feat: add "Anthropic — via Claude CLI" as a first-class provider

Replaces the stray useClaudeCli settings checkbox + onboarding question
with a proper provider-card UX. The card lives next to OAuth + API-key
cards in onboarding and settings, with Enable/Disable + Test actions.

Backend:
 - Vendors rchern/pi-claude-cli@0.3.1 as packages/pi-claude-cli
   (MIT, attribution in UPSTREAM.md). Lets us bump peer-dep on
   pi-coding-agent in lockstep with Fusion (upstream pinned ^0.52.0
   vs ours ^0.62.0) and fix bugs without waiting on upstream.
 - Adds @fusion/pi-claude-cli as a workspace dep of @runfusion/fusion
   so users don't have to `npm install -g pi-claude-cli` manually.
 - serve/daemon/dashboard conditionally load the extension via
   discoverAndLoadExtensions() when GlobalSettings.useClaudeCli is on;
   no side-effects on user ~/.fusion/agent/settings.json.
 - New GET /api/providers/claude-cli/status: claude --version probe
   + toggle state + cached extension resolution.
 - New POST /api/auth/claude-cli: flips useClaudeCli, refuses if the
   claude binary is missing, fires the existing skill-backfill hook.
 - /api/auth/status now injects a synthetic {id:"claude-cli", type:"cli"}
   provider entry so onboarding + settings see a consistent list.

Frontend:
 - New ClaudeCliProviderCard component shared between ModelOnboardingModal
   and SettingsModal's Authentication section.
 - New AuthProvider.type = "cli" variant.
 - Removed the old "Route AI calls through the Claude CLI" checkbox from
   Global Models settings and the opt-in step from the onboarding wizard.
 - ProviderIcon gets a composite Anthropic-mark-plus-terminal glyph for
   the claude-cli provider id.

Tests:
 - 8 unit tests for extension resolution (@fusion/pi-claude-cli is
   workspace-linked so these run in-tree).
 - 2 unit tests for the binary probe.
 - Existing /auth/status tests filter out the new synthetic entry so
   they keep asserting structural OAuth/API-key behavior in isolation.
 - The vendored package's own 296 tests still pass unchanged.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

packages/pi-claude-cli/src/control-handler.ts

@@ -0,0 +1,68 @@
+/**
+ * Control protocol handler for Claude CLI stream-json communication.
+ *
+ * Processes control_request messages from Claude CLI stdout and writes
+ * control_response messages to stdin.
+ *
+ * - Custom MCP tools (mcp__custom-tools__*): DENIED — pi executes these
+ * - Everything else (user MCP tools, internal tools): ALLOWED — Claude handles
+ */
+
+import type { ClaudeControlRequest } from "./types";
+import { CUSTOM_TOOLS_MCP_PREFIX } from "./tool-mapping.js";
+
+export const TOOL_EXECUTION_DENIED_MESSAGE =
+  "Tool execution is unavailable in this environment.";
+
+/** Prefix for MCP (Model Context Protocol) tool names. */
+export const MCP_PREFIX = "mcp__";
+
+interface ControlResponse {
+  type: "control_response";
+  request_id: string;
+  response: {
+    subtype: "success";
+    response: {
+      behavior: "allow" | "deny";
+      message?: string;
+    };
+  };
+}
+
+/**
+ * Handle a control_request from the Claude CLI.
+ *
+ * Denies custom MCP tools (mcp__custom-tools__*) so pi can execute them.
+ * Allows everything else (user MCP tools, internal Claude tools).
+ *
+ * @returns true if the tool was allowed, false if denied
+ */
+export function handleControlRequest(
+  msg: ClaudeControlRequest,
+  stdin: NodeJS.WritableStream,
+): boolean {
+  if (!msg.request_id || !msg.request) {
+    console.error(
+      "[pi-claude-cli] Malformed control_request: missing request_id or request object",
+      msg,
+    );
+    return false;
+  }
+
+  const toolName = msg.request?.tool_name ?? "";
+  const isCustomTool = toolName.startsWith(CUSTOM_TOOLS_MCP_PREFIX);
+
+  const response: ControlResponse = {
+    type: "control_response",
+    request_id: msg.request_id,
+    response: {
+      subtype: "success",
+      response: isCustomTool
+        ? { behavior: "deny", message: TOOL_EXECUTION_DENIED_MESSAGE }
+        : { behavior: "allow" },
+    },
+  };
+
+  stdin.write(JSON.stringify(response) + "\n");
+  return !isCustomTool;
+}

