Feature: /usage command — token usage and limits

[Rustam-Z]

Problem

There's currently no way to see how many tokens Luna has consumed, what the current context window usage is, or whether any rate limits are approaching — without digging into logs manually.

Proposal

Add a /usage owner-only slash command (intercepted by the harness, not Claude) that returns a snapshot of current token and rate-limit state.

Sample output:

📊 Usage snapshot — 2026-05-11 06:02 UTC

Context window:  ~320K / 1M tokens (32%)
Session length:  4h 12m (since last restart)

API usage (rolling 24h):
  Input tokens:   1,842,300
  Output tokens:    214,500
  Total:          2,056,800

Rate limits:
  Requests/min:   12 / 50
  Tokens/min:     48,200 / 200,000

Subagents spawned today: 5
  └ trends-uzbekistan: 20,045 tokens
  └ trends (global):   18,108 tokens
  └ github-watch:      14,225 tokens
  └ ...

Backups: 3 prompt backups in data/prompt_backups/
Memory:  self/learnings.md — 14.2 KiB / 64 KiB

Implementation notes

• Context window % can be derived from the usage field in Claude API responses (input_tokens / max_context)
• Rolling 24h API usage requires the harness to accumulate token counts from each CC response
• Rate limit state is available from Anthropic API response headers (anthropic-ratelimit-*)
• Subagent token usage is already returned in task notifications — harness should aggregate these
• Memory file sizes can be read from disk at command time

Access

Owner-only (same gate as /health, /audit). Should be interceptable by the harness before reaching Luna, so it works even when Luna's context is full.

Related

• /health already returns some state — /usage is a deeper drill-down focused on tokens and cost visibility
• Issue [Bug/Feature] Context window overflow is silent — user not notified when session resets #28 (context compaction) — /usage gives the signal needed to decide when to compact