You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
After every fix ticket blocking this issue has merged into release/0.5.0. Run at the release branch head.
How
Setup: the "Live reruns" and "Reading colors" notes on Map: ctui 0.5.0 live-check fixes. That means Claude Code 2.1.292 with DISABLE_AUTOUPDATER=1, tmux, a scratch CLAUDE_CONFIG_DIR per session, and --plugin-dir ./ctui.
All seven sections. The first pass's results don't carry over.
Run every row, and leave none "not run". Send hovers and clicks, the ✕ click included, as raw mouse escapes through tmux send-keys -l. Use omarchy theme set on this machine and a scratch probe mod. A row that needs something only the human can do, such as dropping the page cache with sudo, goes to the human as a short checklist, and the pass waits for their result.
Run a failing row again in a fresh session before recording it.
Done
The release pull request's Live checks section is rewritten to the current template, with every section ticked and a result per row.
If anything fails, start a new map, "Map: ctui 0.5.0 full-pass fixes", with the same ship bar: fix everything ctui-side before 0.5.0 ships, and send an engine limit to a defer-or-drop grilling ticket. Fix branches are bugfix/<name> cut from release/0.5.0. Don't fix anything inside this issue.
Setup agreed with the human on 2026-10-11, before the pass. The pass waits until every blocking issue has merged (human, 2026-10-11) and for the account's weekly usage reset (it was at 94%).
Session recipe (checked on 2.1.292 at c7766dd): own tmux server per run, tmux -L ctui-<name> -f /dev/null new-session -d -x <cols> -y <rows> -c <repo>, running env CLAUDE_CONFIG_DIR=<scratch> DISABLE_AUTOUPDATER=1 [CLAUDE_CODE_TMUX_TRUECOLOR=1] node_modules/.bin/claude --plugin-dir ./ctui. The scratch dir holds a copy of ~/.claude/.credentials.json, settings.json{"tui":"fullscreen"}, and a .claude.json with hasCompletedOnboarding: true, theme: "dark" and the repo under projects with hasTrustDialogAccepted and hasCompletedProjectOnboarding. With that, the Sidebar docks at start with no dialog.
Visual rows (glyph sizes, the eye, ▶ as text in Alacritty and Ghostty): open the terminal out of sight on a headless Hyprland output and screenshot it. hyprctl output create headless CTUI-TEST (it takes an empty workspace, 7 here), hyprctl eval 'hl.window_rule({ match = { class = "^(ctui\\..*|ctui-.*)$" }, workspace = "7 silent" })', then setsid -f ghostty --class=ctui.probe -e tmux -L <server> attach -t <session> and setsid -f alacritty --class ctui-probe -e …, and grim -o CTUI-TEST shot.png. The human's focus isn't touched. Remove the output after (hyprctl output remove CTUI-TEST). Alacritty is now installed.
The first run of both reports is on #223: live checks 14 PASS / 2 FAIL / 14 NOT RUN (scenarios 1, 2, 5 and 6), and the eval suite's three deltas.
Both FAILs were over-specified expectations, corrected in #296: the frame check asked for the dock's rule in subtle, a cell the engine owns under inherit, now recorded as the engine-pinned expectation frame-rule-colour; the gutter check asked for a last text column of exactly 105, a ceiling no row type reaches, now an assertion that nothing reaches the rule's neighbour with one reply row forced to the ceiling by an unbroken token.
The eval found a real regression, fixed in #295 and closed as #294: ctui's section made the model start a task list for a one-line fix in 3 of 3 runs.
Two re-runs are left before 0.5.0 goes to main, both the human's: scenario 1 alone on the corrected expectations, and the eval suite to confirm the reworded section.
Superseded: 0.5.0 shipped without this pass finishing. chore: release 0.5.0 merged into main as 52da5e6 and is tagged v0.5.0, so the release this pass was meant to gate is past and the Done condition here can no longer be met.
Where it ended, for the record: the last run was scenario 1 only — 7 PASS, 1 FAIL, 22 NOT RUN of 30 (docking 3 of 6, and all of toasts, themes, menu, todo and cost). The single FAIL, rewrites/rule-gutter, failed because the ⎿ column was never measured, not because a row drew wrong. The first full run's two FAILs were over-specified expectations and were corrected before shipping, in the engine-pinned frame rule and the gutter ceiling, and the eval suite's real regression was fixed in #295 and closed as #294.
ready-for-agent removed so no session picks this up and spends the usage window on a shipped release. What it never ran, and the setup recipe in the comments above, carry into #313.
