Export

An open session has two buttons in its header: Markdown and JSON. The file is called buff-code-<session id>.md or .json. The whole session is exported, not just the part loaded into the feed; service calls are always included.

Markdown — for reading

The whole session history as text: a heading from the first line of the developer’s first message, the developer, the device, the project, the start, the number of turns, the amount charged and tokens by type; then, in order:

• Developer: the message;
• Buff Code — time · model · tokens: the model’s reasoning, the answer, each action — verb, object, the tool’s technical name, the arguments and the result in code blocks — and the turn’s errors;
• service calls are marked “(service call)”;
• a tool’s answer whose call did not make it into the record — as a “Tool output” block.

This is how the start of the file looks (a two-turn session from the English workspace; the values are illustrative):

# Why does test_orders fail?

Developer: alice · device alice’s laptop · project shop-web
Started: 2026-09-24T10:12:03+00:00 · turns: 2 · charged: 3.40 ₽
Tokens: 18,240 input · 612 output · 0 reasoning · 9,100 cache read · 0 cache write

**Developer:**

Why does test_orders fail?

### Buff Code — 2026-09-24T10:12:05+00:00 · gpt-5 · 9,120 input · 96 output · 0 reasoning · 0 cache read · 0 cache write

**Ran** `pytest tests/test_orders.py -q` (`bash`):

```
{
  "command": "pytest tests/test_orders.py -q",
  "description": "Run the failing test"
}
```

Result:

```
F                                                                        [100%]
FAILED tests/test_orders.py::test_total - AssertionError: 100 != 90
Command exited with code 1
```

### Buff Code — 2026-09-24T10:12:19+00:00 · gpt-5 · 9,120 input · 516 output · 0 reasoning · 9,100 cache read · 0 cache write

The test expects the total with the discount, but `total()` does not apply it — fixing.

**Edited** `src/orders.py` (`edit`):

```
{
  "filePath": "src/orders.py",
  "oldString": "    return sum(i.price for i in items)",
  "newString": "    return sum(i.price for i in items) - discount"
}
```

Result:

```
Edited src/orders.py
```

It is suitable for attaching to an incident review, sending to a colleague or putting into the knowledge base.

JSON — the whole session as data

An object with three fields:

• session — the session description: identifier, organisation, project, login, device, start and last change, number of turns, model, first message, tokens by type, duration and amount;
• note — an explanation of the file’s contents;
• messages — the session’s messages in order, as Buff Code recorded them. Each has its number in the session (id), who speaks (role: the developer or the model), the time, a service flag (service) and parts (parts): text, reasoning, a tool call with arguments (tool_call), the call result (tool_result), an error. A model answer has answer_of: the turn that produced it — model, status, duration, tokens by type and the amount charged.

It is suitable for your own analysis: count tool calls, break down typical questions, load it into a SIEM.

What the export leaves out

• the system prompt and the tool descriptions that Buff Code adds to every request — neither in Markdown nor in JSON (see what is not shown);
• files on the developer’s disk that were not passed to the model;
• connection secrets — they are not in the record in the first place.

Audit log

Every export — of your own or someone else’s session — is recorded in the audit log as a “Development session exported” event: who, whose session, when, from which address. More — Audit log and Buff Code.

Local export from Buff Code

Buff Code itself has buff session export — a local copy of the session from the developer’s machine, in Buff Code’s format. It has nothing to do with the records in the workspace: the workspace keeps what passed through the server, and the developer cannot change that record.