rose-ash/plans/acl-on-sx.md

acl-on-sx: Access Control on Datalog

rose-ash needs fine-grained, explainable, federation-aware access control. Subjects (users, groups, roles, services) × actions (read, edit, comment, moderate, federate) × resources (pages, posts, threads, peers). Decisions must come with a trace — not just permit/deny, but why.

Datalog's bottom-up rule engine produces transparent permit/deny chains: the proof tree is the audit trail. Inheritance over groups + resource hierarchies is recursive Datalog in one rule. Federation extends naturally — fed-sx replicates ACL facts, peers reason over the union.

End-state: a Datalog-on-SX layer specifically for ACL, with explanation API, audit log, and federation extension. Reuses lib/datalog/ evaluator and term model where possible.

Status (rolling)

bash lib/acl/conformance.sh → 89/89 (Phases 1-3 complete)

Ground rules

• Scope: only touch lib/acl/** and plans/acl-on-sx.md. Do not edit spec/, hosts/, shared/, lib/datalog/**, or other lib/<lang>/. You may import from lib/datalog/ (its public API in lib/datalog/datalog.sx); do not copy or modify Datalog code.
• Shared-file issues go under "Blockers" with a minimal repro; do not fix here.
• SX files: use sx-tree MCP tools only.
• Architecture: thin layer on top of lib/datalog/. Define schema, surface API, audit + federation hooks. The rule engine itself is Datalog's.
• Watch for shared patterns going into lib/guest/ — both acl-sx and mod-sx need rule-engine plumbing. If you find shared shape, flag it for extraction (don't extract yet — wait for mod-sx to start).
• Commits: one feature per commit. Keep Progress log updated and tick boxes.

Architecture sketch

ACL declarations (SX)            User query
        │                            │
        ▼                            ▼
lib/acl/schema.sx             lib/acl/api.sx
  — subject sorts                — (acl/permit? subj act res)
  — resource sorts               — (acl/explain subj act res)
  — action sorts                 — (acl/audit subj act res :allowed?)
  — fact schema                       │
        │                             ▼
        ▼                       lib/acl/engine.sx
lib/acl/facts.sx                  — builds Datalog query
  — actor(id, kind)               — invokes lib/datalog/
  — resource(id, kind)            — extracts proof tree
  — member_of(actor, group)            │
  — child_of(res, parent)              ▼
  — grant(actor, act, res)        lib/acl/audit.sx
  — deny (actor, act, res)          — persistent decision log
                                    — query API

Phase 1 — Direct grants

• lib/acl/schema.sx — sorts: subject {user, group, role, service}, action, resource {page, post, thread, peer}
• lib/acl/facts.sx — actor, resource, grant, deny predicates as Datalog EDB
• lib/acl/engine.sx — (permit? subj act res db) reduces to Datalog query
• lib/acl/api.sx — public (acl/permit? ...) taking implicit current db
• lib/acl/tests/direct.sx — 15+ cases: direct grant, missing grant, explicit deny
• lib/acl/scoreboard.{json,md} baseline
• lib/acl/conformance.sh runs the suite

Phase 2 — Inheritance

• member_of(actor, group) chain — group grants apply to members (transitive)
• child_of(res, parent) chain — parent grants apply to children (transitive)
• role expansion — role contains list of (action, resource) tuples
• deny-overrides — explicit deny wins over inherited allow
• lib/acl/tests/inherit.sx — 25+ cases: nested groups, deep resource trees, conflict resolution, deny precedence
• document the deny-overrides choice in plan

deny-overrides policy (the choice)

Encoded as stratified negation: permit(S,A,R) :- eff_grant(S,A,R), not eff_deny(S,A,R). Both eff_grant and eff_deny inherit through the same member_of (group/role) and child_of (resource) chains. Consequences:

• An explicit deny on the exact (S,A,R) defeats any inherited allow.
• A group-level or ancestor-resource deny inherits down and defeats a member's/descendant's grant — deny is authoritative across the closure, not only at the leaf. This is the fail-safe reading: the most permissive interpretation of "deny wins" would let a narrow grant escape a broad deny; we chose the opposite.
• Deny is dimension-scoped: a deny on (S, edit, R) never blocks (S, read, R).

Stratifiable because neither eff_grant nor eff_deny depends on permit; permit sits in a strictly higher stratum. Termination is guaranteed — recursion is only over EDB member_of/child_of (no function symbols), so cyclic membership/containment reaches a fixpoint rather than looping (tested).

