hotfix(image-generation): fix gpt-image-2 / gpt-image-1.5 failures and stuck pending placeholder (#14578)

### What this PR does Before this PR: - `gpt-image-2` (and any image-edit request routed through AiHubMix / NewAPI / CherryIN) failed with `common.UnmarshalBodyReusable failed: bind request body failed: invalid character '-' in numeric literal`. The providers hard-coded `Content-Type: application/json` in `authHeaders()`, but `OpenAICompatibleImageModel.doGenerate` uses `postFormDataToApi` for `/images/edits` and relies on `fetch` to auto-set `multipart/form-data; boundary=...`. Forcing JSON made the server receive a multipart body under a JSON content-type and choke on the leading `--boundary`. - Even when the request reached the server, image uploads failed with `unsupported mimetype ('image')` because `collectImagesFromMessages` built data URIs using `block.file.type` — which is the `FileType` enum (`"image"`), not a MIME type. - Some saved/pasted images produced `ENOENT: ... <uuid>jpg` because `saveBase64Image` / `savePastedImage` return `ext` without the leading dot, and the caller was concatenating `file.id + file.ext`. - After a successful image generation, a `BeatLoader` placeholder kept spinning next to the finished image. `fetchImageGeneration` only emitted `LLM_RESPONSE_COMPLETE`; the trailing loader in `Blocks/index.tsx` is driven by `isMessageProcessing(message)`, which is only cleared when `onComplete` runs on `BLOCK_COMPLETE`. - If a `file` chunk arrived without a preceding `IMAGE_CREATED`, the initial UNKNOWN placeholder block was never claimed, leaving an orphan spinner below the image. - `gpt-image-2` / `gpt-image-1.5` / `gpt-image-1*` / `chatgpt-image-*` generations through any OpenAI-compatible provider still failed with `400 Unknown parameter: 'response_format'` (see #14485, #14540, #14579). The earlier `@ai-sdk/openai@3.0.53` patch (#14488) only covered `OpenAIImageModel` (direct OpenAI + Azure via `@ai-sdk/azure`); `OpenAICompatibleImageModel.doGenerate` unconditionally added `response_format: "b64_json"` to the `/images/generations` body, so every provider whose image model is wired to `OpenAICompatibleImageModel` kept hitting the 400 — including `type: 'openai-compatible'` and the AiHubMix / NewAPI / CherryIN gateway image models in this repo. After this PR: - `aihubmix-provider`, `newapi-provider`, and `cherryin-provider` no longer hard-code `Content-Type` in auth headers. `postJsonToApi` still defaults JSON for other endpoints, and `postFormDataToApi` is free to let `fetch` set `multipart/form-data; boundary=...` for `/images/edits`. - `collectImagesFromMessages` reads via `window.api.file.base64Image(block.file.name)`, which returns a correct `data:image/<ext>;base64,...` URI (with `jpg → jpeg` normalization) and uses the consistent on-disk filename regardless of whether the stored `ext` has a leading dot. - `fetchImageGeneration` now emits `ChunkType.BLOCK_COMPLETE` before `LLM_RESPONSE_COMPLETE`, so `onComplete` runs and the assistant message transitions out of `PROCESSING`. - `onImageGenerated` reuses the initial UNKNOWN placeholder when one exists, so a `file` chunk without a prior `IMAGE_CREATED` no longer leaves an orphan spinner. - `gpt-image-2` is registered in `IMAGE_ENHANCEMENT_MODELS`. - `patches/@ai-sdk__openai-compatible@2.0.37.patch` is extended with the same `hasDefaultResponseFormat` guard that `@ai-sdk/openai` already uses. `OpenAICompatibleImageModel.doGenerate` now spreads `response_format: "b64_json"` only when the model ID does **not** start with `chatgpt-image-`, `gpt-image-1-mini`, `gpt-image-1.5`, `gpt-image-1`, or `gpt-image-2`. This fixes #14485, #14540, #14579 for `openai-compatible` typed providers and the AiHubMix / NewAPI / CherryIN gateways. Fixes #14485 Fixes #14540 Fixes #14579 ### Why we need it and why it was done in this way The following tradeoffs were made: - The `ext`-with-or-without-dot inconsistency exists in historical DB records, so patching only `FileStorage.ts` would not help already-saved rows. The fix uses `file.name` at the consumer side — always the real on-disk