From ea7d526cad0abedd6f9d8cd15dffdec5aacde564 Mon Sep 17 00:00:00 2001 From: teamchong <25894545+teamchong@users.noreply.github.com> Date: Fri, 3 Jul 2026 19:30:33 -0400 Subject: [PATCH] =?UTF-8?q?docs(readme):=20cut=20to=20217=20lines=20?= =?UTF-8?q?=E2=80=94=20dedupe=20caveats,=20move=20benchmark=20prose=20to?= =?UTF-8?q?=20eval/=20links?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Launch feedback: too long. Structural trims only; FAQ answers, demo captions, and all measured numbers kept verbatim. --- README.md | 292 ++++++++++++++++++++---------------------------------- 1 file changed, 110 insertions(+), 182 deletions(-) diff --git a/README.md b/README.md index f3fed20..d92ed5b 100644 --- a/README.md +++ b/README.md @@ -2,56 +2,43 @@ **Cut Claude Code's input tokens by rendering bulky context as images — the same system prompt, tool docs, and history, in a fraction of the tokens.** -An image's token cost is fixed by its pixel dimensions, not by how much -text is inside it. Dense content (code, JSON, tool output) packs ~3.1 chars per -image-token vs ~1 char per text-token on real Claude Code traffic. pxpipe is a -local proxy that exploits that gap: it rewrites the bulky parts of your -request (system prompt, tool docs, older history) into compact PNGs before -the request leaves your machine. - -Savings are **workload-dependent** — pxpipe wins on token-dense content and -leaves sparse/small requests untouched — so these are measured snapshots, not -constants. The primary, durable result is **input-token reduction**: dense -system prompts, tool docs, and history go in as compact images instead of text -(the example above is ≈25k text tokens rendered as ≈2.7k image tokens), every -request measured against its own `count_tokens` counterfactual. **Dollars are -downstream of that** — at current Fable list prices the token cut lands as a -**~59–70% lower end-to-end bill** (~72–74% on compressed requests; full pricing -math in the FAQ). But list prices can change tomorrow and the token count -won't, so tokens — not dollars — are the number to watch. Reproduce both from +An image's token cost is fixed by its pixel dimensions, not by how much text +is inside it. Dense content (code, JSON, tool output) packs ~3.1 chars per +image-token vs ~1 char per text-token on real Claude Code traffic. pxpipe is +a local proxy that exploits the gap: it rewrites the bulky parts of each +request into compact PNGs before it leaves your machine. At current Fable +list prices that lands as a **~59–70% lower end-to-end bill** — but prices +move and workloads differ, so the durable number is the token cut itself, +measured per-request against a free `count_tokens` counterfactual in `~/.pxpipe/events.jsonl`. This is what the model sees instead of text: ![example: a real `transformRequest` output: system prompt + tool docs reflowed into one dense 1573×1248 page, instruction banner on top, ↵ marking original newlines](https://raw.githubusercontent.com/teamchong/pxpipe/main/docs/assets/example-render.png) -*~48k characters of system prompt + tool docs (this repo's own README, -FINDINGS, and source), ≈25k tokens as text, ≈2.7k image tokens as this page. -Produced by the real `transformRequest` pipeline: whitespace-minified, reflowed -into full rows with ↵ marking original newlines, OCR instruction banner -co-rendered on top. The model reads renders like this at 100/100 on a clean -eval (see benchmarks).* +*~48k chars of system prompt + tool docs: ≈25k tokens as text, ≈2.7k image +tokens as this page. Real pipeline output; the model reads renders like this +at 100/100 (see benchmarks).* ## Demo -**Fable 5 demo (the default, 100/100 reader):** +**Fable 5 (the default, 100/100 reader) — plain left, pxpipe right:** https://github.com/user-attachments/assets/1c8ee63a-fcd7-4958-917b-da788d718349 -* Both demos with both panes on **Fable 5** (plain left, pxpipe right). +pxpipe counts an exact token **10/10** across 39 imaged filler files +(matches `grep` line-for-line), gets the multi-step ledger arithmetic right, +and ends the session at **$6.06** with context to spare (73.5k/1M) vs +**$42.21** at 96% full. One caveat visible in the clip: the pxpipe arm +needed a nudge to match the requested one-line output format. -- **Fable reads what Opus can't.