The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes local bash scripts (scripts/init-pipeline.sh, scripts/show-status.sh) and standard git commands to initialize, track, and commit progress. These operations are core to the skill's functionality for session resilience and are properly scoped to the project environment.
[PROMPT_INJECTION]: The skill reads and processes state data from .claude/pipeline-state.json, creating an indirect prompt injection surface where external modification of the state file could influence agent behavior.
Ingestion points: The state file is ingested via the Read tool and summarized by the show-status.sh script.
Boundary markers: No explicit delimiters are present to separate injected state content from system instructions.
Capability inventory: The skill possesses the ability to execute bash commands, perform git operations, and manage sub-tasks via TaskCreate.
Sanitization: While the JSON is parsed for structure, no semantic sanitization of the values (e.g., phase descriptions) is performed before they are used in the prompt context.

checkpoint-resume