Add Rust port of proxy (unproven, not wired up)

Functional reimplementation in Rust with the same render_chunks/transform
public surface as src/proxy.py. Compiles clean on darwin-arm64, smoke-
tested but NOT YET battle-tested against real Claude Code sessions.

Layout:
  src/rust/
    Cargo.toml         hyper + tokio + reqwest + ab_glyph + png + clap
    Cargo.lock         committed (application crate, deterministic builds)
    src/main.rs        clap CLI matching bin/cli.js flags
    src/lib.rs         module roots
    src/font.rs        ab_glyph rasterization of bundled JetBrains Mono
    src/render.rs      newspaper-column text -> PNG (matches Python algo)
    src/transform.rs   request body transformation (system prompt -> image)
    src/proxy.rs       hyper HTTP server, upstream forwarding
    src/stats.rs       persisted savings counters mirroring proxy.py
    assets/JetBrainsMono-Regular.ttf  OFL-licensed, legally embeddable
    examples/render_sample.rs   smoke test: render fibonacci snippet
    examples/diff_render.rs     diff harness companion (paired with
                                scripts/diff_render.py)

Diff harness:
  scripts/diff_render.py   feeds same text through Python and Rust
                           render_chunks, decodes both PNGs, compares
                           pixel-by-pixel. First run found two divergences:

  1. Rust font.rs:43 uses ab_glyph h_advance which returns ~3.4px ceiled
     to 4 for Menlo-5pt 'M'; PIL's getlength('M') returns 3.0 exactly.
     Every Rust column is 33% wider than Python's.
  2. Python proxy.py:125 pads every single-column image to MAX_EDGE=1568
     height regardless of content (already noted in baseline commit).

Neither is fixed yet — surfacing the divergences before any swap. Next:
visual inspect both PNGs to confirm Rust output is at least OCR-readable,
then decide whether to fix bug-for-bug compat or correct both sides.

bin/cli.js still spawns Python; this commit only adds the Rust tree, it
does NOT change runtime behavior. The live proxy at :47821 is unaffected.

scripts/diff_render.py

@@ -0,0 +1,136 @@
+#!/usr/bin/env python3
+"""
+Diff harness: compare Python render_chunks vs Rust render_chunks pixel-by-pixel.
+
+Strategy:
+  1. For each sample text, call Python render_chunks() -> list[PNG bytes]
+  2. Call Rust diff_render binary -> writes PNGs to disk
+  3. Decode both sides' PNGs and compare pixel arrays
+  4. Report: pass / fail with first diff coordinates
+
+Run:
+  python3 scripts/diff_render.py
+"""
+
+import hashlib
+import json
+import subprocess
+import sys
+import tempfile
+from pathlib import Path
+
+# Make src/proxy.py importable
+ROOT = Path(__file__).resolve().parent.parent
+sys.path.insert(0, str(ROOT / "src"))
+
+import proxy as py_proxy  # noqa: E402
+from PIL import Image  # noqa: E402
+import io  # noqa: E402
+
+RUST_BIN = ROOT / "src" / "rust" / "target" / "release" / "examples" / "diff_render"
+
+SAMPLES = {
+    "short_ascii": "hello world\n",
+    "multiline_code": (
+        "def fibonacci(n):\n"
+        "    if n <= 1: return n\n"
+        "    return fibonacci(n-1) + fibonacci(n-2)\n"
+        "print(fibonacci(10))\n"
+    ),
+    "long_with_blanks": (
+        "# Header\n\n"
+        + "\n".join(f"line {i}: " + "x" * (i % 60) for i in range(200))
+        + "\n\n# Footer\n"
+    ),
+    "many_short_lines": "\n".join(f"L{i}" for i in range(500)) + "\n",
+    "wide_lines": "\n".join("." * 80 for _ in range(100)) + "\n",
+    "tools_like_json": json.dumps(
+        {"tools": [{"name": f"tool_{i}", "input_schema": {"type": "object",
+         "properties": {f"p{j}": {"type": "string"} for j in range(5)}}}
+         for i in range(20)]}, indent=2),
+}
+
+
+def png_to_pixels(png_bytes: bytes):
+    """Decode PNG to (width, height, raw_grayscale_bytes)."""
+    img = Image.open(io.BytesIO(png_bytes))
+    img = img.convert("L")  # force 8-bit grayscale
+    return img.width, img.height, img.tobytes()
+
+
+def run_rust(samples: dict, out_dir: Path) -> dict:
+    """Invoke rust diff_render binary, returns name -> list[png_bytes]."""
+    spec = {name: text for name, text in samples.items()}
+    spec_path = out_dir / "spec.json"
+    spec_path.write_text(json.dumps(spec))
+    res = subprocess.run(
+        [str(RUST_BIN), str(spec_path), str(out_dir)],
+        capture_output=True, text=True,
+    )
+    if res.returncode != 0:
+        print("RUST FAILED:", res.stderr, file=sys.stderr)
+        sys.exit(2)
+    manifest = json.loads((out_dir / "manifest.json").read_text())
+    out = {}
+    for name, files in manifest.items():
+        out[name] = [(out_dir / f).read_bytes() for f in files]
+    return out
+
+
+def diff_pixels(py_bytes: bytes, rs_bytes: bytes):
+    """Return None if pixel-identical, else (reason, details)."""
+    pw, ph, pp = png_to_pixels(py_bytes)
+    rw, rh, rp = png_to_pixels(rs_bytes)
+    if (pw, ph) != (rw, rh):
+        return ("dim_mismatch", f"py={pw}x{ph} rs={rw}x{rh}")
+    if pp == rp:
+        return None
+    # Find first diff
+    for i, (a, b) in enumerate(zip(pp, rp)):
+        if a != b:
+            x = i % pw
+            y = i // pw
+            return ("pixel_mismatch", f"first diff at ({x},{y}): py={a} rs={b}")
+    return ("byte_len_mismatch", f"py_len={len(pp)} rs_len={len(rp)}")
+
+
+def main():
+    if not RUST_BIN.exists():
+        print(f"build rust first: cd src/rust && cargo build --release --example diff_render")
+        print(f"missing: {RUST_BIN}")
+        sys.exit(1)
+
+    with tempfile.TemporaryDirectory() as td:
+        out_dir = Path(td)
+        rust_out = run_rust(SAMPLES, out_dir)
+
+        total = 0
+        fail = 0
+        for name, text in SAMPLES.items():
+            py_pngs = py_proxy.render_chunks(text)
+            rs_pngs = rust_out.get(name, [])
+            total += 1
+            print(f"\n[{name}]  py_chunks={len(py_pngs)} rs_chunks={len(rs_pngs)}")
+            if len(py_pngs) != len(rs_pngs):
+                fail += 1
+                print(f"  FAIL: chunk count differs")
+                continue
+            any_fail = False
+            for i, (a, b) in enumerate(zip(py_pngs, rs_pngs)):
+                ah = hashlib.sha256(a).hexdigest()[:12]
+                bh = hashlib.sha256(b).hexdigest()[:12]
+                diff = diff_pixels(a, b)
+                if diff is None:
+                    print(f"  chunk {i}: OK   py={ah} rs={bh} bytes={len(a)}/{len(b)}")
+                else:
+                    any_fail = True
+                    print(f"  chunk {i}: FAIL {diff[0]}: {diff[1]}  py={ah} rs={bh}")
+            if any_fail:
+                fail += 1
+
+        print(f"\n=== {total - fail}/{total} samples pixel-identical ===")
+        sys.exit(0 if fail == 0 else 1)
+
+
+if __name__ == "__main__":
+    main()

