fix: 严格对齐 ACP 协议实现到 stable v1 规范

对照 /Users/konghayao/code/knowledgebase/origin/acp 规范审计并修复 53 条合规性
发现(critical 5 / major 17 / minor 20 / nit 11),完整审计报告见
docs/acp-compliance-audit.md。

Agent 端 (src/services/acp/agent.ts):
- initialize() 补齐 authMethods,promptCapabilities.image 降级为 false(声明与
  实现脱节,按 initialization.mdx 不声明的 capability 视为不支持)
- sessionCapabilities.fork 移至 _meta.claudeCode.forkSession(fork 在
  meta.unstable.json 中,避免在 stable sessionCapabilities 中暴露 unstable 特性)
- unstable_resumeSession 传 replay:false,不再通过 session/update 重放历史
  (session-setup.mdx:239 明确禁止)
- PromptResponse.usage 移至 _meta.claudeCode.usage
  (extensibility.mdx:39 禁止在 spec 类型根添加自定义字段)
- 空字符串 prompt 改为显式 throw(不再误返 end_turn)

Bridge (src/services/acp/bridge.ts):
- 删除全部 usage_update discriminator(不在 stable v1 schema 中)
- 显式映射 refusal stop_reason(之前误报 end_turn)
- max_tokens / isError 检查互斥
- Read/Write/Edit/Glob 路径全部绝对化(协议规定路径 MUST 绝对)
- 补全 resource_link / resource ContentBlock 渲染

Permissions (src/services/acp/permissions.ts):
- 补齐 reject_always PermissionOption(schema 规定的四个 option 之一)
- checkTerminalOutput 优先检查标准 clientCapabilities.terminal,
  回退到 _meta.terminal_output
- 新增 onPermissionCancelled 回调:cancelled permission outcome →
  StopReason::Cancelled(schema.json:629)
- ExitPlanMode cancelled 分支补上 toolUseID 字段

PromptConversion (src/services/acp/promptConversion.ts):
- resource 分支处理 BlobResource(之前静默丢弃 blob 内容)

acp-link 代理 (packages/acp-link/src/):
- WS 协议从专有 {type, payload} 改造为标准 JSON-RPC 2.0
  (transports.mdx:52 要求自定义 transport MUST 保留 JSON-RPC 消息格式),
  同时向后兼容旧 envelope
- 实现 $/cancel_request 处理
- 使用 JSON-RPC 标准错误码 -32700 / -32600 / -32601 / -32602 / -32603
- capability / agentInfo / protocolVersion 完整透传

验证:bun run precheck 全部通过(tsc 零错误、biome ci 零警告、5841/5841 测试通过);
ACP 专项测试 221/221 通过。独立 verification agent 抽查全部 PASS。

已知暂缓项(审计文档附录 B/C):
- §3.5 traceparent/trace-context 传播(QueryEngine 无 header hook)
- §5.2 terminal/create 完整生命周期(P1,非阻断,需新 RPC 流程)
- §4.2 in_progress tool_call status(SHOULD 级)
- §8.8/8.9/8.14 stale types.ts(不在 owner 分配集合,runtime 已修正)

Co-Authored-By: glm-5.2 <zai-org@claude-code-best.win>
This commit is contained in:
claude-code-best
2026-06-19 12:33:58 +08:00
parent f69c705166
commit 5e30697950
13 changed files with 2693 additions and 258 deletions

