AI coding agent for the terminal

Fork of badlogic/pi-mono by @mariozechner

---

Installation

Via Bun (recommended)

Requires Bun runtime:

bun install -g @oh-my-pi/pi-coding-agent

Via installer script

Linux / macOS:

curl -fsSL https://raw.githubusercontent.com/can1357/oh-my-pi/main/scripts/install.sh | sh

Windows (PowerShell):

irm https://raw.githubusercontent.com/can1357/oh-my-pi/main/scripts/install.ps1 | iex

By default, the installer uses bun if available, otherwise downloads the prebuilt binary.

Options:

• --source / -Source: Install via bun (installs bun first if needed)
• --binary / -Binary: Always use prebuilt binary

# Force bun installation
curl -fsSL .../install.sh | sh -s -- --source

Manual download

Download binaries directly from GitHub Releases.

---

+ LSP Integration (Language Server Protocol)

Full IDE-like code intelligence with automatic formatting and diagnostics:

• Format-on-write: Auto-format code using the language server's formatter (rustfmt, gofmt, prettier, etc.)
• Diagnostics on write/edit: Immediate feedback on syntax errors and type issues after every file change
• Workspace diagnostics: Check entire project for errors (lsp action=workspace_diagnostics)
• 40+ language configs: Out-of-the-box support for Rust, Go, Python, TypeScript, Java, Kotlin, Scala, Haskell, OCaml, Elixir, Ruby, PHP, C#, Lua, Nix, and many more
• Local binary resolution: Auto-discovers project-local LSP servers in node_modules/.bin/, .venv/bin/, etc.
• Hover docs, symbol references, code actions, workspace-wide symbol search

+ Time Traveling Streamed Rules (TTSR)

Zero context-use rules that inject themselves only when needed:

• Pattern-triggered injection: Rules define regex triggers that watch the model's output stream
• Just-in-time activation: When a pattern matches, the stream aborts, the rule injects as a system reminder, and the request retries
• Zero upfront cost: TTSR rules consume no context until they're actually relevant
• One-shot per session: Each rule only triggers once, preventing loops
• Define via ttsrTrigger field in rule files (regex pattern)

Example: A "don't use deprecated API" rule only activates when the model starts writing deprecated code, saving context for sessions that never touch that API.

+ Interactive Code Review

Structured code review with priority-based findings:

• /review command: Interactive mode selection (branch comparison, uncommitted changes, commit review)
• Structured findings: report_finding tool with priority levels (P0-P3: critical → nit)
• Verdict rendering: submit_review aggregates findings into approve/request-changes/comment
• Combined result tree showing verdict and all findings

+ Task Tool (Subagent System)

Parallel execution framework with specialized agents and real-time streaming:

• 5 bundled agents: explore, plan, browser, task, reviewer
• Parallel exploration: Reviewer agent can spawn explore agents for large codebase analysis
• Real-time artifact streaming: Task outputs stream as they're created, not just at completion
• Output tool: Read full agent outputs by ID when truncated previews aren't sufficient
• User-level (~/.omp/agent/agents/) and project-level (.omp/agents/) custom agents
• Concurrency-limited batch execution with progress tracking

+ Model Roles

Configure different models for different purposes with automatic discovery:

• Three roles: default (main model), smol (fast/cheap), slow (comprehensive reasoning)
• Auto-discovery: Smol finds haiku → flash → mini; Slow finds codex → gpt → opus → pro
• Role-based selection: Task tool agents can use model: pi/smol for cost-effective exploration
• CLI args (--smol, --slow) and env vars (OMP_SMOL_MODEL, OMP_SLOW_MODEL)
• Configure via /model selector with keybindings (Enter=default, S=smol, L=slow)

+ Ask Tool (Interactive Questioning)

Structured user interaction with typed options:

• Multiple choice questions: Present options with descriptions for user selection
• Multi-select support: Allow multiple answers when choices aren't mutually exclusive

+ Custom TypeScript Slash Commands

Programmable commands with full API access:

• Create at ~/.omp/agent/commands/[name]/index.ts or .omp/commands/[name]/index.ts
• Export factory returning { name, description, execute(args, ctx) }
• Full access to HookCommandContext for UI dialogs, session control, shell execution
• Return string to send as LLM prompt, or void for fire-and-forget actions
• Also loads from Claude Code directories (~/.claude/commands/, .claude/commands/)

