The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from GitHub issue titles, bodies, and comments in Step 3.
Ingestion points: Step 3 reads mined issue content into the session for clustering and synthesis.
Boundary markers: The skill contains an explicit 'Untrusted-content boundary' section instructing the agent to treat mined content as evidence only and never to execute instructions embedded within it.
Capability inventory: The skill has access to Bash, Write, Edit, and the Skill tool (used to invoke /issue).
Sanitization: A pre-processing step (learn-from-bugs prepare) compresses issue bodies into a compact root-cause digest, which significantly reduces the attack surface by stripping potentially malicious text before it reaches the model's main context.
[COMMAND_EXECUTION]: The skill executes several commands via a vendor-provided script (larch.sh) to prepare data, check proposals, and publish state.
The execution is restricted to specific verbs (resolve-zones, prepare, check-proposals, state-publish, filing-deps) within the vendor's controlled script environment.
Commands are used for functional requirements such as resolving Topical Zones to GitHub search queries and managing durable proposal state.

learn-from-bugs