BYOK Responses retry for previous_response_not_found re-sends previous_response_id because OpenAIEndpoint overwrites ignoreStatefulMarker

[PenguinDOOM]

Type: Bug

Environment

• VS Code Insiders
• Built-in Copilot extension bundled with Insiders
• BYOK OpenAI-compatible endpoint
• Responses API (Codex CLI Backend)
• Reproduced against a local OpenAI-compatible test server

Summary

When a BYOK Responses request fails with HTTP 400:

{
  "code": "previous_response_not_found",
  "message": "No response found for previous_response_id resp_xxx.",
  "error": {
    "code": "previous_response_not_found",
    "message": "No response found for previous_response_id resp_xxx."
  }
}

VS Code appears to retry, but the retry still includes the same previous_response_id.

Expected behavior is to retry once without the stateful marker / previous_response_id.

Expected behavior

After receiving HTTP 400 with root-level:

• code: "previous_response_not_found"
• message: "No response found for previous_response_id ..."

the client should classify the failure as an invalid stateful marker and retry once without previous_response_id.

Actual behavior

A second POST /v1/responses is sent, but it still contains the same previous_response_id, so the request fails again with the same 400 and the user sees an error instead of recovery.

Repro

1. Configure Copilot BYOK against an OpenAI-compatible Responses endpoint.
2. Start a stateful conversation so the client sends previous_response_id.
3. Make the server return HTTP 400 with:

{
  "code": "previous_response_not_found",
  "message": "No response found for previous_response_id resp_xxx.",
  "error": {
    "code": "previous_response_not_found",
    "message": "No response found for previous_response_id resp_xxx."
  }
}

4. Observe that VS Code sends a second request.
5. Observe that the second request still contains the same previous_response_id.

Notes

I first suspected the server payload shape, but I verified the live server already returns root-level code and message as required.

The remaining issue is that the retry request still includes the original previous_response_id.

Evidence

The request log shows:

• first POST /v1/responses with previous_response_id = X
• HTTP 400
• second POST /v1/responses also with previous_response_id = X
• HTTP 400 again

So the retry is happening, but the stateful marker is not being cleared.

Code analysis

The retry path appears to be intended to work:

• extensions/copilot/src/extension/prompt/node/chatMLFetcher.ts
  • recognizes HTTP 400 with jsonData.code === 'previous_response_not_found'
• extensions/copilot/src/platform/endpoint/node/chatEndpoint.ts
  • converts that into InvalidStatefulMarker
  • retries with ignoreStatefulMarker: true

However, the BYOK OpenAIEndpoint appears to override that retry flag.

1) createRequestBody() mutates ignoreStatefulMarker

extensions/copilot/src/extension/byok/node/openAIEndpoint.ts

const zdr = !!this.modelMetadata.zeroDataRetentionEnabled;
options.ignoreStatefulMarker = zdr;
const body = super.createRequestBody(options);

This overwrites ignoreStatefulMarker: true during retry back to false in the normal non-ZDR case.

2) makeChatRequest2() also forces ignoreStatefulMarker: false

extensions/copilot/src/extension/byok/node/openAIEndpoint.ts

const modifiedOptions: IMakeChatRequestOptions = { ...options, ignoreStatefulMarker: false };
const response = await super.makeChatRequest2(modifiedOptions, token);

Even if this is not the primary reason for the observed retry failure, it is also suspicious because it prevents upstream intent from being preserved.

Proposed patch

I believe the BYOK endpoint should preserve an existing retry flag instead of overwriting it.

diff --git a/extensions/copilot/src/extension/byok/node/openAIEndpoint.ts b/extensions/copilot/src/extension/byok/node/openAIEndpoint.ts
index XXXXXXX..YYYYYYY 100644
--- a/extensions/copilot/src/extension/byok/node/openAIEndpoint.ts
+++ b/extensions/copilot/src/extension/byok/node/openAIEndpoint.ts
@@ -240,7 +240,7 @@ export class OpenAIEndpoint extends ChatEndpoint {
 		if (this.useResponsesApi) {
 			// Handle Responses API: customize the body directly
 			const zdr = !!this.modelMetadata.zeroDataRetentionEnabled;
-			options.ignoreStatefulMarker = zdr;
+			options.ignoreStatefulMarker = !!options.ignoreStatefulMarker || zdr;
 			const body = super.createRequestBody(options);
 			body.store = !zdr;
 			body.n = undefined;
@@ -367,10 +367,8 @@ export class OpenAIEndpoint extends ChatEndpoint {
 	}

 	public override async makeChatRequest2(options: IMakeChatRequestOptions, token: CancellationToken): Promise<ChatResponse> {
-		// Apply ignoreStatefulMarker: false for initial request
-		const modifiedOptions: IMakeChatRequestOptions = { ...options, ignoreStatefulMarker: false };
-		const response = await super.makeChatRequest2(modifiedOptions, token);
-		return hydrateBYOKErrorMessages(response);
+		const response = await super.makeChatRequest2(options, token);
+		return hydrateBYOKErrorMessages(response);
 	}
 }

Why this patch

• ChatEndpoint.makeChatRequest2() already computes the default behavior for ignoreStatefulMarker.
• A retry from InvalidStatefulMarker should be allowed to preserve ignoreStatefulMarker: true.
• ZDR should still disable stateful marker reuse, but it should not erase an already-true retry flag.

Suggested regression tests

Please consider adding a BYOK regression test covering this flow:

1. initial Responses request uses a stateful marker
2. server returns HTTP 400 previous_response_not_found
3. retry is triggered
4. retry request body omits previous_response_id

A targeted test around OpenAIEndpoint.createRequestBody() preserving ignoreStatefulMarker: true would also help.

