The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill performs controlled network requests to well-known astronomical services to facilitate scientific calibration.
Evidence: The color.calibrate protocol allows Siril to query the Gaia DR3 catalog via the vizier.cds.unistra.fr mirror or Simbad at simbad.cds.unistra.fr when local catalog extracts are missing or insufficient.
Context: These are domain-standard scientific queries directed at established academic institutions (CDS Strasbourg, Zenodo). The skill's configuration policy (color-calibrate.md) explicitly limits online access to these specific paths.
[COMMAND_EXECUTION]: The skill executes the Siril 1.4.4 CLI to perform image processing tasks, using a secure execution contract.
Evidence: The Python orchastrator script (deep_sky_siril.py) spawns Siril subprocesses to run Agent-authored .ssf scripts.
Sanitization: Every command within the .ssf scripts is validated against a local whitelist and policy file (command-policy.json). Commands not explicitly authorized for a given protocol are blocked before execution.
[INDIRECT_PROMPT_INJECTION]: The skill processes external astronomical data (FITS, XISF) and evaluates logs derived from these files, creating a potential attack surface that is heavily mitigated.
Ingestion points: Primary data ingestion occurs through user-supplied FITS/XISF image files and subsequent parsing of Siril process logs and statistics.
Boundary markers: The skill uses structured JSON receipts (runs/*.json) and a session manifest to strictly separate data from control flow. Cryptographic SHA-256 bindings ensure that inputs and logs cannot be tampered with between stages.
Mitigation: A required human visual review of 'actual pixels' and statistics before finalization prevents the agent from making autonomous decisions based purely on potentially manipulated log data.

starun-siril