filename across every save path — without touching the storage layer. - The provider `authHeaders()` change could theoretically affect non-JSON, non-multipart requests, but all SDK models in use go through `postJsonToApi`, which already defaults `Content-Type: application/json` internally, so behavior is preserved. - Extending the `@ai-sdk/openai-compatible` patch rather than overriding `response_format` per-provider keeps the fix in one place and automatically covers every caller of `OpenAICompatibleImageModel` (current and future). The guard list mirrors the upstream `@ai-sdk/openai` list, so the two stay in sync. The following alternatives were considered: - Normalizing `ext` in `FileStorage.saveBase64Image` / `savePastedImage`: rejected because it would not repair existing records and could regress other callers that have compensated for the dotless form. - Overriding `Content-Type` only inside the image model factory: rejected as more surface area than needed; the global `authHeaders()` already defers to `postJsonToApi` defaults. - Stripping `response_format` in each custom provider (AiHubMix / NewAPI / CherryIN) instead of patching the SDK: rejected — it would leave `type: 'openai-compatible'` users still broken and duplicate the model-prefix list in three places. Links to places where the discussion took place: #14485, #14540, #14579, #14488 ### Breaking changes  None. ### Special notes for your reviewer  Scope is intentionally minimal for the main branch freeze: no refactoring, no new abstractions. Surface changes: - `authHeaders()` in three custom providers (drops one hard-coded header). - `collectImagesFromMessages` (one helper, swaps input building). - `fetchImageGeneration` (one extra chunk emission). - `onImageGenerated` (placeholder-claim fallback). - `IMAGE_ENHANCEMENT_MODELS` (one new entry). - `patches/@ai-sdk__openai-compatible@2.0.37.patch` (adds `hasDefaultResponseFormat` helper + conditional spread of `response_format`; mirrors `@ai-sdk/openai@3.0.53` patch). Drop the openai-compatible patch addition once `@ai-sdk/openai-compatible` ships the equivalent check upstream. ### Checklist This checklist is not enforcing, but it's a reminder of items that could be relevant to every PR. Approvers are expected to review this list. - [x] PR: The PR description is expressive enough and will help future contributors - [x] Code: [Write code that humans can understand](https://en.wikiquote.org/wiki/Martin_Fowler#code-for-humans) and [Keep it simple](https://en.wikipedia.org/wiki/KISS_principle) - [x] Refactor: You have [left the code cleaner than you found it (Boy Scout Rule)](https://learning.oreilly.com/library/view/97-things-every/9780596809515/ch08.html) - [x] Upgrade: Impact of this change on upgrade flows was considered and addressed if required - [ ] Documentation: A [user-guide update](https://docs.cherry-ai.com) was considered and is present (link) or not required. Check this only when the PR introduces or changes a user-facing feature or behavior. - [x] Self-review: I have reviewed my own code (e.g., via [`/gh-pr-review`](/.claude/skills/gh-pr-review/SKILL.md), `gh pr diff`, or GitHub UI) before requesting review from others ### Release note  ```release-note Fix image generation and editing for gpt-image-2 / gpt-image-1.5 / gpt-image-1* / chatgpt-image-* across AiHubMix, NewAPI, CherryIN, Azure OpenAI, and other OpenAI-compatible providers: requests no longer fail with `Unknown parameter: 'response_format'`, multipart content-type, or mimetype errors; generated/pasted images are resolved correctly; and the loading placeholder disappears once the image is rendered. ``` --------- Signed-off-by: suyao <sy20010504@gmail.com>
2026-07-07 15:12:18 +08:00 · 2026-04-26 00:34:41 +08:00
parent 8ed7c0754a
commit 26d877e0fa
10 changed files with 142 additions and 54 deletions
--- a/.changeset/cherryin-image-edit-content-type.md
+++ b/.changeset/cherryin-image-edit-content-type.md
@@ -0,0 +1,7 @@
+---
+'@cherrystudio/ai-sdk-provider': patch
+---
+
+Fix CherryIN image-edit requests failing with `invalid character '-' in numeric literal`.
