The Agent Skills Directory

[SAFE]: The skill is primarily instructional, outlining a methodology for long-term task planning and execution. It promotes best practices such as model switching for cost efficiency and using checkpoints for resumability.
[INDIRECT_PROMPT_INJECTION]: The framework is designed to ingest external task definitions (e.g., jobs.json) and specifications (spec). While this is a vulnerability surface, the skill explicitly mitigates this by requiring independent 'checker' agents and strict 'handoff' contracts to verify the actual state of work before proceeding.
Ingestion points: Reads from jobs.json, templates/job-contract.json, and paths defined in spec fields.
Boundary markers: Uses handoff contracts and actual_end_state fields to delimit task boundaries.
Capability inventory: Involves file system interaction (Git/Knowledge Base) and utility execution via the lp CLI tool.
Sanitization: Includes a validation step (lp job validate) to ensure task files adhere to the expected schema and safety constraints.
[COMMAND_EXECUTION]: The skill references CLI commands lp job validate and lp report. These are used for workflow validation and reporting, which is consistent with the stated purpose of planning and monitoring long-running jobs.

night-loop-planning