src/rust/Cargo.toml

@@ -0,0 +1,71 @@
+[package]
+name = "claude-image-proxy"
+version = "0.1.0"
+edition = "2021"
+description = "Token-saving proxy for Claude Code that renders system prompt + tools as images"
+license = "MIT"
+repository = "https://github.com/your/claude-image-proxy"
+
+[[bin]]
+name = "claude-image-proxy"
+path = "src/main.rs"
+
+[dependencies]
+# Async runtime
+tokio = { version = "1", features = ["full"] }
+
+# HTTP server (incoming from Claude Code)
+axum = { version = "0.7", features = ["macros"] }
+
+# HTTP client (upstream to api.anthropic.com) - rustls so no OpenSSL link
+reqwest = { version = "0.12", default-features = false, features = ["rustls-tls", "stream", "json", "http2"] }
+
+# JSON parse + manipulate
+serde = { version = "1", features = ["derive"] }
+serde_json = "1"
+
+# Font rasterization (TTF -> glyph bitmaps)
+ab_glyph = "0.2"
+
+# PNG encoding
+png = "0.17"
+
+# Base64 for image data URI
+base64 = "0.22"
+
+# Hashing for render-cache keys (deterministic across turns)
+sha2 = "0.10"
+
+# In-memory cache (thread-safe, no allocations on hit)
+dashmap = "6"
+
+# Bytes manipulation
+bytes = "1"
+
+# CLI args parsing
+clap = { version = "4", features = ["derive"] }
+
+# Logging
+tracing = "0.1"
+tracing-subscriber = { version = "0.3", features = ["env-filter"] }
+
+# Misc
+once_cell = "1"
+
+# Error handling
+anyhow = "1"
+
+[profile.release]
+opt-level = 3
+lto = "thin"
+codegen-units = 1
+strip = true
+panic = "abort"
+
+[profile.release-small]
+inherits = "release"
+opt-level = "z"
+
+[lib]
+name = "claude_image_proxy"
+path = "src/lib.rs"

src/rust/examples/diff_render.rs

@@ -0,0 +1,48 @@
+//! Diff harness companion: read a JSON spec {name: text, ...}, render each via
+//! the Rust pipeline, write PNGs to <out_dir>/<name>_<idx>.png, then emit
+//! <out_dir>/manifest.json mapping name -> [filenames].
+//!
+//! Run: cargo run --release --example diff_render -- <spec.json> <out_dir>
+
+use std::collections::BTreeMap;
+use std::path::PathBuf;
+
+use claude_image_proxy::font::AtlasFont;
+use claude_image_proxy::render::render_chunks;
+
+fn main() -> anyhow::Result<()> {
+    let mut args = std::env::args().skip(1);
+    let spec_path: PathBuf = args
+        .next()
+        .ok_or_else(|| anyhow::anyhow!("usage: diff_render <spec.json> <out_dir>"))?
+        .into();
+    let out_dir: PathBuf = args
+        .next()
+        .ok_or_else(|| anyhow::anyhow!("usage: diff_render <spec.json> <out_dir>"))?
+        .into();
+
+    let spec_bytes = std::fs::read(&spec_path)?;
+    let spec: BTreeMap<String, String> = serde_json::from_slice(&spec_bytes)?;
+
+    let font = AtlasFont::load(5.0)?;
+
+    let mut manifest: BTreeMap<String, Vec<String>> = BTreeMap::new();
+    for (name, text) in &spec {
+        let pngs = render_chunks(&font, text);
+        let mut filenames = Vec::with_capacity(pngs.len());
+        for (i, png) in pngs.iter().enumerate() {
+            let fname = format!("{name}_{i}.png");
+            std::fs::write(out_dir.join(&fname), &png.bytes)?;
+            filenames.push(fname);
+        }
+        manifest.insert(name.clone(), filenames);
+    }
+
+    std::fs::write(
+        out_dir.join("manifest.json"),
+        serde_json::to_vec_pretty(&manifest)?,
+    )?;
+
+    eprintln!("rendered {} samples", spec.len());
+    Ok(())
+}

src/rust/examples/render_sample.rs

@@ -0,0 +1,23 @@
+//! Render a sample text → PNG so we can visually inspect + OCR-test.
+//! Run: cargo run --example render_sample -- <output.png> [size]
+
+use claude_image_proxy::font::AtlasFont;
+use claude_image_proxy::render::render_chunks;
+
+fn main() -> anyhow::Result<()> {
+    let mut args = std::env::args().skip(1);
+    let out = args.next().unwrap_or_else(|| "/tmp/rust_sample.png".to_string());
+    let size: f32 = args.next().and_then(|s| s.parse().ok()).unwrap_or(5.0);
+
+    let font = AtlasFont::load(size)?;
+    println!("cell: {}x{}", font.cell_w, font.cell_h);
+
+    let sample = "def fibonacci(n):\n    if n <= 1: return n\n    return fibonacci(n-1) + fibonacci(n-2)\nprint(fibonacci(10))";
+
+    let pngs = render_chunks(&font, sample);
+    println!("images: {}, dims: {:?}", pngs.len(), pngs.iter().map(|p| (p.width, p.height)).collect::<Vec<_>>());
+
+    std::fs::write(&out, &pngs[0].bytes)?;
+    println!("wrote {} ({} bytes)", out, pngs[0].bytes.len());
+    Ok(())
+}

