Policy engine

Hostwright has a local policy module for deterministic and explainable safety decisions. The policy engine evaluates entirely in memory, using only the inputs it is handed.

Every decision records a category, a stable reason code, severity, subject, message, and remediation text, so a blocked plan tells you why it was blocked and what to change.

Where policy runs

• Planner safety checks: desired identities, host ports, bind addresses, privileged ports, mounts, and secret-like environment values are evaluated before becoming plan issues.
• Cleanup classification: ownership-backed and observed-only resources are classified through policy decisions, preserving the dry-run/token gates.
• Image policy: imagePolicy: require-digest failures are explained locally, without registry calls.
• Secret references: unresolved references fail closed, without carrying raw keychain labels in messages.
• Advisory scheduling: the local scheduler reuses policy decisions as explanation and scoring inputs; it never places workloads or reserves capacity.

Categories

Category	Current behavior
Identity	Empty project or service identity blocks planning and mutation.
Ports	Duplicate desired host ports and observed host-port conflicts are blockers. Privileged host ports are warnings (the create path rejects them before mutation anyway).
Exposure	Broad bind addresses are blockers. Tunnels, DNS, cloud, and reverse-proxy scopes are unsupported.
Mounts	Ambiguous references, host-root mounts, and parent-traversal sources are blockers.
Images	require-digest failures are local string-policy blockers, evaluated without registry lookups.
Environment and secrets	Secret-like env values are redaction warnings; unresolved secret references are blockers before mutation.
Cleanup	Only exact Hostwright-owned non-running containers can become eligible; everything else fails closed.
Lifecycle	Create, start, stop, restart, remove, and update actions are allowed only inside a confirmed, dependency-ordered lifecycle plan or cleanup gate. An action outside that plan, or one whose preconditions no longer hold, is a blocker.
Untrusted manifests	Unsupported fields are blockers.
Accelerators	GPU, ANE, Metal, Core ML, MLX, and accelerator scheduling are blockers in current core scope.

Defaults and overrides

The default policy is local, deterministic, and fail-closed. Current overrides are code-level configuration inputs only: the privileged-port warning threshold, the broad-bind block list, the redaction policy, and the image digest policy.

There is no remote policy service, central policy distribution, silent bypass, policy-driven runtime mutation, or automatic remediation. The team workflow layer consumes these deterministic decisions through explicit local profile and approval files; policy evaluation itself stays non-mutating and strict-only.

Reason codes and profile mechanics are documented in the policy reference.