fix: preserve context by retrying with full conversation on invalid previous_response_id #7714

daniel-lxs · 2025-09-05T19:27:48Z

Description

This PR fixes a critical regression in the OpenAI Responses API where conversation context was lost after a continuity failure.

The Problem

When a previous_response_id becomes invalid or expired (400 error), the current code on main:

Removes the invalid ID and retries
BUT still sends only the latest message in the retry
This causes complete context loss and degraded responses

The Solution

This PR ensures that when previous_response_id fails:

Clear the invalid response ID
Remove previous_response_id from the request
Re-prepare the full conversation history for the retry
Send all messages to preserve context

Changes

Modified executeRequest() to re-prepare input with full conversation on retry
Modified makeGpt5ResponsesAPIRequest() with the same fix for SSE fallback
Added comprehensive test coverage verifying retry sends full conversation

Test Procedure

Run the tests:

cd src && npx vitest run api/providers/__tests__/openai-native.spec.ts

Expected: All 36 tests pass, including the new test that verifies:

First attempt with previous_response_id sends only latest message
On 400 error, retry sends the full conversation history
Context is preserved on retry

Impact

Fixes context loss bug that degraded response quality
Ensures conversation continuity is maintained even when response IDs expire
Backward compatible - only affects retry behavior after specific 400 errors

Important

Fixes context loss by retrying with full conversation history on invalid previous_response_id errors in OpenAI Responses API.

Behavior:
- Fixes context loss by retrying with full conversation history when previous_response_id is invalid in executeRequest() and makeGpt5ResponsesAPIRequest().
- On 400 error, clears previous_response_id and resends full conversation.
Tests:
- Added tests in openai-native.spec.ts to verify retry logic sends full conversation on 400 error.
- Tests ensure context is preserved and verify two requests are made: first with previous_response_id, second with full history.
Misc:
- No changes to existing functionality, only affects retry behavior after specific 400 errors.

^{This description was created by}^{for 8d4a77f. You can customize this summary. It will automatically update as commits are pushed.}

…revious_response_id When the OpenAI Responses API returns a 400 error due to an invalid or expired previous_response_id, the code now properly re-prepares the full conversation history for the retry instead of just removing the ID and sending only the latest message. This fixes a critical bug where conversation context was completely lost when continuity failed, leading to degraded responses. Changes: - Modified executeRequest() to re-prepare input with full conversation on retry - Modified makeGpt5ResponsesAPIRequest() with the same fix for SSE fallback - Added comprehensive test coverage for both SDK and SSE retry paths - Tests verify retry sends full conversation, not just latest message

ellipsis-dev · 2025-09-05T19:29:34Z

src/api/providers/openai-native.ts

 				// Clear the stored lastResponseId to prevent using it again
 				this.lastResponseId = undefined

+				// Re-prepare the full conversation without previous_response_id


Good fix: when a 400 error occurs due to an invalid previous_response_id, the code now clears it and re-prepares the full conversation. Consider extracting the duplicated logic for re-preparing the conversation (lines ~327–339 and ~517–521) into a helper method to reduce repetition and ease future maintenance.

roomote

Thank you for this important fix! The solution correctly addresses the context loss issue when previous_response_id becomes invalid. I've reviewed the changes and have some suggestions to improve code maintainability.

roomote · 2025-09-05T19:32:24Z

src/api/providers/openai-native.ts

+				// If we have the original messages, re-prepare the full conversation
+				if (systemPrompt && messages) {
+					const { formattedInput } = this.prepareStructuredInput(systemPrompt, messages, undefined)
+					retryRequestBody.input = formattedInput


I agree with the existing suggestion - this duplicated logic for re-preparing the conversation (here and around line 517-521) should be extracted into a helper method. Consider something like:

Suggested change

retryRequestBody.input = formattedInput

private prepareRetryRequestBody(requestBody: any, systemPrompt?: string, messages?: Anthropic.Messages.MessageParam[]): any {

const retryRequestBody = { ...requestBody };

delete retryRequestBody.previous_response_id;

if (systemPrompt && messages) {

const { formattedInput } = this.prepareStructuredInput(systemPrompt, messages, undefined);

retryRequestBody.input = formattedInput;

}

return retryRequestBody;

}

roomote · 2025-09-05T19:32:24Z

src/api/providers/openai-native.ts


 			if (is400Error && requestBody.previous_response_id && isPreviousResponseError) {
 				// Log the error and retry without the previous_response_id



The comment mentions logging but no actual logging is implemented. Consider adding proper error logging here for debugging purposes:

Suggested change

// Log the error and retry without the previous_response_id

console.warn('Previous response ID not found, retrying with full conversation:', errorMessage);

roomote · 2025-09-05T19:32:24Z

src/api/providers/openai-native.ts


 				if (response.status === 400 && requestBody.previous_response_id && isPreviousResponseError) {
 					// Log the error and retry without the previous_response_id



