The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses the Bash tool to run the command python -m ane_package.qa.word_count to perform word count verification. The file path passed to this command is constructed using a run slug and specialist name derived from the inbound plan and registry. This interpolation could potentially be a vector for command injection if the input strings contain shell metacharacters.
[DATA_EXFILTRATION]: The skill accesses a hardcoded absolute file path on a specific user's OneDrive directory (C:/Users/AGasser/OneDrive/5 ANE CLAUDE work folder/mel_wiki/wiki/). While intended for loading project context, this pattern exposes local directory structures and is non-portable.
[DYNAMIC_EXECUTION]: The orchestrator dynamically spawns subagents using the Agent tool based on a registry (agent_registry.md). It also supports a 'Write-and-bridge' pattern that allows the creation and execution of inline prompts for a 'general-purpose' subagent when a specific specialist is not found in the registry.
[INDIRECT_PROMPT_INJECTION]: As an orchestration layer, the skill ingests content from external plans and subagent returns, which could potentially contain malicious instructions. The skill implements several mitigation strategies, including:
Ingestion points: Inbound prompt blocks (## P1 wiki context), subagent returns files in %TEMP%, and local wiki files.
Boundary markers: The use of specific markdown blocks and slugs to identify and isolate run-specific data.
Capability inventory: Access to Agent (subagent spawning), Read, Write, Edit (file system modification), and Bash (shell execution).
Sanitization: A comprehensive review process involving dedicated qa-reviewer and safeguarding-reviewer subagents, as well as automated sweeps for em-dashes and factual reliability.

vi