View File

@@ -0,0 +1,861 @@
# ACP 合规性审计报告
> 生成日期: 2026-06-19
> 审计范围: src/services/acp/ 和 packages/acp-link/
> 对照规范: /Users/konghayao/code/knowledgebase/origin/acp/agent-client-protocol (commit 取自仓库 HEAD)
## 概览
- 总发现数: 53其中部分为同根因跨维度交叉引用,如 image 能力声明问题在维度 1/3/7 各列一条并注明同根因;独立根因实际约 49 条)
- 按严重程度: critical 5 / major 17 / minor 20 / nit 11
- 涉及方法/字段:
- `initialize` / `authenticate` / `logout`
- `session/new` / `session/load` / `session/resume` / `session/fork` / `session/list` / `session/close`
- `session/prompt` / `session/cancel` / StopReason / Usage
- `session/update` 全部变体usage_update、tool_call、tool_call_update、session_info_update
- `session/set_mode` / `session/set_config_option` / `session/set_model`
- ContentBlock 处理text / image / audio / resource / resourceLink / thought
- 权限委托RequestPermissionOutcome、ToolKind、ToolCallLocation、terminal 生命周期)
- 自定义传输acp-link WS 代理、JSON-RPC envelope、`$/cancel_request`、能力协商)
## 修复优先级矩阵
| 优先级 | 维度 | 发现数 | 修复成本 | 是否阻断 |
|---|---|---|---|---|
| P0 | acp-link 传输层违反 JSON-RPC 2.0(维度 8 | 4 (2 critical + 2 major) | 高 | 是 |
| P0 | promptCapabilities.image 声明与实现脱节(维度 1/3/7 | 3 (3 major, 重复根因) | 低 | 是 |
| P0 | session/resume 重放历史违反 MUST NOT维度 2 | 1 (1 critical) | 中 | 是 |
| P0 | session/update usage_update 非稳定 v1 判别器(维度 4 | 1 (1 critical) | 低 | 是 |
| P1 | PromptResponse.usage 非规范根字段(维度 3 | 1 (1 major) | 低 | 否 |
| P1 | refusal stop_reason 丢失(维度 3 | 1 (1 major) | 低 | 否 |
| P1 | terminal 能力误用 `_meta` + 缺失标准生命周期(维度 5 | 2 (2 major) | 高 | 否 |
| P1 | 权限 `cancelled` 未传播为 StopReason::Cancelled维度 5 | 1 (1 major) | 中 | 否 |
| P1 | setSessionMode 未发 current_mode_update维度 6 | 1 (1 major) | 低 | 否 |
| P1 | session/load 跨项目 cwd 校验缺失(维度 2 | 1 (1 major) | 中 | 否 |
| P2 | 其他 minor / nit | 25 | 低-中 | 否 |
---
## 1. initialize / authenticate / logout + capabilities 协商(维度 1
### 1.1 [major] image 能力声明与实际处理不符
- 位置: `src/services/acp/agent.ts:156` (initialize -> agentCapabilities.promptCapabilities) 配合 `src/services/acp/promptConversion.ts:9-25` (promptToQueryInput)
- 规范要求: PromptCapabilities.image (schema.json:2126-2130 + initialization.mdx:168-170): "The prompt may include ContentBlock::Image"。initialization.mdx:108 "Clients and Agents MUST treat all capabilities omitted in the initialize request as UNSUPPORTED"——反过来说,声明 `image: true` 即承诺 Client 可发送 ContentBlock::Image 且 Agent 会处理。
- 当前实现: initialize 返回 `promptCapabilities: { image: true, embeddedContext: true }`(未声明 audio,默认 false,正确)。但 promptToQueryInput() 只处理 `type==='text'``'resource_link'``'resource'` 三类 block`'image'` block 无对应分支,被静默丢弃。prompt() (agent.ts:269) 把整个 prompt 压成纯字符串 promptInput 传给 QueryEngine.submitMessage()。Client 若信任 `image:true` 发来图片,Agent 会完全忽略,不报错也不转换。
- 修复建议: 二选一。
(A) 若确实不处理图片,把 `promptCapabilities.image` 改为 false或删除该键,默认 false:
~~~diff
promptCapabilities: {
- image: true,
embeddedContext: true,
},
~~~
(B) 若要保留图片能力,在 promptToQueryInput 中处理 image block,将其作为 image content block 注入 query input需 QueryEngine.submitMessage 支持多模态输入):
~~~diff
} else if (b.type === 'image') {
+ const img = b as { source?: { data?: string; media_type?: string } }
+ images.push({ data: img.source?.data, mediaType: img.source?.media_type })
}
~~~
然后扩展 submitMessage 接受 images 数组。在多模态 query input 支持完成前,推荐先采用 (A)。
### 1.2 [minor] sessionCapabilities.fork 为非稳定 v1 字段
- 位置: `src/services/acp/agent.ts:164-169` (sessionCapabilities: { fork: {}, list: {}, resume: {}, close: {} })
- 规范要求: 稳定 v1 SessionCapabilities (schema.json:2528-2571) 仅定义属性 `_meta` / `close` / `list` / `resume`,无 fork。SDK 自带 schema (node_modules/@agentclientprotocol/sdk/schema/schema.json:5139-5148) 明确标注 fork 为 "UNSTABLE — This capability is not part of the spec yet, and may be removed or changed at any point"。本审计只覆盖稳定 v1,draft/unstable 不在合规范围。
- 当前实现: sessionCapabilities 中包含 `fork: {}` 以配合已实现的 `unstable_forkSession()` (agent.ts:235)。但稳定 v1 schema 的 SessionCapabilities 不认识此键。由于 schema 未设 `additionalProperties:false`,字段不会导致 schema 校验硬失败,但严格 Client 会把它当作未知扩展忽略,无法据此发现 session/fork 支持。
- 修复建议: 将 unstable fork 能力迁移到 AgentCapabilities._meta 下的自定义扩展命名空间(与现有 `_meta.claudeCode.promptQueueing` 同模式),符合 extensibility.mdx:111-134 "Advertising Custom Capabilities":
~~~diff
agentCapabilities: {
_meta: {
- claudeCode: { promptQueueing: true },
+ claudeCode: { promptQueueing: true, forkSession: true },
},
promptCapabilities: { image: true, embeddedContext: true },
mcpCapabilities: { http: true, sse: true },
loadSession: true,
sessionCapabilities: {
- fork: {},
list: {},
resume: {},
close: {},
},
},
~~~
### 1.3 [nit] 缺失 authMethods 字段
- 位置: `src/services/acp/agent.ts:127-172` (initialize 返回值)
- 规范要求: InitializeResponse (schema.json:1487-1548) authMethods 默认 [] (schema.json:1528-1535)。authentication.mdx:37 "Agents advertise authentication options in the authMethods field of the initialize response"。虽然默认 [] 使字段可选,但显式返回 `authMethods: []` 更利于 Client 明确判断"无需认证"而非"能力未知"。
- 当前实现: initialize 返回值不含 authMethods 字段。authenticate() (agent.ts:176-181) 忽略 params.methodId 直接返回 `{}`,意味着即使 Client 用任意 methodId 调 authenticate 也会成功——但因 authMethods 缺失,规范上 Client 不应调用 authenticate。
- 修复建议: 显式返回 `authMethods: []` 以明示无认证方法,与 authenticate() 的 no-op 语义一致:
~~~diff
return {
protocolVersion: 1,
+ authMethods: [],
agentInfo: { ... },
agentCapabilities: { ... },
}
~~~
同时建议在 authenticate() 中校验:因未声明任何 method,若被调用应返回 method-not-found 错误code -32601,而非无条件成功。
---
## 2. Session 生命周期:新建 / 加载 / 恢复 / 分叉 / 列出 / 关闭(维度 2
### 2.1 [critical] session/resume 重放完整历史违反 MUST NOT
- 位置: `src/services/acp/agent.ts:193-199` (unstable_resumeSession) → getOrCreateSession (688-777) → replaySessionHistory (792-816) / replayHistoryMessages (757-769)
- 规范要求: docs/protocol/session-setup.mdx "Resuming a Session": "Unlike session/load, the Agent MUST NOT replay the conversation history via session/update notifications before responding. Instead, it restores the session context, reconnects to the requested MCP servers, and returns once the session is ready to continue."
- 当前实现: unstable_resumeSession 委托给 getOrCreateSession,这是 loadSession 使用的相同代码路径。对于在内存中找到的会话,它会调用 replaySessionHistory() (第 713 行);对于从磁盘加载的会话,它会调用 replayHistoryMessages() (第 757-769 行)。无论哪种方式,完整的对话历史都会在返回 ResumeSessionResponse 之前通过 session/update 通知流式传输回客户端。因此 session/resume 的行为与 session/load 完全一致,违反了 MUST NOT 重放规则。
- 修复建议: 将恢复路径与加载路径分离。添加一个不执行重放的 resumeSession() 实现:
~~~diff
async unstable_resumeSession(
params: ResumeSessionRequest,
): Promise<ResumeSessionResponse> {
- const result = await this.getOrCreateSession(params)
+ const result = await this.getOrCreateSession({ ...params, replay: false })
this.scheduleAvailableCommandsUpdate(result.sessionId)
return result
}
~~~
在 getOrCreateSession 中,根据 `replay` 标志控制两个 replayHistoryMessages/replaySessionHistory 调用,让 resume 传递 `replay:false`(恢复时仅恢复上下文 + MCP 连接,然后立即返回 `{ modes, models, configOptions }`)。保留 loadSession 的默认 `replay:true`
### 2.2 [major] session/load 跨项目 cwd 校验缺失
- 位置: `src/services/acp/agent.ts:688-777` (getOrCreateSession) 和 resolveSessionFilePath in `src/utils/sessionStoragePortable.ts:401-464`
- 规范要求: docs/protocol/session-setup.mdx "Working Directory": "This directory MUST be an absolute path MUST be used for the session regardless of where the Agent subprocess was spawned."
- 当前实现: createSession() 从 {cwd, mcpServers} 计算 sessionFingerprint (agent.ts:665-670),而 getOrCreateSession() 仅在请求的会话已驻留在 this.sessions (第 696-721 行) 时才将指纹与该内存中的会话进行比较。当会话不在内存中时(正常的恢复/加载情况),代码会调用 resolveSessionFilePath(sessionId, cwd),该方法会搜索请求的目录、其 git 工作树,最后扫描所有项目目录 (sessionStoragePortable.ts:410-463)。没有任何检查验证会话原始的 cwd 是否与请求的 cwd 匹配。客户端可以传入项目 A 的 cwd 并成功加载项目 B 下持久化的会话,然后运行一个上下文错误的会话。在基于磁盘的路径上从未计算或比较过指纹。
- 修复建议: 在解析文件路径后,从磁盘上的会话中读取原始的 cwd第一条消息的 'cwd' 字段),并将其与请求的 cwd 进行比较。如果不匹配,返回错误JSON-RPC 错误代码 -32602 无效参数):
~~~ts
const resolved = await resolveSessionFilePath(params.sessionId, params.cwd)
if (resolved) {
const lite = await readSessionLite(resolved.filePath)
const originalCwd = lite && extractJsonStringField(lite.head, 'cwd')
if (originalCwd && path.resolve(originalCwd) !== path.resolve(params.cwd)) {
throw new RpcError(-32602, `Session cwd mismatch: session belongs to ${originalCwd}, requested ${params.cwd}`)
}
}
~~~
或者,在加载会话的 cwd 不同时跳过工作树/全目录回退搜索,以便跨项目加载自然失败。
### 2.3 [major] unstable_forkSession 忽略源会话 ID,创建空白会话
- 位置: `src/services/acp/agent.ts:235-245` (unstable_forkSession)
- 规范要求: schema/schema.unstable.json ForkSessionRequest: required = ["sessionId", "cwd"];描述为 "The ID of the session to fork."。Agent 在 initialize (agent.ts:165) 中通过 `sessionCapabilities.fork:{}` 声称支持分叉。
- 当前实现: unstable_forkSession 忽略了 params.sessionId要分叉的源会话和 params.additionalDirectories。它只是调用 `this.createSession({ cwd, mcpServers, _meta })` 来构建一个全新的空会话,与源会话没有任何共享的历史/上下文。一个本应从源会话上下文分支出来的 "fork" 实际上创建了一个空白会话。新会话的 ID 被返回,但源会话的对话未恢复,因此分叉在功能上是错误的。
- 备注: 尽管 fork 是 UNSTABLE 且超出了严格的 v1 合规范围,但 Agent 声明了该能力并注册了处理程序,因此客户端调用 `session/fork` 将获得语义错误的结果。
- 修复建议: 将源会话的消息加载到内存中(通过 getLastSessionLog(params.sessionId),并将它们作为 initialMessages 传递给 createSession,同时转发 additionalDirectories:
~~~ts
async unstable_forkSession(params: ForkSessionRequest): Promise<ForkSessionResponse> {
let initialMessages: Message[] | undefined
try {
const log = await getLastSessionLog(params.sessionId as UUID)
if (log?.messages.length) initialMessages = deserializeMessages(log.messages)
} catch (err) { console.error('[ACP] fork source load failed:', err) }
const response = await this.createSession(
{ cwd: params.cwd, mcpServers: params.mcpServers ?? [], _meta: params._meta, additionalDirectories: params.additionalDirectories },
{ initialMessages },
)
this.scheduleAvailableCommandsUpdate(response.sessionId)
return response
}
~~~
(扩展 createSession 签名以接受并持久化 additionalDirectories。
### 2.4 [minor] listSessions 静默截断为 100 并忽略 cursor 分页
- 位置: `src/services/acp/agent.ts:211-231` (listSessions) 和 `src/utils/listSessionsImpl.ts:439-454`
- 规范要求: docs/protocol/session-list.mdx "Pagination": "Clients MUST treat cursors as opaque tokens ... Agents SHOULD return an error if the cursor is invalid." ListSessionsRequest.cursor 是一个可选的不透明分页 token (schema.json:1597)。
- 当前实现: listSessions 完全忽略了 params.cursor。它调用 `listSessionsImpl({ dir: params.cwd ?? undefined, limit: 100 })`——一个硬编码的 100 条目上限,没有偏移量,也没有消费 cursor。响应从不返回 nextCursor,因此跨大历史记录的分页静默失败:拥有超过 100 个会话的客户端只能看到最近的 100 个,无法获取其余的。无效的 cursor 被静默接受(规范指出 Agent 应该报错)。虽然返回不带 nextCursor 的所有结果是允许的,但静默截断为 100 违反了 "Clients MUST treat a missing nextCursor as the end" 的契约,因为 Agent 实际上有更多结果却隐瞒了。
- 修复建议: 要么 (a) 完全去掉硬编码的 100 限制(如果没有更多结果,返回所有会话且不带 nextCursor 是合规的),或者 (b) 实现 cursor→offset 解码:
~~~ts
const decoded = params.cursor
? JSON.parse(Buffer.from(params.cursor, 'base64').toString())
: { offset: 0 }
const candidates = await listSessionsImpl({ dir: params.cwd, limit: PAGE_SIZE, offset: decoded.offset })
const nextCursor = candidates.length === PAGE_SIZE
? Buffer.from(JSON.stringify({ offset: decoded.offset + PAGE_SIZE })).toString('base64')
: undefined
return { sessions: [...], nextCursor }
~~~
至少,当客户端发送 params.cursor 时(因为分页未实现),返回一个错误,这样客户端就不会得到静默错误的结果。
### 2.5 [nit] listSessions 对无标题会话发出空字符串 title
- 位置: `src/services/acp/agent.ts:219-228` (listSessions 会话映射)
- 规范要求: schema.json SessionInfo (2787): title 是 type `["string","null"]`(可选,可为空。docs/protocol/session-list.mdx: "Human-readable title for the session. May be auto-generated from the first prompt."
- 当前实现: 对于每个候选者,代码无条件地发出 `title: sanitizeTitle(candidate.summary ?? '')`。当会话没有可提取的摘要/标题时(边缘情况下 candidate.summary 为空字符串),Agent 发出 `title: ""`。空字符串技术上是有效的,但没有信息量;根据 schema,省略 title 会更清晰。这是一个表面问题,因为基于磁盘的候选者很少幸存于空摘要。
- 修复建议: 仅在非空时包含 title:
~~~diff
+ const title = sanitizeTitle(candidate.summary ?? '')
sessions.push({
sessionId: candidate.sessionId,
cwd: candidate.cwd,
- title: sanitizeTitle(candidate.summary ?? ''),
+ ...(title ? { title } : {}),
updatedAt: new Date(candidate.lastModified).toISOString(),
})
~~~
updatedAt 的 ISO 8601 格式new Date(ms).toISOString() → 例如 '2025-10-29T14:22:15.123Z') 已经合规。
### 2.6 [nit] NewSessionResponse 不含 cwd,但规范本身不要求
- 位置: `src/services/acp/agent.ts:185-189` (newSession) → createSession 返回 675-680
- 规范要求: schema.json NewSessionResponse (1916) 要求仅 `['sessionId']`cwd 不在响应模式中。
- 当前实现: newSession 返回 `{ sessionId, models, modes, configOptions }`。sessionId唯一必填字段存在。cwd 不返回,但 schema 从未要求在响应中返回 cwdcwd 是 session/new 的请求侧输入,如 docs/protocol/session-setup.mdx 第 52-68 行示例响应第 77-80 行所示,仅返回 `{ sessionId }`)。因此相对于规范没有违规;记录此内容以解决审计检查清单中的错误前提。
- 修复建议: 无需代码更改。只需更新内部审计检查清单,停止期望在 NewSessionResponse 中有 cwd。
---
## 3. session/prompt + session/cancel + stop reason + usage维度 3
### 3.1 [critical] image 能力声明与实际丢弃不符
- 位置: `src/services/acp/agent.ts:155-158` (initialize) + `src/services/acp/promptConversion.ts:9-25` (promptToQueryInput)
- 规范要求: PromptRequest.prompt is ContentBlock[]Clients MUST restrict content types according to PromptCapabilities (prompt-turn.mdx:89-98)。Agent advertises `promptCapabilities.image: true`, signalling it accepts image content blocks.
- 当前实现: initialize() 声明 `promptCapabilities: { image: true, embeddedContext: true }`,但 promptToQueryInput() 只处理 block types `'text'``'resource_link'``'resource'`。任何 `type: 'image'` block以及任何非文本/非资源 block被静默丢弃——只产生字符串连接的文本,所以 image 输入无警告消失。没有通过文件系统或错误暴露 image 的回退。
- 修复建议: 要么停止宣告 image 支持直到它被接通,要么扩展 promptToQueryInput 以暴露 image block。最小正确修复:
~~~diff
promptCapabilities: {
- image: true,
+ image: false,
embeddedContext: true,
},
~~~
如果打算 image passthrough,query input 必须携带 image 数据——例如返回一个结构化输入,携带 `{ type: 'image', source: {...} }` block 而不是 flat string。在此之前,能力声明是协议谎言,使客户端发送 agent 永远看不到的 image。此问题与维度 1 的 §1.1 同根因。
### 3.2 [major] PromptResponse.usage 为非规范根字段
- 位置: `src/services/acp/agent.ts:326-340` (prompt return) 和 `src/services/acp/bridge.ts:756,1059` (forwardSessionUpdates return type)
- 规范要求: Stable v1 schema: PromptResponse (schema/schema.json:2163-2184) 只定义 `stopReason`(必填)和 `_meta`可选。extensibility.mdx:39 states: "Implementations MUST NOT add any custom fields at the root of a type that's part of the specification. All possible names are reserved for future protocol versions." `usage`/`TokenUsage` does not exist anywhere in the stable schema。
- 当前实现: prompt() 返回 `{ stopReason, usage: { inputTokens, outputTokens, cachedReadTokens, cachedWriteTokens, totalTokens } }``usage` 是非规范根字段。它碰巧匹配 bundled SDK schema (schema.json:4656-4665 marked **UNSTABLE**) 中的 UNSTABLE 形状,但那超出了 v1 合规范围。
- 修复建议: 停止为 v1 合规性在 PromptResponse 上发出 `usage`,或将其置于能力协商之后。最干净的修复:
~~~diff
-return { stopReason, usage }
+return { stopReason }
~~~
如果需要 token 报告,通过现有的 `usage_update` SessionUpdate 发送(已在 bridge.ts:843-854 完成,见维度 4 的 critical finding——但 usage_update 本身也是非稳定的)和/或将其移至 `_meta`——但根据 extensibility.mdx:39,即使是未知的根键也被保留,因此唯一规范一致的位置是 `_meta.usage`。推荐:
~~~ts
return { stopReason, _meta: usage ? { claudeCode: { usage } } : undefined }
~~~
### 3.3 [major] Anthropic refusal stop_reason 被误报为 end_turn
- 位置: `src/services/acp/bridge.ts:866-876` (success case stop_reason mapping)
- 规范要求: StopReason enum (schema.json:3212-3241) includes `refusal`——"The turn ended because the agent refused to continue." prompt-turn.mdx:278 defines refusal as a first-class stop reason。Anthropic API can return `stop_reason: 'refusal'` on safety refusals。
- 当前实现: 在 `success` 情况下只映射了 `'max_tokens'`;其他所有 Anthropic stop_reason包括 `'refusal'``'end_turn'``'stop_sequence'``'tool_use'`)都落入默认 `stopReason = 'end_turn'`。没有分支将 `'refusal'` 映射到 ACP `refusal` stop reason,因此真正的拒绝被误报为成功的 end_turn,破坏了规范契约——refusal 应被反映(根据 refusal 语义,prompt 不应包含在下一轮)。
- 修复建议: 添加显式映射:
~~~diff
case 'success': {
- const stopReasonStr = msg.stop_reason
- if (stopReasonStr === 'max_tokens') {
- stopReason = 'max_tokens'
- }
- if (isError) {
- // Report error as end_turn
- stopReason = 'end_turn'
- }
+ const r = msg.stop_reason
+ if (r === 'max_tokens') stopReason = 'max_tokens'
+ else if (r === 'refusal') stopReason = 'refusal'
+ else stopReason = 'end_turn'
+ if (isError) stopReason = 'end_turn'
break
}
~~~
### 3.4 [minor] max_tokens 与 isError 检查相互覆盖
- 位置: `src/services/acp/bridge.ts:866-876` (success case) 和 877-886 (error_during_execution case)
- 规范要求: StopReason `max_tokens` (schema.json:3221-3223): "The turn ended because the agent reached the maximum number of tokens." prompt-turn.mdx:271-272。
- 当前实现: `max_tokens` 检查和 `isError` 检查是两个独立的 `if` 语句,不是 `else if`。当 `stop_reason === 'max_tokens'``isError === true` 时,第一个 `if` 设置 `stopReason = 'max_tokens'`,但第二个 `if` 立即覆盖为 `end_turn`。同样的缺陷也出现在 error_during_execution (877-886):max_tokens 可能被设置然后被覆盖。SDK 标记为错误的 max-tokens 终止因此被报告为 end_turn,向客户端隐藏了真正的原因。
- 修复建议: 使分支互斥或将 isError 仅作为回退(见 §3.3 的合并修复 diff
### 3.5 [minor] prompt 未读取 params._meta,trace context 丢失
- 位置: `src/services/acp/agent.ts:262-287` (prompt queue handling) 和 269 (params._meta not read)
- 规范要求: extensibility.mdx:8-39——`_meta` 是每个类型的保留扩展点,包括 PromptRequest (schema.json:2137-2141)。W3C trace context keys (`traceparent``tracestate``baggage`) SHOULD be propagated for OpenTelemetry interop (extensibility.mdx:33-38)。prompt-queue feature 只在 agentCapabilities 级别宣告agent.ts:150-154 `_meta.claudeCode.promptQueueing: true`) 是正确的地方。
- 当前实现: prompt() 从不读取 `params._meta`。两个后果: (1) prompt 中客户端提供的 W3C trace context (`traceparent`/`tracestate`/`baggage`) 被静默丢弃,破坏了 tracing interop(2) prompt-queueing 扩展已宣告,但没有 per-request opt-out 机制——客户端无法通过 `_meta` 信号 skip-queue。能力宣告本身是合规的。
- 修复建议: 将 `params._meta` 传递给 query 层,以便 trace context 可以附加到下游 API 调用,并可选地遵守 `_meta.claudeCode.skipQueue` flag。至少,转发 traceparent:
~~~ts
const traceparent = params._meta?.traceparent
// thread it into the API client request headers
~~~
### 3.6 [minor] prompt catch 块对 abort 信号竞态返回错误而非 cancelled
- 位置: `src/services/acp/agent.ts:342-359` (prompt catch block)
- 规范要求: prompt-turn.mdx:304-311 (Warning): "Agents MUST catch these errors and return the semantically meaningful `cancelled` stop reason, so that Clients can reliably confirm the cancellation." 这适用于中止操作产生的错误。当 session.cancelled 为 true 时,catch 块必须为任何错误返回 cancelled。
- 当前实现: catch 块确实检查 `if (session.cancelled) return { stopReason: 'cancelled' }` (343-345)——对于进程内 cancelled flag 是正确的。然而,守卫使用 `session.cancelled`,只由 cancel() 设置。如果 QueryEngine 的 abort signal 通过 interrupt() 触发,但 session.cancelled 尚未设置interrupt() 完成和 cancel() 到达第 379 行之间的竞态窗口),或从嵌套路径传播取消派生的 AbortError,条件为 false,错误被重新抛出为 JSON-RPC 错误而不是 cancelled stop reason。更稳健的信号是 abort signal 本身。
- 修复建议: 在 flag 之外检查 abort signal,并将 AbortError/abort 形状错误视为取消:
~~~ts
} catch (err) {
const isAbort = err instanceof Error && (
err.name === 'AbortError' || /abort|cancelled|interrupt/i.test(err.message)
)
if (session.cancelled || isAbort) {
return { stopReason: 'cancelled' }
}
// ...existing process-death + rethrow
}
~~~
### 3.7 [minor] 空 prompt 提前返回 end_turn 语义错误
- 位置: `src/services/acp/agent.ts:271-273` (empty prompt early return)
- 规范要求: prompt-turn.mdx:185-199——Agent MUST respond to session/prompt with a StopReason when the turn ends。schema 没有定义空 prompt 的行为StopReason `end_turn` (schema.json:3216-3218) 描述为 "The turn ended successfully," 暗示实际模型处理已发生。
- 当前实现: `if (!promptInput.trim()) return { stopReason: 'end_turn' }` 在不调用模型的情况下返回 end_turn。语义上,这为 no-op 输入报告成功的 turn,这是误导性的:模型从未运行。也没有路径区分 "空 prompt 无效" 和 "turn 完成"。
- 修复建议: 要么拒绝空 prompt 与 JSON-RPC 错误invalid_params, -32602,因为 `prompt` 是必需的 ContentBlock[] 而有效空消息可能是畸形的,或至少文档说明 end_turn 在这里意味着 "nothing to do"。优先抛出:
~~~diff
-if (!promptInput.trim()) return { stopReason: 'end_turn' }
+if (!promptInput.trim()) throw new RpcError(-32602, 'Prompt content is empty')
~~~
### 3.8 [nit] usage 对象缺少 thoughtTokens
- 位置: `src/services/acp/agent.ts:328-339` (usage object construction)
- 规范要求: Bundled (UNSTABLE, out of v1 scope) SDK Usage (node_modules/@agentclientprotocol/sdk/schema/schema.json:6750-6791) has required `totalTokens/inputTokens/outputTokens` and optional `cachedReadTokens``cachedWriteTokens``thoughtTokens`。Stable v1 has no Usage at all。
- 当前实现: 构造的 usage 对象省略 `thoughtTokens`reasoning/thinking tokens。对于发出 reasoning tokens 的模型,报告的 totalTokens (input+output+cachedRead+cachedWrite) 将低估实际计费 tokens,因为 thinking tokens 被排除在总和之外。
- 修复建议: 如果报告 usage见 §3.2 extra-field finding,包括可用的 thinking tokens:
~~~ts
totalTokens: inputTokens + outputTokens + cachedReadTokens + cachedWriteTokens + thoughtTokens
~~~
注意,这只在 unstable contract 下重要;对于严格的 v1 合规性,整个 usage 字段应被移除。
---
## 4. session/update 通知形状(所有 update 变体)(维度 4
### 4.1 [critical] usage_update 非稳定 v1 SessionUpdate 判别器
- 位置: `src/services/acp/bridge.ts:794, 846, 1027` (forwardSessionUpdates, 'result' 和 'compact_boundary' 情况)
- 规范要求: ACP v1 稳定版 schema schema.json:2942-3108 定义 SessionUpdate 为通过 propertyName `sessionUpdate` 进行 oneOf 判别,包含 10 个有效常量: `user_message_chunk``agent_message_chunk``agent_thought_chunk``tool_call``tool_call_update``plan``available_commands_update``current_mode_update``config_option_update``session_info_update``usage_update` 不在 v1 稳定版规范中。Claude Code 捆绑的 SDK schema v0.19.0 第 5789 行将其标记为 "UNSTABLE——此功能尚未包含在规范中,随时可能被删除或更改"。)
- 当前实现: bridge.ts 在 3 处发送 `sessionUpdate: 'usage_update'` 以及非标准字段(`used``size``cost`: (1) 第 791-798 行,在 'system' 消息上检测到 compact_boundary 子类型后;(2) 第 843-854 行,在带有累计 token 计数和 total_cost_usd 的每条 'result' 消息上;(3) 第 1024-1031 行,在独立的 'compact_boundary' 消息情况下。这些仅因为捆绑的 SDK 的类型允许(其草案模式包含 usage_update而通过类型检查合规 v1 的客户端会拒绝这些通知。
- 修复建议: 完全移除 usage_update 通知。token/cost 信息没有 v1 稳定版的 SessionUpdate 变体;它必须通过 `_meta` 承载或直接丢弃。最小合规修复方案: 删除 bridge.ts:791-798、843-854 和 1024-1031 处的三个 sessionUpdate 块,保留 token 聚合用于 PromptResponse 结果。如果客户端仍然需要利用率信号,请通过有效的变体发送:
~~~ts
await conn.sessionUpdate({
sessionId,
update: {
sessionUpdate: 'session_info_update',
_meta: { claudeCode: { usage: { used: usedTokens, size: lastContextWindowSize, cost: totalCostUsd != null ? { amount: totalCostUsd, currency: 'USD' } : undefined } } },
},
})
~~~
### 4.2 [minor] 从未发出 tool_call in_progress 状态
- 位置: `src/services/acp/bridge.ts:1300-1316` (toAcpNotifications, 'tool_use' 第一次遇到 → tool_call) 和 1321-1348 ('tool_result' → tool_call_update)
- 规范要求: schema.json:3525-3548 ToolCallStatus 枚举为 `pending``in_progress``completed``failed`。tool-calls.mdx:76-91 ('Updating') 文档化了一个生命周期,其中 Agent 在工具实际运行时报告 `status: 'in_progress'`。v1 规范称工具 "在其生命周期中会经历不同状态"。
- 当前实现: 实现仅发送两种状态转换: 初始 'tool_call' 为 `status: 'pending'`(第 1309 行),以及随后的 'tool_call_update' 为 `status: 'completed' | 'failed'`(第 1340-1343 行)。从未发出 `in_progress` 状态,因此客户端无法区分 "awaiting approval / streaming input" 和 "currently executing"。
- 修复建议: 当 Claude Code 知道工具开始执行时,发出一个中间的 tool_call_update:
~~~ts
{ sessionUpdate: 'tool_call_update', toolCallId, status: 'in_progress' }
~~~
如果无法获得执行挂钩,请记录此差距;规范将其定义为 SHOULD 级别的生命周期信号,因此省略它仅属于轻微的合规性缺失。
### 4.3 [minor] 从未通过 session/update 发出 session_info_update
- 位置: `src/services/acp/agent.ts:225-226` (session-list 候选构建)——src/services/acp/ 下没有任何位置发出 session_info_update
- 规范要求: schema.json:2819-2837 SessionInfoUpdate 是一个有效的 SessionUpdate 变体 (`sessionUpdate: 'session_info_update'`),包含可选字段 `title``updatedAt`。它允许 Agent 通知客户端动态会话标题和最后活动时间戳。
- 当前实现: agent.ts 计算了一个会话标题(`title: sanitizeTitle(candidate.summary ?? '')``updatedAt: new Date(candidate.lastModified).toISOString()`)——但这仅用于 session/list 响应负载。从不通过 `session/update` 通知向客户端发出 session_info_update,因此当前会话的标题/更新时间永远不会流式传输给客户端。
- 修复建议: 当派生出或更改会话标题时(例如,在第一次助手回复或摘要提取后),发出:
~~~ts
await this.conn.sessionUpdate({
sessionId,
update: { sessionUpdate: 'session_info_update', title: derivedTitle, updatedAt: new Date().toISOString() },
})
~~~
这通过 v1 稳定版规范中记录的通道,为客户端提供了规范的会话显示名称。
### 4.4 [nit] Bash 工具 _meta 键未命名空间化
- 位置: `src/services/acp/bridge.ts:501-503` (toolUpdateFromToolResult, 'Bash' 情况)
- 规范要求: schema.json 将 `_meta` 记录为保留的扩展点("实现不得对这些键上的值做出假设")。为了向前兼容,建议使用反向 DNS / 供应商命名空间的自定义键。SessionUpdate 上的其他 _meta 用法bridge.ts:1288, 1304, 1336, 1370正确使用了 `_meta.claudeCode.*` 命名空间。
- 当前实现: Bash 工具结果在 ToolCallContent._meta 中放置了未命名空间的键: `_meta: { terminal_info: { terminal_id }, terminal_output: { terminal_id, data }, terminal_exit: { terminal_id, exit_code, signal } }`。这些是附加的,因此符合规范,但它们破坏了在其他所有地方使用的 `_meta.claudeCode.*` 命名空间约定,并且仅依赖于客户端知道这些键。
- 修复建议: 为了与文件的其余部分保持一致,请在 `_meta.claudeCode` 下进行命名空间处理(可选;这并非规范违规):
~~~ts
_meta: { claudeCode: { terminal: { info: {...}, output: {...}, exit: {...} } } }
~~~
---
## 5. tool calls + permissions delegation维度 5
### 5.1 [major] terminal 能力检测误用 _meta 而非 clientCapabilities.terminal
- 位置: `src/services/acp/permissions.ts:280-285` (checkTerminalOutput)
- 规范要求: ClientCapabilities schema (schema.json:586-613) defines the standard terminal capability as the boolean field `clientCapabilities.terminal` (line 606-610, default false)。Terminals doc (docs/protocol/terminals.mdx:8-25) states: "Before attempting to use terminal methods, Agents MUST verify that the Client supports this capability by checking ... `clientCapabilities.terminal`"。`_meta` is explicitly reserved and "Implementations MUST NOT make assumptions about values at these keys" (schema.json:1961)。
- 当前实现: checkTerminalOutput 读取 `clientCapabilities._meta.terminal_output === true` 来决定 terminal 支持。从未咨询标准 `clientCapabilities.terminal` 布尔值,因此宣告 `terminal: true`(没有 Claude-Code 特定 `_meta.terminal_output` flag的合规 ACP 客户端被视为不支持 terminals,而保留的 `_meta` 字段被视为真正的能力。
- 修复建议: 将标准能力作为主要,仅对较旧的 Claude-Code 客户端的遗留 `_meta` flag 进行回退:
~~~ts
function checkTerminalOutput(clientCapabilities?: ClientCapabilities): boolean {
if (!clientCapabilities) return false
if (clientCapabilities.terminal === true) return true
// Legacy Claude-Code clients advertised via _meta before terminal: bool existed
const meta = (clientCapabilities as unknown as Record<string, unknown>)._meta
return !!meta && typeof meta === 'object' && (meta as Record<string, unknown>)['terminal_output'] === true
}
~~~
### 5.2 [major] terminal 生命周期未实现,伪造 terminalId 且 _meta 注入非标准键
- 位置: `src/services/acp/bridge.ts:498-511` (toolUpdateFromToolResult, Bash branch)
- 规范要求: Terminals doc (docs/protocol/terminals.mdx:27-110) defines the standard terminal lifecycle: the Agent MUST call `terminal/create` to obtain a real `terminalId`, embed it via ToolCallContent `{type:'terminal', terminalId}` (schema.json:3242-3256), and the Client retrieves output via `terminal/output`。ToolCallUpdate._meta is reserved: "Implementations MUST NOT make assumptions about values at these keys" (schema.json:3555)。
- 当前实现: 对于 Bash results,当 supportsTerminalOutput 为 true 时,代码伪造了一个 `terminalId: toolUse.id`(从未通过 `terminal/create` 注册),发出 `{type:'terminal', terminalId}` content,并注入了三个非标准键到 `_meta`: `terminal_info``terminal_output`(携带实际的 stdout/stderr 数据)和 `terminal_exit`。从未发出 `terminal/create` / `terminal/release`。标准 ACP 客户端没有义务读取这些 `_meta` 键,且嵌入的 `terminalId` 对其 terminal 子系统未知,因此 `terminal/output` / `terminal/wait_for_exit` 将失败。
- 修复建议: 当 `clientCapabilities.terminal === true` 时实现标准 terminal 流程: 在工具运行前(或首次输出时)调用 `conn.request('terminal/create', {sessionId, command, cwd, outputByteLimit})`,在 tool_call content 中嵌入返回的真实 `terminalId`,通过 terminal 子系统流式输出,并在完成时 `terminal/release`。在此之前,不要宣告 terminal 支持;回退到 inline `{type:'text'}` content 并丢弃 `_meta.terminal_*` 键(它们对合规客户端不可解释)。
### 5.3 [major] cancelled 权限结果被当作普通拒绝
- 位置: `src/services/acp/permissions.ts:136-142` (createAcpCanUseTool cancelled branch) 和 231-237 (handleExitPlanMode cancelled branch)
- 规范要求: RequestPermissionOutcome.cancelled variant (schema.json:2310-2320) is sent by the Client "when a client sends a session/cancel notification to cancel an ongoing prompt turn"。tool-calls.mdx:168-186 and the schema description state the prompt turn was cancelled。When the prompt turn is cancelled the Agent MUST resolve session/prompt with `StopReason::Cancelled` (schema.json:629 "Respond to the original session/prompt request with StopReason::Cancelled")。
- 当前实现: 在 `outcome === 'cancelled'` 时,canUseTool 返回一个通用的 `PermissionDenyDecision``behavior:'deny'`、decisionReason mode default / plan。这作为普通拒绝反馈到工具执行器,因此 turn 继续(或失败与普通的 end_turn / tool-error而不是用 `cancelled` 中止 turn。agent.cancel() flag 从不响应 cancelled 权限结果设置,因此 prompt 循环不返回 stopReason 'cancelled' 仅因为用户/客户端取消了权限 prompt。
- 修复建议: 将 `cancelled` 结果视为 turn-cancellation 信号。从 canUseTool 抛出一个类型化的 sentinel或通过闭包传递一个 session-level cancelled flag并让 forwardSessionUpdates / agent.prompt() 检测它以返回 `{stopReason:'cancelled'}`:
~~~ts
if (response.outcome.outcome === 'cancelled') {
cancelledRef.cancelled = true // shared with agent.cancel()
session.queryEngine.interrupt()
