commandcode-usage (dsh)

Command Code Usage

See how much of your Command Code plan is left — 5-hour, weekly and monthly windows with reset times — inside the coding agent you already have open.

Runs on your machine. No model round-trip, so checking your quota costs no quota.

CC GOAT │ 5h █▎░░░░░░░░ 12% 4h27m后重置 │ 周 █▏░░░░░░░░ 11% 09-27重置 │ 月 ▋░░░░░░░░░ 6% $66.03 10-20重置

简体中文 · What was verified

---

Install

Each agent has its own way of loading plugins, so pick yours. Six of the seven end up running the same core script (core/cc-usage.mjs); the DeepSeek Harness adapter keeps its own data layer, for the reason given under Layout.

Agent	How	Persistent?
Claude Code	/plugin marketplace add Jovan1666/commandcode-usage then /plugin install commandcode-usage@commandcode-usage, then node ~/.claude/plugins/marketplaces/commandcode-usage/plugins/claude-code/scripts/setup.mjs	✅ status line
Codex CLI	plugin marketplace (.agents/plugins/marketplace.json) — the bundled hook runs each turn; Codex will ask you to trust it	⚠️ per-turn line
Grok Build	grok plugin marketplace add Jovan1666/commandcode-usage, then run plugins/grok/scripts/setup.mjs	✅ status line
opencode	run plugins/opencode/scripts/setup.mjs	✅ sidebar
pi	pi install ./commandcode-usage/plugins/pi, or copy index.ts into ~/.pi/agent/extensions/	✅ above the input
DeepSeek Harness	clone this repo, then dsh plugin --profile web add ./commandcode-usage/plugins/dsh (needs dsh ^0.1.5-rc.1)	✅ sidebar
ZCode	/plugin marketplace add Jovan1666/commandcode-usage then /plugin install command-code-usage (its marketplace and plugin name differ)	⚠️ command only

Nothing here asks for your API key up front. The script finds it — see Credentials.

The DeepSeek Harness card has its own three-state screenshot in plugins/dsh/assets/screenshot.png — collapsed badge, light expanded, and dark expanded. The other five platforms all render the same one line shown above, so they get text rather than a picture.

Why some installs need a setup script

Claude Code and Grok keep their status line in the user's config, and neither lets a plugin declare it (settings.json accepts only agent / subagentStatusLine; Grok's plugin format has no status-line slot). So the last step is a script that writes that one setting for you. It backs up what was there, refuses to overwrite someone else's status line unless you pass --force, and --remove puts yours back.

On Windows the Grok script writes one more file: a small cc-usage.cmd beside cc-usage.mjs, with the config naming that batch and nothing else. Grok starts command through CreateProcess, which refuses an argument that is an absolute path (os error 123) and refuses a quoted program name just the same, so a bare executable path is the only form that survives — the node call lives inside the batch, and the config names it unquoted, which also keeps a path containing a space working.

What it shows

Window	Meaning	On GOAT
5-hour	Rolling burst limit — one long session cannot drain the month	$14
Weekly	Rolling 7-day limit	$35
Monthly	The billing period's credit allowance	$70

Each window shows percent used, a bar, and when it resets (a countdown under a day, a date beyond that). The monthly one also shows the credit left.

Colors follow how full the window is — green under 60 %, amber to 85 %, red above. (The terminal and HTML panels use a slightly earlier 50 / 80 split; the status line is the one that had to be tuned for a glance, so it warns later.)

Plans with no rolling windows (Provider, Enterprise) show the balance alone. Plans without API access (Go) render nothing at all in the status line — no error, no empty box. Ask for the terminal panel on such a plan and it does fail loudly, because there you asked a direct question and silence would be the wrong answer.

