fix(tools): forward nested image results in code mode

.agents/notes/implemented/feature/2026-07-20-code-mode-typed-tool-returns.i18n.yaml

@@ -2,5 +2,5 @@
 # side as of the last confirmed-consistent state. Both languages carry equal authority;
 # after editing either side, bring the other along and re-record with:
 #   pnpm run verify-translation-pairing --write .agents/notes/implemented/feature/2026-07-20-code-mode-typed-tool-returns.md
-2026-07-20-code-mode-typed-tool-returns.md: 6b31dbca21a22a24f2bbb0ef61097622884165c3
+2026-07-20-code-mode-typed-tool-returns.md: 6bb4ccdeb81172ec9102a8656d380dfed09b63ae
-2026-07-20-code-mode-typed-tool-returns.zh.md: 1bb88d29ffd96d65ce06033202499d2780814abd
+2026-07-20-code-mode-typed-tool-returns.zh.md: 49084d6104ce2f733810126e4bb7cf78e92740e5

.agents/notes/implemented/feature/2026-07-20-code-mode-typed-tool-returns.md

@@ -14,7 +14,7 @@ The [canonical tool-output contract](../architecture/2026-07-20-canonical-tool-o
 
 ## Decision
 
-Code Mode is a typed projection of the visible tool registry. Each successful binding resolves to the final canonical `JsonValue` after post-execute policy, while a failed binding rejects with a real `ToolCallError`. Intermediate values remain inside the run and cross the worker boundary whole. Only the outer `run_code` logs, completion value, or failure diagnostic enter the configurable output ledger and the model-facing spill pipeline.
+Code Mode is a typed projection of the visible tool registry. Each successful binding resolves to the final canonical `JsonValue` after post-execute policy, while a failed binding rejects with a real `ToolCallError`. Intermediate values remain inside the run and cross the worker boundary whole. The outer `run_code` logs, completion value, or failure diagnostic enter the configurable output ledger and model-facing spill pipeline; a successfully settled sub-call whose final Native content contains an image additionally defers that complete ordered content through the parent result as logged, source-attributed context.
 
 This note owns the return and failure contract layered on the original [Code Mode foundation](2026-06-15-code-mode.md). The unified schema vocabulary is owned by the [JSON-value schema DSL note](../architecture/2026-07-20-unified-json-value-schema-dsl.md), and Native rendering and policy projection remain owned by the canonical-output note.
 
@@ -49,7 +49,7 @@ declare const tools: {
 
 ### Binding values and failures
 
-Before dispatch the bridge snapshots binding arguments as lossless JSON and snapshots the detached value again for an independent durable summary event. Host-side detachment, immutable execution, and output-schema projection all use iterative traversals rather than nested structured clone or recursive freezing. `undefined`, non-finite numbers, `-0`, sparse arrays, cycles, functions, and exotic objects reject that call before the tool runs. Successful dispatch returns `ToolExecutionResult.value`; Native `content`, metadata, and internal error information do not cross to the program.
+Before dispatch the bridge snapshots binding arguments as lossless JSON and snapshots the detached value again for an independent durable summary event. Host-side detachment, immutable execution, and output-schema projection all use iterative traversals rather than nested structured clone or recursive freezing. `undefined`, non-finite numbers, `-0`, sparse arrays, cycles, functions, and exotic objects reject that call before the tool runs. Successful dispatch returns `ToolExecutionResult.value`; Native `content`, metadata, and internal error information do not cross to the program. Image-bearing final content is not a second binding value: the bridge ferries it after the outer result so the next model request can see the durable image, while post-execute block/content replacement remains authoritative and text-only results are not duplicated.
 
 Code Mode declares its rejection capability on the runtime request as `{ name: "ToolCallError", memberNameProperty: "toolName" }`. The runtime Service Definition treats those names as data: the worker materializes and injects the actual constructor used for `tools` binding failures, so `error instanceof ToolCallError` works without making a generic runtime know about tools. The worker constructs failures and defines their public fields through module-captured error and property-definition intrinsics plus null-prototype descriptors, so model mutations cannot replace the promised rejection with a worker failure. The error has the standard `Error` message plus the exact `toolName`; it deliberately omits `ToolFailure.info`, error codes, and Native content. This is an exception contract for control flow, not a failure union for programmatic classification.
 
@@ -73,13 +73,13 @@ Temporary Cordis Plugins follow the same rule: `cordis_mount` returns `{ id, plu
 
 ### Persistence, metadata, and spill
 
-Nested dispatch logs the sub-call's full rendered `content`/`isError` on `tool/code-dispatch` but does not persist canonical values. `tool/result` continues to persist only rendered content, error, and optional metadata. `SESSION_FORMAT_VERSION` remains unchanged (pre-release shape churn does not bump it) and replay cannot recreate intermediate canonical program values.
+Nested dispatch logs the sub-call's full rendered `content`/`isError` on `tool/code-dispatch` but does not persist canonical values. `tool/result` continues to persist only rendered content, error, and optional metadata. A successful final content sequence containing an image is also wrapped in a source-attributed user message and deferred through the outer result; the normal session event makes that model-visible input reconstructable. `SESSION_FORMAT_VERSION` remains unchanged (pre-release shape churn does not bump it) and replay cannot recreate intermediate canonical program values.
 
 The opaque `exec.parent` token marks nested calls. Presentation metadata and generic or tool-owned spill projections skip those calls because they have no direct result card and their canonical values never enter context. The outer `run_code` call alone produces one card and may spill its final post-policy presentation; `run_code` intentionally declares neither a result presenter nor presentation metadata, so UI adapters complete the card through their generic raw-content fallback using durable `tool/result.content`.
 
 ## Testing
 
-Compile-time and snapshot tests pin exact `ToolArgsMap`, `ToolOutputMap`, `ToolName`, schema-to-TypeScript coverage, and exotic names. Registry and real-worker tests cover scalar, array, object, and null values; raw string rendering; absent `undefined`; consumer-declared real rejection classes, including `ToolCallError`; invalid arguments and completions, including intrinsic-looking forged prototypes; model-mutated JSON-boundary globals, prototype methods, constructor slots, and inherited descriptor fields; typed binding failures after those mutations; large uncapped intermediate bindings; nested spill suppression; exact and over-limit 64 MiB accounting; combined logs/value/diagnostic accounting; giant thrown stacks; bounded failure spill; hostile forged traffic; and built-package execution.
+Compile-time and snapshot tests pin exact `ToolArgsMap`, `ToolOutputMap`, `ToolName`, schema-to-TypeScript coverage, exotic names, and assembled Code Mode image forwarding. Registry and real-worker tests cover scalar, array, object, and null values; raw string rendering; absent `undefined`; consumer-declared real rejection classes, including `ToolCallError`; invalid arguments and completions, including intrinsic-looking forged prototypes; model-mutated JSON-boundary globals, prototype methods, constructor slots, and inherited descriptor fields; typed binding failures after those mutations; large uncapped intermediate bindings; nested spill suppression; generic image-bearing context deferral plus post-execute replacement/block precedence; exact and over-limit 64 MiB accounting; combined logs/value/diagnostic accounting; giant thrown stacks; bounded failure spill; hostile forged traffic; and built-package execution.
 
