[eric-engberg]

What

Three fixes to Custom Command output and input:

• Styling after the command's own resets. With preserve colors on, a Custom Command now keeps its background, bold and dim across the whole widget in both render modes, even when the command's output resets its own colors partway through. Before, in plain mode, the text after the command's first \e[0m lost the widget's background, bold and dim. With a widget background or a global background override, that left a stretch of terminal background between the widget's colored padding cells. Powerline mode kept the background but lost bold and dim.
• Output with no visible text. A command with preserve colors on that prints only color codes, such as printf '\033[32m%s\033[0m' "$value" when the value is empty, now drops out of the line the same way it does with preserve colors off. Before, it still took a slot: a doubled separator in plain mode (A | | B instead of A | B), or an empty segment with an arrow on each side in Powerline mode.
• The JSON piped to the command. The command now gets Claude Code's JSON as it arrived, plus terminal_width, as docs/USAGE.md says. Before, it got the copy ccstatusline had checked against its schema. That copy drops nested fields the schema doesn't list (such as workspace.added_dirs, workspace.git_worktree, workspace.repo, and any new key under model, cost, context_window or rate_limits) and turns number strings into numbers. So a status line tool chained in as a Custom Command saw less than it would on its own.

Why

Styling after resets. The plain renderer wraps preserve-colors output once: widget bold, dim and background in front, and their off codes after. Any \e[0m, \e[22m or \e[49m in the command's output turns that styling off for the rest of the widget. Powerline mode already put the segment background back after each such reset, but only the background. A widget with background hex:112233, bold on and the command printf '\033[32m+5\033[0m \033[31m-3\033[0m', piped through main under Bun in plain mode, renders the widget as:

\e[1m\e[48;2;17;34;51m\e[32m+5\e[0m \e[31m-3\e[0m\e[49m\e[22m

Nothing restores the background or bold after +5\e[0m, so -3 is drawn on the terminal's background in normal weight. The same happens with a global background override and no widget background. In Powerline mode the same widget gives \e[1m\e[48;2;17;34;51m \e[32m+5\e[0m\e[48;2;17;34;51m \e[31m-3\e[0m\e[48;2;17;34;51m \e[0m: the background comes back, bold doesn't.

Output with no visible text. The widget strips escape codes only when preserve colors is off, then returns output || null. With it on, \e[32m\e[0m is a non-empty string, so it was returned as the widget's text. The renderer decides whether a widget is there by whether its text is empty, not by its width, so the zero-width widget kept its separators and padding. The line Custom Text "A", separator, that command with preserve colors on, separator, Custom Text "B", piped through main, prints [ A | | B ] (brackets added). With preserve colors off it prints [ A | B ]. In Powerline mode, with backgrounds on the three widgets, main draws an extra segment in the command's background color, with its two arrows, between A and B.

The JSON piped to the command. main parses stdin, validates it with StatusJSONSchema, and keeps only the validated copy. The widget then serializes context.data. The schema keeps unknown keys only at the top level (z.looseObject). Its nested objects are plain z.object, which drops keys it doesn't list, and its number fields convert number strings. With a Custom Command of cat > captured.json; echo ok and this payload:

{"session_id":"s","model":{"id":"claude-opus-4-6","display_name":"Opus","family":"opus"},"workspace":{"current_dir":"/p","project_dir":"/p","added_dirs":["/e"],"git_worktree":"wt","repo":{"name":"app"}},"cost":{"total_cost_usd":"1.5"},"context_window":{"context_window_size":200000,"new_key":1},"rate_limits":{"five_hour":{"used_percentage":42,"resets_at":1774020000,"extra":true}},"top_level_unknown":{"a":1}}

main writes:

{"session_id":"s","model":{"id":"claude-opus-4-6","display_name":"Opus"},"workspace":{"current_dir":"/p","project_dir":"/p"},"cost":{"total_cost_usd":1.5},"context_window":{"context_window_size":200000},"rate_limits":{"five_hour":{"used_percentage":42,"resets_at":1774020000}},"top_level_unknown":{"a":1},"terminal_width":245}

model.family, workspace.added_dirs, workspace.git_worktree, workspace.repo, context_window.new_key and rate_limits.five_hour.extra are gone, and "1.5" became 1.5. This branch writes the payload back unchanged, with terminal_width added.

