mirror of
https://github.com/BigMarc/TokenGem.git
synced 2026-10-13 17:03:10 +02:00
README, changelog, and upstream license/attribution surfaces for the public distribution of TokenGem. DMG binaries ship via GitHub Releases.
2.6 KiB
2.6 KiB
Changelog
All notable changes to TokenGem are documented here. Upstream tool version bumps are called out explicitly, because they are the changes most likely to affect routing behaviour.
[1.0.0] — 2026-09-04
First public release.
The menu-bar app
- Native macOS app (universal binary, macOS 13+): measured net cost-per-turn delta in the menu bar, red when net-negative; panel with session-by-session sparkline against the baseline median, the four deltas verbatim, profile switcher, component health, hook-conflict list, Repair install, and Open Dashboard.
- Starts at login and launches silently; adopts a dashboard server you started yourself and restarts one that dies.
- Upgrades from pre-rebrand builds migrate
~/.codegem→~/.tokengemonce, on first run. A baseline cannot be recaptured, so it is never left behind.
The router
- One
PreToolUsehook per agent, owned by TokenGem. Per-command dispatch: Graphify's index when fresh, RTK where an equivalent exists, pass-through otherwise — never double-processing. - Three profiles: Balanced (default), Maximum savings, Index-only. Switching re-writes the router in place; the ledger records which profile was active for every event.
The measurement engine
- Mandatory baseline capture per agent from real session history before install —
installrefuses to run without it. - Continuous local ledger: command, routing path, raw vs. delivered output bytes, per-session token and cost totals, net delta vs. baseline.
- Output-token watch: when compression provokes longer output, the dashboard says so and names the profile to switch to. Net-negative results are never softened.
- Model pricing table for every provider TokenGem can measure, user-overridable via
~/.tokengem/pricing.json; unknown models are reported UNPRICED, never guessed.
Agents
- Claude Code, Codex, Cursor, OpenCode, Copilot, Gemini CLI, Qwen Code — plus proxy mode for apps that call a model API directly (requests forwarded untouched; only token counts recorded).
Upstream pins (see LICENSES/ and upstream.json in the source tree)
- RTK v0.48.0 (
fde0a8f) · Apache-2.0 - Ponytail v4.9.0 (
0a4dd63) · MIT - Graphify v0.9.53 (
33362d9) · Apache-2.0 (pre-relicense contributions under MIT)
Known limitations
- The binary is ad-hoc signed; Developer-ID notarization is pending, so first launch needs right-click → Open.
- Node 20+ is required for the CLI/dashboard the app supervises.
- Per-tool dollar attribution is reported only where instrumentation supports it; Graphify and Ponytail effects are not inventable and are not invented.