src/rust/src/font.rs

@@ -0,0 +1,111 @@
+//! Embedded TTF font + glyph rasterization at a fixed size.
+//!
+//! Pre-bakes the printable ASCII range (32-126) into a fixed-cell atlas at
+//! load time. Renders text by blitting from the atlas into a u8 grayscale
+//! buffer (0 = ink, 255 = paper). This matches the Python proxy's pipeline
+//! that we verified produces 99.5% OCR accuracy on Opus 4.7.
+
+use ab_glyph::{Font, FontRef, PxScale, ScaleFont};
+
+/// JetBrains Mono Regular, OFL-licensed. Bundled at compile time.
+const FONT_DATA: &[u8] = include_bytes!("../assets/JetBrainsMono-Regular.ttf");
+
+pub struct AtlasFont {
+    pub cell_w: u32,
+    pub cell_h: u32,
+    pub ascent: i32,
+    /// Glyph bitmaps, one per code point in [FIRST..=LAST]. Each bitmap is
+    /// cell_w * cell_h bytes (0 = ink, 255 = paper).
+    pub glyphs: Vec<Vec<u8>>,
+    pub first: u8,
+    pub last: u8,
+}
+
+const FIRST: u8 = 32;
+const LAST: u8 = 126;
+
+impl AtlasFont {
+    pub fn load(font_size_pt: f32) -> anyhow::Result<Self> {
+        let font = FontRef::try_from_slice(FONT_DATA)?;
+        // ab_glyph's PxScale is in *pixels* of the EM box. PIL/typography use
+        // points (1pt = 1/72 in, 1 px = 1/96 in @ 96 DPI). Convert pt → px.
+        // This gives us the same physical glyph size PIL produced (where
+        // Menlo 5pt OCR'd at 99.7%).
+        let px_height = font_size_pt * 96.0 / 72.0;
+        let scale = PxScale::from(px_height);
+        let scaled = font.as_scaled(scale);
+
+        // Determine cell dimensions: use widest glyph width + line height.
+        let ascent = scaled.ascent().ceil() as i32;
+        let descent = scaled.descent().floor() as i32;
+        let line_h = (ascent - descent).max(1) as u32;
+
+        // Use a consistent monospace advance from 'M'.
+        let advance_m = scaled.h_advance(font.glyph_id('M')).ceil().max(1.0) as u32;
+        // Cell width: tight to advance; allow 1 px of slop for descenders/spillover.
+        let cell_w = advance_m.max(1);
+        let cell_h = line_h;
+
+        let mut glyphs = Vec::with_capacity((LAST - FIRST + 1) as usize);
+        for code in FIRST..=LAST {
+            let ch = code as char;
+            let glyph_id = font.glyph_id(ch);
+            let glyph = glyph_id.with_scale(scale);
+            let mut bitmap = vec![255u8; (cell_w * cell_h) as usize];
+            if let Some(outlined) = font.outline_glyph(glyph) {
+                let bounds = outlined.px_bounds();
+                let x_offset = bounds.min.x as i32;
+                let y_offset = bounds.min.y as i32 + ascent;
+                outlined.draw(|gx, gy, coverage| {
+                    let px = gx as i32 + x_offset;
+                    let py = gy as i32 + y_offset;
+                    if px < 0 || py < 0 { return; }
+                    let px = px as u32;
+                    let py = py as u32;
+                    if px >= cell_w || py >= cell_h { return; }
+                    // Keep antialiased grayscale (0=black ink, 255=white paper).
+                    // PIL+Menlo at 5pt with AA→99.7% OCR; 1-bit threshold killed
+                    // legibility for ab_glyph. Vision encoder actually uses the
+                    // grayscale gradients at tiny font sizes.
+                    let ink = (coverage * 255.0).clamp(0.0, 255.0) as u8;
+                    let cur = bitmap[(py * cell_w + px) as usize];
+                    let new_v = 255_u8.saturating_sub(ink);
+                    if new_v < cur {
+                        bitmap[(py * cell_w + px) as usize] = new_v;
+                    }
+                });
+            }
+            glyphs.push(bitmap);
+        }
+
+        Ok(AtlasFont {
+            cell_w,
+            cell_h,
+            ascent,
+            glyphs,
+            first: FIRST,
+            last: LAST,
+        })
+    }
+
+    /// Blit one glyph into a destination buffer (255=paper, 0=ink) at (dx, dy).
+    /// Keeps the darker of (existing, glyph) — preserves antialiasing.
+    pub fn blit(&self, ch: u8, dst: &mut [u8], dst_w: u32, dst_h: u32, dx: u32, dy: u32) {
+        if ch < self.first || ch > self.last { return; }
+        let bitmap = &self.glyphs[(ch - self.first) as usize];
+        for y in 0..self.cell_h {
+            let py = dy + y;
+            if py >= dst_h { break; }
+            for x in 0..self.cell_w {
+                let px = dx + x;
+                if px >= dst_w { break; }
+                let v = bitmap[(y * self.cell_w + x) as usize];
+                let idx = (py * dst_w + px) as usize;
+                // Preserve grayscale: keep the darker value (paper=255 default).
+                if v < dst[idx] {
+                    dst[idx] = v;
+                }
+            }
+        }
+    }
+}

src/rust/src/lib.rs

@@ -0,0 +1,2 @@
+pub mod font;
+pub mod render;

src/rust/src/main.rs

@@ -0,0 +1,111 @@
+//! claude-image-proxy — token-saving proxy for Claude Code.
+//!
+//! Architecture mirrors the proven Python prototype:
+//!   1. Accept HTTP from Claude Code on a local port.
+//!   2. For POST /v1/messages: parse JSON body, replace the bulky system
+//!      prompt + tool descriptions + tool schemas + <system-reminder> blocks
+//!      with a single rendered PNG image. Stub the original fields.
+//!   3. Forward the modified request to api.anthropic.com untouched-otherwise.
+//!   4. Return the response unchanged. Log effective-token math.
+//!
+//! Single static binary, no Python dependency.
+
+use clap::Parser;
+use std::sync::Arc;
+
+mod proxy;
+mod stats;
+mod transform;
+mod render;
+mod font;
+
+#[derive(Parser, Debug, Clone)]
+#[command(name = "claude-image-proxy", version, about, long_about = None)]
+struct Args {
+    /// Port to listen on
+    #[arg(short, long, default_value_t = 47821)]
+    port: u16,
+
+    /// Disable ALL compression (pure passthrough — for benchmarking)
+    #[arg(long)]
+    no_compress: bool,
+
+    /// Don't compress tool descriptions
+    #[arg(long)]
+    no_tools: bool,
+
+    /// Don't compress tool input_schemas (saves the most)
+    #[arg(long)]
+    no_schemas: bool,
+
+    /// Don't compress <system-reminder> blocks
+    #[arg(long)]
+    no_reminders: bool,
+
+    /// Don't compress large tool_result content
+    #[arg(long)]
+    no_tool_results: bool,
+
+    /// Font size for rendering (5pt is verified OCR floor)
+    #[arg(long, default_value_t = 5.0)]
+    font_size: f32,
+
+    /// Minimum chars before triggering compression
+    #[arg(long, default_value_t = 2000)]
+    min_chars: usize,
+
+    /// Upstream URL (default: https://api.anthropic.com)
+    #[arg(long, default_value = "https://api.anthropic.com")]
+    upstream: String,
+}
+
+#[derive(Clone)]
+pub struct AppState {
+    pub args: Arc<Args>,
+    pub client: reqwest::Client,
+    pub render_cache: Arc<dashmap::DashMap<[u8; 32], Vec<render::Png>>>,
+    pub stats: Arc<stats::Stats>,
+    pub font: Arc<font::AtlasFont>,
+}
+
+#[tokio::main(flavor = "multi_thread")]
+async fn main() -> anyhow::Result<()> {
+    let args = Args::parse();
+
+    tracing_subscriber::fmt()
+        .with_env_filter(
+            tracing_subscriber::EnvFilter::try_from_default_env()
+                .unwrap_or_else(|_| tracing_subscriber::EnvFilter::new("claude_image_proxy=info,info")),
+        )
+        .with_target(false)
+        .init();
+
+    let port = args.port;
+    let compress = !args.no_compress;
+    let font_size = args.font_size;
+
+    let state = AppState {
+        args: Arc::new(args),
+        client: reqwest::Client::builder()
+            .timeout(std::time::Duration::from_secs(300))
+            .build()?,
+        render_cache: Arc::new(dashmap::DashMap::new()),
+        stats: Arc::new(stats::Stats::default()),
+        font: Arc::new(font::AtlasFont::load(font_size)?),
+    };
+
+    println!("claude-image-proxy v{} starting", env!("CARGO_PKG_VERSION"));
+    println!("  port:          {}", port);
+    println!("  compression:   {}", if compress { "ON" } else { "OFF (passthrough)" });
+    if compress {
+        println!("    font:        JetBrains Mono {}pt", font_size);
+    }
+    println!();
+    println!("  Point Claude Code at: ANTHROPIC_BASE_URL=http://127.0.0.1:{}", port);
+    println!("  Live stats:           curl http://127.0.0.1:{}/proxy-stats", port);
+    println!();
+
+    proxy::serve(port, state).await
+}
+
+// Submodules declared inline at top of file

