Commit graph

2 commits

Author SHA1 Message Date
d3a9ea7ad9 feat(server): federal-mode reference Rego (parent-deny-is-absolute)
Ship a second parity-tested Rego policy that flips the cascade's
leaf-allow-overrides-parent-deny rule for NIST AC-6 conformance.

Standard cascade (existing access.rego, mirrors internal Go evaluator):
  Bottom-up walk; first explicit match wins; deny-first within a level.
  A leaf-level allow CAN override an ancestor's deny. This is the
  cascade's intentional delegation property — a project-owner who
  re-allows a previously-denied collaborator works as expected.

Federal mode (new access_federal.rego):
  Any deny anywhere along the chain is absolute. An allow only matters
  if no level (any depth) has denied the same email. Required by
  NIST AC-6 default expectations: a central admin's deny at the root
  must be unbypassable by a tenant who controls a subtree's .zddc.

Operators run real OPA with this Rego and point ZDDC_OPA_URL at it;
the internal Go evaluator stays on the commercial cascade. The
toggle is "which policy does your OPA evaluate," not a knob inside
zddc-server.

Surfaced via --print-rego flag:

  zddc-server --print-rego               # standard (default)
  zddc-server --print-rego=standard      # same
  zddc-server --print-rego=federal       # AC-6 strict variant

Parity test (federal_parity_test.go) compiles both Regos and asserts:
  * They AGREE on every cascade scenario where no ancestor-deny
    intersects a leaf-allow (most cases).
  * They DISAGREE — by design — on the three scenarios where the
    AC-6 rule differs:
      - "leaf allows what parent denied" → standard allows, federal denies
      - "deep leaf re-allows after middle deny" → same
      - "glob deny at root + specific allow at leaf" → same

Cross-checks the divergence flag explicitly so any future change that
accidentally collapses the two policies fails the test.

Closes the AC-6 row of the federal-readiness gap analysis (now marked
"partially complete" in zddc/README.md — the full bullet would be a
built-in --policy-mode=federal toggle that also flips the in-process
Go evaluator).

Production binary unchanged at 13.1 MB (Rego files embedded as bytes;
OPA library remains test-only).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
2026-05-04 18:05:44 -05:00
a01315fd00 feat(server): reference Rego, parity test, decision cache, listing ETags
Phase 2 enhancements to the policy decider, plus listing-level ETags
that benefit every deployment regardless of decider mode.

Reference Rego policy
---------------------
internal/policy/rego/access.rego mirrors InternalDecider's semantics
exactly — bottom-up walk, deny-first within a level, default-deny when
HasAnyFile=true, glob matching with @-boundary semantics (special-cased
bare "*" because OPA's glob.match treats empty delimiters
inconsistently for that pattern).

Embedded into the binary via go:embed; --print-rego dumps it to stdout
so federal customers standing up an external OPA can use it as a
parity-tested baseline:

    zddc-server --print-rego > /etc/opa/policies/zddc-access.rego

Parity test runner
------------------
parity_test.go imports the OPA Go module as a TEST-ONLY dependency
(github.com/open-policy-agent/opa@v0.70.0). Every fixture from the
internal Go evaluator's test set runs through both implementations;
any divergence fails CI. The test-only import means production
binaries (built by `go build ./cmd/zddc-server`) stay OPA-free —
release-flag binary size unchanged at ~13 MB.

The parity test caught a real bug on first run: bare "*" patterns
didn't match through OPA's glob.match with empty delimiters. Fixed
in access.rego with a special-case rule. This is exactly the kind of
subtle drift the parity guard exists to catch.

External-mode decision cache
----------------------------
HTTPDecider is now wrapped in a cachingDecider with a default 1s TTL.
Bursty patterns like .archive listings (one OPA round-trip per entry
before, one per (email, decision-input) tuple per TTL window after)
amortize cleanly. Verified: 20 identical /D/ requests produce 1 OPA
hit with cache, 40 hits without (each listing makes 2 ACL queries).

ZDDC_OPA_CACHE_TTL knob (default 1s) lets operators tune. 0 disables.
1s matches the fsnotify watcher debounce window — staleness is
bounded the same way other policy-edit propagation already is.
Internal mode unchanged; the in-process Go evaluator is already
cheaper than a cache lookup would be.

Listing ETags
-------------
GET / (project list) and GET /<dir>/ (directory listing JSON) now
carry content-hash ETag + Cache-Control: private, max-age=0,
must-revalidate. SHA-256 of the rendered JSON, truncated to 16 hex
chars (64 bits — collision risk on a listing of any realistic size
is vanishingly small).

Server-side caching deliberately not added: it would require
mtime-based invalidation, and Azure Files SMB mounts (a common
deployment substrate) don't support fsnotify reliably. The
content-hash ETag delivers the bandwidth savings (304 on identical
fetches) without depending on watcher correctness — the hash is the
actual response, so it can't lie about staleness regardless of
underlying watcher behavior.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
2026-05-04 17:46:24 -05:00