 Keyless real-worker integration tests pin the two handle workflows that prose results could not safely support. A background bash call returns its task id, the outer run settles, and a later run polls that id to completion; separate cases prove pre-abort creates no task, post-publication call abort preserves the task, foreground execution stays signal-coupled, and `task_kill` owns cancellation. A Cordis program reads an active or pending mount's id and `waitingFor` fields directly, unmounts by that id, and confirms removal without parsing rendered text.
 
@@ -93,6 +93,10 @@ Keyless real-worker integration tests pin the two handle workflows that prose re
 
 **Silently inspect or truncate an oversized completion.** Rejected because changing a JSON value into a string is lossy and type-incorrect. The explicit `output-limit` failure lets the model choose a smaller result, while the retained logs and diagnostic can still use normal outer spill.
 
+**Require each rich leaf tool to inspect `exec.parent` and defer itself.** Rejected because it couples leaf tools to Code Mode internals, duplicates policy handling, and misses future rich tools. The dispatch bridge owns generic forwarding from the already settled final result.
+
+**Expose Native rich content as part of every binding's canonical value.** Rejected because a canonical value is lossless JSON and tool-specific; attachment blocks are a model projection with durable lifecycle semantics. Keeping the value and projection separate preserves typed programs without dropping images from later model context.
+
 ## Consequences
 
 Code programs can compose tools through stable values instead of reverse-engineering Native prose. Native and Both Mode retain their existing text and UI presentation, while Code Mode receives output-schema types and exact runtime JSON. Tool authors must treat the canonical value as their programmatic API and put display-only formatting in the renderer.
@@ -107,6 +111,6 @@ The worker performs bounded-depth flat-wire transport and lossless validation bu
 - Intermediate values have no byte cap and can exhaust process or worker memory through retention, flat-wire copies, or structured-clone cost.
 - The 64 MiB hard cap applies only to the outer variable payloads, excluding fixed result-envelope syntax and presentation whitespace; spill cannot recover bytes rejected beyond that cap.
 - Provider or executor acquisition limits may already have discarded source data before a canonical value reaches Code Mode.
-- Unsupported MCP output schemas fall back to `JsonValue`; richer Native multimedia projection is deferred.
+- Unsupported MCP output schemas fall back to `JsonValue`; admitted MCP images use the generic deferred projection, while audio and embedded-resource payloads remain diagnostic-only.
 - There is one result card per outer `run_code`, never per nested call.
 - Code failures expose `ToolCallError` message and tool name only, without a programmatic error-code union.

.agents/notes/implemented/feature/2026-07-20-code-mode-typed-tool-returns.zh.md

@@ -14,7 +14,7 @@ Code Mode 过去会把每个嵌套工具的结果从 `ContentBlock[]` 重新投
 
 ## 决策
 
-Code Mode 是可见工具注册表的类型化投影。每个成功的绑定调用都会解析为 post-execute 策略处理后的最终规范 `JsonValue`，失败的绑定调用则会以真正的 `ToolCallError` 拒绝 Promise。中间值只存在于本次运行中，并完整跨越 worker 边界。只有外层 `run_code` 的日志、完成值或失败诊断会进入可配置的输出账本以及面向模型的输出落盘流水线。
+Code Mode 是可见工具注册表的类型化投影。每个成功的绑定调用都会解析为 post-execute 策略处理后的最终规范 `JsonValue`，失败的绑定调用则会以真正的 `ToolCallError` 拒绝 Promise。中间值只存在于本次运行中，并完整跨越 worker 边界。外层 `run_code` 的日志、完成值或失败诊断会进入可配置的输出账本以及面向模型的输出落盘流水线；如果成功结算的子调用最终 Native 内容包含图片，其完整有序内容还会经父结果延后为写入日志且带来源归属的上下文。
 
 本文档定义叠加在原始 [Code Mode 基础](2026-06-15-code-mode.md)之上的返回值与失败约定。统一 schema 词汇由 [JSON 值 schema DSL Agent Note](../architecture/2026-07-20-unified-json-value-schema-dsl.md)负责定义；Native 渲染与策略投影仍由规范输出 Agent Note 负责定义。
 
@@ -49,7 +49,7 @@ declare const tools: {
 
 ### 绑定值与失败
 
-分发前，桥接层会把绑定参数快照为无损 JSON，再对分离后的值生成一次快照，供独立的持久摘要事件使用。宿主侧的值分离、执行数据的不可变处理与输出 schema 投影均采用迭代遍历，而不使用嵌套结构化克隆或递归冻结。`undefined`、非有限数、`-0`、稀疏数组、循环引用、函数和非普通对象都会使该调用在工具运行前被拒绝。成功分发会返回 `ToolExecutionResult.value`；Native `content`、元数据和内部错误信息不会传入程序。
+分发前，桥接层会把绑定参数快照为无损 JSON，再对分离后的值生成一次快照，供独立的持久摘要事件使用。宿主侧的值分离、执行数据的不可变处理与输出 schema 投影均采用迭代遍历，而不使用嵌套结构化克隆或递归冻结。`undefined`、非有限数、`-0`、稀疏数组、循环引用、函数和非普通对象都会使该调用在工具运行前被拒绝。成功分发会返回 `ToolExecutionResult.value`；Native `content`、元数据和内部错误信息不会传入程序。含图片的最终内容不是第二份绑定值：桥接层会在外层结果之后转运它，使下一次模型请求可以看到持久图片；post-execute 阻止／内容替换仍具有权威性，纯文本结果不会重复。
 
 Code Mode 通过运行时请求中的 `{ name: "ToolCallError", memberNameProperty: "toolName" }` 声明其以异常拒绝 Promise 的能力。运行时 Service Definition 只把这些名称视为数据：worker 会动态生成并注入真正用于 `tools` 绑定失败的构造函数，因此无需让通用运行时了解工具，`error instanceof ToolCallError` 也能成立。worker 使用模块初始化时捕获的 Error 构造函数与属性定义内建方法，配合原型为 null 的属性描述符，构造失败对象并定义其公开字段，因此模型代码的修改不会把约定承诺的 reject 变成 worker 失败。该错误包含标准的 `Error` 消息和确切的 `toolName`，并有意省略 `ToolFailure.info`、错误代码与 Native 内容。这是一项用于控制流的异常约定，而不是供程序分类的失败联合。
 
@@ -73,13 +73,13 @@ Code Mode 通过运行时请求中的 `{ name: "ToolCallError", memberNameProper
 
 ### 持久化、元数据与输出落盘
 
-嵌套分发在 `tool/code-dispatch` 上记录子调用完整渲染后的 `content`/`isError`，但不会持久化规范值。`tool/result` 继续只持久化渲染后的内容、错误和可选元数据。`SESSION_FORMAT_VERSION` 保持不变（预发布阶段的形状变动不递增版本号），回放也无法重建程序的规范中间值。
+嵌套分发在 `tool/code-dispatch` 上记录子调用完整渲染后的 `content`/`isError`，但不会持久化规范值。`tool/result` 继续只持久化渲染后的内容、错误和可选元数据。包含图片的成功最终内容序列还会包装成带来源归属的用户消息，并经外层结果延后；普通会话事件使该模型可见输入可以重建。`SESSION_FORMAT_VERSION` 保持不变（预发布阶段的形状变动不递增版本号），回放也无法重建程序的规范中间值。
 
