fix: v4.0.2 - Fix compaction context loss and agent creation

**Fixed**:
- Compaction losing context: Only filter orphaned function_call_output items, preserve matched pairs
- Agent creation failing: Properly detect streaming vs non-streaming requests
- SSE/JSON response handling: Convert SSE→JSON for generateText(), passthrough for streamText()

**Added**:
- gpt-5.1-chat-latest model support (normalizes to gpt-5.1)

**Technical**:
- Capture original stream value before transformation
- API always gets stream=true, but response handling based on original intent
- Orphan detection: only remove function_call_output without matching function_call

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <[email protected]>

Author: numman-ali

CHANGELOG.md

@@ -2,6 +2,23 @@
 
 All notable changes to this project are documented here. Dates use the ISO format (YYYY-MM-DD).
 
+## [4.0.2] - 2025-11-27
+
+**Bugfix release**: Fixes compaction context loss, agent creation, and SSE/JSON response handling.
+
+### Fixed
+- **Compaction losing context**: v4.0.1 was too aggressive in filtering tool calls - it removed ALL `function_call`/`function_call_output` items when tools weren't present. Now only **orphaned** outputs (without matching calls) are filtered, preserving matched pairs for compaction context.
+- **Agent creation failing**: The `/agent create` command was failing with "Invalid JSON response" because we were returning SSE streams instead of JSON for `generateText()` requests.
+- **SSE/JSON response handling**: Properly detect original request intent - `streamText()` requests get SSE passthrough, `generateText()` requests get SSE→JSON conversion.
+
+### Added
+- **`gpt-5.1-chat-latest` model support**: Added to model map, normalizes to `gpt-5.1`.
+
+### Technical Details
+- Root cause of compaction issue: OpenCode sends `item_reference` with `fc_*` IDs for function calls. We filter these for stateless mode, but v4.0.1 then removed ALL tool items. Now we only remove orphaned `function_call_output` items (where no matching `function_call` exists).
+- Root cause of agent creation issue: We were forcing `stream: true` for all requests and returning SSE for all responses. Now we capture original `stream` value before transformation and convert SSE→JSON only when original request wasn't streaming.
+- The Codex API always receives `stream: true` (required), but response handling is based on original intent.
+
 ## [4.0.1] - 2025-11-27
 
 **Bugfix release**: Fixes API errors during summary/compaction and GitHub rate limiting.

README.md

@@ -60,7 +60,7 @@ Follow me on [X @nummanthinks](https://x.com/nummanthinks) for future updates an
 #### Recommended: Pin the Version
 
 ```json
-"plugin": ["[email protected].1"]
+"plugin": ["[email protected].2"]
 ```
 
 **Why pin versions?** OpenCode uses Bun's lockfile which pins resolved versions. If you use `"opencode-openai-codex-auth"` without a version, it resolves to "latest" once and **never updates** even when new versions are published.
@@ -74,7 +74,7 @@ Simply change the version in your config and restart OpenCode:
 "plugin": ["[email protected]"]
 
