The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes platform hooks (moru-hook) to perform operations such as repair-verify, gate, and gate-batch to manage the task lifecycle and verify instruction changes.
[DYNAMIC_EXECUTION]: The workflow instructs the agent to execute py_compile and unittest on files it has modified or authored. This involves runtime compilation and execution of dynamically generated content during the maintenance phase.
[INDIRECT_PROMPT_INJECTION]: The core logic involves ingesting evidence from completed tasks (which may include untrusted input from tickets, plan execution, or results) to update the agent's canonical instructions (SKILL.md). This creates a feedback loop susceptible to persistent rule drift or malicious manipulation, with no specified external sanitization for the ingested data.
[PRIVILEGE_ESCALATION]: The skill explicitly grants itself "standing local_write self-evolution authority," instructing the agent to modify its own canonical instruction files within specific directories (.agents/shared/llm-skills/ and .agents/local/skills/) without requesting user permission or pausing for approval.

retrospective-learning