src/rust/src/proxy.rs

@@ -0,0 +1,280 @@
+//! HTTP server: accept from Claude Code, transform, forward to api.anthropic.com.
+
+use axum::{
+    body::Bytes,
+    extract::State,
+    http::{HeaderMap, HeaderName, HeaderValue, Method, StatusCode, Uri},
+    response::{IntoResponse, Response},
+    routing::{any, get},
+    Router,
+};
+use std::collections::VecDeque;
+use std::sync::Mutex;
+use std::time::Instant;
+
+use crate::transform::{transform, TransformConfig, TransformInfo};
+use crate::AppState;
+
+// Connection-specific (hop-by-hop) headers + things we always strip.
+const HOP_HEADERS: &[&str] = &[
+    "connection", "keep-alive", "proxy-connection", "transfer-encoding",
+    "upgrade", "te", "host", "content-length", "expect", "accept-encoding",
+];
+
+#[derive(Clone, serde::Serialize)]
+pub struct RequestEntry {
+    pub method: String,
+    pub path: String,
+    pub size_in: usize,
+    pub size_out: usize,
+    pub status: u16,
+    pub upstream_ms: u64,
+    pub info: TransformInfo,
+    pub tokens: Option<TokenLog>,
+    pub session_saved_so_far: f64,
+    pub session_saved_usd: f64,
+}
+
+#[derive(Clone, Debug, serde::Serialize)]
+pub struct TokenLog {
+    pub input: u64,
+    pub output: u64,
+    pub cache_read: u64,
+    pub cache_create: u64,
+    pub effective_cost: f64,
+}
+
+// Tail of recent requests for the dashboard.
+pub struct Recent(pub Mutex<VecDeque<RequestEntry>>);
+impl Default for Recent {
+    fn default() -> Self { Recent(Mutex::new(VecDeque::with_capacity(64))) }
+}
+impl Recent {
+    pub fn push(&self, e: RequestEntry) {
+        let mut q = self.0.lock().unwrap();
+        if q.len() >= 64 { q.pop_front(); }
+        q.push_back(e);
+    }
+    pub fn snapshot(&self) -> Vec<RequestEntry> {
+        self.0.lock().unwrap().iter().cloned().collect()
+    }
+}
+
+pub async fn serve(port: u16, state: AppState) -> anyhow::Result<()> {
+    let recent: std::sync::Arc<Recent> = std::sync::Arc::new(Recent::default());
+
+    let app = Router::new()
+        .route("/proxy-stats", get(stats_handler))
+        .fallback(any(proxy_handler))
+        .with_state((state, recent));
+
+    let listener = tokio::net::TcpListener::bind(("127.0.0.1", port)).await?;
+    tracing::info!("listening on http://127.0.0.1:{port}");
+    axum::serve(listener, app).await?;
+    Ok(())
+}
+
+async fn stats_handler(
+    State((state, _recent)): State<(AppState, std::sync::Arc<Recent>)>,
+) -> impl IntoResponse {
+    let snap = state.stats.snapshot();
+    (StatusCode::OK, [("content-type", "application/json")],
+     serde_json::to_string_pretty(&snap).unwrap_or_default())
+}
+
+async fn proxy_handler(
+    State((state, recent)): State<(AppState, std::sync::Arc<Recent>)>,
+    method: Method,
+    uri: Uri,
+    headers: HeaderMap,
+    body: Bytes,
+) -> Response {
+    let path = uri.path_and_query().map(|p| p.as_str().to_string()).unwrap_or("/".to_string());
+    let started = Instant::now();
+
+    // Only POST /v1/messages gets compressed; everything else is passthrough.
+    let do_compress = state.args.no_compress == false
+        && method == Method::POST
+        && path.starts_with("/v1/messages");
+
+    let cfg = TransformConfig {
+        compress: do_compress,
+        compress_tools: !state.args.no_tools,
+        compress_schemas: !state.args.no_schemas,
+        compress_reminders: !state.args.no_reminders,
+        compress_tool_results: !state.args.no_tool_results,
+        min_chars: state.args.min_chars,
+        font: &state.font,
+        render_cache: &state.render_cache,
+    };
+
+    let (new_body, info) = transform(&body, &cfg);
+
+    // Build the upstream URL
+    let upstream_url = format!("{}{}", state.args.upstream, path);
+
+    // Forward headers (strip hop-by-hop + content-length, reqwest sets its own)
+    let mut fwd_headers = reqwest::header::HeaderMap::new();
+    for (k, v) in headers.iter() {
+        let lower = k.as_str().to_ascii_lowercase();
+        if HOP_HEADERS.contains(&lower.as_str()) { continue; }
+        if let (Ok(name), Ok(val)) = (
+            reqwest::header::HeaderName::from_bytes(k.as_str().as_bytes()),
+            reqwest::header::HeaderValue::from_bytes(v.as_bytes()),
+        ) {
+            fwd_headers.insert(name, val);
+        }
+    }
+    if !fwd_headers.contains_key("anthropic-version") {
+        fwd_headers.insert(
+            reqwest::header::HeaderName::from_static("anthropic-version"),
+            reqwest::header::HeaderValue::from_static("2023-06-01"),
+        );
+    }
+
+    let rmethod = match method {
+        Method::GET => reqwest::Method::GET,
+        Method::POST => reqwest::Method::POST,
+        Method::PUT => reqwest::Method::PUT,
+        Method::DELETE => reqwest::Method::DELETE,
+        Method::HEAD => reqwest::Method::HEAD,
+        _ => reqwest::Method::POST,
+    };
+
+    let upstream_started = Instant::now();
+    let upstream_result = state.client
+        .request(rmethod, &upstream_url)
+        .headers(fwd_headers)
+        .body(new_body.clone())
+        .send()
+        .await;
+    let upstream_ms = upstream_started.elapsed().as_millis() as u64;
