Skip to content

AGENTS.md

Latest commit

 

History

History
32 lines (22 loc) · 1.39 KB

AGENTS.md

File metadata and controls

32 lines (22 loc) · 1.39 KB
Raw

AGENTS.md

Bun + TypeScript monorepo. Keep changes small, readable, and boring.

Goal

Investigate with AI, but use as little AI as possible. Prefer deterministic logic, parsing, filtering, scoring, and evidence reduction before calling an agent.

Structure

  • packages/core — investigation pipeline, prompts, diagnosis, validation, evidence rendering.
  • packages/tools — agent tools. Tools gather evidence only; no mutation of cluster/cloud.
  • packages/clients — real + fixture clients for GCP, Prometheus, Jaeger, Kubernetes.
  • packages/cli — thin command shell: onboard, investigate, bench, input/output wiring.
  • packages/bench — fixture scenarios and scoring.
  • packages/notify — Lark card build/send.

Keep features in core, tools, clients, or notify when possible. CLI should stay thin.

Rules

  • investigate is read-only. propose_remediation records suggestions, never executes them.
  • Agent may run against production kube contexts. Do not add write-capable tools unless explicitly asked.
  • Preserve subpath exports used across packages, especially @opsremedy/core/types and @opsremedy/clients/*.
  • Match nearby code when touching clients, tools, or notifications.

Style

  • Comments: short WHY only. No banners. No obvious comments.
  • Prefer descriptive names over comments.
  • Prefer simple code over clever abstractions.
  • No new dependency unless user approves.