@bitkyc08/opencodex 2.88.0 → 2.90.0

package/src/providers/registry/entries-extended.ts

@@ -54,7 +54,7 @@ import {
   DEEPSEEK_VISION_PREVIEW_MODEL,
   COMMAND_CODE_MIMO_CONTEXT_WINDOWS,
   COMMAND_CODE_MODEL_INPUT_MODALITIES,
-  OPENCODE_FREE_DEEPSEEK_MODELS,
+  OPENCODE_ZEN_FREE_DEEPSEEK_MODELS,
   OPENCODE_ZEN_TEXT_ONLY_MODELS,
   OPENCODE_ZEN_IMAGE_MODELS,
   deepseekThinkingEffortsFor,
@@ -1020,7 +1020,12 @@ export const PROVIDER_REGISTRY_EXTENDED: readonly ProviderRegistryEntry[] = [
     },
   },
   // FREEZE 2026-07-10: codestral-latest is unconfirmed behind auth. Evidence: devlog/_plan/260710_provider_hardening/003_research_aggregators.md.
-  { id: "mistral", label: "Mistral", baseUrl: "https://api.mistral.ai/v1", adapter: "openai-chat", authKind: "key", dashboardUrl: "https://console.mistral.ai/api-keys", defaultModel: "codestral-latest" },
+  {
+    id: "mistral", label: "Mistral", baseUrl: "https://api.mistral.ai/v1", adapter: "openai-chat", authKind: "key",
+    dashboardUrl: "https://console.mistral.ai/api-keys", defaultModel: "codestral-latest", promptCacheKey: true,
+    // Native text-only input enrolls this model in the existing vision sidecar.
+    modelInputModalities: { "zai-glm-5-3": ["text"] },
+  },
   {
     id: "minimax", label: "MiniMax — Coding Plan", baseUrl: "https://api.minimax.io/v1", adapter: "openai-chat", authKind: "key",
     dashboardUrl: "https://platform.minimax.io", defaultModel: "MiniMax-M3", models: MINIMAX_MODELS,
@@ -1089,20 +1094,20 @@ export const PROVIDER_REGISTRY_EXTENDED: readonly ProviderRegistryEntry[] = [
   },
   {
     id: "opencode-zen", label: "opencode zen", baseUrl: "https://opencode.ai/zen/v1", adapter: "openai-chat", authKind: "key", dashboardUrl: "https://opencode.ai/auth",
-    // Same opencode.ai/zen/v1 gateway as `opencode-free` (keyed tier): DeepSeek thinking mode
+    // Keyed opencode.ai/zen/v1 gateway: DeepSeek thinking mode
     // requires the assistant's original reasoning_content to be replayed on tool-call
     // continuations, or the gateway answers HTTP 400 (issues #950/#994). Mirror the DeepSeek
     // reasoning + thinking metadata so `opencode-zen/deepseek-v4-flash-free` — and the other
     // Zen DeepSeek thinking models — never serialize a bare tool-call turn.
-    note: "Keyed OpenCode Zen gateway. Free models on this tier are often short-window rate-limited at roughly 15-20 requests/minute (community-measured; OpenCode does not publish RPM). Zen may return generic 429s without Retry-After / X-RateLimit headers; when Retry-After is omitted, opencodex adds a synthetic backoff hint (upstream Retry-After still wins). Distinct from the keyless opencode-free desktop quota (~200 Big Pickle/free-model requests per 5 hours). Docs: https://opencode.ai/docs/zen/. Free-model prompts may be retained for training — do not send confidential material.",
+    note: "Keyed OpenCode Zen gateway. Free models on this tier are often short-window rate-limited at roughly 15-20 requests/minute (community-measured; OpenCode does not publish RPM). Zen may return generic 429s without Retry-After / X-RateLimit headers; when Retry-After is omitted, opencodex adds a synthetic backoff hint (upstream Retry-After still wins). Docs: https://opencode.ai/docs/zen/. Free-model prompts may be retained for training — do not send confidential material.",