packages/pi-claude-cli/src/event-bridge.ts

@@ -0,0 +1,385 @@
+import type { ClaudeApiEvent, TrackedContentBlock } from "./types";
+import { calculateCost } from "@mariozechner/pi-ai";
+import type {
+  AssistantMessage,
+  AssistantMessageEventStream,
+  Model,
+  TextContent,
+  ThinkingContent,
+  ToolCall,
+} from "@mariozechner/pi-ai";
+import {
+  mapClaudeToolNameToPi,
+  translateClaudeArgsToPi,
+  isPiKnownClaudeTool,
+} from "./tool-mapping.js";
+
+/**
+ * Extended tracking for tool_use content blocks during streaming.
+ * Stores the Claude tool name for argument translation at block_stop.
+ */
+interface TrackedToolBlock {
+  type: "tool_use";
+  index: number;
+  id: string;
+  name: string; // Already mapped to pi name
+  claudeName: string; // Original Claude name for arg translation
+  arguments: Record<string, unknown>;
+  partialJson: string;
+}
+
+/** Union of tracked block types for the blocks array. */
+type TrackedBlock = TrackedContentBlock | TrackedToolBlock;
+
+/**
+ * The event bridge interface returned by createEventBridge.
+ * handleEvent processes each Claude API streaming event and pushes
+ * the appropriate pi events to the stream.
+ * getOutput returns the accumulated AssistantMessage.
+ */
+export interface EventBridge {
+  handleEvent(event: ClaudeApiEvent): void;
+  getOutput(): AssistantMessage;
+}
+
+/**
+ * Map Claude API stop reasons to pi's stop reason format.
+ */
+function mapStopReason(
+  reason: string | undefined,
+): "stop" | "length" | "toolUse" {
+  switch (reason) {
+    case "tool_use":
+      return "toolUse";
+    case "max_tokens":
+      return "length";
+    case "end_turn":
+    default:
+      return "stop";
+  }
+}
+
+/**
+ * Create an event bridge that translates Claude API streaming events
+ * into pi's AssistantMessageEventStream events.
+ *
+ * The bridge maintains internal state to track content blocks and
+ * accumulate the final AssistantMessage. It handles:
+ * - text content blocks (start/delta/stop -> text_start/text_delta/text_end)
+ * - message lifecycle (message_start for usage, message_delta for stop reason, message_stop for done)
+ * - unsupported block types (tool_use, thinking) with warnings
+ */
+export function createEventBridge(
+  stream: AssistantMessageEventStream,
+  model: Model<any>,
+): EventBridge {
+  // Tracked content blocks indexed by Claude's content_block index
+  const blocks: TrackedBlock[] = [];
+
+  // The accumulated output message
+  const output: AssistantMessage = {
+    role: "assistant" as const,
+    content: [] as (TextContent | ThinkingContent | ToolCall)[],
+    api: "pi-claude-cli",
+    provider: model.provider,
+    model: model.id,
+    usage: {
+      input: 0,
+      output: 0,
+      cacheRead: 0,
+      cacheWrite: 0,
+      totalTokens: 0,
+      cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },
+    },
+    stopReason: "stop" as const,
+    timestamp: Date.now(),
+  };
+
+  let started = false;
+
+  function handleEvent(event: ClaudeApiEvent): void {
+    // Emit start event on first message — tells pi to begin incremental rendering
+    if (!started) {
+      stream.push({ type: "start", partial: output });
+      started = true;
+    }
+
+    switch (event.type) {
+      case "message_start":
+        handleMessageStart(event);
+        break;
+      case "content_block_start":
+        handleContentBlockStart(event);
+        break;
+      case "content_block_delta":
+        handleContentBlockDelta(event);
+        break;
+      case "content_block_stop":
+        handleContentBlockStop(event);
+        break;
+      case "message_delta":
+        handleMessageDelta(event);
+        break;
+      case "message_stop":
+        handleMessageStop();
+        break;
+      // Unknown event types are silently ignored
+    }
+  }
+
+  function handleMessageStart(event: ClaudeApiEvent): void {
+    const usage = event.message?.usage;
+    if (usage) {
+      output.usage.input = usage.input_tokens ?? 0;
+      output.usage.output = usage.output_tokens ?? 0;
+      output.usage.cacheRead = usage.cache_read_input_tokens ?? 0;
+      output.usage.cacheWrite = usage.cache_creation_input_tokens ?? 0;
+      output.usage.totalTokens =
+        output.usage.input +
+        output.usage.output +
+        output.usage.cacheRead +
+        output.usage.cacheWrite;
+      calculateCost(model, output.usage);
+    }
+  }
+
+  function handleContentBlockStart(event: ClaudeApiEvent): void {
+    const blockType = event.content_block?.type;
+
+    if (blockType === "text") {
+      const block: TrackedContentBlock = {
+        type: "text",
+        text: "",
+        index: event.index ?? 0,
+      };
+      blocks.push(block);
+      output.content.push({ type: "text" as const, text: "" });
+
+      stream.push({
+        type: "text_start",
+        contentIndex: output.content.length - 1,
+        partial: output,
+      });
+    } else if (blockType === "thinking") {
+      const block: TrackedContentBlock = {
+        type: "thinking",
+        text: "",
+        index: event.index ?? 0,
+      };
+      blocks.push(block);
+      output.content.push({
+        type: "thinking" as const,
+        thinking: "",
+        thinkingSignature: "",
+      });
+
+      stream.push({
+        type: "thinking_start",
+        contentIndex: output.content.length - 1,
+        partial: output,
+      });
+    } else if (blockType === "tool_use") {
+      const claudeName = event.content_block!.name!;
+
+      // Skip internal Claude Code tools (ToolSearch, Task, Agent, etc.)
+      // that pi cannot execute — only emit pi-known tools
+      if (!isPiKnownClaudeTool(claudeName)) {
+        return;
+      }
+
+      const piName = mapClaudeToolNameToPi(claudeName);
+      const id = event.content_block!.id!;
+
+      const block: TrackedToolBlock = {
+        type: "tool_use",
+        index: event.index ?? 0,
+        id,
+        name: piName,
+        claudeName,
+        arguments: {},
+        partialJson: "",
+      };
+      blocks.push(block);
+      output.content.push({
+        type: "toolCall" as const,
+        id,
+        name: piName,
+        arguments: {},
+      } as ToolCall);
+
+      stream.push({
+        type: "toolcall_start",
+        contentIndex: output.content.length - 1,
+        partial: output,
+      });
+    }
+    // Unknown block types silently ignored
+  }
+
+  function handleContentBlockDelta(event: ClaudeApiEvent): void {
+    const deltaType = event.delta?.type;
+
+    if (deltaType === "text_delta" && event.delta!.text != null) {
+      const idx = blocks.findIndex((b) => b.index === event.index);
+      if (idx === -1) return;
+
+      const block = blocks[idx];
+      if (block.type === "text") {
+        block.text += event.delta!.text;
+        const contentBlock = output.content[idx] as TextContent;
+        contentBlock.text = block.text;
+
+        stream.push({
+          type: "text_delta",
+          contentIndex: idx,
+          delta: event.delta!.text,
+          partial: output,
+        });
+      }
+    } else if (
+      deltaType === "thinking_delta" &&
+      event.delta!.thinking != null
+    ) {
+      const idx = blocks.findIndex((b) => b.index === event.index);
+      if (idx === -1) return;
+
+      const block = blocks[idx];
+      if (block.type === "thinking") {
+        block.text += event.delta!.thinking;
+        const contentBlock = output.content[idx] as ThinkingContent;
+        contentBlock.thinking = block.text;
+
+        stream.push({
+          type: "thinking_delta",
+          contentIndex: idx,
+          delta: event.delta!.thinking,
+          partial: output,
+        });
+      }
+    } else if (
+      deltaType === "input_json_delta" &&
+      event.delta!.partial_json != null
+    ) {
+      const idx = blocks.findIndex((b) => b.index === event.index);
+      if (idx === -1) return;
+
+      const block = blocks[idx];
+      if (block.type === "tool_use") {
+        block.partialJson += event.delta!.partial_json;
+
+        // Try to parse accumulated JSON -- on success update args, on failure keep previous
+        try {
+          block.arguments = JSON.parse(block.partialJson);
+          (output.content[idx] as any).arguments = block.arguments;
+        } catch {
+          // Partial JSON not yet parseable -- keep previous arguments
+        }
+
+        stream.push({
+          type: "toolcall_delta",
+          contentIndex: idx,
+          delta: event.delta!.partial_json,
+          partial: output,
+        });
+      }
+    } else if (
+      deltaType === "signature_delta" &&
+      event.delta!.signature != null
+    ) {
+      // Accumulate signature on the thinking block
+      const idx = blocks.findIndex((b) => b.index === event.index);
+      if (idx === -1) return;
+
+      const block = blocks[idx];
+      if (block.type === "thinking") {
+        const contentBlock = output.content[idx] as ThinkingContent;
+        contentBlock.thinkingSignature =
+          (contentBlock.thinkingSignature || "") + event.delta!.signature;
+      }
+    }
+  }
+
+  function handleContentBlockStop(event: ClaudeApiEvent): void {
+    const idx = blocks.findIndex((b) => b.index === event.index);
+    if (idx === -1) return;
+
+    const block = blocks[idx];
+    // Clean up the tracking index from the block (no longer needed)
+    delete (block as any).index;
+
+    if (block.type === "text") {
+      stream.push({
+        type: "text_end",
+        contentIndex: idx,
+        content: block.text,
+        partial: output,
+      });
+    } else if (block.type === "thinking") {
+      stream.push({
+        type: "thinking_end",
+        contentIndex: idx,
+        content: block.text,
+        partial: output,
+      });
+    } else if (block.type === "tool_use") {
+      // Final JSON parse with fallback to raw string
+      let finalArgs: Record<string, unknown> | string;
+      try {
+        const parsed = JSON.parse(block.partialJson);
+        finalArgs = translateClaudeArgsToPi(block.claudeName, parsed);
+      } catch {
+        finalArgs = block.partialJson;
+      }
+
+      // Update output.content with final arguments
+      const contentBlock = output.content[idx] as ToolCall;
+      (contentBlock as any).arguments = finalArgs;
+
+      // ToolCall.arguments is typed as Record<string, any> in pi-ai, but we
+      // intentionally emit a raw string when JSON parse fails completely.
+      // Pi handles string arguments gracefully at runtime.
+      const toolCall = {
+        type: "toolCall" as const,
+        id: block.id,
+        name: block.name,
+        arguments: finalArgs,
+      } as ToolCall;
+
+      stream.push({
+        type: "toolcall_end",
+        contentIndex: idx,
+        toolCall,
+        partial: output,
+      });
+    }
+  }
+
+  function handleMessageDelta(event: ClaudeApiEvent): void {
+    if (event.delta?.stop_reason) {
+      output.stopReason = mapStopReason(event.delta.stop_reason);
+    }
+
+    const usage = event.usage;
+    if (usage) {
+      if (usage.input_tokens != null) output.usage.input = usage.input_tokens;
+      if (usage.output_tokens != null)
+        output.usage.output = usage.output_tokens;
+      output.usage.totalTokens =
+        output.usage.input +
+        output.usage.output +
+        output.usage.cacheRead +
+        output.usage.cacheWrite;
+      calculateCost(model, output.usage);
+    }
+  }
+
+  function handleMessageStop(): void {
+    // No-op: done event is pushed by the provider after readline closes.
+    // Pushing done here (synchronously) prevents pi from executing tools.
+  }
+
+  return {
+    handleEvent,
+    getOutput: () => output,
+  };
+}

packages/pi-claude-cli/src/mcp-config.ts

@@ -0,0 +1,93 @@
+/**
+ * Custom tool discovery and MCP config file generation.
+ *
+ * Discovers non-built-in tools from pi, writes their schemas to a temp file,
+ * and generates an MCP config that points to the schema-only MCP server.
+ */
+
+import { writeFileSync } from "node:fs";
+import { join, dirname } from "node:path";
+import { tmpdir } from "node:os";
+import { fileURLToPath } from "node:url";
+
+/** The 6 built-in tools that pi handles natively (match pi tool names). */
+const BUILT_IN_TOOL_NAMES = new Set([
+  "read",
+  "write",
+  "edit",
+  "bash",
+  "grep",
+  "find",
+]);
+
+/** A custom tool definition with MCP-compatible schema. */
+export interface McpToolDef {
+  name: string;
+  description: string;
+  inputSchema: Record<string, unknown>;
+}
+
+/**
+ * Get custom tool definitions from pi, filtering out built-in tools.
+ *
+ * @param pi - The pi ExtensionAPI instance
+ * @returns Array of custom tool definitions (empty if all tools are built-in)
+ */
+export function getCustomToolDefs(pi: any): McpToolDef[] {
+  const allTools = pi.getAllTools();
+
+  if (!Array.isArray(allTools)) {
+    return [];
+  }
+
+  return allTools
+    .filter((tool: any) => !BUILT_IN_TOOL_NAMES.has(tool.name))
+    .map((tool: any) => ({
+      name: tool.name,
+      description: tool.description,
+      inputSchema: tool.parameters,
+    }));
+}
+
+/**
+ * Write MCP config and tool schemas to temp files.
+ *
+ * Creates two temp files:
+ * 1. Schema file: JSON array of tool definitions
+ * 2. Config file: MCP config pointing to the schema-only server
+ *
+ * @param toolDefs - Array of custom tool definitions
+ * @returns Path to the MCP config file
+ */
+export function writeMcpConfig(toolDefs: McpToolDef[]): string {
+  // Write tool schemas to temp file
+  const schemaFilePath = join(
+    tmpdir(),
+    `pi-claude-mcp-schemas-${process.pid}.json`,
+  );
+  writeFileSync(schemaFilePath, JSON.stringify(toolDefs));
+
+  // Resolve path to the schema server .cjs file (sibling of this module)
+  const __filename = fileURLToPath(import.meta.url);
+  const __dirname = dirname(__filename);
+  const serverPath = join(__dirname, "mcp-schema-server.cjs");
+
+  // Build MCP config
+  const config = {
+    mcpServers: {
+      "custom-tools": {
+        command: "node",
+        args: [serverPath, schemaFilePath],
+      },
+    },
+  };
+
+  // Write config to temp file
+  const configFilePath = join(
+    tmpdir(),
+    `pi-claude-mcp-config-${process.pid}.json`,
+  );
+  writeFileSync(configFilePath, JSON.stringify(config));
+
+  return configFilePath;
+}

