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 Reads 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

npx pxpipe-proxy export --json --out <scratchpad> <file-or-dir>   # or --git / --diff <ref> / --stdin

Output dir pxpipe-export-<hash>/ contains page-*.png, factsheet.txt, manifest.json (token report). Then:

1. Read every page-*.png and factsheet.txt. If instead you pass the

renders to a model over the raw API (subagent prompt, panel), include the export's prompt.txt banner — naked dense renders can trip stop_reason: refusal on Fable-class models; the banner prevents it.

1. Treat the manifest's percentSaved as the go/no-go — if it printed low,

read the text instead.

1. Before acting on any exact string seen only in an image, verify it against

the source with Grep.

Cheaper still is not reading at all — Grep/offset/limit (Lever 3b) and eson crush (CCR) come first. PX is for when you genuinely need the whole bulk in view.