ai-sdk-provider-codex-cli 0.1.0-ai-sdk-v4 → 0.2.0

package/README.md

@@ -4,80 +4,18 @@
 [![npm downloads](https://img.shields.io/npm/dm/ai-sdk-provider-codex-cli.svg)](https://www.npmjs.com/package/ai-sdk-provider-codex-cli)
 [![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](LICENSE)
 ![Node >= 18](https://img.shields.io/badge/node-%3E%3D18-43853d?logo=node.js&logoColor=white)
-![AI SDK v4](https://img.shields.io/badge/AI%20SDK-v4-000?logo=vercel&logoColor=white)
+![AI SDK v5](https://img.shields.io/badge/AI%20SDK-v5-000?logo=vercel&logoColor=white)
 ![Modules: ESM + CJS](https://img.shields.io/badge/modules-ESM%20%2B%20CJS-3178c6)
 ![TypeScript](https://img.shields.io/badge/TypeScript-blue)
 [![PRs welcome](https://img.shields.io/badge/PRs-welcome-brightgreen.svg)](https://github.com/ben-vargas/ai-sdk-provider-codex-cli/issues)
 [![Latest Release](https://img.shields.io/github/v/release/ben-vargas/ai-sdk-provider-codex-cli?display_name=tag)](https://github.com/ben-vargas/ai-sdk-provider-codex-cli/releases/latest)
 
-A community provider for Vercel AI SDK v4 that uses OpenAI’s Codex CLI (non‑interactive `codex exec`) to talk to GPT‑5 class models with your ChatGPT Plus/Pro subscription. The provider spawns the Codex CLI process, parses its JSONL output, and adapts it to the AI SDK LanguageModelV1 interface.
+A community provider for Vercel AI SDK v5 that uses OpenAI’s Codex CLI (non‑interactive `codex exec`) to talk to GPT‑5 class models (`gpt-5` and the Codex-specific `gpt-5-codex` slug) with your ChatGPT Plus/Pro subscription. The provider spawns the Codex CLI process, parses its JSONL output, and adapts it to the AI SDK LanguageModelV2 interface.
 
-Note: For AI SDK v5 support, see the `main` branch and install the `latest` tag from npm. This branch targets AI SDK v4 and is published under the `ai-sdk-v4` dist-tag.
-
-- Works with `generateText`, `streamText`, and `generateObject` (JSON schemas via prompt engineering)
+- Works with `generateText`, `streamText`, and `generateObject` (native JSON Schema support via `--output-schema`)
 - Uses ChatGPT OAuth from `codex login` (tokens in `~/.codex/auth.json`) or `OPENAI_API_KEY`
 - Node-only (spawns a local process); supports CI and local dev
-
-## Version Compatibility
-
-- Provider (v5 line, main branch): published as `latest`, compatible with AI SDK v5.
-- Provider (v4 line, this branch): published as `ai-sdk-v4`, compatible with AI SDK v4.
-
-| Provider Version | AI SDK Version | NPM Tag     | Status       | Branch     |
-|------------------|----------------|-------------|--------------|------------|
-| 0.1.x            | v5             | `latest`    | Active       | `main`     |
-| 0.1.0-ai-sdk-v4  | v4             | `ai-sdk-v4` | Maintenance  | `ai-sdk-v4`|
-
-### Installing the Right Version
-
-- AI SDK v5 (recommended): `npm i ai-sdk-provider-codex-cli ai` (or `@latest`)
-- AI SDK v4 (this branch): `npm i ai-sdk-provider-codex-cli@ai-sdk-v4 ai@^4.3.16`
-
-## AI SDK v4 vs v5
-
-- For AI SDK v4 (this branch):
-  - Install: `npm i ai-sdk-provider-codex-cli@ai-sdk-v4 ai@^4.3.16`
-  - Usage (non-streaming):
-    
-    ```ts
-    import { generateText } from 'ai';
-    import { codexCli } from 'ai-sdk-provider-codex-cli';
-
-    const { text } = await generateText({
-      model: codexCli('gpt-5', { allowNpx: true, skipGitRepoCheck: true }),
-      prompt: 'Hello from v4',
-    });
-    ```
-
-  - Usage (streaming):
-
-    ```ts
-    import { streamText } from 'ai';
-    import { codexCli } from 'ai-sdk-provider-codex-cli';
-
-    const { textStream } = await streamText({
-      model: codexCli('gpt-5', { allowNpx: true, skipGitRepoCheck: true }),
-      prompt: 'Stream a short reply',
-    });
-    for await (const chunk of textStream) process.stdout.write(chunk);
-    ```
-
-- For AI SDK v5 (main branch):
-  - Install: `npm i ai-sdk-provider-codex-cli ai` (or `@latest`)
-  - Usage (streaming):
-
-    ```ts
-    import { streamText } from 'ai';
-    import { codexCli } from 'ai-sdk-provider-codex-cli';
-
-    const result = streamText({
-      model: codexCli('gpt-5'),
-      prompt: 'Hello from v5',
-    });
-    const text = await result.text;
-    ```
-
-  - See the `main` branch README for full v5 docs and examples.
+- **v0.2.0 Breaking Changes**: Switched to `--experimental-json` and native schema enforcement (see [CHANGELOG](CHANGELOG.md))
 
 ## Installation
 
@@ -88,14 +26,16 @@ npm i -g @openai/codex
 codex login   # or set OPENAI_API_KEY
 ```
 
+> **⚠️ Version Requirement**: Requires Codex CLI **>= 0.42.0** for `--experimental-json` and `--output-schema` support. Check your version with `codex --version` and upgrade if needed:
+>
+> ```bash
+> npm i -g @openai/codex@latest
+> ```
+
 2. Install provider and AI SDK
 
 ```bash
-# AI SDK v4 users (this branch)
-npm i ai@^4.3.16 ai-sdk-provider-codex-cli@ai-sdk-v4
-
-# AI SDK v5 users (use main branch/latest tag)
-# npm i ai ai-sdk-provider-codex-cli
+npm i ai ai-sdk-provider-codex-cli
 ```
 
 ## Quick Start
@@ -106,7 +46,7 @@ Text generation
 import { generateText } from 'ai';
 import { codexCli } from 'ai-sdk-provider-codex-cli';
 
-const model = codexCli('gpt-5', {
+const model = codexCli('gpt-5-codex', {
   allowNpx: true,
   skipGitRepoCheck: true,
   approvalMode: 'on-failure',
@@ -126,8 +66,10 @@ Streaming
 import { streamText } from 'ai';
 import { codexCli } from 'ai-sdk-provider-codex-cli';
 
+// The provider works with both `gpt-5` and `gpt-5-codex`; use the latter for
+// the Codex CLI specific slug.
