feat: Inference Policy — per-sandbox enforcement of model endpoint access #16

Overview

Per-sandbox inference policy enforcement for Spark Agent Dev Kit. Keep sensitive work local by default; only allow external models when explicitly approved.

Full feature definition: architecture/plans/inference-policy.md (MR !27)

Owner: Alex Watson | Eng: Piotr Mlocek | Target: v0 (~30 days, GTC-critical)

v0 Scope

• Safe defaults: new sandboxes route to local/private endpoints only
• Explicit allowlisting: specifying a provider (e.g., --provider claude) automatically allowlists its required external endpoints
• Per-sandbox policy: each sandbox has its own routing policy, enforced at the inference service layer
• Audit logging: structured logs showing which endpoints were called, by which sandbox, and whether the request was allowed or denied
• Works end-to-end with a local endpoint + at least one external endpoint

Out of Scope (v0)

• Dynamic routing based on prompt content or PII detection
• Automatic redaction or rewriting in-flight
• Enterprise multi-tenant policy enforcement
• Org-wide compliance dashboards

Roadmap (v1+)

• Semantic / PII-aware routing
• Privacy transformations (synthetic rewriting)
• Richer audit and compliance
• Hot policy update without sandbox restart

Links

• Feature definition: architecture/plans/inference-policy.md
• MR: !27

Originally by @zredlined on 2026-02-17T11:21:47.838-08:00