The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests data from external artifacts and subagent responses that could potentially contain instructions intended to override the orchestrator's behavior.
Ingestion points: Subagent response text captured during the 'settle' step, concerns blocks extracted from agent output, and architect specifications fetched via artifact_read.
Boundary markers: The skill does not implement delimiters or safety instructions when interpolating these external strings into subsequent commands or tool calls.
Capability inventory: The skill utilizes shell execution (git, sam_cli), subagent spawning (Agent), and external state modification via backlog tools.
Sanitization: No explicit escaping or validation of the ingested content is described before it is used in command arguments or tool parameters.
[DATA_EXFILTRATION]: The skill identifies and accesses platform-level session transcript data to perform health checks on subagents.
Evidence: The health check procedure in references/agent-health-check.md involves locating and reading .jsonl transcript files within the ~/.claude/projects/ directory.
Context: This access is specific to the tool's purpose of managing subagent lifecycles and monitors platform metadata rather than user-provided credentials.

implement-feature