 const { textStream } = await streamText({
-  model: codexCli('gpt-5', { allowNpx: true, skipGitRepoCheck: true }),
+  model: codexCli('gpt-5-codex', { allowNpx: true, skipGitRepoCheck: true }),
   prompt: 'Write two short lines of encouragement.',
 });
 for await (const chunk of textStream) process.stdout.write(chunk);
@@ -142,7 +84,7 @@ import { codexCli } from 'ai-sdk-provider-codex-cli';
 
 const schema = z.object({ name: z.string(), age: z.number().int() });
 const { object } = await generateObject({
-  model: codexCli('gpt-5', { allowNpx: true, skipGitRepoCheck: true }),
+  model: codexCli('gpt-5-codex', { allowNpx: true, skipGitRepoCheck: true }),
   schema,
   prompt: 'Generate a small user profile.',
 });
@@ -153,22 +95,34 @@ console.log(object);
 
 - AI SDK v5 compatible (LanguageModelV2)
 - Streaming and non‑streaming
-- JSON object generation with Zod schemas (prompt‑engineered)
+- **Native JSON Schema support** via `--output-schema` (API-enforced with `strict: true`)
+- JSON object generation with Zod schemas (100-200 fewer tokens per request vs prompt engineering)
 - Safe defaults for non‑interactive automation (`on-failure`, `workspace-write`, `--skip-git-repo-check`)
 - Fallback to `npx @openai/codex` when not on PATH (`allowNpx`)
+- Usage tracking from experimental JSON event format
 
 ### Streaming behavior
 
-When using `codex exec --json`, the Codex CLI intentionally suppresses token/assistant deltas in its JSON event stream. Instead, it writes the final assistant message to the file you pass via `--output-last-message` and then signals completion. This provider:
+**Status:** Incremental streaming not currently supported with `--experimental-json` format (expected in future Codex CLI releases)
 
-- emits `response-metadata` early (as soon as the session is configured), and
-- returns the final text as a single `text-delta` right before `finish`.
+The `--experimental-json` output format (introduced Sept 25, 2025) currently only emits `item.completed` events with full text content. Incremental streaming via `item.updated` or delta events is not yet implemented by OpenAI.
 
-This is expected behavior for JSON mode in Codex exec, so streaming typically “feels” like a final chunk rather than a gradual trickle.
+**What this means:**
+- `streamText()` works functionally but delivers the entire response in a single chunk after generation completes
+- No incremental text deltas—you wait for the full response, then receive it all at once
+- The AI SDK's streaming interface is supported, but actual incremental streaming is not available
+
+**Future support:** The Codex CLI commit (344d4a1d) introducing experimental JSON explicitly notes: "or other item types like `item.output_delta` when we need streaming" and states "more event types and item types to come."
+
+When OpenAI adds streaming support, this provider will be updated to handle those events and enable true incremental streaming.
 
 ## Documentation
 
-- Getting started, configuration, and troubleshooting live in `docs/` (v4 docs coming soon in this branch).
+- Getting started, configuration, and troubleshooting live in `docs/`:
+  - [docs/ai-sdk-v5/guide.md](docs/ai-sdk-v5/guide.md) – full usage guide and examples
+  - [docs/ai-sdk-v5/configuration.md](docs/ai-sdk-v5/configuration.md) – all settings and how they map to CLI flags
+  - [docs/ai-sdk-v5/troubleshooting.md](docs/ai-sdk-v5/troubleshooting.md) – common issues and fixes
+  - [docs/ai-sdk-v5/limitations.md](docs/ai-sdk-v5/limitations.md) – known constraints and behavior differences
 - See [examples/](examples/) for runnable scripts covering core usage, streaming, permissions/sandboxing, and object generation.
 
 ## Authentication
@@ -192,15 +146,25 @@ See [docs/ai-sdk-v5/configuration.md](docs/ai-sdk-v5/configuration.md) for the f
 
 ## Zod Compatibility
 
-- Peer supports `zod@^3`
-- Validation logic normalizes zod v3/v4 error shapes
+- Peer supports `zod@^3 || ^4`
+- Validation logic normalizes v3/v4 error shapes
 
 ## Limitations
 
 - Node ≥ 18, local process only (no Edge)
-- Codex JSON mode (`codex exec --json`) suppresses mid‑response deltas; streaming typically yields a final chunk. The CLI writes the final assistant text via `--output-last-message`, which this provider reads and emits at the end.
+- Codex `--experimental-json` mode emits events rather than streaming deltas; streaming typically yields a final chunk. The CLI provides the final assistant text in the `item.completed` event, which this provider reads and emits at the end.
 - Some AI SDK parameters are unsupported by Codex CLI (e.g., temperature/topP/penalties); the provider surfaces warnings and ignores them
 
+### JSON Schema Limitations (v0.2.0+)
+
+**⚠️ Important:** OpenAI strict mode has limitations:
+
+- **Optional fields NOT supported**: All fields must be required (no `.optional()`)
+- **Format validators stripped**: `.email()`, `.url()`, `.uuid()` are removed (use descriptions instead)
+- **Pattern validators stripped**: `.regex()` is removed (use descriptions instead)
+
+See [LIMITATIONS.md](LIMITATIONS.md) for comprehensive details and migration guidance.
+
 ## Disclaimer
 
 This is a community provider and not an official OpenAI or Vercel product. You are responsible for complying with all applicable terms and ensuring safe usage.