+
+The JSON headers getter hard-coded `Content-Type: application/json`, which leaked into `OpenAICompatibleImageModel`'s `/images/edits` call. That path uses `postFormDataToApi` and relies on `fetch` to auto-set `multipart/form-data; boundary=...`; forcing JSON made the server try to parse the multipart body as JSON and choke on the leading `--boundary`. Removed the explicit `Content-Type` — `postJsonToApi` still defaults it for JSON endpoints.
--- a/.changeset/openai-compatible-image-response-format.md
+++ b/.changeset/openai-compatible-image-response-format.md
@@ -0,0 +1,8 @@
+---
+'@cherrystudio/ai-core': patch
+'@cherrystudio/ai-sdk-provider': patch
+---
+
+Fix `Unknown parameter: 'response_format'` when generating images with `gpt-image-2`, `gpt-image-1.5`, `gpt-image-1`, `gpt-image-1-mini`, or `chatgpt-image-*` through any provider that routes the image model via `OpenAICompatibleImageModel` (e.g. `openai-compatible` typed providers and the AiHubMix / NewAPI / CherryIN gateways in this repo).
+
+`OpenAICompatibleImageModel.doGenerate` unconditionally added `response_format: "b64_json"` to `/images/generations` bodies. The previous `@ai-sdk/openai@3.0.53` patch only covered `OpenAIImageModel` (direct OpenAI + Azure via `@ai-sdk/azure`), so users on the compatible route (AiHubMix, NewAPI, CherryIN, generic openai-compatible) kept hitting this 400 on newer gpt-image models. Extended `patches/@ai-sdk__openai-compatible@2.0.37.patch` with the same `hasDefaultResponseFormat` guard used upstream by `@ai-sdk/openai`. Drop the addition once `@ai-sdk/openai-compatible` ships the equivalent check.
--- a/packages/ai-sdk-provider/src/cherryin-provider.ts
+++ b/packages/ai-sdk-provider/src/cherryin-provider.ts
@@ -137,7 +137,6 @@ const toBearerToken = (authorization?: string) => (authorization ? authorization
 const createJsonHeadersGetter = (options: CherryInProviderSettings): (() => Record<string, HeaderValue>) => {
  return () => ({
    Authorization: `Bearer ${resolveApiKey(options)}`,
-    'Content-Type': 'application/json',
    ...resolveConfiguredHeaders(options.headers)
  })
 }
--- a/patches/@ai-sdk__openai-compatible@2.0.37.patch
+++ b/patches/@ai-sdk__openai-compatible@2.0.37.patch
@@ -1,11 +1,11 @@
 diff --git a/dist/index.js b/dist/index.js
-index d73573b013e61903991cad2074187a3ade8921ab..919cbb2d14c0019716080c1bf041dde96c2b6cab 100644
 --- a/dist/index.js
 +++ b/dist/index.js
-@@ -572,6 +572,17 @@ var OpenAICompatibleChatLanguageModel = class {
+@@ -571,6 +571,17 @@
+         type: "reasoning",
         text: reasoning
       });
-     }
+    }
 +    if (choice.message.images) {
 +      for (const image of choice.message.images) {
 +        const match1 = image.image_url.url.match(/^data:([^;]+)/);
@@ -16,11 +16,10 @@ index d73573b013e61903991cad2074187a3ade8921ab..919cbb2d14c0019716080c1bf041dde9
 +          data: match2 ? match2[1] : image.image_url.url,
 +        });
 +      }
-+    }
+     }
     if (choice.message.tool_calls != null) {
       for (const toolCall of choice.message.tool_calls) {
