feat: keep image-heavy sessions within provider request-size limits (#1508)

* feat(kosong): classify HTTP 413 request-body-too-large as a dedicated error type

* feat(agent-core): lower default image downscale cap to 2000px and make it configurable

* feat(agent-core): strip media to text markers and retry when the compaction request is too large

* feat(agent-core): cap model-initiated image reads with a configurable byte budget

* feat(agent-core): resend with degraded media when the provider rejects the request body as too large

* test(agent-core): add explicit timeouts to encode-heavy image budget tests

* feat: add WebP decoding support with wasm integration

- Introduced a new WebP decoding module using @jsquash/webp's wasm decoder.
- Implemented functions to decode WebP images and check for animated WebP formats.
- Updated image compression tests to include scenarios for WebP handling, including encoding and decoding.
- Enhanced error handling for API request size limits to accommodate various error messages.
- Updated pnpm lockfile to include new dependencies for WebP encoding and decoding.

* chore(changeset): consolidate this PR's entries into one

* fix(nix): update pnpmDeps hash for merged lockfile

* feat(agent-core): refuse HEIC/HEIF reads with platform-matched conversion guidance

Author: RealKai42

.changeset/image-request-size-limits.md

@@ -0,0 +1,6 @@
+---
+"@moonshot-ai/kimi-code": patch
+---
+
+Keep image-heavy sessions within provider request-size limits: model-read images now honor a 256 KB per-image budget and a 2000px downscale cap (configurable via `[image]` in config.toml or `KIMI_IMAGE_*` env vars), oversized WebP is compressed as well, HEIC/HEIF reads are refused with a platform-matched conversion command instead of poisoning the session, and a request-too-large rejection (HTTP 413) now recovers automatically — the request and /compact both retry with older media replaced by text markers instead of failing the session.
+

apps/kimi-code/test/tui/controllers/editor-keyboard-image-paste.test.ts

@@ -141,14 +141,14 @@ describe('clipboard image paste compression', () => {
     if (att?.kind !== 'image') throw new Error('expected image attachment');
 
     // Stored metadata reflects the compressed size.
-    expect(Math.max(att.width, att.height)).toBeLessThanOrEqual(3000);
-    expect(att.placeholder).toContain('3000×1500');
+    expect(Math.max(att.width, att.height)).toBeLessThanOrEqual(2000);
+    expect(att.placeholder).toContain('2000×1000');
 
     // The stored bytes decode to the compressed dimensions — the thumbnail and
     // the submitted image both read from these bytes, so they cannot diverge.
     const dims = parseImageMeta(att.bytes);
     expect(dims).not.toBeNull();
-    expect(Math.max(dims!.width, dims!.height)).toBeLessThanOrEqual(3000);
+    expect(Math.max(dims!.width, dims!.height)).toBeLessThanOrEqual(2000);
   });
 
   it('records and persists the pre-compression original for an oversized paste', async () => {

docs/en/configuration/config-files.md

@@ -87,11 +87,12 @@ Fields in the config file fall into two categories: **top-level scalars** that d
 | `thinking` | `table` | — | Default parameters for Thinking mode → [`thinking`](#thinking) |
 | `loop_control` | `table` | — | Agent loop control parameters → [`loop_control`](#loop_control) |
 | `background` | `table` | — | Background task runtime parameters → [`background`](#background) |
+| `image` | `table` | — | Image compression parameters → [`image`](#image) |
 | `services` | `table` | — | Built-in external service configuration → [`services`](#services) |
 | `permission` | `table` | — | Initial permission rules → [`permission`](#permission) |
 | `hooks` | `array<table>` | — | Lifecycle hooks; see [Hooks](../customization/hooks.md) |
 
-The following sections cover each of the nested tables in turn: `providers`, `models`, `thinking`, `loop_control`, `background`, `services`, and `permission`.
+The following sections cover each of the nested tables in turn: `providers`, `models`, `thinking`, `loop_control`, `background`, `image`, `services`, and `permission`.
 
 ## `providers`
 
@@ -202,6 +203,17 @@ You can also switch models temporarily without touching the config file — by s
 
 In print mode (`kimi -p "<prompt>"`), Kimi Code runs a single non-interactive turn and exits as soon as the main agent finishes. If you launch background tasks (for example, concurrent subagents via `Agent(run_in_background=true)`) and need them to run to completion, set `keep_alive_on_exit = true`: the process then waits for every background task to reach a terminal state before exiting, bounded by `print_wait_ceiling_s`. Without it, the single turn ending tears background tasks down with the process.
 
