pymc-labs File handling
What you read becomes permanent. Every read is inlined into this conversation
and replayed to the model on every later turn — you cannot remove it, edit it, or
apologise for it. Two kinds of read are unrecoverable, and both are avoidable in
one extra command.
Images: longest edge under 2000px, always
The failure. The API caps how large an image may be. That cap is not fixed: once a conversation carries more than 20 images it tightens to 2000px on the longest edge, and it is applied to the whole conversation — so an image that was legal when you read it becomes illegal later, when some unrelated 21st image pushes the count over. The request is then rejected outright, every following turn replays the same illegal image, and the conversation is dead for good. Not slow, not degraded — dead, with no way to remove the offending block.
The rule: never read an image whose longest edge is 2000px or more. Under that size nothing can retroactively invalidate it, whatever else the conversation accumulates.
Check before you read, whenever you did not create the file at a known size:
python3 -c "from PIL import Image; print(Image.open('shot.png').size)"
Downscale in place if it is over — 1600px is a safe target that stays sharp:
python3 -c "from PIL import Image; p='shot.png'; im=Image.open(p); im.thumbnail((1600,1600)); im.save(p)"
Screenshots are the usual culprit
Two habits produce oversized files without looking like they do:
| Habit | What you get | Fix |
|---|---|---|
device_scale_factor=2 |
doubles both edges — a 1800×1400 viewport becomes 3600×2800 | use device_scale_factor=1 |
full_page=True on a long page |
width fine, height enormous — e.g. 980×4262 | capture viewport sections and read them one at a time |
Note the second one: a full-page capture is usually within the limit on width and far over it on height. Checking only the width will not save you.
# safe defaults for a page screenshot
page = browser.new_page(viewport={"width": 1440, "height": 900}) # scale factor 1
page.screenshot(path="shot.png") # not full_page
Keep the count down too
The cap only tightens past 20 images, so fewer images is its own protection — and each one costs tokens on every subsequent turn.
- Re-read one file as you revise it, rather than saving
v1.png,v2.png,v3.pngand reading each. - Crop to the region you are actually checking instead of reading the whole frame.
- When you have inspected an image and acted on it, do not read it again to confirm — you still have your own description of it.
Large text files: look before you read
A whole-file read of something large is not fatal, but it is irreversible and it
crowds out everything else for the rest of the conversation. Size it first:
wc -lc path/to/file
For anything big, retrieve only what you need — grep for the symbol, head/
tail for structure, or read with a line range. Reach for a whole-file read
when the file is genuinely small or you genuinely need all of it.
Files you produce for the user
Write deliverables to /mnt/session/outputs/ — files there are captured and can
be handed back to the user. Do not read a file back purely to prove you wrote it;
ls -la confirms it exists and costs nothing.