-         const thoughtSignature = (_c = (_b = toolCall.extra_content) == null ? void 0 : _b.google) == null ? void 0 : _c.thought_signature;
-@@ -732,6 +743,17 @@ var OpenAICompatibleChatLanguageModel = class {
+@@ -732,6 +743,17 @@
                 delta: delta.content
               });
             }
@@ -38,10 +37,11 @@ index d73573b013e61903991cad2074187a3ade8921ab..919cbb2d14c0019716080c1bf041dde9
             if (delta.tool_calls != null) {
               if (isActiveReasoning) {
                 controller.enqueue({
-@@ -928,6 +950,14 @@ var OpenAICompatibleChatResponseSchema = import_v43.z.looseObject({
+@@ -927,6 +949,14 @@
+                 thought_signature: import_v43.z.string().nullish()
               }).nullish()
             }).nullish()
-           })
+          })
 +        ).nullish(),
 +        images: import_v43.z.array(
 +          import_v43.z.object({
@@ -49,14 +49,14 @@ index d73573b013e61903991cad2074187a3ade8921ab..919cbb2d14c0019716080c1bf041dde9
 +            image_url: import_v43.z.object({
 +              url: import_v43.z.string(),
 +            })
-+          })
+           })
         ).nullish()
       }),
-       finish_reason: import_v43.z.string().nullish()
-@@ -964,6 +994,14 @@ var chunkBaseSchema = import_v43.z.looseObject({
+@@ -963,6 +993,14 @@
+                 thought_signature: import_v43.z.string().nullish()
               }).nullish()
             }).nullish()
-           })
+          })
 +        ).nullish(),
 +        images: import_v43.z.array(
 +          import_v43.z.object({
@@ -64,18 +64,43 @@ index d73573b013e61903991cad2074187a3ade8921ab..919cbb2d14c0019716080c1bf041dde9
 +            image_url: import_v43.z.object({
 +              url: import_v43.z.string(),
 +            })
-+          })
+           })
         ).nullish()
       }).nullish(),
-       finish_reason: import_v43.z.string().nullish()
+@@ -1529,6 +1567,16 @@
+ // src/image/openai-compatible-image-model.ts
+ var import_provider_utils5 = require("@ai-sdk/provider-utils");
+ var import_v48 = require("zod/v4");
+var defaultResponseFormatPrefixes = [
+  "chatgpt-image-",
+  "gpt-image-1-mini",
+  "gpt-image-1.5",
+  "gpt-image-1",
+  "gpt-image-2"
+];
+function hasDefaultResponseFormat(modelId) {
+  return defaultResponseFormatPrefixes.some((prefix) => modelId.startsWith(prefix));
+}
+ var OpenAICompatibleImageModel = class {
+   constructor(modelId, config) {
+     this.modelId = modelId;
+@@ -1625,7 +1673,7 @@
+         n,
+         size,
+         ...args,
+-        response_format: "b64_json"
+        ...(!hasDefaultResponseFormat(this.modelId) ? { response_format: "b64_json" } : {})
+       },
+       failedResponseHandler: (0, import_provider_utils5.createJsonErrorResponseHandler)(
+         (_e = this.config.errorStructure) != null ? _e : defaultOpenAICompatibleErrorStructure
 diff --git a/dist/index.mjs b/dist/index.mjs
-index 237c9e209da9da2d4d557eb309e1adc2157c1426..2e140f053bb9d7e88ed6b54238d355e7f5aae49b 100644
 --- a/dist/index.mjs
