red verified: 6 new tests fail without the fix.
Fixes pushed: both Majors addressed (head now 27b48ff)Thanks for both reviews. I agree with the finding and the shape of the fix: anchor on the frame, not the glyph. M1 (PS2 / node REPL reads as a clean composer) - fixed. A bare M2 ( Minor (dialog case in ACTIVE_PANE_RE comment) - not taken in this PR; it is a comment clarification only and out of scope for a fix commit. Noted. Tests. New regression cases, all of which fail on the pre-fix head (revert-to-red verified, 6 failures):
Evidence on 27b48ff: Not tested: no live agy pane was probed (no agy window available in this worktree); the agy frame shape is taken from the existing fixture and the review's own example. Worth one live agy wake before merge if a window is free. Review legs on this head: Sol or self leg required; Sonnet + Aggy legs are being run by Io. |
Review, Sonnet, 2026-10-03PR 41 (inkie-deck), fix commit Method: read the fix diff and surrounding code from FETCH_HEAD, then ran both versions of Are the Majors fixed?M1 (PS2 / node REPL read as clean): fixed. M2 ( Is the frame anchoring robust?Live agy pane (80x24) shows: FindingsMinor 1. A wrapped statusline can still hijack the scan in a narrow pane
Minor 2. Border character set is narrow
Minor 3. Test coverage is one-directional in places, and the commit message overstates it
Minor 4. PS2 or REPL directly under a box-drawing line is still acceptedHarness: Minor 5. Stale doc comment contradicts the new behaviour
Minor 6. Pre-existing, out of scope: starship/pure shells use U+276FA bare Public-repo hygiene
PositiveNarrow, readable fix. Keeping Merge-order note for PR 43If PR 43 reuses |
Deck PR 41 Fix Review (Aggy - Independent Review Leg)
1. Scope & ObjectiveAudit the follow-up fix for Deck PR 41 (
2. Findings (Ranked)[Observation / Good Practice] Positive Frame Corroboration for agy
[Observation / Good Practice] Selective Start-of-Line Anchoring in
|
Follow-up to 27b48ff from review notes, no behavioural change beyond the doc fix: - isAgentPaneIdle's doc comment still listed an agy prompt as refused; it now describes the real rule (an uncorroborated > is refused). - Adds three cases the review asked for: a dirty agy composer, a > in a Codex statusline below the hint (must not hijack the dim-hint path), and a narrow pane whose wrapped continuation row contains a glyph. Accuracy correction on the 27b48ff commit body: of its six new tests, four failed on the pre-fix head, not six. The agy frame wake and the Claude dirty-with-footer cases pass on the old head too (they are negative assertions), so the earlier message overclaimed. On the old head this commit's Codex-statusline case also fails; the dirty-agy and narrow-pane cases pass there as negative assertions.
Follow-up review, Sonnet, 2026-10-03PR 41 follow-up a29f179 (on top of 27b48ff, "close review notes on the agy frame anchor"). Behaviour changeNone. TestsRan
The commit message corrects an earlier overclaim (four of six, not six). It also says the Codex-statusline case fails on the old head; on d1767a2 it passes. I did not test dfa2cbe, so that claim may refer to the intermediate head; it should name the SHA. FindingsMinor
Nit
Public-repo hygieneDiff and commit message contain no absolute home-directory paths or agent names ("Sonnet" is a model name in a comment, acceptable). Did not check
|
Deck PR 41 Follow-Up Review (Aggy - Independent Review Leg)
1. Scope & ObjectiveAudit PR 41 follow-up commit
2. Findings (Ranked)[Observation / Good Practice] Accurate Documentation Alignment
[Observation / Good Practice] Test Coverage for Edge Cases
3. Test Suite Verification & Public Hygiene
|
What
An idle agy/Gemini pane was never woken by a live push: the idle-composer probe recognised only Claude Code's
U+276Fand Codex'sU+203A, so the composer could not be confirmed idle and the wake was skipped. The message then sat unread until the agent happened to take a turn.Two changes:
IDLE_PROMPT_REandCOMPOSER_PROMPT_GLYPHSaccept agy's bare>composer as a third idle glyph.ACTIVE_PANE_REacceptsesc to cancelalongsideesc to interrupt, so a busy agy pane is skipped rather than woken.A dirty agy composer (
> git status) still fails the bare-prompt check, and a bare>still falls through the shell-frame guard, so neither opens a new wake path into a shell or a half-typed line.Review legs
check:leaksclean (385 files, 30 rules),check:dashesclean,lib-version-gateclean,tmux.test.ts84/84.Not tested
OSS-clean
Public repo. No personal or customer names, no internal agent names, no
/homepaths, no keys, no commercial or pricing content.