Same logging issue here - consider adding actual error logging:

Suggested change

// Log the error and retry without the previous_response_id

console.warn('Previous response ID not found (SSE path), retrying with full conversation:', errorDetails);

roomote · 2025-09-05T19:32:24Z

src/api/providers/__tests__/openai-native.spec.ts

+					role: "user",
+					content: [{ type: "input_text", text: "What number did I ask you to remember?" }],
+				},
+			])


Great test coverage! Consider adding one more test case for when both the initial request AND the retry fail, to ensure proper error propagation in that scenario.

daniel-lxs requested review from cte, jr and mrubens as code owners September 5, 2025 19:27

github-project-automation bot added this to Roo Code Roadmap and Roo Code Roadmap Sep 5, 2025

github-project-automation bot moved this to New in Roo Code Roadmap Sep 5, 2025

github-project-automation bot moved this to Triage in Roo Code Roadmap Sep 5, 2025

dosubot bot added size:L This PR changes 100-499 lines, ignoring generated files. bug Something isn't working labels Sep 5, 2025

ellipsis-dev bot reviewed Sep 5, 2025

View reviewed changes

daniel-lxs moved this from Triage to PR [Needs Review] in Roo Code Roadmap Sep 5, 2025

mrubens approved these changes Sep 5, 2025

View reviewed changes

dosubot bot added the lgtm This PR has been approved by a maintainer label Sep 5, 2025

roomote bot reviewed Sep 5, 2025

View reviewed changes

hannesrudolph added Issue/PR - Triage New issue. Needs quick review to confirm validity and assign labels. PR - Needs Review and removed Issue/PR - Triage New issue. Needs quick review to confirm validity and assign labels. labels Sep 5, 2025

mrubens merged commit 49b50c8 into main Sep 5, 2025
26 checks passed

mrubens deleted the fix/openai-context-loss-on-retry branch September 5, 2025 20:09

github-project-automation bot moved this from New to Done in Roo Code Roadmap Sep 5, 2025

github-project-automation bot moved this from PR [Needs Review] to Done in Roo Code Roadmap Sep 5, 2025

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

fix: preserve context by retrying with full conversation on invalid previous_response_id #7714

fix: preserve context by retrying with full conversation on invalid previous_response_id #7714

Uh oh!

daniel-lxs commented Sep 5, 2025 •

edited by ellipsis-dev bot

Loading

Uh oh!

ellipsis-dev bot Sep 5, 2025

Uh oh!

roomote bot left a comment

Uh oh!

roomote bot Sep 5, 2025

Uh oh!

roomote bot Sep 5, 2025

Uh oh!

roomote bot Sep 5, 2025

Uh oh!

roomote bot Sep 5, 2025

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

-					retryRequestBody.input = formattedInput
+private prepareRetryRequestBody(requestBody: any, systemPrompt?: string, messages?: Anthropic.Messages.MessageParam[]): any {
+  const retryRequestBody = { ...requestBody };
+  delete retryRequestBody.previous_response_id;
+  if (systemPrompt && messages) {
+    const { formattedInput } = this.prepareStructuredInput(systemPrompt, messages, undefined);
+    retryRequestBody.input = formattedInput;
+  }
+  return retryRequestBody;
+}


		if (is400Error && requestBody.previous_response_id && isPreviousResponseError) {
		// Log the error and retry without the previous_response_id


	// Log the error and retry without the previous_response_id
	console.warn('Previous response ID not found, retrying with full conversation:', errorMessage);


		if (response.status === 400 && requestBody.previous_response_id && isPreviousResponseError) {
		// Log the error and retry without the previous_response_id

fix: preserve context by retrying with full conversation on invalid previous_response_id #7714

fix: preserve context by retrying with full conversation on invalid previous_response_id #7714

Uh oh!

Conversation

daniel-lxs commented Sep 5, 2025 • edited by ellipsis-dev bot Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Description

The Problem

The Solution

Changes

Test Procedure

Impact

Uh oh!

ellipsis-dev bot Sep 5, 2025

Choose a reason for hiding this comment

Uh oh!

roomote bot left a comment

Choose a reason for hiding this comment

Uh oh!

roomote bot Sep 5, 2025

Choose a reason for hiding this comment

Uh oh!

roomote bot Sep 5, 2025

Choose a reason for hiding this comment

Uh oh!

roomote bot Sep 5, 2025

Choose a reason for hiding this comment

Uh oh!

roomote bot Sep 5, 2025

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

daniel-lxs commented Sep 5, 2025 •

edited by ellipsis-dev bot

Loading