pi-mono agent framework reference (github.com/badlogic/pi-mono). TRIGGER when: writing agent code using pi-ai/pi-agent-core/pi-coding-agent packages, defining tools with TypeBox schemas, implementing TUI or Web UI over an agent core, building extensions that hook into agent lifecycle, working with session/compaction/retry logic, implementing LLM provider abstractions, or any code that imports from @mariozechner/* packages.
설치
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
pi-mono agent framework reference (github.com/badlogic/pi-mono). TRIGGER when: writing agent code using pi-ai/pi-agent-core/pi-coding-agent packages, defining tools with TypeBox schemas, implementing TUI or Web UI over an agent core, building extensions that hook into agent lifecycle, working with session/compaction/retry logic, implementing LLM provider abstractions, or any code that imports from @mariozechner/* packages.
Strict layer isolation: each layer has ZERO knowledge of layers above. pi-agent-core does NOT contain MCP, skills, or sessions — pi-mono uses a custom Extension system instead of MCP.
Gotchas
These are the most common pitfalls. Read before writing any pi-mono code.
TypeBox is mandatory for tool schemas — @sinclair/typebox for JSON Schema + TypeScript inference. No raw JSON Schema anywhere.
ToolCall.arguments is Record<string, any>, not string — arguments are parsed objects, not raw JSON strings.
ThinkingContent field is thinking, not text — and redacted is optional (redacted?: boolean).
ThinkingLevel has no "off" value — only "minimal" | "low" | "medium" | "high" | "xhigh".
convertToLlm and transformContext are async — convertToLlm returns Message[] | Promise<Message[]>, transformContext takes (messages, signal?) => Promise<AgentMessage[]>.
beforeToolCall/afterToolCall are async with signal — signature is (ctx, signal?) => Promise<Result | undefined>. Can return undefined.
Tool preparation is always sequential even in parallel mode — beforeToolCall hooks run one at a time. Only execution is concurrent.
TUI render(width) lines MUST NOT exceed width — overflow crashes the TUI.
Web UI connects to Agent directly, not AgentSession — it uses IndexedDB for its own storage. TUI uses the higher-level AgentSession.
Default model is Gemini Flash Lite, not an Anthropic model.
No persistence in agent-core — session management is entirely in pi-coding-agent.
Cross-provider normalization is critical — tool call IDs, thinking blocks, and orphaned tool calls must be handled when switching providers. Use transform-messages.ts.
OAuth tokens trigger Claude Code impersonation — sk-ant-oat prefix activates special headers, system prompt prefix, and tool name translation.
Compaction uses characters / 4 heuristic — not actual tokenizer. Cut point is always at user/assistant boundaries, never mid-turn.
Extension emit methods have DIFFERENT chaining semantics — emitToolResult chains results, emitContext deep-clones via structuredClone, emitToolCall short-circuits on block. Don't assume one generic dispatch.
Extensions can't call action methods during loading — runtime stubs throw. Provider registrations are queued and flushed after binding.
Register tool with same name to override built-in — no special API, just pi.registerTool({ name: "read", ... }).
