@ai-sdk/harness-codex 1.0.126 → 1.0.129

package/src/codex-harness.ts

@@ -7,6 +7,7 @@ import {
   type HarnessV1,
   type HarnessV1DebugConfig,
   type HarnessV1BuiltinTool,
+  type HarnessV1BuiltinToolFiltering,
   type HarnessV1ContinueTurnState,
   type HarnessV1CredentialForwarding,
   type HarnessV1MintBridgeTokenCallback,
@@ -29,6 +30,7 @@ import {
   createSandboxCredentialEnvironment,
   createBridgeErrorHandler,
   createBridgeStartupError,
+  experimental_createBridgeUserMessageSubmitter,
   drainBridgeProcessStream,
   forwardBridgeProcessStream,
   getRestrictedSandboxSession,
@@ -43,6 +45,7 @@ import {
   type SandboxChannelReconnectOptions,
 } from '@ai-sdk/harness/utils';
 import {
+  tool,
   type Experimental_SandboxProcess,
   type Experimental_SandboxSession as SandboxSession,
 } from '@ai-sdk/provider-utils';
@@ -67,7 +70,6 @@ import {
   type InboundMessage,
   type OutboundMessage,
 } from './codex-bridge-protocol';
-import { CLI_SHIM_FILENAME } from './bridge/cli-relay';
 import { VERSION } from './version';
 
 type CodexChannel = SandboxChannel<OutboundMessage, InboundMessage>;
@@ -147,14 +149,13 @@ export type CodexHarnessSettings = {
 };
 
 /*
- * Every native tool the Codex CLI can invoke as a model-callable tool,
+ * Every native tool Codex app-server can invoke as a model-callable tool,
  * declared as a `ToolSet` keyed by what the bridge emits as `toolName` on
- * the wire (`commonName ?? nativeName`). Schemas reflect the `ThreadItem`
- * union in `@openai/codex-sdk`'s `dist/index.d.ts`.
+ * the wire (`commonName ?? nativeName`). Schemas reflect the corresponding
+ * app-server item payloads.
  *
- * Codex's other native operations (`apply_patch`, todo planning) surface
- * only as side-effect events (`file_change`, `todo_list`) and are not
- * model-callable tools — they don't appear here.
+ * Other native operations such as todo planning surface as side-effect
+ * events and are not model-callable tools.
  */
 const CODEX_BUILTIN_TOOLS = {
   bash: commonTool('bash', {
@@ -169,6 +170,24 @@ const CODEX_BUILTIN_TOOLS = {
     description: 'Search the web',
     inputSchema: z.object({ query: z.string() }),
   }),
+  apply_patch: {
+    ...tool({
+      description: 'Apply a patch to files',
+      inputSchema: z.string(),
+    }),
+    toolUseKind: 'edit',
+  },
+  view_image: {
+    ...tool({
+      description: 'View a local image file',
+      inputSchema: z.object({
+        path: z.string(),
+        detail: z.enum(['high', 'original']).optional(),
+        environment_id: z.string().optional(),
+      }),
+    }),
+    toolUseKind: 'readonly',
+  },
 } as const satisfies Record<string, HarnessV1BuiltinTool<any, any>>;
 
 /**
@@ -185,7 +204,7 @@ const codexBridgeCoordsSchema = z.object({
 
 /**
  * Schema for the adapter-specific lifecycle `data` payload Codex produces.
- * `threadId` is what `codex.resumeThread(...)` requires for the replay/rerun
+ * `threadId` is what app-server's `thread/resume` method requires for the replay/rerun
  * rungs; the sandbox lookup is handled separately via
  * `provider.resumeSession({ sessionId })`. `bridge` carries live coordinates
  * for cross-process `attach` (present on `doDetach()` and `doSuspendTurn()`
@@ -207,18 +226,12 @@ export function createCodex(
     specificationVersion: 'harness-v1',
     harnessId: 'codex',
     builtinTools: CODEX_BUILTIN_TOOLS,
+    supportsBuiltinToolFiltering: true,
     supportsBuiltinToolApprovals: false,
     lifecycleStateSchema: codexResumeStateSchema,
     getBootstrap: getCodexBootstrap,
     doStart: async startOpts => {
       const model = DEFAULT_CODEX_MODEL;
-      if (startOpts.builtinToolFiltering != null) {
-        throw new HarnessCapabilityUnsupportedError({
-          message:
-            "Harness 'codex' does not support built-in tool filtering controls.",
-          harnessId: 'codex',
-        });
-      }
       if (
         startOpts.permissionMode != null &&
         startOpts.permissionMode !== 'allow-all'
@@ -336,8 +349,6 @@ export function createCodex(
         sessionId: startOpts.sessionId,
       });
       const bridgeStateDir = `${sessionDataDir}/bridge`;
-      const cliShimDir = `${sessionDataDir}/codex`;
-      const cliShimPath = `${cliShimDir}/${CLI_SHIM_FILENAME}`;
       const timeoutMs = settings.startupTimeoutMs ?? 120_000;
 