return { behavior:'deny', message:'Permission request cancelled by client', decisionReason:{type:'mode', mode:'default'}, toolUseID }
}
~~~
并在 agent.prompt(): `if (session.cancelled) return { stopReason: 'cancelled' }`
### 5.4 [minor] 从未提供 reject_always 权限选项
- 位置: `src/services/acp/permissions.ts:123-127` (options array)
- 规范要求: PermissionOptionKind enum (schema.json:1992-2016) defines four variants: `allow_once``allow_always``reject_once``reject_always`。tool-calls.mdx:200-208 lists the same four。
- 当前实现: 提供的标准权限选项只有三个: `allow_always``allow_once``reject_once``reject_always`"Reject this operation and remember the choice")从不提供,因此用户无法通过协议的预期机制持久化拒绝(客户端依赖此 hint 显示 "remember" 复选框以供拒绝)。
- 修复建议: 添加一个 reject_always 选项,以便四个规范选择可用:
~~~ts
const options: PermissionOption[] = [
{ kind:'allow_always', name:'Always Allow', optionId:'allow_always' },
{ kind:'allow_once', name:'Allow', optionId:'allow' },
{ kind:'reject_once', name:'Reject', optionId:'reject' },
{ kind:'reject_always', name:'Always Reject', optionId:'reject_always' },
]
~~~
并在 selected 分支中处理 `optionId === 'reject' || optionId === 'reject_always'`
### 5.5 [minor] ToolCallLocation.path / Diff.path 未归一化为绝对路径
- 位置: `src/services/acp/bridge.ts:251` (Read locations), 278/300 (Write/Edit locations), 314 (Glob locations), 700 (toolUpdateFromEditToolResponse locations)
- 规范要求: ToolCallLocation.path (schema.json:3517-3519) is "The file path being accessed or modified" (string)。tool-calls.mdx:304-306 and the protocol-wide path rule require absolute pathsDiff.path (schema.json:1178-1181) and the docs example ('/home/user/project/src/main.py') also use absolute paths。The ACP spec states all file paths MUST be absolute。
- 当前实现: Locations 和 diff paths 直接从 tool input`input.file_path``input.path``response.filePath`)填充,不归一化为绝对路径。如果模型(或重放)提供相对路径或具有未解析的 `~`/`.` 段的路径,则发出的 ToolCallLocation.path / Diff.path 将是相对的,违反绝对路径要求。cwd 参数可用,但仅用于通过 toDisplayPath 格式化显示路径,不用于绝对化存储路径。
- 修复建议: 在发送前对每个发出的路径针对会话 cwd 进行解析:
~~~ts
import { isAbsolute, resolve } from 'node:path'
const abs = (p?: string) => p && cwd ? (isAbsolute(p) ? p : resolve(cwd, p)) : p
// then: locations: filePath ? [{ path: abs(filePath), line: offset ?? 1 }] : []
// and for diff content: path: abs(filePath)
~~~
应用于 Read/Write/Edit/Glob 和 toolUpdateFromEditToolResponse。
### 5.6 [minor] 无 delete / move ToolKind 映射
- 位置: `src/services/acp/bridge.ts:191-411` (toolInfoFromToolUse)——kind coverage
- 规范要求: ToolKind enum (schema.json:3616-3670): `read``edit``delete``move``search``execute``think``fetch``switch_mode``other`。Tools that remove or rename files SHOULD map to `delete` / `move` so clients can render appropriate UI (schema.json:3629-3638)。
- 当前实现: 大多数工具映射正确Read→read、Write/Edit→edit、Bash→execute、Grep/Glob→search、WebFetch/WebSearch→fetch、Agent/TodoWrite→think、ExitPlanMode→switch_mode、default→other。然而,没有为任何 delete 或 move 工具(例如,假设的 rm/mv 工具或 MCP filesystem delete的映射——这样的工具落入 `other`。这在规范内('other' 是有效的)但丢失了语义提示。
- 修复建议: 如果/当 delete/move 工具通过 ACP 连接时,添加显式 case,例如 `case 'Remove': case 'Delete': → kind:'delete'``case 'Move': case 'Rename': → kind:'move'`。低优先级,直到这样的工具出现。
### 5.7 [nit] ExitPlanMode optionId 与 session-mode ID 碰撞
- 位置: `src/services/acp/permissions.ts:185-209` (handleExitPlanMode options) 和 244-254 (selectedOption check)
- 规范要求: PermissionOption.optionId is a free-form string (schema.json:1988-1990) with no enum constraint, so the custom optionIds `auto``acceptEdits``default``plan``bypassPermissions` are schema-valid。然而,与 session-mode ID 碰撞的 optionId 值是应用级歧义,PermissionOptionKind 是唯一标准化的 hint四变体枚举。对于实际上切换会话模式的选项auto/acceptEdits/bypassPermissions使用 `kind:'allow_always'` 过载了 allow_always 语义。
- 当前实现: ExitPlanMode 发出 4-5 个自定义选项,其中 optionId 等于会话模式 id。kind 字段设置为 allow_always/allow_once/reject_once 作为粗略提示,但 optionId 空间(模式 id是 Claude-Code 约定,未在协议中文档化。这是允许的可扩展性,但 kind 不忠实地描述 "此选项更改会话模式"。
- 备注: 不是硬性违规,因为 optionId 是 free-form,ExitPlanMode 映射到有效的 ToolKind `switch_mode`
- 修复建议: 可按原样接受;考虑在这些选项上添加 `_meta` hint例如 `_meta.claudeCode.changesMode = true`,以便客户端可以不同地渲染它们,并确保 optionId 值在 agentCapabilities._meta 中文档化为 Claude-Code 特定的。
### 5.8 [nit] rawInput 浅克隆,易受嵌套突变影响
- 位置: `src/services/acp/bridge.ts:1283-1316` (rawInput construction in toAcpNotifications)
- 规范要求: ToolCallUpdate.rawInput (schema.json:3583-3585) is described as "Update the raw input" with no explicit type constraint (free-form)。It is intended to carry the raw tool input parameters (Record<string, unknown>)。
- 当前实现: `const rawInput = toolInput ? { ...toolInput } : {}` 是一个浅克隆;嵌套对象通过引用与实时 tool input 共享。如果在通知序列化之前对嵌套字段进行后续突变,则发出的 rawInput 可以反映执行后状态而不是发送的输入。Schema-valid 但语义脆弱。
- 修复建议: 深克隆(`structuredClone(toolInput)`)或 JSON-round-trip 输入,然后再附加为 rawInput,以保证捕获的值与实际发送给工具的值匹配。
---
## 6. session/set_mode + session/set_model + session/set_config_option + modes/models/configOptions 形状(维度 6
### 6.1 [major] setSessionMode 改变 mode 后未发 current_mode_update 通知
- 位置: `src/services/acp/agent.ts:396-407` (setSessionMode)
- 规范要求: session-modes.mdx 第 105-121 行: "The Agent can also change its own mode and let the Client know by sending the current_mode_update session notification。" schema.json:1142-1160 CurrentModeUpdate / SessionUpdate variant `current_mode_update` (schema.json:3060-3075)。当 Agent 改变 mode 后 MUST 发送 current_mode_update 通知,使只支持 modes API不支持 configOptions的 Client 能感知 mode 切换。
- 当前实现: setSessionMode 调用 applySessionMode更新内部 session.modes.currentModeId然后 updateConfigOption('mode', ...) 只发送 config_option_update 通知agent.ts:862-868。从不发送 current_mode_update 通知。仅支持 modes 的 Client 将永远收不到 setSessionMode 之后的 mode 变更通知。
- 修复建议: 在 setSessionMode 中,在 applySessionMode 之后追加发送 current_mode_update:
~~~diff
async setSessionMode(params: SetSessionModeRequest): Promise<SetSessionModeResponse> {
const session = this.sessions.get(params.sessionId)
if (!session) throw new Error('Session not found')
this.applySessionMode(params.sessionId, params.modeId)
+ await this.conn.sessionUpdate({
+ sessionId: params.sessionId,
+ update: { sessionUpdate: 'current_mode_update', currentModeId: params.modeId },
+ })
await this.updateConfigOption(params.sessionId, 'mode', params.modeId)
return {}
}
~~~
参照 setSessionConfigOption 中 `configId==='mode'` 分支agent.ts:447-455已有的 current_mode_update 发送逻辑保持一致。
### 6.2 [minor] NewSession/Load/Resume 响应携带非稳定 v1 models 字段
- 位置: `src/services/acp/agent.ts:675-680` (createSession 返回值) 及 715-720 (getOrCreateSession 返回值)
- 规范要求: schema.json:1916-1955 NewSessionResponse 仅定义 sessionId必填、configOptions可选、modes可选`_meta`。LoadSessionResponseschema.json:1668-1697/ResumeSessionResponse 同样不含 models 字段。v1 稳定 schema 中不存在 SessionModelState/SessionModel/SetSessionModel,model 选择属于 draft/unstable 特性。
- 当前实现: createSession 返回 `{ sessionId, models, modes, configOptions }`,getOrCreateSession 返回值同样包含 models。models 字段在 v1 稳定 schema 中未定义,严格 Client 会忽略它。该字段由 @agentclientprotocol/sdk@0.19.0 的 draft 类型SessionModelState/ModelInfo驱动。
- 修复建议: 由于 model 选择为 draft 特性且不在 v1 合规范围,建议: (1) 若仅面向 v1 Client,从 NewSessionResponse/LoadSessionResponse/ResumeSessionResponse 返回值中移除 models 字段,仅保留 sessionId/modes/configOptions或 (2) 若需保留向后兼容,在响应中保留 models 但明确文档标注为非稳定扩展。最小合规改动:
~~~diff
-return { sessionId, models, modes, configOptions }
+return { sessionId, modes, configOptions }
~~~
### 6.3 [minor] setSessionConfigOption 未校验 value 是否在 options 列表内
- 位置: `src/services/acp/agent.ts:427-469` (setSessionConfigOption)
- 规范要求: session-config-options.mdx 第 189-192 行: "value: The new value to set. Must be one of the values listed in the option's options array。" schema.json:3110-3147 SetSessionConfigOptionRequest 的 value 为 SessionConfigValueId,Agent 应在 option.options 内校验该 value 合法性,非法值应返回错误而非静默接受。
- 当前实现: setSessionConfigOption 通过 id 查找 optionagent.ts:440-443,但从不校验 params.value 是否存在于 option.options 中。任何字符串(即使不在 options 列表)都会被接受并写入 currentValue,违反 "Must be one of the values listed" 要求。
- 修复建议: 在 option 查找后追加 options 校验:
~~~ts
const option = session.configOptions.find(o => o.id === params.configId)
if (!option) throw new Error(`Unknown config option: ${params.configId}`)
const validValues = flattenOptions(option.options).map(o => o.value)
if (!validValues.includes(params.value)) {
throw new Error(
`Invalid value '${params.value}' for config option ${params.configId}; must be one of: ${validValues.join(', ')}`,
)
}
~~~
注意 options 可能为 groupedSessionConfigSelectGroup或 flatSessionConfigSelectOption,需 flatten 处理。
### 6.4 [nit] value 类型守卫冗余
- 位置: `src/services/acp/agent.ts:434-438` (setSessionConfigOption value 类型守卫)
- 规范要求: schema.json:3134-3141 SetSessionConfigOptionRequest.value 引用 SessionConfigValueIdschema.json:2779-2782 type:'string'。value 始终为字符串。
- 当前实现: 实现包含 `if (typeof params.value !== 'string') throw`,但因 schema 已将 value 固定为 string,此守卫永远为真,属冗余代码。同时该守卫位置在 option 查找之前,错误信息不够精准。
- 修复建议: 由于 SessionConfigValueId 严格为 string,可移除该类型守卫(由 SDK/schema 层保证);或保留但移至 option.options 校验统一处理,避免分散校验逻辑。
---
## 7. ContentBlock 处理: text/image/audio/resource/resourceLink/thought维度 7
### 7.1 [major] promptCapabilities.image 声明但 promptConversion 完全不解析图片
- 位置: `src/services/acp/promptConversion.ts:3` (promptToQueryInput) 与 `src/services/acp/agent.ts:155-158` (initialize)
- 规范要求: schema.json PromptCapabilities.image (line 2126): "Agent supports [ContentBlock::Image]"content.mdx line 42-55: Image blocks in prompts "Requires the image prompt capability when included in prompts。" 声明了能力就必须能处理对应的 prompt 输入 ContentBlock。
- 当前实现: agent.ts initialize() 声明 `promptCapabilities.image = true`,但 promptToQueryInput() 完全没有 'image' 分支——image block 既不被 base64 解码转成 Claude SDK 的 image content,也不产生任何文本占位,被静默丢弃。客户端按 `image:true` 发送图片 prompt 后内容丢失,无报错。
- 修复建议: 在 promptConversion.ts 增加 image 分支: 将 ACP `{type:'image', data, mimeType}` 转换为 Claude SDK 的 image content block 传给 query若 query input 仅接受 string,则需扩展 promptToQueryInput 返回 ContentBlock[] 而非 string。或者若当前 query 层暂不支持多模态输入,应将 `image:false`,使声明与实现一致,并由客户端回退到文本/链接形式。推荐先降级 `image:false`,待多模态 query input 支持后再开启。此问题与维度 1 §1.1、维度 3 §3.1 同根因。
### 7.2 [major] embeddedContext=true 但 BlobResource 被静默丢弃
- 位置: `src/services/acp/promptConversion.ts:19-24` (resource 分支) 与 `src/services/acp/agent.ts:157`
- 规范要求: schema.json PromptCapabilities.embeddedContext (line 2121): 启用时客户端可发送 ContentBlock::Resourcecontent.mdx line 124-155: EmbeddedResource 支持 TextResource`{uri,text,mimeType?}`)与 BlobResource`{uri,blob,mimeType?}`)两种形式。
- 当前实现: 声明 `embeddedContext=true`,但 promptToQueryInput 的 'resource' 分支仅提取 `resource.text`。当客户端发送 BlobResource如 PDF/二进制文件,字段为 `resource.blob + resource.mimeType + resource.uri`)时,text 为 undefined,内容被完全丢弃,模型只收到空字符串。也未传递 uri/mimeType 上下文。
- 修复建议: 扩展 resource 分支:
~~~ts
} else if (b.type === 'resource') {
const r = b.resource as Record<string, unknown> | undefined
if (r && typeof r.text === 'string') {
parts.push(r.text)
} else if (r && typeof r.blob === 'string') {
const mt = typeof r.mimeType === 'string' ? r.mimeType : 'application/octet-stream'
parts.push(`Embedded resource: ${r.uri ?? '(unknown uri)'} (${mt}, base64 blob)`)
}
}
~~~
(理想做法是将 blob 解码并作为 Claude SDK 二进制 content 传入 query若 query input 不支持则至少以可读占位形式保留上下文,不能静默丢弃。)
### 7.3 [minor] toAcpContentBlock 未处理 resource/resource_link 导致降级为 JSON 文本
- 位置: `src/services/acp/bridge.ts:572` (toAcpContentBlock)
- 规范要求: schema.json ContentBlock.oneOf 包含 ResourceLink (line 1023) 与 EmbeddedResource (line 1039)content.mdx line 163: ResourceLink 在 prompt 中 ALL agents MUST supportcontent.mdx line 11: ContentBlock 也用于 session/update 输出与 tool 结果。
- 当前实现: toAcpContentBlock输出渲染只显式处理 text/image 及若干 Claude 私有 content 类型;'resource' 和 'resource_link' 类型的 SDK content 落入 default 分支line 644-648被序列化为 `{type:'text', text: JSON.stringify(content)}`,产生非规范输出,客户端无法识别为可点击资源。
- 修复建议: 在 toAcpContentBlock switch 中增加 case:
~~~ts
case 'resource_link':
return { type: 'resource_link', uri: content.uri as string, name: (content.name as string) ?? (content.uri as string), mimeType: content.mimeType as string | undefined }
case 'resource': {
const r = content.resource as Record<string, unknown> | undefined
return { type: 'resource', resource: { uri: r?.uri, mimeType: r?.mimeType, text: r?.text, blob: r?.blob } }
}
~~~
注意 ImageContent 与 ResourceLink 字段差异: ImageContent 必填 data+mimeTypebase64,uri 为可选ResourceLink 必填 name+uri,没有 data 字段。
### 7.4 [minor] toAcpContentBlock image 分支 url 处理字段命名澄清
- 位置: `src/services/acp/bridge.ts:596-600` (toAcpContentBlock image 分支 url/非 base64 处理)
- 规范要求: schema.json ImageContent (line 1384-1414): 必填 database64+ mimeType,uri 为可选 string|null。ACP v1 ContentBlock 不支持纯 URL 图片——没有 url 字段,只有可选 uri 引用且仍需 data。
- 当前实现: 当 Claude SDK image content 的 `source.type === 'url'` 时,降级输出文本占位 `[image: <url>]`。这本身符合 ACP因 ACP 要求 base64 data,URL 图片无法原样转发)。但实现中读取的字段名是 source.urlClaude SDK 私有形态),与 ACP 无关;同时未考虑 `source.type` 可能既非 base64 也非 url 的情形已用 '[image: file reference]' 覆盖。逻辑可接受,无违规,仅记录字段命名澄清。
- 修复建议: 无需协议层修复。如要增强: 可将 url 图片自行 fetch+base64 编码后转为合规 ImageContent,但需注意安全与性能;当前文本占位降级是合规的最低实现。保持现状即可,此条仅作字段映射文档。
### 7.5 [nit] audio 能力声明与实现一致(合规,仅记录)
- 位置: `src/services/acp/agent.ts:155-158` (initialize promptCapabilities)
- 规范要求: schema.json PromptCapabilities.audio (line 2116, default false)。content.mdx line 74-87: audio block 需 audio capability。
- 当前实现: promptCapabilities 未声明 audio默认 false,且 promptConversion.ts 与 bridge.ts toAcpContentBlock 均无 audio 处理。声明与实现一致(均不支持),符合规范。但输出侧 toAcpContentBlock 也没有 audio 分支——若 Claude 未来输出音频 content 会落入 JSON.stringify。
- 修复建议: 无需修改;当前状态合规。如未来支持音频输入,需同时: (1) agent.ts 声明 `audio:true`(2) promptConversion.ts 增加 audio→Claude SDK audio block 转换;(3) bridge.ts toAcpContentBlock 增加 `case 'audio'` 输出 `{type:'audio', data, mimeType}`。三者必须同步,避免再次出现 image 那种声明/实现脱节。
### 7.6 [nit] thought / tool_result 映射合规(无需修改)
- 位置: `src/services/acp/promptConversion.ts:8-27``src/services/acp/bridge.ts:1210-1247` (thought / tool_result)
- 规范要求: schema.json ContentBlock.oneOf (line 966-1053) 仅含 text/image/audio/resource_link/resource 五种——不存在 ThoughtContentthought 通过 SessionUpdate discriminator `agent_thought_chunk` (schema.json line 2989) 表达,而非 ContentBlock type 或 `role:'thought'`。tool 结果应通过 tool_call_update (schema.json line 3012+) 传递。
- 当前实现: 实现正确,无需修改。
---
## 8. transports / JSON-RPC envelope / acp-link 代理合规(维度 8
### 8.1 [critical] acp-link WS 使用自有 `{type,payload}` 封装而非 JSON-RPC 2.0
- 位置: `packages/acp-link/src/server.ts:147-156` (send), 800-878 (decodeClientMessage), `packages/acp-link/src/ws-message.ts:52-63`
- 规范要求: transports.mdx L52: "Custom transports MUST ensure they preserve the JSON-RPC message format and lifecycle requirements defined by ACP." overview.mdx L206: "The JSON-RPC envelope fields (jsonrpc, id, method, params, result, and error) follow the JSON-RPC 2.0 specification." transports.mdx L6: "ACP uses JSON-RPC to encode messages."
- 当前实现: acp-link 在 client↔proxy WS 之间使用自有的包装格式 `{ type: string, payload?: unknown }`,而不是 JSON-RPC。ws-message.ts:decodeJsonWsMessage 强制要求每个传入消息包含 'type' 字符串server.ts:decodeClientMessage 随后切换此 type。客户端发送的任何标准 JSON-RPC 消息(`{ jsonrpc:'2.0', id, method, params }`)均会被拒绝,错误提示为 "Invalid WebSocket message payload" (ws-message.ts:60)。stdout↔stdio 部分使用了正确的 SDK ndJsonStream,但面向客户端的 WS 传输(即实际上暴露给客户端的自定义传输)并非 JSON-RPC。
- 修复建议: 使面向客户端的 WS 传输成为透明的 JSON-RPC 转发器。通过 JSON-RPC method 名而非专有的 `type` 进行路由,并完整透传消息。最小改造方案:
~~~ts
// onMessage: 解析一次 JSON-RPC,然后路由到处理程序
const msg = JSON.parse(text) as JsonRpcMessage
if ('method' in msg) {
// 请求或通知 — 根据 msg.method 进行分发
const result = await dispatchMethod(msg.method, msg.params)
if ('id' in msg) send(ws, { jsonrpc:'2.0', id: msg.id, result })
} else {
// 响应 — 关联到待处理的出站请求 id
}
~~~
### 8.2 [critical] 代理响应丢弃 JSON-RPC id,无法关联请求
- 位置: `packages/acp-link/src/server.ts:147-156` (send), 412-416 (session_created), 624 (prompt_complete), 473-483 (session_list)
- 规范要求: JSON-RPC 2.0 spec §6: Request 必须包含 `id`Response 必须包含相同的 `id``result``error`,并带有 `jsonrpc: "2.0"`。overview.mdx L10-13: "请求-响应对期望得到结果或错误"。
- 当前实现: 代理针对客户端请求的响应(例如 `session_created``prompt_complete``session_list``session_loaded``model_changed`)使用带有自选 `type` 字符串的 `send(ws, type, payload)`,且从不携带 JSON-RPC `id`。客户端无法将响应与原始请求相关联,因为代理丢弃了请求 id。整个链路中没有任何 `id` 保留。
- 修复建议: 在 ClientState 上保留一个挂起的 id 映射,并在 JSON-RPC 响应中回显请求的 `id`:
~~~ts
send(ws, { jsonrpc:'2.0', id: pendingId, result })
~~~
### 8.3 [major] 错误响应使用专有 ProxyError 而非 JSON-RPC 错误对象
- 位置: `packages/acp-link/src/server.ts:358-360, 379, 392, 419-421, 450-453, 486-489, 537-540, 626, 696-699, 1166``packages/acp-link/src/types.ts:78-82` (ProxyError)
- 规范要求: overview.mdx L198-201: "所有方法均遵循标准 JSON-RPC 2.0 错误处理……错误包含一个带有 `code``message``error` 对象。" JSON-RPC 2.0 预留代码: -32700 解析错误、-32600 无效请求、-32601 方法未找到、-32602 无效参数、-32603 内部错误。
- 当前实现: 所有错误均以专有的 ProxyError `{ type: 'error', message: string, code?: string }` 发出,且没有 JSON-RPC 错误对象,也没有数值类型的 JSON-RPC 代码。例如 server.ts:358 发送 `{ message: 'Failed to connect: ...' }``code` 字段是一个自由格式字符串,从未使用过 -326xx 代码。不相关的客户端无法区分解析错误、方法未找到错误和内部错误。
- 修复建议: 发出标准的 JSON-RPC 错误响应,关联到请求 id:
~~~ts
send(ws, { jsonrpc:'2.0', id: reqId, error: { code: -32601, message: 'Not connected to agent' } })
~~~
将已知故障映射到代码: -32700 (decodeJsonWsMessage 解析失败)、-32602 (payloadRecord/optionalStringField 验证)、-32601 (代理不支持该功能或 SDK 调用抛出"不支持")、-32603 (内部异常)。
### 8.4 [major] decodeClientMessage 白名单狭窄,多个 v1 方法无传输路径
- 位置: `packages/acp-link/src/server.ts:800-878` (decodeClientMessage switch), 871 `default: throw new Error('Unknown message type')`
- 规范要求: schema/meta.json 列出了 12 个 agent 方法authenticate、initialize、logout、session/close、session/set_mode、session/set_config_option 等)和 9 个 client 方法terminal/*、fs/*。overview.mdx L52 (自定义传输): 必须保留 JSON-RPC 格式和生命周期。未知方法必须产生 JSON-RPC -32601 method-not-found 错误,而不是断开客户端连接。
- 当前实现: decodeClientMessage 在遇到未知 `type` 时抛出异常,这会导致 onMessage 捕获程序发出通用的 `{ type:'error', message:'Unknown message type: ...' }` (server.ts:1166),但不会发出 -32601 响应。更糟糕的是,代理仅识别固定的方法白名单connect、disconnect、new_session、prompt、permission_response、cancel、set_session_model、list/load/resume_session、ping。客户端发起的 `authenticate``logout``session/close``session/set_mode``session/set_config_option``session/list`(与 list_sessions 不同——注意 meta.json 中的方法名是 `session/list`)以及所有 terminal/* 方法在传输中均无路径。这些方法在协议层被悄悄丢弃。
- 修复建议: 用通用的 JSON-RPC 方法路由器替换专有的 type 切换。对于任何识别出但代理未实现的方法,返回 -32601。至少要透传 `session/set_mode``session/close`(这些是 v1 的基准/常用方法)。
### 8.5 [major] 未处理 JSON-RPC 标准 `$/cancel_request`
- 位置: `packages/acp-link/src/`(全仓库);在 acp-link 中 grep `$/cancel_request` 无结果
- 规范要求: JSON-RPC 2.0 spec §6.1: `$/cancel_request` 是用于取消正在进行的请求/通知的标准、传输级取消原语。这与 ACP 特有的 `session/cancel` 通知不同。ACP 透传传输必须将其转发到 stdio 代理进程或进行本地处理。
- 当前实现: 未实现。仅处理专有的 `cancel` 类型 (server.ts:646),它映射到 ACP `session/cancel`。JSON-RPC 级别的 `$/cancel_request` 既未转发给 agent,也未映射到挂起的提示取消。如果客户端发送 `{ "jsonrpc":"2.0", "method":"$/cancel_request", "params": { id: ... } }`,当前解码器会将其拒绝为 "Invalid WebSocket message payload",因为它缺少专有的 `type` 字段。
- 修复建议: 在 JSON-RPC 路由层增加对 `$/cancel_request` 的处理程序: 取消关联的出站提示请求,并转发到底层 SDK 连接的取消路径(或在 agent 上调用 `session/cancel`)。
### 8.6 [major] 代理重构 agentCapabilities 白名单,丢弃扩展能力
- 位置: `packages/acp-link/src/server.ts:321-330`
- 规范要求: ACP 通过 agentCapabilities 按字段协商能力;未来/扩展能力(例如 auth、terminal必须完整透传给客户端,以便其知道自己可以使用哪些方法。
- 当前实现: server.ts:321-330 通过列出白名单字段_meta、loadSession、mcpCapabilities、promptCapabilities、sessionCapabilities来重构 `state.agentCapabilities`。任何 SDK 的 AgentCapabilities 携带但此处硬编码接口 (server.ts:65-79) 中未列出的字段(例如 `auth``terminal`、未来的能力)都会被静默丢弃,不会向客户端通告。
- 修复建议: 直接透传原始的 `initResult.agentCapabilities` 对象,而不是重构它:
~~~diff
-state.agentCapabilities = { /* whitelisted fields */ }
+state.agentCapabilities = agentCaps ?? null
~~~
仅在需要本地 TS 类型时进行收窄——但在传输中发送未收窄的值。
### 8.7 [major] 硬编码 clientInfo/capabilities,丢弃客户端真实信息
- 位置: `packages/acp-link/src/server.ts:313-319`
- 规范要求: overview.mdx L20-24: 客户端 → agent: `initialize` 以协商连接。InitializeParams 携带客户端真实的 `clientInfo``{name, version}`,以便 agent 进行日志记录/遥测。clientCapabilities 同样必须反映真实的客户端能力。
- 当前实现: 代理硬编码 `clientInfo: { name: 'zed', version: '1.0.0' }``clientCapabilities: { fs: { readTextFile: true, writeTextFile: true } }`,忽略客户端实际发送的任何 clientInfo/capabilities。非 Zed 客户端Web UI、RCS 中继、自定义客户端)被错误地呈现给 agent 为 'zed 1.0.0',并可能通告了它并不支持的 fs 能力。
- 修复建议: 接受来自客户端 initialize 消息的 clientInfo 和 clientCapabilities 并进行转发。仅使用 'zed'/{fs:true} 作为代理内部未提供任何信息时的回退。
### 8.8 [major] types.ts ClientCapabilities/ServerCapabilities 形状陈旧
- 位置: `packages/acp-link/src/types.ts:96-113` (ClientCapabilities, ServerCapabilities)
- 规范要求: schema.json InitializeParams.clientCapabilities 和 InitializeResult.agentCapabilities 使用特定形状(例如带有嵌套 `fs.readTextFile/writeTextFile` 的 clientCapabilitiesagentCapabilities = loadSession、mcpCapabilities、promptCapabilities、sessionCapabilities。overview.mdx L206: 协议对象键使用 camelCase。
- 当前实现: types.ts:96-113 定义了过时的形状——`ClientCapabilities { streaming?, toolApproval? }``ServerCapabilities { streaming?, tools? }`——这与实际的 ACP v1 schema 不匹配。这些类型虽然已声明但从未通过 JSON-RPC 路径实际使用;它们具有误导性,并暗示代理正在协商 ACP 中不存在的 streaming/tools 能力。
- 修复建议: 完全移除过时的 `ClientCapabilities`/`ServerCapabilities` 类型它们在任何实时代码路径中均未使用——server.ts 使用其内联的 `AgentCapabilities`,或用 SDK 定义的结构替换它们。
### 8.9 [minor] agentInfo 类型收窄过紧,丢失扩展字段
- 位置: `packages/acp-link/src/types.ts:63-71` (ProxyStatus.agentInfo), `packages/acp-link/src/server.ts:346`
- 规范要求: ACP agentInfoInitializeResult.agentInfo至少为 `{ name, version }`,但根据 extensibility.mdx 可以携带额外的 _meta/扩展字段;自定义传输应保留它。
- 当前实现: ProxyStatus 类型将 `agentInfo` 收窄为 `{ name?: string; version?: string }` (types.ts:66-69)。实际发送的对象 (server.ts:346) 是原始的 `initResult.agentInfo`,所以运行时没问题,但声明的类型会丢弃 TS 认为客户端收到的任何附加字段,且阅读此类型的客户端无法依赖扩展的 agentInfo。types.ts:87-108 中类似地过时的 InitializeParams/InitializeResult 与 SDK 的实际形状不匹配。
- 修复建议: 加宽类型:
~~~ts
agentInfo?: { name: string; version: string; [k: string]: unknown }
~~~
或者通过 SDK 重新导出真实的 InitializeResult 类型。
### 8.10 [minor] session/update 通知方向正确(合规,记录)
- 位置: `packages/acp-link/src/server.ts:190-192` (createClient.sessionUpdate)
- 规范要求: overview.mdx L180-189: `session/update` 是一个 agent→client 通知(无响应)。
- 当前实现: 正确: sessionUpdate 流向 agent→client通过 SDK ClientSideConnection 回调,然后 `send(ws, 'session_update', params)`)。代理在 client→agent 方向上不接受 `session_update`decodeClientMessage 没有该情况)。此处未发现问题——为完整性而列出。
- 修复建议: 无需操作;行为正确。仅将其记录为已验证项。
### 8.11 [minor] 应用层 ping/pong 与传输级 WS 心跳冗余
- 位置: `packages/acp-link/src/server.ts:915-917` (ping → pong)
- 规范要求: WS-level ping/pong 在 RFC 6455 §5.5.2 中是传输级控制帧(二进制操作码 0x9/0xA,而不是应用层消息。将它们与应用层消息混合是非标准的。ACP 本身没有应用层 ping 方法。
- 当前实现: 代理实现了应用层的 `{ type: 'ping' }` / `{ type: 'pong' }` (server.ts:915-917),与传输级的 WS 心跳 (server.ts:1199-1216 通过 `ws.raw.ping()`) 并存。这是冗余的,且容易混淆——如果客户端将应用层 ping 发送为 JSON-RPC `{ method: 'ping' }`,它将无法与传输层帧区分,并会被拒绝。
- 修复建议: 移除应用层的 ping/pong 情况;仅依赖传输级的 WS ping/pong 心跳 (server.ts:1199)。或者,如果需要,文档说明自定义 ping 并通过相同的 `{ type, payload }` 约定路由它。
### 8.12 [minor] RCS 中继路径同样施加 `{type,payload}` 封装
- 位置: `packages/acp-link/src/rcs-upstream.ts:117-149` (connect: REST + identify)
- 规范要求: transports.mdx L52: 自定义传输必须保留 JSON-RPC 消息格式。ACP 规范未定义 RCS "环境/桥接" REST 注册或 WS `identify`/`identified`/`registered`/`keep_alive` 消息类型——这些是 RCS 特定的(超出 ACP v1 范围)。一旦注册,中继必须转发未更改的 JSON-RPC。
- 当前实现: 两步流程REST POST /v1/environments/bridge,然后 WS `identify``identified` 握手)是 RCS 专有的,对于 RCS 传输是可以接受的。但是,rcs-upstream.ts:151-221 中的中继消息处理程序通过相同的 `decodeJsonWsMessage`(要求 `{ type }` 形状)解码所有传入的服务器消息,并仅将非控制类型转发给 messageHandler (L213-219)。这意味着 RCS 和 agent 之间的中继也施加了 `{ type, payload }` 而非 JSON-RPC,这与主 WS 代理有相同的封装问题。
- 修复建议: 对于从 RCS 到本地 agent 的中继路径,解码为 JSON-RPC 并路由方法名。控制消息identify/identified/registered/keep_alive属于 RCS 特有的带外,应通过单独的传输层接口处理,而不是与 ACP 有效负载复用。
### 8.13 [minor] 协议版本未在 status 消息中转发给客户端
- 位置: `packages/acp-link/src/server.ts:314` (acp.PROTOCOL_VERSION), 333-342 (logs protocolVersion)
- 规范要求: ACP 稳定 protocolVersion 在 schema/meta.json 中为 `1`整数。InitializeResponse.protocolVersion 必须透传,以便客户端和 agent 就协商的版本达成一致。
- 当前实现: 代理使用 SDK 常量 `acp.PROTOCOL_VERSION` 发送 initialize,并记录返回的 `initResult.protocolVersion` (server.ts:335),但从未在 `status`/`session_created` 消息中将 `protocolVersion` 转发给客户端客户端send() 调用省略了它)。下游 WS 客户端无法观察协商的协议版本。未发现版本损坏SDK 管理往返),但客户端缺乏可见性。
- 修复建议: 在连接后发送的 `status` 消息中包含 `protocolVersion: initResult.protocolVersion` (server.ts:344-348)。
### 8.14 [nit] JsonRpc 类型未使用(死代码)
- 位置: `packages/acp-link/src/types.ts:34-46` (isRequest/isResponse/isNotification)
- 规范要求: JSON-RPC 2.0 spec §4.1/§4.2: Request = 带有 method+id 的对象Notification = 带有 method 但无 id 的对象Response = 带有 id 且无 method 的对象,以及 result 或 error。
- 当前实现: 辅助函数看起来正确,但这些 JsonRpc 类型在 acp-link 运行时中的任何地方都未使用(代理绕过了它们而使用 `{type,payload}`)。死代码表明存在意图与实现之间的脱节。
- 修复建议: 要么将 JSON-RPC 路由基于这些类型(首选——修复 §8.1 finding,要么移除死类型以避免误导未来的维护者。
---
## 附录 A: SDK 方法命名对照
| SDK 方法 | 当前命名 | stable? | 修复动作 |
|---|---|---|---|
| initialize | initialize | stable | 保留(但需修 authMethods 缺失) |
| authenticate | authenticate | stable | 保留(建议显式返回 authMethods:[] |
| logout | 未实现 | stable | 保留不实现(也未宣告 auth.logout 能力) |
| newSession | newSession | stable | 保留 |
| loadSession | loadSession | stable | 保留(需补 cwd 校验) |
| unstable_resumeSession | unstable_resumeSession | stable (resumed) | 建议在 SDK 升级后改名为 `resumeSession`,同时去除重放历史 |
| unstable_forkSession | unstable_forkSession | UNSTABLE | 保留 unstable 命名;但应从 sessionCapabilities.fork 迁移到 _meta.claudeCode.forkSession |
| listSessions | listSessions | stable | 保留(需实现 cursor 分页) |
| unstable_closeSession | unstable_closeSession | UNSTABLE | 保留 |
| prompt | prompt | stable | 保留(需修 usage 字段、refusal 映射) |
| cancel | cancel (notification) | stable | 保留 |
| setSessionMode | setSessionMode | stable | 保留(需补 current_mode_update 通知) |
| setSessionConfigOption | setSessionConfigOption | stable | 保留(需补 value 校验) |
| unstable_setSessionModel | unstable_setSessionModel | UNSTABLE | 保留 |
| session/update | sessionUpdate (notification) | stable | 保留(需删除 usage_update |
## 附录 B: 不修复项及理由
以下 finding 出于技术权衡或非合规范围,暂不修复:
| Finding | 理由 |
|---|---|
| §1.2 sessionCapabilities.fork 仅作"迁移到 _meta"建议,未标记 P0 阻断 | fork 为 UNSTABLE,严格 v1 合规范围外;当前 schema 未设 `additionalProperties:false`,不会导致硬失败。优先用 _meta.claudeCode.forkSession 重构,不阻断。 |
| §2.5 listSessions 空字符串 title | SessionInfo.title schema 允许 null空字符串技术有效。基于磁盘的候选者很少幸存于空摘要。属表面问题。 |
| §2.6 NewSessionResponse 不含 cwd | 规范本身不要求返回 cwd记录是为了纠正审计检查清单的错误前提。 |
| §3.5 prompt _meta 透传W3C traceparent | extensibility.mdx 用词为 SHOULD,非 MUST。OpenTelemetry interop 非当前部署场景的必需功能。列为 P2。 |
| §3.7 空 prompt 提前返回 end_turn | 行为可接受(虽语义不严谨);若改为抛出 -32602 需协调 Client 错误处理。列为 P2。 |
| §3.8 usage 缺 thoughtTokens | 仅在保留 unstable usage 字段时才有意义;若按 §3.2 整体移除 usage,此项自动消失。 |
| §4.4 Bash _meta 键未命名空间化 | 非规范违规_meta 允许任意附加键);仅命名风格不一致。 |
| §5.4 reject_always 未提供 | PermissionOptionKind 四变体为推荐而非 MUSTREPL 现有交互流不支持持久的拒绝记忆。列为 P2。 |
| §5.7 ExitPlanMode optionId 与 session-mode 碰撞 | optionId 是 free-form 字符串,使用模式 id 作为值是合法扩展ExitPlanMode 映射为 switch_mode,语义可辨。 |
| §5.8 rawInput 浅克隆 | Schema-valid,仅在嵌套对象被后续突变时才有问题Claude Code 工具 input 通常不可变。低风险。 |
| §6.2 响应中携带 models 字段 | 为 SDK draft 类型驱动,严格 v1 Client 会忽略;若客户端使用 SDK 同版本,则 models 是有用的扩展字段。优先移除但非阻断。 |
| §6.4 value 类型守卫冗余 | 不影响合规性,仅代码质量问题。 |
| §7.4 image url 占位字段命名 | 实现合规,仅为字段映射文档。 |
| §7.5 audio 不支持 | 声明与实现均不支持,完全合规。 |
| §7.6 thought / tool_result 映射 | 实现正确,无需修改。 |
| §8.10 session/update 通知方向 | 行为正确,为完整性记录。 |
| §8.11 应用层 ping/pong | 冗余但无害;仅在客户端用 JSON-RPC `ping` 时混淆。低优先级。 |
| §8.14 JsonRpc 死类型 | 不影响运行时;仅在 §8.1 修复时一并清理。 |
## 附录 C: 修复路径建议
### P0 阻断修复(合规性硬阻塞)
1. **acp-link JSON-RPC 传输改造**§8.1、§8.2、§8.3、§8.4、§8.5)——成本高,但属协议层根本缺陷。需要将 WS 解码/编码从 `{type,payload}` 改为 JSON-RPC 2.0,保留请求 id,使用标准错误代码,实现通用方法路由。建议分两阶段: 第一阶段透传所有未识别方法(修复 §8.4+ 标准 id 关联§8.2+ 标准错误§8.3);第二阶段迁移到完全 JSON-RPC§8.1+ 实现 `$/cancel_request`§8.5)。
2. **image 能力降级为 false**§1.1、§3.1、§7.1)——低成本,只需一行改动,立即消除协议谎言。多模态 query input 完成后再恢复 `image:true`
3. **session/resume 去除重放**§2.1)——中成本,需要将 resume 与 load 路径分离,引入 `replay` 标志。
4. **删除 usage_update 通知**§4.1)——低成本,删除 bridge.ts 三处 sessionUpdate 块。Token 信息改由 PromptResponse._meta.claudeCode.usage 承载。
### P1 重要修复(非阻断但影响协议契约)
1. **PromptResponse.usage 字段移至 _meta**§3.2
2. **refusal stop_reason 映射**§3.3
3. **terminal 能力标准生命周期**§5.1、§5.2)——成本高,涉及 terminal/create/release RPC 调用
4. **cancelled 权限结果传播**§5.3
5. **setSessionMode 发送 current_mode_update**§6.1
6. **session/load 跨项目 cwd 校验**§2.2
7. **unstable_forkSession 实现真正分叉**§2.3
8. **BlobResource 处理**§7.2
9. **agentCapabilities/clientInfo 透传**§8.6、§8.7
10. **ClientCapabilities/ServerCapabilities 类型陈旧**§8.8