packages/pi-claude-cli/src/mcp-schema-server.cjs

@@ -0,0 +1,49 @@
+#!/usr/bin/env node
+// Schema-only MCP server. Reads tool schemas from a JSON file.
+// Only implements initialize + tools/list. tools/call is never reached
+// because the parent process kills the Claude subprocess at message_stop
+// before tool execution (break-early pattern).
+"use strict";
+
+const fs = require("fs");
+const readline = require("readline");
+
+const schemaPath = process.argv[2];
+if (!schemaPath) {
+  process.exit(1);
+}
+
+let tools = [];
+try {
+  tools = JSON.parse(fs.readFileSync(schemaPath, "utf-8"));
+} catch {
+  process.exit(1);
+}
+
+const rl = readline.createInterface({ input: process.stdin });
+rl.on("line", (line) => {
+  let msg;
+  try {
+    msg = JSON.parse(line);
+  } catch {
+    return;
+  }
+
+  if (msg.method === "initialize") {
+    const resp = {
+      jsonrpc: "2.0",
+      id: msg.id,
+      result: {
+        protocolVersion: "2024-11-05",
+        capabilities: { tools: {} },
+        serverInfo: { name: "custom-tools", version: "1.0.0" },
+      },
+    };
+    process.stdout.write(JSON.stringify(resp) + "\n");
+  } else if (msg.method === "tools/list") {
+    const resp = { jsonrpc: "2.0", id: msg.id, result: { tools } };
+    process.stdout.write(JSON.stringify(resp) + "\n");
+  }
+  // notifications/initialized: no response needed (notification)
+  // tools/call: never reached (break-early kills subprocess first)
+});

packages/pi-claude-cli/src/process-manager.ts

@@ -0,0 +1,218 @@
+/**
+ * Process manager for spawning and managing Claude CLI subprocesses.
+ *
+ * Handles subprocess lifecycle: spawn with correct CLI flags, write NDJSON
+ * messages to stdin, force-kill after result (CLI hangs bug), and stderr capture.
+ * Also provides startup validation for CLI presence and authentication.
+ */
+
+import spawn from "cross-spawn";
+import { execSync } from "node:child_process";
+import { writeFileSync, unlinkSync } from "node:fs";
+import { join } from "node:path";
+import { tmpdir } from "node:os";
+import type { ChildProcess } from "node:child_process";
+
+/**
+ * Spawn a Claude CLI subprocess with all required flags for stream-json communication.
+ *
+ * @param modelId - The model ID to pass via --model flag
+ * @param systemPrompt - Optional system prompt appended via --append-system-prompt
+ * @param options - Optional cwd, AbortSignal, and effort level
+ * @returns The spawned ChildProcess with piped stdin/stdout/stderr
+ */
+export function spawnClaude(
+  modelId: string,
+  systemPrompt?: string,
+  options?: {
+    cwd?: string;
+    signal?: AbortSignal;
+    effort?: string;
+    mcpConfigPath?: string;
+    resumeSessionId?: string;
+    newSessionId?: string;
+  },
+): ChildProcess {
+  const args = [
+    "-p",
+    "--input-format",
+    "stream-json",
+    "--output-format",
+    "stream-json",
+    "--verbose",
+    "--include-partial-messages",
+    "--model",
+    modelId,
+    "--permission-prompt-tool",
+    "stdio",
+  ];
+
+  if (options?.resumeSessionId) {
+    // Resume an existing session — CLI loads prior conversation from disk
+    args.push("--resume", options.resumeSessionId);
+  } else if (options?.newSessionId) {
+    // First turn: create session with this ID so subsequent turns can --resume it
+    args.push("--session-id", options.newSessionId);
+  }
+
+  if (systemPrompt) {
+    // Write system prompt to a temp file to avoid ENAMETOOLONG on Windows.
+    // Claude CLI's --append-system-prompt accepts a file path or literal text.
+    const tmpFile = join(
+      tmpdir(),
+      `pi-claude-cli-sysprompt-${process.pid}.txt`,
+    );
+    writeFileSync(tmpFile, systemPrompt, "utf-8");
+    args.push("--append-system-prompt", tmpFile);
+  }
+
+  if (options?.effort) {
+    args.push("--effort", options.effort);
+  }
+
+  if (options?.mcpConfigPath) {
+    args.push("--mcp-config", options.mcpConfigPath);
+  }
+
+  const proc = spawn("claude", args, {
+    stdio: ["pipe", "pipe", "pipe"],
+    cwd: options?.cwd ?? process.cwd(),
+  });
+
+  return proc as ChildProcess;
+}
+
+/**
+ * Clean up the temp system prompt file created by spawnClaude.
+ * Safe to call multiple times or when no file exists.
+ */
+export function cleanupSystemPromptFile(): void {
+  try {
+    unlinkSync(join(tmpdir(), `pi-claude-cli-sysprompt-${process.pid}.txt`));
+  } catch {
+    // File doesn't exist or already deleted — ignore
+  }
+}
+
+/**
+ * Write a user message to the subprocess stdin as NDJSON.
+ * Does NOT call stdin.end() -- stdin stays open for control_response in Phase 2.
+ *
+ * Accepts both string (text-only prompt) and array (ContentBlock[] with images)
+ * content. JSON.stringify handles both natively. The stream-json protocol
+ * supports either format in the content field.
+ *
+ * @param proc - The Claude subprocess
+ * @param prompt - The prompt text or ContentBlock[] to send
+ */
+export function writeUserMessage(
+  proc: ChildProcess,
+  prompt: string | any[],
+): void {
+  const message = {
+    type: "user",
+    message: {
+      role: "user",
+      content: prompt,
+    },
+  };
+  proc.stdin!.write(JSON.stringify(message) + "\n");
+}
+
+/**
+ * Force-kill a subprocess immediately via SIGKILL.
+ * No-ops if the process is already dead (killed or exited).
+ * Cross-platform safe: Node.js treats SIGKILL as forceful termination on Windows.
+ *
+ * @param proc - The subprocess to force-kill
+ */
+export function forceKillProcess(proc: ChildProcess): void {
+  if (proc.killed || proc.exitCode !== null) return;
+  proc.kill("SIGKILL");
+}
+
+/** Registry of active subprocesses for cleanup on teardown. */
+const activeProcesses = new Set<ChildProcess>();
+
+/**
+ * Register a subprocess in the global process registry.
+ * The process is automatically removed from the registry when it exits.
+ *
+ * @param proc - The subprocess to track
+ */
+export function registerProcess(proc: ChildProcess): void {
+  activeProcesses.add(proc);
+  proc.on("exit", () => activeProcesses.delete(proc));
+}
+
+/**
+ * Force-kill all registered subprocesses and clear the registry.
+ * Safe to call multiple times -- no-ops on already-dead processes.
+ */
+export function killAllProcesses(): void {
+  for (const proc of activeProcesses) {
+    forceKillProcess(proc);
+  }
+  activeProcesses.clear();
+}
+
+/**
+ * Force-kill the subprocess after a 500ms grace period.
+ * The Claude CLI hangs after emitting the result message (known bug).
+ * Brief grace period allows final stdout flushing before force-kill.
+ *
+ * @param proc - The Claude subprocess to clean up
+ */
+export function cleanupProcess(proc: ChildProcess): void {
+  setTimeout(() => {
+    forceKillProcess(proc);
+  }, 500);
+}
+
+/**
+ * Attach a data listener to stderr and accumulate output into a buffer.
+ *
+ * @param proc - The Claude subprocess
+ * @returns A function that returns the accumulated stderr string
+ */
+export function captureStderr(proc: ChildProcess): () => string {
+  let buffer = "";
+  proc.stderr!.on("data", (data: Buffer) => {
+    buffer += data.toString();
+  });
+  return () => buffer;
+}
+
+/**
+ * Validate that the Claude CLI is installed and on PATH.
+ * Throws with install instructions if not found.
+ */
+export function validateCliPresence(): void {
+  try {
+    execSync("claude --version", { stdio: "pipe", timeout: 5000 });
+  } catch {
+    throw new Error(
+      "Claude Code CLI not found. Install it: npm install -g @anthropic-ai/claude-code\n" +
+        "Then authenticate: claude auth login",
+    );
+  }
+}
+
+/**
+ * Validate that the Claude CLI is authenticated.
+ * Returns false and warns if not authenticated.
+ *
+ * @returns true if authenticated, false otherwise
+ */
+export function validateCliAuth(): boolean {
+  try {
+    execSync("claude auth status", { stdio: "pipe", timeout: 5000 });
+    return true;
+  } catch {
+    console.warn(
+      "[pi-claude-cli] Claude CLI is not authenticated. " +
+        "Run 'claude auth login' to authenticate.",
+    );
+    return false;
+  }
+}