       // Normalize each forwarded bridge diagnostics frame into the general
@@ -390,12 +401,12 @@ export function createCodex(
           return createSession({
             sessionId: startOpts.sessionId,
             channel: attachChannel,
-            cliShimPath,
             // The live bridge was spawned by another process; no process handle.
             proc: undefined,
             model,
             reasoningEffort: settings.reasoningEffort,
             webSearch: settings.webSearch,
+            builtinToolFiltering: startOpts.builtinToolFiltering,
             codexConfig: settings.codexConfig,
             mcpServers: settings.mcpServers,
             headers: startOpts.headers,
@@ -424,7 +435,7 @@ export function createCodex(
        * log is replayed *from*. `resumeFrom` is a between-turn resume; even when
        * it carries bridge coordinates, replaying the previous turn would
        * re-deliver stale events into the next turn. Those resumes always `rerun`
-       * via `codex.resumeThread(threadId)` when attach is unavailable.
+       * via app-server's `thread/resume` method when attach is unavailable.
        */
       let respawnStrategy: CodexRespawnStrategy | undefined = isResume
         ? 'rerun'
@@ -490,7 +501,7 @@ export function createCodex(
       });
 
       const proc = await toolSafeSandboxSession.spawn({
-        command: `node ${shellQuote(`${bootstrapDir}/bridge.mjs`)} --workdir ${shellQuote(workDir)} --bridge-state-dir ${shellQuote(bridgeStateDir)} --cli-shim-dir ${shellQuote(cliShimDir)}`,
+        command: `node ${shellQuote(`${bootstrapDir}/bridge.mjs`)} --workdir ${shellQuote(workDir)} --bridge-state-dir ${shellQuote(bridgeStateDir)}`,
         env,
         abortSignal: startOpts.abortSignal,
       });
@@ -556,11 +567,11 @@ export function createCodex(
       return createSession({
         sessionId: startOpts.sessionId,
         channel,
-        cliShimPath,
         proc,
         model,
         reasoningEffort: settings.reasoningEffort,
         webSearch: settings.webSearch,
+        builtinToolFiltering: startOpts.builtinToolFiltering,
         codexConfig: settings.codexConfig,
         mcpServers: settings.mcpServers,
         headers: startOpts.headers,
@@ -707,11 +718,11 @@ function openWebSocket({
 function createSession({
   sessionId,
   channel,
-  cliShimPath,
   proc,
   model,
   reasoningEffort,
   webSearch,
+  builtinToolFiltering,
   codexConfig,
   mcpServers,
   headers,
@@ -731,12 +742,12 @@ function createSession({
 }: {
   sessionId: string;
   channel: CodexChannel;
-  cliShimPath: string;
   /** Undefined on `attach` — the live bridge was spawned by another process. */
   proc: Experimental_SandboxProcess | undefined;
   model: string | undefined;
   reasoningEffort: 'low' | 'medium' | 'high' | 'xhigh' | 'max' | undefined;
   webSearch: boolean | undefined;
+  builtinToolFiltering: HarnessV1BuiltinToolFiltering | undefined;
   codexConfig: Record<string, unknown> | undefined;
   mcpServers: Record<string, unknown> | undefined;
   headers: Readonly<Record<string, string>> | undefined;
@@ -758,21 +769,13 @@ function createSession({
   let stopPromise: Promise<void> | undefined;
   /*
    * Send the persisted threadId on the first prompt only when the bridge was
-   * respawned (rerun/replay) so it takes the `codex.resumeThread(...)` branch.
+   * respawned (rerun/replay) so it takes the app-server `thread/resume` branch.
    * An `attach`ed bridge already holds its threadState in memory and continues
    * on its own, so it needs no seed.
    */
   let pendingResumeThreadId = seedResumeThreadOnFirstPrompt
     ? resumeThreadId
     : undefined;
-  /*
-   * Host-tool relay guidance is prepended to the first user message of a fresh
-   * session only. A resumed session (attach/replay/rerun) already carried it in
-   * its original first message (preserved in the persisted thread), so it
-   * starts "applied".
-   */
-  let initialPromptGuidanceApplied = isResume;
-
   /*
    * Latest codex thread id, cached from the bridge's `bridge-thread`
    * announcements. Seeded from lifecycle state so `doDetach()` and `doStop()`
@@ -783,7 +786,6 @@ function createSession({
   channel.on('bridge-thread', msg => {
     latestThreadId = msg.threadId;
   });
-
   const synchronizeTurnConfiguration = async ({
     skills,
     instructions,
@@ -821,13 +823,6 @@ function createSession({
           latestTurnConfigurationFingerprint !== nextFingerprint));
     latestTurnConfigurationFingerprint = nextFingerprint;
     if (restartThread) {
-      /*
-       * `codex exec resume` retains the native thread's original developer
-       * instructions and skill catalog. A fresh native thread is therefore
-       * required for replacement semantics. Host-tool guidance is framed
-       * again because that guidance also belongs to the new native thread.
-       */
-      initialPromptGuidanceApplied = false;
       pendingResumeThreadId = undefined;
     }
     return { restartThread };
@@ -853,6 +848,11 @@ function createSession({
       pendingResolve = resolve;
       pendingReject = reject;
     });
+    const userMessageSubmitter = experimental_createBridgeUserMessageSubmitter({
+      send: message => channel.send(message),
+      onResponse: listener => channel.on('user-message-response', listener),
+      onReconnect: listener => channel.onReconnect(listener),
+    });
 