+## `image`
+
+`image` controls how images are compressed before being sent to the model, across every ingestion point (pasted images, `ReadMediaFile` reads, images in MCP tool results, and so on).
+
+| Field | Type | Default | Description |
+| --- | --- | --- | --- |
+| `max_edge_px` | `integer` | `2000` | Longest-edge ceiling in pixels. Larger images are scaled down proportionally to fit; raising it preserves more detail at the cost of larger request bodies |
+| `read_byte_budget` | `integer` | `262144` (256 KB) | Per-image byte budget for images the model reads for itself (`ReadMediaFile` default reads). It bounds the accumulated request-body size when the model keeps screenshotting and reading images; fine detail stays reachable through the `region` parameter, which reads a crop back at full fidelity (`region` and `full_resolution` are not subject to this budget) |
+
+`max_edge_px` can be overridden by the `KIMI_IMAGE_MAX_EDGE_PX` environment variable and `read_byte_budget` by `KIMI_IMAGE_READ_BYTE_BUDGET`; both take higher priority than `config.toml`.
+
 <!--
 ## `experimental`
 

docs/en/configuration/env-vars.md

@@ -122,6 +122,8 @@ Switches that control the behavior of subsystems such as telemetry, background t
 | --- | --- | --- |
 | `KIMI_DISABLE_TELEMETRY` | Disable anonymous telemetry reporting | `1`, `true`, `yes`, `y` (case-insensitive) |
 | `KIMI_CODE_BACKGROUND_KEEP_ALIVE_ON_EXIT` | Whether to keep background tasks when the session closes; takes higher priority than `config.toml`. The default is to stop them on exit | Truthy: `1`/`true`/`yes`/`on`; falsy: `0`/`false`/`no`/`off` |
+| `KIMI_IMAGE_MAX_EDGE_PX` | Longest-edge ceiling (px) for image compression; takes higher priority than `[image] max_edge_px` in `config.toml` (default `2000`) | Positive integer; invalid values are ignored |
+| `KIMI_IMAGE_READ_BYTE_BUDGET` | Per-image byte budget for model-initiated image reads (`ReadMediaFile` default reads); takes higher priority than `[image] read_byte_budget` in `config.toml` (default `262144`, i.e. 256 KB) | Positive integer; invalid values are ignored |
 | `KIMI_CODE_PLUGIN_MARKETPLACE_URL` | Override the plugin marketplace JSON loaded by `/plugins`; useful for dev loopback servers, staging CDN files, or alternate marketplace directories | `https://code.kimi.com/kimi-code/plugins/marketplace.json`; also accepts `http://`, `file://` URLs, and local paths |
 | `KIMI_CODE_AGENT_SWARM_MAX_CONCURRENCY` | Cap how many AgentSwarm subagents run concurrently during the initial ramp; leave unset for no cap | Positive integer; invalid values fail fast |
 | `KIMI_CODE_EXPERIMENTAL_FLAG` | Enable all registered experimental features for this process | `1`, `true`, `yes`, `on` |

docs/zh/configuration/config-files.md

@@ -87,11 +87,12 @@ timeout = 5
 | `thinking` | `table` | — | Thinking 模式默认参数 → [`thinking`](#thinking) |
 | `loop_control` | `table` | — | Agent 循环控制参数 → [`loop_control`](#loop_control) |
 | `background` | `table` | — | 后台任务运行参数 → [`background`](#background) |
+| `image` | `table` | — | 图片压缩参数 → [`image`](#image) |
 | `services` | `table` | — | 内置外部服务配置 → [`services`](#services) |
 | `permission` | `table` | — | 初始权限规则 → [`permission`](#permission) |
 | `hooks` | `array<table>` | — | 生命周期 hook，详见 [Hooks](../customization/hooks.md) |
 
-以下各节对 `providers`、`models`、`thinking`、`loop_control`、`background`、`services`、`permission` 等嵌套表逐一展开。
+以下各节对 `providers`、`models`、`thinking`、`loop_control`、`background`、`image`、`services`、`permission` 等嵌套表逐一展开。
 
 ## `providers`
 
@@ -202,6 +203,17 @@ display_name = "Kimi for Coding (custom)"
 
 在 print 模式（`kimi -p "<prompt>"`）下，Kimi Code 只跑一个非交互的单轮 turn，主 agent 一结束就退出。如果你启动了后台任务（例如通过 `Agent(run_in_background=true)` 并发子代理）并希望它们跑完，请设置 `keep_alive_on_exit = true`：进程会在退出前等待所有后台任务进入终态，最长不超过 `print_wait_ceiling_s`。否则，单轮 turn 结束时后台任务会随进程一起被清理。
 
