Imported from hongyime/theprawnfeeds (AGENTS.md). Install upstream with npx skills add hongyime/theprawnfeeds. Copyright stays with the author.

AI Agents Configuration

This file defines how AI coding agents (Cursor, Antigravity, Claude Code, GitHub Copilot, etc.) should behave across all repositories in this workspace.

Cross-Session State (Read This First)

Before doing anything else, read .agents/STATE.md if it exists. It is the handover from whoever worked here last, possibly from a different tool or machine.

This workspace is worked on by multiple agents: Gemini CLI, Antigravity, Mistral CLI, Kiro, Codex, Pi, Hermes, opencode, Claude Code, Cursor, and others. Sessions can end abruptly when a free-tier limit is hit. No harness can read every other harness's private session store, so durable state lives in plain files that all of them can read.

The state directory is:

.agents/
  STATE.md      current task, progress, next steps
  JOURNAL.md    append-only decisions and rationale
  handoffs/     detailed handoff documents

.agents/ is committed deliberately so state reaches the other machines. Never write secrets or personal details there. These files are permanent and world-readable in public repositories.

Treat .agents/ as shared project state, not personal scratch space. In team repos, write only what another teammate or agent needs to resume the work: current task, decisions, blockers, linked issues or PRs, and safe next steps. Do not write private notes, credentials, customer data, local machine paths that should stay private, or anything that would be unsafe in a public repository.

Your obligations:

1. On start, read .agents/STATE.md if it exists.
2. As you work, update .agents/STATE.md after each meaningful step.
3. For durable decisions, append one dated line to .agents/JOURNAL.md.
4. Before a long or risky stretch, write a handoff in .agents/handoffs/.

Keep STATE.md short and current. Put durable decisions in JOURNAL.md. Put detailed resume notes in timestamped files under .agents/handoffs/ so multiple people do not overwrite one another.

Memory tools such as cognee or cavemem are optional local aids. Trust .agents/STATE.md and git log over an empty memory-tool result.

Agent Types and Roles

1. Primary Coding Agent (Cursor/Claude Code)

Purpose: Main development assistant for code changes Capabilities: Full codebase access, file editing, terminal commands Behavior: Follows all conventions in this file

2. Review Agent (GitHub PR Review)

Purpose: Automated code review on pull requests Capabilities: Read-only access to PR changes Behavior: Strict adherence to coding standards

3. Documentation Agent

Purpose: Maintains and updates documentation Capabilities: Can modify markdown files Behavior: Never creates new docs without explicit request

4. Security Agent (TruffleHog, CodeQL)

Purpose: Security scanning and vulnerability detection Capabilities: Full codebase scan Behavior: Blocks PRs on security issues

Universal Rules (Apply to All Agents)

1. File Operations

• NEVER create new markdown files without explicit user request
• ALWAYS update existing documentation when possible
• NEVER delete files without confirmation
• ALWAYS preserve file history and git history

2. Code Style

• Follow language-specific conventions (see skill files)
• Use consistent naming patterns
• Keep functions small and focused
• Add type hints for Python, TypeScript interfaces for JS

3. Git Conventions

• Commit messages: type: description
• Types: feat, fix, docs, refactor, test, chore, perf, style
• Keep commits atomic and focused
• Never force push to main branches

4. Communication

• Be concise and direct
• Explain complex changes briefly
• Admit uncertainty when present
• Ask for clarification when requirements are ambiguous

Repository-Specific Overrides

source-repo-code (Template/Source)

• Purpose: Source of truth for shared configurations
• Special Rules: Changes here should be synced to all repos

Skill System

Do not assume project-level skills are present in target repositories. The current sync intentionally removes skills/, skills-lock.json, and docs/ from target repos, and .claude/ remains local-only because it can contain session databases and credentials.

If a harness has local skills installed, it may use them, but shared repo instructions must live in AGENTS.md and shared handoff state must live under .agents/.

MCP Configuration

MCP support varies by harness. Use local MCP configuration if present, but do not assume one shared MCP config is read by every agent.

Syncing Strategy

Source of Truth

• Primary: sourcerepo 仓库作为唯一真源（Skills、MCP、通用配置、仓库设置）
• Secondary: 目标仓库保留仓库特定覆盖（需在 sync 后自行维护）

Sync Workflows

• 仓库设置与通用配置：[sync-repo-settings.yml](.github/workflows/sync-repo-settings.yml) 负责传播 GitHub Actions、Dependabot、labels、AGENTS.md 等

Sync Process

1. 在 sourcerepo 中修改通用配置
2. 推送到 main 分支自动触发相应 workflow
3. workflow 会遍历所有非 archive/fork 仓库，复制变更并提交/开 PR
4. 定时任务每日/每日执行，覆盖未来新仓库

New Repository Setup

当创建新仓库时：

1. 将 sourcerepo 中的配置同步过去（由定时任务或手动触发完成）
2. 如需仓库特定覆盖，同步后手动维护
3. Read AGENTS.md, then read .agents/STATE.md if present

Maintenance

Quarterly Reviews

• Review and update skill files
• Check for outdated action versions
• Verify agent behavior consistency
• Update this configuration as needed

Automated Monitoring

• Use Dependabot for dependency updates
• Use TruffleHog for secret scanning
• Use custom scripts to detect outdated GitHub Actions

Troubleshooting

Agent Not Following Conventions

1. 检查仓库特定覆盖在 AGENTS.md 中
2. 检查 .agents/STATE.md 中是否有最新任务上下文
3. 审阅近期约定变更
4. 从源重新同步配置

Sync Failures

1. 使用相应 workflow 的 workflow_dispatch 手动触发以观察日志
2. 检查目标仓库中的 git 冲突
3. 验证文件权限
4. 查看同步日志中的具体错误

References

• GitHub Skills Documentation
• Cursor AI Documentation
• Claude Code Documentation
• Dependabot Configuration