Parallel coding agents with tmux and Markdown specs
8 hours ago
- #productivity
- #coding-agents
- #tmux
- The author uses tmux, Markdown files, bash aliases, and slash commands to manage parallel coding agents.
- Agents are categorized into roles: Planner (designs specs), Worker (implements specs), and PM (backlog grooming).
- Feature Designs (FDs) are Markdown files detailing problems, solutions, implementation plans, and verification steps.
- FDs go through 8 stages from 'Planned' to 'Closed', tracked in an index file.
- Six slash commands manage the FD lifecycle: /fd-new, /fd-status, /fd-explore, /fd-deep, /fd-verify, /fd-close.
- The /fd-init command bootstraps the FD system in any repository.
- A typical FD includes problem statements, solutions, files to modify, and verification steps.
- The system supports 4-8 parallel agents, beyond which decision quality suffers.
- The author uses inline annotations (%% prefix) for precise feedback during FD refinement.
- Complex problems are tackled with /fd-deep, launching 4 parallel agents to explore different angles.
- Verification is streamlined with /fd-verify, which commits code and executes verification plans.
- The dev guide (docs/dev_guide/) provides coding standards and best practices.
- The author uses tmux for window management, with custom shortcuts and color-coded idle notifications.
- Challenges include cognitive load, context window limits, and deny list anxiety for destructive commands.
- The system creates a traceable decision history with 300+ FDs in the original project.