packages/pi-claude-cli/src/prompt-builder.ts

@@ -0,0 +1,514 @@
+/**
+ * Prompt builder for flattening pi conversation history into a labeled text prompt.
+ *
+ * Follows the reference project's buildPromptBlocks() pattern:
+ * - USER: / ASSISTANT: / TOOL RESULT: labels
+ * - Content blocks serialized by type
+ * - Images in the final user message are translated to Anthropic API format (HIST-02)
+ * - Images in non-final messages get placeholder text with console.warn
+ */
+
+import { existsSync, readFileSync } from "node:fs";
+import { resolve, join, dirname } from "node:path";
+import { homedir } from "node:os";
+import {
+  mapPiToolNameToClaude,
+  translatePiArgsToClaude,
+  isCustomToolName,
+} from "./tool-mapping.js";
+
+/**
+ * Anthropic API content block types for image passthrough.
+ * Used when the final user message contains images that need to be
+ * translated from pi-ai format to Anthropic format.
+ */
+type AnthropicContentBlock =
+  | { type: "text"; text: string }
+  | {
+      type: "image";
+      source: { type: "base64"; media_type: string; data: string };
+    };
+
+// We use `any` for Context to avoid requiring @mariozechner/pi-ai at dev time.
+// At runtime, pi provides the real Context type.
+
+/**
+ * Flattens a pi conversation context's messages array into a labeled text prompt
+ * suitable for sending to the Claude CLI subprocess.
+ *
+ * Each message is labeled with its role:
+ * - USER: for user messages
+ * - ASSISTANT: for assistant messages
+ * - TOOL RESULT (historical {toolName}): for tool result messages
+ */
+/** Module-level counter for placeholder images, reset per buildPrompt call. */
+let placeholderImageCount = 0;
+
+/**
+ * Translate a pi-ai image block to Anthropic API format.
+ * Returns null if the block is missing required data/mimeType fields.
+ *
+ * pi-ai format:  { type: "image", data: string (base64), mimeType: string }
+ * Anthropic format: { type: "image", source: { type: "base64", media_type: string, data: string } }
+ */
+function translateImageBlock(piBlock: any): AnthropicContentBlock | null {
+  if (piBlock.data && piBlock.mimeType) {
+    return {
+      type: "image",
+      source: {
+        type: "base64",
+        media_type: piBlock.mimeType,
+        data: piBlock.data,
+      },
+    };
+  }
+  return null; // Invalid image block, will fall back to placeholder
+}
+
+/**
+ * Build content blocks for the final user message, translating images
+ * from pi-ai format to Anthropic API format.
+ *
+ * @returns Array of AnthropicContentBlock with text and translated images
+ */
+function buildFinalUserContent(
+  content: string | any[],
+): AnthropicContentBlock[] {
+  if (typeof content === "string") {
+    return [{ type: "text", text: content }];
+  }
+  if (!Array.isArray(content)) {
+    return [{ type: "text", text: "" }];
+  }
+
+  const blocks: AnthropicContentBlock[] = [];
+  for (const block of content) {
+    if (block.type === "text") {
+      blocks.push({ type: "text", text: block.text ?? "" });
+    } else if (block.type === "image") {
+      const translated = translateImageBlock(block);
+      if (translated) {
+        blocks.push(translated);
+      } else {
+        // Invalid image block: fall back to placeholder text
+        blocks.push({
+          type: "text",
+          text: "[An image was shared here but could not be included]",
+        });
+        placeholderImageCount++;
+      }
+    }
+    // Unknown block types silently skipped
+  }
+  return blocks;
+}
+
+/**
+ * Check if a message content array contains image blocks.
+ */
+function contentHasImages(content: string | any[]): boolean {
+  if (typeof content === "string" || !Array.isArray(content)) return false;
+  return content.some((block) => block.type === "image");
+}
+
+/**
+ * Check if the conversation ends with a custom tool result.
+ * If so, build a simplified prompt that presents the result directly
+ * instead of replaying the full conversation history with tool labels.
+ */
+function buildCustomToolResultPrompt(messages: any[]): string | null {
+  if (messages.length < 3) return null;
+
+  const last = messages[messages.length - 1];
+  if (last.role !== "toolResult") return null;
+  if (!last.toolName || !isCustomToolName(last.toolName)) return null;
+
+  // Find the original user message (scan backwards past assistant + toolResult)
+  let userMessage: string | null = null;
+  for (let i = messages.length - 3; i >= 0; i--) {
+    if (messages[i].role === "user") {
+      userMessage = userContentToText(messages[i].content);
+      break;
+    }
+  }
+  if (!userMessage) return null;
+
+  const toolResult = toolResultContentToText(last.content);
+  return `${userMessage}\n\n[The ${last.toolName} tool was called and returned the following result]\n${toolResult}\n\nRespond to the user using the tool result above.`;
+}
+
+/**
+ * Build a prompt for a resumed session.
+ *
+ * When resuming via --resume, the CLI already has the full conversation history.
+ * We only need to send the new content since the last turn: the last assistant
+ * response's tool results (if any) followed by the latest user message.
+ *
+ * For tool_use flows: pi sends [user, assistant(toolCall), toolResult, ...]
+ * We need to include tool results so the resumed session sees them, plus the
+ * final user message.
+ *
+ * Falls back to full prompt if the message structure is unexpected.
+ */
+export function buildResumePrompt(context: {
+  messages: any[];
+}): string | AnthropicContentBlock[] {
+  const messages = context.messages;
+  if (messages.length === 0) return "";
+
+  // Find the last user message
+  const finalUserIndex = findFinalUserMessageIndex(messages);
+  if (finalUserIndex < 0) return "";
+
+  // Collect new messages: everything from the last assistant turn onwards
+  // (tool results from the last assistant + the new user message)
+  const newMessages: any[] = [];
+
+  // Walk backwards from finalUserIndex to find where new content starts.
+  // Include trailing toolResult messages that follow the last assistant turn.
+  let startIdx = finalUserIndex;
+  for (let i = finalUserIndex - 1; i >= 0; i--) {
+    if (messages[i].role === "toolResult") {
+      startIdx = i;
+    } else {
+      break;
+    }
+  }
+
+  for (let i = startIdx; i < messages.length; i++) {
+    newMessages.push(messages[i]);
+  }
+
+  // If there are only tool results + one user message, build a combined prompt
+  const parts: string[] = [];
+  for (const msg of newMessages) {
+    if (msg.role === "toolResult") {
+      if (msg.toolName && isCustomToolName(msg.toolName)) {
+        parts.push(`TOOL RESULT (${msg.toolName}):`);
+      } else {
+        const claudeToolName = msg.toolName
+          ? mapPiToolNameToClaude(msg.toolName)
+          : "unknown";
+        parts.push(`TOOL RESULT (historical ${claudeToolName}):`);
+      }
+      parts.push(toolResultContentToText(msg.content));
+    } else if (msg.role === "user") {
+      // Check for images in the final user message
+      if (contentHasImages(msg.content)) {
+        const textSoFar = parts.join("\n");
+        const userContent = buildFinalUserContent(msg.content);
+        const result: AnthropicContentBlock[] = [];
+        if (textSoFar) {
+          result.push({ type: "text", text: textSoFar });
+        }
+        result.push(...userContent);
+        return result;
+      }
+      parts.push(userContentToText(msg.content));
+    }
+  }
+
+  return parts.join("\n") || "";
+}
+
+export function buildPrompt(context: {
+  messages: any[];
+}): string | AnthropicContentBlock[] {
+  // Reset placeholder counter for each call
+  placeholderImageCount = 0;
+
+  // Special case: when conversation ends with a custom tool result,
+  // present it directly instead of complex history replay
+  const customToolPrompt = buildCustomToolResultPrompt(context.messages);
+  if (customToolPrompt) {
+    // customToolPrompt calls userContentToText which may increment placeholderImageCount
+    if (placeholderImageCount > 0) {
+      console.warn(
+        `[pi-claude-cli] ${placeholderImageCount} image(s) in conversation history could not be included in the prompt`,
+      );
+    }
+    return customToolPrompt;
+  }
+
+  // Determine if any message has images worth passing through
+  const finalUserIndex = findFinalUserMessageIndex(context.messages);
+  const finalUserHasImages =
+    finalUserIndex >= 0 &&
+    contentHasImages(context.messages[finalUserIndex].content);
+  const anyToolResultHasImages = context.messages.some(
+    (m: any) => m.role === "toolResult" && toolResultHasImages(m.content),
+  );
+
+  if (finalUserHasImages || anyToolResultHasImages) {
+    // Build history as text (all messages except the final user message)
+    const historyParts: string[] = [];
+    const toolResultImageBlocks: AnthropicContentBlock[] = [];
+    for (let i = 0; i < context.messages.length; i++) {
+      if (i === finalUserIndex) continue; // Skip final user message -- handled separately
+      const message = context.messages[i];
+      if (message.role === "user") {
+        historyParts.push("USER:");
+        historyParts.push(userContentToText(message.content));
+      } else if (message.role === "assistant") {
+        historyParts.push("ASSISTANT:");