package/dist/index.cjs

@@ -13,22 +13,6 @@ var zod = require('zod');
 var _documentCurrentScript = typeof document !== 'undefined' ? document.currentScript : null;
 // src/codex-cli-provider.ts
 
-// src/extract-json.ts
-function extractJson(text) {
-  const start = text.indexOf("{");
-  if (start === -1) return text;
-  let depth = 0;
-  for (let i = start; i < text.length; i++) {
-    const ch = text[i];
-    if (ch === "{") depth++;
-    else if (ch === "}") {
-      depth--;
-      if (depth === 0) return text.slice(start, i + 1);
-    }
-  }
-  return text;
-}
-
 // src/logger.ts
 var defaultLogger = {
   warn: (m) => console.warn(m),
@@ -108,7 +92,7 @@ function isToolItem(p) {
   if (out.type === "text" && typeof out.value !== "string") return false;
   return true;
 }
-function mapMessagesToPrompt(prompt, mode = { type: "regular" }, jsonSchema) {
+function mapMessagesToPrompt(prompt) {
   const warnings = [];
   const parts = [];
   let systemText;
@@ -151,17 +135,6 @@ function mapMessagesToPrompt(prompt, mode = { type: "regular" }, jsonSchema) {
   let promptText = "";
   if (systemText) promptText += systemText + "\n\n";
   promptText += parts.join("\n\n");
-  if (mode.type === "object-json" && jsonSchema) {
-    const schemaStr = JSON.stringify(jsonSchema, null, 2);
-    promptText = `CRITICAL: You MUST respond with ONLY a JSON object. NO other text.
-Your response MUST start with { and end with }
-The JSON MUST match this EXACT schema:
-${schemaStr}
-
-Now, based on the following conversation, generate ONLY the JSON object:
-
-${promptText}`;
-  }
   return { promptText, ...warnings.length ? { warnings } : {} };
 }
 function createAPICallError({
@@ -209,11 +182,12 @@ function resolveCodexPath(explicitPath, allowNpx) {
   }
 }
 var CodexCliLanguageModel = class {
-  specificationVersion = "v1";
+  specificationVersion = "v2";
   provider = "codex-cli";
   defaultObjectGenerationMode = "json";
   supportsImageUrls = false;
-  supportsStructuredOutputs = false;
+  supportedUrls = {};
+  supportsStructuredOutputs = true;
   modelId;
   settings;
   logger;
@@ -228,9 +202,9 @@ var CodexCliLanguageModel = class {
     const warn = validateModelId(this.modelId);
     if (warn) this.logger.warn(`Codex CLI model: ${warn}`);
   }
-  buildArgs(promptText) {
+  buildArgs(promptText, responseFormat) {
     const base = resolveCodexPath(this.settings.codexPath, this.settings.allowNpx);
-    const args = [...base.args, "exec", "--json"];
+    const args = [...base.args, "exec", "--experimental-json"];
     if (this.settings.fullAuto) {
       args.push("--full-auto");
     } else if (this.settings.dangerouslyBypassApprovalsAndSandbox) {
@@ -250,6 +224,20 @@ var CodexCliLanguageModel = class {
     if (this.modelId) {
       args.push("-m", this.modelId);
     }
+    let schemaPath;
+    if (responseFormat?.type === "json" && responseFormat.schema) {
+      const dir = fs.mkdtempSync(path.join(os.tmpdir(), "codex-schema-"));
+      schemaPath = path.join(dir, "schema.json");
+      const schema = typeof responseFormat.schema === "object" ? responseFormat.schema : {};
+      const sanitizedSchema = this.sanitizeJsonSchema(schema);
+      const schemaWithAdditional = {
+        ...sanitizedSchema,
+        // OpenAI strict mode requires additionalProperties=false even if user requested otherwise.
+        additionalProperties: false
+      };
+      fs.writeFileSync(schemaPath, JSON.stringify(schemaWithAdditional, null, 2));
+      args.push("--output-schema", schemaPath);
+    }
     args.push(promptText);
     const env = {
       ...process.env,
@@ -262,7 +250,34 @@ var CodexCliLanguageModel = class {
       lastMessagePath = path.join(dir, "last-message.txt");
     }
     args.push("--output-last-message", lastMessagePath);
-    return { cmd: base.cmd, args, env, cwd: this.settings.cwd, lastMessagePath };
+    return { cmd: base.cmd, args, env, cwd: this.settings.cwd, lastMessagePath, schemaPath };
+  }
+  sanitizeJsonSchema(value) {
+    if (typeof value !== "object" || value === null) {
+      return value;
+    }
+    if (Array.isArray(value)) {
+      return value.map((item) => this.sanitizeJsonSchema(item));
+    }
+    const obj = value;
+    const result = {};
+    for (const [key, val] of Object.entries(obj)) {
+      if (key === "properties" && typeof val === "object" && val !== null && !Array.isArray(val)) {
+        const props = val;
+        const sanitizedProps = {};
+        for (const [propName, propSchema] of Object.entries(props)) {
+          sanitizedProps[propName] = this.sanitizeJsonSchema(propSchema);
+        }
+        result[key] = sanitizedProps;
+        continue;
+      }
+      if (key === "$schema" || key === "$id" || key === "$ref" || key === "$defs" || key === "definitions" || key === "title" || key === "examples" || key === "default" || key === "format" || // OpenAI strict mode doesn't support format
+      key === "pattern") {
+        continue;
+      }
+      result[key] = this.sanitizeJsonSchema(val);
+    }
+    return result;
   }
   mapWarnings(options) {
     const unsupported = [];
@@ -275,7 +290,6 @@ var CodexCliLanguageModel = class {
         });
     };
     add(options.temperature, "temperature");
-    add(options.maxTokens, "maxTokens");
     add(options.topP, "topP");
     add(options.topK, "topK");
     add(options.presencePenalty, "presencePenalty");
@@ -284,11 +298,34 @@ var CodexCliLanguageModel = class {
     add(options.seed, "seed");
     return unsupported;
   }
-  parseJsonLine(line) {
+  parseExperimentalJsonEvent(line) {
     try {
-      return JSON.parse(line);
+      const evt = JSON.parse(line);
+      const result = {};
+      switch (evt.type) {
+        case "session.created":
+          result.sessionId = evt.session_id;
+          break;
+        case "turn.completed":
