The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to process external tasks and specification files as primary inputs (SKILL.md). It mitigates the risk of indirect injection by enforcing a specialized 'ai' lens (references/lens-catalog.md) that audits for untrusted content reaching sensitive sinks.
Ingestion points: The skill accepts a spec path or task text as arguments in SKILL.md.
Boundary markers: Guidelines in references/lens-catalog.md (item 8) explicitly instruct agents to trace untrusted content and flag missing iteration/cost caps.
Capability inventory: The skill dispatches subagents with the ability to write to the file system and execute shell commands via project gates (e.g., pytest, npm test).
Sanitization: The auditing protocol requires charter validators to trace untrusted content to prompts and sinks, ensuring validation before merging or execution.
[UNVERIFIABLE_DEPENDENCIES]: The preflight protocol described in references/preflight.md allows the agent to automatically attempt to install missing dependencies (e.g., via npm or pip) to resolve 'environmental blockers'. This is intended to unblock existing validation tools and is limited to two attempts.
[DYNAMIC_EXECUTION]: As part of its core functionality, the orchestrator dispatches subagents to perform code changes and subsequently runs project-defined gate commands (tests, builds, linters) resolved from repository manifests like package.json or pyproject.toml (references/stack-adapters.md). Execution is performed by fresh validation contexts to maintain independence.

nuke-exec