 +++ b/dist/index.mjs
-@@ -559,6 +559,17 @@ var OpenAICompatibleChatLanguageModel = class {
+@@ -558,6 +558,17 @@
+         type: "reasoning",
         text: reasoning
       });
-     }
+    }
 +    if (choice.message.images) {
 +      for (const image of choice.message.images) {
 +        const match1 = image.image_url.url.match(/^data:([^;]+)/);
@@ -86,11 +111,10 @@ index 237c9e209da9da2d4d557eb309e1adc2157c1426..2e140f053bb9d7e88ed6b54238d355e7
 +          data: match2 ? match2[1] : image.image_url.url,
 +        });
 +      }
-+    }
+     }
     if (choice.message.tool_calls != null) {
       for (const toolCall of choice.message.tool_calls) {
-         const thoughtSignature = (_c = (_b = toolCall.extra_content) == null ? void 0 : _b.google) == null ? void 0 : _c.thought_signature;
-@@ -719,6 +730,17 @@ var OpenAICompatibleChatLanguageModel = class {
+@@ -719,6 +730,17 @@
                 delta: delta.content
               });
             }
@@ -108,10 +132,11 @@ index 237c9e209da9da2d4d557eb309e1adc2157c1426..2e140f053bb9d7e88ed6b54238d355e7
             if (delta.tool_calls != null) {
               if (isActiveReasoning) {
                 controller.enqueue({
-@@ -915,6 +937,14 @@ var OpenAICompatibleChatResponseSchema = z3.looseObject({
+@@ -914,6 +936,14 @@
+                 thought_signature: z3.string().nullish()
               }).nullish()
             }).nullish()
-           })
+          })
 +        ).nullish(),
 +        images: z3.array(
 +          z3.object({
@@ -119,14 +144,14 @@ index 237c9e209da9da2d4d557eb309e1adc2157c1426..2e140f053bb9d7e88ed6b54238d355e7
 +            image_url: z3.object({
 +              url: z3.string(),
 +            })
-+          })
+           })
         ).nullish()
       }),
-       finish_reason: z3.string().nullish()
-@@ -951,6 +981,14 @@ var chunkBaseSchema = z3.looseObject({
+@@ -950,6 +980,14 @@
+                 thought_signature: z3.string().nullish()
               }).nullish()
             }).nullish()
-           })
+          })
 +        ).nullish(),
 +        images: z3.array(
 +          z3.object({
@@ -134,7 +159,32 @@ index 237c9e209da9da2d4d557eb309e1adc2157c1426..2e140f053bb9d7e88ed6b54238d355e7
 +            image_url: z3.object({
 +              url: z3.string(),
 +            })
-+          })
+           })
         ).nullish()
       }).nullish(),
-       finish_reason: z3.string().nullish()
+@@ -1543,6 +1581,16 @@
+   postJsonToApi as postJsonToApi4
+ } from "@ai-sdk/provider-utils";
+ import { z as z8 } from "zod/v4";
+var defaultResponseFormatPrefixes = [
+  "chatgpt-image-",
+  "gpt-image-1-mini",
+  "gpt-image-1.5",
+  "gpt-image-1",
+  "gpt-image-2"
+];
+function hasDefaultResponseFormat(modelId) {
+  return defaultResponseFormatPrefixes.some((prefix) => modelId.startsWith(prefix));
+}
+ var OpenAICompatibleImageModel = class {
+   constructor(modelId, config) {
+     this.modelId = modelId;
+@@ -1639,7 +1687,7 @@
+         n,
+         size,
+         ...args,
+-        response_format: "b64_json"
