grith.aidocs

[proxy]

Scoring thresholds and per-filter sub-configs.

The [proxy] section configures the scoring engine — the thresholds, the cold-start behaviour, and the queue timeout.

[proxy]

KeyTypeDefaultDescription
auto_allow_thresholdfloat3
auto_deny_thresholdfloat8
cold_start_callsinteger0
cold_start_escalation_highfloat10
cold_start_escalation_lowfloat2
review_timeout_secondsinteger300auto-deny queued calls after 5 minutes

Per-filter sub-sections

Each filter has its own sub-section under [proxy.filters.*]:

[proxy.filters.reputation]

KeyTypeDefaultDescription
domains_filestring"config/filters/domains.toml"
enabledbooleantrue
ip_address_scorefloat2
malicious_domain_scorefloat4
safe_domain_scorefloat-1
unknown_domain_scorefloat1

[proxy.filters.behavioural]

KeyTypeDefaultDescription
enabledbooleantrue
mild_deviation_scorefloat1
min_calls_for_baselineinteger200
significant_deviation_scorefloat3

[proxy.filters.taint]

KeyTypeDefaultDescription
enabledbooleantrue
high_risk_sinksarray["network", "shell"]
medium_risk_sinksarray["file_write"]
sensitive_pathsarray[".env", ".env.local", "credentials.json", "secrets.yaml", "id_rsa", "id_ed25519"]

[proxy.filters.rate_limit]

KeyTypeDefaultDescription
burst_multiplierfloat3
cooldown_secondsinteger30
enabledbooleantrue
file_writes_per_minuteinteger30
network_requests_per_minuteinteger60
shell_execs_per_minuteinteger20

[proxy.filters.semantic]

KeyTypeDefaultDescription
enabledbooleanfalseStub in v1.5, requires local embedding model
modelstring"all-MiniLM-L6-v2"

[proxy.filters.egress]

KeyTypeDefaultDescription
config_filestring"config/filters/egress.toml"
enabledbooleantrue

[proxy.filters.session_containment]

KeyTypeDefaultDescription
config_filestring"config/filters/containment.toml"
enabledbooleantrue

Tuning thresholds

The most important pair is auto_allow_threshold and auto_deny_threshold:

  • Lower auto-allow = more calls go to queue (paranoid).
  • Lower auto-deny = more calls auto-deny (very paranoid).
  • Higher auto-allow = more calls auto-allow (quiet, less safe).
  • Higher auto-deny = fewer calls auto-deny (chatty queue).

See Tuning scoring thresholds for the walkthrough.

See also

Last updated: 2026-05-14Edit this page on GitHub →
© 2026 grith. All rights reserved.