omnigraph/docs/user/operations/policy.md

Authorization (Cedar policy)

OmniGraph integrates AWS Cedar (cedar-policy = 4.9) for ABAC.

Policy actions

Per-graph actions (bind to Omnigraph::Graph::"<graph_id>"):

1. read — query / snapshot / list branches & commits
2. export — NDJSON export
3. change — mutations
4. schema_apply — apply schema migrations
5. branch_create
6. branch_delete
7. branch_merge
8. admin — reserved for policy-management surfaces (hot reload, audit log, approvals). No call site today.
9. invoke_query — gates invoking a server-side stored query (the queries: registry). Graph-scoped (like admin) — per-branch access is enforced by the inner read / change gate, so a rule that sets branch_scope on invoke_query is rejected. Coarse in this release: an invoke_query allow rule permits any stored query on the graph; a future, additive refinement adds an optional per-query-name scope without changing rules written against the coarse action. Enforced at POST /queries/{name} (see server). A stored mutation is double-gated: invoke_query to reach the tool, plus change for the write itself (the engine _as writers still enforce per the query body).

Server-scoped action (v0.6.0+; binds to Omnigraph::Server::"root"):

10. graph_list — GET /graphs registry enumeration (multi-graph mode)

Server-scoped actions cannot use branch_scope or target_branch_scope — they operate on the registry, not on a graph's branches. A rule cannot mix server-scoped and per-graph actions; split into separate rules. (Runtime graph_create / graph_delete over HTTP are reserved but not shipped; operators add/remove graphs by editing the cluster's cluster.yaml, running omnigraph cluster apply, and restarting the server.)

Scope kinds

• branch_scope — applied to source branch (read, export, change)
• target_branch_scope — applied to destination (schema_apply, branch ops, run ops)
• protected_branches — named list with special rules; rule scopes are any | protected | unprotected

Per-graph vs. server-level policy

A server boots from a cluster (--cluster <dir>), and the cluster's cluster.yaml declares its policy bundles in a policies: section. Each bundle names the scopes it applies_to: a graph id (per-graph rules — read, change, branch_*, schema_apply) or the literal cluster (server-level rules — graph_list).

# cluster.yaml
policies:
  base:
    file: base.policy.yaml
    applies_to: [cluster, knowledge]   # cluster-level + the `knowledge` graph
  alpha:
    file: policies/alpha.yaml
    applies_to: [alpha]                # per-graph: alpha only

A graph with no bundle bound to it has no engine-layer Cedar enforcement. Each graph's HTTP request flows through its bound bundle; the management endpoint (GET /graphs) flows through the cluster-scoped bundle. When no bundle binds cluster, GET /graphs is denied in every runtime state, including --unauthenticated; with bearer tokens configured it returns 403 after admission control because graph_list is not a read-equivalent action. The operator must bind a cluster-scoped bundle granting graph_list to expose /graphs.

Example cluster-scoped bundle:

version: 1
groups:
  admins: [act-andrew]
rules:
  - id: admins-can-list-graphs
    allow:
      actors: { group: admins }
      actions: [graph_list]

Each per-graph rule may use at most one of branch_scope or target_branch_scope. Server-scoped rules (graph_list) take neither — they have no branch context.

Actor for direct-engine writes

The default actor identity for CLI direct-engine (--store) writes is operator.actor in ~/.omnigraph/config.yaml. Override per-invocation with --as <ACTOR> — --as wins, otherwise operator.actor, otherwise no actor. Remote HTTP writes ignore both — they resolve their actor server-side from the bearer token. (Direct-store access carries no Cedar policy; policy lives in the cluster/server.)

CLI

Policy tooling reads a cluster's applied policy bundles: pass --cluster <dir>, and --graph <id> to pick a graph's bundle when several apply.