+          if (evt.usage) {
+            result.usage = {
+              inputTokens: evt.usage.input_tokens || 0,
+              outputTokens: evt.usage.output_tokens || 0
+            };
+          }
+          break;
+        case "item.completed":
+          if (evt.item?.item_type === "assistant_message" || evt.item?.item_type === "reasoning") {
+            result.text = evt.item.text;
+          }
+          break;
+        case "error":
+          result.error = evt.message;
+          break;
+      }
+      return result;
     } catch {
-      return void 0;
+      return {};
     }
   }
   handleSpawnError(err, promptExcerpt) {
@@ -306,20 +343,19 @@ var CodexCliLanguageModel = class {
     });
   }
   async doGenerate(options) {
-    const mode = options.mode?.type === "object-json" ? { type: "object-json" } : { type: "regular" };
-    const { promptText, warnings: mappingWarnings } = mapMessagesToPrompt(
-      options.prompt,
-      mode,
-      options.mode?.type === "object-json" ? options.mode.schema : void 0
-    );
+    const { promptText, warnings: mappingWarnings } = mapMessagesToPrompt(options.prompt);
     const promptExcerpt = promptText.slice(0, 200);
     const warnings = [
       ...this.mapWarnings(options),
       ...mappingWarnings?.map((m) => ({ type: "other", message: m })) || []
     ];
-    const { cmd, args, env, cwd, lastMessagePath } = this.buildArgs(promptText);
+    const responseFormat = options.responseFormat?.type === "json" ? { type: "json", schema: options.responseFormat.schema } : void 0;
+    const { cmd, args, env, cwd, lastMessagePath, schemaPath } = this.buildArgs(
+      promptText,
+      responseFormat
+    );
     let text = "";
-    const usage = { promptTokens: 0, completionTokens: 0 };
+    const usage = { inputTokens: 0, outputTokens: 0, totalTokens: 0 };
     const finishReason = "stop";
     const child = child_process.spawn(cmd, args, { env, cwd, stdio: ["ignore", "pipe", "pipe"] });
     let onAbort;
@@ -339,15 +375,13 @@ var CodexCliLanguageModel = class {
         child.stdout.on("data", (chunk) => {
           const lines = chunk.split(/\r?\n/).filter(Boolean);
           for (const line of lines) {
-            const evt = this.parseJsonLine(line);
-            if (!evt) continue;
-            const msg = evt.msg;
-            const type = msg?.type;
-            if (type === "session_configured" && msg) {
-              this.sessionId = msg.session_id;
-            } else if (type === "task_complete" && msg) {
-              const last = msg.last_agent_message;
-              if (typeof last === "string") text = last;
+            const parsed = this.parseExperimentalJsonEvent(line);
+            if (parsed.sessionId) this.sessionId = parsed.sessionId;
+            if (parsed.text) text = parsed.text;
+            if (parsed.usage) {
+              usage.inputTokens = parsed.usage.inputTokens;
+              usage.outputTokens = parsed.usage.outputTokens;
+              usage.totalTokens = usage.inputTokens + usage.outputTokens;
             }
           }
         });
@@ -367,6 +401,13 @@ var CodexCliLanguageModel = class {
       });
     } finally {
       if (options.abortSignal && onAbort) options.abortSignal.removeEventListener("abort", onAbort);
+      if (schemaPath) {
+        try {
+          const schemaDir = path.dirname(schemaPath);
+          fs.rmSync(schemaDir, { recursive: true, force: true });
+        } catch {
+        }
+      }
     }
     if (!text && lastMessagePath) {
       try {
@@ -381,38 +422,35 @@ var CodexCliLanguageModel = class {
       } catch {
       }
     }
-    if (options.mode?.type === "object-json" && text) {
-      text = extractJson(text);
-    }
+    const content = [{ type: "text", text }];
     return {
-      text: text || void 0,
+      content,
       usage,
       finishReason,
-      warnings: warnings.length ? warnings : void 0,
+      warnings,
       response: { id: providerUtils.generateId(), timestamp: /* @__PURE__ */ new Date(), modelId: this.modelId },
       request: { body: promptText },
       providerMetadata: {
         "codex-cli": { ...this.sessionId ? { sessionId: this.sessionId } : {} }
-      },
-      rawCall: { rawPrompt: promptText, rawSettings: { model: this.modelId, settings: this.settings } }
+      }
     };
   }
   async doStream(options) {
-    const mode = options.mode?.type === "object-json" ? { type: "object-json" } : { type: "regular" };
-    const { promptText, warnings: mappingWarnings } = mapMessagesToPrompt(
-      options.prompt,
-      mode,
-      options.mode?.type === "object-json" ? options.mode.schema : void 0
-    );
+    const { promptText, warnings: mappingWarnings } = mapMessagesToPrompt(options.prompt);
     const promptExcerpt = promptText.slice(0, 200);
     const warnings = [
       ...this.mapWarnings(options),
       ...mappingWarnings?.map((m) => ({ type: "other", message: m })) || []
     ];
-    const { cmd, args, env, cwd, lastMessagePath } = this.buildArgs(promptText);
+    const responseFormat = options.responseFormat?.type === "json" ? { type: "json", schema: options.responseFormat.schema } : void 0;
+    const { cmd, args, env, cwd, lastMessagePath, schemaPath } = this.buildArgs(
+      promptText,
+      responseFormat
+    );
     const stream = new ReadableStream({
       start: (controller) => {
         const child = child_process.spawn(cmd, args, { env, cwd, stdio: ["ignore", "pipe", "pipe"] });
+        controller.enqueue({ type: "stream-start", warnings });
         let stderr = "";
         let accumulatedText = "";
         const onAbort = () => {
@@ -431,32 +469,37 @@ var CodexCliLanguageModel = class {
         child.stdout.on("data", (chunk) => {
           const lines = chunk.split(/\r?\n/).filter(Boolean);
           for (const line of lines) {