** The imaged phrase-count that Opus refuses (see the Opus demo below): the pxpipe arm counts the exact token **10/10** across 39 imaged filler files (matches `grep` ground truth line-for-line) and gets the multi-step ledger arithmetic right (8037 → … → 15,021). -- **Same answers, ~7× cheaper.** Session totals after both demos: plain **$42.21**, context **96% full** (964.5k/1M — one task away from forced compaction) vs pxpipe **$6.06** with context to spare (73.5k/1M). -- **Honest caveat, visible in the clip:** the pxpipe arm answered the count first and needed one follow-up nudge to also print the ledger balance in the requested one-line format; the plain arm followed the format on the first try. Legibility is solved on Fable — single-reply format compliance is the remaining rough edge. - -**Opus 4.8 demo (Opus disabled by default):** +**Opus 4.8 (disabled by default) — same layout:** https://github.com/user-attachments/assets/f4e50137-31b5-426f-a6ed-b83f829b4a2c -*Side-by-side — plain Claude (left) vs pxpipe (right), both on **Opus 4.8** (opt-in; pxpipe is tuned for Fable — see the Fable clip above). Click the image to watch (Google Drive).* - -- **Demo 1 — fix a failing test suite:** both pass; the dashboard shows pxpipe cut the request to a fraction of the tokens (real, server-measured **context/token reduction**). -- **Demo 2 — a big file-context (40 files, ~382k tokens) plus a math question and a "count this phrase" task:** the math answer (a small **text** needle) reads on both. The phrase-count needs reading the **imaged** filler — so pxpipe-on-Opus can't read it and **honestly surfaces that it won't fabricate a number** (the documented lossy limit: exact values stay text). Plain, meanwhile, bogs down counting file-by-file. +Text needles read fine on both arms; the imaged phrase-count doesn't read on +Opus — and pxpipe **says so instead of fabricating a number**. That misread +rate is why Opus is opt-in. ## Try it (30 seconds) @@ -60,57 +47,36 @@ npx pxpipe-proxy # proxy on 127.0.0.1:47821 ANTHROPIC_BASE_URL=http://localhost:47821 claude # point Claude Code at it ``` -Open for a live dashboard: tokens saved, per-session -stats, every text→image conversion side by side, a global kill switch, and -runtime model chips including GPT 5.6 and GPT 5.5. +Dashboard at : tokens saved, every text→image +conversion side by side, kill switch, live model chips. Responses stream +normally — pxpipe compresses the *request* only, never the model's output. +Recent turns stay text; the system prompt, tool docs, and older bulk history +are imaged. -Nothing else changes. Responses stream normally; pxpipe only compresses the -*request* (your context going up), never the model's output. Recent turns stay -text; the system prompt, tool docs, and older bulk history are imaged. +## The honest part -## The honest part, read before relying on it - -**It is lossy.** pxpipe is a *gist* tier, not a lossless store. In a -needle-in-haystack eval, exact 12-char hex strings inside dense imaged content -came back **0/15** on Opus and 13/15 on Fable 5, and the failure mode is -*silent confabulation*: a plausible wrong value, not an error. Anything you -need back byte-exact (IDs, hashes, secrets, exact numbers) must stay text. -Recent turns do; a dedicated verbatim-risk guard is not built yet. - -**Exact-recall escape hatch.** pxpipe only images Fable requests -(`PXPIPE_MODELS=claude-fable-5`), so any subagent on a non-Fable model passes -through as text. Route work that needs byte-exact values to one — globally with -`CLAUDE_CODE_SUBAGENT_MODEL=claude-sonnet-4-6`, or per-agent with `model: sonnet` -in the agent frontmatter. It reads from source (file/JSONL), not the imaged -history. This covers exact-recall you route on purpose; it does **not** catch a -silent misread you did not expect — that is the unbuilt guard above. - -**Does it break real work?