This release fixes the long-standing issue where Task agents would not
reliably persist learning data. Now learnings are captured automatically.
## Major Feature: Automatic Learning Capture
- New `scripts/hooks/capture-task-learning.js` PostToolUse hook
- Captures agent type, task output, duration, token usage
- Calculates reward based on output quality indicators
- Stores to `learning_experiences` table in memory.db
- Deduplication: skips if agent already stored via MCP (60s window)
- Updated `aqe init` to copy hook scripts to user projects
- New phase: "Hook Scripts" in initialization pipeline
- Settings.json includes automatic learning capture hook
## Clean QE-Only Configuration
- Removed all claude-flow and agentdb dependencies from QE agents
- Updated settings.json with clean AQE env vars
- Removed agentdb.db references (deprecated)
- All persistence unified to `.agentic-qe/memory.db`
## Agent Learning Instructions Audit
- Added MANDATORY `<learning_protocol>` sections to all 30 QE agents
- 19 main QE agents updated
- 11 QE subagents updated
- Instructions include: query past learnings, store experiences, store patterns
## Files Changed
- 30 agent/subagent .md files with learning protocols
- scripts/hooks/capture-task-learning.js (new)
- src/cli/init/claude-config.ts - clean QE-only hooks
- src/cli/init/helpers.ts - copyHookScripts() function
- src/cli/init/index.ts - Hook Scripts phase
- .claude/settings.json - removed claude-flow hooks
- Version bumped to 2.3.0 in all required files
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Profa