fix(reasoning): preserve compatible response state

changelog.d/fixes/10550-responses-reasoning-transport.md

@@ -0,0 +1 @@
+- Preserve portable plaintext reasoning by default across streaming and non-streaming Chat Completions and Responses routes while keeping provider-bound opaque state target-compatible. Combos now drop incompatible continuation reasoning by default and can explicitly skip incompatible targets, while known providers no longer show redundant encrypted-reasoning controls. (#10550)

config/quality/file-size-baseline.json

@@ -1,4 +1,5 @@
 {
+  "_rebaseline_2026_08_20_10574_reasoning_transport_fallback": "PR #10574 (jackjinke, fix/responses-reasoning-transport, fixes #10550) own growth: src/sse/handlers/chatHelpers.ts 1017->1019 (+2 = the new reasoningTransportFallback option threaded through executeChatWithBreaker's options destructure and its downstream handleSingleModel call, at the existing per-attempt options-passthrough chokepoint; not extractable without splitting the option-forwarding call itself). Covered by the PR's own reasoning-policy test suite (tests/unit/chatcore-translation-paths.test.ts, tests/unit/combo-attempt-body-isolation-7847.test.ts, tests/unit/reasoning-cache.test.ts, tests/unit/strip-reasoning-blobs-agentic-context-1599.test.ts among others), 446/446 focused tests passing.",
   "_rebaseline_2026_08_18_10517_zed_hosted_oauth_callback_port": "PR #10517 (phatchau036, fix/zed-hosted-oauth-callback-port) own growth: src/shared/components/OAuthModal.tsx 1131->1148 (wc -l; check-file-size.mjs counts via split(\"\\n\").length so the gate sees 1134->1149, +15/+18, crosses the frozen 1134 cap). Wires the zed-hosted native-app callback auto-complete: forceManual gating on isTrueLocalhost for zed-hosted, the loopback-redirect-URI comment block, and the exchangeToken full-URL-as-code branch, all at the existing provider-switch chokepoints this modal already carries growth for (seventh bump: 969->989->993->998->1030->1056->1100->1149; structural shrink tracked in #3501). The actual port-derivation logic lives in src/lib/oauth/providers/zed-hosted.ts (not frozen here) and was hardened during pre-merge review to use the server's own getRuntimePorts() instead of a browser-guessed scheme/port, covered by the new tests/unit/zed-hosted-loopback-port-derivation.test.ts (8/8 passing).",
   "_rebaseline_2026_08_13_10243_codex_fingerprint_merge": "PR #10243 (xz-dev, Codex OAuth fingerprint convergence) merge into release/v3.8.50: src/app/(dashboard)/dashboard/providers/[id]/providerPageHelpers.ts crossed the 1000-line new-file cap for the first time (974 on base, 997 on the PR's own branch, 1013 after merging + prettier reflow) purely from combining two independent, already-legitimate feature additions that landed on the same shared UI-helper file — this PR's own Codex fingerprint-mode select/toggle wiring (CODEX_FINGERPRINT_MODE_VALUES, getCodexFingerprintModeLabel, CodexFingerprintModeValue) plus #8949's unrelated Codex account-service-tier helpers merged concurrently on release/v3.8.50. Neither addition alone crosses the cap; git's line-level auto-merge does not detect a threshold crossing. Not modularized as part of this conflict-resolution merge commit (out of scope — this is a merge, not a feature change). Covered by the PR's own tests/unit/codex-fingerprint-convergence.test.ts, tests/unit/executor-codex.test.ts, tests/unit/provider-specific-data-schema.test.ts (all passing post-merge).",
   "_rebaseline_2026_08_09_8984_api_key_cache_mode": "PR #8984 own growth during the 2026-08-09 rebase: src/lib/db/apiKeys.ts 1529->1545 (+16 = the per-key apiKeys.cacheDefaultMode  column + its row parsers and cascade wiring; additive at the existing connection write/read chokepoints). Covered by tests/unit/chatcore-semantic-cache.test.ts. (chatCore.ts stays at the pre-existing base-red ceiling — upstream tip already exceeds the frozen 5042, this PR only adds +2 on top; not re-bumped per the no-inherit-ratchet rule.)",
@@ -450,7 +451,7 @@
     "src/lib/modelCapabilities.ts": 1006,
     "src/app/(dashboard)/dashboard/providers/[id]/providerPageHelpers.ts": 1014,
     "open-sse/config/imageRegistry.ts": 1034,
-    "src/sse/handlers/chatHelpers.ts": 1017,
+    "src/sse/handlers/chatHelpers.ts": 1019,
     "src/shared/middleware/chatBodyAdmission.ts": 1005,
     "_rebaseline_2026_08_20_10668_tabitoken_gateway": "#10668 (yawar-aquil) own catalog growth: src/shared/constants/providers/apikey/gateways.ts 1268->1283 (+15, entirely this PR diff -- one new tabitoken gateway entry, data lines only; base moved from 1255 to 1268 via other merges since the PR forked). Not combination drift: reproducible on the PR branch alone, so the WS5.5 release-captain rule does not apply. Extraction is not available -- the file is pure data (own header: \"Pure data; merged by apikey/index.ts via spread\") and already split into 6 family files under apikey/. Same precedent as _rebaseline_2026_08_14_imagetotext_servicekinds (#10275/#10291, gateways.ts 1250->1255, data lines only) and _rebaseline_2026_08_11_v3850_merge_storm_provider_registry (owner-authorized for this same file)."
   },

