claude-code-config: Skill for Claude Code

Report how much of each registered account's allowance is gone, when it comes back, and what continuing past it would cost — one row per account per window.

The cheapest-next hint never acts. It never switches accounts, never purchases anything, and never gates dispatch. It names an account and stops. Buying a Codex reset, enabling Claude usage credits, and raising a Cursor spend limit are the owner's actions, taken in the provider's own UI — do not offer to perform any of them, and do not treat "cheapest to continue on" as an instruction to re-route work. The prices it quotes are a checked-in table with last verified dates, so it is only as fresh as those dates.

The recorded history is display data too, and so is the projection over it. Every run appends one line per readable row to ~/.claude/ai-quotas-history.jsonl, and the START / %/DAY / LEFT columns are read back out of it. No dispatch gate reads that file or those columns, and none may start: they are not an input to credit-budget.sh, not a reason to pause, defer, downgrade, or re-route anything. The projection says how fast an account is being drained, and that answer is printed to the owner, never acted on. Do not hand-edit the file, and do not compute your own verdict from it.

Display only — never a gate. Nothing this skill produces may gate dispatch, pause work, downgrade a model, defer a launch, or feed credit-budget.sh. Quota and spend authority belongs to Anthropic's own in-app UI and upstream harness signals — see .claude/rules/safety.md §"Anthropic Quota & Spend Authority". This is a third, purely observational surface, distinct from the two that legitimately gate: the user-configured daily_credit_budget_usd budget and the #1427 usage-horizon counter. The earlier /quota skill was rolled back in issue #499 for exactly this — gating agent decisions on locally-read numbers. Answer the user's question and stop.

No credential value ever passes through you. The helper borrows each account's live token in place for one request — for Cursor, the token the Cursor IDE already holds for the signed-in user, turned into a session cookie in memory — and never prints any of them. Do not run a command that reads a credential out of the Keychain, a profile directory, or the Cursor IDE's state store yourself, and never repeat a token, cookie, or auth.json body into the conversation.

Step 1 — Resolve the helper

resolve_script() {
  local name="$1" candidate
  for candidate in \
    "$HOME/.claude/skills-worktree/.claude/scripts/$name" \
    "$HOME/.claude/scripts/$name" \
    ".claude/scripts/$name"; do
    if [[ -x "$candidate" ]]; then echo "$candidate"; return 0; fi
  done
  return 1
}
AI_QUOTAS_SH=$(resolve_script ai-quotas.sh || true)
[[ -n "$AI_QUOTAS_SH" ]] || { echo "ERROR: ai-quotas.sh not found (checked all three paths) — quota reporting unavailable" >&2; exit 1; }

STOP if the helper does not resolve: print that one line and end. Do not hand-roll a reader — a second caller of the provider APIs is how two answers start disagreeing, and reading a credential by hand is exactly what this design forbids.