 // To:
-"plugin": ["[email protected].1"]
+"plugin": ["[email protected].2"]
 ```
 
 OpenCode will detect the version mismatch and install the new version automatically.
@@ -108,7 +108,7 @@ Check [releases](https://github.com/numman-ali/opencode-openai-codex-auth/releas
 {
   "$schema": "https://opencode.ai/config.json",
   "plugin": [
-    "[email protected].1"
+    "[email protected].2"
   ],
   "provider": {
     "openai": {
@@ -517,7 +517,7 @@ Apply settings to all models:
 ```json
 {
   "$schema": "https://opencode.ai/config.json",
-  "plugin": ["[email protected].1"],
+  "plugin": ["[email protected].2"],
   "model": "openai/gpt-5-codex",
   "provider": {
     "openai": {
@@ -537,7 +537,7 @@ Create your own named variants in the model selector:
 ```json
 {
   "$schema": "https://opencode.ai/config.json",
-  "plugin": ["[email protected].1"],
+  "plugin": ["[email protected].2"],
   "provider": {
     "openai": {
       "models": {

config/full-opencode.json

@@ -1,7 +1,7 @@
 {
   "$schema": "https://opencode.ai/config.json",
   "plugin": [
-    "[email protected].1"
+    "[email protected].2"
   ],
   "provider": {
     "openai": {

index.ts

@@ -162,13 +162,17 @@ export const OpenAIAuthPlugin: Plugin = async ({ client }: PluginInput) => {
 
 						// Step 3: Transform request body with model-specific Codex instructions
 						// Instructions are fetched per model family (codex-max, codex, gpt-5.1)
+						// Capture original stream value before transformation
+						// generateText() sends no stream field, streamText() sends stream=true
+						const originalBody = init?.body ? JSON.parse(init.body as string) : {};
+						const isStreaming = originalBody.stream === true;
+
 						const transformation = await transformRequestForCodex(
 							init,
 							url,
 							userConfig,
 							codexMode,
 						);
-						const hasTools = transformation?.body.tools !== undefined;
 						const requestInit = transformation?.updatedInit ?? init;
 
 						// Step 4: Create headers with OAuth and ChatGPT account info
@@ -203,7 +207,7 @@ export const OpenAIAuthPlugin: Plugin = async ({ client }: PluginInput) => {
 							return await handleErrorResponse(response);
 						}
 
-						return await handleSuccessResponse(response, hasTools);
+						return await handleSuccessResponse(response, isStreaming);
 					},
 				};
 			},

lib/request/fetch-helpers.ts

@@ -279,24 +279,24 @@ export async function handleErrorResponse(
 
 /**
  * Handles successful responses from the Codex API
- * Converts SSE to JSON for non-tool requests
+ * Converts SSE to JSON for non-streaming requests (generateText)
+ * Passes through SSE for streaming requests (streamText)
  * @param response - Success response from API
- * @param hasTools - Whether the request included tools
- * @returns Processed response (SSE→JSON for non-tool, stream for tool requests)
+ * @param isStreaming - Whether this is a streaming request (stream=true in body)
+ * @returns Processed response (SSE→JSON for non-streaming, stream for streaming)
  */
 export async function handleSuccessResponse(
     response: Response,
-    hasTools: boolean,
+    isStreaming: boolean,
 ): Promise<Response> {
     const responseHeaders = ensureContentType(response.headers);
 
-	// For non-tool requests (compact/summarize), convert streaming SSE to JSON
-	// generateText() expects a non-streaming JSON response, not SSE
-	if (!hasTools) {
+	// For non-streaming requests (generateText), convert SSE to JSON
+	if (!isStreaming) {
 		return await convertSseToJson(response, responseHeaders);
 	}
 
-	// For tool requests, return stream as-is (streamText handles SSE)
+	// For streaming requests (streamText), return stream as-is
 	return new Response(response.body, {
 		status: response.status,
 		statusText: response.statusText,

lib/request/helpers/model-map.ts

@@ -43,6 +43,7 @@ export const MODEL_MAP: Record<string, string> = {
 	"gpt-5.1-low": "gpt-5.1",
 	"gpt-5.1-medium": "gpt-5.1",
 	"gpt-5.1-high": "gpt-5.1",
+	"gpt-5.1-chat-latest": "gpt-5.1",
 
 	// ============================================================================
 	// GPT-5 Codex Models (LEGACY - maps to gpt-5.1-codex as gpt-5 is being phased out)

lib/request/request-transformer.ts

@@ -398,6 +398,7 @@ export async function transformRequestBody(
 	// Codex required fields
 	// ChatGPT backend REQUIRES store=false (confirmed via testing)
 	body.store = false;
+	// Always set stream=true for API - response handling detects original intent
 	body.stream = true;
 	body.instructions = codexInstructions;
 
@@ -441,12 +442,22 @@ export async function transformRequestBody(
 			body.input = addToolRemapMessage(body.input, !!body.tools);
 		}
 
+		// Filter orphaned function_call_output items (where function_call was an item_reference that got filtered)
+		// Keep matched pairs for compaction context
 		if (!body.tools && body.input) {
-			body.input = body.input.filter(
-				(item) =>
-					item.type !== "function_call" &&
-					item.type !== "function_call_output",
+			// Collect all call_ids from function_call items
+			const functionCallIds = new Set(
+				body.input
+					.filter((item) => item.type === "function_call" && item.call_id)
+					.map((item) => item.call_id),
 			);
+			// Only filter function_call_output items that don't have a matching function_call
+			body.input = body.input.filter((item) => {
+				if (item.type === "function_call_output") {
+					return functionCallIds.has(item.call_id);
+				}
+				return true;
+			});
 		}
 	}
 

package.json

@@ -1,6 +1,6 @@
 {
   "name": "opencode-openai-codex-auth",
-  "version": "4.0.1",
+  "version": "4.0.2",
   "description": "OpenAI ChatGPT (Codex backend) OAuth auth plugin for opencode - use your ChatGPT Plus/Pro subscription instead of API credits",
   "main": "./dist/index.js",
   "types": "./dist/index.d.ts",

test/request-transformer.test.ts

@@ -847,13 +847,12 @@ describe('Request Transformer Module', () => {
 			expect(result.reasoning?.effort).toBe('medium');
 		});
 
-		it('should drop function_call items when no tools present', async () => {
+		it('should drop orphaned function_call_output when no tools present', async () => {
 			const body: RequestBody = {
 				model: 'gpt-5-codex',
 				input: [
 					{ type: 'message', role: 'user', content: 'hello' },
-					{ type: 'function_call', role: 'assistant', name: 'write', arguments: '{}' } as any,
-					{ type: 'function_call_output', role: 'assistant', call_id: 'call_1', output: '{}' } as any,
+					{ type: 'function_call_output', role: 'assistant', call_id: 'orphan_call', output: '{}' } as any,
 				],
 			};
 
@@ -862,7 +861,24 @@ describe('Request Transformer Module', () => {
 			expect(result.tools).toBeUndefined();
 			expect(result.input).toHaveLength(1);
 			expect(result.input![0].type).toBe('message');
-			expect(result.input![0].role).toBe('user');
+		});
+
+		it('should keep matched function_call pairs when no tools present (for compaction)', async () => {
+			const body: RequestBody = {
+				model: 'gpt-5-codex',
+				input: [
+					{ type: 'message', role: 'user', content: 'hello' },
+					{ type: 'function_call', call_id: 'call_1', name: 'write', arguments: '{}' } as any,
+					{ type: 'function_call_output', call_id: 'call_1', output: 'success' } as any,
+				],
+			};
+
+			const result = await transformRequestBody(body, codexInstructions);
+
+			expect(result.tools).toBeUndefined();
+			expect(result.input).toHaveLength(3);
+			expect(result.input![1].type).toBe('function_call');
+			expect(result.input![2].type).toBe('function_call_output');
 		});
 
 		describe('CODEX_MODE parameter', () => {