experimental devbox support

server/README.md

@@ -46,6 +46,14 @@ If `ADMIN_TOKEN` is not set, the server runs in open mode (dev).
 - `EXA_API_KEY`
 - `CHAT_WEB_SEARCH_ENGINE` (`exa` by default, or `searxng` for chat tool calls only)
 - `SEARXNG_BASE_URL` (required when `CHAT_WEB_SEARCH_ENGINE=searxng`; instance must allow `format=json`)
+- `CHAT_CODEX_TOOL_ENABLED` (`false` by default; enables the `codex_exec` chat tool for OpenAI/xAI)
+- `CHAT_CODEX_REMOTE_HOST` (required when Codex tool is enabled; SSH host/IP or `user@host`)
+- `CHAT_CODEX_REMOTE_USER` (optional SSH user when host does not include one)
+- `CHAT_CODEX_REMOTE_PORT` (`22` by default)
+- `CHAT_CODEX_REMOTE_WORKDIR` (`/workspace/sybil-codex` by default; created and reused on the devbox)
+- `CHAT_CODEX_SSH_KEY_PATH` (recommended: path to a read-only mounted private key)
+- `CHAT_CODEX_SSH_PRIVATE_KEY_B64` (optional fallback private key delivery)
+- `CHAT_CODEX_EXEC_TIMEOUT_MS` (`600000` by default)
 
 ## API
 - `GET /health`

server/src/env.ts

@@ -20,6 +20,34 @@ const ChatWebSearchEngineSchema = z.preprocess(
   z.enum(["exa", "searxng"]).default("exa")
 );
 
+const BooleanFlagSchema = z.preprocess((value) => {
+  if (typeof value !== "string") return value;
+  const normalized = value.trim().toLowerCase();
+  if (!normalized) return undefined;
+  if (["1", "true", "yes", "on"].includes(normalized)) return true;
+  if (["0", "false", "no", "off"].includes(normalized)) return false;
+  return value;
+}, z.boolean().default(false));
+
+const OptionalTrimmedStringSchema = z.preprocess(
+  (value) => (typeof value === "string" && value.trim() === "" ? undefined : value),
+  z.string().trim().min(1).optional()
+);
+
+function defaultedPositiveInt(defaultValue: number) {
+  return z.preprocess(
+    (value) => (typeof value === "string" && value.trim() === "" ? undefined : value),
+    z.coerce.number().int().positive().default(defaultValue)
+  );
+}
+
+function defaultedTrimmedString(defaultValue: string) {
+  return z.preprocess(
+    (value) => (typeof value === "string" && value.trim() === "" ? undefined : value),
+    z.string().trim().min(1).default(defaultValue)
+  );
+}
+
 const EnvSchema = z.object({
   PORT: z.coerce.number().int().positive().default(8787),
   HOST: z.string().default("0.0.0.0"),
@@ -36,6 +64,17 @@ const EnvSchema = z.object({
   // Chat-mode web_search tool configuration. Search mode remains Exa-only for now.
   CHAT_WEB_SEARCH_ENGINE: ChatWebSearchEngineSchema,
   SEARXNG_BASE_URL: OptionalUrlSchema,
+
+  // Optional chat-mode Codex tool. When enabled, the server SSHes into a remote
+  // devbox and runs `codex exec` in a persistent scratch directory there.
+  CHAT_CODEX_TOOL_ENABLED: BooleanFlagSchema,
+  CHAT_CODEX_REMOTE_HOST: OptionalTrimmedStringSchema,
+  CHAT_CODEX_REMOTE_USER: OptionalTrimmedStringSchema,
+  CHAT_CODEX_REMOTE_PORT: defaultedPositiveInt(22),
+  CHAT_CODEX_REMOTE_WORKDIR: defaultedTrimmedString("/workspace/sybil-codex"),
+  CHAT_CODEX_SSH_KEY_PATH: OptionalTrimmedStringSchema,
+  CHAT_CODEX_SSH_PRIVATE_KEY_B64: OptionalTrimmedStringSchema,
+  CHAT_CODEX_EXEC_TIMEOUT_MS: defaultedPositiveInt(600_000),
 }).superRefine((value, ctx) => {
   if (value.CHAT_WEB_SEARCH_ENGINE === "searxng" && !value.SEARXNG_BASE_URL) {
     ctx.addIssue({
@@ -44,6 +83,14 @@ const EnvSchema = z.object({
       message: "SEARXNG_BASE_URL is required when CHAT_WEB_SEARCH_ENGINE=searxng",
     });
   }
+
+  if (value.CHAT_CODEX_TOOL_ENABLED && !value.CHAT_CODEX_REMOTE_HOST) {
+    ctx.addIssue({
+      code: "custom",
+      path: ["CHAT_CODEX_REMOTE_HOST"],
+      message: "CHAT_CODEX_REMOTE_HOST is required when CHAT_CODEX_TOOL_ENABLED=true",
+    });
+  }
 });
 
