BUDGET — Session token consumption check

You produce a quick estimate of where the current session stands on token usage, what's costing the most, and what to do next.

What you can actually check

Claude Code does not expose live token counters directly inside the SKILL. You can:

1. Check session stats cache:

   cat ~/.claude/stats-cache.json 2>/dev/null | python3 -m json.tool | head -60

2. Check recent conversation length proxies:
   • Number of tool calls so far (visible to you as the calling LLM)
   • Approximate size of files Read so far
   • Whether any background agents have run (TaskOutput sizes)
3. Look at file-history size in ~/.claude/file-history/ — recent activity intensity.

Output

## Session budget check

**Approximate context window usage:** [low / medium / high / near-limit]
**Heaviest consumers in this session:**
1. [What — e.g. "/auto pipeline with 6 workers (~75K tokens)"]
2. [What]
3. [What]

**Recommendations:**
- ✅ Healthy — proceed normally
- 🟡 Medium — consider `/compact` before next big task
- 🔴 High — `/compact` now, or `/clear` if topic-switching
- 🆘 Near-limit — `/clear` and start fresh with `@~/.claude/MEMORY.md`

**To save tokens going forward:**
- Use `/lean` for routine tasks
- Use Haiku model in `/auto haiku <task>`
- Avoid Reading files >500 lines (use Grep+offset Read)
- Delegate exploration to subagents (their context is isolated)

**Anthropic data points (for context):**
- Multi-agent ≈ 15× tokens of chat ([source](https://www.anthropic.com/engineering/multi-agent-research-system))
- 80% of perf variance in browsing comes from token usage
- Prompt caching read = 0.1× write cost — large stable CLAUDE.md is essentially free after first cache hit

Hard rules

1. No precise dollar amounts unless the user has provided per-token pricing in their config.
2. No alarmist language — calibrate severity honestly.
3. Always end with concrete next action.