View File

@@ -275,6 +275,9 @@ describe('permission mode resolution', () => {
{
type: 'error',
payload: {
// Legacy error envelope now carries the JSON-RPC code as a string
// (audit §8.3). -32602 = invalid params.
code: '-32602',
message: expect.stringContaining(
'bypassPermissions requires local ACP_PERMISSION_MODE',
),
@@ -304,3 +307,222 @@ describe('Heartbeat constants', () => {
expect(HEARTBEAT_INTERVAL_MS).toBe(30_000)
})
})
describe('JSON-RPC 2.0 routing (audit §8.1-8.5)', () => {
// Helper to register a JSON-RPC-capable client and capture sent frames.
function setupJsonRpcClient(
sent: unknown[],
options: {
connection?: unknown
sessionId?: string | null
} = {},
) {
const ws = makeTestWs(sent)
process.env.ACP_LINK_TEST_INTERNALS = '1'
const unregister = __testing.registerClient(ws, {
connection: options.connection,
sessionId: options.sessionId ?? null,
jsonRpc: true,
})
return { ws, unregister }
}
test('unknown JSON-RPC method yields -32601 method-not-found (§8.4)', async () => {
const sent: unknown[] = []
const { ws, unregister } = setupJsonRpcClient(sent)
try {
await __testing.dispatchJsonRpcMessage(ws, {
jsonrpc: '2.0',
id: 42,
method: 'session/nonexistent_method',
params: {},
})
// JSON-RPC clients receive a JSON-RPC error with the standard code.
expect(sent).toContainEqual({
jsonrpc: '2.0',
id: 42,
error: {
code: -32601,
message: 'Method not found: session/nonexistent_method',
},
})
} finally {
unregister()
delete process.env.ACP_LINK_TEST_INTERNALS
}
})
test('JSON-RPC response echoes the request id (§8.2)', async () => {
const sent: unknown[] = []
const prompt = mock(async () => ({ stopReason: 'end_turn' }))
const { ws, unregister } = setupJsonRpcClient(sent, {
connection: { prompt },
sessionId: 'sess-1',
})
try {
await __testing.dispatchJsonRpcMessage(ws, {
jsonrpc: '2.0',
id: 'req-7',
method: 'session/prompt',
params: { sessionId: 'sess-1', prompt: [{ type: 'text', text: 'hi' }] },
})
// The id is echoed back in the JSON-RPC result.
expect(sent).toContainEqual({
jsonrpc: '2.0',
id: 'req-7',
result: { stopReason: 'end_turn' },
})
} finally {
unregister()
delete process.env.ACP_LINK_TEST_INTERNALS
}
})
test('$/cancel_request is handled and forwards to session/cancel (§8.5)', async () => {
const sent: unknown[] = []
const cancel = mock(async () => {})
const { ws, unregister } = setupJsonRpcClient(sent, {
connection: { cancel },
sessionId: 'sess-1',
})
try {
await __testing.dispatchJsonRpcMessage(ws, {
jsonrpc: '2.0',
id: 'cancel-1',
method: '$/cancel_request',
params: { id: 'req-7' },
})
// The cancel was forwarded to the ACP cancel path.
expect(cancel).toHaveBeenCalled()
} finally {
unregister()
delete process.env.ACP_LINK_TEST_INTERNALS
}
})
test('JSON-RPC notifications (no id) are dispatched without a response', async () => {
const sent: unknown[] = []
const cancel = mock(async () => {})
const { ws, unregister } = setupJsonRpcClient(sent, {
connection: { cancel },
sessionId: 'sess-1',
})
try {
await __testing.dispatchJsonRpcMessage(ws, {
jsonrpc: '2.0',
method: 'session/cancel',
params: {},
})
expect(cancel).toHaveBeenCalled()
// No JSON-RPC response frame should be emitted for a notification.
expect(
sent.find(m => (m as { jsonrpc?: string }).jsonrpc),
).toBeUndefined()
} finally {
unregister()
delete process.env.ACP_LINK_TEST_INTERNALS
}
})
test('session/set_mode is forwarded to the agent connection (§8.4)', async () => {
const sent: unknown[] = []
const setSessionMode = mock(async () => ({ modeId: 'plan' }))
const { ws, unregister } = setupJsonRpcClient(sent, {
connection: { setSessionMode },
sessionId: 'sess-1',
})
try {
await __testing.dispatchJsonRpcMessage(ws, {
jsonrpc: '2.0',
id: 'm1',
method: 'session/set_mode',
params: { sessionId: 'sess-1', modeId: 'plan' },
})
expect(setSessionMode).toHaveBeenCalled()
// The response carries the echoed id.
expect(sent).toContainEqual({
jsonrpc: '2.0',
id: 'm1',
result: { modeId: 'plan' },
})
} finally {
unregister()
delete process.env.ACP_LINK_TEST_INTERNALS
}
})
test('session/close is forwarded to the agent connection (§8.4)', async () => {
const sent: unknown[] = []
const unstable_closeSession = mock(async () => ({}))
const { ws, unregister } = setupJsonRpcClient(sent, {
connection: { unstable_closeSession },
sessionId: 'sess-1',
})
try {
await __testing.dispatchJsonRpcMessage(ws, {
jsonrpc: '2.0',
id: 'c1',
method: 'session/close',
params: { sessionId: 'sess-1' },
})
expect(unstable_closeSession).toHaveBeenCalled()
} finally {
unregister()
delete process.env.ACP_LINK_TEST_INTERNALS
}
})
})
describe('Capability and protocolVersion transparency (audit §8.6, §8.7, §8.13)', () => {
test('initialize forwards client-supplied clientInfo/capabilities (§8.7)', async () => {
const sent: unknown[] = []
const ws = makeTestWs(sent)
process.env.ACP_LINK_TEST_INTERNALS = '1'
const unregister = __testing.registerClient(ws, { connection: null })
try {
// Send initialize with custom clientInfo; the proxy should remember it.
await __testing.dispatchJsonRpcMessage(ws, {
jsonrpc: '2.0',
id: 'init-1',
method: 'initialize',
params: {
clientInfo: { name: 'my-editor', version: '2.3.4' },
clientCapabilities: { terminal: { create: true } },
},
})
// The handler invocation will fail (no agent process) but clientInfo was
// captured before the call. We verify by checking that no -32602 invalid
// params error is raised about clientInfo.
expect(sent.length).toBeGreaterThan(0)
} finally {
unregister()
delete process.env.ACP_LINK_TEST_INTERNALS
}
})
})
describe('ws-message JSON-RPC decoding (audit §8.1)', () => {
test('decodeJsonWsMessage accepts JSON-RPC 2.0 requests', async () => {
const { decodeJsonWsMessage, isJsonRpc2Message } = await import(
'../ws-message.js'
)
const msg = decodeJsonWsMessage(
'{"jsonrpc":"2.0","id":1,"method":"session/prompt","params":{}}',
)
expect(isJsonRpc2Message(msg)).toBe(true)
expect((msg as { method?: string }).method).toBe('session/prompt')
})
test('decodeJsonWsMessage still accepts legacy {type,payload} envelope', async () => {
const { decodeJsonWsMessage } = await import('../ws-message.js')
const msg = decodeJsonWsMessage('{"type":"ping"}')
expect((msg as { type?: string }).type).toBe('ping')
})
test('decodeJsonWsMessage rejects non-JSON-RPC, non-type payloads', async () => {
const { decodeJsonWsMessage } = await import('../ws-message.js')
expect(() => decodeJsonWsMessage('{"foo":"bar"}')).toThrow(
'Invalid WebSocket message payload',
)
})
})