     modelReasoningEfforts: Object.fromEntries(
-      [...DEEPSEEK_GATEWAY_THINKING_MODELS, ...OPENCODE_FREE_DEEPSEEK_MODELS].map(id => [id, deepseekThinkingEffortsFor(id)]),
+      [...DEEPSEEK_GATEWAY_THINKING_MODELS, ...OPENCODE_ZEN_FREE_DEEPSEEK_MODELS].map(id => [id, deepseekThinkingEffortsFor(id)]),
     ),
     modelReasoningEffortMap: Object.fromEntries(
-      [...DEEPSEEK_GATEWAY_THINKING_MODELS, ...OPENCODE_FREE_DEEPSEEK_MODELS].map(id => [id, deepseekReasoningMapFor(id)]),
+      [...DEEPSEEK_GATEWAY_THINKING_MODELS, ...OPENCODE_ZEN_FREE_DEEPSEEK_MODELS].map(id => [id, deepseekReasoningMapFor(id)]),
     ),
-    preserveReasoningContentModels: [...DEEPSEEK_GATEWAY_THINKING_MODELS, ...OPENCODE_FREE_DEEPSEEK_MODELS],
-    // Same Zen gateway as opencode-free: the DeepSeek vision preview id
+    preserveReasoningContentModels: [...DEEPSEEK_GATEWAY_THINKING_MODELS, ...OPENCODE_ZEN_FREE_DEEPSEEK_MODELS],
+    // The DeepSeek vision preview id
     // (merges into deepseek-v4-flash later).
     modelContextWindows: {
       [DEEPSEEK_VISION_PREVIEW_MODEL]: 1_048_576,
@@ -1114,7 +1119,7 @@ export const PROVIDER_REGISTRY_EXTENDED: readonly ProviderRegistryEntry[] = [
     noVisionModels: [...OPENCODE_ZEN_TEXT_ONLY_MODELS, ...DEEPSEEK_GATEWAY_THINKING_MODELS],
     // Same DeepSeek routes as the Go preset above, behind the same vendor, so they carry
     // the same json_schema rejection (#1338 / #1415).
-    noJsonSchemaModels: [...DEEPSEEK_GATEWAY_THINKING_MODELS, ...OPENCODE_FREE_DEEPSEEK_MODELS],
+    noJsonSchemaModels: [...DEEPSEEK_GATEWAY_THINKING_MODELS, ...OPENCODE_ZEN_FREE_DEEPSEEK_MODELS],
     // Muse Spark on Zen can sit silent during prolonged reasoning and close without a protocol terminal.
     modelResponsesTerminalRepair: {
       "muse-spark-1.2-contributor-free": { graceMs: 5_000 },
@@ -1266,50 +1271,6 @@ export const PROVIDER_REGISTRY_EXTENDED: readonly ProviderRegistryEntry[] = [
     models: ["glm-5.3-flash", "claude-sonnet-5-5", "gpt-6.1-sol", "deepseek-v4-flash"],
     note: "OpenAI-compatible relay for Claude, GPT, GLM and DeepSeek chat models. The seeded models were verified with chat, streaming and tool calls; deepseek-v4.1-flash is listed upstream but currently answers 503 and is left out. An ordinary key only reaches its own routing group's models; a Smart key reaches all seeded models. Image and video models are not served through this Chat preset. Upstream latency varies widely (a few seconds to about 30 s) with occasional transient 503s.",
   },
-  {
-    id: "opencode-free",
-    label: "OpenCode Free",
-    adapter: "openai-chat",
-    baseUrl: "https://opencode.ai/zen/v1",
-    authKind: "key",
-    keyOptional: true,
-    featured: true,
-    liveModels: true,