Phase 3 — Explanation + audit

• (acl/explain subj act res) → {:allowed? T :proof <tree>}
• proof tree extracts from Datalog's derivation
• lib/acl/audit.sx — append-only decision log (in-memory + serializer for disk)
• (acl/audit-tail n) for recent decisions
• lib/acl/tests/explain.sx — proof correctness, audit completeness

proof reconstruction (the choice)

lib/datalog/ records derived facts but not provenance, so lib/acl/explain.sx reconstructs the proof by goal-directed search over the saturated db: for a ground goal, find the first ACL rule (in acl-rules order) whose body holds, take the first dl-query solution binding the rest, recurse on each body literal; negated literals become verified :neg-ok leaves. The Datalog derivation graph is a DAG (a fact may hold many ways) — we pick ONE canonical proof: first-rule, first-solution, with EDB/direct rules ordered first so proofs bottom out quickly. A depth cap (64) guards pathological cyclic data. acl-explain returns {:allowed? :proof :reason}; on denial :reason carries the blocking eff_deny proof (explicit or inherited) when one exists, else nil (no grant). Audit log is append-only with monotonic seq numbers (no wall-clock, for determinism); acl-audit-decide! is the logged path, acl-permit? stays pure.

Phase 4 — Federation

• peer trust facts — peer(addr, kind), trust(peer, level)
• delegated grants — delegate(peer, actor, action, resource)
• cross-instance permit chain — query asks local + queries trusted peers via fed-sx
• revocation propagation — fact retraction across federation
• lib/acl/tests/fed.sx — federated grant chains (mock fed-sx transport in tests)

Progress log

• Phase 1 complete (24/24). ACL is a thin layer over lib/datalog/:
  • schema.sx — sorts (subject/resource kinds, well-known actions) + EDB predicate arity table + acl-fact-valid? validator. Schema is data, since Datalog is untyped.
  • facts.sx — acl-actor/acl-resource-fact/acl-grant/acl-deny constructors returning Datalog fact tuples.
  • engine.sx — owns the ruleset acl-phase1-rules and reduces decisions to dl-query. acl-build-db = dl-program-data facts rules; acl-permit? = non-empty (permit S A R) query.
  • api.sx — acl/load! rebuilds an implicit current db; acl/permit? queries it. (Slash-symbols like acl/permit? parse fine as single tokens.)
  • deny-overrides encoded as permit(S,A,R) :- grant(S,A,R), not deny(S,A,R). Stratifies cleanly because deny is EDB-only (no rule derives it). Verified: grant+deny on same (S,A,R) → denied.
  • Conformance: conformance.conf (datalog preloads + acl modules) + thin conformance.sh wrapper over lib/guest/conformance.sh. Scoreboard generated by the shared driver.
  • Shared-plumbing note (for eventual lib/guest/rules/): the build-db = dl-program-data(facts, rules) + decide = non-empty ground query shape is exactly what mod-sx (Prolog moderation) will also need. The reusable seam is engine.sx's two functions — facts→db and ground-query→bool — both pure pass-throughs to the rule engine. Not extracting yet (wait for mod-sx as second consumer per ground rules).
• Phase 2 complete (54/54, +30 inherit). Extended acl-rules with eff_grant/eff_deny derived relations; member_of carries both group and role membership, child_of carries resource trees, role_grant confers a role's (action,resource) capabilities. Direct grants unchanged (base case of eff_grant), Phase 1 suite still green. Constructors acl-member-of, acl-child-of, acl-role-grant added; schema arity table extended. See the deny-overrides policy section above. Verified cyclic membership terminates.
  • Shared-plumbing update: the reusable seam is still just engine.sx's facts -> db + ground-query -> bool. The inheritance rules are ACL-specific (group/resource/role vocabulary); mod-sx will have its own. So the lib/guest/rules/ extraction stays at the build/decide level, not the ruleset level.
• Phase 3 complete (89/89, +35 explain). Added explain.sx (proof reconstruction, see policy section above), audit.sx (append-only log), and extended api.sx with acl/explain/acl/audit/acl/audit-tail. No engine changes — explanation reads the same saturated db the decision uses.
  • Substrate gotcha: the host = compares symbols by interned identity, which is unstable across dl-query saturation/substitution within a single image — the same two structurally-equal symbol-lists compared = true once and false moments later in the REPL. Conformance runs in a fresh process per suite so it's deterministic there, but test assertions now use a name-based acl-et-eq? (compare symbols via symbol->string), matching the datalog suite's dl-api-deep=? convention. Worth flagging to the kernel owners but out of acl scope.
  • Tooling note: sx-tree path-based edit tools (sx_replace_node, sx_read_subtree with a path) ignored the path argument in this worktree (always resolved to index 0 / [0,..]), in BOTH (a b c) and (a,b,c) forms. sx_write_file, sx_validate, sx_find_all, sx_summarise, sx_eval all work; used full-file rewrites instead of path edits throughout.

Blockers

(loop fills this in)