 不透明的 `exec.parent` token 用于标识嵌套调用。由于这些调用没有直接对应的结果卡片，而且其规范值永远不会进入上下文，展示元数据以及通用或工具自有的输出落盘投影都会跳过它们。只有外层 `run_code` 调用会生成一张卡片，并且可能将 post-policy 处理后的最终展示写入落盘文件；`run_code` 有意既不声明结果展示器，也不声明展示元数据，因此 UI 适配器会通过通用的原始内容回退机制，使用持久化的 `tool/result.content` 补全该卡片。
 
 ## 测试
 
-编译期测试与快照测试锁定了精确的 `ToolArgsMap`、`ToolOutputMap`、`ToolName`、schema 到 TypeScript 的覆盖范围以及特殊名称。注册表与真实 worker 测试覆盖标量、数组、对象和 null 值；字符串原文渲染；缺席的 `undefined`；消费方声明、实际用于拒绝 Promise 的异常类，包括 `ToolCallError`；无效参数与完成值，包括伪装为内建原型的伪造原型；模型代码修改过的 JSON 边界全局对象、原型方法、构造函数槽位，以及继承而来的属性描述符字段；上述修改后的类型化绑定失败；不设上限的大型中间绑定值；嵌套输出落盘抑制；64 MiB 上限内外的精确计量；日志、值与诊断的组合计量；抛出的超大堆栈；有界失败的输出落盘；不可信对端伪造的流量；以及构建后包的执行。
+编译期测试与快照测试锁定了精确的 `ToolArgsMap`、`ToolOutputMap`、`ToolName`、schema 到 TypeScript 的覆盖范围、特殊名称，以及组装后的 Code Mode 图片转发。注册表与真实 worker 测试覆盖标量、数组、对象和 null 值；字符串原文渲染；缺席的 `undefined`；消费方声明、实际用于拒绝 Promise 的异常类，包括 `ToolCallError`；无效参数与完成值，包括伪装为内建原型的伪造原型；模型代码修改过的 JSON 边界全局对象、原型方法、构造函数槽位，以及继承而来的属性描述符字段；上述修改后的类型化绑定失败；不设上限的大型中间绑定值；嵌套输出落盘抑制；通用含图片上下文延后以及 post-execute 替换／阻止优先级；64 MiB 上限内外的精确计量；日志、值与诊断的组合计量；抛出的超大堆栈；有界失败的输出落盘；不可信对端伪造的流量；以及构建后包的执行。
 
 无密钥的真实 worker 集成测试锁定了自然语言结果无法安全支持的两种句柄工作流。后台 bash 调用返回 task id，外层运行结束，之后的运行再根据该 id 轮询直至任务完成；其他用例分别证明，预先中止不会创建任务、发布后的调用取消会保留任务、前台执行仍与信号耦合，并且取消归 `task_kill` 所有。Cordis 程序会直接读取 active 或 pending 挂载的 id 和 `waitingFor` 字段，按该 id 卸载，并在不解析渲染文本的情况下确认挂载已移除。
 
@@ -93,6 +93,10 @@ Code Mode 通过运行时请求中的 `{ name: "ToolCallError", memberNameProper
 
 **静默检查格式化或截断过大的完成值：**不予采纳。把 JSON 值改成字符串既有损又违反类型。显式的 `output-limit` 失败让模型可以选择返回更小的结果，而保留的日志和诊断仍可使用普通的外层输出落盘机制。
 
+**要求每个丰富叶子工具检查 `exec.parent` 并自行延后。** 不予采用，因为这会把叶子工具与 Code Mode 内部机制耦合、重复策略处理，并遗漏未来丰富工具。分发桥接层负责从已经结算的最终结果通用转发。
+
+**把 Native 丰富内容暴露为每个绑定规范值的一部分。** 不予采用，因为规范值是无损 JSON 且由工具定义；附件块是具有持久生命周期语义的模型投影。保持值与投影分离，既能保留类型化程序，也不会从后续模型上下文中丢弃图片。
+
 ## 后果
 
 Code Mode 程序可以通过稳定值组合工具，无需逆向解析 Native 自然语言。Native 和 Both Mode 保留现有文本与 UI 展示，Code Mode 则获得输出 schema 类型和精确的运行时 JSON。工具作者必须把规范值视为程序化 API，并将仅用于展示的格式化放入渲染器。
@@ -107,6 +111,6 @@ worker 会以嵌套深度有界的扁平协议格式传输数据并执行无损
 - 中间值没有字节上限，可能因值的保留、扁平协议格式副本或结构化克隆开销而耗尽进程或 worker 内存。
 - 64 MiB 硬上限只适用于外层可变负载，不计固定的结果封装语法与展示空白；输出落盘无法恢复超出该上限后被拒绝的字节。
 - 提供方或执行器的采集上限可能在规范值到达 Code Mode 前就已丢弃部分源数据。
-- 不支持的 MCP 输出 schema 会回退为 `JsonValue`；更丰富的 Native 多媒体投影留待后续实现。
+- 不支持的 MCP 输出 schema 会回退为 `JsonValue`；已准入的 MCP 图片使用通用延后投影，而音频和嵌入资源载荷仍只提供诊断。
 - 每个外层 `run_code` 只有一张结果卡片，嵌套调用不会各自生成卡片。
 - Code Mode 失败只暴露 `ToolCallError` 的消息与工具名，不提供程序可用的错误代码联合。

docs/tool-catalog.i18n.yaml

@@ -2,5 +2,5 @@
 # side as of the last confirmed-consistent state. Both languages carry equal authority;
 # after editing either side, bring the other along and re-record with:
 #   pnpm run verify-translation-pairing --write docs/tool-catalog.md
-tool-catalog.md: 898c5700eddfe49083b2ce0e3e04761298b28bbb
+tool-catalog.md: 135020e4fbf41f010e32f21647502d57494bd3c4
-tool-catalog.zh.md: 3b17cf4e3b1b74b0735783cfe899c9c693146c38
+tool-catalog.zh.md: 9d1c7c0fa20e45c1a447b915a2d34adbd3551b38

docs/tool-catalog.md

@@ -116,7 +116,7 @@ ask_user_question pauses the tool call until the active UI provider returns a hu
 
 ### `run_code`
 
-Execute a TypeScript program against the available tools. Write the BODY of an async function (erasable syntax only; top-level `await` and `return` work) and call tools as `await tools.name(args)` per the declarations in the system prompt. Only what you print or return comes back — curate it.
+Execute a TypeScript program against the available tools. Write the BODY of an async function (erasable syntax only; top-level `await` and `return` work) and call tools as `await tools.name(args)` per the declarations in the system prompt. Only what you print or return is program output; image-bearing subtool results are attached after the run.
 
 ```json
 {

docs/tool-catalog.zh.md

@@ -118,7 +118,7 @@ ask_user_question 会暂停工具调用，直到当前 UI 提供方返回人类
 
 ### `run_code`
 
-针对可用工具执行 TypeScript 程序。请编写异步函数的**函数体**（仅使用可擦除语法；支持顶层 `await` 和 `return`），并根据系统提示词中的声明，以 `await tools.name(args)` 形式调用工具。只有打印或返回的内容会传回，请谨慎筛选。
+针对可用工具执行 TypeScript 程序。请编写异步函数的**函数体**（仅使用可擦除语法；支持顶层 `await` 和 `return`），并根据系统提示词中的声明，以 `await tools.name(args)` 形式调用工具。只有打印或返回的值属于程序输出；含图片的子工具结果会在运行结束后附加。
 
