Context-Engineering AGENTS.md

AGENTS.md — Project-Wide Persistent Context

Identity

This is the Context Engineering + Closed-Loop Agent Benchmarks template repository, created by Belkis Aslani. It provides a production-grade scaffold for AI-assisted development workflows with built-in quality gates and benchmark-driven feedback loops.

Hard Rules

1. Benchmark-first policy: Every change must pass npm run benchmark before merge. Score must meet or exceed the target in benchmarks/thresholds.json.
2. No secrets in repo: The CI pipeline scans for leaked keys/tokens. Any match fails the build immediately.
3. ESM only: All TypeScript uses ES modules. No require(), no CommonJS.
4. Strict TypeScript: strict: true in tsconfig. No any without justification.
5. Minimal dependencies: Every new dependency must be justified. The dep scanner enforces a maximum count and a blocklist.
6. Deterministic benchmarks: All benchmark runners must produce identical scores given identical input. No network calls, no randomness.

Decision Rules

• Adding a runner: Create a new file in benchmarks/runners/, implement the BenchmarkRunner interface from benchmarks/types.ts, register it in benchmarks/run.ts.
• Changing weights: Edit benchmarks/scorecard.json. Weights must sum to 100.
• Changing thresholds: Edit benchmarks/thresholds.json. Document the reason in docs/decisions.md.
• Adding orchestrator steps: Extend tools/orchestrator/controller.ts state machine. Document in docs/orchestrator.md.

Where to Find Things

What	Where
Project docs	docs/
Copy-paste prompts	prompts/
Benchmark runners	benchmarks/runners/
Benchmark config	benchmarks/scorecard.json
Score thresholds	benchmarks/thresholds.json
Orchestrator	tools/orchestrator/
Reports	reports/
Example project	examples/context-minimal/
Lessons learned	memory/lessons.json
Patterns catalog	memory/patterns.md