-    note: "No key needed, but OpenCode now gates this tier to its own client: Zen refuses any request that arrives without an x-opencode-session header (error type MissingSessionID, \"OpenCode's free tier can only be used in OpenCode\"). opencodex does not mint that header or claim an OpenCode client identity, because no upstream contract authorizes a third-party agent to present itself as OpenCode. Until OpenCode publishes a third-party integration path for the keyless tier, use the keyed opencode-zen provider instead (https://opencode.ai/auth). Quota figures for when the tier admitted a request: OpenCode advertises about 200 Big Pickle/free-model requests per 5 hours, and the same Zen gateway can short-window rate-limit free models at roughly 15-20 requests/minute, and may return generic 429s without Retry-After (opencodex synthesizes backoff only when that header is omitted). Free models are discovered live from Zen. Data use: per OpenCode's Zen docs (https://opencode.ai/docs/zen/), prompts sent to free models may be retained and used for training/improvement — do not send confidential material through this provider.",
-    dashboardUrl: "https://opencode.ai",
-    staticHeaders: {
-      // Zen answers a bare runtime User-Agent (Bun/x.y.z) more aggressively than a client
-      // that identifies itself, which is what the 429 in #2067 traced to. The value is
-      // deliberately unversioned: a pinned "opencode-cli/<version>" is a claim about an
-      // install we do not have and goes stale on the vendor's schedule, not ours.
-      // Corroboration, not authority: OmniRoute — an independent open-source broker against
-      // the same Zen upstream — defaults to exactly this pair (userAgent "opencode", client
-      // "desktop") in open-sse/executors/opencode.ts, and got there by RETREATING from its
-      // own earlier "opencode-cli/1.0.0" pin. An operator can still override either value
-      // through the provider headers API; user headers win case-insensitively at route time.
-      "User-Agent": "opencode",
-      "x-opencode-client": "desktop",
-    },
-    modelReasoningEfforts: Object.fromEntries(OPENCODE_FREE_DEEPSEEK_MODELS.map(id => [id, deepseekThinkingEffortsFor(id)])),
-    modelReasoningEffortMap: Object.fromEntries(OPENCODE_FREE_DEEPSEEK_MODELS.map(id => [id, deepseekReasoningMapFor(id)])),
-    preserveReasoningContentModels: OPENCODE_FREE_DEEPSEEK_MODELS,
-    // The DeepSeek vision preview id is preemptive metadata for when Zen starts
-    // serving it (merges into v4-flash later).
-    modelContextWindows: {
-      [DEEPSEEK_VISION_PREVIEW_MODEL]: 1_048_576,
-    },
-    modelInputModalities: {
-      [DEEPSEEK_VISION_PREVIEW_MODEL]: ["text", "image"],
-      ...Object.fromEntries(OPENCODE_ZEN_IMAGE_MODELS.map(id => [id, ["text", "image"] as string[]])),
-    },
-    // Same Zen roster behind the same base URL, so it carries the same measured
-    // text-only list rather than only its DeepSeek member (#1043).
-    noVisionModels: OPENCODE_ZEN_TEXT_ONLY_MODELS,
-    // Same reasoning: the free tier is the same Zen roster, so its DeepSeek members get
-    // the keyed tier's json_schema treatment and its reasoning contract rather than a
-    // narrower table that silently falls behind whenever the keyed one is updated.
-    noJsonSchemaModels: [...DEEPSEEK_GATEWAY_THINKING_MODELS, ...OPENCODE_FREE_DEEPSEEK_MODELS],
-  },
   // Xiaomi retires mimo-v2.5 and mimo-v2.5-pro on 2026-10-21 with no redirect
   // (https://mimo.mi.com/docs/en-US/updates/deprecate), so the first-party presets default to V2.6.
   // Saved defaults are not rewritten; V2.5 stays listed until it stops answering.