** Parity in what we measured: a 10-instance -SWE-bench Lite pilot (the easy subset) resolved **10/10 on both arms**, -pxpipe ON at $27 vs OFF at $54 token-equivalent, and 19 SWE-bench Pro -pairs (harder, long-horizon) resolved **14/19 ON vs 15/19 OFF** at -**-60% per-request**: verdicts agree on 18/19, and the single split -(one ON fail) re-resolved 3/3 when replicated, i.e. run-to-run agentic -variance, not compression. Small n, details and caveats below. - -**Savings are workload-dependent.** It wins on token-dense content -(~1 char/token: code, JSON, hashes) and *loses money* on sparse English prose -(~3.5 chars/token). The built-in gate only images content where the math wins, -calibrated against N=391 production rows. - -**Model scope:** one `PXPIPE_MODELS` CSV controls which model bases get imaged -across both families — default `claude-fable-5,gpt-5.6` (GPT 5.5 is opt-in; -it degrades on imaged context). Set -`PXPIPE_MODELS=off` to disable imaging entirely, or use -`~/.config/pxpipe/config.json` with `{ "models": "off" }` (or a list). For GPT, -pxpipe keeps tool definitions in native JSON (only verbose schema prose moves -into the image) so tool-calling stays reliable; unlike the Claude path, the GPT -path does not add or depend on Anthropic `cache_control` prompt-cache markers. -The dashboard chips can flip any model live without changing client configs. -Opus 4.7/4.8 was the original Claude scope but misread ~7% of renders -(`10200`→`9400`), so it was turned off by default once Fable 5 hit 100/100 with -identical image billing — opt it back in at your own risk via `PXPIPE_MODELS` or -the dashboard chips. Everything else passes through untouched. +- **It is lossy.** Exact 12-char hex strings in dense imaged content: + **13/15** on Fable 5, **0/15** on Opus — and misses are *silent + confabulations*, not errors. Byte-exact values (IDs, hashes, secrets) + must stay text; recent turns do. A dedicated verbatim-risk guard is not + built yet. +- **Escape hatch:** subagents on non-allowlisted models pass through as + text — route byte-exact work there + (`CLAUDE_CODE_SUBAGENT_MODEL=claude-sonnet-4-6`, or `model: sonnet` in + agent frontmatter). +- **Real work:** SWE-bench Lite pilot **10/10 both arms** at −65% request + size; SWE-bench Pro **14/19 ON vs 15/19 OFF** at −60%, verdicts agree + 18/19, and the single split re-resolved 3/3 on replication — run-to-run + variance, not compression. Small n; receipts in `eval/`. +- **Workload-dependent.** Wins on token-dense content (~1 char/token), + loses money on sparse prose (~3.5 chars/token); a profitability gate + (calibrated on N=391 production rows) images only where the math wins. +- **Model scope:** default `PXPIPE_MODELS=claude-fable-5,gpt-5.6`. Opus + 4.7/4.8 misread ~7% of renders and GPT 5.5 degrades on imaged context, so + both are opt-in via `PXPIPE_MODELS` or the dashboard chips. + `PXPIPE_MODELS=off` disables imaging. Everything else passes through + byte-identical. On the GPT path, tool definitions stay native JSON and no + Anthropic `cache_control` markers are used. ## Benchmarks (reproducible) @@ -126,43 +92,51 @@ Measured with novel random-number problems the model cannot have memorized: | verbatim 12-char hex recall, dense render, Opus | 15 | 15/15 | **0/15** | - | | verbatim 12-char hex recall, dense render, Fable 5 | 15 | - | **13/15** | - | -### SWE-bench Lite pilot (end-to-end task quality) +SWE-bench run totals, receipts, and caveats: +[`eval/swe-bench/`](eval/swe-bench/) · +[`eval/swe-bench-pro/`](eval/swe-bench-pro/) · +[`eval/needle-haystack/`](eval/needle-haystack/) · +[`eval/gist-recall/`](eval/gist-recall/) · analysis in +[`FINDINGS.md`](FINDINGS.md). (GSM8K scored 96% imaged, but it's in training +data — memorized answers survive misreads — so we lead with the novel-number +evals.) -10 SWE-bench Lite instances, Claude Code + Fable 5, paired runs through -pxpipe ON vs OFF, graded with the official `swebench` Docker harness: +## How it works -| | pxpipe ON | OFF | -|---|---:|---:| -| resolved | **10/10** | 10/10 | -| request size vs own uncompressed body | **−65%** | ±0 | +``` +tool_result string ──► wrap at 1928px-wide columns ──► pack ~92,000 chars/page ──► PNG[] +``` -The −65% is per-request (`count_tokens` probe of each body before -compression), so it has no turn-count confound. n=10/arm, Lite skews easy. -Run totals, receipts, caveats: [`eval/swe-bench/`](eval/swe-bench/). +The proxy intercepts `/v1/messages`, rewrites eligible bulk into image +blocks, splices them back cache-friendly (static prefix preserved, prompt +caching keeps working), and forwards. A 1928×1928 image costs ≈4,761 vision +tokens and holds ≈92,000 chars, so text wins only above ~19 chars/token — +Claude Code traffic runs ~1.91 (N=391). A per-request estimator decides; +sparse prose stays text. Events log to `~/.pxpipe/events.jsonl`. -### SWE-bench Pro bench (harder, long-horizon) +## Library use (no proxy) -19 completed pairs across two runs (2 dropped: checkout failed both -arms), same setup, official `SWE-bench_Pro-os` Docker harness: +```ts +import { renderTextToPngs, transformAnthropicMessages } from "pxpipe"; -| | pxpipe ON | OFF | -|---|---:|---:| -| resolved | 14/19 | 15/19 | -| request size vs own uncompressed body | **−60%** | ±0 | +const imgs = await renderTextToPngs(toolResultText); // RenderedImage[] +const { body, applied, info } = await transformAnthropicMessages({ + body: requestBytes, + model: "claude-fable-5", +}); +``` -Verdicts agree on 18/19 (three instances failed both arms, one with -byte-identical patches across arms). The single split (navidrome, ON -fail) was replicated 3x on the ON arm: all three runs produced an -identical patch and **resolved**, so the original loss was run-to-run -agentic variance, not compression. Receipts: -[`eval/swe-bench-pro/`](eval/swe-bench-pro/). +`options.keepSharp(block)` pins blocks as text; `options.emitRecoverable` +returns the originals of imaged blocks. Pure-JS runtime (Node and +edge/Workers); `@napi-rs/canvas` is build-time only. Full API: +`src/core/index.ts`. -We also ran GSM8K: 96% imaged. But GSM8K is in training data, so the model -recalls memorized answers through its own misreads, inflating the score, so we -lead with the clean novel-number eval instead. Reproduce: -[`eval/gsm8k/`](eval/gsm8k/) · [`eval/needle-haystack/`](eval/needle-haystack/) · -[`eval/gist-recall/`](eval/gist-recall/) · -full analysis in [`FINDINGS.md`](FINDINGS.md). +## Development + +```bash +pnpm install && pnpm test +pnpm run build # regenerates dist/ +``` ## FAQ @@ -199,7 +173,11 @@ Three kinds of *input* blocks, each behind a profitability gate: Everything else passes through byte-identical: your messages, recent turns, the model's output (it is the response, the proxy never touches it), sparse -prose, and anything too small to win. Non-Fable models pass through entirely. +prose, and anything too small to win. Models outside the allowlist pass +through entirely — the default scope is Fable 5 and GPT 5.6 only. Opus 4.8 +and GPT 5.5 read imaged content measurably worse (FINDINGS.md 2026-06-16), +so they are deliberately opt-in via the dashboard or `PXPIPE_MODELS`, never +silently imaged. **Has it ever failed for real, outside the benchmarks?** Yes, once in weeks of daily use: the model recalled a person's name from @@ -207,82 +185,32 @@ imaged chat history and got it confidently wrong. No error, just a plausible wrong name. That is the documented failure mode: exact strings in imaged content are not byte-safe. Coding sessions tolerate this because the agent re-reads files before editing; pure chat recall has no such check. +This failure mode is measured, not anecdotal: +[the legibility audit](docs/LEGIBILITY-AUDIT-2026-07-01.md) quantifies +exact-string recall off rendered pages (blind reads top out at 63% on dense +identifiers, with every miss predicted by a glyph-confusability matrix) and +documents the shipped mitigations — page geometry clamped to the API's +resample cap so billed pixels actually reach the vision encoder, and exact +identifiers (SHAs, numbers) riding alongside