The engine already honored requireAllowlist + allowlistMiss, and the
commerce example configured a seller allowlist by hand -- but there was no
reusable, named "only show allowlisted publishers" preset, the symmetric
twin of the denylist. Integrators (Tessera) punted on allowlist mode for
lack of one.
Add ALLOWLIST_POLICY (registry name "allowlist"): a publisher is shown only
when an authorized actor has contributed a user-allow for them, and every
unlisted publisher misses and is hidden across all four dimensions.
Administrative denials still apply on top, and protected actors (root,
moderators) are never gated -- the people who curate the list can always
see. Populate userAllow exactly as the denylist: signed contributions
resolved from relays (e.g. NIP-32 allow labels).
Exported from @bitgate/core, registered in POLICY_PRESETS, documented in the
README, and covered by presets.test.js (unlisted hidden, allowlisted shown,
protected bypasses).
Addresses the "allowlist / whitelist mode" item of the Tessera integration
feedback (docs/TODO.md).
Co-Authored-By: Claude Opus 4.8 <[email protected]>
Claude-Session: https://claude.ai/code/session_01BXQCvJPA6sVSQ5j8uBd9bB
A second adversarial pass over areas the first audit did not reach: codecs,
transport internals, and the verify/outbox/private-mute paths added since.
Reliability/CI (most urgent)
- The previous commit's Node-safety base-class fallback lost the HTMLElement
type, so defineElement() and this-typing broke. It only appeared with dist
cleared, so it passed locally with stale declarations and would fail CI from
a clean checkout. Annotated the fallback; cold lint/test/build now pass.
Security
- Relay URLs from a followed contact's kind:10002 flow straight to
new WebSocket(url). normalizeRelayUrl only checked the ws(s) prefix, so a URL
containing CR/LF or spaces passed through to the socket constructor — a
CRLF-injection shape. Now validated with the URL constructor and rejected for
any control character or whitespace, with a second guard at the socket
boundary regardless of caller.
- canonicalStringify recursed without cycle detection: a self-referential value
was a stack-overflow DoS rather than a wrong answer. Now guarded with a
WeakSet; a repeat serializes as a sentinel.
- The viewer's private mute payload was JSON.parsed with no size bound. Capped
at 128 KiB, checked before spending a decrypt on it.
What held up under this pass: address-coordinate parsing, the outbox per-author
relay cap (3), verifier handling of NaN/Infinity/string timestamps, and mute
tag robustness against non-string entries.
861 tests; audit clean; cold gates green.
Critical
- Any pubkey could publish a kind:30078 roles document and become root of a
deployment, granting itself every capability and denying anyone. Roles are now
accepted only from the configured root, compared against the constructor
value rather than current authority state so a bad ingest cannot move the
goalposts. Reproduced end to end before the fix; regression test added.
- Administrative state now fails closed without signature verification. Without
it, event.pubkey is an unauthenticated claim, so an authorship check alone
proves nothing — a relay can put the root's key on an event the root never
wrote. trustUnsignedEvents opts in for local and test use.
High
- Report categories could evade moderation entirely: resolveThresholds used
bracket truthiness, so "constructor" resolved to the Object constructor (all
gates undefined) and "__proto__" to Object.prototype, both bypassing the
default threshold. Own-property lookups and null-prototype tables throughout
core — policy, evaluator, evidence, authority, presets.
- Live subscriptions bypassed signature verification entirely; an attacker only
had to wait for the subscription rather than the load.
- The default administrative query asked relays for every kind:30078 and
kind:30000 event in existence, with no author, identifier, or limit. Now
filtered to governance identifiers with a bounded limit.
Medium
- Cached state was trusted as authority: a cache naming a different root is now
rejected, so same-origin tampering cannot persist a takeover.
- Provider attribute changes silently stopped working after the first one, and
concurrent starts leaked a transport. Generation-guarded.
- Unbounded growth in the report store, mute store, decision cache, and the
transport's per-subscription dedupe set — all now capped.
Also restores the on-wire `d` identifier to <namespace>:governance:<scope>:v1.
The BitGate rename replaced "governance:" too broadly and silently changed the
event format; both encoder and parser changed together, so round-trip tests
never noticed. The identifier names the kind of document, not the library.
