Files
gbrain/test/operations-trust-boundary.test.ts
T
0bbaed2e48 v0.42.67.0 feat(migrate): provider-agnostic embedding migration — the path off ZeroEntropy (#3390, fixes #3391) (#3459)
* feat(migrate): provider-agnostic embedding migration service — the path off ZeroEntropy (#3390)

- gbrain migrate embeddings --to <provider:model> (alias: retrieval-upgrade):
  plan + cost preflight, consent gate (--yes / TTY confirm / non-TTY exit 2),
  live probe against the target provider before any mutation, env-override
  gate, schema dimension transition via the shared runSchemaTransition,
  dual-plane config write, NULL-signature-inclusive invalidation, query-cache
  purge, resumable re-embed through the standard embed pipeline (single-flight
  locks, backoff, pacing, stderr progress). Killed runs resume by re-running
  the same command; the NULL-embedding column is the checkpoint.
- #3391 root-cause fix (both engines): countStaleChunks / sumStaleChunkChars /
  invalidateStaleSignatureEmbeddings accept includeNullSignature to lift the
  v108 grandfather clause; embed --stale warns loudly when a model swap
  leaves NULL-signature pages in the old embedding space, and
  --include-null-signature re-embeds them. Default sweep behavior unchanged.
- knobs_hash v=12 → v=13 (prov=default legacy callers must not be served
  pre-migration cache rows).
- migrate_embeddings op: scope admin, localOnly, hidden cliHints, hard
  remote refusal, needs_confirmation without yes=true.
- One-shot post-upgrade ZE-sunset banner (ze_sunset_notice_shown) for brains
  resolving to a zeroentropyai:* embedding model or reranker.
- doctor's dimension-mismatch repair hint now names the real command.
- Docs: docs/guides/embedding-migration.md, KEY_FILES entries, spend-controls
  gate row. Tests: PGLite unit + full-lifecycle flow (interrupted-run resume),
  real-Postgres e2e (pgvector DDL path + #3391 predicate parity).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* fix(test): satisfy check:test-isolation + bump the remaining knobs_hash pins

- test/migrate-embeddings-flow.test.ts → .serial.test.ts: the file holds a
  temp GBRAIN_HOME + an installed fake embed transport for its whole
  lifecycle (beforeAll → afterAll), which withEnv() can't wrap. This also
  fixes the CI shard-pollution failure in
  test/ai/recipes-existing-regression.test.ts (that file passes solo on both
  master and this branch; the flow test's configureGateway + provider-key
  deletion was leaking into it inside the same shard process).
- test/embedding-migration.test.ts: env-override case now uses withEnv().
- Bump the three remaining KNOBS_HASH_VERSION pins to 13
  (cross-modal-phase1, search-alias-resolved-boost, search/knobs-hash-reranker).
- Docs + llms bundles follow the test rename.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* chore(test): wire the new Postgres e2e into the smart e2e selector map

Changes to embed.ts / embedding-migration.ts / retrieval-upgrade-planner.ts /
postgres-engine.ts now trigger test/e2e/migrate-embeddings-postgres.test.ts —
the #3391 stale predicates and runSchemaTransition's DDL path behave
differently on real pgvector than on PGLite, so the smart selector has to know.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* feat(migrate): consult spend.posture in the embedding-migration consent gate

The brief asked the gate to honor spend.posture; it previously didn't read it
at all. Now it does — but deliberately does NOT bypass on tokenmax: posture
waives the spend CEILING, and this gate also guards a destructive schema
rebuild (existing vectors dropped, retrieval degraded until the re-embed
finishes). Under tokenmax the dollar figure is marked informational on stderr
and the confirmation is still asked; --yes stays the single scripted bypass.

Pinned by a new case in the flow test so a later refactor can't quietly turn
posture into a bypass. Guide + spend-controls table updated to match.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* wip: blocker fixes

---------

Co-authored-by: Garry Tan <garrytan@gmail.com>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
2026-07-27 17:48:37 -07:00

