Versioned stream consumed by the ADEEP evaluation pipeline. The delivery is a library API (adeep.New(writer) driven through Start/Consume/Finish), exposed by bits run --delivery adeep, which writes one JSON object per line to stdout with diagnostics on stderr. The schema discriminator is bits.delivery.adeep and this document describes version 1. Bump Version in jsonl.go and update this document whenever a record gains a field or changes meaning; consumers must reject unknown versions.
Sign in with bits login, then run one assistant turn:
bits run --prompt "Summarize this conversation" --conversation YOUR_CONVERSATION_ID --delivery adeepOmit --conversation to start a new conversation. --prompt is required and
literal; positional and stdin prompts are not supported. The command requires
a working login and never opens an interactive login flow. For API-key
authentication, see Contributing.
Actions requiring approval are denied by default (--permissions deny). Pass
--permissions skip-permissions to allow them without prompting. The final
assistant text is in run.finished.response.
| Exit status | Meaning |
|---|---|
0 |
Turn completed. |
1 |
Startup, runtime, or delivery failure. |
2 |
Invalid command or flags. |
3 |
At least one approval was denied, even if the assistant subsequently finished. |
Every record carries:
| Field | Type | Meaning |
|---|---|---|
schema |
string | Always bits.delivery.adeep. |
version |
int | Contract version, currently 1. |
type |
string | Record type below. |
round |
int | Omitted when 0. 1-based backend send the record belongs to. |
| Type | Round | Fields |
|---|---|---|
run.started |
0 | requested_model?, started_at? (RFC3339) |
round.started |
N | — |
round.finished |
N | — |
conversation.updated |
N | conversation_id |
tool.call |
N | tool{id, name, side: client|server, arguments} — arguments is the exact wire string, including malformed JSON |
tool.result |
N | tool{id, status, result} — result is the exact wire string |
usage |
N | usage{tokens_used, max_tokens, input_tokens?, output_tokens?} |
run.finished |
0 | outcome, conversation_id?, response?, rounds, ended_at? (RFC3339), error{message}? |
tool.call and tool.result correlate by tool.id. A call whose input never arrives is still emitted at Finish, with the exact (possibly empty) arguments and no invented result. Tool definitions opt into provisional input events with stream_input: true. Arguments come from the authoritative final tool call; provisional input stays in the transcript preview and is not emitted as complete arguments.
| Status | Meaning |
|---|---|
success |
The tool ran and returned a result. |
error |
The tool or its handler failed. |
denied |
The approval flow refused the call. |
cancelled |
The call was ended by a user denial of another call, a stop, or cancellation — not a tool failure. |
outcomeiscompleted,approval_denied,failed,canceled, ortimed_out. A completed turn that denied at least one gate reportsapproval_denied.roundscounts backend sends, including rounds whose content a stopped-round drain discarded, soround.started/round.finishedpairs may exist without any content records between them. A send canceled before producing events is not counted.responseis the final assistant Markdown, drawn only from text blocks observed during this run; prior or restored history is never included.
- Exactly one
run.finishedends a healthy stream;Finishis rejected afterwards. ConsumebeforeStartor afterrun.finishedis rejected, as is a secondStart.- The first writer failure is latched: every later write, including the terminal record, fails, so a recovering writer can never emit a well-formed but incomplete stream.
headless.Runcancels the turn on the consumer error and reports it as afailedoutcome;bits runmaps it to exit1.