     const unsubs: Array<() => void> = [];
     const forward = (event: HarnessV1StreamPart) => {
@@ -880,12 +880,14 @@ function createSession({
     const settleSuccess = () => {
       if (isSettled) return;
       isSettled = true;
+      userMessageSubmitter.close();
       for (const u of unsubs) u();
       pendingResolve!();
     };
     const settleError = (err: unknown) => {
       if (isSettled) return;
       isSettled = true;
+      userMessageSubmitter.close(err);
       for (const u of unsubs) u();
       pendingReject!(err);
     };
@@ -963,6 +965,9 @@ function createSession({
           reason: input.reason,
         });
       },
+      submitUserMessage: async text => {
+        await userMessageSubmitter.submit(text);
+      },
       done,
     };
 
@@ -1017,24 +1022,9 @@ function createSession({
         abortSignal: promptOpts.abortSignal,
       });
 
-      let promptText = extractUserText(promptOpts.prompt);
-      if (!initialPromptGuidanceApplied) {
-        promptText = frameInitialPromptGuidance({
-          toolUsageBlock:
-            tools.length > 0
-              ? composeToolUsageInstructions({
-                  tools,
-                  cliShimPath,
-                })
-              : undefined,
-          userText: promptText,
-        });
-      }
-      initialPromptGuidanceApplied = true;
-
       const startMessage = {
         type: 'start' as const,
-        prompt: promptText,
+        prompt: extractUserText(promptOpts.prompt),
         tools,
         ...(promptOpts.responseFormat == null
           ? {}
@@ -1045,6 +1035,7 @@ function createSession({
         model: promptOpts.model ?? model,
         reasoningEffort,
         webSearch,
+        ...(builtinToolFiltering == null ? {} : { builtinToolFiltering }),
         ...(codexConfig == null ? {} : { codexConfig }),
         ...(mcpServers == null ? {} : { mcpServers }),
         ...(headers == null ? {} : { headers }),
@@ -1123,6 +1114,7 @@ function createSession({
             model: continueOpts.model ?? model,
             reasoningEffort,
             webSearch,
+            ...(builtinToolFiltering == null ? {} : { builtinToolFiltering }),
             ...(codexConfig == null ? {} : { codexConfig }),
             ...(mcpServers == null ? {} : { mcpServers }),
             ...(headers == null ? {} : { headers }),
@@ -1139,10 +1131,9 @@ function createSession({
     doCompact: async () => {
       /*
        * Codex compacts its context automatically inside the core turn loop
-       * (~90% of the model context window), but the `codex exec` transport this
-       * adapter drives exposes no manual compaction trigger and emits no
-       * compaction event. Manual `compact()` is therefore unsupported; Codex's
-       * own auto-compaction continues to run regardless.
+       * (~90% of the model context window). This adapter does not yet expose a
+       * manual compaction trigger or compaction event, so manual `compact()` is
+       * unsupported while Codex's own auto-compaction continues to run.
        */
       throw new HarnessCapabilityUnsupportedError({
         message:
@@ -1229,7 +1220,7 @@ function createSession({
