The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill reads and injects the contents of project files like task_plan.md and progress.md into the agent's prompt context during every tool call. The skill documentation explicitly identifies this as an injection surface and provides an attestation mechanism to ensure that only user-approved plans are processed. Injected data is framed with nonces and SHA-256 hashes to prevent delimiter confusion.
[COMMAND_EXECUTION]: The skill uses lifecycle hooks to execute local shell and PowerShell scripts for environment resolution and context injection. These scripts further execute system utilities such as sha256sum, awk, and sed.
[DYNAMIC_EXECUTION]: The Python injector script scripts/inject-plan.py uses subprocess.run to invoke ledger-summary.sh. This process is used to summarize project progress from JSONL ledger files.
[DATA_EXPOSURE]: The scripts/session-catchup.py script accesses local session history files from Claude Code and other platforms to aid in task resumption. While this data is local, it involves reading transcripts that may contain sensitive project metadata.

planning-with-files-zh