Description

The context pruner's summarizeToolCall function always falls back to the generic "Read docs" summary for read_docs tool calls because it reads input.query which doesn't exist on the ReadDocsParams schema.

Root Cause

In agents/context-pruner.ts (line 253-256 in the exported function and the duplicated version inside handleSteps):

case 'read_docs': {
  const query = input.query as string | undefined  // ← wrong field name
  return query ? `Read docs: "${query}"` : 'Read docs'
}

But ReadDocsParams (defined in agents/types/tools.ts:207-214) has these fields:

export interface ReadDocsParams {
  libraryTitle: string
  topic: string
  max_tokens?: number
}

There is no query field. input.query is always undefined, so every read_docs call is summarized as the generic 'Read docs' instead of e.g. 'Read docs: "Next.js - routing"'.

Impact

After context pruning, the model loses information about which libraries/topics were already fetched. This can cause redundant re-fetches of documentation that was already retrieved, wasting tokens and time.

Suggested Fix

case 'read_docs': {
  const libraryTitle = input.libraryTitle as string | undefined
  const topic = input.topic as string | undefined
  if (libraryTitle && topic) {
    return `Read docs: "${libraryTitle} - ${topic}"`
  }
  return libraryTitle ? `Read docs: "${libraryTitle}"` : 'Read docs'
}

Note: this fix needs to be applied in both copies — the exported summarizeToolCall function at the top of the file and the duplicated version inside handleSteps (which is serialized to a string for the agent runtime).