-            const evt = this.parseJsonLine(line);
-            if (!evt) continue;
-            const msg = evt.msg;
-            const type = msg?.type;
-            if (type === "session_configured" && msg) {
-              this.sessionId = msg.session_id;
+            const parsed = this.parseExperimentalJsonEvent(line);
+            if (parsed.sessionId) {
+              this.sessionId = parsed.sessionId;
               controller.enqueue({
                 type: "response-metadata",
                 id: crypto.randomUUID(),
                 timestamp: /* @__PURE__ */ new Date(),
                 modelId: this.modelId
               });
-            } else if (type === "task_complete" && msg) {
-              const last = msg.last_agent_message;
-              if (typeof last === "string") {
-                accumulatedText = last;
-              }
+            }
+            if (parsed.text) {
+              accumulatedText = parsed.text;
             }
           }
         });
+        const cleanupSchema = () => {
+          if (!schemaPath) return;
+          try {
+            const schemaDir = path.dirname(schemaPath);
+            fs.rmSync(schemaDir, { recursive: true, force: true });
+          } catch {
+          }
+        };
         child.on("error", (e) => {
           if (options.abortSignal) options.abortSignal.removeEventListener("abort", onAbort);
+          cleanupSchema();
           controller.error(this.handleSpawnError(e, promptExcerpt));
         });
         child.on("close", (code) => {
           if (options.abortSignal) options.abortSignal.removeEventListener("abort", onAbort);
+          cleanupSchema();
           if (code !== 0) {
             controller.error(
               createAPICallError({
@@ -481,18 +524,12 @@ var CodexCliLanguageModel = class {
             }
           }
           if (finalText) {
-            if (options.mode?.type === "object-json") {
-              finalText = extractJson(finalText);
-            }
-            controller.enqueue({ type: "text-delta", textDelta: finalText });
+            controller.enqueue({ type: "text-delta", id: crypto.randomUUID(), delta: finalText });
           }
           controller.enqueue({
             type: "finish",
             finishReason: "stop",
-            usage: { promptTokens: 0, completionTokens: 0 },
-            providerMetadata: {
-              "codex-cli": { ...this.sessionId ? { sessionId: this.sessionId } : {} }
-            }
+            usage: { inputTokens: 0, outputTokens: 0, totalTokens: 0 }
           });
           controller.close();
         });
@@ -500,11 +537,7 @@ var CodexCliLanguageModel = class {
       cancel: () => {
       }
     });
-    return {
-      stream,
-      rawCall: { rawPrompt: promptText, rawSettings: { model: this.modelId, settings: this.settings } },
-      warnings: warnings.length ? warnings : void 0
-    };
+    return { stream, request: { body: promptText } };
   }
 };
 

package/dist/index.d.cts

@@ -1,4 +1,4 @@
-import { ProviderV1, LanguageModelV1 } from '@ai-sdk/provider';
+import { ProviderV2, LanguageModelV2 } from '@ai-sdk/provider';
 
 interface Logger {
     warn: (message: string) => void;
@@ -25,10 +25,10 @@ interface CodexCliProviderSettings {
     defaultSettings?: CodexCliSettings;
 }
 
-interface CodexCliProvider extends ProviderV1 {
-    (modelId: string, settings?: CodexCliSettings): LanguageModelV1;
-    languageModel(modelId: string, settings?: CodexCliSettings): LanguageModelV1;
-    chat(modelId: string, settings?: CodexCliSettings): LanguageModelV1;
+interface CodexCliProvider extends ProviderV2 {
+    (modelId: string, settings?: CodexCliSettings): LanguageModelV2;
+    languageModel(modelId: string, settings?: CodexCliSettings): LanguageModelV2;
+    chat(modelId: string, settings?: CodexCliSettings): LanguageModelV2;
     textEmbeddingModel(modelId: string): never;
     imageModel(modelId: string): never;
 }
@@ -39,23 +39,25 @@ interface CodexLanguageModelOptions {
     id: string;
     settings?: CodexCliSettings;
 }
-declare class CodexCliLanguageModel implements LanguageModelV1 {
-    readonly specificationVersion: "v1";
+declare class CodexCliLanguageModel implements LanguageModelV2 {
+    readonly specificationVersion: "v2";
     readonly provider = "codex-cli";
     readonly defaultObjectGenerationMode: "json";
     readonly supportsImageUrls = false;
-    readonly supportsStructuredOutputs = false;
+    readonly supportedUrls: {};
+    readonly supportsStructuredOutputs = true;
     readonly modelId: string;
     readonly settings: CodexCliSettings;
     private logger;
     private sessionId?;
     constructor(options: CodexLanguageModelOptions);
     private buildArgs;
+    private sanitizeJsonSchema;
     private mapWarnings;
-    private parseJsonLine;
+    private parseExperimentalJsonEvent;
     private handleSpawnError;
-    doGenerate(options: Parameters<LanguageModelV1['doGenerate']>[0]): Promise<Awaited<ReturnType<LanguageModelV1['doGenerate']>>>;
-    doStream(options: Parameters<LanguageModelV1['doStream']>[0]): Promise<Awaited<ReturnType<LanguageModelV1['doStream']>>>;
+    doGenerate(options: Parameters<LanguageModelV2['doGenerate']>[0]): Promise<Awaited<ReturnType<LanguageModelV2['doGenerate']>>>;
+    doStream(options: Parameters<LanguageModelV2['doStream']>[0]): Promise<Awaited<ReturnType<LanguageModelV2['doStream']>>>;
 }
 
 declare function isAuthenticationError(err: unknown): boolean;

package/dist/index.d.ts

@@ -1,4 +1,4 @@
-import { ProviderV1, LanguageModelV1 } from '@ai-sdk/provider';
+import { ProviderV2, LanguageModelV2 } from '@ai-sdk/provider';
 
 interface Logger {
     warn: (message: string) => void;
@@ -25,10 +25,10 @@ interface CodexCliProviderSettings {
     defaultSettings?: CodexCliSettings;
 }
 
-interface CodexCliProvider extends ProviderV1 {