open-sse/config/providers/registry/chatgpt-web-codex/index.ts

@@ -14,6 +14,7 @@ export const chatgpt_web_codexProvider: RegistryEntry = {
   format: "openai-responses",
   executor: "chatgpt-web-codex",
   baseUrl: "https://chatgpt.com",
+  reasoningTransport: "opaque",
   authType: "apikey",
   authHeader: "cookie",
   forceStream: true,

open-sse/config/providers/registry/codex/index.ts

@@ -12,6 +12,7 @@ export const codexProvider: RegistryEntry = {
   format: "openai-responses",
   executor: "codex",
   baseUrl: "https://chatgpt.com/backend-api/codex/responses",
+  reasoningTransport: "opaque",
   authType: "oauth",
   authHeader: "bearer",
   defaultContextLength: 400000,

open-sse/config/providers/registry/grok-cli/index.ts

@@ -14,6 +14,7 @@ export const grok_cliProvider: RegistryEntry = {
   // Keep the generic translate-path contract stable. GrokCliExecutor owns the
   // official Grok Build upstream URL and always dispatches to /v1/responses.
   baseUrl: "https://cli-chat-proxy.grok.com/v1/chat/completions",
+  reasoningTransport: "opaque",
   modelsUrl: GROK_BUILD_MODELS_URL,
   clientVersion: getGrokBuildClientVersion(),
   authType: "oauth",

open-sse/config/providers/registry/muse-code/index.ts

@@ -14,6 +14,7 @@ export const muse_codeProvider: RegistryEntry = buildOpenAiCompatibleRegistryEnt
   id: "muse-code",
   alias: "mc",
   passthroughModels: true,
+  reasoningTransport: "opaque",
   defaultContextLength: 200000,
   models: [
     {

open-sse/config/providers/registry/openai/index.ts

@@ -7,6 +7,7 @@ export const openaiProvider: RegistryEntry = {
   format: "openai",
   executor: "default",
   baseUrl: "https://api.openai.com/v1/chat/completions",
+  reasoningTransport: "opaque",
   authType: "apikey",
   authHeader: "bearer",
   defaultContextLength: 128000,

open-sse/config/providers/registry/xai/index.ts

@@ -12,6 +12,7 @@ export const xaiProvider: RegistryEntry = {
   // XaiExecutor.buildUrl (open-sse/executors/xai.ts) for models tagged
   // targetFormat: "openai-responses" below.
   responsesBaseUrl: "https://api.x.ai/v1/responses",
+  reasoningTransport: "opaque",
   authType: "apikey",
   authHeader: "bearer",
   models: [
@@ -54,6 +55,7 @@ export const xai_oauthProvider: RegistryEntry = {
   executor: "xai-oauth",
   baseUrl: xaiProvider.baseUrl,
   responsesBaseUrl: xaiProvider.responsesBaseUrl,
+  reasoningTransport: "opaque",
   authType: "oauth",
   authHeader: xaiProvider.authHeader,
   passthroughModels: true,

open-sse/config/providers/shared.ts

@@ -108,6 +108,8 @@ export interface RegistryOAuth {
   pollUrlBase?: string;
 }
 
+export type ReasoningTransport = "plaintext" | "opaque" | "none";
+
 export interface RegistryEntry {
   id: string;
   alias?: string;
@@ -120,6 +122,8 @@ export interface RegistryEntry {
   /** Override models URL used only for API key validation, not catalog discovery. */
   testKeyModelsUrl?: string;
   responsesBaseUrl?: string;
+  /** Provider-bound replay format; omitted providers accept portable plaintext reasoning. */
+  reasoningTransport?: ReasoningTransport;
   /** Anthropic-native /v1/messages endpoint (e.g. GitHub Copilot's shim) used
    *  for models tagged `targetFormat: "claude"` on an otherwise openai-format
    *  provider — see registry/github/index.ts. */

