canopy-agent-factory

Stamp out a new Claude Code agent from the canopy operating model — a persona, the turn orchestrator, reads-free / writes-gated guarding via a config-driven PreToolUse hook, and a canopy-web-ready layout. Extracted from the canopy distribution's src/orchestrator/agent_factory.py so a second consumer (canopy-web, a Django app) can scaffold an agent repo without inheriting canopy's video-tooling dependency footprint (Pillow, NumPy).

Zero runtime dependencies. Versions independently of canopy.

Public API

from canopy_agent_factory import AgentSpec, create_agent, normalize_slug, AgentFactoryError
from canopy_agent_factory import templates, gating_config  # canonical stamp-table baseline

• AgentSpec — dataclass describing the agent to scaffold.
• create_agent(spec, target_dir, *, force=False) -> list[Path] — writes the templated files into target_dir (caller-supplied; this function does no git init and no other I/O beyond writing the scaffold) and returns the paths it wrote.
• normalize_slug(raw) -> str — validate + normalize an agent slug.
• AgentFactoryError — raised for bad input or an unsafe target directory.
• templates() -> Mapping[str, str] — the canonical stamp table (repo-relative path → rendered-template source), read-only. Ground truth for "what a factory-stamped agent looks like"; used by canopy's fleet-alignment lens to diff real agent repos against the baseline.
• gating_config() -> str — the canonical config/gating.json body a stamped agent ships with.

Development

uv run --with pytest pytest -q

Runs standalone — no orchestrator import, no Django.