-    (modelId: string, settings?: CodexCliSettings): LanguageModelV1;
-    languageModel(modelId: string, settings?: CodexCliSettings): LanguageModelV1;
-    chat(modelId: string, settings?: CodexCliSettings): LanguageModelV1;
+interface CodexCliProvider extends ProviderV2 {
+    (modelId: string, settings?: CodexCliSettings): LanguageModelV2;
+    languageModel(modelId: string, settings?: CodexCliSettings): LanguageModelV2;
+    chat(modelId: string, settings?: CodexCliSettings): LanguageModelV2;
     textEmbeddingModel(modelId: string): never;
     imageModel(modelId: string): never;
 }
@@ -39,23 +39,25 @@ interface CodexLanguageModelOptions {
     id: string;
     settings?: CodexCliSettings;
 }
-declare class CodexCliLanguageModel implements LanguageModelV1 {
-    readonly specificationVersion: "v1";
+declare class CodexCliLanguageModel implements LanguageModelV2 {
+    readonly specificationVersion: "v2";
     readonly provider = "codex-cli";
     readonly defaultObjectGenerationMode: "json";
     readonly supportsImageUrls = false;
-    readonly supportsStructuredOutputs = false;
+    readonly supportedUrls: {};
+    readonly supportsStructuredOutputs = true;
     readonly modelId: string;
     readonly settings: CodexCliSettings;
     private logger;
     private sessionId?;
     constructor(options: CodexLanguageModelOptions);
     private buildArgs;
+    private sanitizeJsonSchema;
     private mapWarnings;
-    private parseJsonLine;
+    private parseExperimentalJsonEvent;
     private handleSpawnError;
-    doGenerate(options: Parameters<LanguageModelV1['doGenerate']>[0]): Promise<Awaited<ReturnType<LanguageModelV1['doGenerate']>>>;
-    doStream(options: Parameters<LanguageModelV1['doStream']>[0]): Promise<Awaited<ReturnType<LanguageModelV1['doStream']>>>;
+    doGenerate(options: Parameters<LanguageModelV2['doGenerate']>[0]): Promise<Awaited<ReturnType<LanguageModelV2['doGenerate']>>>;
+    doStream(options: Parameters<LanguageModelV2['doStream']>[0]): Promise<Awaited<ReturnType<LanguageModelV2['doStream']>>>;
 }
 
 declare function isAuthenticationError(err: unknown): boolean;

package/dist/index.js

@@ -2,30 +2,14 @@ import { NoSuchModelError, LoadAPIKeyError, APICallError } from '@ai-sdk/provide
 import { spawn } from 'child_process';
 import { randomUUID } from 'crypto';
 import { createRequire } from 'module';
-import { mkdtempSync, readFileSync, rmSync } from 'fs';
+import { mkdtempSync, writeFileSync, rmSync, readFileSync } from 'fs';
 import { tmpdir } from 'os';
-import { join } from 'path';
+import { join, dirname } from 'path';
 import { generateId } from '@ai-sdk/provider-utils';
 import { z } from 'zod';
 
 // src/codex-cli-provider.ts
 
-// src/extract-json.ts
-function extractJson(text) {
-  const start = text.indexOf("{");
-  if (start === -1) return text;
-  let depth = 0;
-  for (let i = start; i < text.length; i++) {
-    const ch = text[i];
-    if (ch === "{") depth++;
-    else if (ch === "}") {
-      depth--;
-      if (depth === 0) return text.slice(start, i + 1);
-    }
-  }
-  return text;
-}
-
 // src/logger.ts
 var defaultLogger = {
   warn: (m) => console.warn(m),
@@ -105,7 +89,7 @@ function isToolItem(p) {
   if (out.type === "text" && typeof out.value !== "string") return false;
   return true;
 }
-function mapMessagesToPrompt(prompt, mode = { type: "regular" }, jsonSchema) {
+function mapMessagesToPrompt(prompt) {
   const warnings = [];
   const parts = [];
   let systemText;
@@ -148,17 +132,6 @@ function mapMessagesToPrompt(prompt, mode = { type: "regular" }, jsonSchema) {
   let promptText = "";
   if (systemText) promptText += systemText + "\n\n";
   promptText += parts.join("\n\n");
-  if (mode.type === "object-json" && jsonSchema) {
-    const schemaStr = JSON.stringify(jsonSchema, null, 2);
-    promptText = `CRITICAL: You MUST respond with ONLY a JSON object. NO other text.
-Your response MUST start with { and end with }
-The JSON MUST match this EXACT schema:
-${schemaStr}
-
-Now, based on the following conversation, generate ONLY the JSON object:
-
-${promptText}`;
-  }
   return { promptText, ...warnings.length ? { warnings } : {} };
 }
 function createAPICallError({
@@ -206,11 +179,12 @@ function resolveCodexPath(explicitPath, allowNpx) {
   }
 }
 var CodexCliLanguageModel = class {
-  specificationVersion = "v1";
+  specificationVersion = "v2";
   provider = "codex-cli";
   defaultObjectGenerationMode = "json";
   supportsImageUrls = false;
-  supportsStructuredOutputs = false;
+  supportedUrls = {};
+  supportsStructuredOutputs = true;
   modelId;
   settings;
   logger;
@@ -225,9 +199,9 @@ var CodexCliLanguageModel = class {
     const warn = validateModelId(this.modelId);
     if (warn) this.logger.warn(`Codex CLI model: ${warn}`);
   }
-  buildArgs(promptText) {
+  buildArgs(promptText, responseFormat) {
     const base = resolveCodexPath(this.settings.codexPath, this.settings.allowNpx);
-    const args = [...base.args, "exec", "--json"];
+    const args = [...base.args, "exec", "--experimental-json"];
     if (this.settings.fullAuto) {
       args.push("--full-auto");
     } else if (this.settings.dangerouslyBypassApprovalsAndSandbox) {
@@ -247,6 +221,20 @@ var CodexCliLanguageModel = class {
     if (this.modelId) {
       args.push("-m", this.modelId);
     }
+    let schemaPath;
+    if (responseFormat?.type === "json" && responseFormat.schema) {
+      const dir = mkdtempSync(join(tmpdir(), "codex-schema-"));
+      schemaPath = join(dir, "schema.json");
+      const schema = typeof responseFormat.schema === "object" ? responseFormat.schema : {};
+      const sanitizedSchema = this.sanitizeJsonSchema(schema);
+      const schemaWithAdditional = {
+        ...sanitizedSchema,
+        // OpenAI strict mode requires additionalProperties=false even if user requested otherwise.
