claude-code-token-optimizer

A session can spend real money without any obvious moment where a read was wasteful or a search should have been scoped. Claude Code writes every turn to a local JSONL transcript including full usage metadata — input tokens, output tokens, cache reads/writes, tool calls, compactions — but surfaces none of it.

This extension reads those transcripts locally and answers two questions: where is the money going, and what specifically should I do differently?

Ten detectors

Each rule is one file, each carries its own token and cost estimate:

1. Context compactions
2. Context-pressure warnings before an auto-compaction
3. Broad or unscoped grep / rg / find searches
4. Uncapped cat / git diff / git log output dumps
5. Oversized or repeated reads of the same file
6. Reading a file back immediately after writing it
7. Redundant read-only inspection commands
8. Failed or interrupted tool calls whose output was billed and discarded
9. Low cache reuse — the signature of a volatile prompt prefix
10. High-input, low-output round trips

Findings roll up into recoverable-dollar totals by category, split into one-off fixes versus habits, alongside a daily efficiency series (cache-reuse percentage, cost per day, compactions) that is persisted so history survives transcript pruning.

Decisions worth noting

The data layer has no vscode dependency — parser, model, pricing, rules, and aggregation are pure TypeScript tested against fixtures and a temporary home directory tree. That’s what makes the rules engine testable at all.

Cost estimates are honest. Claude Code’s local stats report costUSD: 0, so the extension ships an editable rate table with cache reads at 0.1x and cache writes at 1.25–2x. The README says plainly that on an already-efficient setup recoverable waste can legitimately be small. I didn’t inflate the numbers.

Regression alerts require at least three active sessions and 20k tokens per comparison window, so a quiet week doesn’t trigger a false alarm.

Fully local. No telemetry, no external calls, zero runtime dependencies, 38 unit tests.

← all projects