Function classifyYoloAction

src/utils/permissions/yoloClassifier.ts:1012–1306

(
  messages: Message[],
  action: TranscriptEntry,
  tools: Tools,
  context: ToolPermissionContext,
  signal: AbortSignal,
)

 * @param signal - Abort signal
 */
export async function classifyYoloAction(
  messages: Message[],
  action: TranscriptEntry,
  tools: Tools,
  context: ToolPermissionContext,
  signal: AbortSignal,
): Promise<YoloClassifierResult> {
  const lookup = buildToolLookup(tools)
  const actionCompact = toCompact(action, lookup)
  // '' = "no security relevance" (Tool.toAutoClassifierInput contract). Without
  // this guard the empty action block + cache_control below hits an API 400.
  if (actionCompact === '') {
    return {
      shouldBlock: false,
      reason: 'Tool declares no classifier-relevant input',
      model: getClassifierModel(),
    }
  }

  const systemPrompt = await buildYoloSystemPrompt(context)
  const transcriptEntries = buildTranscriptEntries(messages)
  const claudeMdMessage = buildClaudeMdMessage()
  const prefixMessages: Anthropic.MessageParam[] = claudeMdMessage
    ? [claudeMdMessage]
    : []

  let toolCallsLength = actionCompact.length
  let userPromptsLength = 0
  const userContentBlocks: Anthropic.TextBlockParam[] = []
  for (const entry of transcriptEntries) {
    for (const block of entry.content) {
      const serialized = toCompactBlock(block, entry.role, lookup)
      if (serialized === '') continue
      switch (entry.role) {
        case 'user':
          userPromptsLength += serialized.length
          break
        case 'assistant':
          toolCallsLength += serialized.length
          break
        default: {
          const _exhaustive: never = entry.role
          void _exhaustive
        }
      }
      userContentBlocks.push({ type: 'text' as const, text: serialized })
    }
  }

  const userPrompt = userContentBlocks.map(b => b.text).join('') + actionCompact
  const promptLengths = {
    systemPrompt: systemPrompt.length,
    toolCalls: toolCallsLength,
    userPrompts: userPromptsLength,
  }

  // Compare main-loop context vs classifier transcript to track projection
  // divergence. tokenCountWithEstimation is cheap (walks back to last API