Qalatra Prompts：面向生产环境的智能代码代理工作流框架

Qalatra Prompts is an open-source AI code agent workflow framework designed for production environments. It uses a 3-layer prompt engineering architecture to enable end-to-end automation from task planning to PR merge and deployment. The framework supports unified management of multi-framework projects (NestJS, Shopify, Electron) and emphasizes clear human-AI collaboration—humans handle key decisions while AI agents execute repetitive tasks. GitHub repo: https://github.com/pirateandfox/qalatra-prompts

Open-sourced by the Pirate and Fox organization, Qalatra Prompts is positioned as a 'standardized workflow prompt library for code agents'. Unlike traditional AI coding tools, it is a multi-agent collaboration framework for managing complex software project automation pipelines. Its design philosophy focuses on layered human-AI governance: humans are responsible for critical decisions (需求 planning, code review, final approval), while AI agents take on execution tasks (coding, testing, PR creation, deployment) to balance quality control and efficiency.

Qalatra Prompts uses a 3-layer architecture:

1. Canonical Layer: The kernel, storing cross-project通用 logic (e.g., pipeline-agent.md for full lifecycle, plan-agent.md for planning workflow). Only truly universal logic is kept here.
2. Per-Repo Config Layer: Each repo has a pipeline-config.md defining project-specific params (framework type, github_slug, auto_merge, deploy command, etc.), enabling adaptation to different tech stacks.
3. Deployment Wrapper Layer: Per-machine CLAUDE.md defining monitored repos, canonical layer references, and rare deployment-level overrides. This ensures reuse while retaining flexibility.

The framework defines a 6-stage task lifecycle:

1. Planning: Plan Agent converts raw需求 to executable plans (read-only, clarifies doubts, outputs plan docs to plans/ directory for review).
2. Execution: Direct task execution (simple tasks) or plan-based execution (complex tasks), integrated with Claude Code sessions.
3. In Flight: Monitors Qalatra (task management), FlightDesk (PR/QA tracking), and Claude sessions every 30 mins; links branches to FlightDesk before PR creation.
4. Review & QA: Monitors SonarCloud/CI status, handles Copilot comments, runs 'intelligence checks' to detect issues.
5. Merged: Auto-merges PR to target branch, runs deployment commands, updates task status to completed.
6. Archived: Cleans up after task completion.

Qalatra Prompts supports various frameworks:

• NestJS/Prisma: Built-in code generation (pnpm db-update, nx build) and database migration (local PostgreSQL container, prisma migrate dev) workflows.
• Shopify: Simplified flow (diff checks, direct PR creation, relies on Shopify pre-push hooks for sync).
• Electron: Focuses on build and test validation (specific flows defined on first use).

The framework has a self-training protocol with 4-layer fix strategy:

1. General Fix: Update canonical layer (e.g., pipeline-agent.md) to affect all projects.
2. Framework Fix: Update framework-specific blocks in the canonical layer.
3. Repo Fix: Modify repo's pipeline-config.md (only affects the repo).
4. Deployment Fix: Adjust deployment layer's CLAUDE.md (only affects the machine). Key principle: Avoid fixing canonical issues in deployment layers to prevent behavior drift across instances.

Qalatra Prompts has been used in real projects: BizToBiz (B2B platform), FlightDesk (PR tracking), Muzebook (music app), TMI Shopify3.0 (e-commerce theme), Moceanic AI (AI project). Key insights: Treat prompts as version-controlled code assets; use layered governance to balance generality and specificity; clear human-AI boundaries; multi-system orchestration. Future: Agent orchestration frameworks like Qalatra will become essential in developer toolchains as LLM capabilities grow. Related resources: Core docs (pipeline-agent.md, pipeline-architecture.md, plan-agent.md); supported frameworks (NestJS/Prisma, Shopify, Electron); integrated systems (Qalatra, FlightDesk, Claude Code, GitHub, Notion, Linear).