        * payload — the workdir is still captured by the sandbox snapshot
        * during the subsequent `sandboxSession.stop()`, so the next turn can
        * resume the filesystem state. The trade-off: we lose
-       * `threadId`, so the codex CLI starts a fresh thread on the
+       * `threadId`, so Codex app-server starts a fresh thread on the
        * preserved workdir rather than resuming the prior conversation
        * inside Codex's runtime. Ability to continue beats throwing.
        */
@@ -1365,62 +1356,10 @@ function fingerprintCodexTurnConfiguration({
 }
 
 /*
- * Frame host-tool relay guidance and the user's text so Codex treats the
- * prepended block as operating guidance rather than user prose. Applied only
- * to the first user message of a fresh session.
- */
-function frameInitialPromptGuidance({
-  toolUsageBlock,
-  userText,
-}: {
-  toolUsageBlock: string | undefined;
-  userText: string;
-}): string {
-  const blocks: string[] = [];
-  if (toolUsageBlock) blocks.push(toolUsageBlock);
-  if (blocks.length === 0) return userText;
-  return `${blocks.join('\n\n')}\n\n<user-message>\n${userText}\n</user-message>`;
-}
-
-function composeToolUsageInstructions({
-  tools,
-  cliShimPath,
-}: {
-  tools: ReadonlyArray<{
-    name: string;
-    description?: string;
-    inputSchema?: unknown;
-  }>;
-  cliShimPath: string;
-}): string {
-  const lines: string[] = [
-    '<host-tool-instructions>',
-    'You have access to the following host-provided tools. To use one, run the following command via your built-in `bash` tool:',
-    '',
-    `  node ${cliShimPath} <toolName> '<jsonInput>'`,
-    '',
-    'The script prints the JSON result to stdout. Do not invent another way to call these tools — only this CLI invocation will work. Pass the JSON input as a single-quoted argument.',
-    'For every user request that depends on a host-provided tool, run a separate CLI invocation for each needed tool call in the current turn before answering. Do not reuse previous tool results, and do not say you used a host tool unless the command has completed in the current turn.',
-    '',
-  ];
-  for (const toolSpec of tools) {
-    lines.push(
-      `- **${toolSpec.name}**${toolSpec.description ? ': ' + toolSpec.description : ''}`,
-    );
-    lines.push(
-      `  - Input schema: \`${JSON.stringify(toolSpec.inputSchema ?? {})}\``,
-    );
-  }
-  lines.push('</host-tool-instructions>');
-  return lines.join('\n');
-}
-
-/*
- * Reduce a `HarnessV1Prompt` to the plain user text the bridge forwards
- * to the Codex SDK. File and image parts on the message are not yet
- * supported by the underlying runtime — throw rather than silently drop
- * them so callers learn about the gap instead of seeing mysteriously
- * truncated prompts.
+ * Reduce a `HarnessV1Prompt` to the plain user text the bridge forwards to
+ * Codex. The adapter does not yet map file and image parts into app-server
+ * turn input, so throw rather than silently dropping them and producing a
+ * truncated prompt.
  */
 function extractUserText(prompt: HarnessV1Prompt): string {
   if (typeof prompt === 'string') return prompt;

