Observed behavior

On 2026-10-13, an existing Chat used Remote Desktop Commander to read an authorized local text file. The call returned running with a call_id. The same Chat performed its first get_tool_result and received completed metadata only. It reported no text, content blocks, result body, truncation notice, or response-resource reference. It stopped immediately; the read was not repeated and no downstream business command was submitted.

The original Chat was the only result receiver assigned by us. The coordinating Codex instance did not retrieve that call. This excludes our coordinator retrieving first as a cause in this attempt; it does not audit other platform consumers.

The local device reports version 0.2.51 and has existing local transport/permission customizations. We have not established which deployed cloud handler or Chat adapter version constructs this result.

Useful field differential, with limits

An earlier, different diagnostic received by Codex exposed a full MCP result shaped like the synthetic example below. Its text was in content, while structuredContent contained async_call metadata. The latest Chat-visible response has the same field structure as that structuredContent subobject. These were different calls and file sizes, so this is a lead, not proof that the latest raw response contained a body or that a particular component dropped it.

Synthetic structural example — not the latest call's network capture:

{
"content": [{"type": "text", "text": "OFFLINE_TEXT_BODY"}],
"structuredContent": {
"async_call": {
"call_id": "00000000-0000-4000-8000-000000000001",
"tool_name": "read_file",
"status": "completed",
"retry_original": false
}
},
"isError": false
}

Observed model-visible shape, with identifying values removed:

{
"async_call": {
"call_id": "",
"tool_name": "read_file",
"status": "completed",
"device_id": "",
"created_at": "",
"completed_at": "",
"retry_original": false
}
}

What has been checked

• The installed @modelcontextprotocol/sdk 1.29.0 CallToolResultSchema and Client.callTool preserve both content and structuredContent in a synthetic offline test. Its request transport was a local stub; this is not a cloud-handler or Chat-adapter test.
• A separate synchronous Chat read returned complete text. Its smaller input and different call make it an imperfect control.
• The Chat record preserves all model-visible messages, not raw network RPC. The exact wall-clock time of the get invocation is not independently available.
• No service, credentials, installed SDK, or production tool implementation was changed for this diagnosis. No stopped call has been replayed.

Requested investigation

Please identify the existing cloud get_tool_result terminal-result builder and the connector/Chat output adapter, or provide their non-sensitive output contract:

1. When status is completed, where is the original tool's content placed? Is structuredContent merged with async_call or replaced?
2. Does any adapter choose structuredContent instead of retaining sibling content? What outputSchema is declared?
3. Can retention remove the result body while completed metadata remains available? What are the deployed rules?
4. Which component owns the raw MCP-to-model-visible conversion for this integration?

Acceptance for a targeted fix

• Preserve the original readable body and appropriate status/error fields through the existing result path. If content contains existing resource or non-text blocks, preserve their supported representation rather than inventing text.
• Do not expose hidden _meta or credentials while fixing visible fields.
• A metadata-only completed read must remain distinguishable from a successfully recovered body. Failed, expired and unknown cases must not be upgraded to success.
• Regression-check a full content-plus-structured result, a metadata-only negative case and a synchronous text result. Any later live verification needs a separately bounded original-caller-only attempt; this report does not request retries of the stopped call.

Reference: https://developers.openai.com/plugins/reference#tool-results
The documented result contract allows both content and structuredContent to be surfaced. It does not establish which deployed layer caused this incident.

Privacy: This draft contains no local paths, business text, user/account identifiers, live call/device identifiers, authentication material, or original Kernel content. No production logs or conversation exports are attached.