+        historyParts.push(contentToText(message.content));
+      } else if (message.role === "toolResult") {
+        if (message.toolName && isCustomToolName(message.toolName)) {
+          historyParts.push(`TOOL RESULT (${message.toolName}):`);
+        } else {
+          const claudeToolName = message.toolName
+            ? mapPiToolNameToClaude(message.toolName)
+            : "unknown";
+          historyParts.push(`TOOL RESULT (historical ${claudeToolName}):`);
+        }
+        // Extract text portion of tool result
+        historyParts.push(toolResultContentToText(message.content));
+        // Collect image blocks from tool results for passthrough
+        if (Array.isArray(message.content)) {
+          for (const block of message.content) {
+            if (block.type === "image") {
+              const translated = translateImageBlock(block);
+              if (translated) {
+                toolResultImageBlocks.push(translated);
+                // Undo the placeholder count from toolResultContentToText since we're passing through
+                placeholderImageCount--;
+              }
+            }
+          }
+        }
+      }
+    }
+
+    // Build final user message content blocks
+    const finalUserContent =
+      finalUserIndex >= 0
+        ? buildFinalUserContent(context.messages[finalUserIndex].content)
+        : [];
+
+    // Combine: history text + tool result images + final user content blocks
+    const result: AnthropicContentBlock[] = [];
+    const historyText = historyParts.join("\n");
+    if (historyText) {
+      result.push({ type: "text", text: historyText });
+    }
+    // Insert tool result images after history text (Claude sees them in context)
+    result.push(...toolResultImageBlocks);
+    result.push(...finalUserContent);
+
+    if (placeholderImageCount > 0) {
+      console.warn(
+        `[pi-claude-cli] ${placeholderImageCount} image(s) in conversation history could not be included in the prompt`,
+      );
+    }
+
+    return result;
+  }
+
+  // No images in final user message: standard text-only path
+  const parts: string[] = [];
+
+  for (const message of context.messages) {
+    if (message.role === "user") {
+      parts.push("USER:");
+      parts.push(userContentToText(message.content));
+    } else if (message.role === "assistant") {
+      parts.push("ASSISTANT:");
+      parts.push(contentToText(message.content));
+    } else if (message.role === "toolResult") {
+      if (message.toolName && isCustomToolName(message.toolName)) {
+        // Custom tools: don't reference MCP tool name. Present result plainly.
+        parts.push(`TOOL RESULT (${message.toolName}):`);
+      } else {
+        const claudeToolName = message.toolName
+          ? mapPiToolNameToClaude(message.toolName)
+          : "unknown";
+        parts.push(`TOOL RESULT (historical ${claudeToolName}):`);
+      }
+      parts.push(toolResultContentToText(message.content));
+    }
+  }
+
+  if (placeholderImageCount > 0) {
+    console.warn(
+      `[pi-claude-cli] ${placeholderImageCount} image(s) in conversation history could not be included in the prompt`,
+    );
+  }
+
+  return parts.join("\n") || "";
+}
+
+/**
+ * Find the index of the last user message in the messages array.
+ * Returns -1 if no user message found.
+ */
+function findFinalUserMessageIndex(messages: any[]): number {
+  for (let i = messages.length - 1; i >= 0; i--) {
+    if (messages[i].role === "user") return i;
+  }
+  return -1;
+}
+
+/**
+ * Builds the system prompt from the context's systemPrompt field,
+ * appending AGENTS.md content if found (walking up from cwd, then global fallback).
+ * Sanitizes .pi references to .claude for Claude Code compatibility.
+ */
+export function buildSystemPrompt(
+  context: { systemPrompt?: string; messages: any[] },
+  cwd: string,
+): string {
+  const parts: string[] = [];
+
+  if (context.systemPrompt) {
+    parts.push(context.systemPrompt);
+  }
+
+  // Look for AGENTS.md
+  const agentsPath = resolveAgentsMdPath(cwd);
+  if (agentsPath) {
+    try {
+      const content = readFileSync(agentsPath, "utf-8");
+      const sanitized = sanitizeAgentsContent(content);
+      parts.push(sanitized);
+    } catch {
+      // If we can't read it, skip silently
+    }
+  }
+
+  // When conversation history has tool results, instruct Claude to use them
+  // instead of trying to re-call tools (which may not be available).
+  if (context.messages?.some((m: any) => m.role === "toolResult")) {
+    parts.push(
+      "IMPORTANT: The conversation history below contains tool results from previously executed tools. " +
+        "Use these results to answer the user's question. Do NOT attempt to re-call tools that already have results.",
+    );
+  }
+
+  return parts.join("\n\n");
+}
+
+/**
+ * Converts user message content to text.
+ * Handles string content and array of content blocks.
+ * Image blocks are replaced with placeholder text (HIST-02).
+ * Increments the module-level placeholderImageCount for each image.
+ */
+function userContentToText(content: string | any[]): string {
+  if (typeof content === "string") return content;
+  if (!Array.isArray(content)) return "";
+
+  const texts: string[] = [];
+  for (const block of content) {
+    if (block.type === "text") {
+      texts.push(block.text ?? "");
+    } else if (block.type === "image") {
+      texts.push("[An image was shared here but could not be included]");
+      placeholderImageCount++;
+    }
+    // Unknown block types silently skipped
+  }
+  return texts.join("\n");
+}
+
+/**
+ * Converts assistant message content to text.
+ * Handles string content and array of content blocks (text, thinking, toolCall).
+ */
+function contentToText(content: string | any[]): string {
+  if (typeof content === "string") return content;
+  if (!Array.isArray(content)) return "";
+
+  return content
+    .map((block) => {
+      if (block.type === "text") return block.text ?? "";
+      if (block.type === "thinking") return ""; // Skip thinking — internal reasoning, not conversation
+      if (block.type === "toolCall") {
+        const isCustom = isCustomToolName(block.name);
+        if (isCustom) {
+          // Custom tools: don't reference the MCP tool name — Claude might try to re-call it.
+          // Just note what was done. The result follows as a TOOL RESULT message.
+          const argsStr = block.arguments
+            ? JSON.stringify(block.arguments)
+            : "{}";
+          return `[Used ${block.name} tool with args: ${argsStr}]`;
+        }
+        const claudeName = mapPiToolNameToClaude(block.name);
+        const claudeArgs =
+          block.arguments && typeof block.arguments === "object"
+            ? translatePiArgsToClaude(
+                block.name,
+                block.arguments as Record<string, unknown>,
+              )
+            : block.arguments;
+        const argsStr = claudeArgs ? JSON.stringify(claudeArgs) : "{}";
+        return `Historical tool call (non-executable): ${claudeName} args=${argsStr}`;
+      }
+      // Unknown block types are represented as a placeholder
+      return `[${block.type}]`;
+    })
+    .join("\n");
+}
+
+/**
+ * Converts tool result content to text.
+ * Handles string content and array of content blocks.
+ * Image blocks get placeholder text (actual image passthrough handled separately).
+ */
+function toolResultContentToText(content: string | any[]): string {
+  if (typeof content === "string") return content;
+  if (!Array.isArray(content)) return "";
+
+  const texts: string[] = [];
+  for (const block of content) {
+    if (block.type === "text") {
+      texts.push(block.text ?? "");
+    } else if (block.type === "image") {
+      texts.push("[An image was shared here but could not be included]");
+      placeholderImageCount++;
+    }
+  }
+  return texts.join("\n");
+}
+
+/**
+ * Check if a tool result content array contains image blocks.
+ */
+function toolResultHasImages(content: string | any[]): boolean {
+  if (typeof content === "string" || !Array.isArray(content)) return false;
+  return content.some((block) => block.type === "image");
+}
+
+/**
+ * Walk up from cwd looking for AGENTS.md, fall back to ~/.pi/agent/AGENTS.md.
+ */
+function resolveAgentsMdPath(cwd: string): string | undefined {
+  let current = resolve(cwd);
+  while (true) {
+    const candidate = join(current, "AGENTS.md");
+    if (existsSync(candidate)) return candidate;
+    const parent = dirname(current);
+    if (parent === current) break;
+    current = parent;
+  }
+
+  // Fall back to global path
+  const globalPath = join(homedir(), ".pi", "agent", "AGENTS.md");
+  if (existsSync(globalPath)) return globalPath;
+
+  return undefined;
+}
+
+/**
+ * Sanitize .pi references to .claude in AGENTS.md content
+ * for Claude Code compatibility.
+ */
+function sanitizeAgentsContent(content: string): string {
+  let sanitized = content;
+  // ~/.pi -> ~/.claude
+  sanitized = sanitized.replace(/~\/\.pi\b/gi, "~/.claude");
+  // .pi/ -> .claude/ (at word boundary or after whitespace/quotes)
+  sanitized = sanitized.replace(/(^|[\s'"`])\.pi\//g, "$1.claude/");
+  // Remaining standalone .pi references
+  sanitized = sanitized.replace(/\b\.pi\b/gi, ".claude");
+  return sanitized;
+}