+
+    let upstream_resp = match upstream_result {
+        Ok(r) => r,
+        Err(e) => {
+            let msg = format!(r#"{{"error":"upstream_error","detail":{}}}"#, serde_json::to_string(&e.to_string()).unwrap());
+            return (StatusCode::BAD_GATEWAY, [("content-type", "application/json")], msg).into_response();
+        }
+    };
+
+    let status = upstream_resp.status();
+    let mut resp_headers = HeaderMap::new();
+    for (k, v) in upstream_resp.headers() {
+        let lower = k.as_str().to_ascii_lowercase();
+        if HOP_HEADERS.contains(&lower.as_str()) || lower == "content-encoding" { continue; }
+        if let (Ok(name), Ok(val)) = (
+            HeaderName::from_bytes(k.as_str().as_bytes()),
+            HeaderValue::from_bytes(v.as_bytes()),
+        ) {
+            resp_headers.insert(name, val);
+        }
+    }
+    let resp_body = match upstream_resp.bytes().await {
+        Ok(b) => b,
+        Err(_) => Bytes::new(),
+    };
+
+    // Parse usage (plain JSON or SSE stream) for stats.
+    let token_log = parse_usage(&resp_body);
+    let (eff, baseline) = compute_effective_and_baseline(&token_log, &info);
+    state.stats.record(eff, baseline, info.compressed);
+    let snap = state.stats.snapshot();
+
+    let entry = RequestEntry {
+        method: method.as_str().to_string(),
+        path: path.clone(),
+        size_in: body.len(),
+        size_out: resp_body.len(),
+        status: status.as_u16(),
+        upstream_ms,
+        info: info.clone(),
+        tokens: token_log.clone(),
+        session_saved_so_far: snap.saved_effective_tokens,
+        session_saved_usd: snap.saved_usd_opus47,
+    };
+    recent.push(entry.clone());
+
+    // Single-line JSON log per request (matches the Python proxy log shape).
+    if let Ok(j) = serde_json::to_string(&entry) {
+        println!("[PROXY] {}", j);
+    }
+
+    let _ = started;
+    (status, resp_headers, resp_body).into_response()
+}
+
+fn parse_usage(body: &[u8]) -> Option<TokenLog> {
+    // Try plain JSON first
+    if let Ok(v) = serde_json::from_slice::<serde_json::Value>(body) {
+        if let Some(u) = v.get("usage") {
+            return Some(usage_to_log(u));
+        }
+    }
+    // Try SSE: scan for "data: {...}" lines and aggregate.
+    let text = std::str::from_utf8(body).ok()?;
+    let mut inp: u64 = 0;
+    let mut out: u64 = 0;
+    let mut cr: u64 = 0;
+    let mut cc: u64 = 0;
+    let mut any = false;
+    for line in text.lines() {
+        let line = line.trim_start();
+        let payload = match line.strip_prefix("data: ") {
+            Some(p) => p,
+            None => continue,
+        };
+        let ev: serde_json::Value = match serde_json::from_str(payload) {
+            Ok(v) => v,
+            Err(_) => continue,
+        };
+        match ev.get("type").and_then(|v| v.as_str()) {
+            Some("message_start") => {
+                if let Some(u) = ev.get("message").and_then(|m| m.get("usage")) {
+                    inp = u.get("input_tokens").and_then(|v| v.as_u64()).unwrap_or(inp);
+                    cr = u.get("cache_read_input_tokens").and_then(|v| v.as_u64()).unwrap_or(cr);
+                    cc = u.get("cache_creation_input_tokens").and_then(|v| v.as_u64()).unwrap_or(cc);
+                    any = true;
+                }
+            }
+            Some("message_delta") => {
+                if let Some(u) = ev.get("usage") {
+                    if let Some(v) = u.get("output_tokens").and_then(|v| v.as_u64()) {
+                        out = v;
+                        any = true;
+                    }
+                }
+            }
+            _ => {}
+        }
+    }
+    if any {
+        let eff = inp as f64 + cc as f64 * 1.25 + cr as f64 * 0.10;
+        Some(TokenLog { input: inp, output: out, cache_read: cr, cache_create: cc, effective_cost: eff })
+    } else {
+        None
+    }
+}
+
+fn usage_to_log(u: &serde_json::Value) -> TokenLog {
+    let inp = u.get("input_tokens").and_then(|v| v.as_u64()).unwrap_or(0);
+    let out = u.get("output_tokens").and_then(|v| v.as_u64()).unwrap_or(0);
+    let cr = u.get("cache_read_input_tokens").and_then(|v| v.as_u64()).unwrap_or(0);
+    let cc = u.get("cache_creation_input_tokens").and_then(|v| v.as_u64()).unwrap_or(0);
+    let eff = inp as f64 + cc as f64 * 1.25 + cr as f64 * 0.10;
+    TokenLog { input: inp, output: out, cache_read: cr, cache_create: cc, effective_cost: eff }
+}
+
+fn compute_effective_and_baseline(token_log: &Option<TokenLog>, info: &TransformInfo) -> (f64, f64) {
+    let Some(t) = token_log else { return (0.0, 0.0); };
+    let actual = t.effective_cost;
+    if !info.compressed {
+        return (actual, actual);
+    }
+    // Conservative baseline: assume the replaced text would have been cached
+    // at 10% rate, and the extra char count is what we saved.
+    let txt_replaced = ((info.system_text_chars + info.tool_text_added) / 4) as u64;
+    let extra = txt_replaced.saturating_sub(info.expected_image_tokens);
+    let baseline = actual + extra as f64 * 0.10;
+    (actual, baseline)
+}