package/src/providers/registry/model-seeds.ts

@@ -477,7 +477,7 @@ export const COMMAND_CODE_MODEL_INPUT_MODALITIES: Record<string, ["text"] | ["te
   ...Object.fromEntries(COMMAND_CODE_IMAGE_MODELS.map(id => [id, ["text", "image"] as ["text", "image"]])),
   ...Object.fromEntries(COMMAND_CODE_TEXT_ONLY_MODELS.map(id => [id, ["text"] as ["text"]])),
 };
-export const OPENCODE_FREE_DEEPSEEK_MODELS = ["deepseek-v4-flash-free"];
+export const OPENCODE_ZEN_FREE_DEEPSEEK_MODELS = ["deepseek-v4-flash-free"];
 /*
  * Zen free models that reject `image_url` upstream (#1043, and the reproducible
  * half of #1024).

package/src/providers/registry.ts

@@ -49,8 +49,7 @@ export function getProviderRegistryEntry(id: string): ProviderRegistryEntry | un
  * was only ever true for a freshly seeded config: `providerConfigSeed` copies the block once
  * (`derive.ts`), `enrichProviderFromCatalog` fills it only when the whole block is absent, and
  * nothing merged it at request time. So an install that predates a header — or that saved any
- * header of its own — never received the new one, which is exactly what #2067 would have
- * shipped for every existing opencode-free user.
+ * header of its own — never received the new one.
  *
  * The comparison is case-insensitive on purpose. HTTP header names are case-insensitive, but a
  * plain object spread is not: merging a registry `User-Agent` over a user's `user-agent`

package/src/providers/routed-config.ts

@@ -134,7 +134,7 @@ export function routedProviderConfig(providerName: string, provider: OcxProvider
   const repairLegacyMimoFreeAuth = providerName === "mimo-free"
     && staticModelCatalog
     && (provider.authMode === undefined || provider.authMode === "local");
-  const explicitKeyOverride = providerUsesKeyAuthOverride(registryEntry, provider, resolvedApiKey);
+  const explicitKeyOverride = providerUsesKeyAuthOverride(registryEntry, provider);
   const canonicalAuthMode = explicitKeyOverride
     ? "key"
     : repairLegacyMimoFreeAuth

package/src/providers/stale-vision-classification-migration.ts

@@ -39,7 +39,7 @@
  * Scope is deliberately one entry. The Go gateway's `deepseek-v4.1-flash` was declared text-only
  * from jawcode metadata and was measured natively multimodal on 2026-09-19 (see the note at that
  * entry in `registry/entries-core.ts`). Its sibling `deepseek-v4-flash` still rejects images
- * upstream, and the sibling Zen tiers (`opencode-zen`, `opencode-free`) could not be probed at
+ * upstream, and the keyed Zen tier (`opencode-zen`) could not be probed at
  * all — an unverified tier is not evidence, so neither is touched here.
  */
 import { PROVIDER_REGISTRY, providerMatchesRegistryTransport } from "./registry";

package/src/providers/xai-transport.ts

@@ -3,6 +3,7 @@ import type { OcxProviderConfig } from "../types";
 import { isEgressTransparentExecutor, markEgressTransparentExecutor } from "../lib/provider-egress";
 import { configuredOutboundFetch } from "../lib/proxy-env";
 import { resolveGithubCopilotTransport } from "./github-copilot-transport";
+import { resolveMistralTransport } from "./mistral-transport";
 
 export const XAI_GROK_CLI_BASE_URL = "https://cli-chat-proxy.grok.com/v1";
 
@@ -121,7 +122,7 @@ export function resolveProviderTransport(
   if (providerName === "github-copilot") {
     return resolveGithubCopilotTransport(provider, apiBaseUrl);
   }
-  if (providerName !== "xai") return provider;
+  if (providerName !== "xai") return resolveMistralTransport(provider);
 
   const cacheKey = promptCacheKey?.trim();
   const affinity = cacheKey ? deriveXaiConvId(cacheKey) : undefined;

package/src/responses/audio.ts

@@ -0,0 +1,46 @@
+import type { OcxAudioContent } from "../types";
+
+export type ChatAudioFormat = OcxAudioContent["format"];
+
+const MIME_FORMATS: Readonly<Record<string, ChatAudioFormat>> = {
+  "audio/mp3": "mp3",
+  "audio/mpeg": "mp3",
+  "audio/wav": "wav",
+  "audio/wave": "wav",
+  "audio/x-wav": "wav",
+};
+
+function normalizedFormat(value: unknown): ChatAudioFormat | undefined {
+  if (typeof value !== "string") return undefined;
+  const format = value.trim().toLowerCase();
+  return format === "mp3" || format === "wav" ? format : undefined;
+}
+
+function validBase64(value: string): boolean {
+  return value.length > 0
+    && value.length % 4 === 0
+    && /^[A-Za-z0-9+/]+={0,2}$/.test(value);
+}
+
+/**
+ * Parse the inline audio shape Codex puts in a Responses content block. Chat Completions
+ * needs bare base64 plus `wav`/`mp3`, while Codex supplies an RFC 2397 data URL. A raw
+ * base64 value with an explicit supported format is accepted as well; remote URLs are not.
+ */
+export function audioContentFromRaw(audioUrl: unknown, declaredFormat?: unknown): OcxAudioContent | undefined {
+  if (typeof audioUrl !== "string" || audioUrl.length === 0) return undefined;
+  const dataUrl = /^data:([^;,]+);base64,([\s\S]*)$/.exec(audioUrl);
+  let data: string;
+  let format: ChatAudioFormat | undefined;
+  if (dataUrl) {
+    const declared = normalizedFormat(declaredFormat);
+    const mime = dataUrl[1].trim().toLowerCase();
+    format = MIME_FORMATS[mime] ?? (mime === "application/octet-stream" ? declared : undefined);
+    data = dataUrl[2].replace(/\s+/g, "");
+  } else {
+    format = normalizedFormat(declaredFormat);
+    data = audioUrl.replace(/\s+/g, "");
+  }
+  if (!format || !validBase64(data)) return undefined;
+  return { type: "audio", data, format, text: `[audio: ${format}]` };
+}

