Decision: one endpoint table, one policy engine, one dispatch pipeline

Status: implemented

Problem

The server had three overlapping mechanisms: the route table plus a capability-group gate, a per-endpoint policy engine on top of it, and an operation layer (core/, POST /ops/{id}) that wrapped both with a second id space. Capability switches existed at two layers (group gate and endpoint rule), handlers had an unscoped branch for "no policy", and the same decision was written three times.

Decision

• server/endpoints.py holds the only declaration: ENDPOINTS (id, method, path, group, scopes, limits). RouteTable.mount(name, handler) binds handlers to it. Policy, audit, /capabilities and the client use endpoint ids.
• One engine (PolicyEngine). Layers: default < code (enable_group/disable_group) < preset < policy < env (ABX_SERVER_ENABLED_GROUPS, enable only). A group is an alias that expands to endpoint ids; the most specific selector wins; every denial names the rule.
• One pipeline (SandboxRequestHandler._pipeline): match, authenticate, admit, duplicate-query check, body limit, handler, audit. The engine is always present (no policy means an empty policy) and handlers always receive an EndpointContext.
• Deleted: core/ and POST /ops/{id}, enable_builtin/disable_builtin, ABX_SERVER_DISABLED_GROUPS, the unscoped handler mode, legacy_codes.

Alternatives considered

• Keep /ops with dotted ids as bindings over the table. Rejected: a second id space needs a mapping layer and nothing used it.
• Keep group switches as a separate layer. Rejected: two places to look when a call is denied.

Test Strategy

tests/test_server/test_single_layer.py, test_policy_pipeline.py, test_policy_config.py, test_routes_agents_local.py::TestEveryRoute (every agents route).