Release v0.5.0

New Feature: Review Task Type

This release introduces independent review tasks for objective evaluation of AI-generated outputs.

Why Separate Reviews?

LLMs cannot reliably review their own outputs within the same context. The same assumptions and blind spots carry over into any "self-review." Running reviews in a separate session (or isolated sub-agent) provides a fresh perspective that catches issues the original session would miss.

New Review Tasks

Task	Purpose
code-review	Validate implementation against Design Doc, check code quality and architecture
technical-document-review	Check structure and content quality of Design Docs, ADRs, PRDs
integration-test-review	Verify test implementation quality, skeleton compliance, AAA structure

How to Use

Start a new session after implementation and request a review:

You: "Review the implementation in src/auth/ against docs/design/auth-design.md"

Cursor Users: MCP Integration

Cursor users can run reviews in isolated contexts without switching sessions using sub-agents-mcp. See README for setup instructions.

Other Changes

• Enhanced metacognition with "known unknowns" and "kill criteria" checkpoints
• Added Design Summary (Meta) section to Design Doc template
• Added Decision Details table to ADR template
• Improved README language and documentation

Internal: Responsibility Clarification

Clarified responsibility boundaries between framework files for better maintainability:

File	Responsibility
AGENTS.md	Entry point only, delegates to task-analysis.md
task-analysis.md	Task classification, references task-rule-matrix.yaml
task-rule-matrix.yaml	Single source of truth for task-rule mappings
metacognition.md	Runtime decisions, dynamic rule loading triggers

This reduces redundancy and makes the framework easier to extend.