fix: avoid replaying file attachment content

.agents/skills/deepchat-sdd/SKILL.md

@@ -42,6 +42,8 @@ Resolve every `[NEEDS CLARIFICATION]` marker before implementation. If a request
 5. Implement the change after the SDD artifacts are complete.
 6. Update `tasks.md` as work lands.
 7. Run `pnpm run format`, `pnpm run i18n`, and `pnpm run lint` before handoff.
+8. After implementation is accepted and validation passes, delete `plan.md` and `tasks.md` for
+   that goal; keep `spec.md` as the durable contract.
 
 ## Documentation Hygiene
 

docs/issues/file-attachment-token-bloat/spec.md

@@ -0,0 +1,73 @@
+# File Attachment Token Bloat Spec
+
+> Status: Implemented
+> Date: 2026-07-02
+> Related: https://github.com/ThinkInAIXYZ/deepchat/issues/1864
+
+## Problem
+
+DeepChat currently treats chat input files as both UI attachments and model-ready context. For
+non-image files, `file.prepareFile` returns LLM-friendly `content`, the user message can store that
+content, and the agent runtime context builder can replay it into every later request while the
+message remains inside history. For images, uploaded/pasted files are represented with image data
+URLs so vision models can see them, but historical image turns can also keep resending large inline
+image payloads. Audio is different from document files: when the selected model supports native
+audio input, keeping audio as structured base64 input is intentional model behavior, not a text
+attachment fallback.
+
+Issue #1864 reports the visible failure mode: attaching an image, PDF, or document makes later API
+calls resend the old file data, increasing request size, cache misses, token use, and latency.
+
+## User Need
+
+Users need file attachments to remain visually correct in the conversation while keeping later model
+requests lean. The model should still receive enough path and metadata context to read local files
+through tools instead of embedding their contents into chat history.
+
+## Goals
+
+- Keep image attachments as message media for the submitted turn so vision behavior remains correct.
+- Avoid resending historical file bytes or extracted document text on unrelated follow-up turns.
+- Route non-image file understanding through existing filesystem tooling when the agent needs the
+  content.
+- Preserve native audio input for models that support audio attachments.
+- Keep message rendering stable: existing attachment chips, thumbnails, names, paths, and metadata
+  stay visible in the transcript.
+
+## Acceptance Criteria
+
+- A newly submitted image attachment is sent as an image part when the active model supports vision.
+- Historical image attachments are not resent as inline image data by default on later unrelated
+  turns; history includes enough metadata/path context for the model to request a read when tools are
+  available.
+- Non-image attachments are replayed as path plus metadata by default, not full extracted content.
+- Existing `agent-filesystem.read` remains the primary path for reading text, PDF, Office, and other
+  supported non-image files during an agent turn.
+- PDF, Word, Excel, PowerPoint, text, and code files continue to rely on the existing
+  `FilePresenter.prepareFileCompletely(..., 'llm-friendly')` pipeline behind
+  `agent-filesystem.read`.
+- Audio attachments still become structured `input_audio` parts when the model supports audio input.
+- Attachment chips do not need a separate path insertion action; path/metadata is included
+  automatically in the model-visible attachment context.
+- Existing sessions with stored attachment content do not require a database migration; request
+  construction strips or ignores stale inline content where needed.
+
+## Non-Goals
+
+- Redesigning the full attachment data model in one step.
+- Replacing `agent-filesystem.read` with a new file reading tool.
+- Automatically invoking file reads before every provider call.
+- Deleting attachment content from existing user databases during this change.
+- Removing structured audio input for audio-capable models.
+- Solving provider-specific cache-key behavior beyond stopping repeated inline payloads.
+
+## Constraints
+
+- Reuse existing main-process tools and permission flow for file reads.
+- Keep privileged filesystem access behind preload/typed route clients.
+- Keep UI copy in i18n keys when implementation starts.
+- Keep the first implementation small enough to review as a focused issue fix.
+
+## Open Questions
+
+None.

src/main/presenter/agentRuntimePresenter/contextBuilder.ts

@@ -37,6 +37,12 @@ type TokenizedTurn = {
   tokens: number
 }
 
+type UserMessageContentBuildOptions = {
+  includeFileContent?: boolean
+  includeImageData?: boolean
+  includeAudioData?: boolean
+}
+
 export type HistoryTurn = {
   records: ChatMessageRecord[]
   messages: ChatMessage[]
@@ -208,28 +214,44 @@ export function isContextHistoryRecord(record: ChatMessageRecord): boolean {
   return record.role === 'assistant' && record.status === 'error'
 }
 