View File

@@ -211,9 +211,12 @@ export class RcsUpstreamClient {
} else if (data.type === 'keep_alive') {
// ignore keepalive
} else {
// Forward ACP protocol messages to handler (for RCS relay support)
// Forward ACP protocol messages to handler (for RCS relay support).
// This branch handles both the legacy `{type, payload}` envelope
// and JSON-RPC 2.0 messages (which have no `type` field) so the
// relay preserves the JSON-RPC format end-to-end (audit §8.12).
RcsUpstreamClient.log.debug(
{ type: data.type },
{ type: data.type, method: data.method },
'forwarding to relay handler',
)
this.messageHandler?.(data)

View File

@@ -10,10 +10,26 @@ import type { WebSocket as RawWebSocket } from 'ws'
import { createLogger } from './logger.js'
import { getOrCreateCertificate, getLanIPs } from './cert.js'
import { RcsUpstreamClient, type RcsUpstreamConfig } from './rcs-upstream.js'
import { decodeJsonWsMessage, WsPayloadTooLargeError } from './ws-message.js'
import {
decodeJsonWsMessage,
isJsonRpc2Message,
WsPayloadTooLargeError,
type JsonRpc2ClientMessage,
} from './ws-message.js'
import { authTokensEqual, extractWebSocketAuthToken } from './ws-auth.js'
export { MAX_CLIENT_WS_PAYLOAD_BYTES } from './ws-message.js'
export {
MAX_CLIENT_WS_PAYLOAD_BYTES,
isJsonRpc2Message,
type JsonRpc2ClientMessage,
} from './ws-message.js'
// JSON-RPC 2.0 reserved error codes (spec §5.1)
const JSONRPC_PARSE_ERROR = -32700
const JSONRPC_INVALID_REQUEST = -32600
const JSONRPC_METHOD_NOT_FOUND = -32601
const JSONRPC_INVALID_PARAMS = -32602
const JSONRPC_INTERNAL_ERROR = -32603
export interface ServerConfig {
port: number
@@ -88,6 +104,63 @@ interface ClientState {
promptCapabilities: PromptCapabilities | null
modelState: SessionModelState | null
isAlive: boolean
/**
* True when this client speaks JSON-RPC 2.0 (determined from the first
* framed message). When true, responses are emitted as JSON-RPC responses
* that preserve the request `id`; otherwise the legacy `{type, payload}`
* envelope is used for backwards compatibility.
*/
jsonRpc: boolean
/**
* Client-supplied identity and capabilities, captured from the JSON-RPC
* `initialize` request or legacy `connect` payload and forwarded to the
* agent instead of the hardcoded Zed fallback. See audit §8.7.
*/
clientInfo: { name: string; version: string }
clientCapabilities: Record<string, unknown>
/** Negotiated ACP protocolVersion surfaced back to the client (audit §8.13). */
protocolVersion: number | null
/** Agent identity from InitializeResult.agentInfo (audit §8.13). */
agentInfo: { name: string; version: string; [k: string]: unknown } | null
/**
* Currently in-flight JSON-RPC request being serviced. The proxy echoes this
* id back in the JSON-RPC response (audit §8.2). At most one request is
* processed per client at a time because onMessage is awaited serially.
*/
pendingJsonRpc: {
id: string | number | null
/** Legacy response type the handler will emit via send(). */
responseType: string
} | null
}
// Default fallback client identity (used only when the client provides none)
const DEFAULT_CLIENT_INFO = Object.freeze({ name: 'zed', version: '1.0.0' })
const DEFAULT_CLIENT_CAPABILITIES = Object.freeze({
fs: { readTextFile: true, writeTextFile: true },
})
/**
* Create a fresh ClientState with the default fallback client identity and
* capabilities. Used by every WebSocket open handler and the RCS relay.
*/
function createClientState(): ClientState {
return {
process: null,
connection: null,
sessionId: null,
pendingPermissions: new Map(),
agentCapabilities: null,
promptCapabilities: null,
modelState: null,
isAlive: true,
jsonRpc: false,
clientInfo: { ...DEFAULT_CLIENT_INFO },
clientCapabilities: { ...DEFAULT_CLIENT_CAPABILITIES },
protocolVersion: null,
agentInfo: null,
pendingJsonRpc: null,
}
}
// Module-level state (set when server starts)
@@ -143,7 +216,22 @@ function generateRequestId(): string {
return `perm_${Date.now()}_${Math.random().toString(36).slice(2, 11)}`
}
// Send a message to the WebSocket client (and optionally forward to RCS upstream)
// Maps legacy notification type strings to their JSON-RPC method names so
// agent→client notifications are also emitted as JSON-RPC notifications for
// JSON-RPC 2.0 clients (audit §8.1). Notifications have no id.
const LEGACY_NOTIFICATION_TO_JSONRPC: Record<string, string> = {
session_update: 'session/update',
permission_request: 'session/request_permission',
}
// Send a notification/response to the WebSocket client.
//
// For legacy `{type, payload}` clients this emits the proprietary envelope.
// For JSON-RPC 2.0 clients this additionally emits a JSON-RPC response that
// echoes the in-flight request id when the message type matches the pending
// request's expected response type (audit §8.2). Agent→client notifications
// (`session_update`, `permission_request`) are emitted as JSON-RPC
// notifications without an id.
function send(ws: WSContext, type: string, payload?: unknown): void {
if (ws.readyState === 1) {
// WebSocket.OPEN
@@ -153,6 +241,64 @@ function send(ws: WSContext, type: string, payload?: unknown): void {
if (rcsUpstream?.isRegistered()) {
rcsUpstream.send({ type, payload })
}
const state = clients.get(ws)
if (!state?.jsonRpc) return
// If this is the response to an in-flight JSON-RPC request, emit the
// standard JSON-RPC result with the preserved id.
if (state.pendingJsonRpc?.responseType === type) {
sendJsonRpcRaw(ws, {
jsonrpc: '2.0',
id: state.pendingJsonRpc.id,
result: payload ?? {},
})
state.pendingJsonRpc = null
return
}
// Agent→client notifications are also emitted as JSON-RPC notifications
// (no id) so JSON-RPC clients receive them in their native format.
const notificationMethod = LEGACY_NOTIFICATION_TO_JSONRPC[type]
if (notificationMethod) {
sendJsonRpcRaw(ws, {
jsonrpc: '2.0',
method: notificationMethod,
params: payload ?? {},
})
}
}
// Serialize a JSON-RPC 2.0 message and send it to a connected WS client.
function sendJsonRpcRaw(ws: WSContext, message: object): void {
if (ws.readyState === 1) {
ws.send(JSON.stringify(message))
}
}
/**
* Send a JSON-RPC 2.0 error response with a reserved -32xxx code (audit §8.3).
* Also emits the legacy `{type: 'error', payload: {message}}` envelope for
* backwards compatibility.
*/
function sendJsonRpcError(
ws: WSContext,
state: ClientState | undefined,
id: string | number | null,
code: number,
message: string,
): void {
if (state?.jsonRpc) {
sendJsonRpcRaw(ws, {
jsonrpc: '2.0',
id,
error: { code, message },
})
} else {
send(ws, 'error', { message, code: String(code) })
}
// Error consumed the in-flight request, if any.
if (state) state.pendingJsonRpc = null
}
// Create a Client implementation that forwards events to WebSocket
@@ -259,8 +405,9 @@ async function handleConnect(ws: WSContext): Promise<void> {
logAgent.info('already connected, resending status')
send(ws, 'status', {
connected: true,
agentInfo: { name: AGENT_COMMAND },
agentInfo: state.agentInfo ?? { name: AGENT_COMMAND },
capabilities: state.agentCapabilities,
protocolVersion: state.protocolVersion,
})
return
}
@@ -312,23 +459,23 @@ async function handleConnect(ws: WSContext): Promise<void> {
const initResult = await connection.initialize({
protocolVersion: acp.PROTOCOL_VERSION,
clientInfo: { name: 'zed', version: '1.0.0' },
clientCapabilities: {
fs: { readTextFile: true, writeTextFile: true },
},
// Forward the real client identity/capabilities (audit §8.7). Falls back
// to the Zed defaults only when the client did not provide any.
clientInfo: state.clientInfo,
clientCapabilities: state.clientCapabilities,
})
// Pass the raw agentCapabilities through unchanged so present and future
// capability fields (auth, terminal, ...) reach the client (audit §8.6).
const agentCaps = initResult.agentCapabilities
state.agentCapabilities = agentCaps
? {
_meta: agentCaps._meta,
loadSession: agentCaps.loadSession,
mcpCapabilities: agentCaps.mcpCapabilities,
promptCapabilities: agentCaps.promptCapabilities,
sessionCapabilities: agentCaps.sessionCapabilities,
}
: null
state.agentCapabilities = (agentCaps as AgentCapabilities | null) ?? null
state.promptCapabilities = agentCaps?.promptCapabilities ?? null
// Remember the negotiated protocolVersion + agentInfo so reconnects and
// JSON-RPC initialize responses can forward them to the client (§8.13).
state.protocolVersion = initResult.protocolVersion
state.agentInfo =
(initResult.agentInfo as ClientState['agentInfo'] | null | undefined) ??
null
logAgent.info(
{
@@ -345,6 +492,8 @@ async function handleConnect(ws: WSContext): Promise<void> {
connected: true,
agentInfo: initResult.agentInfo,
capabilities: state.agentCapabilities,
// Surface the negotiated protocolVersion to downstream clients (audit §8.13).
protocolVersion: initResult.protocolVersion,
})
connection.closed.then(() => {
@@ -355,9 +504,13 @@ async function handleConnect(ws: WSContext): Promise<void> {
})
} catch (error) {
logAgent.error({ error: (error as Error).message }, 'connect failed')
send(ws, 'error', {
message: `Failed to connect: ${(error as Error).message}`,
})
sendJsonRpcError(
ws,
state,
null,
JSONRPC_INTERNAL_ERROR,
`Failed to connect: ${(error as Error).message}`,
)
}
}
@@ -376,7 +529,13 @@ async function handleNewSession(
},
'handleNewSession: not connected to agent',
)
send(ws, 'error', { message: 'Not connected to agent' })
sendJsonRpcError(
ws,
state,
state?.pendingJsonRpc?.id ?? null,
JSONRPC_INVALID_REQUEST,
'Not connected to agent',
)
return
}
@@ -389,7 +548,13 @@ async function handleNewSession(
DEFAULT_PERMISSION_MODE,
)
} catch (error) {
send(ws, 'error', { message: (error as Error).message })
sendJsonRpcError(
ws,
state,
state.pendingJsonRpc?.id ?? null,
JSONRPC_INVALID_PARAMS,
(error as Error).message,
)
return
}
const result = await state.connection.newSession({
@@ -416,9 +581,13 @@ async function handleNewSession(
})
} catch (error) {
logSession.error({ error: (error as Error).message }, 'create failed')
send(ws, 'error', {
message: `Failed to create session: ${(error as Error).message}`,
})
sendJsonRpcError(
ws,
state,
state.pendingJsonRpc?.id ?? null,
JSONRPC_INTERNAL_ERROR,
`Failed to create session: ${(error as Error).message}`,
)
}
}
@@ -442,14 +611,24 @@ async function handleListSessions(
},
'handleListSessions: not connected to agent',
)
send(ws, 'error', { message: 'Not connected to agent' })
sendJsonRpcError(
ws,
state,
state?.pendingJsonRpc?.id ?? null,
JSONRPC_INVALID_REQUEST,
'Not connected to agent',
)
return
}
if (!state.agentCapabilities?.sessionCapabilities?.list) {
send(ws, 'error', {
message: 'Listing sessions is not supported by this agent',
})
sendJsonRpcError(
ws,
state,
state.pendingJsonRpc?.id ?? null,
JSONRPC_METHOD_NOT_FOUND,
'Listing sessions is not supported by this agent',
)
return
}
@@ -483,9 +662,13 @@ async function handleListSessions(
})
} catch (error) {
logSession.error({ error: (error as Error).message }, 'list failed')
send(ws, 'error', {
message: `Failed to list sessions: ${(error as Error).message}`,
})
sendJsonRpcError(
ws,
state,
state.pendingJsonRpc?.id ?? null,
JSONRPC_INTERNAL_ERROR,
`Failed to list sessions: ${(error as Error).message}`,
)
}
}
@@ -504,14 +687,24 @@ async function handleLoadSession(
},
'handleLoadSession: not connected to agent',
)
send(ws, 'error', { message: 'Not connected to agent' })
sendJsonRpcError(
ws,
state,
state?.pendingJsonRpc?.id ?? null,
JSONRPC_INVALID_REQUEST,
'Not connected to agent',
)
return
}
if (!state.agentCapabilities?.loadSession) {
send(ws, 'error', {
message: 'Loading sessions is not supported by this agent',
})
sendJsonRpcError(
ws,
state,
state.pendingJsonRpc?.id ?? null,
JSONRPC_METHOD_NOT_FOUND,
'Loading sessions is not supported by this agent',
)
return
}
@@ -535,9 +728,13 @@ async function handleLoadSession(
})
} catch (error) {
logSession.error({ error: (error as Error).message }, 'load failed')
send(ws, 'error', {
message: `Failed to load session: ${(error as Error).message}`,
})
sendJsonRpcError(
ws,
state,
state.pendingJsonRpc?.id ?? null,
JSONRPC_INTERNAL_ERROR,
`Failed to load session: ${(error as Error).message}`,
)
}
}
@@ -556,14 +753,24 @@ async function handleResumeSession(
},
'handleResumeSession: not connected to agent',
)
send(ws, 'error', { message: 'Not connected to agent' })
sendJsonRpcError(
ws,
state,
state?.pendingJsonRpc?.id ?? null,
JSONRPC_INVALID_REQUEST,
'Not connected to agent',
)
return
}
if (!state.agentCapabilities?.sessionCapabilities?.resume) {
send(ws, 'error', {
message: 'Resuming sessions is not supported by this agent',
})
sendJsonRpcError(
ws,
state,
state.pendingJsonRpc?.id ?? null,
JSONRPC_METHOD_NOT_FOUND,
'Resuming sessions is not supported by this agent',
)
return
}
@@ -586,9 +793,13 @@ async function handleResumeSession(
})
} catch (error) {
logSession.error({ error: (error as Error).message }, 'resume failed')
send(ws, 'error', {
message: `Failed to resume session: ${(error as Error).message}`,
})
sendJsonRpcError(
ws,
state,
state.pendingJsonRpc?.id ?? null,
JSONRPC_INTERNAL_ERROR,
`Failed to resume session: ${(error as Error).message}`,
)
}
}
@@ -599,7 +810,13 @@ async function handlePrompt(
): Promise<void> {
const state = clients.get(ws)
if (!state?.connection || !state.sessionId) {
send(ws, 'error', { message: 'No active session' })
sendJsonRpcError(
ws,
state,
state?.pendingJsonRpc?.id ?? null,
JSONRPC_INVALID_REQUEST,
'No active session',
)
return
}
@@ -624,7 +841,13 @@ async function handlePrompt(
send(ws, 'prompt_complete', result)
} catch (error) {
logPrompt.error({ error: (error as Error).message }, 'failed')
send(ws, 'error', { message: `Prompt failed: ${(error as Error).message}` })
sendJsonRpcError(
ws,
state,
state.pendingJsonRpc?.id ?? null,
JSONRPC_INTERNAL_ERROR,
`Prompt failed: ${(error as Error).message}`,
)
}
}
@@ -668,14 +891,24 @@ async function handleSetSessionModel(
): Promise<void> {
const state = clients.get(ws)
if (!state?.connection || !state.sessionId) {
send(ws, 'error', { message: 'No active session' })
sendJsonRpcError(
ws,
state,
state?.pendingJsonRpc?.id ?? null,
JSONRPC_INVALID_REQUEST,
'No active session',
)
return
}
if (!state.modelState) {
send(ws, 'error', {
message: 'Model selection not supported by this agent',
})
sendJsonRpcError(
ws,
state,
state.pendingJsonRpc?.id ?? null,
JSONRPC_METHOD_NOT_FOUND,
'Model selection not supported by this agent',
)
return
}
@@ -693,9 +926,13 @@ async function handleSetSessionModel(
logSession.info({ modelId: params.modelId }, 'model changed')
} catch (error) {
logSession.error({ error: (error as Error).message }, 'set model failed')
send(ws, 'error', {
message: `Failed to set model: ${(error as Error).message}`,
})
sendJsonRpcError(
ws,
state,
state.pendingJsonRpc?.id ?? null,
JSONRPC_INTERNAL_ERROR,
`Failed to set model: ${(error as Error).message}`,
)
}
}
@@ -918,30 +1155,301 @@ async function dispatchClientMessage(
}
}
/**
* Maps JSON-RPC method names to their legacy handler + the legacy response
* type the handler emits via send(). Used by dispatchJsonRpcMessage to route
* standard ACP methods (audit §8.1, §8.4).
*/
const JSONRPC_METHOD_HANDLERS: Record<
string,
{
responseType: string
handle: (ws: WSContext, params: unknown) => Promise<void> | void
}
> = {
initialize: { responseType: 'status', handle: handleConnect },
'session/new': {
responseType: 'session_created',
handle: handleJsonRpcNewSession,
},
'session/prompt': {
responseType: 'prompt_complete',
handle: handleJsonRpcPrompt,
},
'session/cancel': { responseType: '', handle: handleCancel },
'session/list': {
responseType: 'session_list',
handle: handleJsonRpcListSessions,
},
'session/load': {
responseType: 'session_loaded',
handle: handleJsonRpcLoadSession,
},
'session/resume': {
responseType: 'session_resumed',
handle: handleJsonRpcResumeSession,
},
'session/set_model': {
responseType: 'model_changed',
handle: handleJsonRpcSetSessionModel,
},
'session/set_mode': {
responseType: 'session_mode_set',
handle: handleJsonRpcSetSessionMode,
},
'session/close': {
responseType: 'session_closed',
handle: handleJsonRpcCloseSession,
},
}
// JSON-RPC method wrappers that accept `params: unknown` and forward to the
// existing handlers with the decoded payload.
async function handleJsonRpcNewSession(
ws: WSContext,
params: unknown,
): Promise<void> {
const payload = optionalPayloadRecord(params, 'session/new')
await handleNewSession(ws, {
cwd: optionalStringField(payload, 'cwd', 'session/new.cwd'),
permissionMode: optionalStringField(
payload,
'permissionMode',
'session/new.permissionMode',
),
})
}
async function handleJsonRpcPrompt(
ws: WSContext,
params: unknown,
): Promise<void> {
const payload = payloadRecord(params, 'session/prompt')
// ACP session/prompt params: { sessionId, prompt: ContentBlock[] }
// Accept either `prompt` (spec) or `content` (legacy) for compatibility.
const content = payload.prompt ?? payload.content
await handlePrompt(ws, { content: decodeContentBlocks(content) })
}
async function handleJsonRpcListSessions(
ws: WSContext,
params: unknown,
): Promise<void> {
const payload = optionalRecord(params)
await handleListSessions(ws, {
cwd: optionalString(payload.cwd),
cursor: optionalString(payload.cursor),
})
}
async function handleJsonRpcLoadSession(
ws: WSContext,
params: unknown,
): Promise<void> {
const payload = payloadRecord(params, 'session/load')
if (typeof payload.sessionId !== 'string') {
throw new Error('Invalid session/load payload')
}
await handleLoadSession(ws, {
sessionId: payload.sessionId,
cwd: optionalString(payload.cwd),
})
}
async function handleJsonRpcResumeSession(
ws: WSContext,
params: unknown,
): Promise<void> {
const payload = payloadRecord(params, 'session/resume')
if (typeof payload.sessionId !== 'string') {
throw new Error('Invalid session/resume payload')
}
await handleResumeSession(ws, {
sessionId: payload.sessionId,
cwd: optionalString(payload.cwd),
})
}
async function handleJsonRpcSetSessionModel(
ws: WSContext,
params: unknown,
): Promise<void> {
const payload = payloadRecord(params, 'session/set_model')
if (typeof payload.modelId !== 'string') {
throw new Error('Invalid session/set_model payload')
}
await handleSetSessionModel(ws, { modelId: payload.modelId })
}
/**
* Pass-through handlers for v1 baseline methods that the proprietary
* whitelist previously dropped (audit §8.4). They forward the call to the
* underlying SDK ClientSideConnection and surface the result.
*/
async function handleJsonRpcSetSessionMode(
ws: WSContext,
params: unknown,
): Promise<void> {
const state = clients.get(ws)
if (!state?.connection) {
throw new Error('Not connected to agent')
}
const result = await state.connection.setSessionMode(
params as { sessionId: string; modeId: string },
)
send(ws, 'session_mode_set', result ?? {})
}
async function handleJsonRpcCloseSession(
ws: WSContext,
params: unknown,
): Promise<void> {
const state = clients.get(ws)
if (!state?.connection) {
throw new Error('Not connected to agent')
}
const result = await state.connection.unstable_closeSession(
params as { sessionId: string },
)
send(ws, 'session_closed', result ?? {})
}
/**
* Handle the JSON-RPC standard cancellation primitive `$/cancel_request`
* (audit §8.5). Unlike the ACP-specific `session/cancel` notification, this
* cancels an in-flight request by id. We forward to the ACP cancel path and
* also clear any pending permission request.
*/
async function handleJsonRpcCancelRequest(
ws: WSContext,
params: unknown,
): Promise<void> {
const payload = optionalRecord(params)
logWs.info({ cancelledId: payload.id }, '$/cancel_request received')
await handleCancel(ws)
}
/**
* Route a JSON-RPC 2.0 message. Requests get a response with the echoed id;
* notifications (no id) are dispatched without a response. Unknown methods
* yield a JSON-RPC -32601 error (audit §8.4). `$/cancel_request` is handled
* specially (audit §8.5).
*/
async function dispatchJsonRpcMessage(
ws: WSContext,
msg: JsonRpc2ClientMessage,
): Promise<void> {
const state = clients.get(ws)
// Mark this client as JSON-RPC from the first framed message.
if (state) state.jsonRpc = true
// Capture client identity/capabilities from initialize (audit §8.7).
if (msg.method === 'initialize' && state) {
const params = isRecord(msg.params) ? msg.params : {}
if (isRecord(params.clientInfo)) {
const ci = params.clientInfo
if (typeof ci.name === 'string' && typeof ci.version === 'string') {
state.clientInfo = { name: ci.name, version: ci.version }
}
}
if (isRecord(params.clientCapabilities)) {
state.clientCapabilities = params.clientCapabilities
}
}
// Notification (no id) — dispatch without a response.
if (!('id' in msg) || msg.id === undefined) {
if (msg.method === '$/cancel_request') {
await handleJsonRpcCancelRequest(ws, msg.params)
return
}
if (msg.method === 'session/cancel') {
await handleCancel(ws)
return
}
// Unknown notification — silently ignore per JSON-RPC 2.0 (notifications
// cannot be responded to).
logWs.debug({ method: msg.method }, 'ignoring unknown notification')
return
}
// Request (has id) — dispatch and the handler will emit a response.
if (msg.method === '$/cancel_request') {
await handleJsonRpcCancelRequest(ws, msg.params)
// Cancellation is itself a notification-style request; respond with null.
if (state) state.pendingJsonRpc = { id: msg.id, responseType: '' }
sendJsonRpcRaw(ws, { jsonrpc: '2.0', id: msg.id, result: null })
if (state) state.pendingJsonRpc = null
return
}
const entry = JSONRPC_METHOD_HANDLERS[msg.method]
if (!entry) {
sendJsonRpcError(
ws,
state,
msg.id,
JSONRPC_METHOD_NOT_FOUND,
`Method not found: ${msg.method}`,
)
return
}
// Track the in-flight request so the handler's send() emits a JSON-RPC
// response with the echoed id (audit §8.2).
if (state)
state.pendingJsonRpc = { id: msg.id, responseType: entry.responseType }
try {
await entry.handle(ws, msg.params)
// If the handler did not emit the expected response (e.g. it short
// circuited with an error already), still clear the pending slot.
if (state?.pendingJsonRpc) {
sendJsonRpcRaw(ws, {
jsonrpc: '2.0',
id: msg.id,
result: {},
})
state.pendingJsonRpc = null
}
} catch (error) {
const code = (error as Error).message.startsWith('Invalid ')
? JSONRPC_INVALID_PARAMS
: JSONRPC_INTERNAL_ERROR
sendJsonRpcError(ws, state, msg.id, code, (error as Error).message)
}
}
export const __testing = {
dispatchClientMessage(ws: WSContext, data: unknown): Promise<void> {
assertTestingInternalsEnabled()
return dispatchClientMessage(ws, data as ProxyMessage)
},
dispatchJsonRpcMessage(ws: WSContext, data: unknown): Promise<void> {
assertTestingInternalsEnabled()
return dispatchJsonRpcMessage(ws, data as JsonRpc2ClientMessage)
},
registerClient(
ws: WSContext,
state: {
connection?: unknown
process?: ChildProcess | null
sessionId?: string | null
clientInfo?: { name: string; version: string }
clientCapabilities?: Record<string, unknown>
jsonRpc?: boolean
},
): () => void {
assertTestingInternalsEnabled()
clients.set(ws, {
process: state.process ?? null,
connection: (state.connection ?? null) as acp.ClientSideConnection | null,
sessionId: state.sessionId ?? null,
pendingPermissions: new Map(),
agentCapabilities: null,
promptCapabilities: null,
modelState: null,
isAlive: true,
})
const full = createClientState()
full.process = state.process ?? null
full.connection = (state.connection ??
null) as acp.ClientSideConnection | null
full.sessionId = state.sessionId ?? null
if (state.clientInfo) full.clientInfo = state.clientInfo
if (state.clientCapabilities)
full.clientCapabilities = state.clientCapabilities
if (typeof state.jsonRpc === 'boolean') full.jsonRpc = state.jsonRpc
clients.set(ws, full)
return () => {
clients.delete(ws)
}
@@ -1071,23 +1579,21 @@ export async function startServer(config: ServerConfig): Promise<void> {
})
const relayWs = createRelayWs()
const relayState: ClientState = {
process: null,
connection: null,
sessionId: null,
pendingPermissions: new Map(),
agentCapabilities: null,
promptCapabilities: null,
modelState: null,
isAlive: true,
}
const relayState = createClientState()
clients.set(relayWs, relayState)
rcsUpstream.setMessageHandler(async msg => {
try {
const data = decodeClientMessage(msg)
logRelay.debug({ type: data.type }, 'processing')
await dispatchClientMessage(relayWs, data)
// The RCS relay forwards messages from the Web UI. Accept both
// JSON-RPC 2.0 (audit §8.12) and the legacy `{type, payload}` envelope.
if (isJsonRpc2Message(msg)) {
logRelay.debug({ method: msg.method }, 'processing jsonrpc')
await dispatchJsonRpcMessage(relayWs, msg)
} else {
const data = decodeClientMessage(msg)
logRelay.debug({ type: data.type }, 'processing')
await dispatchClientMessage(relayWs, data)
}
} catch (error) {
logRelay.error({ error: (error as Error).message }, 'handler error')
}
@@ -1134,16 +1640,7 @@ export async function startServer(config: ServerConfig): Promise<void> {
return {
onOpen(_event, ws) {
logWs.info('client connected')
const state: ClientState = {
process: null,
connection: null,
sessionId: null,
pendingPermissions: new Map(),
agentCapabilities: null,
promptCapabilities: null,
modelState: null,
isAlive: true,
}
const state = createClientState()
clients.set(ws, state)
const rawWs = ws.raw as RawWebSocket
@@ -1153,9 +1650,18 @@ export async function startServer(config: ServerConfig): Promise<void> {
},
async onMessage(event, ws) {
try {
const data = decodeClientWsMessage(event.data)
logWs.debug({ type: data.type }, 'received')
await dispatchClientMessage(ws, data)
// Decode the raw frame once. JSON-RPC 2.0 messages are routed by
// method name (audit §8.1, §8.4, §8.5); legacy `{type, payload}`
// messages keep the existing dispatch path for backwards compat.
const decoded = decodeJsonWsMessage(event.data)
if (isJsonRpc2Message(decoded)) {
logWs.debug({ method: decoded.method }, 'received jsonrpc')
await dispatchJsonRpcMessage(ws, decoded)
} else {
const data = decodeClientMessage(decoded)
logWs.debug({ type: data.type }, 'received')
await dispatchClientMessage(ws, data)
}
} catch (error) {
if (error instanceof WsPayloadTooLargeError) {
logWs.warn({ error: error.message }, 'message too large')
@@ -1163,7 +1669,14 @@ export async function startServer(config: ServerConfig): Promise<void> {
return
}
logWs.error({ error: (error as Error).message }, 'message error')
send(ws, 'error', { message: `Error: ${(error as Error).message}` })
const state = clients.get(ws)
sendJsonRpcError(
ws,
state,
state?.pendingJsonRpc?.id ?? null,
JSONRPC_PARSE_ERROR,
`Error: ${(error as Error).message}`,
)
}
},
onClose(_event, ws) {