 export type Env = z.infer<typeof EnvSchema>;

server/src/llm/chat-tools.ts

@@ -1,3 +1,8 @@
+import { execFile } from "node:child_process";
+import { mkdtemp, rm, writeFile } from "node:fs/promises";
+import os from "node:os";
+import path from "node:path";
+import { promisify } from "node:util";
 import { convert as htmlToText } from "html-to-text";
 import type OpenAI from "openai";
 import { z } from "zod";
@@ -13,6 +18,12 @@ const MAX_WEB_RESULTS = 10;
 const DEFAULT_FETCH_MAX_CHARACTERS = 12_000;
 const MAX_FETCH_MAX_CHARACTERS = 50_000;
 const FETCH_TIMEOUT_MS = 12_000;
+const MAX_CODEX_PROMPT_CHARACTERS = 60_000;
+const DEFAULT_CODEX_MAX_OUTPUT_CHARACTERS = 24_000;
+const MAX_CODEX_MAX_OUTPUT_CHARACTERS = 80_000;
+const CODEX_EXEC_MAX_BUFFER_BYTES = 1_000_000;
+
+const execFileAsync = promisify(execFile);
 
 const WebSearchArgsSchema = z
   .object({
@@ -33,7 +44,43 @@ const FetchUrlArgsSchema = z
   })
   .strict();
 
-const CHAT_TOOLS: any[] = [
+const CodexExecArgsSchema = z
+  .object({
+    prompt: z.string().trim().min(1).max(MAX_CODEX_PROMPT_CHARACTERS),
+    maxCharacters: z.coerce.number().int().min(1_000).max(MAX_CODEX_MAX_OUTPUT_CHARACTERS).optional(),
+  })
+  .strict();
+
+type CodexExecArgs = z.infer<typeof CodexExecArgsSchema>;
+
+const CODEX_EXEC_TOOL = {
+  type: "function",
+  function: {
+    name: "codex_exec",
+    description:
+      "Delegate a coding, terminal, or multi-step software task to a persistent remote Codex CLI workspace. Use for complex code changes, repository inspection, running programs/tests, debugging build failures, or other tasks that need a real shell. Return the remote Codex summary and relevant stdout/stderr.",
+    parameters: {
+      type: "object",
+      properties: {
+        prompt: {
+          type: "string",
+          description:
+            "A complete, self-contained instruction for the remote Codex instance. Include the goal, relevant context, constraints, and what result to report back.",
+        },
+        maxCharacters: {
+          type: "integer",
+          minimum: 1_000,
+          maximum: MAX_CODEX_MAX_OUTPUT_CHARACTERS,
+          description: "Maximum stdout/stderr characters returned to the model (default 24000).",
+        },
+      },
+      required: ["prompt"],
+      additionalProperties: false,
+    },
+  },
+};
+
+const BASE_CHAT_TOOLS: any[] = [
   {
     type: "function",
     function: {
@@ -95,10 +142,15 @@ const CHAT_TOOLS: any[] = [
   },
 ];
 