+        ...(!hasDefaultResponseFormat(this.modelId) ? { response_format: "b64_json" } : {})
+       },
+       failedResponseHandler: createJsonErrorResponseHandler4(
+         (_e = this.config.errorStructure) != null ? _e : defaultOpenAICompatibleErrorStructure
--- a/pnpm-lock.yaml
+++ b/pnpm-lock.yaml
@@ -44,7 +44,7 @@ patchedDependencies:
    hash: 3dbdd2fbbe5a44bc23a04b1195329dd8d78043447e925342ad5ed53c8c3b96ec
    path: patches/@ai-sdk__google@3.0.64.patch
  '@ai-sdk/openai-compatible@2.0.37':
-    hash: d24ef2c45b84eab205e51894a90c5c36e93e6e0b17f56533aa37145129539a1a
+    hash: 50ffd57b3f7d2de99f0315f0b3bca05cde1dc3d9f760d4f1be4143d8c6f260c7
    path: patches/@ai-sdk__openai-compatible@2.0.37.patch
  '@ai-sdk/openai@3.0.53':
    hash: 83a4d486601eff47ad8b5e22024878be2ced05b5b86e09884199a3f47fd53530
@@ -216,7 +216,7 @@ importers:
        version: 3.0.53(patch_hash=83a4d486601eff47ad8b5e22024878be2ced05b5b86e09884199a3f47fd53530)(zod@4.3.4)
      '@ai-sdk/openai-compatible':
        specifier: ^2.0.37
-        version: 2.0.37(patch_hash=d24ef2c45b84eab205e51894a90c5c36e93e6e0b17f56533aa37145129539a1a)(zod@4.3.4)
+        version: 2.0.37(patch_hash=50ffd57b3f7d2de99f0315f0b3bca05cde1dc3d9f760d4f1be4143d8c6f260c7)(zod@4.3.4)
      '@ai-sdk/perplexity':
        specifier: ^3.0.29
        version: 3.0.29(zod@4.3.4)
@@ -1306,7 +1306,7 @@ importers:
        version: 3.0.53(patch_hash=83a4d486601eff47ad8b5e22024878be2ced05b5b86e09884199a3f47fd53530)(zod@4.3.6)
      '@ai-sdk/openai-compatible':
        specifier: ^2.0.37
-        version: 2.0.37(patch_hash=d24ef2c45b84eab205e51894a90c5c36e93e6e0b17f56533aa37145129539a1a)(zod@4.3.6)
+        version: 2.0.37(patch_hash=50ffd57b3f7d2de99f0315f0b3bca05cde1dc3d9f760d4f1be4143d8c6f260c7)(zod@4.3.6)
      '@ai-sdk/provider':
        specifier: ^3.0.8
        version: 3.0.8
@@ -1343,7 +1343,7 @@ importers:
        version: 3.0.53(patch_hash=83a4d486601eff47ad8b5e22024878be2ced05b5b86e09884199a3f47fd53530)(zod@4.3.4)
      '@ai-sdk/openai-compatible':
        specifier: ^2.0.37
-        version: 2.0.37(patch_hash=d24ef2c45b84eab205e51894a90c5c36e93e6e0b17f56533aa37145129539a1a)(zod@4.3.4)
+        version: 2.0.37(patch_hash=50ffd57b3f7d2de99f0315f0b3bca05cde1dc3d9f760d4f1be4143d8c6f260c7)(zod@4.3.4)
      '@ai-sdk/provider':
        specifier: ^3.0.8
        version: 3.0.8
@@ -13036,13 +13036,13 @@ snapshots:
      '@ai-sdk/provider-utils': 4.0.23(zod@4.3.4)
      zod: 4.3.4

-  '@ai-sdk/openai-compatible@2.0.37(patch_hash=d24ef2c45b84eab205e51894a90c5c36e93e6e0b17f56533aa37145129539a1a)(zod@4.3.4)':
+  '@ai-sdk/openai-compatible@2.0.37(patch_hash=50ffd57b3f7d2de99f0315f0b3bca05cde1dc3d9f760d4f1be4143d8c6f260c7)(zod@4.3.4)':
    dependencies:
      '@ai-sdk/provider': 3.0.8
      '@ai-sdk/provider-utils': 4.0.23(zod@4.3.4)
      zod: 4.3.4

-  '@ai-sdk/openai-compatible@2.0.37(patch_hash=d24ef2c45b84eab205e51894a90c5c36e93e6e0b17f56533aa37145129539a1a)(zod@4.3.6)':
+  '@ai-sdk/openai-compatible@2.0.37(patch_hash=50ffd57b3f7d2de99f0315f0b3bca05cde1dc3d9f760d4f1be4143d8c6f260c7)(zod@4.3.6)':
