Transcript format¶

A transcript is conversation history in OpenAI chat format, wrapped in a small versioned envelope. Applications send it in with each turn and receive the turn's new messages back in the same format.

Media type: application/vnd.astralvessel.transcript+json

The envelope¶

{
  "version": 1,
  "messages": [
    {"role": "user", "content": "What is here?"},
    {"role": "assistant", "content": "Looking. ",
     "tool_calls": [{"id": "call_1", "type": "function", "function": {"name": "list", "arguments": "{}"}}]},
    {"role": "tool", "tool_call_id": "call_1", "content": "species/"},
    {"role": "assistant", "content": "There is one folder, species/.", "reasoning_content": "..."}
  ]
}

Messages¶

Role	Allowed fields	Rules
`user`	`role`, `content`	`content` is a string, or a non-empty list of `{"type": "text", "text": ...}` and `{"type": "image_url", "image_url": {"url": ..., "detail": ...}}` parts
`assistant`	`role`, `content`, `tool_calls`, `reasoning_content`	`content` is a string or null; each tool call has exactly `id`, `type: "function"` and `function` with string `name` and `arguments`
`tool`	`role`, `tool_call_id`, `content`	Must answer an open tool call from the preceding assistant message; `content` is a string

system messages are refused: the briefing is fixed by the operator.

The vessel also checks that:

the envelope's version is 1 and it holds at most 2,000 messages;
no other field appears on a message;
every tool call is answered before the next user or assistant message, and the history does not end with an unanswered call;
images are inline data: URLs of PNG, JPEG, GIF or WebP, real base64 with the right leading bytes, at most 5 MB each and 32 in the whole history, and only for agents whose model accepts images;
detail, when present, is auto, low or high.

A history that fails a check rejects the turn with a message naming the problem.

What comes back¶

With the transcript media type in acceptedOutputModes, the transcript artifact streams each completed message of the turn, in order: the user's message first, then assistant messages, tool results, and the final answer. Each message is sent once, as soon as it is complete, so a dropped stream still leaves a usable prefix. The last chunk adds a summary:

{"version": 1, "messages": [],
 "turn": {"status": "ok", "steps": 2, "tokens": {"input": 30, "output": 13}, "elapsed_ms": 3, "model": "your-model-name"}}

Append the messages to your stored history and send them with the next turn.

Situation	What the transcript holds
The turn ended before the model answered	A closing assistant message `[this turn ended early: <status>]`
A tool call never ran	A tool result `[not run: the turn ended first]`
The model sent tool arguments that are not JSON	The call's `arguments` are stored as `{}`, so the history stays replayable; the error result stays

reasoning_content is kept in the transcript; it is sent back to the model on later turns only when the agent sets send_reasoning = true.