parallel-dev-cycle_skill

Overview

This skill implements a multi-agent parallel development cycle that runs four specialized worker agents inline in the main flow: Requirements Analysis (RA), Exploration & Planning (EP), Code Development (CD), and Validation & Archival Summary (VAS). It orchestrates phases, state, and iteration without spawning a separate orchestrator, preserving an auditable session directory with main documents and append-only logs. The design favors reproducible iterations, shared discovery via an NDJSON board, and automatic archival of previous versions.

How this skill works

On trigger (parallel-dev-cycle) the main flow initializes or resumes a cycle, creates the session state and progress directories, and launches the four agents in parallel. Each agent rewrites one main document per iteration and appends to its change logs; all agents read and append to a shared discoveries.ndjson to coordinate findings. After agents complete, the main flow aggregates results, detects issues, and either loops for another iteration with feedback or finalizes a summary report and updates cycle state.

When to use it

• When you need structured, repeatable development iterations with clear audit trails.
• For tasks requiring simultaneous requirement refinement, exploration, implementation and validation.
• When you want lightweight multi-agent parallelism without a separate orchestrator process.
• To run automated cycles with optional --auto mode for unattended convergence.
• When you must preserve discovery history and versioned main documents for compliance or review.

Best practices

• Keep tasks concise or provide a prep-package.json to guide per-iteration behavior and convergence rules.
• Treat each agent's main document as authoritative for the current iteration; use changes.log for history and recovery.
• Append discoveries immediately (no batching) and deduplicate by the defined keys to avoid redundant work.
• Use --auto only when comfortable with default decisions and iteration limits to avoid runaway loops.
• Limit parallel agents to the supported maximum (4) so dependency blocking behavior remains predictable.

Example use cases

• Bootstrapping a TypeScript service: RA detects tech stack, EP proposes architecture, CD implements features, VAS runs tests and summarizes results.
• Iterative bugfix cycles: CD writes debug hypotheses to debug-log.ndjson while RA/EP refine requirements and impact.
• Refactoring sprint: EP discovers patterns and integration points while CD implements and VAS validates regression coverage.
• Onboarding contribution: run a cycle to generate plan.json, exploration.md and a guided implementation outline for new contributors.