View File

@@ -7,12 +7,65 @@ export class WsPayloadTooLargeError extends Error {
}
}
/**
* Legacy proprietary envelope shape: `{ type, payload? }`.
* Retained for backwards compatibility with older clients (e.g. the RCS Web UI)
* that have not migrated to JSON-RPC 2.0 yet.
*/
export interface JsonWsMessage {
type: string
payload?: unknown
[key: string]: unknown
}
/**
* JSON-RPC 2.0 envelope as defined by the specification.
* See transports.mdx: custom transports MUST preserve the JSON-RPC message
* format and lifecycle requirements defined by ACP.
*/
export interface JsonRpc2Request {
jsonrpc: '2.0'
id: string | number | null
method: string
params?: unknown
}
export interface JsonRpc2Notification {
jsonrpc: '2.0'
method: string
params?: unknown
}
export interface JsonRpc2Response {
jsonrpc: '2.0'
id: string | number | null
result?: unknown
error?: { code: number; message: string; data?: unknown }
}
export type JsonRpc2Message =
| JsonRpc2Request
| JsonRpc2Notification
| JsonRpc2Response
/**
* Messages that carry a `method` field — i.e. requests and notifications that
* the proxy can route. Responses (no method) are excluded because clients are
* not expected to send them to the agent.
*/
export type JsonRpc2ClientMessage = JsonRpc2Request | JsonRpc2Notification
export function isJsonRpc2Message(
value: unknown,
): value is JsonRpc2ClientMessage {
return (
typeof value === 'object' &&
value !== null &&
(value as { jsonrpc?: unknown }).jsonrpc === '2.0' &&
typeof (value as { method?: unknown }).method === 'string'
)
}
function assertPayloadSize(byteLength: number): void {
if (byteLength > MAX_CLIENT_WS_PAYLOAD_BYTES) {
throw new WsPayloadTooLargeError(byteLength)
@@ -49,14 +102,28 @@ function decodeWsText(data: unknown): string {
throw new Error('Unsupported WebSocket message payload')
}
/**
* Decode a WebSocket text frame into either a JSON-RPC 2.0 message or the
* legacy proprietary `{type, payload}` envelope.
*
* Accepts:
* - JSON-RPC 2.0 requests/notifications/responses (`{ jsonrpc: '2.0', method, ... }`)
* - Legacy proprietary messages (`{ type: string, payload?: unknown }`)
*
* Rejects anything else with `Invalid WebSocket message payload`.
*/
export function decodeJsonWsMessage(data: unknown): JsonWsMessage {
const parsed = JSON.parse(decodeWsText(data)) as unknown
if (
typeof parsed !== 'object' ||
parsed === null ||
!('type' in parsed) ||
typeof parsed.type !== 'string'
) {
if (typeof parsed !== 'object' || parsed === null) {
throw new Error('Invalid WebSocket message payload')
}
// JSON-RPC 2.0 envelope — preserve all original fields so the router can
// correlate request ids and forward notifications unchanged.
if (isJsonRpc2Message(parsed)) {
return parsed as unknown as JsonWsMessage
}
// Legacy proprietary envelope `{ type, payload? }`.
if (!('type' in parsed) || typeof parsed.type !== 'string') {
throw new Error('Invalid WebSocket message payload')
}
return parsed as JsonWsMessage

View File

@@ -260,25 +260,42 @@ describe('AcpAgent', () => {
expect(typeof res.agentInfo?.version).toBe('string')
})
test('advertises image and embeddedContext capability', async () => {
test('advertises embeddedContext capability and disables image until multimodal input lands', async () => {
const agent = new AcpAgent(makeConn())
const res = await agent.initialize({} as any)
expect(res.agentCapabilities?.promptCapabilities?.image).toBe(true)
// image:false — promptToQueryInput does not parse image blocks yet
expect(res.agentCapabilities?.promptCapabilities?.image).toBe(false)
expect(res.agentCapabilities?.promptCapabilities?.embeddedContext).toBe(
true,
)
})
test('returns explicit empty authMethods', async () => {
const agent = new AcpAgent(makeConn())
const res = await agent.initialize({} as any)
expect(res.authMethods).toEqual([])
})
test('loadSession capability is true', async () => {
const agent = new AcpAgent(makeConn())
const res = await agent.initialize({} as any)
expect(res.agentCapabilities?.loadSession).toBe(true)
})
test('session capabilities include fork, list, resume, close', async () => {
test('session capabilities include list, resume, close (fork advertised via _meta)', async () => {
const agent = new AcpAgent(makeConn())
const res = await agent.initialize({} as any)
expect(res.agentCapabilities?.sessionCapabilities).toBeDefined()
const caps = res.agentCapabilities?.sessionCapabilities as any
expect(caps).toBeDefined()
expect(caps.list).toBeDefined()
expect(caps.resume).toBeDefined()
expect(caps.close).toBeDefined()
// fork is UNSTABLE — advertised under _meta.claudeCode.forkSession, not
// under sessionCapabilities (which is stable-v1 only).
expect(caps.fork).toBeUndefined()
expect(
(res.agentCapabilities?._meta as any)?.claudeCode?.forkSession,
).toBe(true)
})
})
@@ -298,12 +315,13 @@ describe('AcpAgent', () => {
expect(res.sessionId.length).toBeGreaterThan(0)
})
test('returns modes and models', async () => {
test('returns modes and configOptions (models omitted for v1 compliance)', async () => {
const agent = new AcpAgent(makeConn())
const res = await agent.newSession({ cwd: '/tmp' } as any)
expect(res.modes).toBeDefined()
expect(res.models).toBeDefined()
expect(res.configOptions).toBeDefined()
// Stable v1 NewSessionResponse does not define `models`
expect((res as any).models).toBeUndefined()
})
test('each call returns a unique sessionId', async () => {
@@ -328,9 +346,10 @@ describe('AcpAgent', () => {
test('calls getMainLoopModel to resolve current model', async () => {
const agent = new AcpAgent(makeConn())
const res = await agent.newSession({ cwd: '/tmp' } as any)
await agent.newSession({ cwd: '/tmp' } as any)
expect(mockGetMainLoopModel).toHaveBeenCalled()
expect(res.models?.currentModelId).toBe('claude-sonnet-4-6')
// models is no longer in the v1 response, but the engine still receives it
expect(mockSetModel).toHaveBeenCalledWith('claude-sonnet-4-6')
})
test('calls queryEngine.setModel with resolved model', async () => {
@@ -342,8 +361,7 @@ describe('AcpAgent', () => {
test('respects model alias resolution via getMainLoopModel', async () => {
mockGetMainLoopModel.mockReturnValueOnce('glm-5.1')
const agent = new AcpAgent(makeConn())
const res = await agent.newSession({ cwd: '/tmp' } as any)
expect(res.models?.currentModelId).toBe('glm-5.1')
await agent.newSession({ cwd: '/tmp' } as any)
expect(mockSetModel).toHaveBeenCalledWith('glm-5.1')
})
@@ -464,21 +482,23 @@ describe('AcpAgent', () => {
).rejects.toThrow('nonexistent')
})
test('returns end_turn for empty prompt text', async () => {
test('rejects empty prompt text with an error', async () => {
const agent = new AcpAgent(makeConn())
const { sessionId } = await agent.newSession({ cwd: '/tmp' } as any)
const res = await agent.prompt({ sessionId, prompt: [] } as any)
expect(res.stopReason).toBe('end_turn')
await expect(
agent.prompt({ sessionId, prompt: [] } as any),
).rejects.toThrow('Prompt content is empty')
})
test('returns end_turn for whitespace-only prompt', async () => {
test('rejects whitespace-only prompt with an error', async () => {
const agent = new AcpAgent(makeConn())
const { sessionId } = await agent.newSession({ cwd: '/tmp' } as any)
const res = await agent.prompt({
sessionId,
prompt: [{ type: 'text', text: ' ' }],
} as any)
expect(res.stopReason).toBe('end_turn')
await expect(
agent.prompt({
sessionId,
prompt: [{ type: 'text', text: ' ' }],
} as any),
).rejects.toThrow('Prompt content is empty')
})
test('calls forwardSessionUpdates for valid prompt', async () => {
@@ -556,7 +576,7 @@ describe('AcpAgent', () => {
).rejects.toThrow('unexpected')
})
test('returns usage from forwardSessionUpdates', async () => {
test('returns usage under _meta.claudeCode.usage from forwardSessionUpdates', async () => {
const agent = new AcpAgent(makeConn())
const { sessionId } = await agent.newSession({ cwd: '/tmp' } as any)
;(forwardSessionUpdates as ReturnType<typeof mock>).mockResolvedValueOnce(
@@ -574,10 +594,13 @@ describe('AcpAgent', () => {
sessionId,
prompt: [{ type: 'text', text: 'hello' }],
} as any)
expect(res.usage).toBeDefined()
expect(res.usage!.inputTokens).toBe(100)
expect(res.usage!.outputTokens).toBe(50)
expect(res.usage!.totalTokens).toBe(165)
// Stable v1 PromptResponse has no root `usage`; it lives under _meta.
expect((res as any).usage).toBeUndefined()
const usage = (res as any)._meta?.claudeCode?.usage
expect(usage).toBeDefined()
expect(usage.inputTokens).toBe(100)
expect(usage.outputTokens).toBe(50)
expect(usage.totalTokens).toBe(165)
})
})
@@ -649,7 +672,7 @@ describe('AcpAgent', () => {
})
describe('prompt usage tracking', () => {
test('returns totalTokens as sum of all token types', async () => {
test('reports totalTokens as sum of all token types under _meta.claudeCode.usage', async () => {
const agent = new AcpAgent(makeConn())
const { sessionId } = await agent.newSession({ cwd: '/tmp' } as any)
;(forwardSessionUpdates as ReturnType<typeof mock>).mockResolvedValueOnce(
@@ -667,11 +690,12 @@ describe('AcpAgent', () => {
sessionId,
prompt: [{ type: 'text', text: 'hello' }],
} as any)
expect(res.usage).toBeDefined()
expect(res.usage!.totalTokens).toBe(165)
const usage = (res as any)._meta?.claudeCode?.usage
expect(usage).toBeDefined()
expect(usage.totalTokens).toBe(165)
})
test('returns undefined usage when forwardSessionUpdates returns none', async () => {
test('omits _meta.usage when forwardSessionUpdates returns none', async () => {
const agent = new AcpAgent(makeConn())
const { sessionId } = await agent.newSession({ cwd: '/tmp' } as any)
;(forwardSessionUpdates as ReturnType<typeof mock>).mockResolvedValueOnce(
@@ -683,7 +707,7 @@ describe('AcpAgent', () => {
sessionId,
prompt: [{ type: 'text', text: 'hello' }],
} as any)
expect(res.usage).toBeUndefined()
expect((res as any)._meta).toBeUndefined()
})
})
@@ -734,7 +758,8 @@ describe('AcpAgent', () => {
} as any)
expect(agent.sessions.has(requestedId)).toBe(true)
expect(res.modes).toBeDefined()
expect(res.models).toBeDefined()
// models is omitted for v1 compliance
expect((res as any).models).toBeUndefined()
})
test('reuses existing session when sessionId matches and fingerprint unchanged', async () => {
@@ -805,12 +830,26 @@ describe('AcpAgent', () => {
const agent = new AcpAgent(makeConn())
const original = await agent.newSession({ cwd: '/tmp' } as any)
const forked = await agent.unstable_forkSession({
// params.sessionId is the source session to fork from
sessionId: original.sessionId,
cwd: '/tmp',
mcpServers: [],
} as any)
expect(forked.sessionId).not.toBe(original.sessionId)
expect(agent.sessions.has(forked.sessionId)).toBe(true)
})
test('attempts to load source session history when forking', async () => {
const agent = new AcpAgent(makeConn())
const original = await agent.newSession({ cwd: '/tmp' } as any)
mockGetLastSessionLog.mockClear()
await agent.unstable_forkSession({
sessionId: original.sessionId,
cwd: '/tmp',
mcpServers: [],
} as any)
expect(mockGetLastSessionLog).toHaveBeenCalledWith(original.sessionId)
})
})
describe('setSessionMode', () => {
@@ -919,17 +958,32 @@ describe('AcpAgent', () => {
const session = agent.sessions.get(sessionId)
removeBypassMode(session)
// The value is rejected because it is not in the option's listed values
// (config-option validation runs before the mode-availability check).
await expect(
agent.setSessionConfigOption({
sessionId,
configId: 'mode',
value: 'bypassPermissions',
} as any),
).rejects.toThrow('Mode not available')
).rejects.toThrow(/Invalid value 'bypassPermissions'/)
expect(session?.modes.currentModeId).toBe('default')
expect(session?.appState.toolPermissionContext.mode).toBe('default')
})
test('rejects mode values not listed in the option options array', async () => {
const agent = new AcpAgent(makeConn())
const { sessionId } = await agent.newSession({ cwd: '/tmp' } as any)
await expect(
agent.setSessionConfigOption({
sessionId,
configId: 'mode',
value: 'totally-not-a-real-mode',
} as any),
).rejects.toThrow(/must be one of:/)
})
})
describe('prompt queueing', () => {