+const CHAT_TOOLS: any[] = env.CHAT_CODEX_TOOL_ENABLED ? [...BASE_CHAT_TOOLS, CODEX_EXEC_TOOL] : BASE_CHAT_TOOLS;
+
 export const CHAT_TOOL_SYSTEM_PROMPT =
   "You can use tools to gather up-to-date web information when needed. " +
   "Use web_search for discovery and recent facts, and fetch_url to read the full content of a specific page. " +
   "Prefer tools when the user asks for current events, verification, sources, or details you do not already have. " +
+  (env.CHAT_CODEX_TOOL_ENABLED
+    ? "Use codex_exec when a request needs substantial coding work, repository inspection, shell commands, tests, debugging, or another complex task suited to a persistent Codex workspace. Provide codex_exec a complete prompt with the goal, constraints, and expected report-back format. "
+    : "") +
   "Do not fabricate tool outputs; reason only from provided tool results.";
 
 type ToolRunOutcome = {
@@ -192,6 +244,14 @@ function buildToolSummary(name: string, args: Record<string, unknown>, status: "
     return url ? `Fetching URL ${toSingleLine(url, 140)} failed.${errSuffix}` : `Fetching URL failed.${errSuffix}`;
   }
 
+  if (name === "codex_exec") {
+    const prompt = typeof args.prompt === "string" ? args.prompt.trim() : "";
+    if (status === "completed") {
+      return prompt ? `Ran Codex task: '${toSingleLine(prompt, 120)}'.` : "Ran Codex task.";
+    }
+    return prompt ? `Codex task '${toSingleLine(prompt, 120)}' failed.${errSuffix}` : `Codex task failed.${errSuffix}`;
+  }
+
   if (status === "completed") {
     return `Ran tool '${name}'.`;
   }
@@ -402,9 +462,133 @@ async function runFetchUrlTool(input: unknown): Promise<ToolRunOutcome> {
   };
 }
 