VS Code version: Code - Insiders 1.121.0-insider (f805d63, 2026-05-15T12:52:51-04:00)
OS version: Windows_NT x64 10.0.26200
Modes:

System Info

Item	Value
CPUs	Intel(R) Core(TM) i7-14700F (28 x 2112)
GPU Status	2d_canvas: enabled GPU0: VENDOR= 0x10de, DEVICE=0x2705 [NVIDIA GeForce RTX 4070 Ti SUPER], DRIVER_VENDOR=NVIDIA, DRIVER_VERSION=32.0.15.9649 ACTIVE GPU1: VENDOR= 0x1414, DEVICE=0x008c [Microsoft Basic Render Driver], DRIVER_VERSION=10.0.26100.8328 Machine model name: Machine model version: direct_rendering_display_compositor: disabled_off_ok gpu_compositing: enabled multiple_raster_threads: enabled_on opengl: enabled_on rasterization: enabled raw_draw: disabled_off_ok skia_graphite: disabled_off trees_in_viz: disabled_off video_decode: enabled video_encode: enabled webgl: enabled webgl2: enabled webgpu: enabled webnn: disabled_off
Load (avg)	undefined
Memory (System)	63.84GB (40.50GB free)
Process Argv	--crash-reporter-id 2ec5f96c-c0f3-4a7f-a64c-d65afeffe46b
Screen Reader	no
VM	0%

Extensions (50)

Extension	Author (truncated)	Version
UnityShader	ash	1.0.13
markdown-preview-github-styles	bie	2.2.0
path-intellisense	chr	2.10.0
csharpier-vscode	csh	10.0.2
vscode-eslint	dba	3.0.24
EditorConfig	Edi	0.18.2
prettier-vscode	esb	12.4.0
remotehub	Git	0.64.0
vscode-github-actions	git	0.31.5
vscode-pull-request-github	Git	0.145.2026051504
discord-vscode	icr	5.9.2
markdown-live-editor	jis	0.6.2
zenkaku-hankaku	mas	1.0.0
vscode-copilot-auto-retry	Max	0.3.1
vscode-github-actions	me-	3.0.1
zenkaku	mos	0.0.3
vscode-language-pack-ja	MS-	1.118.2026051522
csdevkit	ms-	3.20.197
csharp	ms-	2.140.8
vscode-dotnet-runtime	ms-	3.0.2
debugpy	ms-	2026.7.11331010
python	ms-	2026.5.2026051501
vscode-pylance	ms-	2026.2.103
vscode-python-envs	ms-	1.33.2026051501
remote-containers	ms-	0.460.0
remote-ssh	ms-	0.123.2026051315
remote-ssh-edit	ms-	0.87.0
remote-wsl	ms-	0.104.3
azure-repos	ms-	0.40.0
cmake-tools	ms-	1.23.52
cpp-devtools	ms-	0.5.13
cpptools	ms-	1.32.2
cpptools-extension-pack	ms-	1.5.1
extension-test-runner	ms-	0.0.14
powershell	ms-	2025.4.0
remote-explorer	ms-	0.6.2026031809
remote-repositories	ms-	0.42.0
vscode-chat-customizations-evaluations	ms-	1.0.2
vscode-github-issue-notebooks	ms-	0.0.134
indent-rainbow	ode	8.3.1
fix-json	oli	0.2.0
chatgpt	ope	26.5513.21555
material-icon-theme	PKi	5.34.0
vscode-yaml	red	1.24.2026050908
trailing-spaces	sha	0.4.1
shader	sle	1.1.5
pdf	tom	1.2.2
native-preview	Typ	0.20260515.1
vstuc	vis	1.2.2
gpg-indicator	wdh	0.7.5

(1 theme extensions excluded)

A/B Experiments

vsliv368cf:30146710
pythonvspyt551:31249597
nativeloc1:31118317
dwcopilot:31158714
dwoutputs:31242946
copilot_t_ci:31333650
g012b348:31231168
pythonrdcb7:31268811
pythonpcpt1cf:31399617
6518g693:31302842
82j33506:31327384
6abeh943:31336334
envsactivate1:31349248
editstats-enabled:31346256
cloudbuttont:31366566
3efgi100_wstrepl:31403338
cp_cls_c_966_ss:31454199
inlinechat_v2_hd992725:31445440
4je02754:31455664
8hhj4413:31478653
ge8j1254_inline_auto_hint_haiku:31490507
cp_cls_c_1081:31454833
conptydll_true:31485575
ia-use-proxy-models-svc:31446143
e9c30283:31453065
test_treatment2:31471001
c9b86496:31447327
idci7584:31454084
nes_chat_context_disabled:31451402
e3e4d672:31454087
ei9d7968:31462942
nes-extended-on:31455475
quick_suggest_off_75197330:31462668
89g7j272:31506658
7e884298:31462391
7e187181:31482583
i2gc6536:31472020
52612955:31508042
h08i8180:31475367
ddid_c:31478205
hmra_i5g22:31509478
getcmakediagnosticsoff:31489825
61f49681:31505879
7df3h592:31491241
pro_large_t:31499377
cp_cls_c_1082:31504161
logging_enabled_new:31490725
jb_cp_cls_c_632:31510883
cg448276_tst_on:31503513
32d76977:31503652
ha629193:31508444
31fi7170_t:31510641
jh5f2457_c:31514653
tco_off:31513904
api_cot_ctrl:31509853
hgf2d445:31510900
rd_file_off:31514889
prpt_ctrl:31513638

[PenguinDOOM]

This occurs not only with OpenAI-compatible but also with OpenAI endpoints that use API.