package/src/responses/input-media.ts

@@ -1,4 +1,5 @@
 import { isInlineDocumentDataUrl } from "./inline-document";
+import { audioContentFromRaw } from "./audio";
 
 /** Input kinds for which the normalized request has no lossless content carrier. */
 export type UntranslatedInputMedia = "audio" | "file";
@@ -10,23 +11,35 @@ function isRecord(value: unknown): value is RecordValue {
 }
 
 /**
- * Whether the position being scanned has a converter that builds a document part.
+ * Whether the position being scanned has a converter that builds a lossless media part.
  *
- * Only user content does. A tool output, a system or assistant message, and a Chat `developer`
- * message are all flattened to text by their converters, so exempting an attachment there would
- * turn today's explicit refusal into the silent drop this scanner exists to prevent.
+ * User content carries documents and inline audio; developer content carries documents only.
+ * Tool output carries inline audio through the Chat adapter's follow-up user message. System and
+ * assistant content have no lossless translated media carrier, so exempting an attachment there
+ * would turn today's explicit refusal into the silent drop this scanner exists to prevent.
  */
-type CarrierPosition = "user-content" | "flattened";
+type CarrierPosition = "user-content" | "developer-content" | "tool-output" | "flattened";
 
-function mediaKind(value: unknown, position: CarrierPosition): UntranslatedInputMedia | undefined {
+function mediaKind(value: unknown, position: CarrierPosition, allowAudio: boolean): UntranslatedInputMedia | undefined {
   if (!isRecord(value)) return undefined;
-  if (value.type === "input_audio" || value.type === "audio") return "audio";
+  if (value.type === "input_audio" || value.type === "audio") {
+    // A Chat content part has a nested input_audio object, while Codex's Responses
+    // block uses audio_url. Only the latter is carried by the translated Chat adapter; the
+    // Chat-shaped `audio` spelling has no parser path and must never pass the guard.
+    return value.type === "input_audio"
+      && allowAudio
+      && (position === "user-content" || position === "tool-output")
+      && audioContentFromRaw(value.audio_url, value.format)
+      ? undefined
+      : "audio";
+  }
   if (value.type === "input_file" || value.type === "file" || value.type === "document") {
     // A document that carries its own bytes has a lossless carrier in user content, so refusing
     // it there would reject the very request #5212 exists to preserve. A reference with no
     // payload has none anywhere: translated adapters cannot dereference a file_id or a remote
     // source.
-    return position === "user-content" && carriesInlineDocumentBytes(value) ? undefined : "file";
+    return (position === "user-content" || position === "developer-content")
+      && carriesInlineDocumentBytes(value) ? undefined : "file";
   }
   // A file-id-only image is not pixels: translated adapters cannot dereference it.
   if (value.type === "input_image" && typeof value.file_id === "string" && value.file_id.length > 0
@@ -46,10 +59,10 @@ function carriesInlineDocumentBytes(value: RecordValue): boolean {
     && value.source.data.length > 0;
 }
 