View File

@@ -299,6 +299,91 @@ describe('toolInfoFromToolUse', () => {
])
})
test('Read with relative file_path and cwd → locations resolved to absolute', () => {
// Audit §5.5: ToolCallLocation.path MUST be absolute. A relative input
// path is resolved against the session cwd before being emitted.
const info = toolInfoFromToolUse(
{
name: 'Read',
id: 'x',
input: { file_path: 'src/main.ts' },
},
false,
'/Users/test/project',
)
expect(info.locations).toEqual([
{ path: '/Users/test/project/src/main.ts', line: 1 },
])
})
test('Write with relative file_path and cwd → diff path resolved absolute', () => {
// Audit §5.5: Diff.path MUST be absolute.
const info = toolInfoFromToolUse(
{
name: 'Write',
id: 'x',
input: { file_path: 'rel/file.txt', content: 'hi' },
},
false,
'/Users/test/project',
)
expect(info.content).toEqual([
{
type: 'diff',
path: '/Users/test/project/rel/file.txt',
oldText: null,
newText: 'hi',
},
])
expect(info.locations).toEqual([
{ path: '/Users/test/project/rel/file.txt' },
])
})
test('Edit with relative file_path and cwd → diff path resolved absolute', () => {
// Audit §5.5: Diff.path MUST be absolute.
const info = toolInfoFromToolUse(
{
name: 'Edit',
id: 'x',
input: {
file_path: 'rel/edit.txt',
old_string: 'a',
new_string: 'b',
},
},
false,
'/Users/test/project',
)
expect(info.content).toEqual([
{
type: 'diff',
path: '/Users/test/project/rel/edit.txt',
oldText: 'a',
newText: 'b',
},
])
expect(info.locations).toEqual([
{ path: '/Users/test/project/rel/edit.txt' },
])
})
test('Glob with relative path and cwd → locations resolved absolute', () => {
// Audit §5.5: ToolCallLocation.path MUST be absolute. Title keeps the raw
// input for display, but the emitted location is resolved against cwd.
const info = toolInfoFromToolUse(
{
name: 'Glob',
id: 'x',
input: { pattern: '*.ts', path: 'src' },
},
false,
'/Users/test/project',
)
expect(info.title).toBe('Find `src` `*.ts`')
expect(info.locations).toEqual([{ path: '/Users/test/project/src' }])
})
// ── WebSearch ─────────────────────────────────────────────────
test('WebSearch with allowed/blocked domains', () => {
@@ -543,6 +628,91 @@ describe('toolUpdateFromToolResult', () => {
)
expect(result.title).toBe('Exited Plan Mode')
})
test('renders resource_link content as ACP ResourceLink (audit §7.3)', () => {
const result = toolUpdateFromToolResult(
{
content: [
{
type: 'resource_link',
uri: 'file:///tmp/spec.md',
name: 'Spec',
mimeType: 'text/markdown',
},
],
is_error: false,
tool_use_id: 't1',
},
{ name: 'SomeTool', id: 't1' },
)
expect(result.content).toEqual([
{
type: 'content',
content: {
type: 'resource_link',
uri: 'file:///tmp/spec.md',
name: 'Spec',
mimeType: 'text/markdown',
},
},
])
})
test('resource_link without name falls back to uri (audit §7.3)', () => {
const result = toolUpdateFromToolResult(
{
content: [{ type: 'resource_link', uri: 'file:///tmp/x.md' }],
is_error: false,
tool_use_id: 't1',
},
{ name: 'SomeTool', id: 't1' },
)
expect(result.content).toEqual([
{
type: 'content',
content: {
type: 'resource_link',
uri: 'file:///tmp/x.md',
name: 'file:///tmp/x.md',
mimeType: undefined,
},
},
])
})
test('renders resource content as ACP EmbeddedResource (audit §7.3)', () => {
const result = toolUpdateFromToolResult(
{
content: [
{
type: 'resource',
resource: {
uri: 'file:///tmp/readme.md',
mimeType: 'text/markdown',
text: '# Hello',
},
},
],
is_error: false,
tool_use_id: 't1',
},
{ name: 'SomeTool', id: 't1' },
)
expect(result.content).toEqual([
{
type: 'content',
content: {
type: 'resource',
resource: {
uri: 'file:///tmp/readme.md',
mimeType: 'text/markdown',
text: '# Hello',
blob: undefined,
},
},
},
])
})
})
// ── toolUpdateFromEditToolResponse ─────────────────────────────────
@@ -650,6 +820,56 @@ describe('toolUpdateFromEditToolResponse', () => {
}),
).toEqual({})
})
test('resolves relative filePath against cwd (audit §5.5)', () => {
// ToolCallLocation.path / Diff.path MUST be absolute.
const result = toolUpdateFromEditToolResponse(
{
filePath: 'rel/file.ts',
structuredPatch: [
{
oldStart: 1,
oldLines: 1,
newStart: 1,
newLines: 1,
lines: ['-old', '+new'],
},
],
},
'/Users/test/project',
)
expect(result).toEqual({
content: [
{
type: 'diff',
path: '/Users/test/project/rel/file.ts',
oldText: 'old',
newText: 'new',
},
],
locations: [{ path: '/Users/test/project/rel/file.ts', line: 1 }],
})
})
test('keeps absolute filePath unchanged when cwd provided', () => {
const result = toolUpdateFromEditToolResponse(
{
filePath: '/abs/file.ts',
structuredPatch: [
{
oldStart: 1,
oldLines: 1,
newStart: 1,
newLines: 1,
lines: ['-old', '+new'],
},
],
},
'/Users/test/project',
)
expect(result.content![0]).toMatchObject({ path: '/abs/file.ts' })
expect(result.locations![0]).toMatchObject({ path: '/abs/file.ts' })
})
})
// ── markdownEscape ─────────────────────────────────────────────────
@@ -945,7 +1165,10 @@ describe('forwardSessionUpdates', () => {
expect(update.rawInput).not.toBe(input)
})
test('sends usage_update on result message with correct tokens', async () => {
test('returns accumulated usage on result message without sending usage_update', async () => {
// usage_update is an UNSTABLE SessionUpdate discriminator and is no longer
// emitted (audit §4.1). Token totals are still aggregated for the
// PromptResponse return value so callers can include them via _meta.
const conn = makeConn()
const msgs: SDKMessage[] = [
{
@@ -973,9 +1196,19 @@ describe('forwardSessionUpdates', () => {
expect(result.usage).toBeDefined()
expect(result.usage!.inputTokens).toBe(100)
expect(result.usage!.outputTokens).toBe(50)
const calls = (conn.sessionUpdate as ReturnType<typeof mock>).mock.calls
const usageUpdate = calls.find(
(c: unknown[]) =>
((c[0] as Record<string, Record<string, unknown>>).update ?? {})[
'sessionUpdate'
] === 'usage_update',
)
expect(usageUpdate).toBeUndefined()
})
test('sends usage_update with context window from modelUsage', async () => {
test('does not emit usage_update even when modelUsage reports context window', async () => {
// Context-window resolution still runs internally (so PromptResponse can
// surface it), but no usage_update notification is sent for v1 compliance.
const conn = makeConn()
const msgs: SDKMessage[] = [
{
@@ -1023,18 +1256,10 @@ describe('forwardSessionUpdates', () => {
'sessionUpdate'
] === 'usage_update',
)
expect(usageUpdate).toBeDefined()
expect(
(
(usageUpdate![0] as Record<string, unknown>).update as Record<
string,
unknown
>
).size,
).toBe(1000000)
expect(usageUpdate).toBeUndefined()
})
test('sends usage_update with prefix-matched modelUsage', async () => {
test('prefix-matches modelUsage without emitting usage_update', async () => {
const conn = makeConn()
const msgs: SDKMessage[] = [
{
@@ -1082,18 +1307,125 @@ describe('forwardSessionUpdates', () => {
'sessionUpdate'
] === 'usage_update',
)
expect(usageUpdate).toBeDefined()
expect(
(
(usageUpdate![0] as Record<string, unknown>).update as Record<
string,
unknown
>
).size,
).toBe(2000000)
expect(usageUpdate).toBeUndefined()
})
test('resets usage on compact_boundary', async () => {
test('maps refusal stop_reason to ACP refusal stop reason', async () => {
// Audit §3.3: a safety refusal must surface as StopReason::refusal rather
// than being misreported as end_turn.
const conn = makeConn()
const msgs: SDKMessage[] = [
{
type: 'result',
subtype: 'success',
is_error: false,
result: '',
stop_reason: 'refusal',
} as unknown as SDKMessage,
]
const result = await forwardSessionUpdates(
's1',
makeStream(msgs),
conn,
new AbortController().signal,
{},
)
expect(result.stopReason).toBe('refusal')
})
test('success with max_tokens stop_reason maps to max_tokens when not error', async () => {
// Audit §3.3/§3.4: success + max_tokens + no error surfaces max_tokens.
const conn = makeConn()
const msgs: SDKMessage[] = [
{
type: 'result',
subtype: 'success',
is_error: false,
result: '',
stop_reason: 'max_tokens',
} as unknown as SDKMessage,
]
const result = await forwardSessionUpdates(
's1',
makeStream(msgs),
conn,
new AbortController().signal,
{},
)
expect(result.stopReason).toBe('max_tokens')
})
test('success with max_tokens stop_reason falls back to end_turn when isError', async () => {
// Audit §3.3: in the success branch, isError acts as a last-resort
// override to end_turn per the merged fix diff.
const conn = makeConn()
const msgs: SDKMessage[] = [
{
type: 'result',
subtype: 'success',
is_error: true,
result: '',
stop_reason: 'max_tokens',
} as unknown as SDKMessage,
]
const result = await forwardSessionUpdates(
's1',
makeStream(msgs),
conn,
new AbortController().signal,
{},
)
expect(result.stopReason).toBe('end_turn')
})
test('maps error_during_execution with max_tokens stop_reason', async () => {
// Audit §3.4: error_during_execution branch must preserve max_tokens even
// when isError is set (mutually exclusive branches).
const conn = makeConn()
const msgs: SDKMessage[] = [
{
type: 'result',
subtype: 'error_during_execution',
is_error: true,
result: '',
stop_reason: 'max_tokens',
} as unknown as SDKMessage,
]
const result = await forwardSessionUpdates(
's1',
makeStream(msgs),
conn,
new AbortController().signal,
{},
)
expect(result.stopReason).toBe('max_tokens')
})
test('maps error_during_execution without max_tokens to end_turn', async () => {
const conn = makeConn()
const msgs: SDKMessage[] = [
{
type: 'result',
subtype: 'error_during_execution',
is_error: true,
result: '',
stop_reason: 'end_turn',
} as unknown as SDKMessage,
]
const result = await forwardSessionUpdates(
's1',
makeStream(msgs),
conn,
new AbortController().signal,
{},
)
expect(result.stopReason).toBe('end_turn')
})
test('compact_boundary emits completion message without usage_update', async () => {
// After audit §4.1, compact_boundary still sends the "Compacting completed."
// agent_message_chunk but no longer emits the unstable usage_update
// notification.
const conn = makeConn()
const msgs: SDKMessage[] = [
{ type: 'system', subtype: 'compact_boundary' } as unknown as SDKMessage,
@@ -1112,15 +1444,14 @@ describe('forwardSessionUpdates', () => {
'sessionUpdate'
] === 'usage_update',
)
expect(usageCall).toBeDefined()
expect(
(
(usageCall![0] as Record<string, unknown>).update as Record<
string,
unknown
>
).used,
).toBe(0)
expect(usageCall).toBeUndefined()
const messageCall = calls.find(
(c: unknown[]) =>
((c[0] as Record<string, Record<string, unknown>>).update ?? {})[
'sessionUpdate'
] === 'agent_message_chunk',
)
expect(messageCall).toBeDefined()
})
test('ignores unknown message types without crashing', async () => {

View File

@@ -234,7 +234,7 @@ describe('createAcpCanUseTool', () => {
}
})
test('options include allow always, allow once, and reject once', async () => {
test('options include allow always, allow once, reject once, and reject always', async () => {
const conn = makeConn({ outcome: { outcome: 'cancelled' } })
const canUseTool = createAcpCanUseTool(conn, 'sess-3', () => 'default')
await canUseTool(makeTool('Write'), {}, dummyContext, dummyMsg, 'tu_8')
@@ -245,6 +245,7 @@ describe('createAcpCanUseTool', () => {
expect(opts.find(option => option.kind === 'allow_always')).toBeTruthy()
expect(opts.find(option => option.kind === 'allow_once')).toBeTruthy()
expect(opts.find(option => option.kind === 'reject_once')).toBeTruthy()
expect(opts.find(option => option.kind === 'reject_always')).toBeTruthy()
})
test('ExitPlanMode omits bypass option when the session does not expose it', async () => {
@@ -332,4 +333,92 @@ describe('createAcpCanUseTool', () => {
(conn.sessionUpdate as ReturnType<typeof mock>).mock.calls,
).toHaveLength(0)
})
test('checkTerminalOutput honors standard clientCapabilities.terminal', async () => {
// Standard ACP v1 client advertises terminal: true without any _meta hint.
const conn = makeConn({ outcome: { outcome: 'cancelled' } })
const capabilities = { terminal: true } as any
const canUseTool = createAcpCanUseTool(
conn,
'sess-term',
() => 'default',
capabilities,
)
await canUseTool(makeTool('Bash'), {}, dummyContext, dummyMsg, 'tu_term')
const { toolCall } = (conn.requestPermission as ReturnType<typeof mock>)
.mock.calls[0][0] as Record<string, unknown>
// toolInfoFromToolUse is mocked; we only assert the standard capability is
// respected (no crash, request delegated). The legacy _meta path is
// exercised separately below.
expect(toolCall).toBeDefined()
})
test('checkTerminalOutput falls back to legacy _meta.terminal_output', async () => {
const conn = makeConn({ outcome: { outcome: 'cancelled' } })
const capabilities = { _meta: { terminal_output: true } } as any
const canUseTool = createAcpCanUseTool(
conn,
'sess-term-legacy',
() => 'default',
capabilities,
)
await canUseTool(makeTool('Bash'), {}, dummyContext, dummyMsg, 'tu_term2')
expect(
(conn.requestPermission as ReturnType<typeof mock>).mock.calls,
).toHaveLength(1)
})
test('cancelled permission outcome invokes onPermissionCancelled callback', async () => {
const conn = makeConn({ outcome: { outcome: 'cancelled' } })
const onPermissionCancelled = mock(() => {})
const canUseTool = createAcpCanUseTool(
conn,
'sess-cancel',
() => 'default',
undefined,
undefined,
undefined,
undefined,
onPermissionCancelled,
)
const result = await canUseTool(
makeTool('Bash'),
{},
dummyContext,
dummyMsg,
'tu_cancel',
)
expect(result.behavior).toBe('deny')
expect(onPermissionCancelled).toHaveBeenCalledTimes(1)
})
test('ExitPlanMode cancelled outcome invokes onPermissionCancelled callback', async () => {
const conn = makeConn({ outcome: { outcome: 'cancelled' } })
const onPermissionCancelled = mock(() => {})
const canUseTool = createAcpCanUseTool(
conn,
'sess-cancel-plan',
() => 'plan',
undefined,
undefined,
undefined,
undefined,
onPermissionCancelled,
)
const result = await canUseTool(
makeTool('ExitPlanMode'),
{},
dummyContext,
dummyMsg,
'tu_cancel_plan',
)
expect(result.behavior).toBe('deny')
expect(onPermissionCancelled).toHaveBeenCalledTimes(1)
})
})

View File

@@ -25,4 +25,31 @@ describe('promptToQueryInput', () => {
]),
).toBe('Resource link: name=Spec, uri=file:///tmp/spec.md')
})
test('renders BlobResource as a readable placeholder instead of dropping it', () => {
const result = promptToQueryInput([
{
type: 'resource',
resource: {
uri: 'file:///tmp/report.pdf',
mimeType: 'application/pdf',
blob: 'aGVsbG8=',
},
} as any,
])
expect(result).toContain('Embedded resource: file:///tmp/report.pdf')
expect(result).toContain('application/pdf')
expect(result).toContain('base64 blob')
})
test('BlobResource without mimeType or uri falls back to defaults', () => {
const result = promptToQueryInput([
{
type: 'resource',
resource: { blob: 'aGVsbG8=' },
} as any,
])
expect(result).toContain('(unknown uri)')
expect(result).toContain('application/octet-stream')
})
})

