teich 0.1.1a2

Generate agent training data from Codex and Pi traces

Teich

v2/ is the experimental trace-first package for collecting raw agent sessions and converting them into training-ready data.

What it does today

• Runs Codex and Pi in a shared Docker runtime with uv, npm, @openai/codex, and @mariozechner/pi-coding-agent
• Configures Codex through a mounted CODEX_HOME/config.toml
• Configures Pi through an isolated mounted ~/.pi/agent/settings.json
• Exports raw session traces from mounted Codex and Pi session directories
• Writes a trace-folder README.md for upload
• Exposes Python conversion helpers for training data preparation

Usage

# Initialize a project
uvx teich init my-project
cd my-project

# Run with the configured agent provider and model settings
uvx teich generate -c config.yaml

Local OSS providers

If you want Codex to talk to a local provider like LM Studio or Ollama, set the provider in config or env:

$env:TEICH_PROVIDER='LMstudio'
$env:TEICH_MODEL='gemma-4'
$env:TEICH_API_KEY='llm'
$env:TEICH_BASE_URL='http://localhost:1234/v1'
python -m teich generate -c test_run/config.yaml

v2 maps LMstudio and ollama onto Codex's native --oss --local-provider ... flow.

Configuration model

Important fields in config.yaml:

agent:
  provider: codex  # or pi

model:
  model: codex-mini-latest
  approval_policy: never
  sandbox: danger-full-access
  reasoning_effort: null

api:
  provider: openai
  base_url: null
  api_key: null

Legacy model.approval_mode is still accepted and normalized internally.

Python conversion API

from pathlib import Path
from teich import convert_traces_to_training_data

examples = convert_traces_to_training_data(Path("./output"))

The converter currently maps example-style raw traces into message/tool records with:

• system/developer instructions
• user messages
• assistant messages
• reasoning_content
• tool calls
• tool results

Development

uv pip install -e ".[dev]"
pytest tests/test_config.py tests/test_cli.py tests/test_runner.py -q

Architecture

• Shared Docker runtime: container image includes Node.js, uv, uvx, @openai/codex, and @mariozechner/pi-coding-agent
• Isolated Pi config: Pi runs with a mounted per-run ~/.pi/agent directory inside the container
• Codex config: generated config.toml under a mounted CODEX_HOME
• Session export: raw JSONL sessions are copied from mounted Codex or Pi session storage into the user output directory
• Upload-first output: traces are preserved in raw form before later conversion
• Provider-aware boundary: agent.provider selects either the Codex or Pi raw-trace path

Project Structure

v2/
├── docker/
│   └── codex-runtime.Dockerfile
├── src/teich/
│   ├── __init__.py
│   ├── __main__.py
│   ├── cli.py
│   ├── config.py
│   ├── converter.py
│   ├── runner.py
│   └── trace_readme.py
└── tests/
    ├── test_cli.py
    ├── test_config.py
    └── test_runner.py