open-sse/executors/codex.ts

@@ -34,7 +34,7 @@ import {
 } from "../config/codexIdentity.ts";
 import { getAccessToken } from "../services/tokenRefresh.ts";
 import { sanitizeResponsesInputItems } from "../services/responsesInputSanitizer.ts";
-import { applyResponsesInputPolicy } from "../services/responsesInputPolicy.ts";
+import { applyReasoningInputPolicy } from "../services/reasoningInputPolicy.ts";
 import { normalizeCodexVerbosity } from "../services/codexVerbosity.ts";
 import { getThinkingBudgetConfig, ThinkingMode } from "../services/thinkingBudget.ts";
 import { CORS_HEADERS } from "../utils/cors.ts";
@@ -1389,10 +1389,12 @@ export class CodexExecutor extends BaseExecutor {
     delete body.session_id;
     delete body.conversation_id;
 
-    applyResponsesInputPolicy(
-      body,
-      credentials?.providerSpecificData?.preserveEncryptedReasoning === true
-    );
+    applyReasoningInputPolicy(body, "responses", {
+      provider: "codex",
+      preserveEncryptedReasoning:
+        credentials?.providerSpecificData?.preserveEncryptedReasoning === true,
+      onIncompatibleReasoning: "drop",
+    });
 
     if (nativeCodexPassthrough) {
       return body;

open-sse/handlers/chatCore.ts

@@ -33,7 +33,7 @@ import { assembleStreamingResponseHeaders } from "./chatCore/streamingResponseHe
 import { storeStreamingSemanticCacheResponse } from "./chatCore/streamingSemanticCacheStore.ts";
 import { assembleStreamingPipeline } from "./chatCore/streamingPipeline.ts";
 import { sanitizeChatRequestBody } from "./chatCore/sanitization.ts";
-import { applyResponsesInputPolicy } from "../services/responsesInputPolicy.ts";
+import { applyReasoningInputPolicy } from "../services/reasoningInputPolicy.ts";
 import {
   createRoutingEvent,
   emitRoutingEvent,
@@ -511,6 +511,7 @@ export async function handleChatCore({
   conversationId = null,
   modelPinned = false,
   skipResourcePressureGuard = false,
+  reasoningTransportFallback = "skip",
   managedLease = null,
 }) {
   let { provider, model, extendedContext } = modelInfo;
@@ -1193,11 +1194,30 @@ export async function handleChatCore({
     return cacheHit;
   }
 
-  if (targetFormat === FORMATS.OPENAI_RESPONSES && body && typeof body === "object") {
-    applyResponsesInputPolicy(
+  const reasoningInputFormat =
+    sourceFormat === FORMATS.OPENAI_RESPONSES
+      ? "responses"
+      : sourceFormat === FORMATS.OPENAI
+        ? "chat"
+        : null;
+  if (reasoningInputFormat && body && typeof body === "object") {
+    const policy = applyReasoningInputPolicy(
       body as Record<string, unknown>,
-      credentials?.providerSpecificData?.preserveEncryptedReasoning === true
+      reasoningInputFormat,
+      {
+        provider,
+        preserveEncryptedReasoning:
+          credentials?.providerSpecificData?.preserveEncryptedReasoning === true,
+        onIncompatibleReasoning: reasoningTransportFallback === "drop" ? "drop" : "reject",
+      }
     );
+    if (policy.incompatibleReasoning) {
+      trackPendingRequest(model, provider, connectionId, false);
+      return createErrorResult(
+        HTTP_STATUS.BAD_REQUEST,
+        "Reasoning continuation is not compatible with the selected target"
+      );
+    }
   }
 
   body = sanitizeChatRequestBody(body, sourceFormat, targetFormat);