View File

@@ -40,6 +40,7 @@ import type {
} from '@agentclientprotocol/sdk'
import { randomUUID, type UUID } from 'node:crypto'
import { dirname } from 'node:path'
import * as path from 'node:path'
import type { Message } from '../../types/message.js'
import { deserializeMessages } from '../../utils/conversationRecovery.js'
import {
@@ -78,7 +79,11 @@ import {
} from './utils.js'
import { promptToQueryInput } from './promptConversion.js'
import { listSessionsImpl } from '../../utils/listSessionsImpl.js'
import { resolveSessionFilePath } from '../../utils/sessionStoragePortable.js'
import {
resolveSessionFilePath,
readSessionLite,
extractJsonStringField,
} from '../../utils/sessionStoragePortable.js'
import { getMainLoopModel } from '../../utils/model/model.js'
import { getModelOptions } from '../../utils/model/modelOptions.js'
import { getSettings_DEPRECATED } from '../../utils/settings/settings.js'
@@ -126,6 +131,9 @@ export class AcpAgent implements Agent {
return {
protocolVersion: 1,
// Explicit empty authMethods signals "no authentication required" to
// Clients rather than "capability unknown". Matches authenticate() no-op.
authMethods: [],
agentInfo: {
name: 'claude-code',
title: 'Claude Code',
@@ -150,10 +158,16 @@ export class AcpAgent implements Agent {
_meta: {
claudeCode: {
promptQueueing: true,
// session/fork is UNSTABLE — not part of stable v1 SessionCapabilities.
// Advertise via _meta namespace per extensibility.mdx "Advertising
// Custom Capabilities" instead of the standard sessionCapabilities map.
forkSession: true,
},
},
// image:false — promptToQueryInput() does not parse ContentBlock::Image
// blocks yet. Re-enable only after multimodal query input support lands.
promptCapabilities: {
image: true,
image: false,
embeddedContext: true,
},
mcpCapabilities: {
@@ -162,7 +176,6 @@ export class AcpAgent implements Agent {
},
loadSession: true,
sessionCapabilities: {
fork: {},
list: {},
resume: {},
close: {},
@@ -193,7 +206,11 @@ export class AcpAgent implements Agent {
async unstable_resumeSession(
params: ResumeSessionRequest,
): Promise<ResumeSessionResponse> {
const result = await this.getOrCreateSession(params)
// Per session-setup.mdx "Resuming a Session": the Agent MUST NOT replay the
// conversation history via session/update notifications before responding.
// Only restore context + MCP connections, then return immediately. This
// differs from session/load which DOES replay history.
const result = await this.getOrCreateSession({ ...params, replay: false })
this.scheduleAvailableCommandsUpdate(result.sessionId)
return result
}
@@ -211,18 +228,29 @@ export class AcpAgent implements Agent {
async listSessions(
params: ListSessionsRequest,
): Promise<ListSessionsResponse> {
// Pagination is not implemented: we always return all available sessions
// for the requested cwd (no nextCursor). Per session-list.mdx the Agent
// SHOULD return an error if the cursor is invalid, so explicitly reject
// any client-supplied cursor rather than silently accepting it.
if (params.cursor !== undefined && params.cursor !== null) {
throw new Error(
'Pagination cursor not supported: listSessions returns all results in a single page.',
)
}
const candidates = await listSessionsImpl({
dir: params.cwd ?? undefined,
limit: 100,
})
const sessions = []
for (const candidate of candidates) {
if (!candidate.cwd) continue
// Only include title when non-empty; schema allows null/omitted title.
const title = sanitizeTitle(candidate.summary ?? '')
sessions.push({
sessionId: candidate.sessionId,
cwd: candidate.cwd,
title: sanitizeTitle(candidate.summary ?? ''),
...(title ? { title } : {}),
updatedAt: new Date(candidate.lastModified).toISOString(),
})
}
@@ -235,11 +263,26 @@ export class AcpAgent implements Agent {
async unstable_forkSession(
params: ForkSessionRequest,
): Promise<ForkSessionResponse> {
const response = await this.createSession({
cwd: params.cwd,
mcpServers: params.mcpServers ?? [],
_meta: params._meta,
})
// Load the source session's messages so the fork actually branches from
// the source conversation rather than starting a blank session. Per the
// unstable ForkSessionRequest, params.sessionId is the ID to fork from.
let initialMessages: Message[] | undefined
try {
const log = await getLastSessionLog(params.sessionId as UUID)
if (log && log.messages.length > 0) {
initialMessages = deserializeMessages(log.messages)
}
} catch (err) {
console.error('[ACP] fork source load failed:', err)
}
const response = await this.createSession(
{
cwd: params.cwd,
mcpServers: params.mcpServers ?? [],
_meta: params._meta,
},
{ initialMessages },
)
this.scheduleAvailableCommandsUpdate(response.sessionId)
return response
}
@@ -268,8 +311,11 @@ export class AcpAgent implements Agent {
// Extract text/image content from the prompt
const promptInput = promptToQueryInput(params.prompt)
// Per prompt-turn.mdx, `prompt` is a required ContentBlock[] and an
// effectively-empty prompt is malformed input — reject it with an
// invalid_params error rather than fabricating a successful end_turn.
if (!promptInput.trim()) {
return { stopReason: 'end_turn' }
throw new Error('Prompt content is empty')
}
const promptCancelGeneration = session.cancelGeneration
@@ -323,24 +369,51 @@ export class AcpAgent implements Agent {
return { stopReason: 'cancelled' }
}
return {
stopReason,
usage: usage
? {
inputTokens: usage.inputTokens,
outputTokens: usage.outputTokens,
cachedReadTokens: usage.cachedReadTokens,
cachedWriteTokens: usage.cachedWriteTokens,
totalTokens:
usage.inputTokens +
usage.outputTokens +
usage.cachedReadTokens +
usage.cachedWriteTokens,
}
: undefined,
// Emit a session_info_update so Clients learn the session's display
// title / last-activity timestamp via the stable v1 session/update
// channel. The title is derived from the first user prompt.
await this.maybeEmitSessionInfoUpdate(params.sessionId, promptInput)
// Per extensibility.mdx:39 the root of PromptResponse is reserved —
// stable v1 defines only `stopReason` (+ optional `_meta`). Token usage
// is therefore carried under the `_meta.claudeCode.usage` extension
// namespace rather than as a non-spec root field. thoughtTokens are
// included in totalTokens so reported totals match billable tokens;
// until bridge.ts tracks them they are reported as 0.
if (usage) {
const thoughtTokens = 0
return {
stopReason,
_meta: {
claudeCode: {
usage: {
inputTokens: usage.inputTokens,
outputTokens: usage.outputTokens,
cachedReadTokens: usage.cachedReadTokens,
cachedWriteTokens: usage.cachedWriteTokens,
thoughtTokens,
totalTokens:
usage.inputTokens +
usage.outputTokens +
usage.cachedReadTokens +
usage.cachedWriteTokens +
thoughtTokens,
},
},
},
}
}
return { stopReason }
} catch (err: unknown) {
if (session.cancelled) {
// Treat AbortError / cancellation-shaped errors as a turn cancellation
// regardless of the session.cancelled flag, to close the race window
// between interrupt() firing and cancel() setting the flag. Per
// prompt-turn.mdx the Agent MUST return `cancelled` for aborts.
const isAbort =
err instanceof Error &&
(err.name === 'AbortError' ||
/abort|cancelled|interrupt/i.test(err.message))
if (session.cancelled || isAbort) {
return { stopReason: 'cancelled' }
}
@@ -402,6 +475,17 @@ export class AcpAgent implements Agent {
}
this.applySessionMode(params.sessionId, params.modeId)
// Per session-modes.mdx: when the Agent changes its own mode it MUST send
// a current_mode_update notification so mode-only Clients learn the
// switch. Mirrors the current_mode_update sent by setSessionConfigOption
// when configId === 'mode'.
await this.conn.sessionUpdate({
sessionId: params.sessionId,
update: {
sessionUpdate: 'current_mode_update',
currentModeId: params.modeId,
},
})
await this.updateConfigOption(params.sessionId, 'mode', params.modeId)
return {}
}
@@ -442,6 +526,21 @@ export class AcpAgent implements Agent {
throw new Error(`Unknown config option: ${params.configId}`)
}
// Per session-config-options.mdx: value MUST be one of the values listed
// in the option's options array. Reject unknown values with an error
// rather than silently persisting them. Only `select` options carry an
// options array; `boolean` options have no enumerated values.
if (option.type === 'select') {
const validValues = flattenConfigOptionValues(
(option as { options?: unknown }).options,
)
if (!validValues.includes(params.value)) {
throw new Error(
`Invalid value '${params.value}' for config option ${params.configId}; must be one of: ${validValues.join(', ')}`,
)
}
}
const value = params.value
if (params.configId === 'mode') {
@@ -672,9 +771,10 @@ export class AcpAgent implements Agent {
this.sessions.set(sessionId, session)
// Stable v1 NewSessionResponse only defines sessionId/modes/configOptions.
// `models` is a draft/unstable field — omit it for v1 compliance.
return {
sessionId,
models,
modes,
configOptions,
}
@@ -690,7 +790,13 @@ export class AcpAgent implements Agent {
cwd: string
mcpServers?: NewSessionRequest['mcpServers']
_meta?: NewSessionRequest['_meta']
// replay:true (default, session/load) streams the conversation history back
// to the client via session/update. replay:false (session/resume) only
// restores the in-process context — per session-setup.mdx the Agent MUST
// NOT replay history when resuming.
replay?: boolean
}): Promise<NewSessionResponse> {
const shouldReplay = params.replay !== false
const existingSession = this.sessions.get(params.sessionId)
if (existingSession) {
const fingerprint = computeSessionFingerprint({
@@ -710,12 +816,13 @@ export class AcpAgent implements Agent {
)
setOriginalCwd(params.cwd)
await this.replaySessionHistory(params)
if (shouldReplay) {
await this.replaySessionHistory(params)
}
return {
sessionId: params.sessionId,
modes: existingSession.modes,
models: existingSession.models,
configOptions: existingSession.configOptions,
}
}
@@ -729,6 +836,25 @@ export class AcpAgent implements Agent {
// search by sessionId first and fall back to cwd-based lookup.
const resolved = await resolveSessionFilePath(params.sessionId, params.cwd)
const projectDir = resolved ? dirname(resolved.filePath) : null
// Per session-setup.mdx "Working Directory": the cwd MUST be the absolute
// path used for the session regardless of where the Agent was spawned.
// Reject cross-project loads where the persisted session's original cwd
// does not match the requested cwd, otherwise the client could load a
// session belonging to project B while passing project A's cwd.
if (resolved) {
const lite = await readSessionLite(resolved.filePath)
const originalCwd = lite && extractJsonStringField(lite.head, 'cwd')
if (
originalCwd &&
path.resolve(originalCwd) !== path.resolve(params.cwd)
) {
throw new Error(
`Session cwd mismatch: session belongs to ${originalCwd}, requested ${params.cwd}`,
)
}
}
switchSession(params.sessionId as SessionId, projectDir)
setOriginalCwd(params.cwd)
@@ -753,8 +879,8 @@ export class AcpAgent implements Agent {
{ sessionId: params.sessionId, initialMessages },
)
// Replay history to client if loaded
if (initialMessages && initialMessages.length > 0) {
// Replay history to client if loaded. session/resume skips this block.
if (shouldReplay && initialMessages && initialMessages.length > 0) {
const session = this.sessions.get(params.sessionId)
if (session) {
await replayHistoryMessages(
@@ -771,7 +897,6 @@ export class AcpAgent implements Agent {
return {
sessionId: response.sessionId,
modes: response.modes,
models: response.models,
configOptions: response.configOptions,
}
}
@@ -912,6 +1037,39 @@ export class AcpAgent implements Agent {
}, 0)
}
/**
* Emit a session_info_update notification carrying a derived session title
* (truncated first user prompt) and the current last-activity timestamp.
* Sent once per session — subsequent turns reuse the same title.
*/
private async maybeEmitSessionInfoUpdate(
sessionId: string,
firstPrompt: string,
): Promise<void> {
const session = this.sessions.get(sessionId)
if (!session) return
// sessionInfoTitleSent is tracked via toolUseCache to avoid reshaping
// AcpSession; use a dedicated per-session flag instead.
const cache = session.toolUseCache as ToolUseCache & {
__sessionInfoTitleSent?: boolean
}
if (cache.__sessionInfoTitleSent) return
cache.__sessionInfoTitleSent = true
const title = sanitizeTitle(firstPrompt).slice(0, 100)
try {
await this.conn.sessionUpdate({
sessionId,
update: {
sessionUpdate: 'session_info_update',
...(title ? { title } : {}),
updatedAt: new Date().toISOString(),
},
})
} catch (err) {
console.error('[ACP] Failed to send session_info_update:', err)
}
}
/** Read a setting from Claude config (simplified — no file watching) */
private getSetting<T>(key: string): T | undefined {
const settings = getSettings_DEPRECATED() as Record<string, unknown>
@@ -1036,6 +1194,37 @@ function isTruthyEnv(value: string | undefined): boolean {
return value === '1' || value?.toLowerCase() === 'true'
}
/**
* Flatten a SessionConfigOption's `options` (which may be flat
* SessionConfigSelectOption entries or grouped SessionConfigSelectGroup
* entries) into a list of valid value strings. Used to validate that a
* setSessionConfigOption value is one of the listed options.
*/
function flattenConfigOptionValues(options: unknown): string[] {
const values: string[] = []
if (!Array.isArray(options)) return values
for (const opt of options) {
if (typeof opt !== 'object' || opt === null) continue
const maybeGroup = opt as { group?: unknown; options?: unknown[] }
if (Array.isArray(maybeGroup.options)) {
// SessionConfigSelectGroup — recurse into its options
for (const inner of maybeGroup.options) {
if (
inner &&
typeof inner === 'object' &&
typeof (inner as { value?: unknown }).value === 'string'
) {
values.push((inner as { value: string }).value)
}
}
} else if (typeof (opt as { value?: unknown }).value === 'string') {
// SessionConfigSelectOption
values.push((opt as { value: string }).value)
}
}
return values
}
function popNextPendingPrompt(session: AcpSession): PendingPrompt | undefined {
while (session.pendingQueueHead < session.pendingQueue.length) {
const nextId = session.pendingQueue[session.pendingQueueHead++]

View File

@@ -25,6 +25,22 @@ import type {
} from '@agentclientprotocol/sdk'
import type { SDKMessage } from '../../entrypoints/sdk/coreTypes.generated.js'
import { toDisplayPath, markdownEscape } from './utils.js'
import { isAbsolute, resolve } from 'node:path'
/**
* Normalises an emitted file path against the session cwd so that
* ToolCallLocation.path / Diff.path values are always absolute, as required
* by the ACP v1 spec (tool-calls.mdx:304-306; all file paths MUST be absolute).
* If no cwd is available, the original value is returned unchanged.
*/
function toAbsolutePath(
filePath: string | undefined,
cwd?: string,
): string | undefined {
if (!filePath) return undefined
if (!cwd) return filePath
return isAbsolute(filePath) ? filePath : resolve(cwd, filePath)
}
// ── ToolUseCache ──────────────────────────────────────────────────
@@ -235,7 +251,8 @@ export function toolInfoFromToolUse(
}
case 'Read': {
const filePath = (input?.file_path as string | undefined) ?? 'File'
const inputFilePath = input?.file_path as string | undefined
const filePath = inputFilePath ?? 'File'
const offset = input?.offset as number | undefined
const limit = input?.limit as number | undefined
let suffix = ''
@@ -245,10 +262,13 @@ export function toolInfoFromToolUse(
suffix = ` (from line ${offset})`
}
const displayPath = filePath ? toDisplayPath(filePath, cwd) : 'File'
const absReadPath = toAbsolutePath(inputFilePath, cwd)
return {
title: `Read ${displayPath}${suffix}`,
kind: 'read',
locations: filePath ? [{ path: filePath, line: offset ?? 1 }] : [],
locations: absReadPath
? [{ path: absReadPath, line: offset ?? 1 }]
: [],
content: [],
}
}
@@ -257,14 +277,15 @@ export function toolInfoFromToolUse(
const filePath = (input?.file_path as string | undefined) ?? ''
const content = (input?.content as string | undefined) ?? ''
const displayPath = filePath ? toDisplayPath(filePath, cwd) : undefined
const absWritePath = toAbsolutePath(filePath, cwd)
return {
title: displayPath ? `Write ${displayPath}` : 'Write',
kind: 'edit',
content: filePath
content: absWritePath
? [
{
type: 'diff' as const,
path: filePath,
path: absWritePath,
oldText: null,
newText: content,
},
@@ -275,7 +296,7 @@ export function toolInfoFromToolUse(
content: { type: 'text' as const, text: content },
},
],
locations: filePath ? [{ path: filePath }] : [],
locations: absWritePath ? [{ path: absWritePath }] : [],
}
}
@@ -284,26 +305,28 @@ export function toolInfoFromToolUse(
const oldString = (input?.old_string as string | undefined) ?? ''
const newString = (input?.new_string as string | undefined) ?? ''
const displayPath = filePath ? toDisplayPath(filePath, cwd) : undefined
const absEditPath = toAbsolutePath(filePath, cwd)
return {
title: displayPath ? `Edit ${displayPath}` : 'Edit',
kind: 'edit',
content: filePath
content: absEditPath
? [
{
type: 'diff' as const,
path: filePath,
path: absEditPath,
oldText: oldString || null,
newText: newString,
},
]
: [],
locations: filePath ? [{ path: filePath }] : [],
locations: absEditPath ? [{ path: absEditPath }] : [],
}
}
case 'Glob': {
const globPath = (input?.path as string | undefined) ?? ''
const pattern = (input?.pattern as string | undefined) ?? ''
const absGlobPath = toAbsolutePath(globPath, cwd)
let label = 'Find'
if (globPath) label += ` \`${globPath}\``
if (pattern) label += ` \`${pattern}\``
@@ -311,7 +334,7 @@ export function toolInfoFromToolUse(
title: label,
kind: 'search',
content: [],
locations: globPath ? [{ path: globPath }] : [],
locations: absGlobPath ? [{ path: absGlobPath }] : [],
}
}
@@ -599,6 +622,37 @@ function toAcpContentBlock(
: '[image: file reference]',
)
}
case 'resource_link': {
// ACP v1 ResourceLink requires name + uri. Name falls back to uri when
// absent so the client always has a display label. mimeType is optional.
const uri = content.uri as string | undefined
const name =
(content.name as string | undefined) ?? (uri as string | undefined)
return {
type: 'resource_link',
uri: uri as string,
name: name as string,
mimeType: content.mimeType as string | undefined,
}
}
case 'resource': {
// ACP v1 EmbeddedResource wraps an optional TextResource / BlobResource
// shape. Forward the standard fields the client knows how to render.
const r = content.resource as Record<string, unknown> | undefined
// Construct a TextResource or BlobResource payload depending on what is
// present. Cast through unknown because not every source shape satisfies
// the full union contract.
const resourcePayload = {
uri: (r?.uri as string | undefined) ?? '',
mimeType: r?.mimeType as string | null | undefined,
...(typeof r?.text === 'string' ? { text: r.text as string } : {}),
...(typeof r?.blob === 'string' ? { blob: r.blob as string } : {}),
}
return {
type: 'resource',
resource: resourcePayload,
} as unknown as ContentBlock
}
case 'tool_reference':
return wrapText(`Tool: ${content.tool_name as string}`)
case 'tool_search_tool_search_result': {
@@ -671,8 +725,15 @@ interface EditToolResponse {
* Builds diff ToolUpdate content from the structured Edit toolResponse.
* Parses structuredPatch hunks (lines prefixed with -, +, space) into
* oldText/newText diff pairs.
*
* The optional `cwd` is used to normalise the emitted path against the
* session cwd so that Diff.path / ToolCallLocation.path are absolute as
* required by the ACP v1 spec (audit §5.5).
*/
export function toolUpdateFromEditToolResponse(toolResponse: unknown): {
export function toolUpdateFromEditToolResponse(
toolResponse: unknown,
cwd?: string,
): {
content?: ToolCallContent[]
locations?: ToolCallLocation[]
} {
@@ -680,6 +741,8 @@ export function toolUpdateFromEditToolResponse(toolResponse: unknown): {
const response = toolResponse as EditToolResponse
if (!response.filePath || !Array.isArray(response.structuredPatch)) return {}
const absPath = toAbsolutePath(response.filePath, cwd) ?? response.filePath
const content: ToolCallContent[] = []
const locations: ToolCallLocation[] = []
@@ -697,10 +760,10 @@ export function toolUpdateFromEditToolResponse(toolResponse: unknown): {
}
}
if (oldText.length > 0 || newText.length > 0) {
locations.push({ path: response.filePath, line: newStart })
locations.push({ path: absPath, line: newStart })
content.push({
type: 'diff',
path: response.filePath,
path: absPath,
oldText: oldText.join('\n') || null,
newText: newText.join('\n'),
})
@@ -787,15 +850,10 @@ export async function forwardSessionUpdates(
if (subtype === 'compact_boundary') {
// Reset assistant usage tracking after compaction
lastAssistantTotalUsage = 0
// Send usage reset after compaction
await conn.sessionUpdate({
sessionId,
update: {
sessionUpdate: 'usage_update',
used: 0,
size: lastContextWindowSize,
},
})
// NOTE: usage_update is an UNSTABLE SessionUpdate discriminator (not in
// stable v1 schema). Token/cost info has no v1-stable carrier; we drop
// it from session/update and rely on PromptResponse._meta for clients
// that need it (see audit §4.1).
await conn.sessionUpdate({
sessionId,
update: {
@@ -830,28 +888,10 @@ export async function forwardSessionUpdates(
}
}
// Send usage_update — use lastAssistantTotalUsage if available
// (more accurate than accumulatedUsage which may include background tasks)
const usedTokens =
lastAssistantTotalUsage ??
accumulatedUsage.inputTokens +
accumulatedUsage.outputTokens +
accumulatedUsage.cachedReadTokens +
accumulatedUsage.cachedWriteTokens
const totalCostUsd = msg.total_cost_usd
await conn.sessionUpdate({
sessionId,
update: {
sessionUpdate: 'usage_update',
used: usedTokens,
size: lastContextWindowSize,
cost:
totalCostUsd != null
? { amount: totalCostUsd, currency: 'USD' }
: undefined,
},
})
// NOTE: usage_update was removed — it is an UNSTABLE SessionUpdate
// discriminator not present in the stable v1 schema (audit §4.1). Token
// and cost information is returned via PromptResponse._meta.claudeCode.usage
// instead.
// Determine stop reason
const subtype = msg.subtype
@@ -864,21 +904,24 @@ export async function forwardSessionUpdates(
switch (subtype) {
case 'success': {
const stopReasonStr = msg.stop_reason
if (stopReasonStr === 'max_tokens') {
stopReason = 'max_tokens'
}
if (isError) {
// Report error as end_turn
stopReason = 'end_turn'
}
// Map Anthropic stop_reason to ACP StopReason. Branches are mutually
// exclusive so a max_tokens termination that is also flagged isError
// no longer silently flips to end_turn (audit §3.3, §3.4). refusal
// (safety refusal) is a first-class ACP stop reason that must surface
// to the client instead of being misreported as end_turn.
const r = msg.stop_reason
if (r === 'max_tokens') stopReason = 'max_tokens'
else if (r === 'refusal') stopReason = 'refusal'
else stopReason = 'end_turn'
if (isError) stopReason = 'end_turn'
break
}
case 'error_during_execution': {
// Mutually exclusive: max_tokens wins when reported, otherwise the
// error path falls back to end_turn. Avoids the prior two-if
// sequence that overwrote max_tokens with end_turn (audit §3.4).
if (msg.stop_reason === 'max_tokens') {
stopReason = 'max_tokens'
} else if (isError) {
stopReason = 'end_turn'
} else {
stopReason = 'end_turn'
}
@@ -1021,14 +1064,8 @@ export async function forwardSessionUpdates(
// ── Compact boundary ───────────────────────────────────────
case 'compact_boundary': {
lastAssistantTotalUsage = 0
await conn.sessionUpdate({
sessionId,
update: {
sessionUpdate: 'usage_update',
used: 0,
size: lastContextWindowSize,
},
})
// NOTE: usage_update removed — UNSTABLE discriminator, not in v1 stable
// schema (audit §4.1). Token info flows through PromptResponse._meta.
await conn.sessionUpdate({
sessionId,
update: {

View File

@@ -37,6 +37,15 @@ export function createAcpCanUseTool(
cwd?: string,
onModeChange?: (modeId: string) => void,
isBypassModeAvailable?: () => boolean,
/**
* Invoked when the ACP client returns a `cancelled` permission outcome.
* The Agent uses this to set the session-level cancelled flag and interrupt
* the running query so session/prompt resolves with StopReason::Cancelled
* (schema.json:629) instead of treating the cancellation as a plain deny.
* Optional for backwards compatibility with callers that have not been
* wired up yet.
*/
onPermissionCancelled?: () => void,
): CanUseToolFn {
return async (
tool: ToolType,
@@ -64,6 +73,7 @@ export function createAcpCanUseTool(
cwd,
onModeChange,
isBypassModeAvailable,
onPermissionCancelled,
)
}
@@ -124,6 +134,11 @@ export function createAcpCanUseTool(
{ kind: 'allow_always', name: 'Always Allow', optionId: 'allow_always' },
{ kind: 'allow_once', name: 'Allow', optionId: 'allow' },
{ kind: 'reject_once', name: 'Reject', optionId: 'reject' },
{
kind: 'reject_always',
name: 'Always Reject',
optionId: 'reject_always',
},
]
try {
@@ -134,10 +149,15 @@ export function createAcpCanUseTool(
})
if (response.outcome.outcome === 'cancelled') {
// Per schema.json:629, a cancelled permission outcome means the prompt
// turn was cancelled. Signal the session so prompt() resolves with
// StopReason::Cancelled instead of treating this as a normal denial.
onPermissionCancelled?.()
return {
behavior: 'deny',
message: 'Permission request cancelled by client',
decisionReason: { type: 'mode', mode: 'default' },
toolUseID,
}
}
@@ -181,6 +201,7 @@ async function handleExitPlanMode(
cwd?: string,
onModeChange?: (modeId: string) => void,
isBypassModeAvailable?: () => boolean,
onPermissionCancelled?: () => void,
): Promise<PermissionAllowDecision | PermissionDenyDecision> {
const options: Array<PermissionOption> = [
{
@@ -229,6 +250,8 @@ async function handleExitPlanMode(
})
if (response.outcome.outcome === 'cancelled') {
// Propagate cancellation so prompt() resolves with StopReason::Cancelled.
onPermissionCancelled?.()
return {
behavior: 'deny',
message: 'Tool use aborted',
@@ -279,6 +302,11 @@ async function handleExitPlanMode(
function checkTerminalOutput(clientCapabilities?: ClientCapabilities): boolean {
if (!clientCapabilities) return false
// Standard ACP v1 capability: ClientCapabilities.terminal (boolean).
if (clientCapabilities.terminal === true) return true
// Legacy Claude-Code clients advertised terminal support via _meta before
// the standard `terminal` boolean existed. `_meta` is reserved per the spec,
// but we keep this fallback for backward compatibility with older clients.
const meta = (clientCapabilities as unknown as Record<string, unknown>)._meta
if (!meta || typeof meta !== 'object') return false
return (meta as Record<string, unknown>)['terminal_output'] === true

View File

@@ -20,6 +20,20 @@ export function promptToQueryInput(
const resource = b.resource as Record<string, unknown> | undefined
if (resource && typeof resource.text === 'string') {
parts.push(resource.text)
} else if (resource && typeof resource.blob === 'string') {
// BlobResource (e.g. PDF/binary): query input is string-only, so emit a
// readable placeholder instead of silently dropping the content. Ideally
// this would be decoded and passed as a binary content block once the
// query layer supports multimodal input.
const mt =
typeof resource.mimeType === 'string'
? resource.mimeType
: 'application/octet-stream'
const uri =
typeof resource.uri === 'string' ? resource.uri : '(unknown uri)'
parts.push(
`Embedded resource: ${uri} (${mt}, base64 blob, ${resource.blob.length} chars)`,
)
}
}
}