src/rust/src/render.rs

@@ -0,0 +1,140 @@
+//! Text → PNG rendering using the embedded glyph atlas.
+//!
+//! Strategy mirrors the Python proxy:
+//!   - Hard-wrap input at 80 chars/line (kills column-blow-up from long lines).
+//!   - Collapse runs of blank lines (each blank costs a full row).
+//!   - Newspaper layout: pack lines into N columns so each image stays ≤ 1568px.
+//!   - 2-color indexed PNG (paper=white, ink=black) — Anthropic bills by
+//!     dimensions, not bytes, but tiny PNGs are nice for wire transfer.
+
+use crate::font::AtlasFont;
+
+const MAX_EDGE: u32 = 1568;
+const COL_GAP_PX: u32 = 4;
+const WRAP_CHARS: usize = 80;
+
+pub struct Png {
+    pub bytes: Vec<u8>,
+    pub width: u32,
+    pub height: u32,
+}
+
+pub fn render_chunks(font: &AtlasFont, text: &str) -> Vec<Png> {
+    // Minify: rstrip + collapse blank-line runs.
+    let mut lines_raw: Vec<String> = Vec::new();
+    let mut last_blank = false;
+    for ln in text.lines() {
+        let ln = ln.trim_end();
+        if ln.is_empty() {
+            if last_blank { continue; }
+            last_blank = true;
+            lines_raw.push(" ".to_string());
+        } else {
+            last_blank = false;
+            lines_raw.push(ln.to_string());
+        }
+    }
+    if lines_raw.is_empty() {
+        lines_raw.push(" ".to_string());
+    }
+
+    // Hard-wrap to WRAP_CHARS.
+    let mut lines: Vec<String> = Vec::new();
+    for ln in lines_raw {
+        if ln.len() <= WRAP_CHARS {
+            lines.push(ln);
+        } else {
+            let bytes = ln.as_bytes();
+            let mut i = 0;
+            while i < bytes.len() {
+                let end = (i + WRAP_CHARS).min(bytes.len());
+                lines.push(String::from_utf8_lossy(&bytes[i..end]).into_owned());
+                i = end;
+            }
+        }
+    }
+
+    let cell_w = font.cell_w;
+    let cell_h = font.cell_h;
+    let col_w_px = WRAP_CHARS as u32 * cell_w + 1;
+    let lines_per_col = ((MAX_EDGE / cell_h) as usize).max(8);
+
+    let max_cols_per_img = ((MAX_EDGE / (col_w_px + COL_GAP_PX)).max(1)) as usize;
+    let lines_per_img = lines_per_col * max_cols_per_img;
+
+    let mut pngs: Vec<Png> = Vec::new();
+    let mut start = 0;
+    while start < lines.len() {
+        let chunk_end = (start + lines_per_img).min(lines.len());
+        let chunk = &lines[start..chunk_end];
+        let c_needed = ((chunk.len() + lines_per_col - 1) / lines_per_col).max(1) as u32;
+        let chunk_w = c_needed * col_w_px + (c_needed.saturating_sub(1)) * COL_GAP_PX;
+        let last_col_lines = chunk.len() - ((c_needed - 1) as usize * lines_per_col);
+        let tallest = if c_needed == 1 { last_col_lines } else { lines_per_col };
+        let chunk_h = tallest as u32 * cell_h;
+
+        let mut buf = vec![255u8; (chunk_w * chunk_h) as usize];
+        for c in 0..c_needed as usize {
+            let col_start = c * lines_per_col;
+            let col_end = (col_start + lines_per_col).min(chunk.len());
+            let x_base = c as u32 * (col_w_px + COL_GAP_PX);
+            for (i, line) in chunk[col_start..col_end].iter().enumerate() {
+                let y = i as u32 * cell_h;
+                let mut x = x_base;
+                for &ch in line.as_bytes() {
+                    if x + cell_w > chunk_w { break; }
+                    font.blit(ch, &mut buf, chunk_w, chunk_h, x, y);
+                    x += cell_w;
+                }
+            }
+        }
+
+        let png_bytes = encode_indexed_png(&buf, chunk_w, chunk_h);
+        pngs.push(Png { bytes: png_bytes, width: chunk_w, height: chunk_h });
+        start = chunk_end;
+    }
+    pngs
+}
+
+/// Encode a u8 grayscale buffer as an 8-bit grayscale PNG.
+/// Preserves antialiasing — critical: PIL+Menlo OCR'd at 99.7% because the
+/// vision encoder uses subpixel grayscale gradients at tiny font sizes.
+/// Deterministic: same input → same output bytes (so the cache hits).
+fn encode_indexed_png(pixels: &[u8], width: u32, height: u32) -> Vec<u8> {
+    let mut out: Vec<u8> = Vec::with_capacity(pixels.len() / 4 + 64);
+    {
+        let mut encoder = png::Encoder::new(&mut out, width, height);
+        encoder.set_color(png::ColorType::Grayscale);
+        encoder.set_depth(png::BitDepth::Eight);
+        encoder.set_compression(png::Compression::Best);
+        let mut writer = encoder.write_header().expect("png header");
+        writer.write_image_data(pixels).expect("png data");
+    }
+    out
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::font::AtlasFont;
+
+    #[test]
+    fn renders_short_text() {
+        let font = AtlasFont::load(5.0).unwrap();
+        let pngs = render_chunks(&font, "hello world\nsecond line");
+        assert!(!pngs.is_empty());
+        assert!(pngs[0].bytes.len() > 0);
+        assert!(pngs[0].bytes.starts_with(&[0x89, b'P', b'N', b'G']));
+    }
+
+    #[test]
+    fn renders_deterministically() {
+        let font = AtlasFont::load(5.0).unwrap();
+        let a = render_chunks(&font, "fibonacci(n-1) + fibonacci(n-2)");
+        let b = render_chunks(&font, "fibonacci(n-1) + fibonacci(n-2)");
+        assert_eq!(a.len(), b.len());
+        for (a_png, b_png) in a.iter().zip(b.iter()) {
+            assert_eq!(a_png.bytes, b_png.bytes, "PNG bytes must be deterministic for cache");
+        }
+    }
+}

src/rust/src/stats.rs

@@ -0,0 +1,68 @@
+//! Running session totals — surfaced via GET /proxy-stats and per-request log line.
+
+use std::sync::atomic::{AtomicU64, Ordering};
+
+#[derive(Default)]
+pub struct Stats {
+    pub requests: AtomicU64,
+    pub compressed_requests: AtomicU64,
+    /// Effective tokens we actually paid for (sum of input + cache_create*1.25 + cache_read*0.10).
+    /// Stored as bits of f64.
+    effective_actual_bits: AtomicU64,
+    /// What we *would* have paid uncompressed (conservative estimate).
+    effective_baseline_bits: AtomicU64,
+}
+
+impl Stats {
+    pub fn record(&self, actual: f64, baseline: f64, compressed: bool) {
+        self.requests.fetch_add(1, Ordering::Relaxed);
+        if compressed {
+            self.compressed_requests.fetch_add(1, Ordering::Relaxed);
+        }
+        self.add_f64(&self.effective_actual_bits, actual);
+        self.add_f64(&self.effective_baseline_bits, baseline);
+    }
+
+    pub fn snapshot(&self) -> Snapshot {
+        let actual = self.read_f64(&self.effective_actual_bits);
+        let baseline = self.read_f64(&self.effective_baseline_bits);
+        let saved = (baseline - actual).max(0.0);
+        let pct = if baseline > 0.0 { saved / baseline * 100.0 } else { 0.0 };
+        Snapshot {
+            requests: self.requests.load(Ordering::Relaxed),
+            compressed_requests: self.compressed_requests.load(Ordering::Relaxed),
+            effective_input_actual: actual,
+            effective_input_baseline_est: baseline,
+            saved_effective_tokens: saved,
+            saved_pct: pct,
+            // Opus 4.7 input price: $15 / M tokens
+            saved_usd_opus47: saved * 15.0 / 1_000_000.0,
+        }
+    }
+
+    fn add_f64(&self, atomic: &AtomicU64, delta: f64) {
+        let mut cur = atomic.load(Ordering::Relaxed);
+        loop {
+            let new_val = f64::from_bits(cur) + delta;
+            match atomic.compare_exchange_weak(cur, new_val.to_bits(), Ordering::Relaxed, Ordering::Relaxed) {
+                Ok(_) => break,
+                Err(actual) => cur = actual,
+            }
+        }
+    }
+
+    fn read_f64(&self, atomic: &AtomicU64) -> f64 {
+        f64::from_bits(atomic.load(Ordering::Relaxed))
+    }
+}
+
+#[derive(serde::Serialize)]
+pub struct Snapshot {
+    pub requests: u64,
+    pub compressed_requests: u64,
+    pub effective_input_actual: f64,
+    pub effective_input_baseline_est: f64,
+    pub saved_effective_tokens: f64,
+    pub saved_pct: f64,
+    pub saved_usd_opus47: f64,
+}