 ```json
 {

examples/acp-agent/tests/snapshots/advanced-toolchain/system-prompt.expected.md

@@ -30,7 +30,7 @@ Pass `run_code` the body of an async TypeScript function (erasable syntax only
 - Call tools as `await tools.name(args)` — quoted access for exotic names: `tools["my-tool"](args)`. Every call resolves to the tool's typed canonical JSON value. Tool arguments must be lossless JSON.
 - A FAILED tool call rejects with `ToolCallError`, whose `toolName` identifies the failed tool and whose `message` is human-readable — `try/catch` it to handle and continue.
 - Independent read-only calls MAY overlap under `Promise.all` (safe calls run concurrently; mutating calls run alone, in submission order). Sequence dependent work with `await`.
-- Emit results with `return` and/or `console.log(...)`. ONLY what you print or return comes back to you — intermediate tool results never enter the conversation, so extract just what you need.
+- Emit results with `return` and/or `console.log(...)`. Only what you print or return is program output. A successful tool result containing an image is attached after the run so you can inspect it on the next step; every other intermediate result stays out of the conversation, so extract just what you need.
 
 The available tools:
 

examples/acp-agent/tests/snapshots/advanced-toolchain/tool-schemas.expected.json

@@ -251,7 +251,7 @@
     },
     {
       "name": "run_code",
-      "description": "Execute a TypeScript program against the available tools. Write the BODY of an async function (erasable syntax only; top-level `await` and `return` work) and call tools as `await tools.name(args)` per the declarations in the system prompt. Only what you print or return comes back — curate it.",
+      "description": "Execute a TypeScript program against the available tools. Write the BODY of an async function (erasable syntax only; top-level `await` and `return` work) and call tools as `await tools.name(args)` per the declarations in the system prompt. Only what you print or return is program output; image-bearing subtool results are attached after the run.",
       "parameters": {
         "type": "object",
         "properties": {

examples/acp-agent/tests/snapshots/both-mode-turn/tool-schemas.expected.json

@@ -194,7 +194,7 @@
     },
     {
       "name": "run_code",
-      "description": "Execute a TypeScript program against the available tools. Write the BODY of an async function (erasable syntax only; top-level `await` and `return` work) and call tools as `await tools.name(args)` per the declarations in the system prompt. Only what you print or return comes back — curate it.",
+      "description": "Execute a TypeScript program against the available tools. Write the BODY of an async function (erasable syntax only; top-level `await` and `return` work) and call tools as `await tools.name(args)` per the declarations in the system prompt. Only what you print or return is program output; image-bearing subtool results are attached after the run.",
       "parameters": {
         "type": "object",
         "properties": {

examples/acp-agent/tests/snapshots/code-mode-turn/system-prompt.expected.md

@@ -32,7 +32,7 @@ Pass `run_code` the body of an async TypeScript function (erasable syntax only
 - Call tools as `await tools.name(args)` — quoted access for exotic names: `tools["my-tool"](args)`. Every call resolves to the tool's typed canonical JSON value. Tool arguments must be lossless JSON.
 - A FAILED tool call rejects with `ToolCallError`, whose `toolName` identifies the failed tool and whose `message` is human-readable — `try/catch` it to handle and continue.
 - Independent read-only calls MAY overlap under `Promise.all` (safe calls run concurrently; mutating calls run alone, in submission order). Sequence dependent work with `await`.
-- Emit results with `return` and/or `console.log(...)`. ONLY what you print or return comes back to you — intermediate tool results never enter the conversation, so extract just what you need.
+- Emit results with `return` and/or `console.log(...)`. Only what you print or return is program output. A successful tool result containing an image is attached after the run so you can inspect it on the next step; every other intermediate result stays out of the conversation, so extract just what you need.
 
 The available tools:
 

examples/acp-agent/tests/snapshots/code-mode-turn/tool-schemas.expected.json

@@ -2,7 +2,7 @@
   "initial": [
     {
       "name": "run_code",
-      "description": "Execute a TypeScript program against the available tools. Write the BODY of an async function (erasable syntax only; top-level `await` and `return` work) and call tools as `await tools.name(args)` per the declarations in the system prompt. Only what you print or return comes back — curate it.",
+      "description": "Execute a TypeScript program against the available tools. Write the BODY of an async function (erasable syntax only; top-level `await` and `return` work) and call tools as `await tools.name(args)` per the declarations in the system prompt. Only what you print or return is program output; image-bearing subtool results are attached after the run.",
       "parameters": {
         "type": "object",
         "properties": {

packages/core/tools/README.i18n.yaml

@@ -2,5 +2,5 @@
 # side as of the last confirmed-consistent state. Both languages carry equal authority;
 # after editing either side, bring the other along and re-record with:
 #   pnpm run verify-translation-pairing --write packages/core/tools/README.md
-README.md: 44eb25b79436a75f08406102fc1e3734e59b1001
+README.md: 88fe6660f169f69a70e5630f853104a7c83a5b3c
-README.zh.md: 35142d8186b21b2930ccc40386bed8cc677d77c3
+README.zh.md: b65892caa86c64971bf4483c21d5fc765183c6e6

packages/core/tools/README.md

@@ -115,12 +115,12 @@ Returning `undefined` selects generic fallback. Presenters depend only on their
 
 ### Code Mode
 
-Under `code` or `both`, the registry exposes the reserved `run_code` transport and a deterministic SDK for the current scope, generated in the loaded runtime's language — the registry selects the renderer by `ctx.codeRuntime.language` (`typescript` → the TypeScript SDK below, `python` → the Python SDK). Only the program's outer logs and return value re-enter model context. The SDK declares exact per-tool argument and canonical-output types for every visible tool (`ToolArgsMap`/`ToolOutputMap` in TypeScript, named `TypedDict`s in Python), and each binding resolves to the tool's canonical JSON value. Each lossless-JSON binding call re-enters the complete tool pipeline under the native scheduling contract (concurrency-safe calls may overlap up to `maxParallelSubCalls`; exclusive calls run alone as ordering barriers) with logged correlation to the outer call. Denials and other failed results reject with the real program-visible `ToolCallError` carrying only `toolName` and `message`; Native content and internal error codes stay outside the Code contract. Ordinary side effects are not rolled back, and sub-call `additionalContexts` are deferred through the parent result to preserve call/result adjacency. Run settlement aborts and drains outstanding bindings; runtime failures surface as `CodeRunFailedError`.
+Under `code` or `both`, the registry exposes the reserved `run_code` transport and a deterministic SDK for the current scope, generated in the loaded runtime's language — the registry selects the renderer by `ctx.codeRuntime.language` (`typescript` → the TypeScript SDK below, `python` → the Python SDK). The SDK declares exact per-tool argument and canonical-output types for every visible tool (`ToolArgsMap`/`ToolOutputMap` in TypeScript, named `TypedDict`s in Python), and each binding resolves to the tool's canonical JSON value. Each lossless-JSON binding call re-enters the complete tool pipeline under the native scheduling contract (concurrency-safe calls may overlap up to `maxParallelSubCalls`; exclusive calls run alone as ordering barriers) with logged correlation to the outer call. Denials and other failed results reject with the real program-visible `ToolCallError` carrying only `toolName` and `message`; Native content and internal error codes stay outside the Code contract. The program's outer logs and return value re-enter model context; when a successfully settled sub-call's final Native content contains an image, the bridge also defers that complete ordered content through the parent result so the image is not lost behind the JSON-only binding. Final post-execute blocking or content replacement is authoritative. Ordinary side effects are not rolled back, and sub-call `additionalContexts` are deferred through the parent result to preserve call/result adjacency. Run settlement aborts and drains outstanding bindings; runtime failures surface as `CodeRunFailedError`.
 