package/src/bridge/cli-relay.ts

@@ -1,268 +0,0 @@
-/*
- * Temporary workaround for upstream codex CLI bug
- * https://github.com/openai/codex/issues/19425 — MCP tools registered via
- * `mcp_servers.*` are not exposed to the model in `codex exec --experimental-json`
- * mode (which is what `@openai/codex-sdk` uses). The MCP handshake completes
- * and `tools/list` succeeds, but codex never registers the tools as
- * model-callable functions.
- *
- * Until that's fixed upstream, this file implements a CLI-based relay:
- *
- *   1. A small Node script is written into harness-owned session state at turn
- *      start (`buildCliShimScript`). It accepts `<toolName> <jsonInput>` as
- *      argv, POSTs to the same HTTP relay the MCP shim uses, and prints the
- *      result to stdout.
- *
- *   2. Tool descriptions and invocation instructions are injected into the
- *      initial user prompt by the host adapter, telling the model to call host
- *      tools by running `node <shim-path> <toolName> '<jsonInput>'` via its
- *      built-in `bash` tool.
- *
- *   3. The bridge's event loop suppresses the matching `command_execution`
- *      events (`isToolRelayCommand`) so callers receive clean `tool-call` /
- *      `tool-result` events from the HTTP relay rather than seeing the
- *      relay invocations as raw bash commands.
- *
- * Once #19425 is fixed and MCP tools are properly exposed in exec mode, the
- * three hookpoints in `bridge/index.ts` can be removed along with this file.
- */
-export const CLI_SHIM_FILENAME = 'harness-tool.mjs';
-
-export function buildCliShimScript({
-  relayPort,
-}: {
-  relayPort: number;
-}): string {
-  return `#!/usr/bin/env node
-const [toolName, inputJson = '{}'] = process.argv.slice(2);
-if (!toolName) {
-  console.error('Usage: harness-tool <tool_name> <json_input>');
-  process.exit(64);
-}
-let input;
-try {
-  input = JSON.parse(inputJson);
-} catch (error) {
-  console.error('Invalid JSON input: ' + (error instanceof Error ? error.message : String(error)));
-  process.exit(64);
-}
-const requestId = 'cli-' + Date.now() + '-' + Math.random().toString(16).slice(2);
-const response = await fetch('http://127.0.0.1:${relayPort}', {
-  method: 'POST',
-  headers: { 'Content-Type': 'application/json' },
-  body: JSON.stringify({ requestId, toolName, input }),
-});
-const text = await response.text();
-let payload;
-try {
-  payload = text ? JSON.parse(text) : {};
-} catch {
-  payload = { error: text };
-}
-if (!response.ok || payload.error) {
-  console.error(String(payload.error ?? ('tool relay failed with HTTP ' + response.status)));
-  process.exit(1);
-}
-console.log(JSON.stringify(payload.result ?? payload, null, 2));
-`;
-}
-
-export function isToolRelayCommand({
-  command,
-  cliShimPath,
-}: {
-  command: string;
-  cliShimPath: string;
-}): boolean {
-  return parseToolRelayCommands({ command, cliShimPath }) !== undefined;
-}
-
-export function parseToolRelayCommand({
-  command,
-  cliShimPath,
-}: {
-  command: string;
-  cliShimPath: string;
-}): { toolName: string; input: unknown } | undefined {
-  const calls = parseToolRelayCommands({ command, cliShimPath });
-  return calls?.length === 1 ? calls[0] : undefined;
-}
-
-export function parseToolRelayCommands({
-  command,
-  cliShimPath,
-}: {
-  command: string;
-  cliShimPath: string;
-}): Array<{ toolName: string; input: unknown }> | undefined {
-  return parseToolRelayCommandsInternal({ command, cliShimPath, depth: 0 });
-}
-
-function parseToolRelayCommandsInternal({
-  command,
-  cliShimPath,
-  depth,
-}: {
-  command: string;
-  cliShimPath: string;
-  depth: number;
-}): Array<{ toolName: string; input: unknown }> | undefined {
-  const commands = splitShellAndCommands(command);
-  if (!commands) return undefined;
-  if (commands.length > 1) {
-    const relayCalls: Array<{ toolName: string; input: unknown }> = [];
-    for (const nestedCommand of commands) {
