devops-refresh-usage

name	devops-refresh-usage
version	0.2.0
description	Fetch live token usage for completion card battery line. Reads the native statusLine-written usage file first (no scrape), falls back to the Edge CDP scraper or cached data. Run silently pre-card, or manually: "refresh usage", "wie viel hab ich verbraucht", "token budget".
allowed-tools	Bash(node *), Read

Refresh Usage Data

Fetch live token usage for the completion card's battery line.

Step 0 — Load Extensions

Check for optional overrides. Use Glob to verify each path exists before reading. Do NOT call Read on files that may not exist — skip missing files silently (no output).

1. Global: ~/.claude/skills/refresh-usage/SKILL.md + reference.md
2. Project: {project}/.claude/skills/refresh-usage/SKILL.md + reference.md
3. Merge: project > global > plugin defaults

Step 1 — Fetch data

1·0. Native source first (no scrape)

~/.claude/usage-live.json is kept minute-fresh by the native statusLine writer (statusline-usage.js, registered by ss.statusline.ensure) — it maps Claude Code's rate_limits JSON onto the usage schema with no browser and no extra Claude turn. If the file's timestamp is recent (≤ a couple of minutes), just read it — you are done; skip the Edge scrape entirely. Only when it is stale/absent (pre-first-API response, an unsupported login, weeklySonnet needed, or a host without the statusLine writer) fall through to the Edge scraper below. The dotclaude-completion MCP server applies this same warm-read-first logic automatically in get_usage / render_completion_card.

Automatic path is zero-interaction. The MCP fallback always runs the scraper with --no-login, so the completion card never opens a login window — a logged-out profile just serves statusLine/cached data. A one-time login is offered only when you run this skill manually (the command in 1a omits --no-login). The card itself renders only the 5h + weekly numbers, which the native statusLine source already provides token-free; weeklySonnet is a manual-summary extra, not a card field.

1·1. Edge CDP fetcher (fallback)

The script spawns a dedicated, isolated Edge instance with its own user-data-dir under ~/.claude/edge-usage-profile. It is completely independent from the user's main Edge — separate cookies, separate processes, no tabs touched. It fetches usage via a cookie-authed in-page call to the internal API (GET /api/organizations/{id}/usage + /rate_limits) — the settings page became an SPA overlay (2026-05) that never renders headless, so DOM parsing is only a single-grab last resort. Runs headless; the hidden instance is reused across runs.

The script path is ${CLAUDE_PLUGIN_ROOT}/scripts/refresh-usage-headless.js (use the plugin root, NOT a relative path).

1a. Run the scraper (manual run — login allowed)

This skill is the manual entry point (the user explicitly asked for usage / weeklySonnet), so the command deliberately omits --no-login: a one-time login window may open here. The automatic card path uses --no-login and never does.

node "${CLAUDE_PLUGIN_ROOT}/scripts/refresh-usage-headless.js" --quiet --summary

Exit codes:

• 0 → success. Usage written to ~/.claude/usage-live.json.
• 2 → scraper profile is not logged in. On this manual run a visible Edge window has been opened at https://claude.ai/login. Tell the user inline, e.g.:

  ⚠️ Der Usage-Scraper hat ein eigenes Edge-Profil unter ~/.claude/edge-usage-profile. Dieses ist noch nicht bei Claude eingeloggt — ein Edge-Fenster wurde gerade geöffnet. Bitte einmal einloggen (Cookies bleiben danach im Scraper-Profil persistent). Danach funktioniert der Scraper dauerhaft headless im Hintergrund. After the message, use cached data for this turn (see 1c).

• 3 / 4 / 5 → scrape/launch failure. Fall through to 1b.

1b. Browser fallback (optional)

If the scraper fails repeatedly, use the Playwright tools:

browser_navigate → https://claude.ai/settings/usage

Parse the page text using the same percentage/reset patterns from Step 2.

1c. Cached data (automatic)

The script already falls back to ~/.claude/usage-live.json with a [cached Xm ago] label when scraping fails and a cache exists. You don't need to do anything extra — just read the file.

1d. No data

Only after 1a–1c all fail: show [no data] with the specific reason.

Step 2 — Parse and store

Write results to ~/.claude/usage-live.json:

{
  "session": { "used_pct": 42, "resetInMinutes": 192 },
  "weekly": { "used_pct": 18, "resetInMinutes": 2885 },
  "timestamp": "2026-03-27T10:30:00Z",
  "_cached": false
}

Step 3 — Compute burn rate

Used by the completion card's battery line:

burn_ratio = token_pct / elapsed_pct

5h window:  elapsed_pct = (300 - resetInMinutes) / 300 * 100
Weekly:     elapsed_pct = (10080 - resetInMinutes) / 10080 * 100

ratio < 0.5  → 🔋 + "Viel Spielraum"
ratio 0.5-1.3 → 🔋 (no hint)
ratio > 1.3  → 🪫 + "Hoher Verbrauch — neue Session oder Haiku empfohlen"

Rules

• Never touch the user's main Edge. The scraper always spawns a dedicated, isolated Edge instance with its own user-data-dir (~/.claude/edge-usage-profile). It reaps only that profile's instances — matched by the --user-data-dir on the live command line, not by a stored PID (which dies on Windows when Edge re-execs into its singleton, the bug that let orphan instances pile up). The user's main Edge is never matched.
• Login windows only on manual runs. The automatic card path passes --no-login, so it never opens a window (a logged-out profile just serves cache). On a manual run (this skill, no --no-login) a one-time login is expected, opened at most once: on first run (or after a profile wipe) the scraper profile has no cookies, so the script opens a visible login window, writes a sticky edge-usage-login-pending.json marker, and exits code 2. While that marker is fresh (≤ 30 min) no session opens another window — so parallel sessions can't stack login windows. Tell the user inline — do NOT retry silently.
• Silent after first login. A successful scrape clears the marker immediately; once the profile has cookies, all subsequent runs are invisible and reuse the one hidden instance.
• Transient render failures never open a window. A slow/unrendered page returns a scrape error (cache fallback), not code 2. Only an explicit /login redirect or login UI counts as logged-out.
• Playwright fallback is acceptable — if the scraper fails, opening a browser tab via Playwright to scrape is fine.
• Delta computation: Read usage-live.json before and after refresh. Delta = new_pct - old_pct.
• Script path: Always use ${CLAUDE_PLUGIN_ROOT}/scripts/refresh-usage-headless.js, never a relative path.