+ Universal Config Discovery

Unified capability-based discovery that loads configuration from 8 AI coding tools:

• Multi-tool support: Claude Code, Cursor, Windsurf, Gemini, Codex, Cline, GitHub Copilot, VS Code
• Discovers everything: MCP servers, rules, skills, hooks, tools, slash commands, prompts, context files
• Native format support: Cursor MDC frontmatter, Windsurf rules, Cline .clinerules, Copilot applyTo globs, Gemini system.md, Codex AGENTS.md
• Provider attribution: See which tool contributed each configuration item
• Discovery settings: Enable/disable individual providers via /config interactive tab
• Priority ordering: Multi-path resolution across .omp, .pi, and .claude directories

+ MCP & Plugin System

Full Model Context Protocol support with external tool integration:

• Stdio and HTTP transports for connecting to MCP servers
• Plugin CLI (omp plugin install/enable/configure/doctor)
• Hot-loadable plugins from ~/.omp/plugins/ with npm/bun integration
• Automatic Exa MCP server filtering with API key extraction

+ Web Search & Fetch

Multi-provider search and full-page scraping:

• Anthropic, Perplexity, and Exa search integration with caching
• HTML-to-markdown conversion with link preservation
• JavaScript rendering support, image handling

+ TUI Overhaul

Modern terminal interface with smart session management:

• Auto session titles: Sessions automatically titled based on first message using smol model
• Welcome screen: Logo, tips, recent sessions with selection
• Powerline footer: Model, cwd, git branch/status, token usage, context %
• LSP status: Shows which language servers are active and ready
• Hotkeys: ? displays shortcuts when editor empty
• Emergency terminal restore: Crash handlers prevent terminal corruption

+ Edit Fuzzy Matching

Handles whitespace and indentation variance automatically:

• High-confidence fuzzy matching for oldText in edit operations
• Fixes the #1 pain point: edits failing due to invisible whitespace differences
• Configurable via edit.fuzzyMatch setting (enabled by default)

... and many more

• Git context: System prompt includes branch, status, recent commits
• Bun runtime: Native TypeScript execution, faster startup, all packages migrated
• Centralized file logging: Debug logs with daily rotation to ~/.omp/logs/
• Clipboard export: /export --copy copies session as formatted text
• Bash interceptor: Optionally block shell commands that have dedicated tools
• Hidden tools: Custom tools can be excluded from default list unless explicitly requested
• @file auto-read: Type @path/to/file in prompts to inject file contents inline
• Additional tools: AST (structural code analysis), Replace (find & replace across files)

---

Packages

Package	Description
@oh-my-pi/pi-ai	Unified multi-provider LLM API (OpenAI, Anthropic, Google, etc.)
@oh-my-pi/pi-agent-core	Agent runtime with tool calling and state management
@oh-my-pi/pi-coding-agent	Interactive coding agent CLI
@oh-my-pi/pi-mom	Slack bot that delegates messages to the pi coding agent
@oh-my-pi/pi-tui	Terminal UI library with differential rendering
@oh-my-pi/pi-web-ui	Web components for AI chat interfaces

---

Development

Setup

bun run install:dev   # Install deps and link all packages
bun run build         # Build all packages
bun run check         # Lint, format, and type check

Note: bun run check requires bun run build first. The web-ui package uses tsc which needs compiled .d.ts files from dependencies.

Watch Mode

bun run dev

Then run directly:

cd packages/coding-agent && bunx tsx src/cli.ts

CI

GitHub Actions runs on push to main and on pull requests. The workflow runs bun run check and bun test for each package in parallel.

Do not add LLM API keys as secrets. Tests requiring LLM access use describe.skipIf() and run locally.

---

Versioning

All packages use lockstep versioning:

bun run version:patch    # 0.7.5 -> 0.7.6
bun run version:minor    # 0.7.5 -> 0.8.0
bun run version:major    # 0.7.5 -> 1.0.0

Never manually edit version numbers.

---

Publishing

bun run release:patch    # Bug fixes
bun run release:minor    # New features
bun run release:major    # Breaking changes

Requires an npm token with "Bypass 2FA on publish" enabled.

---

License

MIT - Original work copyright Mario Zechner