It hides itself when you are not using it

If you configured Command Code but switched to another model, a permanent quota bar is noise. The script decides per turn whether this session is actually routed there:

1. Your local router's own mapping — tools like cc-switch write ANTHROPIC_DEFAULT_OPUS_MODEL / ..._MODEL_NAME pairs into the env; the script reads the pair to learn the real upstream model. This is the router's own configuration, not a guess. (This one is an inference rather than a measurement — see what was verified — so if it ever comes up empty the script falls through to the next level rather than guessing.)
2. The model name the host hands over directly. Codex's hook puts model in as a plain string ("gpt-5.6-terra") — and gives an empty transcript_path, so without this level Codex would never resolve. Claude Code sends an object here, so it skips to level 3.
3. The session transcript — the model each message actually used (Claude Code's message.model, Grok's modelId).
4. Account activity — fallback, only when the three above say nothing.

The resolved model is checked against Command Code's public model catalog (/provider/v1/models, no auth needed). Not in the catalog → hidden.

Some model names are genuinely ambiguous (a bare claude-opus-5 exists both natively and in Command Code's catalog), and those are deliberately not guessed — add your own with --model <substring> instead.

Commands

Claude Code, Codex, Grok and ZCode also get a /quota command that prints the compact panel. That one does go through the model — it is a prompt, so it costs a turn; the status line is the free path, and /quota is for when you want the numbers in the transcript.

pi has /ccq-bar on|off|toggle|refresh|status instead (handled by the extension, no model turn), and the opencode adapter has no command at all — its sidebar is the whole surface.

Credentials

Found automatically, in this order:

1. COMMAND_CODE_API_KEY / COMMANDCODE_API_KEY / CMD_API_KEY
2. any env var whose name contains commandcode
3. ~/.commandcode/auth.json (the official CLI's login state)
4. a Command Code provider route in your agent's own config (~/.zcode/…, ~/.config/opencode/…, ~/.claude/settings.json, ~/.pi/agent/…)
5. dsh / Codex / Grok config.toml or settings.yaml, including apiKeyEnv indirection

If nothing is found the status line simply does not render — it never prints an error into your editor.

One exception: the DeepSeek Harness adapter (plugins/dsh) keeps its own data layer rather than importing core/cc-usage.mjs. Its host/client halves and 141 offline checks are built against that contract, so it stays as-is for now. See plugins/dsh/README.md.

Layout

core/cc-usage.mjs        ← the only implementation; everything else is a synced copy
scripts/sync-core.mjs    ← copies core into every adapter (--check for CI)
docs/FINDINGS.md         ← what was measured, and what to revisit when things change
plugins/<agent>/         ← one thin adapter per agent

The adapters are copies, not imports, so each one is self-contained and can be installed without a package manager. Edit core/, never a copy, then run node scripts/sync-core.mjs.

Requirements

• Node 18+ (for the script; the adapters themselves need nothing else)
• A Command Code plan with API access — the $1 Go tier does not have one
• Windows: Claude Code runs status line commands through Git Bash when it is installed, through PowerShell when it is not. Git Bash itself costs ~29 ms per repaint, and PowerShell roughly three times that — so if repaints feel slow, install Git Bash.

A note on pacing warnings

The script computes a burn-rate projection. The status line and the Codex hook never show it; the terminal panel, --compact, --md and --html still print it, and --json always carries it.

It stays out of the status line for a reason: extrapolating from a short sample says "you will run out" almost every time — 25 minutes into a 5-hour window a normal burst projects to 140 % — and a warning that is always on is not a warning. Where it is printed you asked for a panel, so the extra line costs you nothing.

Contributing

node scripts/check.mjs          # everything: sync, rendering, gating, secrets, dsh's 141 checks
node scripts/check.mjs --quiet  # one line per suite

That is the same script CI runs, so a local pass means a green build. The dsh suite needs React first — see plugins/dsh/README.md.

core/cc-usage.mjs is the only file you edit. The per-adapter copies are generated; after a change to the core run node scripts/sync-core.mjs (or let the check tell you).

License

MIT — see LICENSE.