A file is read into words once, and the words are what travels

[thankywal]

What the owner saw

On a fresh account, a price-list PDF sent in the conversation was read fine on arrival — the assistant knew the file by name — and then every attempt to file it under the business came back:

Hanbit Beauty is set up, but the PDF failed to add to its knowledge again due to a conversion error…

The mechanism, from the code

add_file_to_business had the kept text and handed it to addDocument encoded back into bytes, under the file's own name ….pdf, with contentType: "text/plain":

contentType: "text/plain",
body: new TextEncoder().encode(file.text).buffer

readUpload decides what a file is from its name, so it sent a plain-text "PDF" to env.AI.toMarkdown. Cloudflare's supported formats table has no plain text, so the converter answered format: "error" → unwrapConversion threw conversion failed → the loop gave the model Failed: conversion failed → "conversion error".

• Deterministic → "again".
• Account-independent → the fresh account was a coincidence. The main account had only ever filed files through the console's Add data screen, which sends real bytes.
• Invisible to CI → the covering test mocked addDocument wholesale, so it proved the tool's intent and nothing about what the reader did next.

The change

IngestInput is now IngestFile | IngestText. Words already read travel as text, are never converted and never archived (there is no original to archive), and are recorded as text/plain at the text's own byte size under the owner's filename. The tool hands on text: file.text. Bytes from the console's Add data screen and from Telegram go through the reader exactly as before.

Tests — 956 passing (+5)

a-file-is-read-once.test.ts runs the real ingestDocument with only the platform faked, and its toMarkdown throws the moment it is reached:

• words under a .pdf name are filed without a second reading;
• they are recorded as the text they are, not as the file they came from;
• empty text is refused in the same words a file is;
• a real PDF body still goes to the converter unchanged;
• no tool re-encodes kept text into a body (source-asserted across tools.ts).

The old test's mock now throws if handed bytes, and says where the real check lives.

Not proved here

Workers AI cannot be exercised from the build machine (wrangler whoami: not logged in; no token in the environment), so the live proof is the owner's deployment after this lands: send the PDF in the conversation, say yes to filing it, and the prices should follow.

🤖 Generated with Claude Code

https://claude.ai/code/session_01NDkiHs276KQhSWxQnHQdLg