@xamukavila/pxpipe 0.8.0 → 0.9.0

package/README.md

@@ -16,7 +16,7 @@ measured per-request against a free `count_tokens` counterfactual in
 
 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)
+![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/XamuAvila/pxpipe/main/docs/assets/example-render.png)
 
 *~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
@@ -45,7 +45,7 @@ rate is why Opus is opt-in.
 ## Try it (30 seconds)
 
 ```bash
-npx pxpipe-proxy                                  # proxy on 127.0.0.1:47821
+npx @xamukavila/pxpipe                            # proxy on 127.0.0.1:47821
 ANTHROPIC_BASE_URL=http://127.0.0.1:47821 claude  # point Claude Code at it
 ```
 
@@ -74,8 +74,8 @@ are imaged.
   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.
+  4.6/4.7/4.8 misread ~7% of renders and GPT 5.5 degrades on imaged context, so
+  all 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.
@@ -127,7 +127,7 @@ bench below shows it is cheaper and cache-stable, **not** that recall holds.
 | Library | `transformAnthropicMessages({ …, caveman: true })` | boolean, default `false` |
 
 ```bash
-PXPIPE_CAVEMAN=1 npx pxpipe-proxy      # prose compression ON (experiment)
+PXPIPE_CAVEMAN=1 npx @xamukavila/pxpipe  # prose compression ON (experiment)
 ```
 
 ⚠️ Flipping the flag changes the rendered image bytes, which **busts warm image
@@ -172,6 +172,29 @@ PXPIPE_CAVEMAN=1 bash bench/run.sh --model 'claude-opus-4-6[1m]' --label caveman
 node bench/score.mjs bench/runs/<off-dir> bench/runs/<caveman-dir>
 ```
 
+### Does it hurt recall?
+
+A recall A/B (`eval/caveman-slab/`, pre-registered) put seeded facts into an
+imaged system-prompt slab and compared **plain image vs caveman image** at
+production density, fable-5, N=50 answerable + 10 unanswerable per arm:
+
+| arm | answerable correct | confabulated (never-stated facts) |
+|---|---:|---:|
+| plain image | 31/50 (62%) | 3/10 |
+| caveman image | 32/50 (64%) | 0/10 |
+
+Tied within noise across every fact type (decision, numeric, path, name,
+negation); caveman confabulated *less*. **caveman does not measurably hurt
+recall vs plain imaging** — and the caveman-rendered slabs cost **−6.9% image
+tokens** (10,751 → 10,004 across the 10 slabs), so the saving is free of recall
+cost. Two caveats keep this from being a default-on green light: both arms hit a
+~63% **floor** (the single dense page is hard to read — plain fails the same
+way, so the floor is the render, not caveman), and the eval deliberately omits
+the production "fact sheet" that keeps exact identifiers as text
+(`transform.ts:1627`), so the numeric/path misses it counts would not occur in
+production. Full design, results, and follow-ups:
+[`eval/caveman-slab/DESIGN.md`](eval/caveman-slab/DESIGN.md).
+
 ## How it works
 