269 lines
10 KiB
TypeScript

/**
* v0.39 trust-boundary contract test (GAP 3 of the e2e-test-wave audit).
*
* Hybrid design (D7 — pure + targeted handler invocation):
*
* - Pure assertions over ALL operations (~74 ops): scope annotations
* present + correct; localOnly ops are filtered out of the canonical
* mcpOperations list; hasScope semantics work for the standard tiers.
*
* - Handler-invocation cases for ops that are NOT localOnly but DO
* enforce remote/scope at the handler layer (defense-in-depth where
* it actually fires in production):
*
* * submit_job — name='shell' + ctx.remote=true MUST reject
* (the HTTP MCP shell-job RCE class, F7b)
* * search_by_image — image_path + ctx.remote=true MUST reject
* (D18 P0 source-isolation leak class)
*
* `file_upload` and `sync_brain` are intentionally NOT in the
* handler-invocation set — both are localOnly, so the canonical
* filter removes them from mcpOperations and the HTTP path never
* reaches their handlers. Calling their handlers with remote=true
* tests an impossible production path (codex CMT-3). The defense-
* in-depth strict-mode checks inside those handlers still exist;
* they're proven by the localOnly-filtered-out contract here.
*
* Criterion for the curated sensitive-ops list:
* ops whose HANDLER (not transport) has been broken historically.
* Add an op here when a real exploit class is fixed at the handler
* level; remove only when the handler-level defense becomes
* structurally unreachable (e.g., the op becomes localOnly).
*
* Companion guard at scripts/check-operations-filter-bypass.sh enforces
* the canonical filter site so a future HTTP route can't bypass it.
*/
import { describe, test, expect, beforeAll, afterAll, beforeEach } from 'bun:test';
import { PGLiteEngine } from '../src/core/pglite-engine.ts';
import { resetPgliteState } from './helpers/reset-pglite.ts';
import { operations, type OperationContext } from '../src/core/operations.ts';
import { hasScope } from '../src/core/scope.ts';
let engine: PGLiteEngine;
beforeAll(async () => {
engine = new PGLiteEngine();
await engine.connect({});
await engine.initSchema();
});
afterAll(async () => {
await engine.disconnect();
});
beforeEach(async () => {
await resetPgliteState(engine);
});
// Minimal context factory — every test that invokes a handler builds
// one of these. Defaults to remote=true (untrusted) because that's the
// trust posture the bug-class regressions live in; tests opt back to
// local trust by overriding remote=false.
function makeContext(overrides: Partial<OperationContext> = {}): OperationContext {
return {
engine: engine as any,
config: {} as any,
logger: console as any,
dryRun: false,
remote: true,
sourceId: 'default',
...overrides,
};
}
describe('operations contract — every op has scope + correct mutability shape', () => {
test('every op declares a scope annotation', () => {
for (const op of operations) {
expect(op.scope, `op "${op.name}" missing scope annotation`).toBeDefined();
}
});
test('every mutating op has a write-class scope (not "read")', () => {
const WRITE_CLASS_SCOPES = new Set([
'write',
'admin',
'sources_admin',
'users_admin',
'agent',
]);
for (const op of operations) {
if (op.mutating === true) {
expect(
WRITE_CLASS_SCOPES.has(op.scope ?? 'read'),
`mutating op "${op.name}" has read-tier scope "${op.scope}"; expected one of ${[...WRITE_CLASS_SCOPES].join('/')}`,
).toBe(true);
}
}
});
test('scope is one of the documented enum values', () => {
const KNOWN_SCOPES = new Set([
'read',
'write',
'admin',
'sources_admin',
'users_admin',
'agent',
]);
for (const op of operations) {
expect(
KNOWN_SCOPES.has(op.scope!),
`op "${op.name}" has unknown scope "${op.scope}"`,
).toBe(true);
}
});
});
describe('mcpOperations filter — localOnly ops are excluded from the HTTP-exposed surface', () => {
// This filter is what serve-http.ts uses to build the tools/list response:
// const mcpOperations = operations.filter(op => !op.localOnly);
// A localOnly op that leaks into mcpOperations is exposed via HTTP MCP
// and bypasses the trust boundary. Pin the filter contract here so a
// regression surfaces as a structural test failure.
test('the canonical filter excludes every localOnly op', () => {
const mcpOps = operations.filter(op => !op.localOnly);
const mcpNames = new Set(mcpOps.map(op => op.name));
const localOnlyOps = operations.filter(op => op.localOnly === true);
expect(localOnlyOps.length).toBeGreaterThan(0);
for (const op of localOnlyOps) {
expect(
mcpNames.has(op.name),
`localOnly op "${op.name}" leaked into the HTTP MCP surface`,
).toBe(false);
}
});
test('known historically-sensitive localOnly ops stay filtered', () => {
// Pin every localOnly op by name so a refactor that flips localOnly off
// on any of them fails this test even if the generic contract above
// somehow regresses. Codex /ship review caught the original 4-name
// snapshot was missing purge_deleted_pages, get_recent_transcripts, and
