refactor(mcp): extract connection/manager modules, add discovery caching and tests

src/acp/validation.ts

@@ -0,0 +1,113 @@
+export const MAX_JSON_PAYLOAD_BYTES = 1024 * 1024
+export const MAX_JSON_NESTING_DEPTH = 10
+
+export type JsonPayloadBudgetErrorData = {
+  kind: 'payload_too_large' | 'payload_too_deep' | 'payload_not_serializable'
+  retryable: false
+  label: string
+  sizeBytes?: number
+  maxBytes?: number
+  depth?: number
+  maxDepth?: number
+}
+
+export class JsonPayloadBudgetError extends Error {
+  readonly code = -32602
+  readonly data: JsonPayloadBudgetErrorData
+
+  constructor(message: string, data: JsonPayloadBudgetErrorData) {
+    super(message)
+    this.name = 'JsonPayloadBudgetError'
+    this.data = data
+  }
+}
+
+function getJsonNestingDepth(
+  value: unknown,
+  seen: WeakSet<object> = new WeakSet(),
+): number {
+  if (value === null || typeof value !== 'object') return 0
+
+  if (seen.has(value)) {
+    throw new JsonPayloadBudgetError('JSON payload is not serializable', {
+      kind: 'payload_not_serializable',
+      retryable: false,
+      label: 'payload',
+    })
+  }
+  seen.add(value)
+
+  try {
+    const children = Array.isArray(value)
+      ? value
+      : Object.values(value as Record<string, unknown>)
+
+    if (children.length === 0) return 1
+    return (
+      1 + Math.max(...children.map(child => getJsonNestingDepth(child, seen)))
+    )
+  } finally {
+    seen.delete(value)
+  }
+}
+
+export function getJsonPayloadBudget(value: unknown): {
+  sizeBytes: number
+  depth: number
+} {
+  let serialized: string
+  try {
+    serialized = JSON.stringify(value) ?? 'null'
+  } catch {
+    throw new JsonPayloadBudgetError('JSON payload is not serializable', {
+      kind: 'payload_not_serializable',
+      retryable: false,
+      label: 'payload',
+    })
+  }
+
+  return {
+    sizeBytes: Buffer.byteLength(serialized, 'utf8'),
+    depth: getJsonNestingDepth(value),
+  }
+}
+
+export function assertJsonPayloadBudget(
+  value: unknown,
+  options?: {
+    label?: string
+    maxBytes?: number
+    maxDepth?: number
+  },
+): void {
+  const label = options?.label ?? 'payload'
+  const maxBytes = options?.maxBytes ?? MAX_JSON_PAYLOAD_BYTES
+  const maxDepth = options?.maxDepth ?? MAX_JSON_NESTING_DEPTH
+  const budget = getJsonPayloadBudget(value)
+
+  if (budget.sizeBytes > maxBytes) {
+    throw new JsonPayloadBudgetError(
+      `${label} exceeds maximum serialized size of ${maxBytes} bytes`,
+      {
+        kind: 'payload_too_large',
+        retryable: false,
+        label,
+        sizeBytes: budget.sizeBytes,
+        maxBytes,
+      },
+    )
+  }
+
+  if (budget.depth > maxDepth) {
+    throw new JsonPayloadBudgetError(
+      `${label} exceeds maximum nesting depth of ${maxDepth}`,
+      {
+        kind: 'payload_too_deep',
+        retryable: false,
+        label,
+        depth: budget.depth,
+        maxDepth,
+      },
+    )
+  }
+}

src/app/binaryFeedback.ts

@@ -1,4 +1,8 @@
-import { TextBlock, ToolUseBlock } from '@anthropic-ai/sdk/resources/index.mjs'
+import type {
+  ContentBlock,
+  TextBlock,
+  ToolUseBlock,
+} from '@anthropic-ai/sdk/resources/index.mjs'
 import type { AssistantMessage, BinaryFeedbackResult } from './query'
 import { isEqual, zip } from 'lodash-es'
 
@@ -30,23 +34,25 @@ function textContentBlocksEqual(cb1: TextBlock, cb2: TextBlock): boolean {
   return cb1.text === cb2.text
 }
 
