Skip to content

Releases: CommonstackAI/UncommonRoute

Release list

v0.7.1

v0.7.1

Choose a tag to compare

View all tags
@anjieyang anjieyang released this 21 Apr 01:56
v0.7.1
c1c316b

Highlights

  • Switch Claude Code proxy setup from ANTHROPIC_API_KEY to ANTHROPIC_AUTH_TOKEN, which is more reliable for proxy/gateway mode
  • Update the guided init flow, setup claude-code output, and README quick start to match the new Claude Code auth shim
  • Add a regression test covering the Claude Code shell block and keep the 0.7.1 release notes/docs aligned

Packages

  • Python: uncommon-route==0.7.1
  • npm: @anjieyang/uncommon-route@0.7.1
Loading

v0.7.0

v0.7.0

Choose a tag to compare

View all tags
@anjieyang anjieyang released this 21 Apr 01:24
v0.7.0
cb34b01

Highlights

  • Add guided uncommon-route init onboarding and improve doctor / CLI help flow
  • Add independent request traces, support bundle, and request-level diagnostics lookup
  • Improve selector / routing behavior for tool-selection and system-prompt-heavy requests
  • Update English and Chinese READMEs to make init the recommended first-run path

Packages

  • PyPI: uncommon-route==0.7.0
  • npm: @anjieyang/uncommon-route@0.7.0
Loading

v0.6.0 — Beats Semantic Router

v0.6.0 — Beats Semantic Router

Choose a tag to compare

View all tags
@anjieyang anjieyang released this 16 Apr 06:06
v0.6.0
c6996f4

What's New

Accuracy 79.4% — surpasses Semantic Router (78.7%) on CommonRouterBench.

Signal Improvements

  • Embed+Meta combined features: Embedding vector + 7 metadata features (msg_count, tools, code blocks, etc.) fed into a single classifier. BFCL accuracy jumped from 82% to 96%.
  • Unified hybrid classifier/KNN: Single code path for production and preview. Classifier when confident, KNN fallback when uncertain.
  • Multi-step workflow dampening: MetadataSignal reduces confidence on long tool-heavy conversations so the semantic signal can dominate.
  • Low-confidence escalation: LOW predictions with confidence < 0.55 automatically bump to MID.
  • Conditional Signal B: Structural text classifier activates for longer conversations where it demonstrably helps.

Dashboard (Nothing Design)

  • All 9 pages with unified Doto titles, sidebar navigation groups (MONITOR/CONFIGURE/INTERACT)
  • Page fade-in animations, hover transitions, improved contrast
  • Playground: cold-start indicator, 3-signal readout with adaptive weights
  • Home page redesigned: hero savings + 4-column stat grid + dot-grid live traffic

Benchmark (970 samples, 5 benchmarks)

Metric v0.5.x v0.6.0
Accuracy 73.0% 79.4%
Pass Rate 90.3% 93.4%
Cost Savings 78% 82%
Overall 78.3 84.0

Install

pip install uncommon-route==0.6.0
Loading