Load balance

Report plan usage headroom across Cursor, Codex, and Claude Code. Read-only — no account switching, no spawning agents.

Run the probe

From this skill directory:

python3 scripts/usage.py

Or pass an absolute path to scripts/usage.py when the skill is installed elsewhere.

The script prints normalized JSON to stdout. Parse limits on each agent before rendering.

Render the table

After the probe, show one row per limit (not one row per agent):

Row rules:

• Agent — cursor, codex, or claude
• Limit — label from each limits[] entry (e.g. 5-hour, 7-day Fable, Code review 7-day, Composer / Auto)
• Used% — used_percent when status: ok; otherwise unknown
• Resets — human-readable resets_at (or —)
• Notes — optional tags from JSON: active, severity (e.g. warning), model for scoped Claude limits

Sort agent groups by ranking score (lowest minimum used_percent across that agent's limits first). Within an agent, keep limits in the order returned by JSON.

Mark the recommended agent on its tightest limit row (lowest used_percent for that agent) with recommended in Notes. If recommended is null globally, say no recommendation.

When an agent has error, add one summary row with Auth missing/failed and the error message.

Stop here

This skill does not spawn or resume agents. For cross-agent CLI delegation and session monitoring, use $agent-delegate and $babysit.

Errors

When error is set, include the message and fix hint (codex login, claude /login, sign in to Cursor, or CURSOR_SESSION_TOKEN).

Probes use undocumented vendor endpoints and may break. See references/output-contract.md.

Example

Use $load-balance before I pick where to run this refactor.