 Under `code` — not `both` — the transport is also the only entry the model may use: a model-direct call naming any other visible tool resolves to `UNKNOWN_TOOL` at execution creation, before `tools/pre-execute`, approval `ask`, and guards, so nothing observes or approves a call that can only fail. The denial names the route back (`only \`run_code\` is callable directly — call \`<name>\` from inside a \`run_code\` program instead`), because the same prompt declares that tool and a bare `unknown tool` reads as a broken deployment. SDK sub-dispatches carry the outer execution's `parent` token and are exempt, so programs keep every binding the SDK declared. See the [executor-collapse note](../../../.agents/notes/implemented/bug-fix/2026-08-07-code-mode-executor-collapse.md), the [Code Mode foundation](../../../.agents/notes/implemented/feature/2026-06-15-code-mode.md), [typed-return contract](../../../.agents/notes/implemented/feature/2026-07-20-code-mode-typed-tool-returns.md), and [code-runtime seam](../../code-runtime/README.md). Try `pnpm run demo:code-mode`.
 
 - **The SDK section** (`tools:sdk`, order 150): a lazy prompt section regenerating the language-appropriate SDK text at each assembly. In the TypeScript flavor it emits `JsonValue`, exact `ToolArgsMap` / `ToolOutputMap`, `ToolName`, the `ToolCallError` declaration, and a mapped `tools` namespace for the calling scope's visible end capabilities (exotic names via quoted keys), plus fixed usage instructions; the Python flavor (`ctx.codeRuntime.language === 'python'`) emits the equivalent named `TypedDict`s and a `tools` object with matching usage instructions. Deterministic — lexicographic tool order, byte-identical text for an unchanged tool set (prefix-cache-friendly). Both codegens are exported and never throw during prompt assembly: `jsonSchemaToTs` handles every unified schema construct and degrades unsupported raw constructs to `unknown`; `jsonSchemaToPy` does the same, degrading to `Any` (and a whole object to `dict[str, Any]` when a field name is not a legal `TypedDict` attribute, or whenever it is called outside the SDK render, which supplies the naming context a `TypedDict` declaration needs).
-- **The dispatch bridge** (`run_code`'s execute): every binding call is snapshotted as lossless JSON before dispatch (`undefined`, `BigInt`, cycles, sparse arrays, `-0`, and exotic objects reject that one call), scheduled through a per-run pool that reuses the native concurrency contract — calls start strictly in submission order, consecutive `isConcurrencySafe` calls overlap up to the validated `maxParallelSubCalls` config (default 10; `1` restores serial dispatch), and an exclusive-classified call drains the pool, runs alone, and bars later calls — given the outer execution's opaque token as `parent`, and run through the complete pre-execute → guards → execute → post-execute → result pipeline. A success returns the final canonical value after policy; a failure reaches the worker as one message and becomes `ToolCallError(toolName, message)`. Each started sub-call logs a `tool/code-dispatch-start` event (deterministic id `<parent>:code:<n>`, numbered by submission) at pipeline entry and settles with one `tool/code-dispatch` event carrying the complete model-facing `content`/`isError` outcome (the `tool/result` vocabulary, so UIs render sub-calls through the native path — the pair's `time` fields carry per-sub-call timing); a queued call abandoned by run settlement logs neither. `deriveMessages()` surfaces neither event nor persists the canonical value. Token correlation lets commit-style observers defer an inner success until the final `run_code` result without exposing the live outer execution; ordinary tool side effects are not rolled back. Every sub-call `additionalContexts` entry is deferred through the outer `ToolRunContext` in dispatch order; the loop appends those contexts only after the parent `run_code` result, preserving adjacency and retaining each source/meta even when the program later fails.
+- **The dispatch bridge** (`run_code`'s execute): every binding call is snapshotted as lossless JSON before dispatch (`undefined`, `BigInt`, cycles, sparse arrays, `-0`, and exotic objects reject that one call), scheduled through a per-run pool that reuses the native concurrency contract — calls start strictly in submission order, consecutive `isConcurrencySafe` calls overlap up to the validated `maxParallelSubCalls` config (default 10; `1` restores serial dispatch), and an exclusive-classified call drains the pool, runs alone, and bars later calls — given the outer execution's opaque token as `parent`, and run through the complete pre-execute → guards → execute → post-execute → result pipeline. A success returns the final canonical value after policy; a failure reaches the worker as one message and becomes `ToolCallError(toolName, message)`. Each started sub-call logs a `tool/code-dispatch-start` event (deterministic id `<parent>:code:<n>`, numbered by submission) at pipeline entry and settles with one `tool/code-dispatch` event carrying the complete model-facing `content`/`isError` outcome (the `tool/result` vocabulary, so UIs render sub-calls through the native path — the pair's `time` fields carry per-sub-call timing); a queued call abandoned by run settlement logs neither. `deriveMessages()` surfaces neither event nor persists the canonical value. Token correlation lets commit-style observers defer an inner success until the final `run_code` result without exposing the live outer execution; ordinary tool side effects are not rolled back. Every sub-call `additionalContexts` entry and every successful final content sequence containing an image is deferred through the outer `ToolRunContext` in dispatch order; the loop appends those contexts only after the parent `run_code` result, preserving adjacency and source attribution even when the program later fails.
 - **Settlement discipline**: the bridge owns a run-scoped abort that follows the outer signal in and fires when the run settles for any reason, so a budget expiry aborts an in-flight sub-tool instead of orphaning it; the bridge then drains its queue BEFORE returning, so every `tool/code-dispatch` lands inside the open turn. A failed run throws `CodeRunFailedError` (`code: 'CODE_RUN_FAILED'`, message = the failure kind + captured logs), which the pipeline converts to a structured `isError` the model self-corrects from.
 - **Result size**: intermediate binding values cross the worker process whole and have no per-binding byte cap. `run_code` returns canonical `{ logs: string[], result?: JsonValue }`; strings render raw, every other present JSON root renders through a stack-safe pretty JSON traversal whose total indentation is capped at ten characters (deeper subtrees stay compact), `null` remains explicit, and absent `result` means the program returned `undefined`. The worker's configurable `maxOutputBytes` (default 64 MiB) applies only to the combined serialized outer log-array, completion-value, or failure-message payloads; fixed result-envelope syntax and presentation whitespace are outside that limit. Invalid and over-limit completions fail explicitly, and only this outer result is eligible for ordinary spill.
 