    dependencies:
      '@ai-sdk/provider': 3.0.8
      '@ai-sdk/provider-utils': 4.0.23(zod@4.3.6)
@@ -15977,7 +15977,7 @@ snapshots:

  '@opeoginni/github-copilot-openai-compatible@1.0.0(patch_hash=4bea0e526136ed32d0be4849ec5cbb41bd5de7b7418dd6920598357438cfef25)(zod@4.3.4)':
    dependencies:
-      '@ai-sdk/openai-compatible': 2.0.37(patch_hash=d24ef2c45b84eab205e51894a90c5c36e93e6e0b17f56533aa37145129539a1a)(zod@4.3.4)
+      '@ai-sdk/openai-compatible': 2.0.37(patch_hash=50ffd57b3f7d2de99f0315f0b3bca05cde1dc3d9f760d4f1be4143d8c6f260c7)(zod@4.3.4)
      '@ai-sdk/provider': 3.0.8
      '@ai-sdk/provider-utils': 4.0.23(zod@4.3.4)
    transitivePeerDependencies:
--- a/src/renderer/src/aiCore/provider/custom/aihubmix-provider.ts
+++ b/src/renderer/src/aiCore/provider/custom/aihubmix-provider.ts
@@ -45,9 +45,13 @@ export function createAihubmix(options: AihubmixProviderSettings = {}): Aihubmix
  const resolveApiKey = () =>
    loadApiKey({ apiKey: options.apiKey, environmentVariableName: 'AIHUBMIX_API_KEY', description: 'AiHubMix' })

+  // Note: Do not hard-code `Content-Type: application/json` here. `postJsonToApi`
+  // already defaults it for JSON endpoints, while `postFormDataToApi` (used by
+  // `OpenAICompatibleImageModel` for `/images/edits`) relies on fetch to set
+  // `multipart/form-data; boundary=...` automatically — forcing JSON here breaks
+  // image edits with "invalid character '-' in numeric literal" on the server.
  const authHeaders = (): Record<string, string> => ({
    Authorization: `Bearer ${resolveApiKey()}`,
-    'Content-Type': 'application/json',
    ...APP_CODE_HEADER,
    ...options.headers
  })
--- a/src/renderer/src/aiCore/provider/custom/newapi-provider.ts
+++ b/src/renderer/src/aiCore/provider/custom/newapi-provider.ts
@@ -53,9 +53,13 @@ export function createNewApi(options: NewApiProviderSettings = {}): NewApiProvid
  const resolveApiKey = () =>
    loadApiKey({ apiKey: options.apiKey, environmentVariableName: 'NEWAPI_API_KEY', description: 'NewAPI' })

+  // Note: Do not hard-code `Content-Type: application/json` here. `postJsonToApi`
+  // already defaults it for JSON endpoints, while `postFormDataToApi` (used by
+  // `OpenAICompatibleImageModel` for `/images/edits`) relies on fetch to set
+  // `multipart/form-data; boundary=...` automatically — forcing JSON here breaks
+  // image edits with "invalid character '-' in numeric literal" on the server.
  const authHeaders = (): Record<string, string> => ({
    Authorization: `Bearer ${resolveApiKey()}`,
-    'Content-Type': 'application/json',
    ...options.headers
  })

--- a/src/renderer/src/config/models/vision.ts
+++ b/src/renderer/src/config/models/vision.ts
@@ -118,6 +118,7 @@ const IMAGE_ENHANCEMENT_MODELS = [
  'grok-2-image(?:-[\\w-]+)?',
  'qwen-image-edit',
  'gpt-image-1',
+  'gpt-image-2',
  'gemini-2.5-flash-image(?:-[\\w-]+)?',
  'gemini-2.0-flash-preview-image-generation',
  'gemini-3(?:\\.\\d+)?-(?:flash|pro)-image(?:-[\\w-]+)?'