Adds docs/security.md and 21 regression tests reproducing each finding.
832 tests.
Closes the four spec gaps from the audit, verified against the NIPs repo.
NIP-65 (kind 10002) — the one that was actually breaking things. Mute lists are
published by each contact to their own write relays, so querying a fixed relay
set finds only the subset who happen to write where this deployment reads. The
trust graph was silently under-fed and counts came out low, which looks like
"nobody muted them" rather than like a bug. Adds a relay-list codec, per-author
grouping with a fallback for accounts that have published no list, and
on-demand relay connections in the transport so an outbox query can reach
relays that were never configured.
NIP-51 private entries — the viewer's own private mutes are now decrypted
through their signer (NIP-44) and applied as viewer mutes. Someone who mutes
privately previously got no effect at all. The privacy boundary is asymmetric
and enforced at the codec: another account's private list is never decoded,
and the check lives there rather than in the caller's discipline.
NIP-02 (kind 3) — the runtime can build the trust graph itself, so
<bitgate-provider> no longer needs a host to supply a follow list. Set
trust="manual" to opt out.
NIP-56 vocabulary — SOCIAL_POLICY declared thresholds for "harassment", which
is not a NIP-56 type, so no standard client could ever trigger it. Replaced
with the defined vocabulary and asserted in tests. Commerce keeps its
marketplace-specific categories, now documented as deliberate, with defaults so
unclassified reports are never silently dropped.
Two bugs found by the new tests:
- checkout had no default threshold, so a flood of "other" reports did nothing
at the one place money changes hands. Now earns review, never auto-denial.
- describe() spread diagnostics after a derived field, so a stale zero shadowed
the live relay-list count.
Adds docs/nips.md recording coverage, deliberate choices, and remaining gaps —
NIP-32 labelling being the largest.
811 tests.
Renames the project and makes it as easy to adopt as BitLogin, which was the
stated bar and which it did not previously meet.
Rename (ADR 0005): @nostr-governance/* -> @bitgate/*, governance-* elements ->
bitgate-*, createGovernanceRuntime -> createBitGate. "Governance" reads as DAO
voting to this audience and named a category rather than a product; BitGate
names the benefit and joins the bit- family that BitVid and BitRoad compose
with. Domain types keep their names — the product is BitGate, the subject
matter is still governance. The migration plan is left as written; this ADR
records the change rather than rewriting history.
Drop-in distribution (ADR 0006): adopting BitGate meant supplying a transport,
authoring a policy, and wiring every element in JavaScript. Now:
- createRelayTransport(urls) — WebSocket transport with reconnection, backoff,
subscription replay, and cross-relay dedupe
- policy presets (social, commerce, admin-only) addressable by name; the engine
still carries no thresholds of its own
- <bitgate-provider> — configuration in markup, with descendants finding the
runtime through a bubbling context event rather than DOM walking, which
breaks across shadow boundaries
- npm run build:widget — a self-contained dist/bitgate.js
That last one came from running it in a browser rather than from a test. Bare
specifiers resolve in Node and vitest but not in a browser, so the widget
source worked everywhere except a plain static page — the one place it exists
to serve. The bundle is not committed: consumers build from reviewed source.
Also found by running it: the provider's ready event fires during element
upgrade, before a page that defines elements after parsing can attach a
listener, so it now exposes `ready` as a promise too.
Four bugs fixed along the way, three found by new tests and one by the browser:
admin state stayed silent when a roster change altered capabilities but no
denials; the relay transport sent every subscription twice on connect; the
commerce preset computed a meaningless visibility verdict at checkout and could
hide a listing from its own seller; <bitgate-status> threw on an unusable
decision.
759 tests; lint, typecheck, and both builds clean.
The performance fixture the plan asks for found a real defect on its first run:
evaluation was quadratic in state size. Both the snapshot and its fingerprint
were rebuilt for every target, so each of 5,000 evaluations walked every report
and every mute list. A 5,000-target pass took ~47 seconds.
