A Python port of render-examples/software-factory (baseline
commit fd3731d), with Pydantic AI as the agent loop.
implementation-plan.md is the single delivery plan. spec.md defines the
product and security contract; plan.md is historical and is not an active plan.
The security-bearing core, Workflow registration, sandbox/agent runtime, Blueprint, and setup/diagnostic surface are present. They have local automated coverage. This checkout has not been deployed or live-verified against Render, GitHub, or credential brokering; Blueprint validation alone is not deployment proof.
uv sync --locked
uv run preflight
make checkFor deployment, follow docs/runbooks/setup.md. Scripts dry-run by default, secret values stay in Render environment groups, and no setup command deploys.
uv sync --locked
make check # ruff, shellcheck, pyright (strict), pytestTests use fakes only: no GitHub, Render, or model calls. The publish tests run sandbox/collect-changes.sh against
a local git repository, so they need git and bash. make check requires shellcheck; CI installs it explicitly.
- Quickstart
- Configuration reference
- Architecture and security model
- Deployment/setup runbook
- Agent and placement extension runbook
Local checks validate source, configuration, and fake-backed behavior only. render blueprints validate render.yaml is read-only validation, not a deploy. A live deployment is verified only after gateway health,
Workflow task registration, GitHub App installation, sandbox group, and credential-brokering checks pass.
- There is no cross-run spend ledger or durable conversation carry-over.
- Brokering prevents key theft, not model spend: code in a live, model-enabled sandbox can use the restricted credential until that sandbox terminates.
- In-task tools and lifecycle polling share the workspace Sandbox API rate limit; the process-wide pacer favors safety over maximum concurrency.