as text. -## How it works - -``` -tool_result string ──► wrap at 1928px-wide columns ──► pack ~92,000 chars/page ──► PNG[] -``` - -The proxy intercepts `/v1/messages`, rewrites eligible bulk history into image -blocks, splices them back cache-friendly (static prefix preserved, so prompt -caching keeps working), and forwards. Per-request events log to -`~/.pxpipe/events.jsonl`. - -The economics: a 1928×1928 image costs ≈4,761 vision tokens and holds up to -≈92,000 chars (≈48,000 text tokens at the observed density), so plain text is -cheaper *only* when it runs denser than ~19 chars/token. Claude Code transcripts -are far below that (observed 1.91 chars/token, N=391). The runtime estimator (`estimateImageCount`) plus a chars/token gate -decides per-request; sparse prose is left as text. - -## Library use (no proxy) - -Same engine, no proxy. Render text → PNGs, or run the full cache-safe transform: - -```ts -import { renderTextToPngs, transformAnthropicMessages } from "pxpipe"; - -const imgs = await renderTextToPngs(toolResultText); // RenderedImage[] -const { body, applied, info } = await transformAnthropicMessages({ - body: requestBytes, - model: "claude-fable-5", -}); -``` - -`options.keepSharp(block)` pins blocks as text (override the heuristic for IDs, -hashes, paths); `options.emitRecoverable` returns the originals of imaged blocks -so a stateful caller can recover them — the two halves of the fidelity contract -for the lossy limitation below. Runtime is pure-JS (Node and edge/Workers); -`@napi-rs/canvas` is build-time only. Full API, types, and constants: -`src/core/index.ts`. - -## Development - -```bash -pnpm install && pnpm test # 376 tests -pnpm run build # regenerates dist/ -``` +**Why does the README read like an AI wrote it?** +Because one did. Most of this repo's commits — the code and the docs — were +authored by Opus/Fable agent sessions running behind pxpipe itself, reading +their own collapsed history as image pages while they worked. ## Limitations -* **Lossy**: see "the honest part" above. Verbatim recall from images is unreliable. -* Render latency: encoding PNGs adds time to large requests before they leave - (partly offset by the model ingesting fewer tokens). Responses stream normally. -* ASCII/Latin-1 well tested; CJK works but conservatively. -* Runtime is pure-JS — runs on Node and edge/Workers. `@napi-rs/canvas` is a - build-time-only dev dep (regenerating the glyph atlas), not a runtime dep. -* Fable 5 only. +- Lossy (above); verbatim recall from images is unreliable. +- PNG encoding adds latency to large requests before they leave. +- ASCII/Latin-1 well tested; CJK works but conservatively. ## Roadmap -Everything above is measured. Everything here is not. These are hypotheses, not -claims; they ship as numbers with an n or they get cut. - -* **Sharper glyphs.** The 13/15 verbatim gap is partly font legibility, not just - the model. A per-char confusion matrix across render styles is paused mid-run - (`eval/glyph-matrix/`); if a zero-cost style lowers read error, the gate - compresses harder at the same fidelity. -* **Effective context.** Dense text carries at ~3x fewer tokens as images. If - that holds in the live window and not just the bill, 1M tokens holds ~2x the - real content. Open question: can a task needing ~2M raw context run inside - Fable's 1M once the bulk is imaged? -* **Less active text, sharper model.** Long contexts degrade reasoning as they - fill. Imaging old bulk shrinks what the model actively reads while keeping it - reachable. Hypothesis: same information, smaller active context, better - long-task accuracy. - -One bet: longer effective context and a sharper model on long tasks, from the -same Fable 5. Numbers or retraction, no hype between. +Hypotheses, not claims — they ship as numbers with an n or they get cut: +sharper glyph rendering (`eval/glyph-matrix/`, paused mid-run), whether +imaged bulk stretches effective context (~2x the real content in the same +1M window), and whether a smaller active context improves long-task +accuracy. ## License