packages/pi-claude-cli/src/provider.ts

@@ -0,0 +1,336 @@
+/**
+ * Provider orchestration for bridging pi requests to the Claude CLI subprocess.
+ *
+ * streamViaCli is the core function that:
+ * 1. Builds the prompt from conversation context
+ * 2. Spawns a Claude CLI subprocess with correct flags
+ * 3. Writes the user message to stdin as NDJSON
+ * 4. Reads stdout line-by-line, parsing NDJSON
+ * 5. Routes stream events through the event bridge to pi's stream
+ * 6. Handles result/error messages and cleans up the subprocess
+ * 7. Implements break-early: kills subprocess at message_stop when
+ *    built-in or custom-tools MCP tool_use blocks are seen
+ * 8. Hardened lifecycle: inactivity timeout, subprocess exit handler,
+ *    streamEnded guard, abort via SIGKILL, process registry
+ */
+
+import { createInterface } from "node:readline";
+import {
+  AssistantMessageEventStream,
+  type Model,
+  type SimpleStreamOptions,
+} from "@mariozechner/pi-ai";
+import {
+  buildPrompt,
+  buildSystemPrompt,
+  buildResumePrompt,
+} from "./prompt-builder.js";
+import {
+  spawnClaude,
+  writeUserMessage,
+  cleanupProcess,
+  captureStderr,
+  forceKillProcess,
+  registerProcess,
+  cleanupSystemPromptFile,
+} from "./process-manager.js";
+import { parseLine } from "./stream-parser.js";
+import { createEventBridge } from "./event-bridge.js";
+import { handleControlRequest } from "./control-handler.js";
+import { mapThinkingEffort } from "./thinking-config.js";
+import { isPiKnownClaudeTool } from "./tool-mapping.js";
+/** Inactivity timeout: kill subprocess if no stdout for 180 seconds (3 minutes). */
+const INACTIVITY_TIMEOUT_MS = 180_000;
+
+/** Extended stream options: pi's SimpleStreamOptions plus optional cwd and mcpConfigPath */
+type StreamViaCLiOptions = SimpleStreamOptions & {
+  cwd?: string;
+  mcpConfigPath?: string;
+};
+
+/**
+ * Stream a response from Claude CLI as an AssistantMessageEventStream.
+ *
+ * Orchestrates the full subprocess lifecycle: spawn, write prompt, parse NDJSON,
+ * bridge events, handle result, and clean up. Implements break-early pattern:
+ * at message_stop, if any built-in or custom-tools MCP tool was seen, kills
+ * the subprocess before Claude CLI can auto-execute the tools.
+ *
+ * Hardened with: inactivity timeout (180s), subprocess exit handler with stderr
+ * surfacing, streamEnded guard against double errors, abort via SIGKILL, and
+ * process registry integration for teardown cleanup.
+ *
+ * @param model - The model to use (from pi's model catalog)
+ * @param context - The conversation context with messages and system prompt
+ * @param options - Optional cwd, abort signal, reasoning level, thinking budgets, and mcpConfigPath
+ * @returns An AssistantMessageEventStream that receives bridged events
+ */
+export function streamViaCli(
+  model: Model<any>,
+  context: { messages: any[]; systemPrompt?: string },
+  options?: StreamViaCLiOptions,
+): AssistantMessageEventStream {
+  // @ts-expect-error — tsc can't verify AssistantMessageEventStream is a value
+  // through pi-ai's `export *` re-export chain. The class constructor exists at runtime.
+  const stream = new AssistantMessageEventStream();
+
+  (async () => {
+    let proc: ReturnType<typeof spawnClaude> | undefined;
+    let abortHandler: (() => void) | undefined;
+
+    try {
+      const cwd = options?.cwd ?? process.cwd();
+
+      // Resume if pi provides a session ID AND this isn't the first turn.
+      // Pi passes sessionId on every call (including first), but we can only
+      // --resume a CLI session that already exists on disk from a prior turn.
+      const resumeSessionId =
+        options?.sessionId && context.messages.length > 1
+          ? options.sessionId
+          : undefined;
+
+      // Build prompt: if resuming, only send the latest user turn;
+      // otherwise build the full flattened conversation history
+      const prompt = resumeSessionId
+        ? buildResumePrompt(context)
+        : buildPrompt(context);
+      const systemPrompt = resumeSessionId
+        ? undefined
+        : buildSystemPrompt(context, cwd);
+
+      // Compute effort level from reasoning options
+      const effort = mapThinkingEffort(
+        options?.reasoning,
+        model.id,
+        options?.thinkingBudgets,
+      );
+
+      // Spawn subprocess
+      proc = spawnClaude(model.id, systemPrompt || undefined, {
+        cwd,
+        signal: options?.signal,
+        effort,
+        mcpConfigPath: options?.mcpConfigPath,
+        resumeSessionId,
+        newSessionId: !resumeSessionId ? options?.sessionId : undefined,
+      });
+      const getStderr = captureStderr(proc);
+
+      // Register in global process registry for teardown cleanup
+      registerProcess(proc);
+
+      // Write user message to subprocess stdin
+      writeUserMessage(proc, prompt);
+
+      // Create event bridge (before endStreamWithError so bridge is in scope)
+      const bridge = createEventBridge(stream, model);
+
+      // Guard against double stream.end() and double error events.
+      // First error path wins; subsequent ones are no-ops.
+      let streamEnded = false;
+
+      /**
+       * End the stream with an error, using a "done" event instead of "error".
+       *
+       * Why "done" not "error": AssistantMessageEventStream.extractResult()
+       * returns event.error (a string) for error events, but agent-loop.js
+       * then calls message.content.filter() on the result, crashing because
+       * a string has no .content property. By pushing "done" with a valid
+       * AssistantMessage (content:[]), pi gets a well-formed object.
+       */
+      function endStreamWithError(errMsg: string) {
+        if (streamEnded || broken) return;
+        streamEnded = true;
+        const output = bridge.getOutput();
+        const errorMessage = {
+          ...output,
+          content: output.content?.length
+            ? output.content
+            : [{ type: "text" as const, text: `Error: ${errMsg}` }],
+          stopReason: "stop" as const,
+        };
+        stream.push({
+          type: "done",
+          reason: "stop",
+          message: errorMessage,
+        } as any);
+        stream.end();
+      }
+
+      // Inactivity timeout: kill subprocess if no stdout for INACTIVITY_TIMEOUT_MS
+      let inactivityTimer: ReturnType<typeof setTimeout> | undefined;
+
+      function resetInactivityTimer() {
+        if (inactivityTimer !== undefined) clearTimeout(inactivityTimer);
+        inactivityTimer = setTimeout(() => {
+          forceKillProcess(proc!);
+          endStreamWithError(
+            `Claude CLI subprocess timed out: no output for ${INACTIVITY_TIMEOUT_MS / 1000} seconds`,
+          );
+        }, INACTIVITY_TIMEOUT_MS);
+      }
+
+      // Set up abort signal handler -- uses SIGKILL for immediate force-kill
+      if (options?.signal) {
+        abortHandler = () => {
+          if (proc) {
+            forceKillProcess(proc);
+          }
+        };
+
+        if (options.signal.aborted) {
+          abortHandler();
+          return;
+        }
+        options.signal.addEventListener("abort", abortHandler, { once: true });
+      }
+
+      // Track tool_use blocks for break-early decision at message_stop
+      let sawBuiltInOrCustomTool = false;
+      // Guard against buffered readline lines firing after rl.close()
+      let broken = false;
+
+      // Set up readline for line-by-line NDJSON parsing
+      const rl = createInterface({
+        input: proc.stdout!,
+        crlfDelay: Infinity,
+        terminal: false,
+      });
+
+      // Handle process error -- use endStreamWithError for guard
+      proc.on("error", (err: Error) => {
+        if (broken) return; // Break-early killed the process intentionally
+        const stderr = getStderr();
+        endStreamWithError(stderr || err.message);
+      });
+
+      // Handle subprocess close -- surface crashes with stderr and exit code
+      proc.on("close", (code: number | null, _signal: string | null) => {
+        clearTimeout(inactivityTimer);
+        if (broken) return; // Break-early kill, expected
+        if (code !== 0 && code !== null) {
+          const stderr = getStderr();
+          const message = stderr
+            ? `Claude CLI exited with code ${code}: ${stderr.trim()}`
+            : `Claude CLI exited unexpectedly with code ${code}`;
+          endStreamWithError(message);
+        }
+      });
+
+      // Start inactivity timer after writing user message
+      resetInactivityTimer();
+
+      // Process NDJSON lines from stdout using event-based callback
+      // NOTE: Using 'line' event instead of `for await` because the async
+      // iterator batches lines, breaking real-time streaming to pi.
+      rl.on("line", (line: string) => {
+        if (broken) return; // Guard: ignore buffered lines after break-early
+
+        // Reset inactivity timer on each line of output
+        resetInactivityTimer();
+
+        const msg = parseLine(line);
+        if (!msg) return;
+
+        if (msg.type === "stream_event") {
+          // Only forward top-level events to pi's event bridge.
+          // Sub-agent events (parent_tool_use_id !== null) are internal to the CLI.
+          const isTopLevel = !(msg as any).parent_tool_use_id;
+          if (isTopLevel) {
+            bridge.handleEvent(msg.event);
+          }
+
+          // Track tool_use blocks for break-early decision (top-level only)
+          if (
+            isTopLevel &&
+            msg.event.type === "content_block_start" &&
+            msg.event.content_block?.type === "tool_use"
+          ) {
+            const toolName = msg.event.content_block.name;
+            if (toolName && isPiKnownClaudeTool(toolName)) {
+              // Built-in tool (Read/Write/etc.) OR custom MCP tool (mcp__custom-tools__*)
+              // Internal Claude Code tools (ToolSearch, Task, etc.) are excluded
+              sawBuiltInOrCustomTool = true;
+            }
+          }
+
+          // Break-early at message_stop: kill subprocess before CLI auto-executes tools
+          // Only on top-level message_stop — sub-agent message_stop is internal
+          if (
+            isTopLevel &&
+            msg.event.type === "message_stop" &&
+            sawBuiltInOrCustomTool
+          ) {
+            broken = true; // Set guard BEFORE rl.close() to prevent buffered lines
+            clearTimeout(inactivityTimer);
+            // Pi will execute these tools. Kill subprocess to prevent CLI from executing them.
+            forceKillProcess(proc!);
+            rl.close();
+            return; // Don't process further -- done event already pushed by event bridge
+          }
+        } else if (msg.type === "control_request") {
+          handleControlRequest(msg, proc!.stdin!);
+        } else if (msg.type === "result") {
+          if (msg.subtype === "error") {
+            endStreamWithError(msg.error ?? "Unknown error from Claude CLI");
+          }
+          // For both success and error: clean up the subprocess
+          clearTimeout(inactivityTimer);
+          cleanupProcess(proc!);
+          rl.close();
+        }
+      });
+
+      // Wait for readline to close (result received or process ended)
+      await new Promise<void>((resolve) => {
+        rl.on("close", resolve);
+      });
+
+      // Push done event after readline closes (async). Pushing synchronously
+      // inside handleMessageStop prevents pi from executing tools.
+      // Guard with streamEnded to avoid pushing done after an error was already pushed.
+      if (!streamEnded) {
+        const output = bridge.getOutput();
+
+        // If stopReason is toolUse but there are no pi-known tool calls in content,
+        // it means only user MCP tools were called (filtered by event bridge).
+        // Override to "stop" so pi doesn't try to execute non-existent tools.
+        const piToolCalls = (output.content || []).filter(
+          (c: any) => c.type === "toolCall",
+        );
+        const effectiveReason =
+          output.stopReason === "toolUse" && piToolCalls.length === 0
+            ? "stop"
+            : output.stopReason;
+
+        streamEnded = true;
+        stream.push({
+          type: "done",
+          reason:
+            effectiveReason === "toolUse"
+              ? "toolUse"
+              : effectiveReason === "length"
+                ? "length"
+                : "stop",
+          message: { ...output, stopReason: effectiveReason },
+        });
+        stream.end();
+      }
+    } catch (err: any) {
+      stream.push({
+        type: "error",
+        reason: "error",
+        error: err.message ?? "Unexpected error in streamViaCli",
+      } as any);
+      stream.end();
+    } finally {
+      // Clean up abort listener
+      if (options?.signal && abortHandler) {
+        options.signal.removeEventListener("abort", abortHandler);
+      }
+      cleanupSystemPromptFile();
+    }
+  })();
+
+  return stream;
+}

