Version 0.4.0

agent-devkit 0.4.0

agent-devkit is a lightweight, prompt-driven workflow toolkit for coding
agents. It gives agents a consistent way to understand a repository, plan and
implement changes, debug issues, verify work, and keep an LLM-facing codebase
wiki aligned with the source.

This release packages the same 12 reusable Markdown skills for multiple coding
agent harnesses while keeping the canonical skills portable and dependency-free.

What is included

• using-devkit routes each request to the appropriate workflow.
• setup-codebase creates a project-specific agent contract and documentation
  skeleton.
• setup-openez and read-codebase-context support evidence-based codebase
  exploration, with direct-source fallbacks when OpenEZ is unavailable.
• brainstorm-feature, plan-feature, and implement-task provide a staged
  feature delivery workflow.
• systematic-debugging guides root-cause investigation before implementation.
• review-and-verify provides a final evidence-based review and verification
  gate.
• document-wiki maintains source-grounded LLM documentation under
  docs/llm/.
• context-handoff preserves useful state when work is paused or a session is
  compacted.
• estimate-feature optionally produces per-task effort ranges when an
  estimate is requested.

Native harness packaging

The repository includes optional adapters for Codex, Claude Code, Cursor,
Devin CLI, and OpenCode. These adapters reuse the canonical skills/ tree and
provide harness-specific discovery and bootstrap behavior where supported.

The shared bootstrap loads using-devkit at session start for hook-based
harnesses. The OpenCode adapter registers the skills directory and injects the
bootstrap into the first user message. Skill references work in both
namespaced plugin installations and direct skills installations.

Design goals

• Keep workflow instructions readable, reviewable, and easy to customize at
  the source level.
• Prefer repository evidence over guesses when tracing code or documenting
  behavior.
• Keep changes small, explicit, and verifiable.
• Separate design, implementation, debugging, and review responsibilities.
• Avoid runtime dependencies and build tooling for the canonical Markdown
  skills.

Verification

Version 0.4.0 includes dependency-free smoke checks covering hook output,
OpenCode bootstrap behavior, duplicate protection, skill registration, and
release metadata consistency. The Codex plugin manifest and Claude Code
marketplace manifest also pass their available validators.

The skills can be used directly from an Agent Skills-compatible directory, or
distributed through the native packaging for the supported harnesses.