--- a/src/renderer/src/services/ApiService.ts
+++ b/src/renderer/src/services/ApiService.ts
@@ -38,7 +38,6 @@ import {
  getQuickModel
 } from './AssistantService'
 import { ConversationService } from './ConversationService'
-import FileManager from './FileManager'
 import { injectUserMessageWithKnowledgeSearchPrompt } from './KnowledgeService'
 import type { BlockManager } from './messageStreaming'
 import type { StreamProcessorCallbacks } from './StreamProcessingService'
@@ -312,12 +311,18 @@ async function collectImagesFromMessages(userMessage: Message, assistantMessage?
  const images: string[] = []

  // 收集用户消息中的图像
+  // NOTE: Use `block.file.name` (always the on-disk filename) rather than
+  // `block.file.id + block.file.ext` — some save paths (saveBase64Image,
+  // savePastedImage) store `ext` without the leading dot, so concatenation
+  // produces broken paths like `<uuid>jpg` → ENOENT.
+  // Also note: `block.file.type` is a FileType enum (e.g. "image"), NOT a MIME
+  // type. `base64Image` derives the real MIME from the extension internally
+  // (and normalizes jpg → image/jpeg).
  const userImageBlocks = findImageBlocks(userMessage)
  for (const block of userImageBlocks) {
    if (block.file) {
-      const base64 = await FileManager.readBase64File(block.file)
-      const mimeType = block.file.type || 'image/png'
-      images.push(`data:${mimeType};base64,${base64}`)
+      const { data } = await window.api.file.base64Image(block.file.name)
+      images.push(data)
    }
  }

@@ -401,17 +406,20 @@ export async function fetchImageGeneration({
      image: { type: imageType, images }
    })

-    onChunkReceived({
-      type: ChunkType.LLM_RESPONSE_COMPLETE,
-      response: {
-        usage: { prompt_tokens: 0, completion_tokens: 0, total_tokens: 0 },
-        metrics: {
-          completion_tokens: 0,
-          time_first_token_millsec: 0,
-          time_completion_millsec: Date.now() - startTime
-        }
+    // Emit BLOCK_COMPLETE so the stream processor's onComplete runs and the
+    // assistant message transitions out of "processing". Without this, the
+    // trailing PlaceholderBlock in Blocks/index.tsx stays visible next to the
+    // finished image because `isMessageProcessing(message)` remains true.
+    const imageResponse = {
+      usage: { prompt_tokens: 0, completion_tokens: 0, total_tokens: 0 },
+      metrics: {
+        completion_tokens: 0,
+        time_first_token_millsec: 0,
+        time_completion_millsec: Date.now() - startTime
      }
-    })
+    }
+    onChunkReceived({ type: ChunkType.BLOCK_COMPLETE, response: imageResponse })
+    onChunkReceived({ type: ChunkType.LLM_RESPONSE_COMPLETE, response: imageResponse })
  } catch (error) {
    onChunkReceived({ type: ChunkType.ERROR, error: error as Error })
    throw error
--- a/src/renderer/src/services/messageStreaming/callbacks/imageCallbacks.ts
+++ b/src/renderer/src/services/messageStreaming/callbacks/imageCallbacks.ts
@@ -71,14 +71,21 @@ export const createImageCallbacks = (deps: ImageCallbacksDependencies) => {
        }
      }

+      if (!imageBlockId && blockManager.hasInitialPlaceholder) {
+        imageBlockId = blockManager.initialPlaceholderBlockId
+      }
+
      if (imageBlockId) {
        if (!imageData) {
          const changes: Partial<ImageMessageBlock> = {
            status: MessageBlockStatus.SUCCESS
          }
-          blockManager.smartBlockUpdate(imageBlockId, changes, MessageBlockType.IMAGE)
+          blockManager.smartBlockUpdate(imageBlockId, changes, MessageBlockType.IMAGE, true)
        } else {
-          const changes = await buildImageBlockFields(imageData)
+          const changes = {
+            type: MessageBlockType.IMAGE,
+            ...(await buildImageBlockFields(imageData))
+          }
          blockManager.smartBlockUpdate(imageBlockId, changes, MessageBlockType.IMAGE, true)
        }
        imageBlockId = null