Usage 0.18.0: Opening a new conversation now runs a health check to stop

Release notes

FEATURE

Opening a new conversation now runs a health check to stop quota leaks

Every time you open a new Claude Code conversation, the system now quietly runs a background health check on your logs; if it spots you burning quota pointlessly, the Progress Concierge will slip a one-line reminder at the end of its opening greeting. Just reply "show me," and the model will read the snapshot (updated daily in the background to ~/.claude/usage-diagnosis.json) to explain exactly what's wasting tokens and how to fix it. Once reminded about a specific issue, it won't pester you about it again for 7 days unless a new diagnosis appears; if the snapshot is too old (over 48 hours), it will smartly skip the warning altogether. It's like a mechanic doing a quick inspection under your car before every trip — they only tap your shoulder if there's an oil leak, and they won't nag you about the same puddle every single day.

BUGFIX

Smarter diagnosis rules stop inflating your wasted money

The system now flags your wasted tokens much more precisely and stops scaring you with wildly inflated estimates. We've armed the engine with five core rules (flagging repeated file reads, scanning trash directories like node_modules, anomalous session bloat, massive terminal outputs, and repeated bash commands), ensuring the most pressing waste is always surfaced first. Additionally, when it detects an overly long session, it no longer foolishly bills every single word at the most expensive rate and calls the entire session "waste", which used to inflate estimates by 9 times; it now smartly separates out your cheap cached (temporarily saved to skip recalculations) tokens, and only counts the exact portion that exceeds your project's baseline as actual waste. Like a taxi meter that used to point at the total fare and call the whole trip a loss just because the driver took one wrong turn — now it calculates exactly the few miles of the detour and applies your membership discount, handing you a fair and precise receipt.

Original changelog

Original changelogOpening a new conversation now runs a health check to stop quota leaks

**Health-check diagnosis on every new conversation**: usage now runs a background diagnosis engine against your Claude Code session logs and, when it finds meaningful waste, quietly appends a one-line reminder to the Progress Concierge's opening handoff. Say "show me" and the model reads the full snapshot (`~/.claude/usage-diagnosis.json`) and explains findings with specific suggestions. The reminder is suppressed for 7 days once a fingerprint is seen, re-surfaces when the diagnosis changes, and is skipped entirely when the snapshot is stale (>48 h).
**Daily diagnosis snapshot** (`usage_diagnosis_snapshot.py`): the menu-bar app refreshes `~/.claude/usage-diagnosis.json` once per day in the background so the cost estimate is always fresh when you open a new conversation.

Original changelogSmarter diagnosis rules stop inflating your wasted money

**Five-rule diagnosis engine** (`analyzer/diagnoser.py`): detects repeated file reads, polluter directories (node_modules, .venv, dist, …), anomalous session sizes, noisy Bash output, and repeated Bash commands. Findings are ranked by estimated token waste so the most actionable finding is always surfaced first.
**Anomaly-session waste estimates are no longer inflated ~9×**: the engine previously counted the entire token total of an anomalous session as waste and priced every token at the full $3/MTok input rate. Long sessions are dominated by cache reads billed at a tenth of that ($0.30/MTok), and the work done in the session isn't waste at all — only the excess over the project baseline is. Cost is now split by token type and scaled to the excess share (real-data result: $254 → $27).