packages/pi-claude-cli/src/stream-parser.ts

@@ -0,0 +1,37 @@
+import type { NdjsonMessage } from "./types";
+
+/**
+ * Parse a single NDJSON line from Claude CLI stdout into a typed message.
+ *
+ * This function is deliberately resilient -- it never throws. Debug noise,
+ * empty lines, and malformed JSON all return null so the streaming pipeline
+ * can safely skip them and continue processing.
+ */
+export function parseLine(line: string): NdjsonMessage | null {
+  const trimmed = line.trim();
+
+  // Skip empty lines
+  if (!trimmed) {
+    return null;
+  }
+
+  // Skip non-JSON lines (debug output like "[SandboxDebug] ...")
+  if (!trimmed.startsWith("{")) {
+    return null;
+  }
+
+  let parsed: unknown;
+  try {
+    parsed = JSON.parse(trimmed);
+  } catch {
+    console.error("Failed to parse NDJSON line:", trimmed);
+    return null;
+  }
+
+  // Validate that the parsed result is a non-null object (not array, not primitive)
+  if (parsed === null || typeof parsed !== "object" || Array.isArray(parsed)) {
+    return null;
+  }
+
+  return parsed as NdjsonMessage;
+}

packages/pi-claude-cli/src/thinking-config.ts

@@ -0,0 +1,83 @@
+/**
+ * Thinking effort configuration for mapping pi's ThinkingLevel to Claude CLI --effort flags.
+ *
+ * Maps pi's reasoning levels (minimal/low/medium/high/xhigh) to the CLI's effort
+ * levels (low/medium/high/max). Opus models get an elevated mapping where medium
+ * becomes high and high becomes max, leveraging their superior reasoning capability.
+ *
+ * IMPORTANT: The CLI does NOT support --thinking-budget. Only --effort is supported.
+ */
+
+import type { ThinkingLevel, ThinkingBudgets } from "@mariozechner/pi-ai";
+
+/** CLI effort levels accepted by the --effort flag */
+export type CliEffortLevel = "low" | "medium" | "high" | "max";
+
+/**
+ * Standard model mapping: pi ThinkingLevel -> CLI effort.
+ * Non-Opus models never receive "max" (would cause CLI error).
+ */
+const STANDARD_EFFORT_MAP: Record<ThinkingLevel, CliEffortLevel> = {
+  minimal: "low",
+  low: "low",
+  medium: "medium",
+  high: "high",
+  xhigh: "high", // non-Opus: silently downgrade (max not supported)
+};
+
+/**
+ * Opus model mapping: shifted up for elevated reasoning.
+ * Opus models get max capability at high/xhigh levels.
+ */
+const OPUS_EFFORT_MAP: Record<ThinkingLevel, CliEffortLevel> = {
+  minimal: "low",
+  low: "low",
+  medium: "high", // shifted: standard high
+  high: "max", // shifted: maximum capability
+  xhigh: "max", // Opus gets max
+};
+
+/**
+ * Detect whether a model ID refers to an Opus model.
+ * Uses includes('opus') for forward-compatibility with future Opus versions.
+ *
+ * @param modelId - The model identifier string
+ * @returns true if the model is an Opus variant
+ */
+export function isOpusModel(modelId: string): boolean {
+  return modelId.includes("opus");
+}
+
+/**
+ * Map pi's ThinkingLevel to a CLI effort string.
+ *
+ * When reasoning is undefined, returns undefined so the --effort flag is omitted
+ * entirely, letting the CLI use its default behavior. When thinkingBudgets are
+ * provided, a console.warn is logged because the CLI only supports effort levels,
+ * not token budgets.
+ *
+ * @param reasoning - Pi's thinking level (undefined = omit flag)
+ * @param modelId - Model ID for Opus detection
+ * @param thinkingBudgets - Custom budgets (logged as unsupported, not applied)
+ * @returns CLI effort level string, or undefined if flag should be omitted
+ */
+export function mapThinkingEffort(
+  reasoning?: ThinkingLevel,
+  modelId?: string,
+  thinkingBudgets?: ThinkingBudgets,
+): CliEffortLevel | undefined {
+  if (reasoning === undefined) {
+    return undefined; // omit --effort flag entirely
+  }
+
+  if (thinkingBudgets && Object.keys(thinkingBudgets).length > 0) {
+    console.warn(
+      "[pi-claude-cli] Custom thinkingBudgets are not supported with CLI subprocess. " +
+        "The CLI uses --effort levels instead of token budgets. Budgets will be ignored.",
+    );
+  }
+
+  const isOpus = modelId ? isOpusModel(modelId) : false;
+  const map = isOpus ? OPUS_EFFORT_MAP : STANDARD_EFFORT_MAP;
+  return map[reasoning];
+}

