v0.3.0 — 7B Models, E2E Test Suite, UI Improvements

What's New

7B Model Upgrade

• Default model: Qwen2.5-7B-Instruct — 96% E2E accuracy, 100% JSON reliability
• Available models: Qwen 2.5 7B, Hermes 2 Pro 7B, Llama 3.1 8B
• ~5GB VRAM, ~4.5GB download (one-time, cached via Service Worker)
• Up from 74% accuracy with 3B models in v0.2.0

E2E Browser Test Suite

• 27 tests across 8 categories (L2 basic + L3 advanced agentic)
• Validates full pipeline: user message → LLM reasoning → tool selection → tool execution → response
• Chrome DevTools MCP integration for Claude Code
• Test observability hook (window.__wpAgenticTestHook)
• 4 workflows tested at 100% pass rate

Bug Fixes

• JSON envelope ({"action": "final_answer", ...}) no longer leaks into user-visible responses
• Loading spinner ("Thinking..." / "Running tool...") appears inline in message flow after user's chat bubble
• Spinner dot properly aligned with timeline dots

Documentation

• Updated all docs (README, ARCHITECTURE, IDEAS, CONTRIBUTING) for 7B models
• Added tests/TESTING.md — comprehensive testing guide for unit and E2E tests
• Removed obsolete SLM-STRATEGY.md (7B models made the semantic translation layer unnecessary)
• Fixed branch strategy inconsistency in CONTRIBUTING.md

Model Comparison (E2E Results)

Category	Qwen 2.5 7B	Hermes 2 Pro 7B
Single Tool Selection (5)	100%	80%
Conversational (3)	100%	100%
Multi-Tool Chains (3)	67%	67%
Error Recovery (2)	100%	100%
JSON Reliability (8)	100%	63%
Goal Completion (3)	100%	67%
Complex Multi-Step (2)	100%	50%
Conditional Logic (1)	100%	100%
Total (27)	96%	74%

---

Full Changelog: v0.2.0...v0.3.0