The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process trace data from external observability providers (such as Langfuse or Braintrust), which are untrusted inputs. Malicious instructions or malformed structures in these traces could attempt to influence the agent's behavior during the import process. The skill mitigates this by requiring strict schema validation, record-count limits, and structural depth checks.
Ingestion points: external provider exports described in SKILL.md and references/failure-and-validation.md.
Boundary markers: The parser contract uses specific return types ('ImportedSession', 'ImportFailure') and explicit parser exceptions to isolate failures.
Capability inventory: Local Python script execution and Kitaru CLI commands (scaffold, test, register, import).
Sanitization: Rigorous redaction instructions are provided to remove prompts, credentials, and PII from trace samples before they are used as fixtures or uploaded for smoke tests.
[DYNAMIC_EXECUTION]: The skill generates Python importer scripts and executes them for local fidelity testing. The instructions warn that this is not a security sandbox and require the agent to use credential-free isolated environments. It mandates explicit user approval before installing dependencies or executing generated code.
[COMMAND_EXECUTION]: The skill relies on the 'kitaru' CLI tool to perform various lifecycle tasks such as scaffolding, testing, and registering importers. These operations are governed by the local system's command schema and are initiated only after user confirmation.
[DATA_EXFILTRATION]: The skill manages potentially sensitive trace content. It contains extensive instructions to prevent accidental data exposure or exfiltration by redacting secrets, credentials, and personal identifiers from all representative samples, fixtures, and logs before they leave the local environment.

kitaru-importer-builder