+        additionalProperties: false
+      };
+      writeFileSync(schemaPath, JSON.stringify(schemaWithAdditional, null, 2));
+      args.push("--output-schema", schemaPath);
+    }
     args.push(promptText);
     const env = {
       ...process.env,
@@ -259,7 +247,34 @@ var CodexCliLanguageModel = class {
       lastMessagePath = join(dir, "last-message.txt");
     }
     args.push("--output-last-message", lastMessagePath);
-    return { cmd: base.cmd, args, env, cwd: this.settings.cwd, lastMessagePath };
+    return { cmd: base.cmd, args, env, cwd: this.settings.cwd, lastMessagePath, schemaPath };
+  }
+  sanitizeJsonSchema(value) {
+    if (typeof value !== "object" || value === null) {
+      return value;
+    }
+    if (Array.isArray(value)) {
+      return value.map((item) => this.sanitizeJsonSchema(item));
+    }
+    const obj = value;
+    const result = {};
+    for (const [key, val] of Object.entries(obj)) {
+      if (key === "properties" && typeof val === "object" && val !== null && !Array.isArray(val)) {
+        const props = val;
+        const sanitizedProps = {};
+        for (const [propName, propSchema] of Object.entries(props)) {
+          sanitizedProps[propName] = this.sanitizeJsonSchema(propSchema);
+        }
+        result[key] = sanitizedProps;
+        continue;
+      }
+      if (key === "$schema" || key === "$id" || key === "$ref" || key === "$defs" || key === "definitions" || key === "title" || key === "examples" || key === "default" || key === "format" || // OpenAI strict mode doesn't support format
+      key === "pattern") {
+        continue;
+      }
+      result[key] = this.sanitizeJsonSchema(val);
+    }
+    return result;
   }
   mapWarnings(options) {
     const unsupported = [];
@@ -272,7 +287,6 @@ var CodexCliLanguageModel = class {
         });
     };
     add(options.temperature, "temperature");
-    add(options.maxTokens, "maxTokens");
     add(options.topP, "topP");
     add(options.topK, "topK");
     add(options.presencePenalty, "presencePenalty");
@@ -281,11 +295,34 @@ var CodexCliLanguageModel = class {
     add(options.seed, "seed");
     return unsupported;
   }
-  parseJsonLine(line) {
+  parseExperimentalJsonEvent(line) {
     try {
-      return JSON.parse(line);
+      const evt = JSON.parse(line);
+      const result = {};
+      switch (evt.type) {
+        case "session.created":
+          result.sessionId = evt.session_id;
+          break;
+        case "turn.completed":
+          if (evt.usage) {
+            result.usage = {
+              inputTokens: evt.usage.input_tokens || 0,
+              outputTokens: evt.usage.output_tokens || 0
+            };
+          }
+          break;
+        case "item.completed":
+          if (evt.item?.item_type === "assistant_message" || evt.item?.item_type === "reasoning") {
+            result.text = evt.item.text;
+          }
+          break;
+        case "error":
+          result.error = evt.message;
+          break;
+      }
+      return result;
     } catch {
-      return void 0;
+      return {};
     }
   }
   handleSpawnError(err, promptExcerpt) {
@@ -303,20 +340,19 @@ var CodexCliLanguageModel = class {
     });
   }
   async doGenerate(options) {
-    const mode = options.mode?.type === "object-json" ? { type: "object-json" } : { type: "regular" };
-    const { promptText, warnings: mappingWarnings } = mapMessagesToPrompt(
-      options.prompt,
-      mode,
-      options.mode?.type === "object-json" ? options.mode.schema : void 0
-    );
+    const { promptText, warnings: mappingWarnings } = mapMessagesToPrompt(options.prompt);
     const promptExcerpt = promptText.slice(0, 200);
     const warnings = [
       ...this.mapWarnings(options),
       ...mappingWarnings?.map((m) => ({ type: "other", message: m })) || []
     ];
-    const { cmd, args, env, cwd, lastMessagePath } = this.buildArgs(promptText);
+    const responseFormat = options.responseFormat?.type === "json" ? { type: "json", schema: options.responseFormat.schema } : void 0;
+    const { cmd, args, env, cwd, lastMessagePath, schemaPath } = this.buildArgs(
+      promptText,
+      responseFormat
+    );
     let text = "";
-    const usage = { promptTokens: 0, completionTokens: 0 };
+    const usage = { inputTokens: 0, outputTokens: 0, totalTokens: 0 };
     const finishReason = "stop";
     const child = spawn(cmd, args, { env, cwd, stdio: ["ignore", "pipe", "pipe"] });
     let onAbort;
@@ -336,15 +372,13 @@ var CodexCliLanguageModel = class {
         child.stdout.on("data", (chunk) => {
           const lines = chunk.split(/\r?\n/).filter(Boolean);
           for (const line of lines) {
-            const evt = this.parseJsonLine(line);
-            if (!evt) continue;
-            const msg = evt.msg;
-            const type = msg?.type;
-            if (type === "session_configured" && msg) {
-              this.sessionId = msg.session_id;
-            } else if (type === "task_complete" && msg) {
-              const last = msg.last_agent_message;
-              if (typeof last === "string") text = last;
+            const parsed = this.parseExperimentalJsonEvent(line);
+            if (parsed.sessionId) this.sessionId = parsed.sessionId;
+            if (parsed.text) text = parsed.text;
+            if (parsed.usage) {
+              usage.inputTokens = parsed.usage.inputTokens;
+              usage.outputTokens = parsed.usage.outputTokens;
+              usage.totalTokens = usage.inputTokens + usage.outputTokens;
             }
           }
         });
@@ -364,6 +398,13 @@ var CodexCliLanguageModel = class {
       });
     } finally {
       if (options.abortSignal && onAbort) options.abortSignal.removeEventListener("abort", onAbort);
+      if (schemaPath) {
+        try {
+          const schemaDir = dirname(schemaPath);
+          rmSync(schemaDir, { recursive: true, force: true });
+        } catch {
+        }
+      }
     }
     if (!text && lastMessagePath) {
       try {
@@ -378,38 +419,35 @@ var CodexCliLanguageModel = class {
       } catch {
       }
     }
-    if (options.mode?.type === "object-json" && text) {
-      text = extractJson(text);
-    }