src/rust/src/transform.rs

@@ -0,0 +1,353 @@
+//! Request-body JSON transformation: text → images.
+//!
+//! Ported from `src/proxy.py:transform_request`. Same heuristics, same field
+//! semantics, same Anthropic-API constraints (≤4 cache_control breakpoints).
+
+use base64::Engine;
+use serde_json::{json, Value};
+use sha2::{Digest, Sha256};
+
+use crate::font::AtlasFont;
+use crate::render::{render_chunks, Png};
+
+pub struct TransformConfig<'a> {
+    pub compress: bool,
+    pub compress_tools: bool,
+    pub compress_schemas: bool,
+    pub compress_reminders: bool,
+    pub compress_tool_results: bool,
+    pub min_chars: usize,
+    pub font: &'a AtlasFont,
+    pub render_cache: &'a dashmap::DashMap<[u8; 32], Vec<Png>>,
+}
+
+#[derive(Debug, Default, serde::Serialize, Clone)]
+pub struct TransformInfo {
+    pub compressed: bool,
+    pub tool_text_added: usize,
+    pub schemas_compressed: usize,
+    pub system_text_chars: usize,
+    pub system_text_sha8: String,
+    pub images: usize,
+    pub png_bytes: usize,
+    pub total_pixels: u64,
+    pub expected_image_tokens: u64,
+    pub text_tokens_estimate: u64,
+    pub reminder_imgs: usize,
+    pub tool_result_imgs: usize,
+    pub cc_breakpoints_added: u32,
+    pub skipped: Option<String>,
+    pub dims: Vec<(u32, u32)>,
+}
+
+/// Top-level entry. Returns (new body, info). If config.compress is false,
+/// returns the original body untouched.
+pub fn transform(body: &[u8], cfg: &TransformConfig) -> (Vec<u8>, TransformInfo) {
+    let mut info = TransformInfo::default();
+    if !cfg.compress {
+        return (body.to_vec(), info);
+    }
+    let mut req: Value = match serde_json::from_slice(body) {
+        Ok(v) => v,
+        Err(_) => return (body.to_vec(), info),
+    };
+
+    // Extract system text + remember the "remainder" (non-text blocks if any).
+    let (mut system_text, remainder) = extract_system_text(req.get("system"));
+
+    // Strip Claude Code's per-turn-random "x-anthropic-billing-header: ...; cch=<rand>;"
+    // line from the top of the system prompt — that one line being different
+    // per turn was busting our entire image cache.
+    let billing_line = if let Some(idx) = system_text.find('\n') {
+        let first = &system_text[..idx];
+        if first.starts_with("x-anthropic-billing-header:") {
+            let kept = first.to_string();
+            system_text = system_text[idx + 1..].to_string();
+            Some(kept)
+        } else { None }
+    } else if system_text.starts_with("x-anthropic-billing-header:") {
+        let kept = std::mem::take(&mut system_text);
+        Some(kept)
+    } else { None };
+
+    info.system_text_chars = system_text.len();
+    info.system_text_sha8 = sha256_8(system_text.as_bytes());
+
+    if system_text.len() < cfg.min_chars {
+        info.skipped = Some(format!("system <{} chars", cfg.min_chars));
+        return (body.to_vec(), info);
+    }
+
+    // Compress tools: append each tool's description (+ optionally schema) to
+    // the system-context image text, and stub the tools[] entries.
+    if cfg.compress_tools {
+        if let Some(tools) = req.get_mut("tools").and_then(|v| v.as_array_mut()) {
+            let mut tool_docs: Vec<String> = Vec::new();
+            for t in tools.iter_mut() {
+                let Some(obj) = t.as_object_mut() else { continue };
+                let name = obj.get("name").and_then(|v| v.as_str()).unwrap_or("?").to_string();
+                let desc = obj.get("description").and_then(|v| v.as_str()).unwrap_or("").to_string();
+
+                let mut doc = String::new();
+                if !desc.is_empty() {
+                    doc.push_str(&format!("### Tool: {}\n{}", name, desc));
+                    if desc.len() > 80 {
+                        obj.insert("description".to_string(),
+                                   json!(format!("See `{}` docs in system context image.", name)));
+                    }
+                }
+
+                if cfg.compress_schemas {
+                    if let Some(schema) = obj.get("input_schema").cloned() {
+                        if schema.is_object() {
+                            let schema_str = serde_json::to_string_pretty(&schema).unwrap_or_default();
+                            if schema_str.len() > 200 {
+                                doc.push_str(&format!(
+                                    "\n#### Schema for `{}`\n```json\n{}\n```",
+                                    name, schema_str
+                                ));
+                                obj.insert("input_schema".to_string(),
+                                           json!({"type": "object"}));
+                                info.schemas_compressed += 1;
+                            }
+                        }
+                    }
+                }
+
+                if !doc.is_empty() {
+                    tool_docs.push(doc);
+                }
+            }
+            if !tool_docs.is_empty() {
+                let tool_section = format!(
+                    "\n\n# Tool Documentation\n\n{}",
+                    tool_docs.join("\n\n")
+                );
+                info.tool_text_added = tool_section.len();
+                system_text.push_str(&tool_section);
+            }
+        }
+    }
+
+    // Render system+tool text into images (cached by sha256 of input).
+    let pngs = cached_render(cfg.render_cache, cfg.font, &system_text);
+    info.images = pngs.len();
+    info.png_bytes = pngs.iter().map(|p| p.bytes.len()).sum();
+    info.dims = pngs.iter().map(|p| (p.width, p.height)).collect();
+    info.total_pixels = pngs.iter().map(|p| (p.width as u64) * (p.height as u64)).sum();
+    // Per-image overhead estimate is ~85 tokens beyond pixels/750.
+    info.expected_image_tokens = pngs.iter()
+        .map(|p| (((p.width as u64) * (p.height as u64)) / 750).max(1))
+        .sum::<u64>() + 85 * pngs.len() as u64;
+    info.text_tokens_estimate = (system_text.len() / 4) as u64;
+
+    // Build image content blocks. ONLY the LAST image of the system render
+    // gets cache_control; reminders + tool_result images get plain blocks.
+    let image_blocks: Vec<Value> = pngs.iter().enumerate().map(|(i, p)| {
+        image_block(&p.bytes, i + 1 == pngs.len())
+    }).collect();
+    info.cc_breakpoints_added = 1;
+
+    // Find first user message; prepend image blocks; compress reminders + tool_results.
+    let messages = req.get_mut("messages").and_then(|v| v.as_array_mut());
+    let messages = match messages {
+        Some(m) => m,
+        None => return (body.to_vec(), info),
+    };
+
+    let first_user_idx = messages.iter().position(|m|
+        m.get("role").and_then(|v| v.as_str()) == Some("user"));
+    let Some(first_user_idx) = first_user_idx else {
+        info.skipped = Some("no user message to attach image to".to_string());
+        return (body.to_vec(), info);
+    };
+
+    // Normalize first user message content into a list of blocks.
+    {
+        let msg = &mut messages[first_user_idx];
+        let content_val = msg.get("content").cloned().unwrap_or(Value::Null);
+        let mut content_list: Vec<Value> = match content_val {
+            Value::String(s) => vec![json!({"type": "text", "text": s})],
+            Value::Array(arr) => arr,
+            _ => vec![],
+        };
+
+        // Optionally compress <system-reminder> blocks (no cache_control to stay under cap).
