OpenCode Training Data Exporter

A utility to export agent chat sessions, tool trajectories, and model outputs from the OpenCode local database into a training-ready JSONL dataset in the standard Hugging Face chat format.

Features

• Standard messages format: One record per session, with a messages list of {role, content} objects — the field trainers consume via a chat template (tokenizer.apply_chat_template, TRL SFTTrainer, etc.).
• Tool calls and results (OpenAI convention): Assistant tool use is emitted as tool_calls (with arguments as a JSON string), and each tool result is a role: "tool" message paired back by tool_call_id. Error results are included.
• Inline reasoning: The assistant's reasoning is inlined as <think>...</think> at the start of the assistant content — the convention used by OLMo 3 / Dolci-Think SFT data.
• Idempotent by default: Re-running rebuilds the output atomically, so the same database always yields the same file with no duplicate rows. An optional --append mode adds only new, deduplicated rows to an existing file.
• Clean data: Reads OpenCode's materialized message and part tables (the final consolidated state), not the streaming event log, so text and tool calls are never duplicated by streaming updates. Sessions with no usable assistant content are dropped.
• Metadata: Each record carries id (session id), source, and the agent and model that produced it.
• Configurable: Paths and options via CLI flags or environment variables.

Installation

The exporter is a standalone Python 3 script with no external dependencies beyond the standard library.

git clone https://github.com/imrija95/opencode-training-exporter
cd opencode-training-exporter

Usage

Run the script directly with Python:

python3 exporter.py

By default this rebuilds the output file from scratch (idempotent). To instead append only new, deduplicated rows to an existing output:

python3 exporter.py --append

Options and configuration

CLI flags take precedence over environment variables, which take precedence over the defaults.

CLI flag	Env variable	Description	Default
--db	OPENCODE_DB_PATH	Path to opencode.db	~/.local/share/opencode/opencode.db
--output	OPENCODE_OUTPUT_PATH	Output JSONL path	~/opencode_training_data.jsonl
--source	OPENCODE_SOURCE	Value for each record's source field	opencode
--prompts-dir	OPENCODE_PROMPTS_DIR	Directory of {agent}.md system-prompt files (see below)	unset
--append	—	Append new deduplicated rows instead of rebuilding	rebuild

Example:

OPENCODE_OUTPUT_PATH="/path/to/my/dataset.jsonl" python3 exporter.py

The database is opened read-only and immutable, so the exporter can run safely while OpenCode is in use; it never locks or writes the database.

Data Format

The exporter produces a .jsonl file where each line is one session:

{
  "id": "ses_...",
  "source": "opencode",
  "agent": "coder",
  "model": "provider/model-id",
  "messages": [
    {"role": "system", "content": "..."},
    {"role": "user", "content": "please list usb"},
    {"role": "assistant", "content": "<think>I should run lsusb</think>",
     "tool_calls": [
       {"id": "c1", "type": "function",
        "function": {"name": "bash", "arguments": "{\"command\": \"lsusb\"}"}}
     ]},
    {"role": "tool", "tool_call_id": "c1", "content": "Bus 001 ..."},
    {"role": "assistant", "content": "Found the device."}
  ]
}

Field notes:

• messages is the training field, in the standard Hugging Face chat shape. Roles are system, user, assistant, and tool.
• Reasoning is inlined as <think>...</think> at the start of the assistant content; the answer text follows after </think> (OLMo 3 / Dolci-Think convention). There are no <answer> tags.
• Tool calls use the OpenAI convention: an assistant message carries tool_calls (each with a JSON-string arguments), and every tool result is a role: "tool" message referring back via tool_call_id. A tool-only assistant message has an empty content.
• id, source, agent, and model are metadata columns; trainers that read only messages ignore them.
• Structural markers (step-start, step-finish, compaction) are omitted.

The system message

The system/agent prompt is not stored in the OpenCode database — OpenCode injects it at request time. As a result no system message is emitted by default. As a best-effort convenience, if you point --prompts-dir (or OPENCODE_PROMPTS_DIR) at a directory of {agent}.md files, the exporter prepends the matching agent's prompt as a system message. This only works for agents whose prompt files exist on disk (typically custom agents); built-in agents' prompts live inside OpenCode and cannot be recovered. It is best-effort and reflects the current prompt files, which may differ from the prompt used when the session was recorded.

Note: the top-level OpenAI tools (function schemas) list is not emitted — the schemas are not present in the OpenCode database.

Privacy & Security

This tool reads a local opencode.db and, optionally, local prompt files. Handle the exported .jsonl files securely, as they contain your private chat history. The exporter never reads OpenCode's opencode.json (which may contain API keys) — only the database and any prompt directory you explicitly pass.