-function buildNonImageFileContext(files: MessageFile[], excludeAudio: boolean = false): string {
+function buildNonImageFileContext(
+  files: MessageFile[],
+  options: {
+    excludeAudio?: boolean
+    includeFileContent?: boolean
+  } = {}
+): string {
   const nonImageFiles = files.filter(
-    (file) => !isImageFile(file) && (!excludeAudio || !isAudioFile(file))
+    (file) => !isImageFile(file) && (!options.excludeAudio || !isAudioFile(file))
   )
   if (nonImageFiles.length === 0) {
     return ''
   }
 
   const blocks = nonImageFiles.map((file, index) => {
+    const isAudio = isAudioFile(file)
     const fileName = typeof file.name === 'string' ? file.name : `file-${index + 1}`
     const filePath = typeof file.path === 'string' ? file.path : ''
     const mimeType = resolveFileMimeType(file)
     const fileContent = typeof file.content === 'string' ? file.content : ''
+    const shouldIncludeContent =
+      options.includeFileContent === true ||
+      (isAudio && !fileContent.trim().toLowerCase().startsWith('data:audio/'))
+    const byteSize = resolveFileByteSize(file)
     const metadataLines = [
       `name: ${fileName}`,
       filePath ? `path: ${filePath}` : '',
-      mimeType ? `mime: ${mimeType}` : ''
+      mimeType ? `mime: ${mimeType}` : '',
+      byteSize ? `size: ${byteSize}` : ''
     ]
       .filter(Boolean)
       .join('\n')
     if (!fileContent.trim()) {
-      return `[Attached File ${index + 1}]\n${metadataLines}\ncontent: [empty]`
+      const placeholder = filePath ? '[omitted; use read if needed]' : '[empty]'
+      return `[Attached File ${index + 1}]\n${metadataLines}\ncontent: ${placeholder}`
+    }
+    if (!shouldIncludeContent) {
+      return `[Attached File ${index + 1}]\n${metadataLines}\ncontent: [omitted; use read if needed]`
     }
     return `[Attached File ${index + 1}]\n${metadataLines}\ncontent:\n${fileContent}`
   })
@@ -383,10 +405,13 @@ function buildImageMetadataContext(files: MessageFile[]): string {
 export function buildUserMessageContent(
   input: SendMessageInput,
   supportsVision: boolean,
-  supportsAudioInput: boolean = false
+  supportsAudioInput: boolean = false,
+  options: UserMessageContentBuildOptions = {}
 ): ChatMessage['content'] {
   const text = input.text ?? ''
   const files = Array.isArray(input.files) ? input.files : []
+  const includeImageData = options.includeImageData !== false
+  const includeAudioData = options.includeAudioData !== false
 
   const imageFiles = files.filter((file) => isImageFile(file))
   const audioFiles = files.filter((file) => isAudioFile(file))
@@ -398,37 +423,44 @@ export function buildUserMessageContent(
       filename?: string
       estimated_tokens?: number
     }
-  }> = supportsAudioInput
-    ? audioFiles.flatMap((file) => {
-        const payload = resolveAudioAttachmentPayload(file)
-        if (!payload) {
-          return []
-        }
+  }> =
+    supportsAudioInput && includeAudioData
+      ? audioFiles.flatMap((file) => {
+          const payload = resolveAudioAttachmentPayload(file)
+          if (!payload) {
+            return []
+          }
 
-        return [
-          {
-            type: 'input_audio' as const,
-            input_audio: {
-              data: payload.data,
-              media_type: payload.mediaType,
-              ...(typeof file.name === 'string' && file.name.trim() ? { filename: file.name } : {}),
-              estimated_tokens: estimateAudioInputTokens(file, payload.byteLength)
+          return [
+            {
+              type: 'input_audio' as const,
+              input_audio: {
+                data: payload.data,
+                media_type: payload.mediaType,
+                ...(typeof file.name === 'string' && file.name.trim()
+                  ? { filename: file.name }
+                  : {}),
+                estimated_tokens: estimateAudioInputTokens(file, payload.byteLength)
+              }
             }
-          }
-        ]
-      })
-    : []
+          ]
+        })
+      : []
 