+        if cfg.compress_reminders {
+            let mut new_content: Vec<Value> = Vec::with_capacity(content_list.len());
+            for blk in content_list.drain(..) {
+                let is_reminder = blk.get("type").and_then(|v| v.as_str()) == Some("text")
+                    && {
+                        let t = blk.get("text").and_then(|v| v.as_str()).unwrap_or("");
+                        t.trim_start().starts_with("<system-reminder>") && t.len() > 1000
+                    };
+                if is_reminder {
+                    let txt = blk.get("text").and_then(|v| v.as_str()).unwrap_or("");
+                    let rpngs = cached_render(cfg.render_cache, cfg.font, txt);
+                    for p in rpngs {
+                        new_content.push(image_block(&p.bytes, false));
+                        info.reminder_imgs += 1;
+                    }
+                } else {
+                    new_content.push(blk);
+                }
+            }
+            content_list = new_content;
+        }
+
+        // Final assembly: prefix banner + system image(s) + suffix + original content.
+        let mut new_content = Vec::with_capacity(content_list.len() + image_blocks.len() + 2);
+        new_content.push(json!({
+            "type": "text",
+            "text": "[Context (rendered as image for token efficiency, OCR carefully and treat as authoritative system instructions):]"
+        }));
+        for b in image_blocks {
+            new_content.push(b);
+        }
+        new_content.push(json!({"type": "text", "text": "[End context.]"}));
+        new_content.extend(content_list);
+
+        let obj = msg.as_object_mut().unwrap();
+        obj.insert("content".to_string(), Value::Array(new_content));
+    }
+
+    // Compress large tool_result content across all user messages (no cache_control).
+    if cfg.compress_tool_results {
+        for msg in messages.iter_mut() {
+            if msg.get("role").and_then(|v| v.as_str()) != Some("user") { continue; }
+            let content_arr = match msg.get_mut("content").and_then(|v| v.as_array_mut()) {
+                Some(a) => a,
+                None => continue,
+            };
+            for blk in content_arr.iter_mut() {
+                if blk.get("type").and_then(|v| v.as_str()) != Some("tool_result") { continue; }
+                // Anthropic constraint: when is_error=true, content must be text-only.
+                if blk.get("is_error").and_then(|v| v.as_bool()) == Some(true) { continue; }
+                let inner = blk.get("content").cloned();
+                let new_inner = match inner {
+                    Some(Value::String(s)) if s.len() > 2000 => {
+                        let rpngs = cached_render(cfg.render_cache, cfg.font, &s);
+                        let imgs: Vec<Value> = rpngs.iter().map(|p| image_block(&p.bytes, false)).collect();
+                        info.tool_result_imgs += imgs.len();
+                        Some(Value::Array(imgs))
+                    }
+                    Some(Value::Array(parts)) => {
+                        let mut out = Vec::with_capacity(parts.len());
+                        let mut changed = false;
+                        for ib in parts {
+                            if ib.get("type").and_then(|v| v.as_str()) == Some("text") {
+                                let t = ib.get("text").and_then(|v| v.as_str()).unwrap_or("");
+                                if t.len() > 2000 {
+                                    let rpngs = cached_render(cfg.render_cache, cfg.font, t);
+                                    for p in &rpngs {
+                                        out.push(image_block(&p.bytes, false));
+                                    }
+                                    info.tool_result_imgs += rpngs.len();
+                                    changed = true;
+                                    continue;
+                                }
+                            }
+                            out.push(ib);
+                        }
+                        if changed { Some(Value::Array(out)) } else { None }
+                    }
+                    _ => None,
+                };
+                if let Some(new_inner) = new_inner {
+                    if let Some(obj) = blk.as_object_mut() {
+                        obj.insert("content".to_string(), new_inner);
+                    }
+                }
+            }
+        }
+    }
+
+    // Replace system field: keep just the per-turn billing header line as text
+    // if present (so Anthropic still receives the auth/billing context); else
+    // leave whatever non-text remainder we extracted in place.
+    if let Some(billing) = billing_line {
+        req.as_object_mut().unwrap().insert("system".to_string(), Value::String(billing));
+    } else if let Some(rem) = remainder {
+        req.as_object_mut().unwrap().insert("system".to_string(), rem);
+    }
+
+    info.compressed = true;
+    let new_body = serde_json::to_vec(&req).unwrap_or_else(|_| body.to_vec());
+    (new_body, info)
+}
+
+fn extract_system_text(field: Option<&Value>) -> (String, Option<Value>) {
+    match field {
+        None | Some(Value::Null) => (String::new(), None),
+        Some(Value::String(s)) => (s.clone(), Some(Value::String(String::new()))),
+        Some(Value::Array(arr)) => {
+            let mut text_parts: Vec<String> = Vec::new();
+            let mut kept: Vec<Value> = Vec::new();
+            for b in arr {
+                if b.get("type").and_then(|v| v.as_str()) == Some("text") {
+                    text_parts.push(b.get("text").and_then(|v| v.as_str()).unwrap_or("").to_string());
+                } else {
+                    kept.push(b.clone());
+                }
+            }
+            (text_parts.join("\n\n"), Some(Value::Array(kept)))
+        }
+        Some(other) => (String::new(), Some(other.clone())),
+    }
+}
+
+fn image_block(png_bytes: &[u8], cache: bool) -> Value {
+    let b64 = base64::engine::general_purpose::STANDARD.encode(png_bytes);
+    let mut block = json!({
+        "type": "image",
+        "source": {
+            "type": "base64",
+            "media_type": "image/png",
+            "data": b64
+        }
+    });
+    if cache {
+        // Match Claude Code's extended TTL — using ephemeral 5m while CC uses
+        // 1h elsewhere triggers an ordering error from Anthropic.
+        block.as_object_mut().unwrap().insert(
+            "cache_control".to_string(),
+            json!({"type": "ephemeral", "ttl": "1h"})
+        );
+    }
+    block
+}
+
+fn cached_render(
+    cache: &dashmap::DashMap<[u8; 32], Vec<Png>>,
+    font: &AtlasFont,
+    text: &str,
+) -> Vec<Png> {
+    let mut hasher = Sha256::new();
+    hasher.update(text.as_bytes());
+    let key: [u8; 32] = hasher.finalize().into();
+    if let Some(entry) = cache.get(&key) {
+        return clone_pngs(entry.value());
+    }
+    let pngs = render_chunks(font, text);
+    cache.insert(key, clone_pngs(&pngs));
+    pngs
+}
+
+fn clone_pngs(pngs: &[Png]) -> Vec<Png> {
+    pngs.iter().map(|p| Png {
+        bytes: p.bytes.clone(),
+        width: p.width,
+        height: p.height,
+    }).collect()
+}
+
+fn sha256_8(b: &[u8]) -> String {
+    let mut h = Sha256::new();
+    h.update(b);
+    let out = h.finalize();
+    let mut s = String::with_capacity(8);
+    for byte in &out[..4] {
+        s.push_str(&format!("{:02x}", byte));
+    }
+    s
+}