The Agent Skills Directory

[COMMAND_EXECUTION]: The skill is granted access to the Bash tool and process.run (as defined in host-contract.json) to perform architectural coordination and handle file-native editing operations. These tools are used within the scope of managing project metadata and applying approved configuration changes.
[PERSISTENCE]: The skill implements an 'Assistant Preference' protocol described in references/assistant-preference.md. This protocol allows the agent to persist specific instructions into a host-verified configuration file using an owned block delimited by  and . This persistence is gated by mandatory user approval, diff previews, and a single-offer policy to prevent repeated prompts.
[INDIRECT_PROMPT_INJECTION]: The skill features a surface for indirect prompt injection as it ingests 'User intent' and 'requested extent' (as noted in host-contract.json) which are used to generate the preference blocks described above.
Ingestion points: Data enters via the inputs specified in host-contract.json (intent, extent, host capability inventory) and is processed during coordination workflows in SKILL.md.
Boundary markers: The skill utilizes explicit HTML-style comment markers to delimit the instructions it writes to host files.
Capability inventory: The skill has access to Bash, Write, Read, Glob, and Grep for file and system operations.
Sanitization: The skill requires a 'concrete plan' for substantial tasks, uses byte-level hashes to verify targets, and presents an 'exact-edit approval' gate to the user before modifying persistent records.
[DYNAMIC_EXECUTION]: The skill coordinates 'execution agents' and 'parallel assignments' by delegating to the host's native agent-creation tools. The instructions in SKILL.md explicitly state that the skill/MCP text does not itself create agents, but instead hands approved plans to the main harness for invocation.

norma