session-usage

Files (1)

• SKILL.md

SKILL.md (first lines)

## Read

```sh
{ { security find-generic-password -s "Claude Code-credentials" -w 2>/dev/null \
    || cat ~/.claude/.credentials.json; } \
| jq -r '.claudeAiOauth.accessToken' \
| { read -r T; https --ignore-stdin --check-status --timeout=10 --print=b \
      GET api.anthropic.com/api/oauth/usage \
      "Authorization: Bearer $T" "anthropic-beta: oauth-2025-04-20"; } 2>/dev/null \
| jq -c '[.limits[] | select(.percent != null)
    | {w: (.scope.model.display_name // .kind), pct: .percent, resets: .resets_at}]'; } \
|| echo "usage unknown"
```

Run it verbatim: the token stays unprinted, `--ignore-stdin` keeps httpie off the credential pipe, `--check-status` makes an HTTP error exit non-zero.

`w` is `session` (5h), `weekly_all`, or a model name (that model's own weekly). `resets: null` means the window is not open. On `usage unknown`, carry on unmeasured — the endpoint is undocumented and never blocks work.

## Interpret

Percent of a window and time to its reset, and nothing else. No token or dollar figure is available, so reason in **proportions of a window**; a "how many agents can I afford" number would be invented. (Not `budget.spent()`, which counts per-turn output tokens against a user-set target.)

Windows bind independently. The tightest governs.

A single reading is a level, not a rate. For **burn rate**, read either side of a representative chunk and extrapolate the delta across what remains — early in a long run, while acting on it is still cheap.

## Act (optional)

- **Session (5h) tight** — waiting is cheap, the window refills within the working day. Finish the chunk, then offer the wait with its reset time.
- **Weekly tight** — waiting costs days, so scope is the lever: fewer parallel agents, narrower sweep, cheaper model for mechanical stages.

Surface headroom and propose the smaller scope; the user decides whether to stop.

From fuzzyfox/skills, MIT.

More from fuzzyfox/skills