// code_traversal_cache_clear — additions that already qualified.
//
// When adding a NEW localOnly op: add its name here too. The generic
// contract above proves the filter rule applies; this list proves the
// specific ops we care about haven't silently shed their localOnly flag.
const KNOWN_LOCAL_ONLY = [
'sync_brain',
'file_upload',
'file_list',
'file_url',
'purge_deleted_pages',
'get_recent_transcripts',
'code_traversal_cache_clear',
'migrate_embeddings',
];
const lookup = new Map(operations.map(op => [op.name, op] as const));
for (const name of KNOWN_LOCAL_ONLY) {
const op = lookup.get(name);
expect(op, `expected canonical op "${name}" to still exist`).toBeDefined();
expect(op!.localOnly, `"${name}" must stay localOnly`).toBe(true);
}
});
});
describe('hasScope — read-only token cannot satisfy write or admin scopes', () => {
// The HTTP path computes `requiredScope = op.scope || 'read'` and gates
// every call on `hasScope(authInfo.scopes, requiredScope)`. Pin the
// semantics here so a refactor of the IMPLIES table can't silently
// grant admin via a read-class token.
test('read scope does NOT satisfy write', () => {
expect(hasScope(['read'], 'write')).toBe(false);
});
test('read scope does NOT satisfy admin', () => {
expect(hasScope(['read'], 'admin')).toBe(false);
});
test('write scope satisfies write AND read', () => {
expect(hasScope(['write'], 'write')).toBe(true);
expect(hasScope(['write'], 'read')).toBe(true);
});
test('admin scope satisfies admin, write, AND read (umbrella implies)', () => {
expect(hasScope(['admin'], 'admin')).toBe(true);
expect(hasScope(['admin'], 'write')).toBe(true);
expect(hasScope(['admin'], 'read')).toBe(true);
});
test('unknown scope strings are ignored, do not satisfy anything', () => {
expect(hasScope(['bogus'], 'read')).toBe(false);
expect(hasScope(['bogus'], 'write')).toBe(false);
});
test('every read-scope op accepts a read-only token; every write-scope op rejects it', () => {
// Walk the op surface and assert that a synthetic read-only token
// satisfies every read-scope op but no write/admin op.
const READ_TOKEN_SCOPES = ['read'] as const;
for (const op of operations) {
const required = op.scope ?? 'read';
const accepted = hasScope(READ_TOKEN_SCOPES, required);
if (required === 'read') {
expect(accepted, `read op "${op.name}" should accept read-only token`).toBe(true);
} else {
expect(accepted, `${required} op "${op.name}" must reject read-only token`).toBe(false);
}
}
});
});
describe('handler invocation — historically-broken trust-boundary classes', () => {
// The two non-localOnly ops whose handler-level defense fires in
// production and has been broken historically (F7b HTTP MCP shell-job
// RCE; D18 P0 image_path remote-leak). file_upload and sync_brain are
// omitted because they're localOnly (codex CMT-3 — testing their
// handlers with remote=true tests an impossible production path).
test('submit_job rejects shell with ctx.remote=true (HTTP MCP shell-job RCE class)', async () => {
const submitJob = operations.find(op => op.name === 'submit_job');
expect(submitJob).toBeDefined();
const ctx = makeContext({ remote: true });
let threw = false;
let message = '';
try {
await submitJob!.handler(ctx, { name: 'shell', data: { cmd: 'echo hi' } });
} catch (e) {
threw = true;
message = e instanceof Error ? e.message : String(e);
}
expect(threw, 'submit_job(shell) with remote=true MUST reject').toBe(true);
// Should mention the protected status — "permission_denied" is the
// canonical OperationError code, plus the user-facing string names
// the rejected name.
expect(message.toLowerCase()).toContain('shell');
});
test('submit_job allows shell when ctx.remote=false (local CLI is trusted)', async () => {
// The flip side of the trust boundary: a local trusted caller with
// explicit remote=false MUST be allowed to submit shell jobs (that's
// how the CLI works in production). We don't actually want to run the
// job — pass dryRun so the op short-circuits.
const submitJob = operations.find(op => op.name === 'submit_job');
const ctx = makeContext({ remote: false, dryRun: true });
const result = await submitJob!.handler(ctx, { name: 'shell', data: { cmd: 'echo hi' } });
expect(result).toMatchObject({ dry_run: true, action: 'submit_job', name: 'shell' });
});
test('search_by_image rejects image_path with ctx.remote=true (D18 P0)', async () => {
const searchByImage = operations.find(op => op.name === 'search_by_image');
expect(searchByImage).toBeDefined();
const ctx = makeContext({ remote: true });
let threw = false;
let message = '';
try {
await searchByImage!.handler(ctx, { image_path: '/tmp/some-image.png' });
} catch (e) {
threw = true;
message = e instanceof Error ? e.message : String(e);
}
expect(threw, 'search_by_image(image_path) with remote=true MUST reject').toBe(true);
expect(message.toLowerCase()).toContain('image_path');
expect(message.toLowerCase()).toContain('permission_denied');
});
});