@@ -177,7 +177,7 @@ Prefix-stable while the Code Mode selection, generated SDK, transport schema, an
 
 #### What the model sees
 
-The loop retains model-emitted arguments and the registry's final content. Any thrown or denied call becomes exactly `Error: <message>`. Code Mode returns only the outer program's printed lines and rendered return value, `(run_code completed with no output)` when both are empty, or `Error: code run failed (<kind>): <message>` followed conditionally by `Captured output:` and the captured lines. Inner dispatch events stay log-only; post-execute listeners may append source-attributed context after the result.
+The loop retains model-emitted arguments and the registry's final content. Any thrown or denied call becomes exactly `Error: <message>`. Code Mode renders the outer program's printed lines and return value, `(run_code completed with no output)` when both are empty, or `Error: code run failed (<kind>): <message>` followed conditionally by `Captured output:` and the captured lines. Inner dispatch events stay log-only, while a successful image-bearing sub-result is appended after the outer result as source-attributed context; post-execute listeners may append other source-attributed context at the same boundary.
 
 #### Token effect
 

packages/core/tools/README.zh.md

@@ -115,12 +115,12 @@ ctx.tools.register(defineTool({
 
 ### Code Mode
 
-在 `code` 或 `both` 模式下，注册表为当前作用域公开保留的 `run_code` 传输和按所加载运行时语言生成的确定性 SDK——注册表按 `ctx.codeRuntime.language` 选择渲染器（`typescript` → 下方的 TypeScript SDK，`python` → Python SDK）。只有程序的外层日志与返回值会重新进入模型上下文。SDK 为每个可见工具声明精确的参数与规范输出类型（TypeScript 为 `ToolArgsMap`/`ToolOutputMap`，Python 为具名 `TypedDict`），每个绑定都会解析为该工具的规范 JSON 值。每个无损 JSON 绑定调用都会在原生调度约定下重新进入完整工具流水线（并发安全的调用最多可重叠 `maxParallelSubCalls` 个；独占调用单独运行并构成排序屏障），并在日志中与外层调用建立关联。拒绝及其他失败结果会以程序实际可见的 `ToolCallError` 形式拒绝，且只携带 `toolName` 和 `message`；Native 内容和内部错误码留在 Code 约定之外。普通副作用不会回滚，子调用的 `additionalContexts` 会通过父结果延迟，以保持调用／结果相邻。运行结算会中止并排空尚未完成的绑定；运行时失败以 `CodeRunFailedError` 形式出现。
+在 `code` 或 `both` 模式下，注册表为当前作用域公开保留的 `run_code` 传输和按所加载运行时语言生成的确定性 SDK——注册表按 `ctx.codeRuntime.language` 选择渲染器（`typescript` → 下方的 TypeScript SDK，`python` → Python SDK）。SDK 为每个可见工具声明精确的参数与规范输出类型（TypeScript 为 `ToolArgsMap`/`ToolOutputMap`，Python 为具名 `TypedDict`），每个绑定都会解析为该工具的规范 JSON 值。每个无损 JSON 绑定调用都会在原生调度约定下重新进入完整工具流水线（并发安全的调用最多可重叠 `maxParallelSubCalls` 个；独占调用单独运行并构成排序屏障），并在日志中与外层调用建立关联。拒绝及其他失败结果会以程序实际可见的 `ToolCallError` 形式拒绝，且只携带 `toolName` 和 `message`；Native 内容和内部错误码留在 Code 约定之外。程序的外层日志与返回值会重新进入模型上下文；当成功结算的子调用最终 Native 内容包含图片时，桥接层还会经父结果延后完整有序内容，避免图片被 JSON 专用绑定遮蔽。最终 post-execute 阻止或内容替换具有权威性。普通副作用不会回滚，子调用的 `additionalContexts` 会通过父结果延迟，以保持调用／结果相邻。运行结算会中止并排空尚未完成的绑定；运行时失败以 `CodeRunFailedError` 形式出现。
 
 在 `code`（而非 `both`）下，该传输同时也是模型唯一可用的入口：模型直呼其他任何可见工具名，都会在创建执行时、早于 `tools/pre-execute`、审批 `ask` 和 guards 解析为 `UNKNOWN_TOOL`，因此没有任何一方会观察或批准一个注定失败的调用。拒绝信息会给出正确路径（`only \`run_code\` is callable directly — call \`<name>\` from inside a \`run_code\` program instead`），因为同一份提示词刚刚声明过那个工具，只说 `unknown tool` 会被读成部署损坏。SDK 子分发携带外层执行的 `parent` token，不受此限制，因此程序保留 SDK 声明的全部绑定。参见[执行器塌缩 note](../../../.agents/notes/implemented/bug-fix/2026-08-07-code-mode-executor-collapse.md)、[Code Mode 基础](../../../.agents/notes/implemented/feature/2026-06-15-code-mode.md)、[类型化返回约定](../../../.agents/notes/implemented/feature/2026-07-20-code-mode-typed-tool-returns.md)和[代码运行时 seam](../../code-runtime/README.md)。可以运行 `pnpm run demo:code-mode` 试用。
 
