/** * PermissionPolicy — data-driven role-based authorization for tool dispatch. * * Closes Neo gap #2 (of 8). Permissions are CROSS-CUTTING — they're not * context engineering, they're a guard ON context-engineering operations * (tool dispatch, skill activation, memory writes, output emission). * That's why this lives in `agentfootprint/security`, parallel to the * provider subpaths. * * Two surfaces, one primitive: * 1. `PermissionPolicy.fromRoles({...}, activeRole)` — declarative, * data-driven, auditable. Production governance. * 2. The PermissionPolicy instance satisfies BOTH: * - `PermissionChecker` interface (async check; consumed by Agent * constructor's `permissionChecker` field) * - sync `isAllowed(toolId)` method (consumed by `gatedTools(...)` * from `agentfootprint/providers`) * * **Status: contract-shaped and tested — independently reproduced against a * local harness, 2026-08-13.** A tool declaring `memory_write` and `user_data` * was put in front of two roles: the read-only role passed the tool-NAME * allowlist, failed the `memory_write` capability check, and the tool executed * ZERO times; the admin role cleared `tool_call`, `memory_write` and * `user_data` and executed once. A separate checker that threw `simulated * authorizer outage` failed CLOSED — a denial carrying the operator's error in * its rationale, the model free to continue without the tool, zero executions * again. The reviewer's own wording: *"the permission checker is a real * execution guard, not merely tool hiding"*, which holds exactly as far as the * policy inputs and the role binding are trustworthy — a fact about a * deployment, not about this class. * * What was exercised: the guard, end to end, by somebody who is not this * library's author — on a DETERMINISTIC LOCAL run whose model was a mock and * whose "authorizer outage" was a thrown `Error` (those tests *"were local and * deterministic, so they consumed no GCP credit"*). What was NOT: **no external * authorization service has answered a `check()` from this repository**, so * this is not the field-validated rung — a real IAM/OPA/entitlements backend, * with its own latency, partial outages and stale caches, remains unexercised. * Runnable: `examples/features/03-permissions.ts` (all four shapes). * * Pattern: Strategy (GoF) for the role-allowlist policy + Adapter * (matches `PermissionChecker` interface so it composes with * existing v2.4 Agent constructor). * * Role: Layer-3 cross-cutting guard. Not Injection. Not provider. * Lives in its own subpath (`agentfootprint/security`). * * @example Read-only role for a support agent * const policy = PermissionPolicy.fromRoles( * { * readonly: ['lookup_order', 'get_status', 'list_skills', 'read_skill'], * support: ['lookup_order', 'get_status', 'process_refund', 'list_skills', 'read_skill'], * }, * 'readonly', * ); * * policy.isAllowed('lookup_order'); // → true * policy.isAllowed('process_refund'); // → false (not in readonly role) * * // As a tool-dispatch gate (composes with gatedTools) * const provider = gatedTools(staticTools(allTools), (name) => policy.isAllowed(name)); * * // As an Agent permissionChecker (the v2.4 surface) * const agent = Agent.create({ provider, model, permissionChecker: policy }).build(); * * @example Per-identity role switching at runtime * const policy = PermissionPolicy.fromRoles({ * readonly: [...], * admin: [...], * }, 'readonly'); * * const adminPolicy = policy.withActiveRole('admin'); * // Same allowlist data; different active role. */ import type { PermissionCapability, PermissionChecker, PermissionRequest, PermissionDecision, ToolCapability } from '../adapters/types.js'; /** * Map of role name → list of tool ids that role is allowed to invoke. * The shape consumers extend over time as new tools / roles arrive. */ export type RoleAllowlist = Readonly>; /** * Map of role name → the ids that role may reach, for a rule map beside the * tool allowlist (9.11.0). A role ABSENT from the map has no rule of that kind, * which means the rule does not apply to it — see {@link PermissionPolicyRules}. */ export type RoleIdRules = Readonly>; /** * The optional rules that sit BESIDE the tool allowlist (9.11.0). * * Both are opt-in and both are enforced only where they are configured, because * every one of them would otherwise change what an existing deployment permits. * A `PermissionPolicy.fromRoles(roles, role)` built without this bag behaves * exactly as it did in every earlier release — and reports `governs` as absent, * so the framework never asks it anything new either. */ export interface PermissionPolicyRules { /** * role → the {@link ToolCapability} values that role may exercise. * * Configured (for ANY role) ⇒ the policy declares it governs all four * capabilities, and the framework starts asking about a tool that declared * them. For a role with a rule, a capability the rule does not list is * DENIED. For a role WITHOUT one, capability checks fall through to the tool * allowlist — the tool already passed it, so nothing changes for that