This file overlays CLAUDE.md with local experiments and private context. When a local rule measurably improves outcomes, promote it to CLAUDE.md with an ADR. When it fails, it stays local.
Status: ✅ VALIDATED - Ready for promotion
Hypothesis: Stop hooks returning JSON with {"ok": boolean} field prevents schema validation errors
Result: Schema validation errors eliminated in rebuilt container
Metrics: 0 errors in 5 session stops vs 100% failure rate before
Promotion: Already integrated into entrypoint-unified.sh
Status: ✅ VALIDATED - Promoted to CLAUDE.md Hypothesis: MCP memory tools provide better cross-agent coordination than CLI Result: 19,659+ entries accessible via MCP vs 47 via CLI Metrics: 100% reliability for memory operations across agent boundaries Promotion: Documented in all coordinating CLAUDE.md files
Status: 🔬 TESTING Hypothesis: Typed constitution with task-scoped shards improves long-horizon autonomy Metrics to track:
- Autonomy duration before intervention
- Cost per successful outcome
- Tool/memory operation reliability
- Runaway loop self-termination rate Started: 2026-02-03
- Use external RuVector PostgreSQL for all agent coordination
- CLI memory commands are for debugging only
- Stop hooks MUST return
{"ok": boolean}JSON format - All hooks should use
|| truefallback for resilience
- Container: turbo-flow-unified (rebuilt 2026-02-03)
- External memory: ruvector-postgres:5432
- Docker network: docker_ragflow
- supervisorctl requires sudo in this container
- npx calls are slower than global claude-flow binary