fix(imagegen): align tool inputs with official contract

src/skills/bundled/imagegen.test.ts

@@ -58,9 +58,9 @@ describe('bundled imagegen skill', () => {
     expect(text).toContain('latest selected output as the next turn')
     expect(text).toContain('one call per image')
     expect(text).toContain('omit <code>input_images</code> entirely')
-    expect(text).toContain('Never pass <code>/dev/null</code>')
-    expect(text).toContain('otherwise omit the field')
-    expect(text).toContain('Never pass <code>default</code>')
+    expect(text).toContain('Never invent a path')
+    expect(text).toContain('Provider and image model selection come from')
+    expect(text).toContain('do not add either to the tool arguments')
     expect(text).not.toContain('CC_HAHA_IMAGE_API_KEY')
   })
 })

src/skills/bundled/imagegen/SKILL.md

@@ -11,7 +11,7 @@ Use the built-in `ImageGen` tool. Provider authentication, model routing, output
 ## Decide the request shape
 
 - Treat a brand-new visual as generation. Treat a request that preserves or changes an existing visual as an edit.
-- For generation, omit `input_images` entirely. Never pass `/dev/null`, an empty string, or any other placeholder path to represent no input image.
+- For generation, omit `input_images` entirely. Never invent a path to represent the absence of an input image.
 - One distinct prompt equals one tool call.
 - Use `count` only for multiple variations of the same prompt. For different concepts, make separate calls.
 - For an edit, populate `input_images` with ordered paths surfaced by `[Image source: ...]` in a user attachment or returned by an earlier `ImageGen` call. Never invent, search for, or substitute another filesystem path.
@@ -19,7 +19,7 @@ Use the built-in `ImageGen` tool. Provider authentication, model routing, output
 - For multi-turn editing, use the latest selected output as the next turn's `edit_target`. Repeat all identity, layout, text, and unchanged-region constraints on every turn so edits do not drift.
 - To edit several images independently, make one call per image. Put multiple images in one call only when the user wants them combined or used together as references. A single call accepts at most three source images.
 - Prefer a useful default composition when the user leaves details open. Do not invent branding, logos, or people they did not request.
-- Respect an explicitly requested concrete provider model ID by passing `model`; otherwise omit the field so the configured model is used. Never pass `default` or another placeholder model name.
+- Provider and image model selection come from the current desktop session; do not add either to the tool arguments.
 - If the provider or tool returns an error, do not retry `ImageGen` automatically. Explain the failure and let the user decide whether to retry or change providers.
 
 ## Build the prompt

src/tools/ImageGenTool/ImageGenTool.test.ts

@@ -7,6 +7,7 @@ import {
   IMAGE_GENERATION_PROVIDER_ID_ENV_KEY,
   IMAGE_GENERATION_PROVIDER_KIND_ENV_KEY,
 } from '../../services/imageGeneration/config.js'
+import { zodToJsonSchema } from '../../utils/zodToJsonSchema.js'
 import { ImageGenTool } from './ImageGenTool.js'
 
 const ENV_KEYS = [
@@ -45,10 +46,18 @@ describe('ImageGenTool', () => {
   })
 
   test('defaults to one slot and preserves the structured result for the desktop', () => {
+    expect(ImageGenTool.strict).toBe(false)
+    const jsonSchema = zodToJsonSchema(ImageGenTool.inputSchema)
+    expect(jsonSchema.required).toEqual(['prompt', 'count'])
+    expect(jsonSchema.properties).not.toHaveProperty('model')
     expect(ImageGenTool.inputSchema.parse({ prompt: 'A paper-cut fox' })).toEqual({
       prompt: 'A paper-cut fox',
       count: 1,
     })
+    expect(() => ImageGenTool.inputSchema.parse({
+      prompt: 'A paper-cut fox',
+      model: 'default',
+    })).toThrow()
 
     const output = {
       type: 'image_generation_result' as const,
@@ -96,8 +105,7 @@ describe('ImageGenTool', () => {
       'do not retry ImageGen automatically',
     )
     expect(prompt).toContain('omit input_images entirely')
-    expect(prompt).toContain('never pass /dev/null')
-    expect(prompt).toContain('Omit model unless')
-    expect(prompt).toContain('never pass "default"')
+    expect(prompt).toContain('Provider and image model selection')
+    expect(prompt).toContain('are not tool arguments')
   })
 })