 - **SDK 段**（`tools:sdk`，顺序 150）：一个惰性提示词段，每次组装时都会重新生成与所加载运行时语言相符的 SDK 文本。TypeScript 形态发出 `JsonValue`、精确的 `ToolArgsMap` / `ToolOutputMap`、`ToolName`、`ToolCallError` 声明、面向调用作用域可见最终能力的映射 `tools` 命名空间（特殊名称使用带引号的键），以及固定用法说明；Python 形态（`ctx.codeRuntime.language === 'python'`）发出等价的具名 `TypedDict` 与一个带相同用法说明的 `tools` 对象。其输出具有确定性：工具按字典序排列；工具集合不变时，文本逐字节相同（有利于前缀 cache）。两个代码生成器都已导出，且绝不会在提示词组装期间抛出：`jsonSchemaToTs` 处理统一 schema 的每种构造并将不受支持的原始构造降级为 `unknown`；`jsonSchemaToPy` 同理，降级为 `Any`（当某字段名不是合法的 `TypedDict` 属性时，或在 SDK 渲染之外被调用时——`TypedDict` 声明所需的命名上下文由该渲染提供——整个对象降级为 `dict[str, Any]`）。
-- **分发桥接层**（`run_code` 的 execute）：每个绑定调用都会在分发前快照为无损 JSON（`undefined`、`BigInt`、循环、稀疏数组、`-0` 和特殊对象会使该次调用被拒绝），经由每次运行独有、复用原生并发约定的池调度——调用严格按提交顺序启动，连续的 `isConcurrencySafe` 调用最多可重叠经校验的 `maxParallelSubCalls` 配置个（默认 10；设为 `1` 即恢复串行分发），被分类为独占的调用先排空池、单独运行并阻挡其后的调用——以外层执行的不透明 token 作为 `parent`，并经过完整的 pre-execute → guards → execute → post-execute → result 流水线。成功会返回策略处理后的最终规范值；失败以一条消息到达 worker，并成为 `ToolCallError(toolName, message)`。每个已启动的子调用在进入流水线时记录一条 `tool/code-dispatch-start` 事件（确定性 id `<parent>:code:<n>`，按提交顺序编号），并以一条携带完整模型可见 `content`/`isError` 结果的 `tool/code-dispatch` 事件完结（采用 `tool/result` 词汇，因此 UI 会沿原生路径呈现子调用——这对事件的 `time` 字段承载每个子调用的计时）；因 run 结算而被放弃的排队调用两者都不记录。`deriveMessages()` 既不公开这两个事件，也不持久化规范值。token 关联让以提交为语义的观察器能够把内部成功延迟到最终 `run_code` 结果，而无需公开实时外层执行；普通工具副作用不会回滚。每个子调用的 `additionalContexts` 条目都会按分发顺序通过外层 `ToolRunContext` 延迟；循环只在父级 `run_code` 结果之后追加这些上下文，从而保持相邻关系，并且即使程序后来失败，也会保留各自的来源／元数据。
+- **分发桥接层**（`run_code` 的 execute）：每个绑定调用都会在分发前快照为无损 JSON（`undefined`、`BigInt`、循环、稀疏数组、`-0` 和特殊对象会使该次调用被拒绝），经由每次运行独有、复用原生并发约定的池调度——调用严格按提交顺序启动，连续的 `isConcurrencySafe` 调用最多可重叠经校验的 `maxParallelSubCalls` 配置个（默认 10；设为 `1` 即恢复串行分发），被分类为独占的调用先排空池、单独运行并阻挡其后的调用——以外层执行的不透明 token 作为 `parent`，并经过完整的 pre-execute → guards → execute → post-execute → result 流水线。成功会返回策略处理后的最终规范值；失败以一条消息到达 worker，并成为 `ToolCallError(toolName, message)`。每个已启动的子调用在进入流水线时记录一条 `tool/code-dispatch-start` 事件（确定性 id `<parent>:code:<n>`，按提交顺序编号），并以一条携带完整模型可见 `content`/`isError` 结果的 `tool/code-dispatch` 事件完结（采用 `tool/result` 词汇，因此 UI 会沿原生路径呈现子调用——这对事件的 `time` 字段承载每个子调用的计时）；因 run 结算而被放弃的排队调用两者都不记录。`deriveMessages()` 既不公开这两个事件，也不持久化规范值。token 关联让以提交为语义的观察器能够把内部成功延迟到最终 `run_code` 结果，而无需公开实时外层执行；普通工具副作用不会回滚。每个子调用的 `additionalContexts` 条目以及每份包含图片的成功最终内容序列都会按分发顺序通过外层 `ToolRunContext` 延迟；循环只在父级 `run_code` 结果之后追加这些上下文，从而保持相邻关系和来源归属，即使程序后来失败也不例外。
 - **结算纪律**：桥接层拥有一个运行作用域的中止机制；该中止会跟随传入的外层信号，并在运行因任何原因结算时触发，因此预算耗尽会中止正在运行的子工具，而不会将其遗留。桥接层随后会在返回之前排空队列，使每个 `tool/code-dispatch` 都落在仍打开的轮次内。失败的运行会抛出 `CodeRunFailedError`（`code: 'CODE_RUN_FAILED'`，message = 失败类型 + 已捕获日志），流水线会将其转换为模型可据以自我修正的结构化 `isError`。
 - **结果大小**：中间绑定值会完整传入 worker 进程，且没有逐绑定字节上限。`run_code` 返回规范的 `{ logs: string[], result?: JsonValue }`；字符串原样呈现，其他所有存在的 JSON 根都通过栈安全的美化 JSON 遍历呈现，总缩进最多为 10 个字符（更深的子树保持紧凑），`null` 保持显式，而缺少 `result` 表示程序返回 `undefined`。worker 可配置的 `maxOutputBytes`（默认 64 MiB）只应用于组合序列化后的外层日志数组、完成值或失败消息载荷；固定的结果 envelope 语法和呈现空白不计入该上限。无效和超限的完成会明确失败，只有此外层结果可以使用普通 spill。
 
@@ -177,7 +177,7 @@ The available tools:
 
 #### 模型看到的内容
 
-循环会保留模型发出的参数和注册表的最终内容。任何抛出或被拒绝的调用都会恰好变为 `Error: <message>`。Code Mode 只返回外层程序打印的行和呈现后的返回值；两者都为空时返回 `(run_code completed with no output)`；失败时返回 `Error: code run failed (<kind>): <message>`，并根据是否存在已捕获内容，在其后附加 `Captured output:` 与捕获的行。内部分发事件只保留在日志中；后置执行监听器可以在结果之后追加带来源归属的上下文。
+循环会保留模型发出的参数和注册表的最终内容。任何抛出或被拒绝的调用都会恰好变为 `Error: <message>`。Code Mode 会渲染外层程序打印的行和返回值；两者都为空时返回 `(run_code completed with no output)`；失败时返回 `Error: code run failed (<kind>): <message>`，并根据是否存在已捕获内容，在其后附加 `Captured output:` 与捕获的行。内部分发事件只保留在日志中，而成功且含图片的子结果会在外层结果之后作为带来源归属的上下文追加；后置执行监听器也可以在同一边界追加其他带来源归属的上下文。
 
 #### Token 影响
 

packages/core/tools/src/code-mode.ts

@@ -6,7 +6,7 @@
  * @module @deepseek-ai/dsh-tools/src/code-mode
  */
 
-import { CallId, HarnessError } from '@deepseek-ai/dsh-llm'
+import { CallId, createUserMessage, HarnessError } from '@deepseek-ai/dsh-llm'
 import type { ContentBlock } from '@deepseek-ai/dsh-llm'
 import type { CodeBindingFunction, CodeRunResult, CodeRuntime } from '@deepseek-ai/dsh-code-runtime'
 import { snapshotJsonValue } from '@deepseek-ai/dsh-session'
@@ -48,7 +48,7 @@ const TYPESCRIPT_FLAVOR: RunCodeFlavor = {
     'Execute a TypeScript program against the available tools. Write the BODY of an '
     + 'async function (erasable syntax only; top-level `await` and `return` work) and '
     + 'call tools as `await tools.name(args)` per the declarations in the system prompt. '
-    + 'Only what you print or return comes back — curate it.',
+    + 'Only what you print or return is program output; image-bearing subtool results are attached after the run.',
   codeDescription: 'The program: the body of an async TypeScript function.',
 }
 
@@ -61,8 +61,9 @@ const PYTHON_FLAVOR: RunCodeFlavor = {
   description:
     'Execute a Python program against the available tools. Write the BODY of an '
     + 'async function (top-level `await` and `return` work) and call tools as '
-    + '`await tools.name(args)` per the declarations in the system prompt. Answer '
+    + '`await tools.name(args)` per the declarations in the system prompt. Use '
-    + 'with `print(...)` and/or `return <value>` — only that comes back, so curate it.',
+    + '`print(...)` and/or `return <value>` for program output; image-bearing '
+    + 'subtool results attach after the run.',
   codeDescription: 'The program: the body of an async Python function.',
 }
 