+function shellQuote(value: string) {
+  return `'${value.replace(/'/g, `'\\''`)}'`;
+}
+
+function buildCodexSshTarget() {
+  const host = env.CHAT_CODEX_REMOTE_HOST;
+  if (!host) {
+    throw new Error("CHAT_CODEX_REMOTE_HOST not set");
+  }
+  if (!env.CHAT_CODEX_REMOTE_USER || host.includes("@")) {
+    return host;
+  }
+  return `${env.CHAT_CODEX_REMOTE_USER}@${host}`;
+}
+
+function buildRemoteCodexCommand(prompt: string) {
+  const workdir = env.CHAT_CODEX_REMOTE_WORKDIR.trim();
+  const codexCommand = `codex exec ${shellQuote(prompt)}`;
+  return `mkdir -p ${shellQuote(workdir)} && cd ${shellQuote(workdir)} && ${codexCommand}`;
+}
+
+async function withCodexSshKeyPath<T>(fn: (keyPath?: string) => Promise<T>) {
+  if (env.CHAT_CODEX_SSH_KEY_PATH) {
+    return fn(env.CHAT_CODEX_SSH_KEY_PATH);
+  }
+
+  if (!env.CHAT_CODEX_SSH_PRIVATE_KEY_B64) {
+    return fn(undefined);
+  }
+
+  const tmpDir = await mkdtemp(path.join(os.tmpdir(), "sybil-codex-ssh-"));
+  const keyPath = path.join(tmpDir, "id");
+  try {
+    await writeFile(keyPath, Buffer.from(env.CHAT_CODEX_SSH_PRIVATE_KEY_B64, "base64"), { mode: 0o600 });
+    return await fn(keyPath);
+  } finally {
+    await rm(tmpDir, { recursive: true, force: true });
+  }
+}
+
+function clipCodexOutput(value: string, maxCharacters: number) {
+  if (value.length <= maxCharacters) {
+    return { text: value, truncated: false };
+  }
+  return {
+    text: `${value.slice(0, maxCharacters)}\n\n[truncated ${value.length - maxCharacters} characters]`,
+    truncated: true,
+  };
+}
+
+function bufferOrStringToString(value: unknown) {
+  if (typeof value === "string") return value;
+  if (Buffer.isBuffer(value)) return value.toString("utf8");
+  return "";
+}
+
+async function runCodexExecTool(input: unknown): Promise<ToolRunOutcome> {
+  if (!env.CHAT_CODEX_TOOL_ENABLED) {
+    return { ok: false, error: "codex_exec is disabled." };
+  }
+
+  const args: CodexExecArgs = CodexExecArgsSchema.parse(input);
+  const maxCharacters = args.maxCharacters ?? DEFAULT_CODEX_MAX_OUTPUT_CHARACTERS;
+  const sshTarget = buildCodexSshTarget();
+  const remoteCommand = buildRemoteCodexCommand(args.prompt);
+
+  const run = async (keyPath?: string) => {
+    const sshArgs = [
+      "-o",
+      "BatchMode=yes",
+      "-o",
+      "StrictHostKeyChecking=accept-new",
+      "-o",
+      "UserKnownHostsFile=/tmp/sybil-codex-known-hosts",
+      "-p",
+      String(env.CHAT_CODEX_REMOTE_PORT),
+    ];
+
+    if (keyPath) {
+      sshArgs.push("-i", keyPath);
+    }
+
+    sshArgs.push(sshTarget, remoteCommand);
+
+    try {
+      const result = await execFileAsync("ssh", sshArgs, {
+        timeout: env.CHAT_CODEX_EXEC_TIMEOUT_MS,
+        maxBuffer: CODEX_EXEC_MAX_BUFFER_BYTES,
+      });
+      const stdout = clipCodexOutput(bufferOrStringToString(result.stdout), maxCharacters);
+      const stderr = clipCodexOutput(bufferOrStringToString(result.stderr), Math.min(maxCharacters, 12_000));
+      return {
+        ok: true,
+        host: env.CHAT_CODEX_REMOTE_HOST,
+        workdir: env.CHAT_CODEX_REMOTE_WORKDIR,
+        stdout: stdout.text,
+        stderr: stderr.text,
+        stdoutTruncated: stdout.truncated,
+        stderrTruncated: stderr.truncated,
+      };
+    } catch (err: any) {
+      const stdout = clipCodexOutput(bufferOrStringToString(err?.stdout), maxCharacters);
+      const stderr = clipCodexOutput(bufferOrStringToString(err?.stderr), Math.min(maxCharacters, 12_000));
+      return {
+        ok: false,
+        error: err?.killed
+          ? `Remote Codex command timed out after ${env.CHAT_CODEX_EXEC_TIMEOUT_MS}ms.`
+          : err?.message ?? String(err),
+        exitCode: typeof err?.code === "number" ? err.code : null,
+        signal: typeof err?.signal === "string" ? err.signal : null,
+        host: env.CHAT_CODEX_REMOTE_HOST,
+        workdir: env.CHAT_CODEX_REMOTE_WORKDIR,
+        stdout: stdout.text,
+        stderr: stderr.text,
+        stdoutTruncated: stdout.truncated,
+        stderrTruncated: stderr.truncated,
+      };
+    }
+  };
+
+  return withCodexSshKeyPath(run);
+}
+
 async function executeTool(name: string, args: unknown): Promise<ToolRunOutcome> {
   if (name === "web_search") return runWebSearchTool(args);
   if (name === "fetch_url") return runFetchUrlTool(args);
+  if (name === "codex_exec") return runCodexExecTool(args);
   return { ok: false, error: `Unknown tool: ${name}` };
 }
 
@@ -419,6 +603,16 @@ function parseToolArgs(raw: unknown) {
   }
 }
 
+function buildEventArgs(name: string, args: Record<string, unknown>) {
+  if (name !== "codex_exec" || typeof args.prompt !== "string") {
+    return args;
+  }
+  return {
+    ...args,
+    prompt: clipText(args.prompt, 1_000),
+  };
+}
+
 function mergeUsage(acc: Required<ToolAwareUsage>, usage: any) {
   if (!usage) return false;
   acc.inputTokens += usage.prompt_tokens ?? 0;
@@ -468,12 +662,13 @@ async function executeToolCallAndBuildEvent(
       : undefined;
 
   const completedAtMs = Date.now();
+  const eventArgs = buildEventArgs(call.name, parsedArgs);
   const event: ToolExecutionEvent = {
     toolCallId: call.id,
     name: call.name,
     status,
-    summary: buildToolSummary(call.name, parsedArgs, status, error),
-    args: parsedArgs,
+    summary: buildToolSummary(call.name, eventArgs, status, error),
+    args: eventArgs,
     startedAt,
     completedAt: new Date(completedAtMs).toISOString(),
     durationMs: completedAtMs - startedAtMs,