develop-feature

name	develop-feature
description	Full development workflow from planning through deployment for AI Agent features. Use when developing new features, making significant architecture changes, or starting work on any non-trivial implementation.
<!-- disable-model-invocation	true -->

Develop Feature Workflow

Follow this 5-phase workflow when developing new features or making significant changes.

Phase 1: Planning & Design

1. Define agent goals: Clarify what the feature should accomplish, scope, and constraints
2. Choose architecture pattern:
   • ReAct: Single-agent tool-use loops
   • Plan-and-Execute: Tasks requiring upfront planning
   • Multi-Agent: Complex workflows with specialized agents
3. Design tool integration: Identify APIs needed, define typed schemas
4. Define evaluation criteria: Success metrics before coding
5. Externalize prompts: Draft prompts as separate files
6. Analyze requirements: Check for SOLID violations
7. Create decision record: Create a file in history/ following naming convention {MAJOR}_{MINOR}_{PATCH}_{SHORT_DESCRIPTION}.md

Phase 2: Core Development

1. Start simple: Begin with minimal working implementation
2. Build modularly: Develop each layer in isolation:
   • Prompt templates → LLM client → Tools → Orchestration → Memory
3. Write tests: TDD or alongside implementation
4. Version prompts: Treat prompt changes as code changes
5. Commit frequently: Use Conventional Commits (feat(scope): description)
6. Refactor immediately: Address code smells right away

Phase 3: Testing & Evaluation

1. Unit tests: Test individual components in isolation
2. Integration tests: Verify tool calls, error handling
3. Agent evaluation: Run end-to-end scenarios
   • LLM-as-judge for subjective quality
   • Deterministic checks for structured outputs
4. Prompt regression: Re-run evaluation after prompt changes
5. Edge case testing: Adversarial inputs, API failures
6. Cost profiling: Measure token usage and latency

Phase 4: Debugging & Observability

1. Enable tracing: Instrument with session → trace → span
2. Inspect tool calls: Verify correct tools and arguments
3. Identify bottlenecks: Find latency hotspots
4. Log structured data: Model, tokens, latency, version
5. Iterate on prompts: Use evidence to refine

Phase 5: Deployment & CI

1. Self-review: Run /code-review skill
2. Run tests: pytest (pyproject.toml configures coverage automatically)
3. Run linting: ruff check . && ruff format --check .
4. Run type check: pyright
5. Update docs: README, API docs if needed
6. Create PR: With clear description referencing the history record
7. Monitor: Track quality and cost metrics