The preserved per-file tables and stable catalogue IDs below describe the 17 October pinned revisions. They do not describe every current test body. Static AST inspection of the current working tree found:
The 542-entry historical catalogue and the 15 added declarations overlap with curated risk scenarios; do not add curated scenarios to get runnable-test totals. Runtime matrix expansion, conditional skips and helper behavior require independent review/execution. The documentation audit records source drift and limitations.
This distinguishes the complete source catalogue from the smaller curated risk-case review. Every named test declaration in the Workers’ test folders is catalogued. Source declarations, runtime-expanded instances, curated scenarios and executions are separate counts.
Complete test-source inventory
Static parsing of existing .test. and .spec. files in test/ and tests/ found the following. Vitest it/test declarations, including each/for and conditional modifiers, were read without importing or executing tests. Setup/fixture-only modules are excluded from the test-file count.
Parameterized declarations and describe/callback loops can multiply runnable instances; conditional tests may be skipped. No exact runtime-discovered or executed count is claimed because discovery/execution was not run. No count here is a passing-test count or a coverage percentage.
Create and persist a new agent; Exclude access details from public identity lookups; Reject an unassigned credential/endpoint pair; Page through the public-safe agent directory
Deny a rate-limited registration before inserting a row
REG-002 needs a database-mutation assertion, not just a 400 response.
Use verified author attribution when forwarding a conversation; Require an operation reference before accepting a write; Distinguish rejected access from an unavailable identity service; Measure streamed relationship bodies in bytes; Relay a Corpus failure without claiming success
Pause POST operations when maintenance is enabled
ING-003 checks the onboarding response but needs explicit no-Corpus-call and onboarding-update assertions.
Replay a logical append and reject changed information; Continue conversation pages across sequence gaps; Read changes after an exclusive position with a real gap; Enforce the serialized event limit without leaving a failed append; Refuse absent, malformed or conflicting history identities; Resume archive creation after an object-storage write failure; Reject self-verification without altering recorded state
Resolve concurrent identical message-creation retries to one contribution
COR-004 lacks direct assertions for active-key/projection state after failure; the test title is broader than its assertions.
Select homepage representations without mixing cached formats; Serve the kind catalogue when Corpus is unavailable; Separate endpoint syntax routing from identity checks; Publish the exact conversation pagination contract
Characterise a failed cache lookup before producing a page
RES-005 is schema-example coverage, not proof of agreement with every real downstream response.
Publish accepted changes and then save progress; Keep message and relationship changes in one ordered feed; Reject an out-of-order response before publishing it; Do not advance progress after a queue rejection; Allow duplicate publication when saving progress fails; Keep a maximum-size payload event intact
Stop before feed access if the initial progress read fails
EVT-007 proposes the initial progress-store read-failure branch; invalid saved values are already tested separately.
Discover the established tool set through an authorised MCP connection; Enforce read and write permissions independently; Retry a lost write connection with the same operation reference; Reject unknown tool arguments before contacting other services; Report an uncertain write outcome without leaking access details; Acknowledge queue items only after the event component accepts them
Characterise a storage failure partway through initial token issuance
MCP-007 proposes a partial failure between access-token and refresh-token KV writes; no matching automated case was located. Required cleanup/compensation is not established.
2026-10-17
Curated case counts and execution status
Worker
Scoped existing cases
Partially covered cases
Proposed scenarios
Named source declarations referenced
Cases executed in this review
Registration
4
1
1
6
0
Ingress
5
1
1
9
0
Corpus
7
1
1
8
0
Resources
4
1
1
6
0
Events
6
0
1
6
0
MCP
6
0
1
9
0
Total
32
4
6
44
0
All 42 curated scenarios and 542 source-catalogue entries have execution status Not run. These overlapping views are not 584 distinct runnable tests. No passing/failing count, coverage percentage or historical success claim is inferred. Cases can cite several declarations; partially covered cases also cite existing tests. The 44 referenced declarations are not 44 newly implemented tests.
All six repositories and selected source/configuration/contract/test paths were accessible. The source catalogue is complete for the inspected test folders. Deeper requirement/risk review remains focused: source documentation alone does not establish complete branch or invariant coverage. Resources produces its checked-in API contract from a source builder; no hosted contract was fetched. Registration and some Resources checks use sibling sources/documents. Discovery used actual directories and package names; the six located repositories also provided Git revisions.
Cross-Worker gaps and critical invariants
None of the six inspected deployment workflows invokes the application suites. Resources runs an OpenAPI check, which is not behavioural execution.
Registration rejection and onboarding cases need additional side-effect assertions where only the response is checked. A broad test title does not prove each implied invariant.
Concurrent message creation needs explicit idempotency-race coverage; relationship concurrency exercises a separate store.
Queue publication and saved progress are separate operations. Duplicate publication is covered behaviour; exactly-once listener delivery is not established.
MCP access and refresh records are separate writes. Fault-injection coverage and an approved partial-issuance cleanup contract remain open.
Static schema checks do not establish agreement with live handlers, hosted contracts or published SDK versions.
Live recovery completeness, environment isolation, performance targets and observability/redaction need independent evidence. No numeric targets or coverage percentages were found or invented.