Activity
a1exk-dev commentedon Oct 11, 2026
Notes for this pass, from #248, #249 and #227:
:). Confirm that looks right.Monthly cost 100, and the docked /ctui row now expects the engine's›and field parts in the /theme's colors.CLAUDE_CODE_TMUX_TRUECOLOR, the Sidebar background should be a dark grey on the grey ramp, not48;5;59(#5f5f5f). For example, the Solitude glass should go out as48;5;234(#1c1c1c), and the Everforest Theme as48;5;237(#3a3a3a). With the variable set, it keeps the tinted true-color glass.a1exk-dev commentedon Oct 11, 2026
Setup agreed with the human on 2026-10-11, before the pass. The pass waits until every blocking issue has merged (human, 2026-10-11) and for the account's weekly usage reset (it was at 94%).
tmux -L ctui-<name> -f /dev/null new-session -d -x <cols> -y <rows> -c <repo>, runningenv CLAUDE_CONFIG_DIR=<scratch> DISABLE_AUTOUPDATER=1 [CLAUDE_CODE_TMUX_TRUECOLOR=1] node_modules/.bin/claude --plugin-dir ./ctui. The scratch dir holds a copy of~/.claude/.credentials.json,settings.json{"tui":"fullscreen"}, and a.claude.jsonwithhasCompletedOnboarding: true,theme: "dark"and the repo underprojectswithhasTrustDialogAcceptedandhasCompletedProjectOnboarding. With that, the Sidebar docks at start with no dialog.▶as text in Alacritty and Ghostty): open the terminal out of sight on a headless Hyprland output and screenshot it.hyprctl output create headless CTUI-TEST(it takes an empty workspace, 7 here),hyprctl eval 'hl.window_rule({ match = { class = "^(ctui\\..*|ctui-.*)$" }, workspace = "7 silent" })', thensetsid -f ghostty --class=ctui.probe -e tmux -L <server> attach -t <session>andsetsid -f alacritty --class ctui-probe -e …, andgrim -o CTUI-TEST shot.png. The human's focus isn't touched. Remove the output after (hyprctl output remove CTUI-TEST). Alacritty is now installed.costrow (63): no sudo step for the human. Use the fixture generator from Cold-cache cost live check depends on each contributor's own transcripts #263 (it blocks this issue and ships in 0.5.0) in the scratch config dir, not the human's transcripts, and follow the reworded row.omarchy theme set: the human's theme is Nord; set it back after the Omarchy rows.a1exk-dev commentedon Oct 13, 2026
The first run of both reports is on #223: live checks 14 PASS / 2 FAIL / 14 NOT RUN (scenarios 1, 2, 5 and 6), and the eval suite's three deltas.
Both FAILs were over-specified expectations, corrected in #296: the frame check asked for the dock's rule in
subtle, a cell the engine owns underinherit, now recorded as the engine-pinned expectationframe-rule-colour; the gutter check asked for a last text column of exactly 105, a ceiling no row type reaches, now an assertion that nothing reaches the rule's neighbour with one reply row forced to the ceiling by an unbroken token.The eval found a real regression, fixed in #295 and closed as #294: ctui's section made the model start a task list for a one-line fix in 3 of 3 runs.
Two re-runs are left before 0.5.0 goes to main, both the human's: scenario 1 alone on the corrected expectations, and the eval suite to confirm the reworded section.
a1exk-dev commentedon Oct 14, 2026
Superseded: 0.5.0 shipped without this pass finishing.
chore: release 0.5.0merged intomainas52da5e6and is taggedv0.5.0, so the release this pass was meant to gate is past and the Done condition here can no longer be met.Where it ended, for the record: the last run was scenario 1 only — 7 PASS, 1 FAIL, 22 NOT RUN of 30 (docking 3 of 6, and all of toasts, themes, menu, todo and cost). The single FAIL,
rewrites/rule-gutter, failed because the⎿column was never measured, not because a row drew wrong. The first full run's two FAILs were over-specified expectations and were corrected before shipping, in the engine-pinned frame rule and the gutter ceiling, and the eval suite's real regression was fixed in #295 and closed as #294.ready-for-agentremoved so no session picks this up and spends the usage window on a shipped release. What it never ran, and the setup recipe in the comments above, carry into #313.