packages/pi-claude-cli/src/tool-mapping.ts

@@ -0,0 +1,147 @@
+/**
+ * Single-source-of-truth tool mapping table for bidirectional translation
+ * between Claude CLI tool names/arguments and pi tool names/arguments.
+ *
+ * All lookup tables are derived from the TOOL_MAPPINGS array.
+ * Unknown tools and arguments pass through unchanged.
+ */
+
+/**
+ * A mapping entry for a single tool.
+ * `args` maps Claude argument names to pi argument names (only renamed args).
+ */
+export interface ToolMapping {
+  claude: string;
+  pi: string;
+  args: Record<string, string>;
+}
+
+/**
+ * The canonical tool mapping table. All other lookup structures are derived from this.
+ */
+export const TOOL_MAPPINGS: ToolMapping[] = [
+  { claude: "Read", pi: "read", args: { file_path: "path" } },
+  { claude: "Write", pi: "write", args: { file_path: "path" } },
+  {
+    claude: "Edit",
+    pi: "edit",
+    args: { file_path: "path", old_string: "oldText", new_string: "newText" },
+  },
+  { claude: "Bash", pi: "bash", args: {} },
+  { claude: "Grep", pi: "grep", args: { head_limit: "limit" } },
+  { claude: "Glob", pi: "find", args: {} },
+];
+
+/** Prefix for custom pi tools exposed via MCP. */
+export const CUSTOM_TOOLS_MCP_PREFIX = "mcp__custom-tools__";
+
+/** Set of built-in pi tool names derived from TOOL_MAPPINGS for O(1) lookup. */
+const BUILT_IN_PI_NAMES = new Set(TOOL_MAPPINGS.map((m) => m.pi));
+
+/**
+ * Check if a pi tool name is a custom tool (not one of the 6 built-in tools).
+ * Used by prompt builder to decide whether to add MCP prefix in history replay.
+ */
+export function isCustomToolName(piName: string): boolean {
+  return !BUILT_IN_PI_NAMES.has(piName);
+}
+
+/**
+ * Check if a Claude tool name maps to a pi-known tool.
+ * Returns true for built-in tools (Read, Write, etc.) and custom MCP tools (mcp__custom-tools__*).
+ * Returns false for internal Claude Code tools (ToolSearch, Task, Agent, etc.) that pi cannot execute.
+ * Used by event bridge to filter out internal tool calls.
+ */
+export function isPiKnownClaudeTool(claudeName: string): boolean {
+  if (claudeName.startsWith(CUSTOM_TOOLS_MCP_PREFIX)) return true;
+  return claudeName.toLowerCase() in CLAUDE_TO_PI_NAME;
+}
+
+// Derived lookup maps
+
+/** Lowercase Claude name -> pi name */
+const CLAUDE_TO_PI_NAME: Record<string, string> = {};
+/** Pi name -> PascalCase Claude name */
+const PI_TO_CLAUDE_NAME: Record<string, string> = {};
+/** Lowercase Claude name -> { claudeArgName: piArgName } */
+const CLAUDE_TO_PI_ARGS: Record<string, Record<string, string>> = {};
+/** Pi name -> { piArgName: claudeArgName } */
+const PI_TO_CLAUDE_ARGS: Record<string, Record<string, string>> = {};
+
+for (const m of TOOL_MAPPINGS) {
+  CLAUDE_TO_PI_NAME[m.claude.toLowerCase()] = m.pi;
+  PI_TO_CLAUDE_NAME[m.pi] = m.claude;
+  CLAUDE_TO_PI_ARGS[m.claude.toLowerCase()] = m.args;
+
+  // Build reverse arg map
+  const reverseArgs: Record<string, string> = {};
+  for (const [from, to] of Object.entries(m.args)) {
+    reverseArgs[to] = from;
+  }
+  PI_TO_CLAUDE_ARGS[m.pi] = reverseArgs;
+}
+
+// Handle glob/find asymmetry: pi's "glob" also maps back to Claude's "Glob"
+PI_TO_CLAUDE_NAME["glob"] = "Glob";
+
+/**
+ * Map a Claude tool name to the corresponding pi tool name.
+ * Strips the mcp__custom-tools__ prefix for custom tools first,
+ * then falls back to case-insensitive built-in lookup.
+ * Unknown tool names pass through unchanged.
+ */
+export function mapClaudeToolNameToPi(claudeName: string): string {
+  // Strip custom-tools MCP prefix first (e.g., "mcp__custom-tools__deploy" -> "deploy")
+  if (claudeName.startsWith(CUSTOM_TOOLS_MCP_PREFIX)) {
+    return claudeName.slice(CUSTOM_TOOLS_MCP_PREFIX.length);
+  }
+  // Standard built-in tool mapping (case-insensitive)
+  return CLAUDE_TO_PI_NAME[claudeName.toLowerCase()] ?? claudeName;
+}
+
+/**
+ * Map a pi tool name to the corresponding Claude tool name.
+ * Direct lookup. Unknown tool names pass through unchanged.
+ */
+export function mapPiToolNameToClaude(piName: string): string {
+  return PI_TO_CLAUDE_NAME[piName] ?? piName;
+}
+
+/**
+ * Translate Claude tool arguments to pi format.
+ * Only known renamed arguments are translated; all others pass through unchanged.
+ * This prevents dropping unknown/extra arguments (Pitfall 5).
+ */
+export function translateClaudeArgsToPi(
+  claudeToolName: string,
+  args: Record<string, unknown>,
+): Record<string, unknown> {
+  const renames = CLAUDE_TO_PI_ARGS[claudeToolName.toLowerCase()];
+  if (!renames || Object.keys(renames).length === 0) return args;
+
+  const result: Record<string, unknown> = {};
+  for (const [key, value] of Object.entries(args)) {
+    const newKey = renames[key] ?? key;
+    result[newKey] = value;
+  }
+  return result;
+}
+
+/**
+ * Translate pi tool arguments to Claude format.
+ * Only known renamed arguments are translated; all others pass through unchanged.
+ */
+export function translatePiArgsToClaude(
+  piToolName: string,
+  args: Record<string, unknown>,
+): Record<string, unknown> {
+  const renames = PI_TO_CLAUDE_ARGS[piToolName];
+  if (!renames || Object.keys(renames).length === 0) return args;
+
+  const result: Record<string, unknown> = {};
+  for (const [key, value] of Object.entries(args)) {
+    const newKey = renames[key] ?? key;
+    result[newKey] = value;
+  }
+  return result;
+}

packages/pi-claude-cli/src/types.ts

@@ -0,0 +1,85 @@
+// Wire protocol types for Claude CLI stream-json NDJSON communication
+
+// NDJSON message types from Claude CLI stdout
+
+export interface ClaudeStreamEventMessage {
+  type: "stream_event";
+  event: ClaudeApiEvent;
+}
+
+export interface ClaudeResultMessage {
+  type: "result";
+  subtype: "success" | "error";
+  result?: string;
+  error?: string;
+  session_id?: string;
+}
+
+export interface ClaudeSystemMessage {
+  type: "system";
+  subtype: string;
+  session_id?: string;
+  tools?: unknown[];
+}
+
+export interface ClaudeControlRequest {
+  type: "control_request";
+  request_id: string;
+  request: {
+    subtype: "can_use_tool";
+    tool_name: string;
+    input: Record<string, unknown>;
+  };
+}
+
+export type NdjsonMessage =
+  | ClaudeStreamEventMessage
+  | ClaudeResultMessage
+  | ClaudeSystemMessage
+  | ClaudeControlRequest;
+
+// Claude API event types (inside stream_event wrapper)
+
+export interface ClaudeApiEvent {
+  type: string; // message_start, content_block_start, content_block_delta, content_block_stop, message_delta, message_stop
+  index?: number;
+  message?: {
+    id?: string;
+    type?: string;
+    role?: string;
+    content?: unknown[];
+    model?: string;
+    usage?: ClaudeUsage;
+  };
+  content_block?: {
+    type: string; // "text", "tool_use", "thinking"
+    text?: string;
+    id?: string;
+    name?: string;
+    input?: string;
+  };
+  delta?: {
+    type?: string; // "text_delta", "input_json_delta", "thinking_delta", "signature_delta"
+    text?: string;
+    partial_json?: string;
+    thinking?: string;
+    signature?: string;
+    stop_reason?: string;
+  };
+  usage?: ClaudeUsage;
+}
+
+export interface ClaudeUsage {
+  input_tokens?: number;
+  output_tokens?: number;
+  cache_read_input_tokens?: number;
+  cache_creation_input_tokens?: number;
+}
+
+// Content block tracking during stream processing
+
+export interface TrackedContentBlock {
+  type: "text" | "thinking";
+  text: string;
+  index: number; // Claude's content_block index
+}