open-sse/handlers/responseSanitizer.ts

@@ -8,7 +8,10 @@ import {
   collapseExcessiveNewlines,
   extractThinkingFromContent,
 } from "./responseSanitizer/reasoning.ts";
-import { applyCacheHitTokensToUsage, applyCacheHitTokensToResponsesUsage } from "./responseSanitizer/cacheHitTokens.ts";
+import {
+  applyCacheHitTokensToUsage,
+  applyCacheHitTokensToResponsesUsage,
+} from "./responseSanitizer/cacheHitTokens.ts";
 export {
   extractThinkingFromContent,
   shouldParseTextualReasoningTags,
@@ -31,7 +34,9 @@ const ALLOWED_USAGE_FIELDS = new Set([
   "total_tokens",
   "cached_tokens",
   "prompt_tokens_details",
-  "completion_tokens_details", "cache_read_input_tokens", "cache_creation_input_tokens",
+  "completion_tokens_details",
+  "cache_read_input_tokens",
+  "cache_creation_input_tokens",
   // Keep through sanitize → applyClientUsageBuffer so heuristic web usage is
   // not inflated by the default USAGE_TOKEN_BUFFER (2000).
   "estimated",
@@ -550,7 +555,7 @@ function sanitizeResponsesUsage(usage: unknown): unknown {
     !(toRecord(normalized.input_tokens_details) ?? {}).cached_tokens
   ) {
     normalized.input_tokens_details = {
-      ...(normalized.input_tokens_details as Record<string, unknown> || {}),
+      ...((normalized.input_tokens_details as Record<string, unknown>) || {}),
       cached_tokens: normalized.prompt_cache_hit_tokens,
     };
   }
@@ -562,7 +567,7 @@ function sanitizeResponsesUsage(usage: unknown): unknown {
     !(toRecord(normalized.input_tokens_details) ?? {}).cached_tokens
   ) {
     normalized.input_tokens_details = {
-      ...(normalized.input_tokens_details as Record<string, unknown> || {}),
+      ...((normalized.input_tokens_details as Record<string, unknown>) || {}),
       cached_tokens: normalized.cache_read_input_tokens,
     };
   }
@@ -863,6 +868,7 @@ function sanitizeResponsesOutputItem(item: unknown, index: number): JsonRecord |
       : [];
 
     return {
+      ...itemRecord,
       id: toString(itemRecord.id) || `rs_${index}`,
       type: "reasoning",
       summary,

open-sse/handlers/responseTranslator.ts

@@ -10,6 +10,7 @@ import {
   caseInsensitiveToolNameLookup,
   restoreOpenAIToolNames,
 } from "../translator/helpers/toolCallHelper.ts";
+import { extractReplayableResponsesReasoningText } from "../services/reasoningInputPolicy.ts";
 import { sanitizeToolId } from "../translator/helpers/schemaCoercion.ts";
 
 type JsonRecord = Record<string, unknown>;
@@ -178,7 +179,8 @@ export function translateNonStreamingResponse(
 
     const messageSelection = findBestMessageText(output);
     let textContent = messageSelection.text;
-    let reasoningContent = "";
+    let replayableReasoningContent = "";
+    let reasoningSummary = "";
     const toolCalls: JsonRecord[] = [];
 
     for (const item of output) {
@@ -192,16 +194,22 @@ export function translateNonStreamingResponse(
           if (partObj.type === "summary_text" && typeof partObj.text === "string") {
             // #9500 — reasoning summary parts are discrete segments; join with "\n\n"
             // (matches extractThinkingFromContent convention) so they don't glue back-to-back.
-            reasoningContent += reasoningContent ? `\n\n${partObj.text}` : partObj.text;
+            reasoningSummary += reasoningSummary ? `\n\n${partObj.text}` : partObj.text;
           }
         }
-      } else if (itemObj.type === "reasoning" && Array.isArray(itemObj.summary)) {
-        for (const part of itemObj.summary) {
-          const partObj = toRecord(part);
-          if (partObj.type === "summary_text" && typeof partObj.text === "string") {
-            // #9500 — reasoning summary parts are discrete segments; join with "\n\n"
-            // (matches extractThinkingFromContent convention) so they don't glue back-to-back.
-            reasoningContent += reasoningContent ? `\n\n${partObj.text}` : partObj.text;
+      } else if (itemObj.type === "reasoning") {
+        const replayable = extractReplayableResponsesReasoningText(itemObj);
+        if (replayable) {
+          replayableReasoningContent += replayableReasoningContent
+            ? `\n\n${replayable}`
+            : replayable;
+        }
+        if (Array.isArray(itemObj.summary)) {
+          for (const part of itemObj.summary) {
+            const partObj = toRecord(part);
+            if (partObj.type === "summary_text" && typeof partObj.text === "string") {
+              reasoningSummary += reasoningSummary ? `\n\n${partObj.text}` : partObj.text;
+            }
           }
         }
       } else if (itemObj.type === "function_call") {
@@ -238,8 +246,11 @@ export function translateNonStreamingResponse(
     if (textContent) {
       message.content = textContent;
     }
-    if (reasoningContent) {
-      message.reasoning_content = reasoningContent;
+    if (replayableReasoningContent) {
+      message.reasoning_content = replayableReasoningContent;
+    }
+    if (reasoningSummary) {
+      message.reasoning_summary = [{ type: "summary_text", text: reasoningSummary }];
     }
     if (toolCalls.length > 0) {
       message.tool_calls = toolCalls;