role. * * @example a support role that may read but never write * PermissionPolicy.fromRoles(roles, 'support', { * capabilities: { support: ['memory_read', 'external_net'], admin: [...] }, * }) */ readonly capabilities?: Readonly>; /** * role → the SKILL ids that role may activate with `read_skill`. * * Configured (for ANY role) ⇒ the policy declares it governs `'skill_read'`, * and the agent starts asking per skill: a refused skill's row disappears * from the `read_skill` menu the model reads, and an activation of it is * refused with this policy's own rationale. For a role WITHOUT a rule, every * skill stays visible — silence is not a denial. * * Ids are the bare skill ids (`'refunds'`), not the `skill:` targets: the * prefix is the wire convention that keeps a skill and a tool of the same * name apart, and it is this policy's job to know that, not yours. */ readonly skills?: RoleIdRules; } export interface PermissionPolicyOptions { /** * The role allowlist. Each role maps to the tool ids it can invoke. * Tool ids match the `name` field of `Tool.schema.name` exactly. */ readonly roles: RoleAllowlist; /** * Which role is active for this policy instance. Calls to * `.isAllowed(toolId)` check against this role's allowlist. * Use `.withActiveRole(name)` to derive a sibling policy with a * different active role. */ readonly activeRole: string; /** Optional capability + skill rules (9.11.0). See {@link PermissionPolicyRules}. */ readonly rules?: PermissionPolicyRules; } /** * Data-driven role-based permission policy. Satisfies the v2.4 * `PermissionChecker` interface AND exposes a sync `isAllowed` method * for use with `gatedTools` from `agentfootprint/providers`. */ export declare class PermissionPolicy implements PermissionChecker { private readonly opts; readonly name = "PermissionPolicy"; /** * Which capabilities this policy asks to be consulted about beyond * `'tool_call'` (9.11.0) — DERIVED from what was configured, never a * separate switch that could disagree with the rules. * * Absent when no rules were given, which is what makes every policy built * the old way byte-identical: the framework asks it nothing new. */ readonly governs?: readonly PermissionCapability[]; private constructor(); /** * Factory: build a role-based policy from a role → tool-ids map and * the role active for this instance. * * Throws if `activeRole` isn't a key in `roles` — fail loud at * config time, not at first denied call. */ static fromRoles(roles: RoleAllowlist, activeRole: string, rules?: PermissionPolicyRules): PermissionPolicy; /** * Sync allowlist check. Use as a predicate with `gatedTools`: * * gatedTools(staticTools(allTools), (toolId) => policy.isAllowed(toolId)) * * Returns true iff `toolId` is in the active role's allowlist. * Closes-fail by design: missing role membership = denied. */ isAllowed(toolId: string): boolean; /** * Async check matching the `PermissionChecker` interface — consumed * by `Agent.create({ permissionChecker })`. Wraps `isAllowed` with * the structured `PermissionDecision` envelope (allow / deny + a * `policyRuleId` so observability can trace which role decided). * * ## What each capability is judged by (9.11.0) * * - `'tool_call'` — the tool allowlist. Unchanged since v2.4. * - `'skill_read'` — the active role's `rules.skills` list, when it has one. * Without one, allowed: a policy that never mentioned skills has not * denied any, and silence is not a refusal. * - a {@link ToolCapability} — the active role's `rules.capabilities` list, * when it has one. Without one, it falls through to the tool allowlist * below, which the tool already passed on its `'tool_call'` check — so a * role with no capability rule behaves exactly as it did before. * * The framework only ever sends the last two to a policy that declared * {@link PermissionPolicy.governs}, which this class derives from the rules * themselves — so "unconfigured" and "never asked" cannot drift apart. */ check(request: PermissionRequest): Promise; /** * Skill activation, judged on its own map (9.11.0). * * Deliberately NOT routed through the tool allowlist: a `skill:` target * is not a tool name, so `isAllowed` would deny every skill for every role * the moment a policy was attached — hiding a whole catalogue nobody asked * to hide. A role with no skill rule sees every skill; a role WITH one sees * exactly what it lists. */ private checkSkill; /** * Derive a sibling policy with a different active role. Same role * map; different active role. Useful for per-identity routing * (one policy instance per request, varying active role per caller). * * Returns a NEW PermissionPolicy — original is unchanged. */ withActiveRole(activeRole: string): PermissionPolicy; /** The role name currently active. Useful for observability. */ get activeRole(): string; /** All defined role names. Stable order = registration order. */ get roles(): readonly string[]; /** All tool ids allowed under the current active role. */ allowedToolIds(): readonly string[]; } //# sourceMappingURL=PermissionPolicy.d.ts.map