   const excludeAudioFromFallback = supportsAudioInput && audioParts.length > 0
-  const nonImageContext = buildNonImageFileContext(files, excludeAudioFromFallback)
+  const nonImageContext = buildNonImageFileContext(files, {
+    excludeAudio: excludeAudioFromFallback,
+    includeFileContent: options.includeFileContent === true
+  })
   const audioMetadata = excludeAudioFromFallback ? buildAudioMetadataContext(audioFiles) : ''
-  const baseText = [text, nonImageContext, audioMetadata]
+  const shouldBuildImageParts = supportsVision && includeImageData && imageFiles.length > 0
+  const imageMetadata = shouldBuildImageParts ? '' : buildImageMetadataContext(imageFiles)
+  const baseText = [text, nonImageContext, audioMetadata, imageMetadata]
     .filter((value) => value.trim())
     .join('\n\n')
 
   if ((!supportsVision || imageFiles.length === 0) && audioParts.length === 0) {
-    const imageMetadata = buildImageMetadataContext(imageFiles)
-    return [baseText, imageMetadata].filter((value) => value.trim()).join('\n\n')
+    return baseText
   }
 
   const parts: Array<
@@ -450,7 +482,7 @@ export function buildUserMessageContent(
     image_url: { url: string; detail?: 'auto' | 'low' | 'high' }
   }> = []
 
-  if (supportsVision) {
+  if (supportsVision && includeImageData) {
     for (const file of imageFiles) {
       const primaryData = typeof file.content === 'string' ? file.content : ''
       const fallbackData = typeof file.thumbnail === 'string' ? file.thumbnail : ''
@@ -466,12 +498,18 @@ export function buildUserMessageContent(
   }
 
   const hasStructuredParts = imageParts.length > 0 || audioParts.length > 0
+  const structuredText = [
+    baseText,
+    shouldBuildImageParts && imageParts.length === 0 ? buildImageMetadataContext(imageFiles) : ''
+  ]
+    .filter((value) => value.trim())
+    .join('\n\n')
   if (!hasStructuredParts) {
-    const imageMetadata = buildImageMetadataContext(imageFiles)
-    return [baseText, imageMetadata].filter((value) => value.trim()).join('\n\n')
+    return structuredText
   }
 
-  const textPart = baseText || buildStructuredAttachmentText(imageParts.length, audioParts.length)
+  const textPart =
+    structuredText || buildStructuredAttachmentText(imageParts.length, audioParts.length)
   parts.push({ type: 'text', text: textPart })
   parts.push(...imageParts, ...audioParts)
 
@@ -486,7 +524,11 @@ export function createUserChatMessage(
   const normalizedInput = normalizeUserInput(input)
   return {
     role: 'user',
-    content: buildUserMessageContent(normalizedInput, supportsVision, supportsAudioInput)
+    content: buildUserMessageContent(normalizedInput, supportsVision, supportsAudioInput, {
+      includeImageData: true,
+      includeAudioData: true,
+      includeFileContent: false
+    })
   }
 }
 
@@ -639,7 +681,11 @@ export function recordToChatMessages(
     const parsed = parseUserRecordContent(record.content)
     const message: ChatMessage = {
       role: 'user',
-      content: buildUserMessageContent(parsed, supportsVision, supportsAudioInput)
+      content: buildUserMessageContent(parsed, supportsVision, supportsAudioInput, {
+        includeImageData: false,
+        includeAudioData: true,
+        includeFileContent: false
+      })
     }
     return hasPromptMessageContent(message) ? [message] : []
   }

test/main/presenter/agentRuntimePresenter/compactionService.test.ts

@@ -209,7 +209,7 @@ describe('CompactionService', () => {
     vi.clearAllMocks()
   })
 
-  it('preserves file bodies and image metadata in user summary blocks', async () => {
+  it('preserves attachment metadata without replaying file bodies in user summary blocks', async () => {
     const { service, messageStore } = createService()
 
     messageStore.getMessages.mockReturnValue([
@@ -247,7 +247,10 @@ describe('CompactionService', () => {
     })
 
     expect(intent).not.toBeNull()
-    expect(intent?.summaryBlocks[0]).toContain('Detailed file body')
+    expect(intent?.summaryBlocks[0]).toContain('[Attached File 1]')
+    expect(intent?.summaryBlocks[0]).toContain('path: /tmp/spec.md')
+    expect(intent?.summaryBlocks[0]).toContain('content: [omitted; use read if needed]')
+    expect(intent?.summaryBlocks[0]).not.toContain('Detailed file body')
     expect(intent?.summaryBlocks[0]).toContain('[Attached Image 1]')
     expect(intent?.summaryBlocks[0]).not.toContain('data:image/png')
   })