v1.10.0

jcode-routed Codex adapter (Antigravity/Cursor stayed on their dedicated adapters per Phase 0's empirical spike - jcode 429s where the native agy/cursor-agent binaries work), agency-agents subagent picker synced into Claude Code's own agents dir, and a full Apple HIG sidebar rework of the app UI (fixes a dark-mode color byte-order bug, adds a macOS System Settings-style sidebar nav, consistent button/card styling). 146 tests green.

v1.9.1

Rolling context proxy, unified model router, cache management diagnostics, provider fixes (Groq/OpenCode/Unsloth).

app-v1.9.0

fix: OpenCode Go needs only an API key; wire Unsloth CLI install, pre…

app-v1.8.0

fix: installer no longer bundles removed run_benchmarks.py

Benchmark execution/reporting infra was already stripped out of the app
earlier this session, but the installer script and Product.wxs still
tried to copy/harvest run_benchmarks.py, which no longer exists at repo
root - this hard-failed the app-v1.7.0 MSI build in CI. Drops the
copy step and the WiX component/ComponentRef for it.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

app-v1.6.0

chore: bump installer version to 1.6.0

app-v1.5.0

TokenOptimizer App v1.5.0: VS Code extension fully ported off legacy …

app-v1.4.0

TokenOptimizer App v1.4.0: resource-aware benchmark model gating, har…

app-v1.3.0

TokenOptimizer App v1.3.0: tabbed UI, live install progress, real imp…

app-v1.2.0

TokenOptimizer App v1.2.0: install/build audit fixes, CI release work…

TokenOptimizer App v1.1.0 (Windows installer)

Adds benchmark reporting/export tooling to the in-app benchmark runner, plus fixes to the underlying benchmark pipeline surfaced by this release's testing.

Install: download TokenOptimizer.msi and run it.

New in the app

• Export for AI Review - zips every benchmark result plus a calibrated review prompt, copies the prompt to your clipboard, and reveals the zip in Explorer so it's ready to hand to an AI for quality scoring.
• Generate Report - writes BENCHMARK_REPORT.md (per-model scoring matrix + averages across all test prompts) and opens it.

Pipeline fixes (run_benchmarks.py)

• Fixed a bug where Hugging-Face-URL-sourced models downloaded successfully but then always failed to load (lms load needs the resolved catalog key, not the source URL).
• Fixed a purge bug that could delete the wrong model's files when two downloads landed at a similar size.
• Per-model generation configs (temperature/context/reasoning budget) sourced from each model's own docs, replacing a one-size-fits-all config that starved reasoning models of the budget they need to produce a real answer.
• The automated quality score no longer gives full credit to responses that never reached a final answer.

Unsigned build - Windows SmartScreen may warn on first run ("More info" -> "Run anyway").