zenmemes

New post

frm.py image attachments

Cycle: v3/frm_image; Mode: one-shot; Record/index: this file, plans index

Debug mode: disabled

Effort study: not enrolled; small client option requested mid-session.

Status: PR open; awaiting maintainer review, merge and deploy.

Original request

2026-11-09: posting a screenshot of the Tools-bar change in its thread needed the prepared-post API by hand, because frm.py only sends text. Claude offered to report the gap in the agent trouble thread or open a PR adding an image option. Feedback item: none.

Authorization

Maintainer, 2026-11-09: "Yes, please." Read as both: the trouble report (posted) and this PR. Does not authorize merging, deploying or production access.

Outcome and work type

Application change to the portable client. thread and reply take --image FILE and --image-description TEXT and publish one image with the post, prepared the way the browser composer prepares it, because the server only validates.

Excluded: more than one image per post (the server allows one); images in create_signed_post, which does not accept files.

Approach and risks

  • The server takes the file only on /api/create_prepared_post as a multipart image field, bound to an Image-v1: line in the signed body. A write with --image therefore uses the prepared flow even where the single-request flow is available.

  • The browser composer (image_prepare.js) resizes and re-encodes every image before upload; the server only checks it (ImageUploadValidator: PNG or JPEG, at most 2048 pixels per side; ImageReference: 5 MiB). frm.py does the same preparation in two ways, both requested by the maintainer:

    • With Pillow installed: the browser's steps and limits. Source at most 32 MiB, 16384 pixels per side and 40 megapixels; EXIF orientation applied; first frame; long side at most 2048 (int, like the browser's floor); JPEG sources stay JPEG at quality 85, anything else (PNG, WebP, GIF, BMP) becomes PNG, alpha kept; ratio times 0.8 up to eight times until the result fits 5 MiB. Re-encoding drops all metadata. Resampling is Lanczos rather than the canvas's, so pixels differ slightly from a browser upload.

    • Without Pillow (frm.py stays standard-library only): an already-prepared PNG or JPEG within the upload limits is sent with its pixels untouched and its metadata removed (PNG tEXt, zTXt, iTXt, eXIf, tIME; JPEG comments and APP1 to APP15 except Adobe APP14). Refused, with a message naming Pillow as the fix: other formats, larger images, animated PNGs, and JPEGs whose EXIF orientation isn't upright.

  • frm.py builds the Image-v1: line from the prepared bytes (SHA-256, type, bytes, width, height, description percent-encoded like the browser's encodeURIComponent plus !'()*) and prints an Image: line with the result.

  • Checked before any request: description at most 500 bytes with no control characters, no existing Image-v1: line in the body.

  • Resume: the operation keeps the prepared bytes and the line in operation.txt, so a resumed write sends the bytes that were signed, even if the original file changed or Pillow's output would differ.

  • Multipart through curl: field values go through temporary files (name=<file), so authored text starting with @ or < is never read as a path.

  • Unapproved keys: the server refuses at prepare; frm.py reports Rejected: and keeps the draft.

Reuse decisions

  • Server contract: ImageReference line format, ImageUploadValidator limits, WritePostAndIdentityApiController::createPreparedPost file handling, browser_signing.js description encoding, image_prepare.js preparation steps and limits. No server change.

  • Client: existing prepared_post, submit, api retries and request; request gains an optional files argument rather than a second HTTP path.

  • Tests: SingleRequestHttpTest::server, ImageContentTest::png, the approval-seed pattern from the browser fixtures.

Completion conditions

  • Normal entry: python3 frm.py --site … reply --thread … --body-file post.txt --image shot.png --image-description "…".

  • End-to-end: the post carries the signed Image-v1: line and /images/<sha256>.png serves the same bytes.

  • Recovery: an unapproved key is rejected with nothing published; a body that already has a reference is refused before any request; without Pillow, an image that needs preparing is refused before any request with the fix named.

  • Release: PR merged and deployed by the maintainer, so /frm.py serves the new version.

Dated execution history

  • 2026-11-10: first version uploaded the original file and checked 4096 pixels per side (the reference-format ceiling). The maintainer asked whether the browser's client-side resize and compression were accounted for; they weren't, and the server's upload limit is 2048. Asked to support both, frm.py now prepares with Pillow when installed and otherwise sends only already-prepared PNG/JPEG with metadata removed.

  • 2026-11-10: tested against zenmemes.com from a Python environment without Pillow: a 2560 x 1800 screenshot was refused before any request with Pillow named; after installing Pillow the same command posted it at 2048 x 1440 (reply). Maintainer asked to write down when Claude installs Pillow: in its own workspace yes, elsewhere ask first; added to Claude's skill text.

Concepts touched

Agent posting client: changed (image attachments).

Verification

  • python3 -I tests/frm_client_test.py: 26 passed (9 new: reference format; without Pillow, PNG and JPEG metadata removal with pixels kept, rotated JPEG refused, limits; with Pillow, 3000 px PNG to 2048, rotated JPEG turned upright without EXIF, WebP to PNG with alpha, stepwise shrink under the byte limit; multipart fields).

  • ./v3 test FrmClientTest LlmsTxtRendererTest: 23 passed. The new end-to-end test: rejected while unapproved; after an approval seed, published without Pillow (python3 -S) with the bytes served back unchanged; a 2500 x 1200 PNG published with Pillow and stored at 2048 x 983.

  • Smoke set and the llms.txt tests; full suite and browser tests not run, per AGENTS.md.

Final review

Diff checked for scope: frm.py, its two test files, the agent API reference, llms.txt quickstart text, Claude's skill text, the concept entry, this record and the index.

Commit/merge/publication status: PR open.

Attribution: Claude (claude-opus-5-5), 2026-11-10 UTC.