-function contentMedia(content: unknown, position: CarrierPosition): UntranslatedInputMedia | undefined {
+function contentMedia(content: unknown, position: CarrierPosition, allowAudio: boolean): UntranslatedInputMedia | undefined {
   if (!Array.isArray(content)) return undefined;
   for (const part of content) {
-    const kind = mediaKind(part, position);
+    const kind = mediaKind(part, position, allowAudio);
     if (kind) return kind;
   }
   return undefined;
@@ -60,20 +73,29 @@ function contentMedia(content: unknown, position: CarrierPosition): Untranslated
  * arguments, schema properties, or arbitrary nested objects. No payload is copied,
  * decoded, fetched or included in the returned value.
  */
-export function untranslatedResponsesInputMedia(body: unknown): UntranslatedInputMedia | undefined {
+export function untranslatedResponsesInputMedia(
+  body: unknown,
+  options: { allowAudio?: boolean } = {},
+): UntranslatedInputMedia | undefined {
   if (!isRecord(body) || !Array.isArray(body.input)) return undefined;
   for (const item of body.input) {
     if (!isRecord(item)) continue;
-    const direct = mediaKind(item, "flattened");
+    const direct = mediaKind(item, "flattened", options.allowAudio === true);
     if (direct) return direct;
     if (item.type === "function_call_output" || item.type === "custom_tool_call_output") {
-      const kind = contentMedia(item.output, "flattened");
+      const kind = contentMedia(item.output, "tool-output", options.allowAudio === true);
       if (kind) return kind;
     } else if (item.type === "message" || item.type === undefined) {
       // `inputContentParts` runs for user and developer messages only; a system message is
       // flattened to text by the parser.
       const role = item.role;
-      const kind = contentMedia(item.content, role === "user" || role === "developer" ? "user-content" : "flattened");
+      const kind = contentMedia(
+        item.content,
+        role === "user" ? "user-content"
+          : role === "developer" ? "developer-content"
+          : "flattened",
+        options.allowAudio === true,
+      );
       if (kind) return kind;
     }
   }
@@ -87,7 +109,7 @@ export function untranslatedChatInputMedia(body: unknown): UntranslatedInputMedi
     if (!isRecord(message)) continue;
     // Only the `user` branch of the Chat projection builds content blocks. `system`,
     // `developer`, `assistant` and `tool` all reduce their content to a string.
-    const kind = contentMedia(message.content, message.role === "user" ? "user-content" : "flattened");
+    const kind = contentMedia(message.content, message.role === "user" ? "user-content" : "flattened", false);
     if (kind) return kind;
   }
   return undefined;

package/src/responses/parser-content.ts

@@ -1,4 +1,5 @@
 import type { OcxContentPart, OcxTextContent } from "../types";
+import { audioContentFromRaw } from "./audio";
 import { inlineDocumentFromDataUrl } from "./inline-document";
 
 export function isObj(v: unknown): v is Record<string, unknown> {
@@ -65,23 +66,15 @@ export function inputContentParts(blocks: unknown): string | OcxContentPart[] {
       const processing = nonEmptyString((block as { processing?: string }).processing);
       if (videoUrl) parts.push({ type: "video", videoUrl, ...(processing ? { processing } : {}) });
     } else if (block.type === "input_audio") {
-      // Upstream Codex sends input_audio with an audio_url (codex-rs
-      // protocol/src/models.rs). The IR has no audio carrier and no adapter consumes
-      // one, so this part used to vanish with no trace at all.
-      //
-      // This records PRESENCE only and is NOT audio support: never the payload, which
-      // is large base64 and would explode the token count, and never the URL, which
-      // can carry a signed token. This parser must stay non-throwing — the native
-      // Responses passthrough also runs through parseRequest before the adapter
-      // forwards _rawBody, so refusing here would regress legitimate raw passthrough.
-      //
-      // No adapter refuses audio at its wire today: by this point the part is already a
-      // text marker, so downstream adapters see text and continue. A typed unsupported-
-      // modality carrier that survives to final adapter dispatch is a separate, recorded
-      // residual — do not describe this branch as a refusal.
       const b = block as { audio_url?: string; format?: string };
-      const format = safeAudioFormat(b.format);
-      if (nonEmptyString(b.audio_url)) {
+      // Keep bytes only when they have a Chat-compatible inline form. Unsupported or
+      // remote references retain a marker for text-only wires; the final adapter guard
+      // refuses them rather than allowing a silent drop.
+      const audio = audioContentFromRaw(b.audio_url, b.format);
+      if (audio) {
+        parts.push(audio);
+      } else if (nonEmptyString(b.audio_url)) {
+        const format = safeAudioFormat(b.format);
         parts.push({ type: "text", text: format ? `[audio: ${format}]` : "[audio]" });
       }
     } else if (block.type === "input_file") {
@@ -104,7 +97,7 @@ export function inputContentParts(blocks: unknown): string | OcxContentPart[] {
       // fabricating a "[file: ...]" marker for an attachment that was never sent.
     }
   }
-  // Collapse to a plain string only for a single TEXT part; images must stay structured.
+  // Collapse to a plain string only for a single TEXT part; media must stay structured.
   if (parts.length === 1 && parts[0].type === "text") return parts[0].text;
   return parts;
 }
@@ -129,15 +122,15 @@ export function outputTextOf(blocks: unknown): OcxTextContent[] {
 }
 
