auto-engineer: Skill for Claude Code

usage

The authoritative source for session-quota state is the set of anthropic-ratelimit-unified-* response headers Anthropic attaches to every /v1/messages call. Claude Code reads them into memory and feeds them to its statusline — they are the real numbers, not an estimate.

Since that state lives only in the running Claude Code process, this skill fetches it by making a minimal 1-token probe call and parsing the headers. Result is cached for 60s at ~/.claude/cache/usage-probe.json so repeat invocations are free.

Primary command

bash .claude/skills/usage/probe.sh

(Add --force to bypass the 60s cache.)

Output is a single-line JSON like:

{"http":200,"fetched_at":1776054016,
 "5h":{"utilization":0.38,"resets_at":1776056400,"status":"allowed"},
 "7d":{"utilization":0.33,"resets_at":1776567600,"status":"allowed"},
 "overage":{"utilization":0.0,"resets_at":1777593600,"status":"allowed"}}

Fields: utilization is 0..1+ (can exceed 1 when the window is over-consumed), resets_at is Unix epoch seconds, status is allowed / allowed_warning / rejected.

Output format

Always emit these two machine-readable lines first (auto-engineer step 9 parses them):

remaining_pct: <max(0, 100 - 5h.utilization*100), rounded to 1 decimal>
reset_ts: <ISO-8601 of 5h.resets_at>

Then a short human summary: 5h %, 7d %, overage %, each window's status, time-until-reset, and a note if status=rejected (Claude Code is currently blocked on that limit until reset).

Fallback

If the probe returns an error (no credentials, non-200, missing headers), fall back to ccusage:

npx -y ccusage blocks --active --token-limit max --json

Label fallback output clearly as estimated — ccusage benchmarks against the user's all-time largest block, not the plan's real cap, so its percentage is routinely off vs. the probe. Use projection.remainingMinutes and endTime from .blocks[0].

Caveats

• The probe costs one tiny Haiku call (~22 in / 1 out, well under $0.001). The 60s cache keeps this negligible.
• overage appears only on plans where extra-usage is enabled; absent otherwise.
• If the user is fully rate-limited (5h.status == "rejected"), the probe itself still succeeds — Anthropic returns headers on 200s and rejects only subsequent inference calls. If it ever fails with 429, report remaining_pct: 0 and surface the cached reset_ts from a prior probe if one exists.