-function contentBlocksEqual(
-  cb1: TextBlock | ToolUseBlock,
-  cb2: TextBlock | ToolUseBlock,
-): boolean {
+function contentBlocksEqual(cb1: ContentBlock, cb2: ContentBlock): boolean {
   if (cb1.type !== cb2.type) {
     return false
   }
   if (cb1.type === 'text') {
     return textContentBlocksEqual(cb1, cb2 as TextBlock)
   }
-  cb2 = cb2 as ToolUseBlock
-  return cb1.name === cb2.name && isEqual(cb1.input, cb2.input)
+  if (cb1.type === 'tool_use') {
+    const toolUseBlock = cb2 as ToolUseBlock
+    return (
+      cb1.name === toolUseBlock.name && isEqual(cb1.input, toolUseBlock.input)
+    )
+  }
+  return isEqual(cb1, cb2)
 }
 
 function allContentBlocksEqual(
-  content1: (TextBlock | ToolUseBlock)[],
-  content2: (TextBlock | ToolUseBlock)[],
+  content1: ContentBlock[],
+  content2: ContentBlock[],
 ): boolean {
   if (content1.length !== content2.length) {
     return false

src/commands/agents/generation.ts

@@ -2,6 +2,7 @@ import { randomUUID } from 'crypto'
 import type { AgentConfig } from '@utils/agent/loader'
 import { debug as debugLogger } from '@utils/log/debugLogger'
 import { logError } from '@utils/log'
+import { extractTextFromContent } from '@utils/ai/anthropic'
 
 export type GeneratedAgent = {
   identifier: string
@@ -33,17 +34,7 @@ Make the agent highly specialized and effective for the described use case.`
     ] as any
     const response = await queryModel('main', messages, [systemPrompt])
 
-    let responseText = ''
-    if (typeof response.message?.content === 'string') {
-      responseText = response.message.content
-    } else if (Array.isArray(response.message?.content)) {
-      const textContent = response.message.content.find(
-        (c: any) => c.type === 'text',
-      )
-      responseText = textContent?.text || ''
-    } else if (response.message?.content?.[0]?.text) {
-      responseText = response.message.content[0].text
-    }
+    const responseText = extractTextFromContent(response.message?.content) ?? ''
 
     if (!responseText) {
       throw new Error('No text content in model response')

src/commands/compact.ts

@@ -9,6 +9,7 @@ import { getCodeStyle } from '@utils/config/style'
 import { clearTerminal } from '@utils/terminal'
 import { resetReminderSession } from '@services/systemReminder'
 import { resetFileFreshnessSession } from '@services/fileFreshness'
+import { createAnthropicUsage } from '@utils/ai/anthropic'
 
 const COMPRESSION_PROMPT = `Please provide a comprehensive summary of our conversation structured as follows:
 
@@ -88,12 +89,9 @@ const compact = {
       throw new Error(summary)
     }
 
-    summaryResponse.message.usage = {
-      input_tokens: 0,
+    summaryResponse.message.usage = createAnthropicUsage({
       output_tokens: summaryResponse.message.usage.output_tokens,
-      cache_creation_input_tokens: 0,
-      cache_read_input_tokens: 0,
-    }
+    })
 
     await clearTerminal()
     getMessagesSetter()([])

src/entrypoints/cli.tsx

@@ -11,8 +11,5 @@ initSentry()
 ensurePackagedRuntimeEnv()
 ensureYogaWasmPath(import.meta.url)
 
-import * as dontcare from '@anthropic-ai/sdk/shims/node'
-Object.keys(dontcare)
-
 installProcessHandlers()
 void runCli()

src/entrypoints/mcp.ts

@@ -19,6 +19,7 @@ import { Command } from '@commands'
 import review from '@commands/review'
 import { lastX } from '@utils/text/generators'
 import { MACRO } from '@constants/macros'
+import { assertJsonPayloadBudget } from '../acp/validation'
 type ToolInput = Record<string, unknown>
 
 const state: {
@@ -74,6 +75,10 @@ export async function startMCPServer(cwd: string): Promise<void> {
       }
 
       try {
+        assertJsonPayloadBudget(args ?? {}, {
+          label: `MCP tool ${name} arguments`,
+        })
+
         if (!(await tool.isEnabled())) {
           throw new Error(`Tool ${name} is not enabled`)
         }

src/services/ai/adapters/responsesStreaming.ts

@@ -1,6 +1,7 @@
 import { StreamingEvent } from './base'
 import { AssistantMessage } from '@query'
 import { setRequestStatus } from '@utils/session/requestStatus'
+import { createAnthropicUsage } from '@utils/ai/anthropic'
 
 export async function processResponsesStream(
   stream: AsyncGenerator<StreamingEvent>,
@@ -74,9 +75,16 @@ export async function processResponsesStream(
   const assistantMessage: AssistantMessage = {
     type: 'assistant',
     message: {
+      id: responseId,
+      container: null,
+      model: '<responses-stream>',
       role: 'assistant',
       content: contentBlocks,
-      usage: {
+      stop_details: null,
+      stop_reason: 'end_turn',
+      stop_sequence: null,
+      type: 'message',
+      usage: createAnthropicUsage({
         input_tokens: usage.prompt_tokens ?? 0,
         output_tokens: usage.completion_tokens ?? 0,
         prompt_tokens: usage.prompt_tokens ?? 0,
@@ -85,7 +93,7 @@ export async function processResponsesStream(
           usage.totalTokens ??
           (usage.prompt_tokens || 0) + (usage.completion_tokens || 0),
         reasoningTokens: usage.reasoningTokens,
-      },
+      }),
     },
     costUSD: 0,
     durationMs: Date.now() - startTime,