+## `image`
+
+`image` 控制图片发送给模型前的压缩行为，对所有图片入口生效（粘贴图片、`ReadMediaFile` 读图、MCP 工具结果里的图片等）。
+
+| 字段 | 类型 | 默认值 | 说明 |
+| --- | --- | --- | --- |
+| `max_edge_px` | `integer` | `2000` | 图片最长边上限（像素）。超过时按比例缩小到该值以内；调大可保留更多细节，代价是更大的请求体积 |
+| `read_byte_budget` | `integer` | `262144`（256 KB） | 模型自行读取的图片（`ReadMediaFile` 默认读取）的单图字节预算。会话中模型反复截图、读图时，累计请求体大小由它控制；细节可通过 `region` 参数按原图坐标全保真回读（`region` 与 `full_resolution` 不受此预算限制） |
+
+`max_edge_px` 可被环境变量 `KIMI_IMAGE_MAX_EDGE_PX` 覆盖，`read_byte_budget` 可被 `KIMI_IMAGE_READ_BYTE_BUDGET` 覆盖，优先级均高于配置文件。
+
 <!--
 ## `experimental`
 

docs/zh/configuration/env-vars.md

@@ -122,6 +122,8 @@ kimi
 | --- | --- | --- |
 | `KIMI_DISABLE_TELEMETRY` | 关闭匿名遥测上报 | `1`、`true`、`yes`、`y`（不区分大小写） |
 | `KIMI_CODE_BACKGROUND_KEEP_ALIVE_ON_EXIT` | 会话关闭时是否保留后台任务，优先级高于 `config.toml`。默认会在退出时停止后台任务 | 真值：`1`/`true`/`yes`/`on`；假值：`0`/`false`/`no`/`off` |
+| `KIMI_IMAGE_MAX_EDGE_PX` | 图片压缩的最长边上限（像素），优先级高于 `config.toml` 的 `[image] max_edge_px`（默认 `2000`） | 正整数；非法值被忽略 |
+| `KIMI_IMAGE_READ_BYTE_BUDGET` | 模型自行读图（`ReadMediaFile` 默认读取）的单图字节预算，优先级高于 `config.toml` 的 `[image] read_byte_budget`（默认 `262144`，即 256 KB） | 正整数；非法值被忽略 |
 | `KIMI_CODE_PLUGIN_MARKETPLACE_URL` | 覆盖 `/plugins` 加载的 plugin marketplace JSON，适合 dev loopback server、测试 CDN 文件或替换 marketplace 目录 | `https://code.kimi.com/kimi-code/plugins/marketplace.json`；也接受 `http://`、`file://` URL 和本地路径 |
 | `KIMI_CODE_AGENT_SWARM_MAX_CONCURRENCY` | 限制 AgentSwarm 初始提升并发阶段可同时运行的子 Agent 数量；不设置表示不限制 | 正整数；非法值会立即失败 |
 | `KIMI_CODE_EXPERIMENTAL_FLAG` | 在当前进程启用所有已注册的实验功能 | `1`、`true`、`yes`、`on` |

flake.nix

@@ -152,7 +152,7 @@
               inherit (finalAttrs) pname version src pnpmWorkspaces;
               inherit pnpm;
               fetcherVersion = 3;
-              hash = "sha256-hUn5Srn3HnEEzU5DLxgjIzFjI0ukM3iSP4QagftEXdE=";
+              hash = "sha256-i3T30dxNQ7DTWOmUX47kmqpIVtDYby0buPo58+Jhhmw=";
             };
 
             nativeBuildInputs = [

packages/agent-core/package.json

@@ -59,6 +59,7 @@
   },
   "dependencies": {
     "@antfu/utils": "^9.3.0",
+    "@jsquash/webp": "^1.5.0",
     "@modelcontextprotocol/sdk": "^1.29.0",
     "@moonshot-ai/kaos": "workspace:^",
     "@moonshot-ai/kimi-code-oauth": "workspace:^",

packages/agent-core/scripts/generate-webp-dec-wasm.mjs

@@ -0,0 +1,36 @@
+/**
+ * Regenerate `src/tools/support/webp-dec-wasm.ts` from the installed
+ * `@jsquash/webp` package.
+ *
+ * The WebP decoder wasm is committed as a base64 string module because the
+ * published CLI bundles every dependency into a single file with no runtime
+ * node_modules — a file-path lookup for the .wasm would break there, while a
+ * string constant survives every packaging (vitest on sources, tsdown
+ * bundling, nix builds) unchanged. Run this after bumping @jsquash/webp:
+ *
+ *   node scripts/generate-webp-dec-wasm.mjs
+ */
+import { createRequire } from 'node:module';
+import { readFileSync, writeFileSync } from 'node:fs';
+import { resolve } from 'node:path';
+
+const packageRoot = resolve(import.meta.dirname, '..');
+const require = createRequire(resolve(packageRoot, 'package.json'));
+
+const wasmPath = require.resolve('@jsquash/webp/codec/dec/webp_dec.wasm');
+const version = require('@jsquash/webp/package.json').version;
+const wasm = readFileSync(wasmPath);
+
+const target = resolve(packageRoot, 'src/tools/support/webp-dec-wasm.ts');
+writeFileSync(
+  target,
+  `// GENERATED FILE — do not edit by hand.
+// WebP decoder wasm from @jsquash/webp@${version} (codec/dec/webp_dec.wasm),
+// base64-encoded so the bundled CLI needs no on-disk wasm asset.
+// Regenerate with: node scripts/generate-webp-dec-wasm.mjs
+
+export const WEBP_DECODER_WASM_BASE64 =
+  '${wasm.toString('base64')}';
+`,
+);
+console.log(`Wrote ${target} (${wasm.length} bytes of wasm)`);

