Configure universal tool-call ingress scanning

Scan every tool's arguments in the central permission gate before execution. Default rules target high-confidence data serialization, encoded output, network exfiltration, shell display of data files, and direct previews of registered protected dataset names. This is defense-in-depth; egress remains the primary model-data boundary.

Usage

shield_ingress(
  langs = c("r", "python", "bash"),
  patterns = NULL,
  include_defaults = TRUE,
  on_fail = c("block", "ask"),
  ignore_case = TRUE
)

Arguments

langs

Any of "r", "python", and "bash"; controls which language-specific default rules are included.

patterns

Optional named regular expressions using PCRE (Perl-Compatible Regular Expression) syntax. A name matching a built-in rule (e.g. py_pandas_export) replaces that rule; a new name is added. Hosts wanting file-managed blacklists read their own file (e.g. yaml::read_yaml()) into a named vector and pass it here.

include_defaults

Include built-in serialization/encoding, network transfer, shell data-file display, and protected-name preview rules.

on_fail

"block" rejects the tool call; "ask" forces the existing permission approval callback/UI (including the tool-call id).

ignore_case

Apply case-insensitive matching.

Value

A Data Shield ingress strategy specification.

Examples

strict_calls <- shield_ingress(on_fail = "block")
supervised <- shield_ingress(langs = c("r", "python", "bash"), on_fail = "ask")