- the runtime memoizes the materialized snapshot and its fingerprint until a
store reports a change
- evaluateMany computes the fingerprint once and shares it across the batch
- decisions are cached per (profile, target); report and mute changes
invalidate only the targets they touch, while administrative, trust, policy,
and viewer changes drop the whole cache
- cached decisions are deeply frozen, since the cache hands the same object to
every caller
Same pass now takes ~85ms cold, ~6ms warm.
Adds tests/performance (5,000 targets, 500 authors, 100 muters per author) and
tests/security (one case per enumerated requirement testable at this layer).
605 tests; lint, typecheck, and the declaration build pass from a cold checkout.
Proves the engine did not stay video-specific during extraction.
- application adapter contract in core (createApplicationAdapter,
evaluateObject, collectTargets): an object names the targets that govern it
and the strictest verdict reaching any of them wins, so denying a seller
takes their whole catalogue down with no cascade logic in the application
- examples/commerce: a marketplace on the public API covering all ten
proof-of-fit behaviors the plan requires, including that no core change was
needed. One product, one snapshot, four surfaces, four different verdicts.
- README, integration guide, per-example README, CHANGELOG, four ADRs, and
GPL-3.0-or-later LICENSE (the license was listed in Phase 1 but missing)
- checkpoints rewritten to reflect actual state and to say plainly which
migration-plan phases belong to the consumer repos rather than here
Fixes a real gap found by the proof-of-fit: composeDecisions kept only the
first decision's evidence, so an object governed by several targets lost the
evidence that drove its verdict. Evidence now merges — counts summed,
contributor lists unioned, flags or-ed.
556 tests; lint, typecheck, and per-package type build clean.
Stores (Phase 4): admin, trust graph, reports, trusted mutes, policy, and
overrides. Each owns one slice of state and emits only when effective state
changes, so a consumer can treat every emission as a real invalidation.
- report storage is keyed target -> reporter -> category, making deduplication
structural: relays redeliver constantly and a redelivery must not inflate a count
- contribution lists replace by (actor, kind) and ignore out-of-order older
documents, so a late delivery cannot resurrect removed entries
- mute lists prune against an injected clock
- viewer switching clears blocks, mutes, and overrides but keeps operator seeds,
which are configuration rather than viewer data
Runtime: event routing, chunked report subscriptions (200 per filter, not one
per target), viewer lifecycle, evaluation, diagnostics, idempotent teardown.
Transport, signer, and storage are injected; in-memory doubles ship for tests.
Commands (Phase 5): every mutation checks capability locally before the signer
is invoked, so an unauthorized actor cannot produce a signed event at all.
Protected actors are refused at the command boundary as well as during
reduction. Partial relay acceptance succeeds while keeping failures visible.
Also fixes the type build: every package's "types" field pointed at a path the
build never produced, so consumers got no types at all. Declarations now emit
per package and resolve.
103 runtime tests; 505 total.
The evaluator was a skeleton with invented thresholds (5/3/10/5) and two
unimplemented branches. Rebuild it against the migration plan's spec and the
behavior characterized from the pinned reference commit.
Core:
- four independent decision dimensions (ranking, visibility, interaction,
transaction) with severity ladders and max-merge composition
- stable reason identifiers replacing free-text strings
- structured evidence, redacted unless a profile opts in
- administrative-state reduction with capability filtering at merge time, so
revoking a role drops that actor's entries immediately
- protected actors stripped from every denial set
- snapshot fingerprints for cache invalidation
- generic policy engine: applications supply profiles and category thresholds;
the core carries none of its own
Characterization:
- 34 generated fixtures covering every behavior the plan's Phase 0 requires
- conformance suite that actually evaluates them, replacing a placeholder that
asserted a list's length
- reference thresholds (3/3/20/5, 60-day mute window) live in bitvid-compat as
an application profile, not as engine defaults
Fixes found while building:
- reduceAdminState accepted malformed target identifiers, because getTargetKey
formats anything it is given
- community-user-deny fired for any contributor rather than only for
community-curated sources
273 tests pass; typecheck and the reference-import guard are clean.
Replace pnpm-only workspace: protocol with npm workspace deps, pin BitVid and
BitRoad reference commits, add Node types, and make the reference-import check
scan package graphs and source for forbidden bitvid/bitroad imports.