+    const content = [{ type: "text", text }];
     return {
-      text: text || void 0,
+      content,
       usage,
       finishReason,
-      warnings: warnings.length ? warnings : void 0,
+      warnings,
       response: { id: generateId(), timestamp: /* @__PURE__ */ new Date(), modelId: this.modelId },
       request: { body: promptText },
       providerMetadata: {
         "codex-cli": { ...this.sessionId ? { sessionId: this.sessionId } : {} }
-      },
-      rawCall: { rawPrompt: promptText, rawSettings: { model: this.modelId, settings: this.settings } }
+      }
     };
   }
   async doStream(options) {
-    const mode = options.mode?.type === "object-json" ? { type: "object-json" } : { type: "regular" };
-    const { promptText, warnings: mappingWarnings } = mapMessagesToPrompt(
-      options.prompt,
-      mode,
-      options.mode?.type === "object-json" ? options.mode.schema : void 0
-    );
+    const { promptText, warnings: mappingWarnings } = mapMessagesToPrompt(options.prompt);
     const promptExcerpt = promptText.slice(0, 200);
     const warnings = [
       ...this.mapWarnings(options),
       ...mappingWarnings?.map((m) => ({ type: "other", message: m })) || []
     ];
-    const { cmd, args, env, cwd, lastMessagePath } = this.buildArgs(promptText);
+    const responseFormat = options.responseFormat?.type === "json" ? { type: "json", schema: options.responseFormat.schema } : void 0;
+    const { cmd, args, env, cwd, lastMessagePath, schemaPath } = this.buildArgs(
+      promptText,
+      responseFormat
+    );
     const stream = new ReadableStream({
       start: (controller) => {
         const child = spawn(cmd, args, { env, cwd, stdio: ["ignore", "pipe", "pipe"] });
+        controller.enqueue({ type: "stream-start", warnings });
         let stderr = "";
         let accumulatedText = "";
         const onAbort = () => {
@@ -428,32 +466,37 @@ var CodexCliLanguageModel = class {
         child.stdout.on("data", (chunk) => {
           const lines = chunk.split(/\r?\n/).filter(Boolean);
           for (const line of lines) {
-            const evt = this.parseJsonLine(line);
-            if (!evt) continue;
-            const msg = evt.msg;
-            const type = msg?.type;
-            if (type === "session_configured" && msg) {
-              this.sessionId = msg.session_id;
+            const parsed = this.parseExperimentalJsonEvent(line);
+            if (parsed.sessionId) {
+              this.sessionId = parsed.sessionId;
               controller.enqueue({
                 type: "response-metadata",
                 id: randomUUID(),
                 timestamp: /* @__PURE__ */ new Date(),
                 modelId: this.modelId
               });
-            } else if (type === "task_complete" && msg) {
-              const last = msg.last_agent_message;
-              if (typeof last === "string") {
-                accumulatedText = last;
-              }
+            }
+            if (parsed.text) {
+              accumulatedText = parsed.text;
             }
           }
         });
+        const cleanupSchema = () => {
+          if (!schemaPath) return;
+          try {
+            const schemaDir = dirname(schemaPath);
+            rmSync(schemaDir, { recursive: true, force: true });
+          } catch {
+          }
+        };
         child.on("error", (e) => {
           if (options.abortSignal) options.abortSignal.removeEventListener("abort", onAbort);
+          cleanupSchema();
           controller.error(this.handleSpawnError(e, promptExcerpt));
         });
         child.on("close", (code) => {
           if (options.abortSignal) options.abortSignal.removeEventListener("abort", onAbort);
+          cleanupSchema();
           if (code !== 0) {
             controller.error(
               createAPICallError({
@@ -478,18 +521,12 @@ var CodexCliLanguageModel = class {
             }
           }
           if (finalText) {
-            if (options.mode?.type === "object-json") {
-              finalText = extractJson(finalText);
-            }
-            controller.enqueue({ type: "text-delta", textDelta: finalText });
+            controller.enqueue({ type: "text-delta", id: randomUUID(), delta: finalText });
           }
           controller.enqueue({
             type: "finish",
             finishReason: "stop",
-            usage: { promptTokens: 0, completionTokens: 0 },
-            providerMetadata: {
-              "codex-cli": { ...this.sessionId ? { sessionId: this.sessionId } : {} }
-            }
+            usage: { inputTokens: 0, outputTokens: 0, totalTokens: 0 }
           });
           controller.close();
         });
@@ -497,11 +534,7 @@ var CodexCliLanguageModel = class {
       cancel: () => {
       }
     });
-    return {
-      stream,
-      rawCall: { rawPrompt: promptText, rawSettings: { model: this.modelId, settings: this.settings } },
-      warnings: warnings.length ? warnings : void 0
-    };
+    return { stream, request: { body: promptText } };
   }
 };
 

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "ai-sdk-provider-codex-cli",
-  "version": "0.1.0-ai-sdk-v4",
-  "description": "AI SDK v4 provider for spawning OpenAI Codex CLI (gpt-5 via ChatGPT OAuth)",
+  "version": "0.2.0",
+  "description": "AI SDK v5 provider for OpenAI Codex CLI with native JSON Schema support",
   "keywords": [
     "ai-sdk",
     "codex",
@@ -56,8 +56,8 @@
     "test:coverage": "vitest run --coverage"
   },
   "dependencies": {
-    "@ai-sdk/provider": "1.1.3",
-    "@ai-sdk/provider-utils": "2.2.8",
+    "@ai-sdk/provider": "2.0.0",
+    "@ai-sdk/provider-utils": "3.0.3",
     "jsonc-parser": "^3.3.1"
   },
   "optionalDependencies": {
@@ -67,17 +67,17 @@
     "@eslint/js": "^9.14.0",
     "@types/node": "20.17.24",
     "@vitest/coverage-v8": "^3.2.4",
-    "ai": "4.3.16",
+    "ai": "5.0.14",
     "eslint": "^9.14.0",
     "prettier": "^3.3.3",
     "tsup": "8.5.0",
     "typescript": "5.6.3",
     "vitest": "^3.2.4",
     "typescript-eslint": "^8.6.0",
-    "zod": "3.24.1"
+    "zod": "^4.0.17"
   },
   "peerDependencies": {
-    "zod": "^3.0.0"
+    "zod": "^3.0.0 || ^4.0.0"
   },
   "engines": {
     "node": ">=18"