To the Pickaxe Engineering Team,
I am encountering a reproducible payload handoff issue with the Pickaxe vision integration. When a user starts a brand new chat session and sends an image AND text in their very first message, Pickaxe’s backend is dropping the image from the payload before it reaches the Anthropic API.
The AI receives the text prompt, but registers zero image attachments.
My Environment & Setup:
-
Model: claude-sonnet-4-6
-
Capabilities: “Allow users to upload files” is set to “This thread only”
-
Token Allocations: * Input Length: 45,000
-
Uploaded Documents: 45,000
-
User Memories: 10,000
-
Memory Buffer: 11,410
-
Output: 1,800
-
-
Total Context Load: ~119,589 tokens. This is only 12% of Claude’s 999,000 capacity. This definitively rules out token truncation or context window overflow as the cause.
Steps to Reproduce:
-
Open a brand new chat session via a live share link (Incognito window).
-
Click the paperclip and attach an image (e.g., a standard PNG screenshot).
-
Type a text prompt in the exact same message box (e.g., “Give me a transcript of this conversation”).
-
Hit Send.
Actual Behavior: Claude responds stating that it cannot see an image or that no image was provided (e.g., “No image or screenshot attached to your message”). The token metrics verify that the image file was not processed or sent to the model.
Expected Behavior: The AI should process the image via its native vision capabilities and execute the text prompt based on the visual context.
Diagnostic Proof (Why this is a Pickaxe payload assembly issue): I can completely bypass this bug using the exact same prompt and token settings if I separate the actions into two API calls.
-
If the user uploads the image by itself on turn 1, it doesn’t acknowledge anything was sent (see attached image)
-
It asks for image then if I upload image it reads it.
-
If the user uploads an image + text on turn 3 of an already active chat, it works flawlessly.
Hypothesis: There appears to be a race condition or payload assembly error occurring only on the creation of a new chat session. The Pickaxe system prioritizes sending the lightweight text prompt to establish the session with Anthropic, and fails to append the heavier image file to the context window before API handoff.
Could your engineering team look into how initial multipart (text + image) payloads are being routed to the Anthropic API on session start? Let me know if you need my specific Pickaxe workspace ID or share links to test.