| name | pipeline |
| description | Deep analysis of the FactHarbor CB pipeline. Use for debugging, architecture decisions, evidence quality issues, or reviewing changes that span multiple pipeline stages. |
| allowed-tools | Read Glob Grep Bash |
ultrathink
Analyze the FactHarbor ClaimAssessmentBoundary pipeline for: $ARGUMENTS
For pipeline fix decisions, isolate root cause before patches, prioritize quality over speed over cost, and use review/debate when the issue crosses stage boundaries or affects prompt/config behavior.
Before forming any conclusion, read every file relevant to the concern. Key files by area:
| Area | Files |
|---|
| Orchestration / data flow | claimboundary-pipeline.ts, research-orchestrator.ts |
| Claim extraction | claim-extraction-stage.ts |
| Query generation | research-query-stage.ts |
| Source acquisition | research-acquisition-stage.ts |
| Evidence extraction | research-extraction-stage.ts |
| Evidence filtering | evidence-filter.ts |
| Boundary clustering | boundary-clustering-stage.ts |
| Verdict generation | verdict-generation-stage.ts |
| Aggregation | aggregation-stage.ts |
| Types / contracts | types.ts, pipeline-utils.ts |
All files are under apps/web/src/lib/analyzer/.
Apply AGENTS.md rules throughout:
- No deterministic text-analysis logic — LLM intelligence for any meaning decision
- No hardcoded keywords or domain-specific logic
- Read fully before editing — do not propose changes based on partial reads
- After a focused validation failure (for example
npm test, npm -w apps/web run build, or the explicitly described manual check), classify the current attempt as keep, quarantine, or revert before proposing a broader multi-stage change
- Treat that as bounded backtracking, not blanket rollback-first behavior