 /**
- * Tool-call output content. Preserves images (e.g. Codex `view_image` returns
- * `input_image` items): returns content parts when any image is present, else a plain joined string.
- * Never inlines an image_url as text (that would explode the token count).
+ * Tool-call output content. Preserves images (e.g. Codex `view_image` returns)
+ * and inline audio: returns content parts when either is present, else a plain joined string.
+ * Never inlines an image_url or audio payload as text (that would explode the token count).
  */
 export function outputToToolResultContent(output: string | unknown[] | undefined): string | OcxContentPart[] {
   if (typeof output === "string") return output;
   if (!Array.isArray(output)) return "";
   const parts: OcxContentPart[] = [];
-  let hasImage = false;
+  let hasMedia = false;
   for (const raw of output) {
     if (!isObj(raw)) continue;
     if (raw.type === "output_text" || raw.type === "text" || raw.type === "input_text") {
@@ -149,15 +142,17 @@ export function outputToToolResultContent(output: string | unknown[] | undefined
       const fileId = nonEmptyString(raw.file_id);
       if (imageUrl) {
         parts.push({ type: "image", imageUrl, ...(typeof raw.detail === "string" ? { detail: normalizeImageDetail(raw.detail) } : {}) });
-        hasImage = true;
+        hasMedia = true;
       } else if (fileId) {
         parts.push({ type: "text", text: `[image: ${fileId}]` });
       }
     } else if (raw.type === "input_audio") {
-      // Same presence-only contract as the user-content branch above: Codex returns
-      // audio in tool output too, and it previously disappeared without trace.
-      const format = safeAudioFormat(raw.format);
-      if (nonEmptyString(raw.audio_url)) {
+      const audio = audioContentFromRaw(raw.audio_url, raw.format);
+      if (audio) {
+        parts.push(audio);
+        hasMedia = true;
+      } else if (nonEmptyString(raw.audio_url)) {
+        const format = safeAudioFormat(raw.format);
         parts.push({ type: "text", text: format ? `[audio: ${format}]` : "[audio]" });
       }
     } else if (raw.type === "encrypted_content") {
@@ -165,7 +160,7 @@ export function outputToToolResultContent(output: string | unknown[] | undefined
       parts.push({ type: "text", text: "[encrypted content omitted]" });
     }
   }
-  if (!hasImage) return parts.map(p => (p.type === "text" ? p.text : "")).join("");
+  if (!hasMedia) return parts.map(p => (p.type === "text" ? p.text : "")).join("");
   return parts;
 }
 

package/src/responses/parser.ts

@@ -22,6 +22,7 @@ import { previousResponseReplayPrefixLength } from "./state";
 import { decodeReasoningEnvelope } from "./reasoning-envelope";
 import { hasRoutedIdentity, nameRoutedIdentity } from "../adapters/identity";
 import { extractHostedWebSearch, WEB_SEARCH_TOOL_NAME } from "../web-search/synthetic-tool";
+import { rewriteReplayedWebSearchCells } from "./web-search-replay";
 import { buildImageTool, extractHostedImageGeneration, IMAGE_GEN_TOOL_NAME } from "../images/synthetic-tool";
 import { toolSearchDescription, toolSearchParameters } from "./tool-search-compat";
 
