statusline-generator

name: statusline-generator disable-model-invocation: true description: >- Installs, configures, customizes, or troubleshoots the Claude Code statusline (cwd, model, token counts). Use when the user wants to set up or change the statusline, switch minimal vs full layouts, show absolute token counts instead of a percentage, add cost via ccusage, or fix a statusline that is blank, silent, stuck, or stopped updating. Trigger phrases: 状态栏, 状态栏不显示, 状态栏空白, "statusline blank", "statusline broken".

Statusline Generator

A single-source-of-truth statusline for Claude Code. One script, two layouts, end-to-end self-verification.

Quick health check (start here when something is wrong)

Run this first whenever the statusline misbehaves. It catches the silent failures that account for most "configured but not working" reports:

bash scripts/health_check.sh

It validates four layers:

1. ~/.claude/statusline.sh exists and is executable. Missing chmod +x is the single most common silent-failure cause — Claude Code runs the script, exec fails, statusline goes blank.
2. ~/.claude/settings.json has a valid statusLine block pointing at the script.
3. Mock stdin tests covering complete data, zero tokens, missing fields, $HOME path shortening, and zero-fork git-branch rendering (via a synthetic .git/HEAD — no git binary required).
4. Real stdin replay from /tmp/.claude-statusline-last-stdin.json if you previously ran with CLAUDE_STATUSLINE_DEBUG=1.

Each failure prints a one-line fix command — you don't have to read documentation to recover.

Quick install

bash scripts/install_statusline.sh

This script:

• Backs up any existing ~/.claude/statusline.sh and settings.json.
• Copies generate_statusline.sh to ~/.claude/statusline.sh and chmod +xs it.
• Updates settings.json statusLine block via jq (preserves other settings).
• Mandatorily runs health_check.sh and shows the result — installation is not "complete" until verification passes.

Restart Claude Code (or send any new message) to see the statusline update.

What you get

Default — minimal one-line layout

~/code/myproject  [main]  Opus 4.7 (1M context)  ctx: 108K / 1M

Just the essentials: short path, git branch, model name, absolute token counts. No colors, no cost, no percentage. The branch is read zero-fork — the script reads .git/HEAD as a plain file (with worktree/submodule gitdir: indirection and detached-HEAD short-sha handling) instead of spawning git, so it costs nothing and works even on hosts without a git binary.

Full — multi-line with cost and git

Set CLAUDE_STATUSLINE_LAYOUT=full in your shell profile to enable:

alex (Sonnet 4.6) [$0.42/$25.93]  ctx: 108K/1M (11%)
~/code/myproject
[git:main*+]

• Line 1: user, model, ccusage session/daily costs, color-coded ctx (green ≤50%, yellow 51–80%, red >80%).
• Line 2: short path.
• Line 3: git branch with * for modified, + for untracked.

Layouts: how to switch

The script reads layout from environment, not flags (Claude Code passes JSON on stdin, so flags would conflict). Set in ~/.zshrc or ~/.bashrc:

# Minimal (default — same as not setting it)
export CLAUDE_STATUSLINE_LAYOUT=minimal

# Full
export CLAUDE_STATUSLINE_LAYOUT=full

Restart your shell (or source the rc file) so Claude Code inherits the change, then send a message — statusline refreshes within 300ms.

Debug stdin capture

To see exactly what JSON Claude Code sends your script:

export CLAUDE_STATUSLINE_DEBUG=1

Each invocation writes its stdin to /tmp/.claude-statusline-last-stdin.json (overwriting on every refresh). Inspect with jq .. Useful for:

• Diagnosing why a field doesn't render the way you expect.
• Re-running the script against real input: cat /tmp/.claude-statusline-last-stdin.json | ~/.claude/statusline.sh.
• Filing bug reports — paste the dump as ground truth.

Authoring rules (why this skill is shaped this way)

Three production failure modes drove the current design. All are sealed in code, not just docs:

Rule 1 — Always chmod +x, always verify by running

The single biggest silent-failure cause of any statusline is a script without the executable bit: Claude Code's exec fails silently and the bar goes blank with no error. install_statusline.sh always chmod +xs; health_check.sh flags the bit if missing. If you hand-write or hand-edit a statusline script, mock-test it before declaring done: echo '{}' | bash your-script.sh.

Rule 2 — "Configuration complete" is meaningless without evidence

"Wrote the file and updated settings.json" is not the same as "the script runs and produces the expected output." install_statusline.sh therefore always runs health_check.sh at the end and exits non-zero if any check fails. Treat any "complete!" report from any agent that lacks evidence as suspect.

Rule 3 — The statusline is a hot path: budget subprocesses, not just correctness

A statusline script looks like UI polish, but it executes on every refresh in every concurrent agent session. Whatever it spawns gets multiplied by refresh rate × number of live sessions, all day. Measured on a machine running many concurrent sessions: a package-runner statusline (bunx <pkg>@latest-style, which re-resolves the registry and re-writes a lockfile per refresh, then runs git status + git branch) cost ~0.4s CPU per refresh; this script costs ~0.01s. Across many sessions that difference is a measurable share of machine-wide process churn, heat, and battery — the statusline was one of the top contributors found in a real battery-drain investigation (2026-07).

Concretely:

• Never resolve packages at refresh time. No bunx/npx @latest in statusLine.command — pin and install once, or use a local script.
• Don't spawn git for the branch. Read .git/HEAD as a file (see git_branch_fast in the script) — same answer, zero subprocesses.
• Treat git status (dirty state) as a luxury. It walks the worktree on every refresh; only the full layout runs it, and only when explicitly enabled.
• Budget: a statusline refresh should cost single-digit milliseconds and a handful of forks at most (one jq + one awk here).

For field-level traps (used_percentage null at session start, total_input_tokens semantics across Claude Code versions, hardcoded context_window_size), see references/context-window-schema.md.

Customization

For colors, custom segments (hostname, time, etc.), and disabling cost tracking, see references/customization.md.

Dependencies

The script auto-detects available tools and degrades gracefully:

Install on macOS: brew install jq. On Debian/Ubuntu: apt install jq.

Troubleshooting

For symptom-by-symptom diagnostics, see references/troubleshooting-decision-tree.md. It walks through:

1. Statusline blank or never updates (chmod cause)
2. ctx segment missing or wrong (field traps)
3. Want token counts not percentages (layout switch)
4. Colors render as raw escape codes (terminal compatibility)
5. Git segment missing (full layout)
6. Cost segment missing (ccusage / cache)
7. Edits have no effect (path mismatch)
8. Slow refresh (jq vs python3)

Resources