The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes shell commands such as git branch, git log, and git status to retrieve workspace metadata. It also invokes a local Python script to validate file integrity. These operations are limited to project-local state retrieval and validation.
[INDIRECT_PROMPT_INJECTION]: The skill processes data from .artifacts/HANDOFF.md which creates an attack surface for indirect prompt injection.
Ingestion points: Untrusted data enters the agent context in instructions/load.md when the handoff file is read.
Boundary markers: The skill uses a structured markdown template with specific headers (Focus, Context, Current state) to delimit data.
Capability inventory: The skill can execute git commands, run a local validation script, and write to the filesystem within the project scope.
Sanitization: Content is passed through a secret-detection script (check_handoff.py) and the agent is instructed to manually redact sensitive strings and verify all loaded claims against current workspace evidence.
[DYNAMIC_EXECUTION]: The skill runs a bundled validation script, scripts/check_handoff.py, using a relative path. The script is designed to detect and prevent the persistence of hardcoded secrets like API keys and tokens.
[CREDENTIALS_SAFE]: The skill includes a dedicated scanner for credentials, covering patterns for AWS, GitHub, Slack, and generic API keys. It mandates that the agent replaces any identified secrets with a {redacted} placeholder.

handoff