-      const nestedCalls = parseToolRelayCommandsInternal({
-        command: nestedCommand,
-        cliShimPath,
-        depth,
-      });
-      if (!nestedCalls) return undefined;
-      relayCalls.push(...nestedCalls);
-    }
-    return relayCalls;
-  }
-
-  const argv = parseShellWords(command);
-  if (!argv) return undefined;
-
-  const relayCall = parseDirectToolRelayArgv({ argv, cliShimPath });
-  if (relayCall) return [relayCall];
-
-  const innerCommand = extractShellEvalCommand(argv);
-  if (!innerCommand || depth >= 2) return undefined;
-  return parseToolRelayCommandsInternal({
-    command: innerCommand,
-    cliShimPath,
-    depth: depth + 1,
-  });
-}
-
-function parseDirectToolRelayArgv({
-  argv,
-  cliShimPath,
-}: {
-  argv: string[];
-  cliShimPath: string;
-}): { toolName: string; input: unknown } | undefined {
-  if (argv.length < 3 || argv.length > 4) return undefined;
-  if (argv[0] !== 'node' || argv[1] !== cliShimPath) return undefined;
-  const toolName = argv[2];
-  if (!toolName) return undefined;
-  try {
-    return { toolName, input: JSON.parse(argv[3] ?? '{}') };
-  } catch {
-    return undefined;
-  }
-}
-
-function extractShellEvalCommand(argv: string[]): string | undefined {
-  if (argv.length !== 3) return undefined;
-  const shellName = argv[0].split('/').at(-1);
-  if (shellName !== 'bash' && shellName !== 'sh' && shellName !== 'zsh') {
-    return undefined;
-  }
-  if (argv[1] !== '-c' && argv[1] !== '-lc') return undefined;
-  return argv[2];
-}
-
-function splitShellAndCommands(command: string): string[] | undefined {
-  const commands: string[] = [];
-  let start = 0;
-  let quote: '"' | "'" | undefined;
-
-  for (let i = 0; i < command.length; i++) {
-    const char = command[i];
-    if (quote === "'") {
-      if (char === "'") quote = undefined;
-      continue;
-    }
-    if (quote === '"') {
-      if (char === '"') {
-        quote = undefined;
-      } else if (char === '\\') {
-        i++;
-      }
-      continue;
-    }
-    if (char === '"' || char === "'") {
-      quote = char;
-      continue;
-    }
-    if (char === '\\') {
-      i++;
-      continue;
-    }
-    if (char === '&' && command[i + 1] === '&') {
-      const nestedCommand = command.slice(start, i).trim();
-      if (!nestedCommand) return undefined;
-      commands.push(nestedCommand);
-      start = i + 2;
-      i++;
-    }
-  }
-
-  if (quote) return undefined;
-  const nestedCommand = command.slice(start).trim();
-  if (!nestedCommand) return undefined;
-  commands.push(nestedCommand);
-  return commands;
-}
-
-function parseShellWords(command: string): string[] | undefined {
-  const words: string[] = [];
-  let current = '';
-  let quote: '"' | "'" | undefined;
-  let hasCurrent = false;
-
-  const pushCurrent = () => {
-    if (!hasCurrent) return;
-    words.push(current);
-    current = '';
-    hasCurrent = false;
-  };
-
-  for (let i = 0; i < command.length; i++) {
-    const char = command[i];
-    if (quote === "'") {
-      if (char === "'") {
-        quote = undefined;
-      } else {
-        current += char;
-      }
-      hasCurrent = true;
-      continue;
-    }
-    if (quote === '"') {
-      if (char === '"') {
-        quote = undefined;
-      } else if (char === '\\' && i + 1 < command.length) {
-        current += command[++i];
-      } else {
-        current += char;
-      }
-      hasCurrent = true;
-      continue;
-    }
-    if (/\s/.test(char)) {
-      pushCurrent();
-      continue;
-    }
-    if (char === '"' || char === "'") {
-      quote = char;
-      hasCurrent = true;
-      continue;
-    }
-    if (char === '\\' && i + 1 < command.length) {
-      current += command[++i];
-      hasCurrent = true;
-      continue;
-    }
-    if (/[;&|<>()`$]/.test(char)) return undefined;
-    current += char;
-    hasCurrent = true;
-  }
-  if (quote) return undefined;
-  pushCurrent();
-  return words;
-}