packages/agent-core/src/agent/compaction/full.ts

@@ -8,10 +8,12 @@ import {
   APIEmptyResponseError,
   inputTotal,
   isRetryableGenerateError,
+  type ContentPart,
   type GenerateResult,
   type Message,
   type TokenUsage,
   APIContextOverflowError,
+  APIRequestTooLargeError,
   APIStatusError,
   createUserMessage,
 } from '@moonshot-ai/kosong';
@@ -430,6 +432,7 @@ export class FullCompaction {
       // prefix-race check and `compactedCount`.
       let historyForModel: readonly ContextMessage[] = stripDynamicToolContext(originalHistory);
       let droppedCount = 0;
+      let mediaStripAttempted = false;
       let overflowShrinkCount = 0;
       let emptyOrTruncatedShrinkCount = 0;
       while (true) {
@@ -465,14 +468,34 @@ export class FullCompaction {
           summary = extractCompactionSummary(response);
           break;
         } catch (error) {
+          // A request-body-size rejection (HTTP 413) is first retried with
+          // media parts replaced by text markers: accumulated base64 payloads
+          // are the usual culprit, and a text summary does not need them —
+          // the conversation already narrates what was seen, and the
+          // ReadMediaFile `<image path="...">` text wrapper survives. Only
+          // the summarizer input copy is rewritten; the real history keeps
+          // its media. A 413 after the strip (or with no media to strip)
+          // falls through to the overflow shrink below — dropping oldest
+          // messages shrinks the body too.
+          if (error instanceof APIRequestTooLargeError && !mediaStripAttempted) {
+            mediaStripAttempted = true;
+            const stripped = replaceMediaPartsWithMarkers(historyForModel);
+            if (stripped !== historyForModel) {
+              historyForModel = stripped;
+              retryCount = 0;
+              continue;
+            }
+          }
           const isContextOverflow = this.shouldRecoverFromContextOverflow(
             error,
             estimatedCompactionRequestTokens,
           );
           if (isContextOverflow) {
             this.observeContextOverflow(estimatedCompactionRequestTokens);
           }
-          if (isContextOverflow && historyForModel.length > 1) {
+          const shouldShrinkAfterOverflow =
+            isContextOverflow || error instanceof APIRequestTooLargeError;
+          if (shouldShrinkAfterOverflow && historyForModel.length > 1) {
             overflowShrinkCount += 1;
             if (overflowShrinkCount > MAX_COMPACTION_OVERFLOW_SHRINK_ATTEMPTS) {
               throw error;
@@ -639,6 +662,40 @@ export class FullCompaction {
 const MAX_COMPACTION_OVERFLOW_SHRINK_ATTEMPTS = 3;
 const COMPACTION_OVERFLOW_SHRINK_RATIOS = [0.7, 0.5, 0.35] as const;
 
+const MEDIA_PART_MARKERS = {
+  image_url: '[image]',
+  audio_url: '[audio]',
+  video_url: '[video]',
+} as const;
+
+function isMediaPart(part: ContentPart): part is ContentPart & { type: keyof typeof MEDIA_PART_MARKERS } {
+  return part.type in MEDIA_PART_MARKERS;
+}
+
+/**
+ * Replace media parts (image/audio/video) with text markers in the summarizer
+ * input, for the 413 strip-and-retry above. Messages without media are
+ * returned by reference (keeping the per-message token-estimate cache warm),
+ * and when nothing changed the input array itself is returned so the caller
+ * can tell there was no media to strip.
+ */
+function replaceMediaPartsWithMarkers(
+  messages: readonly ContextMessage[],
+): readonly ContextMessage[] {
+  let changed = false;
+  const out = messages.map((message) => {
+    if (!message.content.some(isMediaPart)) return message;
+    changed = true;
+    return {
+      ...message,
+      content: message.content.map((part): ContentPart =>
+        isMediaPart(part) ? { type: 'text', text: MEDIA_PART_MARKERS[part.type] } : part,
+      ),
+    };
+  });
+  return changed ? out : messages;
+}
+
 function shrinkCompactionHistoryAfterOverflow<T extends Message>(
   messages: readonly T[],
   attempt: number,