/feature
You are orchestrating the sigcli agent team to deliver a feature end-to-end. The user's request: $ARGUMENTS
$ npx -y skills add sigcli/sigcli --agent claude-codeHow it fires
How this command gets triggered: by you, by Claude, or both.
- Fires itselfClaude auto-loads it when your prompt matches the work.
- You can call itInvoke it directly when you want it.
- Slash command
/feature
Context preview
What this command does when you run it.
You are orchestrating the sigcli agent team to deliver a feature end-to-end. The user's request: $ARGUMENTS
Command definition
feature.md/feature — End-to-End Feature Delivery
You are orchestrating the sigcli agent team to deliver a feature end-to-end. The user's request: $ARGUMENTS
Workflow
Execute these phases in order. Do NOT skip phases. Present results to the user between phases.
Phase 1: Architecture & Design
Spawn the **architect** agent with this prompt: > Analyze the sigcli codebase and design an implementation plan for: $ARGUMENTS > > Read the relevant interfaces from src/core/interfaces/, study reference implementations, identify reusable code, and produce a concrete plan with: files to create/modify, interfaces to implement, patterns to follow, wiring steps, and test plan.
Present the architect's plan to the user. **Wait for user approval before proceeding.**
Phase 2: Implementation
Spawn the **dev** agent with this prompt: > Implement the following plan for the sigcli project: > [Include the architect's approved plan here] > > Follow all project conventions: Result<T,E> pattern, .js import extensions, Factory pattern for strategies, three-class pattern for adapters, register*Handler for MCP tools. Wire new components in server.ts, handlers/index.ts, and index.ts as needed. Run npm run build when done.
Phase 3: Testing
Spawn the **tester** agent with this prompt: > Write unit tests for the code just implemented: > [Describe what was implemented and which files were created/modified] > > Follow existing test patterns in tests/unit/. Use vitest, MemoryStorage for isolation, isOk/isErr assertions. Cover happy paths, error paths, and edge cases. Run npm test to verify all tests pass.
Phase 4: Review
Spawn the **reviewer** agent with this prompt: > Review the complete changeset for: > [Describe what was implemented] > > Check: type safety (run tsc --noEmit), Result pattern compliance, ESM imports with .js extensions, pattern compliance (Factory/three-class/registerHandler), wiring (server.ts, handlers/index.ts, index.ts), test coverage (exist, cover error paths). Run npm test. Report structured findings.
Phase 5: Report
Summarize the full delivery to the user:
- What was built
- Files created/modified
- Tests added
- Review findings (if any issues need attention)
Read more
/feature — End-to-End Feature Delivery
You are orchestrating the sigcli agent team to deliver a feature end-to-end. The user's request: $ARGUMENTS
Workflow
Execute these phases in order. Do NOT skip phases. Present results to the user between phases.
Phase 1: Architecture & Design
Spawn the **architect** agent with this prompt: > Analyze the sigcli codebase and design an implementation plan for: $ARGUMENTS > > Read the relevant interfaces from src/core/interfaces/, study reference implementations, identify reusable code, and produce a concrete plan with: files to create/modify, interfaces to implement, patterns to follow, wiring steps, and test plan.
Present the architect's plan to the user. **Wait for user approval before proceeding.**
Phase 2: Implementation
Spawn the **dev** agent with this prompt: > Implement the following plan for the sigcli project: > [Include the architect's approved plan here] > > Follow all project conventions: Result<T,E> pattern, .js import extensions, Factory pattern for strategies, three-class pattern for adapters, register*Handler for MCP tools. Wire new components in server.ts, handlers/index.ts, and index.ts as needed. Run npm run build when done.
Phase 3: Testing
Spawn the **tester** agent with this prompt: > Write unit tests for the code just implemented: > [Describe what was implemented and which files were created/modified] > > Follow existing test patterns in tests/unit/. Use vitest, MemoryStorage for isolation, isOk/isErr assertions. Cover happy paths, error paths, and edge cases. Run npm test to verify all tests pass.
Phase 4: Review
Spawn the **reviewer** agent with this prompt: > Review the complete changeset for: > [Describe what was implemented] > > Check: type safety (run tsc --noEmit), Result pattern compliance, ESM imports with .js extensions, pattern compliance (Factory/three-class/registerHandler), wiring (server.ts, handlers/index.ts, index.ts), test coverage (exist, cover error paths). Run npm test. Report structured findings.
Phase 5: Report
Summarize the full delivery to the user:
- What was built
- Files created/modified
- Tests added
- Review findings (if any issues need attention)
Sign in your way. AI works on your behalf. AI agents need access to your work systems — Jira, wikis, calendars, internal APIs. But passing credentials through shell history, environment variables, and agent context windows is a security nightmare.

