feat: resize and compress images before base64 encoding

[chaodu-agent]

Problem

Large Discord image attachments encoded as base64 exceed JSON-RPC transport limits, causing Internal Error (code: -32603) when sent via ACP to coding CLI agents.

Survey: How other ACP agents handle images

CLI Agent	ACP Support	Image Input	Large Image Issues
Claude Code	✅ Native ACP	base64 ImageContent	Multiple issues: empty base64, format errors (#34338, #4534, #6300)
Codex CLI	⚠️ Via codex-acp bridge	Native --image flag reads local file; ACP uses base64	Model-dependent vision support
Gemini CLI	❌ Not ACP	Local file / paste base64	1MB+ base64 hangs (#3955)

Key findings:

• All ACP agents use base64 — no file path standard exists in the ACP spec
• Large images are a shared pain point — Gemini CLI hangs at 1MB+, Claude Code has multiple base64-related bugs
• Compression before encoding is necessary regardless of which agent is used

Decision

Follow OpenClaw's approach: resize + compress images before base64 encoding on the openab side.

• Resize: longest side capped at 1200px (matching OpenClaw's DEFAULT_IMAGE_MAX_DIMENSION_PX)
• Compress: re-encode as JPEG at quality 75
• Result: ~200-400KB after base64, well within transport limits
• GIFs: pass through unchanged to preserve animation

Changes

• Cargo.toml: add image crate (jpeg, png, gif, webp features)
• src/discord.rs: replace raw base64 encoding with resize → compress → encode pipeline

References

• Issue: v0.6.4: Large image base64 payload causes Internal Error (-32603) via ACP #209
• Original feature: feat: support Discord image attachments via ACP ImageContent blocks #158 by @the3mi
• OpenClaw image sanitization: image-sanitization.ts, image-ops.ts
• Hermes Agent: downloads to local file (non-ACP, not applicable)

Fixes #209

[the3mi]

🔴 Critical: resize() distorts aspect ratio

File: src/discord.rs, line ~320

let img = img.resize(
    IMAGE_MAX_DIMENSION_PX,   // 1200
    IMAGE_MAX_DIMENSION_PX,   // ← forces image to 1200×1200 regardless of aspect ratio
    image::imageops::FilterType::Lanczos3,
);

image::Image::resize(w, h, filter) resizes the image to the exact given dimensions — it does not preserve aspect ratio automatically. A 4000×2000 landscape image gets squashed into 1200×1200, completely distorted.

---

Fix

let (w, h) = (img.width(), img.height());
let img = if w > IMAGE_MAX_DIMENSION_PX || h > IMAGE_MAX_DIMENSION_PX {
    let ratio = f64::from(IMAGE_MAX_DIMENSION_PX) / f64::max(w, h);
    let new_w = (f64::from(w) * ratio) as u32;
    let new_h = (f64::from(h) * ratio) as u32;
    img.resize(new_w, new_h, image::imageops::FilterType::Lanczos3)
} else {
    img
};

The test portrait_image_respects_aspect_ratio also has the same bug — it asserts result.width() == 1200 but should be 600 to match the correct aspect ratio (2000:4000 → 600:1200).

---

🟡 Minor: compression fallback does not re-check file size

When resize_and_compress fails, the code falls back to the original bytes. If the original image was >10MB, it still gets sent through, potentially exceeding transport limits. Add a size check on the fallback path.

---

🟢 Info

JPEG quality is 75, while OpenClaw uses progressive 85→35. If quality complaints come in later, this difference may be the cause.

[chaodu-agent]

Review Feedback

1. Post-download size check 被移除了

原本有 defense-in-depth 的下載後二次檢查 (bytes.len() > MAX_SIZE)，現在只剩 pre-download 的 attachment.size 檢查。attachment.size 是 Discord API 回報的值，理論上可信，但如果 CDN 回傳的實際 bytes 不一致（redirect、proxy 等），就少了一層保護。建議保留 post-download 的 bytes.len() 檢查。

2. Error log structured fields

簡化後的 error log 像 error!("download failed {url}: {e}") 少了 structured fields。原本用的格式比較適合 structured logging / grep：

// Before
error!("failed to download image {}: {}", url, e);

// Suggested
error!(url = %url, error = %e, "download failed");

建議統一用 tracing 的 structured field 風格，跟 codebase 其他地方一致。

[chaodu-agent]

Thanks @the3mi for the thorough review! All three points addressed in the latest commit (0e49dc4):

🔴 Fixed: aspect ratio preserved

let ratio = f64::from(IMAGE_MAX_DIMENSION_PX) / f64::from(max_side);
let new_w = (f64::from(w) * ratio) as u32;
let new_h = (f64::from(h) * ratio) as u32;
img.resize(new_w, new_h, ...)

4000×2000 → 1200×600 ✅, 2000×4000 → 600×1200 ✅

🟡 Fixed: fallback size check

When resize_and_compress fails, original bytes >1MB are now rejected instead of sent through.

🟢 Noted: JPEG quality 75 vs OpenClaw's progressive 85→35

Good callout. We can revisit if quality complaints come in — easy to add progressive reduction later.

Tests updated to match correct aspect ratio assertions.

[chaodu-agent]

🟡 Minor (non-blocking)

1. Post-download size check 仍未恢復

下載後的 bytes.len() 二次檢查還是沒加回來。目前只有 pre-download 的 attachment.size 檢查。雖然 resize 後通常會變小，但 defense-in-depth 多一層不嫌多：

let bytes = match response.bytes().await {
    Ok(b) => b,
    Err(e) => { error!(url = %url, error = %e, "read failed"); return None; }
};
if bytes.len() as u64 > MAX_SIZE {
    error!(filename = %attachment.filename, size = bytes.len(), "downloaded image exceeds limit");
    return None;
}

2. Error log 建議改用 structured fields

目前 inline format 如 error!("download failed {url}: {e}") 跟 codebase 其他地方的 tracing structured field 風格不一致，建議統一：

// 現在
error!("download failed {url}: {e}");
// 建議
error!(url = %url, error = %e, "download failed");

---

兩點都是 minor，不 block merge 👍

[chaodu-agent]

Minor feedback addressed in 1321060:

1. ✅ Post-download bytes.len() size check restored (defense-in-depth)
2. ✅ Error logs switched to structured fields: error!(url = %url, error = %e, "download failed")

[chaodu-agent]

@the3mi All your review feedback has been addressed across 3 commits. Mind taking another look and approving when you get a chance? 🦞🙏

• 🔴 Aspect ratio: fixed ✅
• 🟡 Fallback size check: added ✅
• 🟡 Post-download size check: restored ✅
• 🟡 Structured logging: fixed ✅
• 🟢 JPEG quality 75 vs progressive: noted for future