honey-px

# Honey PX An image's token cost is fixed by its pixel area, not its character count. Dense text packs ~3 chars per image-token vs ~1 as text, so a 90k-char file read as PNG pages costs a fraction of the text read. `pxpipe export` (teamchong/pxpipe) does the rendering; the agent `Read`s the pages. **Lossy by design** — misreads of exact strings are *silent confabulations*, not errors. The export writes a `factsheet.txt` carrying the verbatim precision tokens (paths, SHAs, ids, numbers) as text alongside the images; always read it too. ## When to use - Bulk is **read-only reference**: skimming a big vendored file, a large diff for review context, docs, generated code — content you reason over, not byte-copy. - **Dense** (code, JSON, logs) and **big** — worth it above ~6k chars; the export prints `% saved`, skip if it's low. Sparse prose loses money. - The reader is **Fable-class**. Opus misreads ~7% of renders; other models degrade — read as text there. ## Don't use - Files you will **Edit** — `old_string` must be byte-exact; imaged reads aren't. Read the real file (with `offset`/`limit`) before any edit. - **Secrets, hashes, ids, numeric values you'll act on** — treat anything exact recalled from an image as unverified; re-`Grep` the source first. - Anything the user asked to see verbatim, or that gets written back. - Small or sparse content — the fixed pixel cost loses. ## How ```sh npx pxpipe-proxy export --json --out <scratchpad> <file-or-dir> # or --...