Instruction file imported from charles-rl/supreme-octo-fiesta (
.cursor/rules/task-dependencies.mdc). Copyright stays with the author.
Task Dependencies
- Never edit
base/for one task's dependencies. - Agent/runtime OS packages:
environment/apt.txt. - Agent/runtime Python packages:
environment/requirements.txt. - Grader-only packages:
scorer/requirements.txtunder root-only/mcp_server/grading_deps. - Hidden-env-only packages:
scorer/env-requirements.txtunder root-only/mcp_server/env_deps. - Placement is a security boundary: a private reference/simulator package that leaks the intended approach must not appear in an agent-visible channel.
- Do not duplicate packages across public/private channels. Avoid mutable VCS, local path, or unpinned network installs.
- Install through the shared task dependency installer before copying/sealing private scorer data. Recheck root ownership and modes afterward.
- Use preloaded mounts for large datasets/models instead of baking or downloading them during an episode.
- Raw Dockerfile installs require an explicit reviewed allow marker only when no channel can express the dependency.
See docs/AUTHORING.md (Task Dependencies) and ml-tasks.