 ```
@@ -188,7 +211,7 @@ sparse prose stays text. Events log to `~/.pxpipe/events.jsonl`.
 ## Library use (no proxy)
 
 ```ts
-import { renderTextToImages, transformAnthropicMessages } from "pxpipe-proxy";
+import { renderTextToImages, transformAnthropicMessages } from "@xamukavila/pxpipe";
 
 const { pages } = await renderTextToImages(toolResultText);     // pages[i].png: Uint8Array
 const { body, applied, info } = await transformAnthropicMessages({
@@ -245,10 +268,10 @@ 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. 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.
+through entirely — the default scope is Fable 5 and GPT 5.6 only. Opus
+4.6/4.7/4.8 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

package/dist/core/applicability.js

@@ -7,11 +7,11 @@ function baseModelId(model) {
 /** Dashboard runtime override; null = fall back to PXPIPE_MODELS env / built-in default. In-memory only. */
 let runtimeModelBases = null;
 /** Built-in default scope when PXPIPE_MODELS is unset: Fable 5 (Claude) plus
- *  GPT 5.6. GPT 5.5 and Opus 4.8 are intentionally off — same pipeline but
+ *  GPT 5.6. GPT 5.5 and Opus 4.6/4.7/4.8 are intentionally off — same pipeline but
  *  measurably worse at reading imaged content (FINDINGS.md 2026-06-16: Opus 4.8
  *  ~2pp arithmetic, 6/15 dense-hex recall vs Fable's 100/100; GPT 5.5 likewise
  *  degrades on imaged history/context) — so silently imaging them is the wrong
- *  default. Both stay opt-in via the dashboard chips or PXPIPE_MODELS. */
+ *  default. All stay opt-in via the dashboard chips or PXPIPE_MODELS. */
 const DEFAULT_MODEL_BASES = ['claude-fable-5', 'gpt-5.6'];
 function falsey(v) {
     return /^(0|false|no|off|none)$/i.test(v.trim());

package/dist/core/export.js

@@ -8,7 +8,7 @@
  */
 import { DENSE_CONTENT_CHARS_PER_IMAGE, DENSE_CONTENT_COLS, MAX_HEIGHT_PX, PAD_X, CELL_W, } from './render.js';
 // Dogfood the public SDK: render via the same `./transform` entry external
-// consumers import (pxpipe-proxy/transform → renderTextToImages), not the
+// consumers import (@xamukavila/pxpipe/transform → renderTextToImages), not the
 // internal leaf renderer.
 import { renderTextToImages } from './library.js';
 import { estimateImageCount, ANTHROPIC_PIXELS_PER_TOKEN, IMAGE_COST_SAFETY_MARGIN, REPORT_CHARS_PER_TOKEN } from './transform.js';
@@ -324,7 +324,7 @@ export async function runExportCore(sourceText, opts) {
     // Render to PNG pages via the public SDK primitive. shrink=true sizes the canvas
     // to the widest line so short-line code isn't padded to full width; multiCol='auto'
     // packs as many columns side-by-side as fit. Same render surface shipped to external
-    // SDK consumers (pxpipe-proxy/transform).
+    // SDK consumers (@xamukavila/pxpipe/transform).
     const { pages: images } = await renderTextToImages(sourceText, {
         cols: opts.cols,
         shrink: true,

package/dist/dashboard/fragments.js

@@ -60,6 +60,7 @@ const MODEL_CATALOG = [
     { id: 'claude-fable-5', label: 'Fable 5' },
     { id: 'claude-opus-4-8', label: 'Opus 4.8' },
     { id: 'claude-opus-4-7', label: 'Opus 4.7' },
+    { id: 'claude-opus-4-6', label: 'Opus 4.6' },
     { id: 'claude-sonnet-5', label: 'Sonnet 5' },
     { id: 'claude-sonnet-4-6', label: 'Sonnet 4.6' },
 ];

package/dist/node.js

@@ -6749,6 +6749,7 @@ var MODEL_CATALOG = [
   { id: "claude-fable-5", label: "Fable 5" },
   { id: "claude-opus-4-8", label: "Opus 4.8" },
   { id: "claude-opus-4-7", label: "Opus 4.7" },
+  { id: "claude-opus-4-6", label: "Opus 4.6" },
   { id: "claude-sonnet-5", label: "Sonnet 5" },
   { id: "claude-sonnet-4-6", label: "Sonnet 4.6" }
 ];
@@ -8349,7 +8350,7 @@ Use with OpenAI-compatible GPT clients:
 `);
 }
 function printVersion() {
-  const injected = true ? "0.8.0" : void 0;
+  const injected = true ? "0.9.0" : void 0;
   console.log(injected ?? process.env.npm_package_version ?? "unknown");
 }
 function toWebRequest(req) {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@xamukavila/pxpipe",
-  "version": "0.8.0",
+  "version": "0.9.0",
   "description": "Token-saving proxy for Claude Code: renders bulky context (system prompt, tool docs, old history) as dense PNGs to cut input tokens. Runs on Node and Cloudflare Workers.",
   "type": "module",
   "bin": {