src/services/ai/llm.ts

@@ -1,4 +1,3 @@
-import '@anthropic-ai/sdk/shims/node'
 import Anthropic, { APIConnectionError, APIError } from '@anthropic-ai/sdk'
 import { StreamingEvent } from './adapters/base'
 import { AnthropicBedrock } from '@anthropic-ai/bedrock-sdk'
@@ -74,6 +73,10 @@ import {
   NO_CONTENT_MESSAGE,
   PROMPT_TOO_LONG_ERROR_MESSAGE,
 } from './llmConstants'
+import {
+  createAnthropicUsage,
+  normalizeAnthropicUsage,
+} from '@utils/ai/anthropic'
 
 function isGPT5Model(modelName: string): boolean {
   return modelName.startsWith('gpt-5')
@@ -613,6 +616,7 @@ function convertOpenAIResponseToAnthropic(
         input: toolArgs,
         name: toolName,
         id: toolCall.id?.length > 0 ? toolCall.id : nanoid(),
+        caller: { type: 'direct' },
       })
     }
   }
@@ -1417,13 +1421,15 @@ async function queryAnthropicNative(
         return {
           type: 'text' as const,
           text: block.text,
+          citations: block.citations ?? null,
         }
       } else if (block.type === 'tool_use') {
         return {
           type: 'tool_use' as const,
           id: block.id,
           name: block.name,
           input: block.input,
+          caller: block.caller,
         }
       }
       return block
@@ -1432,9 +1438,11 @@ async function queryAnthropicNative(
     const assistantMessage: AssistantMessage = {
       message: {
         id: response.id,
+        container: (response as any).container ?? null,
         content,
         model: response.model,
         role: 'assistant',
+        stop_details: (response as any).stop_details ?? null,
         stop_reason: response.stop_reason,
         stop_sequence: response.stop_sequence,
         type: 'message',
@@ -1909,16 +1917,27 @@ function buildAssistantMessageFromUnifiedResponse(
         input: toolArgs,
         name: toolName,
         id: toolCall.id?.length > 0 ? toolCall.id : nanoid(),
+        caller: { type: 'direct' },
       })
     }
   }
 
   return {
     type: 'assistant',
     message: {
+      id:
+        unifiedResponse.responseId ??
+        unifiedResponse.id ??
+        `resp_${Date.now()}`,
+      container: null,
+      model: unifiedResponse.model ?? '<unified>',
       role: 'assistant',
       content: contentBlocks,
-      usage: {
+      stop_details: null,
+      stop_reason: unifiedResponse.stopReason ?? 'end_turn',
+      stop_sequence: null,
+      type: 'message',
+      usage: createAnthropicUsage({
         input_tokens:
           unifiedResponse.usage?.promptTokens ??
           unifiedResponse.usage?.input_tokens ??
@@ -1951,7 +1970,7 @@ function buildAssistantMessageFromUnifiedResponse(
             (unifiedResponse.usage?.completionTokens ??
               unifiedResponse.usage?.output_tokens ??
               0),
-      },
+      }),
     },
     costUSD: 0,
     durationMs: Date.now() - startTime,
@@ -1961,34 +1980,7 @@ function buildAssistantMessageFromUnifiedResponse(
 }
 
 function normalizeUsage(usage?: any) {
-  if (!usage) {
-    return {
-      input_tokens: 0,
-      output_tokens: 0,
-      cache_read_input_tokens: 0,
-      cache_creation_input_tokens: 0,
-    }
-  }
-
-  const inputTokens =
-    usage.input_tokens ?? usage.prompt_tokens ?? usage.inputTokens ?? 0
-  const outputTokens =
-    usage.output_tokens ?? usage.completion_tokens ?? usage.outputTokens ?? 0
-  const cacheReadInputTokens =
-    usage.cache_read_input_tokens ??
-    usage.prompt_token_details?.cached_tokens ??
-    usage.cacheReadInputTokens ??
-    0
-  const cacheCreationInputTokens =
-    usage.cache_creation_input_tokens ?? usage.cacheCreatedInputTokens ?? 0
-
-  return {
-    ...usage,
-    input_tokens: inputTokens,
-    output_tokens: outputTokens,
-    cache_read_input_tokens: cacheReadInputTokens,
-    cache_creation_input_tokens: cacheCreationInputTokens,
-  }
+  return normalizeAnthropicUsage(usage)
 }
 
 function getModelInputTokenCostUSD(model: string): number {