• omnigraph policy validate — parse + count actors, exit 1 on parse error.
• omnigraph policy test --tests <file> — run the declarative cases in <file> against the selected bundle, exit 1 on any expectation mismatch.
• omnigraph policy explain --actor … --action … [--branch …] [--target-branch …] — show decision and matched rule.
• omnigraph --as <ACTOR> <subcommand> — set the actor for the duration of one invocation. Effective for change, load (and its deprecated ingest alias), branch create|delete|merge, and schema apply against a direct (--store) graph. Rejected on a served write (--server): the actor is bearer-token-resolved server-side, so --as can't set it there.

Enforcement

Policy is a property of the engine, not the transport. Every mutating write — mutate_as, load_as (the deprecated ingest_as shims route through it), apply_schema_as, branch_create_as, branch_create_from_as, branch_delete_as, branch_merge_as — consults the policy gate at the head of the method. The gate fires identically whether the call originates from the HTTP server, the CLI, or an embedded SDK consumer. When no policy is installed (the dev/embedded default) the gate is a strict no-op; when one is installed and the call site forgets to thread an actor through, the gate fails closed rather than silently bypassing.

Server runtime states

The HTTP server classifies its startup configuration into one of three states based on whether bearer tokens are configured and whether a policy file is set. The state determines what happens to a request that reaches the authorization gate without a matching policy permit.

State	Tokens	Policy file	Behavior
Open	no	no	Every request is permitted. Refuses to start unless --unauthenticated or OMNIGRAPH_UNAUTHENTICATED=1 is set — the operator must explicitly opt in.
DefaultDeny	yes	no	Every authenticated request for an action other than read is rejected with HTTP 403. Closes the "tokens but forgot the policy file" trap — an operator who sets up auth and forgot to point at a policy file used to ship the illusion of protection.
PolicyEnabled	yes	yes	Authenticated requests that reach a configured policy engine are evaluated by Cedar. Server-scoped actions still require a cluster-scoped policy bundle.

The server refuses to start for the "no tokens, no policy, no flag" cell and for "policy file, no tokens" — instead of silently shipping an open instance or a policy-protected server that can only 401.

Server-side, request authorization still runs at the HTTP boundary — that's where actor identity is resolved from the bearer token and where admission control / per-actor rate limits live. Engine-layer enforcement is the defense in depth layer: it catches CLI direct-engine writes, embedded SDK consumers, and any future transport that hasn't (or won't) re-implement the HTTP boundary's authorization. Both layers consult the same Cedar policy, so decisions cannot disagree.

Coarse vs. fine enforcement

There are two enforcement points, each with non-overlapping responsibilities:

Layer	Question it answers	Where it fires
Engine-layer (coarse)	Can this actor invoke this action against this branch / branch-transition?	The policy gate at the head of every _as writer; one Cedar decision per call.
Query-layer (fine)	For the rows / types this action actually touches, which can the actor see or modify?	Per-row predicates pushed into the query plan. Not yet implemented.

The engine-layer gate keeps its resource scope deliberately at branch granularity (graph, branch, target branch, branch transition). Per-type and per-row authority is the query-layer's job; conflating them in the engine-layer scope would create two places per-type policy could be evaluated and a drift surface between them.

Actor identity (signed-claim-only)

The actor identity used for every policy decision comes from the matched bearer token — never from a client-supplied request header, query parameter, or body field. The server resolves the token at the auth middleware boundary, looks up the actor it was minted for, and overwrites whatever the handler may have placed in the policy request. Clients cannot set actor_id directly.

This is intentional. Trusting client-supplied identity for authorization is "asking the attacker if they're an admin" — Supabase's RLS history names the same footgun. The chokepoint lives at the server's auth boundary: a request with Authorization: Bearer <token-for-actor-A> plus X-Actor-Id: actor-B always evaluates as actor A, never as actor B.

If you find yourself wanting to let clients override actor_id for impersonation, delegation, or service-account flows — that's a feature, but it needs explicit design (e.g., signed delegation claims, an On-Behalf-Of audit trail). It is not a convenience knob.