@@ -232,6 +233,13 @@ export function parseRequest(
   // Built before the item loop: a custom_tool_call echoed back in `input` needs the
   // namespace from the request's own tool catalog to survive the round trip.
   const customToolNamespacesByName = customToolNamespaces(data.tools);
+  // Stash the hosted web_search config (if Codex enabled it) so the proxy can run searches via the
+  // sidecar for routed providers. buildTools still drops the hosted tool; the sidecar path
+  // re-injects a synthetic function tool only when it will actually handle the call. The same
+  // declaration decides how replayed search cells are rebuilt below.
+  const webSearch = extractHostedWebSearch(data.tools as unknown[] | undefined);
+  // With search declared, replayed hosted cells become web_search call/output pairs (#7269).
+  const webSearchReplay = webSearch && Array.isArray(data.input) ? rewriteReplayedWebSearchCells(data.input) : undefined;
   const systemPrompt: string[] = [];
   // Responses reasoning siblings belong to the following assistant, including across call items.
   // Keep them off the message list until that assistant arrives; turn boundaries clear the array.
@@ -274,9 +282,12 @@ export function parseRequest(
     if (data.previous_response_id) continuationConversationMessageIndex = messages.length;
     messages.push({ role: "user", content: data.input, timestamp: now });
   } else if (data.input) {
-    for (let inputIndex = 0; inputIndex < data.input.length; inputIndex++) {
-      const item = data.input[inputIndex];
-      if (inputIndex > 0 && isAnthropicProgressMirror(item, anthropicProgressEnvelope(data.input[inputIndex - 1]))) continue;
+    const inputItems = (webSearchReplay?.items ?? data.input) as typeof data.input;
+    for (let itemIndex = 0; itemIndex < inputItems.length; itemIndex++) {
+      const item = inputItems[itemIndex];
+      // Boundaries are measured on the caller's input, before any replayed-search rewrite.
+      const inputIndex = webSearchReplay?.sourceIndexes[itemIndex] ?? itemIndex;
+      if (itemIndex > 0 && isAnthropicProgressMirror(item, anthropicProgressEnvelope(inputItems[itemIndex - 1]))) continue;
       const effectiveType = (item as { type?: string }).type ?? ("role" in item ? "message" : undefined);
       const itemRole = (item as { role?: string }).role;
       const externalTaskInput = effectiveType === "function_call_output" ? externalTaskInputContent(item) : undefined;
@@ -529,9 +540,9 @@ export function parseRequest(
       }
 
       if (effectiveType === "web_search_call") {
-        // Replayed hosted web-search evidence has no paired result payload that routed providers can
-        // consume. Keep it out of assistant-visible text: the old marker was useful as an internal
-        // loop hint, but when no sidecar is available the model can echo it as a fake answer.
+        // Without a search declaration the cell stays out of routed history: a call to an undeclared
+        // tool, or assistant-visible marker text the model could echo as a fake answer, would be
+        // worse than the gap. With one, the cell was already rewritten into a web_search pair.
         pendingReasoning.length = 0;
         continue;
       }
@@ -689,10 +700,6 @@ export function parseRequest(
   if (data.service_tier !== undefined) options.serviceTier = data.service_tier;
   if (data.prompt_cache_key !== undefined) options.promptCacheKey = data.prompt_cache_key;
 
-  // Stash the hosted web_search config (if Codex enabled it) so the proxy can run searches via the
-  // gpt-mini sidecar for routed providers. buildTools still drops the hosted tool; the sidecar path
-  // re-injects a synthetic function tool only when it will actually handle the call.
-  const webSearch = extractHostedWebSearch(data.tools as unknown[] | undefined);
   const imageGen = extractHostedImageGeneration([
     ...(data.tools as unknown[] ?? []),
     ...loadedToolSpecs,
@@ -718,6 +725,7 @@ export function parseRequest(
       ? { _continuationConversationMessageIndex: continuationConversationMessageIndex }
       : {}),
     ...(webSearch ? { _webSearch: webSearch } : {}),
+    ...(webSearchReplay ? { _replayedWebSearchCells: webSearchReplay.cells } : {}),
     ...(imageGen ? { _imageGeneration: imageGen } : {}),
     ...(textFormat ? { _structuredOutput: true } : {}),
     ...(compactionRequest ? { _compactionRequest: true } : {}),