devops-bot: Instructions file for Codex

DevOps Bot

Chat-driven AI coding agent. Users communicate via IM group chat (Feishu or Slack), a two-layer AI system classifies intent and executes code changes automatically. Supports multiple AI providers (Anthropic, OpenAI, OpenAI-compatible).

Commands

Always run pnpm check before committing.

Critical Conventions

• ESM + NodeNext: All imports must use .js extension (e.g., import { foo } from './bar.js')
• Node.js builtins: Use node: protocol (e.g., import { readFile } from 'node:fs/promises')
• Biome for formatting and linting — not ESLint, not Prettier
• Single quotes, no semicolons, trailing commas (enforced by Biome)
• pnpm as package manager — do not use npm or yarn
• No new dependencies without justification; prefer Node.js builtins

Project Context

• Supports multi-project mode — projects added via chat (add_project intent), auto-cloned to ~/.devops-bot/repos/
• Supports workspace mode — register a workspace meta-repo (add_workspace intent) with workspace.json manifest; dispatcher AI sees all sub-projects and clones on demand via targetGitUrl
• Falls back to single-project mode if TARGET_PROJECT_PATH is set
• IM bot (Feishu or Slack) is the primary user interface — there is no web frontend
• Two AI layers: fast model (dispatcher) and powerful model (task executor) — provider-agnostic
• Three-tier task execution: execute_task (low risk, immediate), propose_task (medium risk, needs approval via Issue AI synthesis), create_issue (high risk, discussion only)
• PR Review: AI-powered code review via review_pr intent, self-review after task completion (with auto-fix loop — up to 2 rounds), or polling/webhook triggers; Memory namespace isolation (task vs review) with selective cross-injection; full PR discussion context (issue comments + review summaries) injected into review and fix prompts
• Issue AI: Independent AI layer that reads full issue context (body + comments) and synthesizes actionable tasks; also scans external issues with configured labels; in workspace mode, uses two-phase cross-repo triage (quality gate + routing to sub-projects, with sub-issue creation)
• GitHub App authentication for GitHub operations (PRs, Issues, git push); PAT fallback supported
• Skills stored at workspace level (~/.devops-bot/skills/), shared across all projects
• Memory persists in data/memory/ using SQLite (primary) + JSONL exports (AI browsing); features semantic dedup (LLM-driven ADD/UPDATE/NOOP/DELETE decision cycle), change audit trail (memory_history table), periodic pruning of stale memories, and per-project custom extraction prompts via .devops-bot.json
• Configuration lives in .env.local (never committed)