Zoom Codex connector: paginate transcript retrieval to avoid model-facing truncation

[kushagrasharma]

Summary

The Zoom Codex connector currently returns meeting transcripts inline from zoom.get_meeting_assets. For longer transcripts, the raw MCP/tool result can contain the full transcript, but the model-facing function_call_output may be truncated before the agent can reason over the full meeting.

This looks like a connector/API-shape issue rather than a Zoom transcript data-loss issue: the connector can retrieve the full structured transcript, but Codex receives a truncated inline blob.

No private transcript text, attendee names, customer data, patient/clinical details, or company-sensitive details are included in this report.

Sanitized evidence

Observed on a meeting transcript with generic characteristics:

• Meeting length: about 33 minutes
• Raw connector result included the full transcript in structuredContent
• Raw result had 1271 transcript items
• First transcript timestamp: 00:00:55.000
• Last transcript timestamp: 00:33:46.000
• Model-facing function_call_output was truncated around 00:06:25 and resumed around 00:26:05
• The output included a truncation marker similar to ...16153 tokens truncated...

Because the agent only sees the truncated model-facing output, it cannot reliably summarize, extract action items, answer questions about the omitted middle of the meeting, or verify completeness.

Current behavior

zoom.get_meeting_assets accepts only:

{ "meetingId": "..." }

It returns transcript data inline as a large JSON payload when a transcript is available. There are no cursor, page, time-range, search, or max-size parameters for transcript retrieval.

zoom.get_recording_resource is not an adequate fallback for transcript-only meetings. In the observed case, the meeting had has_transcript=true and has_recording=false; attempting the recording-resource path returned 404.

Requested behavior

Please add a dedicated transcript retrieval API/tool that allows agents to fetch transcripts incrementally and verify completeness without relying on one huge inline payload.

Suggested tool shape:

get_meeting_transcript({
  meetingId: string,
  start_time?: string,
  end_time?: string,
  cursor?: string,
  page_size?: number,
  max_chars?: number,
  format?: "items" | "text" | "vtt" | "srt"
})

Suggested response metadata:

{
  transcript_source: string,
  total_items: number,
  returned_items: number,
  next_cursor?: string,
  truncated: boolean,
  items?: TranscriptItem[],
  text?: string
}

It would also help to expose one or both of:

• search_meeting_transcript({ meetingId, query, start_time?, end_time?, cursor?, page_size? })
• get_meeting_action_items({ meetingId, transcript_range?, source? })

Compatibility suggestion

Consider changing get_meeting_assets so it returns transcript metadata/reference only, rather than embedding the full transcript inline. For example:

{
  has_transcript: true,
  transcript: {
    available: true,
    total_items: 1271,
    first_timestamp: "00:00:55.000",
    last_timestamp: "00:33:46.000",
    retrieval_tool: "get_meeting_transcript"
  }
}

That would keep the asset summary compact while directing agents to a paginated transcript tool when full transcript reasoning is required.

Impact

Without pagination/search/range retrieval, Codex agents can silently miss large parts of transcript-only meetings even though the connector has access to the complete transcript. This can produce incomplete summaries, missed action items, and false confidence in meeting analysis.