Skip to content

release-v0.1.0.md

Latest commit

 

History

History
92 lines (65 loc) · 2.63 KB

release-v0.1.0.md

File metadata and controls

92 lines (65 loc) · 2.63 KB

v0.1.0 Release Notes Draft

Summary

agent-debugger is an early open-source MVP for turning repeated backend troubleshooting habits into executable runbooks for AI agents.

Repository: debug-runbook

This release focuses on one narrow but reusable layer:

  • runbook selection
  • ordered evidence collection
  • evidence normalization
  • evidence-based conclusion generation
  • MCP-based integration surface

It does not ship the original internal production DAG, permission system, or repair workflows that inspired the project.

Why This Exists

Many online incidents are investigated with a repeatable sequence:

  1. compare actual behavior with the expected result
  2. inspect cache
  3. inspect persistence
  4. inspect trace / execution flow
  5. decide which evidence-backed conclusion is most likely

This repository packages that style of investigation into a replayable, inspectable framework.

What Is Included In v0.1.0

  • zero-config replay demo
  • benchmark over built-in fixture cases
  • metadata consistency checks
  • built-in runbooks for:
    • missing expected side effects
    • stale cache views
    • abnormal persisted state
  • custom runbook loading through config
  • read-only adapter model for Langfuse, Postgres, and Redis

Context Coverage In This Release

Built-in runbooks intentionally cover a limited set of locator types:

  • request_not_effective: request_id, order_id
  • cache_stale: order_id, task_id
  • state_abnormal: order_id, task_id

Broader locator support should be added through custom runbooks or later releases.

What Changed Before This Release

  • rewrote the README around a real incident pattern and a zero-config demo path
  • reduced installation friction by adding pnpm demo
  • fixed config interpolation so unresolved env vars are preserved safely
  • made runbooks: config entries work at runtime for custom runbooks
  • tightened package boundaries with engines, files, and prepack
  • documented MVP boundaries more explicitly

Suggested Validation Commands

pnpm test -- --run
pnpm build
pnpm demo
pnpm benchmark
pnpm check
npm pack --dry-run

Known Limits

  • built-in runbooks are intentionally narrow
  • benchmark results validate replay cases, not every real integration path
  • no write-path or repair automation is shipped in this repository
  • real production observability and orchestration setup must be provided by the user

Suggested Launch Framing

Use wording like:

  • early open-source MVP
  • research preview
  • replayable incident investigation framework

Avoid wording like:

  • production-ready self-healing platform
  • general autonomous debugging engine
  • safe automatic repair system