@@ -557,6 +558,12 @@ export function createRunCodeTool(registry: ToolRegistry, options: RunCodeBridge
               const result = parked.kind === 'post-result'
                 ? await scheduler.finalize(parked.exec, parked.result)
                 : scheduler.finish(parked.exec, parked.result)
+              if (!result.isError && result.content.some(block => block.type === 'image')) {
+                exec.deferContext(createUserMessage({
+                  content: result.content,
+                  source: { kind: 'plugin', plugin: 'tools-code-mode' },
+                }))
+              }
               for (const context of result.additionalContexts ?? []) {
                 exec.deferContext(context)
               }

packages/core/tools/src/py-types.ts

@@ -738,7 +738,7 @@ Pass \`run_code\` the body of an async Python function (top-level \`await\` and
 - Call tools as \`await tools.name(args)\` — subscript access for exotic, reserved, or underscore-leading names: \`await tools["my-tool"](args)\`. Every call resolves to the tool's typed canonical JSON value (each method's return type below). Tool arguments must be lossless JSON.
 - A FAILED tool call raises \`ToolCallError\`, whose \`toolName\` identifies the failed tool and whose message is human-readable — wrap in \`try/except\` to handle and continue.
 - Independent read-only calls MAY overlap under \`asyncio.gather\` (safe calls run concurrently; mutating calls run alone, in submission order). Sequence dependent work with \`await\`.
-- Emit the run's answer with \`print(...)\` and/or a top-level \`return <value>\`; the returned value must be lossless JSON. ONLY what you print and the returned value come back — intermediate tool results never enter the conversation, so extract just what you need.
+- Emit the run's answer with \`print(...)\` and/or a top-level \`return <value>\`; the returned value must be lossless JSON. Only what you print and return is program output. A successful tool result containing an image is attached after the run so you can inspect it on the next step; every other intermediate result stays out of the conversation, so extract just what you need.
 
 The available tools:`
 

packages/core/tools/src/ts-types.ts

@@ -254,7 +254,7 @@ Pass \`run_code\` the body of an async TypeScript function (erasable syntax only
 - Call tools as \`await tools.name(args)\` — quoted access for exotic names: \`tools["my-tool"](args)\`. Every call resolves to the tool's typed canonical JSON value. Tool arguments must be lossless JSON.
 - A FAILED tool call rejects with \`ToolCallError\`, whose \`toolName\` identifies the failed tool and whose \`message\` is human-readable — \`try/catch\` it to handle and continue.
 - Independent read-only calls MAY overlap under \`Promise.all\` (safe calls run concurrently; mutating calls run alone, in submission order). Sequence dependent work with \`await\`.
-- Emit results with \`return\` and/or \`console.log(...)\`. ONLY what you print or return comes back to you — intermediate tool results never enter the conversation, so extract just what you need.
+- Emit results with \`return\` and/or \`console.log(...)\`. Only what you print or return is program output. A successful tool result containing an image is attached after the run so you can inspect it on the next step; every other intermediate result stays out of the conversation, so extract just what you need.
 
 The available tools:`
 

packages/core/tools/tests/code-mode.spec.ts

@@ -1082,6 +1082,73 @@ describe('the run_code dispatch bridge', () => {
     ])
   })
 
+  it('defers image-bearing final sub-call content onto the outer run_code result', async () => {
+    const { ctx, runtime } = await setup({ mode: 'code' })
+    ctx.tools.register(defineContentToolFixture({
+      name: 'image_result',
+      description: 'Return one durable image.',
+      parameters: {},
+      execute: () => Promise.resolve([
+        { type: 'text', text: 'image result' },
+        {
+          type: 'image',
+          attachment: {
+            attachmentId: 'sha256:aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa' as never,
+            mediaType: 'image/png', bytes: 1, width: 1, height: 1,
+          },
+        },
+      ]),
+    }))
+    runtime.behavior = async (request) => {
+      await request.bindings[0]!.functions.image_result!({})
+      return { logs: [], value: 'done' }
+    }
+
+    const result = await runCode(ctx, 'program')
+
+    expect(result.additionalContexts).toMatchObject([{
+      role: 'user',
+      source: { kind: 'plugin', plugin: 'tools-code-mode' },
+      content: [
+        { type: 'text', text: 'image result' },
+        { type: 'image', attachment: { mediaType: 'image/png', bytes: 1, width: 1, height: 1 } },
+      ],
+    }])
+  })
+
+  it('does not defer images removed by a nested post-execute decision', async () => {
+    for (const decision of ['block', 'replace'] as const) {
+      const { ctx, runtime } = await setup({ mode: 'code' })
+      ctx.tools.register(defineContentToolFixture({
+        name: 'image_result',
+        description: 'Return one durable image.',
+        parameters: {},
+        execute: () => Promise.resolve([{
+          type: 'image',
+          attachment: {
+            attachmentId: 'sha256:bbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbb' as never,
+            mediaType: 'image/png', bytes: 1, width: 1, height: 1,
+          },
+        }]),
+      }))
+      ctx.on('tools/post-execute', (exec, _result, next): Promise<PostToolDecision> => {
+        if (exec.name !== 'image_result') return next()
+        return Promise.resolve(decision === 'block'
+          ? { kind: 'block', feedback: [{ type: 'text', text: 'blocked' }] }
+          : { kind: 'accept', content: [{ type: 'text', text: 'replaced' }] })
+      })
+      runtime.behavior = async (request) => {
+        await request.bindings[0]!.functions.image_result!({}).catch(() => undefined)
+        return { logs: [], value: 'done' }
+      }
+
+      const result = await runCode(ctx, 'program')
+
+      expect(result.additionalContexts).toBeUndefined()
+      await ctx.fiber.dispose()
+    }
+  })
+
   it('keeps sub-call contexts when run_code fails after the nested dispatch', async () => {
     const { ctx, runtime } = await setup({ mode: 'both' })
     registerEcho(ctx)

packages/fs/tool-fs/src/read-image.ts

@@ -15,7 +15,6 @@ import { basename, extname } from 'node:path'
 import type { Context } from '@deepseek-ai/cordis'
 import { AttachmentError, AttachmentId } from '@deepseek-ai/dsh-attachment'
 import type { ImageAttachmentRef, ImageMediaType } from '@deepseek-ai/dsh-attachment'
-import { createUserMessage } from '@deepseek-ai/dsh-llm'
 import type { ContentBlock } from '@deepseek-ai/dsh-llm'
 import { defineTool } from '@deepseek-ai/dsh-tools'
 import type { GenericCallView, ToolExecution } from '@deepseek-ai/dsh-tools'
@@ -209,12 +208,6 @@ export function applyReadImageTool(ctx: Context): void {
           ...ref.name === undefined ? {} : { name: ref.name },
         },
       }
-      if (exec.parent !== undefined) {
-        exec.deferContext(createUserMessage({
-          content: imageReadContent(value),
-          source: { kind: 'plugin', plugin: 'tool-fs' },
-        }))
-      }
       return value
     },
     // Pure display: a generic card in the read family with a follow-along