How

• restoreBackgroundAfterResets in ansi.ts becomes restoreStyleAfterResets, which takes the background code plus the widget's bold and dim. For each color sequence in the text, it reads the parameters in order and tracks the background, bold and dim separately. 0 (or an empty parameter) clears all three, 49 the background, and 22 bold and dim. A later 1, 2 or background color in the same sequence counts as the command's own setting, so it stays. After the sequence, it re-applies whatever was cleared and not set again. Extended color arguments are still skipped, so the 22 in 38;5;22 is not read as a reset.
• Powerline mode passes the segment's bold and dim along with its background. Plain mode now calls the same function before wrapping the widget, with the widget's background (or the global background override, through a new resolveBackgroundColor helper that applyColorsWithOverride now uses too), globalBold or the widget's bold, and dim when it is on for the whole widget (not the parentheses-only setting). Claude Status's history strip also preserves its colors, but it only emits foreground codes, so this changes nothing for it. The sentence in docs/USAGE.md that said only Powerline backgrounds are restored now says both modes restore background, bold and dim.
• CustomCommandWidget.render returns null when the output has no visible text, whether or not colors are preserved. It checks with getVisibleText, which also drops OSC 8 links with no text.
• The piped entry point keeps the parsed JSON next to the validated copy and passes it as rawData in the render context. Custom Command serializes rawData (falling back to data) plus terminal_width. All other widgets still read the validated data.

Left as is:

• A command that turns bold off for part of its output, inside a widget the user set to bold, gets bold back after its \e[22m. The widget's bold, dim and background cover the whole widget. Powerline mode already did this for the background.
• Output whose only visible text is spaces, such as a colored block, still renders. Those spaces take cells, and preserve colors off behaves the same.
• The JSON is re-serialized from the parsed object, as before, so whitespace and number formatting can differ from the bytes Claude Code sent. The values are the same.
• The TUI preview shows [cmd: …] and never runs the command, so it isn't affected.

Demo

Each case quits the TUI, then pipes a payload through the CLI. The dim text above the status line says what the command does.

1. A preserve-colors command with a gray background and bold whose output resets its colors twice: printf '\033[32m12 added\033[0m, \033[31m3 removed\033[0m, 2 files'. Before, plain mode drops the background and bold after 12 added, and Powerline drops the bold. After, the whole widget is gray and bold in both.
2. printf '\033[32m%s\033[0m' "${VIRTUAL_ENV##*/}" with no virtualenv active, with preserve colors off on line 1 and on on line 2. Before, line 2 has a doubled separator in plain mode, and in Powerline an empty segment that also moves Ctx on to the next theme color. After, both lines match.
3. A command that reads .workspace.git_worktree and .workspace.added_dirs from its stdin with jq, fed a payload that sets both. Before, it prints worktree: none, added dirs: 0. After, it prints worktree: feature-login, added dirs: 1.

Powerline

Styling after the command's resets, before:

Styling after the command's resets, after:

Output with no visible text, before:

Output with no visible text, after:

JSON piped to the command, before:

JSON piped to the command, after:

Plain

Styling after the command's resets, before:

Styling after the command's resets, after:

Output with no visible text, before:

Output with no visible text, after:

JSON piped to the command, before:

JSON piped to the command, after:

Testing

• On main as of [low] perf(claude-status): hold a lock while the status page fetch is in flight #635 and [medium] perf(git-review): key the PR cache through the cached runGit #639 (62a371b): merges cleanly, bun run lint is clean, and bun test gives 3730 pass, 0 fail. The built CLI under Bun 1.4.4 and Node 26.13.0 matches main in plain and Powerline mode with the baseline configs, and the TUI opens and exits cleanly.
• bun test: 3195 pass, 0 fail across 178 files (main: 3168 pass, 0 fail). There are 27 new tests, and all 27 fail against main's source:
  • ansi.test.ts, 20: bold and dim restored after 0, an empty parameter and 22, each attribute tracked on its own, and sequences that leave bold on, so it must not be re-applied after them (22;1, 2, 49, 38;5;22 and others). Each of those comes after a \e[0m that must bring bold back, so it fails on main too. main has no restoreStyleAfterResets, so these ran against main through a wrapper under the new name that passes the style's background to main's restoreBackgroundAfterResets. The existing background tests move to the new name and pass on both.
  • renderer-ansi.test.ts, 2: plain mode keeps a widget background with bold, and a global background override, after the command's resets.
  • renderer-powerline-separators.test.ts, 2: Powerline re-applies bold and dim in a segment with a background, and bold in one without.
  • CustomCommand.test.ts, 3: color codes around an empty value, and an OSC 8 link with no text, render nothing with preserve colors on. The command gets the JSON as it arrived, not the validated copy.
• bun run lint: clean.
• Node 26.10.0, the 4 changed test files through vitest: 119 pass, 2 fail (main: 92 pass, 2 fail). The 2 failures are the same on both: existing CustomCommand.test.ts cache tests whose temp-home helper calls vi.spyOn(os, 'homedir'), which Node's ESM doesn't allow. All 27 new tests pass under Node. Against main's source, with the same wrapper, Node gives 92 pass and 29 fail: the 27 new tests plus the same 2.
• Built dist under Bun 1.4.2 and Node 26.10.0:
  • With the baseline plain and Powerline configs, the runtimes agree and the output is the same as main.
  • With each demo's configs (and the demo payload for the stdin case), the runtimes agree with each other and differ from main as intended.
  • The TUI smoke test passes on both runtimes.
• Manual piped renders under a scratch HOME, main against this branch: the outputs quoted under Why, plus the global background override case in plain mode, which now re-applies the override after each reset.