This private E28/E29 draft exposes two existing Module families. Discovery collects commissioned public Sources and re-verifies Citations. Service access builds names-only maps and reports readiness. It implements no scheduling, Research workflow, Knowledge, authentication, provider operation or publication. Independent Astra review passed, and it was accepted locally on 1 October 2026 (owner-designated acceptance reviewer, after independent Opus 5.5 and Fable 5.1 reviews).
Use the Journal adoption guide to obtain and install the private tarball. The demonstrated environment is Node.js 26.8.1 on macOS 27 arm64; other platforms are unverified. Runtime dependencies are built-in Node modules only. Discovery's declaration entry explicitly requires caller development @types/node 26.6.3, tested with undici-types 8.9.0. The three Service declaration entries require no Node type package. Imports start no process, request, authentication or storage write.
Discovery with owned storage
software-factory/discovery-sources exports DiscoverySources, CaptureCache, the brief/collection/Citation parsers, limits, transport interface and errors. CaptureCache(root) requires an absolute normalised caller path, other than /, without a trailing slash. The caller must create that real directory before collection. There is no default machine cache in this entry.
import { CommandJournal } from 'software-factory/command-journal';
import {
CaptureCache, DiscoverySources, parseResearchBrief, type Transport,
} from 'software-factory/discovery-sources';
// Supplied by this application. An offline transport can implement the same interface.
declare const transport: Transport;
declare const cacheRoot: string;
const journal = CommandJournal.open('caller-discovery.sqlite');
const sources = new DiscoverySources({
journal, cache: new CaptureCache(cacheRoot), transport,
now: () => new Date('2026-09-29T10:00:00.000Z'),
});
const brief = parseResearchBrief({
kind: 'sf-research-brief/1', productId: 'caller', revision: 1,
allowedDomains: ['corpus.example'],
sources: [{ id: 'page', url: 'https://corpus.example/page', rhythm: 'basic', extractor: 'page' }],
});
const result = await sources.collect('discovery/caller/basic/2026-09-29', brief);
const capture = result.collection.captures[0];
if (capture) {
const quote = sources.cite(capture, 'a verbatim quote from the recorded text');
if (quote.ok) sources.verify(quote.citation);
}
journal.close();
The transport is required; the clock and per-exchange timeout are optional. The public httpsTransport() remains explicitly opt-in and follows the existing allow-list, public-address, fixed-header, no-cookie, no-redirect and robots rules. A caller transport must honour the supplied request deadline and close contract. The package proof uses an in-memory offline transport, not live collection.
An occurrence spends one nonrenewable request/byte/deadline Budget. An undeclared-length body may cross the byte limit by its last received chunk; that chunk is counted honestly, collection stops partial and no later request is admitted. Repeating a recorded occurrence replays; an active claim refuses. After an interrupted claim's deadline, collection records unknown request/byte usage (null) and exhausted Sources without retrying them. Captures and Citations retain their existing provenance checks. Quotes are at most 15 words, verbatim and free of screened personal data. Cached text expires after 35 days: a Citation becomes unverifiable, not failed. Later changed content reports drift; it never rewrites an old Citation. Cache corruption and corrupt journal records are refused without repair. The cache root belongs to the caller, and collection prunes expired entries across its Products.
Factory source imports keep src/discovery-sources.ts. That facade reexports the identical collector, error and transport classes and supplies a compatible CaptureCache subclass with its unchanged Factory cache default. A Factory cache is also a core cache and passes the existing identity check. No existing journal, receipt, claim, Capture or Citation protocol changes.
Service names and readiness
One Service-access family has three public entry points:
| Import | Responsibility |
|---|---|
software-factory/service-discovery | scanTree, readListing, readSignIn, discoverServices and typed names-only inputs/results |
software-factory/service-kinds | CATALOGUE, defineKind, defineCatalogue, bounds, grammars and shared errors |
software-factory/service-access | ServiceAccess, binding/discovery/reference/route commands, map and readiness |
The application supplies a TreeSource containing a declared commit identity and bounded list/read methods. Denied paths are never opened. Allowed file content is reduced to names, locations and fixed findings; values are not retained. Listing and sign-in parsers consume already captured caller data and make no authentication or provider call. They refuse unexpected fields and preserve unavailable, unauthenticated, corrupt and partial coverage. A name-shaped secret cannot be distinguished from a legitimate name; this existing screen is not a general secret detector.
Create new ServiceAccess(journal, { catalogue }), bind the caller's repository identity through bindLocation, then pass a scan/listing DiscoveryInput to recordDiscovery using the IDs and versions from next(productId). The repository is a supplied identity; no Git checkout or Portfolio registration is required. The optional catalogue uses the existing defineKind/defineCatalogue validation and bounds. Changing relevant catalogue pins makes readiness unknown until new recorded facts support them; replay retains its original meaning. Adding a kind changes the detection digest: a map recorded before cloudflare/1 reads unchanged under the catalogue that recorded it and as catalogue-drift under the current one, and its recorded discovery still replays. A revised kind (posthog/2 and sentry/2 supersede posthog/1 and sentry/1) keeps its first revision in SUPERSEDED_KIND_SPECS, so a caller can rebuild the catalogue that recorded a map; under the current one a need naming the superseded id is refused KIND, and its successor reads as catalogue-drift until rediscovery.
Readiness is ready, missing, unknown or refused. An owner-executed ready route has actionable: false; a declared brokered recipe and vault reference still cannot make a brokered route ready: without a recorded listing of the bound vault the item's presence is unknown (vault-not-listed), and with one its field stays field-unverified. Vault listings (increment 4, recorded through recordVaultListing) are names only, not yet accepted, and newer than this draft's package evidence; the op reader that captures one is Factory source, tested against a fake op only, and not packaged. Readiness consumers are a fixed, declared Factory vocabulary (CONSUMERS, exported by software-factory/service-kinds): a caller's catalogue may name only those consumers and cannot add new ones. Producer and delivery consumers remain refused. No result authenticates a credential, grants authority, starts CI or reads a secret. Unknown coverage never becomes exact zero. Count and CoverageStatus are shared data types; Portfolio keeps compatible type reexports, and the Service package closure does not include Portfolio.
Evidence and limits
The tarball proof exercises useful offline collection, Citation expiry/drift, robots and hostile-input refusal, fixed Budgets, interrupted-claim recovery, names-only Service discovery, custom catalogue readiness, restart/replay, corruption and Factory compatibility. A captured pre-change fixture retains actual original aggregates, receipts, events and outbox rows; packed readers replay them unchanged. Separate clean TypeScript consumers prove Service declarations work without Node types and Discovery requires them explicitly.
The repository's docs/discovery-services-adoption-draft.json records the exact artefact and remaining gates. It is deliberately excluded from the package to avoid an artefact self-hash cycle. This evidence supports local draft review only; neither adoption epic nor a Workflow is accepted by it.
