Providers & auth

The turnkey brain runs on one of four provider backends. Select it with the --provider flag or the PYOCD_TURNKEY_PROVIDER environment variable.

# openai-api | anthropic-api | codex-cli | claude-cli
PYOCD_TURNKEY_PROVIDER=codex-cli

The four backends

Provider	Auth	Notes
codex-cli	Local codex CLI auth	Uses whatever Codex auth you configured — a ChatGPT/Codex subscription or Codex’s own API key.
claude-cli	Local claude CLI auth	A user-controlled local Claude Code CLI adapter — a Claude subscription or ANTHROPIC_API_KEY. Not bundled Claude access.
openai-api	OPENAI_API_KEY	Requires an explicit model via --model or PYOCD_TURNKEY_MODEL.
anthropic-api	ANTHROPIC_API_KEY	Requires an explicit model. Stateless at the Messages API layer, so continuity uses the brain’s memory layer.

Note

The CLI backends are your own local dependencies. Hard Fault does not collect, proxy, or repackage any subscription credentials — you authenticate the official vendor CLI yourself.

Choosing a model

The API backends require an explicit model:

uv run pyocd-debug-brain benchmark \
  --provider anthropic-api \
  --model <model-id> \
  --case-id nrf52833dk__k001_reference_green

The CLI backends inherit their default model from the local CLI unless you override it with --model / PYOCD_TURNKEY_MODEL.

Session continuity

Inside a single run, continuity is unified across backends: the brain always persists compact local memory, and remote-capable providers (OpenAI Responses, Claude CLI resume, Codex CLI thread resume) use native continuation with the compact memory injected periodically as a safety sync. A resume failure on a real provider handle stops a headless run rather than silently starting a fresh session.