src/tools/ImageGenTool/ImageGenTool.ts

@@ -46,7 +46,7 @@ const inputSchema = lazySchema(() =>
         path: z
           .string()
           .min(1)
-          .describe('Absolute path from an [Image source: ...] attachment or a prior ImageGen result; omit input_images entirely for new generation and never use /dev/null as a placeholder'),
+          .describe('Absolute path from an [Image source: ...] attachment or a prior ImageGen result; omit input_images entirely for new generation'),
         role: z
           .enum(['edit_target', 'reference', 'style_reference', 'composite_source'])
           .describe('How this ordered image should influence the edit'),
@@ -55,11 +55,6 @@ const inputSchema = lazySchema(() =>
       .max(3)
       .optional()
       .describe('Ordered source images for editing, compositing, or visual reference'),
-    model: z
-      .string()
-      .min(1)
-      .optional()
-      .describe('Concrete image model ID to override the configured model only when the user explicitly asks; otherwise omit this field and never use "default" as a placeholder'),
     aspect_ratio: z
       .enum(ASPECT_RATIOS)
       .optional()
@@ -104,13 +99,13 @@ export const ImageGenTool = buildTool({
   name: IMAGE_GEN_TOOL_NAME,
   searchHint: 'generate images or artwork from a visual prompt',
   maxResultSizeChars: 100_000,
-  strict: true,
+  strict: false,
   shouldDefer: true,
   async description() {
     return 'Generate one or more images with the image provider configured for this desktop session.'
   },
   async prompt() {
-    return `Use this tool when the user asks to generate or edit an image. For a brand-new image, omit input_images entirely; never pass /dev/null or another placeholder path. Omit model unless the user explicitly requests a concrete image model ID; never pass "default" as a placeholder. For edits, pass ordered input_images using only paths surfaced by [Image source: ...] in the current conversation or returned by a prior ImageGen call; repeat preservation constraints in every edit prompt. One call represents one distinct prompt; use count only for variations of that same prompt. The tool saves finished raster images locally and returns their absolute paths. If a provider call fails, do not retry ImageGen automatically; explain the error and wait for the user to decide.`
+    return `Use this tool when the user asks to generate or edit an image. For a brand-new image, omit input_images entirely. For edits, pass ordered input_images using only paths surfaced by [Image source: ...] in the current conversation or returned by a prior ImageGen call; repeat preservation constraints in every edit prompt. Provider and image model selection come from the current desktop session and are not tool arguments. One call represents one distinct prompt; use count only for variations of that same prompt. The tool saves finished raster images locally and returns their absolute paths. If a provider call fails, do not retry ImageGen automatically; explain the error and wait for the user to decide.`
   },
   get inputSchema(): InputSchema {
     return inputSchema()

src/tools/ImageGenTool/backend.test.ts

@@ -246,50 +246,7 @@ describe('ImageGen backend', () => {
     expect(result.images).toHaveLength(2)
   })
 
-  test('treats a model-supplied /dev/null image placeholder as generation', async () => {
-    outputDir = await mkdtemp(join(tmpdir(), 'imagegen-output-'))
-    const calls: string[] = []
-    const fetchImpl = async (input: string | URL | Request) => {
-      calls.push(String(input))
-      return Response.json({
-        data: [{ b64_json: PNG_BYTES.toString('base64') }],
-      })
-    }
-
-    const result = await generateImages({
-      prompt: 'A paper-cut fox poster',
-      count: 1,
-      input_images: [{ path: '/dev/null', role: 'reference' }],
-    }, customConfig, { fetchImpl, outputDir })
-
-    expect(calls).toEqual([
-      'https://relay.example.test/v1/images/generations',
-    ])
-    expect(result.operation).toBe('generate')
-    expect(result.inputImageCount).toBe(0)
-  })
-
-  test('uses the configured image model for a model-supplied default placeholder', async () => {
-    outputDir = await mkdtemp(join(tmpdir(), 'imagegen-output-'))
-    let requestBody: Record<string, unknown> | undefined
-    const fetchImpl = async (_input: string | URL | Request, init?: RequestInit) => {
-      requestBody = JSON.parse(String(init?.body))
-      return Response.json({
-        data: [{ b64_json: PNG_BYTES.toString('base64') }],
-      })
-    }
-
-    const result = await generateImages({
-      prompt: 'A paper-cut fox poster',
-      count: 1,
-      model: 'default',
-    }, customConfig, { fetchImpl, outputDir })
-
-    expect(requestBody).toMatchObject({ model: 'relay-image-model' })
-    expect(result.model).toBe('relay-image-model')
-  })
-
-  test('uses the configured ChatGPT OAuth image model for a default placeholder', async () => {
+  test('uses the configured ChatGPT OAuth image model', async () => {
     outputDir = await mkdtemp(join(tmpdir(), 'imagegen-openai-oauth-'))
     const tokenPath = join(outputDir, 'openai-oauth.json')
     const previousTokenPath = process.env[OPENAI_CODEX_OAUTH_FILE_ENV_KEY]
@@ -318,7 +275,6 @@ describe('ImageGen backend', () => {
       const result = await generateImages({
         prompt: 'A paper-cut fox poster',
         count: 1,
-        model: 'default',
       }, {
         kind: 'openai_oauth',
         providerId: 'openai-official',

src/tools/ImageGenTool/backend.ts

@@ -30,7 +30,6 @@ export type ImageGenerationInput = {
     path: string
     role: 'edit_target' | 'reference' | 'style_reference' | 'composite_source'
   }>
-  model?: string
   aspect_ratio?: string
   resolution?: '1k' | '2k'
   size?: 'auto' | '1024x1024' | '1024x1536' | '1536x1024'
@@ -94,11 +93,7 @@ export async function generateImages(
   options: GenerateOptions = {},
 ): Promise<ImageGenerationOutput> {
   const startedAt = Date.now()
-  const requestedModel = input.model?.trim()
-  // Some tool-calling models use "default" to mean no model override.
-  const model = requestedModel && requestedModel.toLowerCase() !== 'default'
-    ? requestedModel
-    : config.model
+  const model = config.model
   const fetchImpl = options.fetchImpl ?? fetch
   const { signal, cleanup } = createCombinedAbortSignal(options.signal, {
     timeoutMs: IMAGE_REQUEST_TIMEOUT_MS,
@@ -611,10 +606,7 @@ async function prepareInputImages(
   input: ImageGenerationInput,
   overrideRootDirs?: string[],
 ): Promise<PreparedInputImage[]> {
-  // Some tool-calling models use the POSIX null device to mean "no image".
-  const requested = (input.input_images ?? []).filter(
-    ({ path: inputPath }) => inputPath.trim() !== '/dev/null',
-  )
+  const requested = input.input_images ?? []
   if (requested.length === 0) return []
   if (requested.length > 3) {
     throw new Error('Image editing supports at most 3 input images per call.')