@canonical/pragma-cli 0.33.0 → 0.35.0

package/src/capabilities/resources/provider.ts

@@ -3,143 +3,362 @@
  *
  * Listing and autocomplete are STORELESS: they read the enriched `index.json`
  * (readPackIndex) off disk, so the resource surface costs nothing until an agent
- * reads one. A read is STORE-BACKED through the SHARED {@link readEntity} — the
- * same reader `graph inspect` uses — so a resource read and a `graph inspect` of
- * the same URI return identical content (the mirror contract).
+ * reads one. They read it for the pack the BOOT DECISION names, so the listing
+ * can never advertise entities whose reads then fail STORE_UNAVAILABLE. A read
+ * is STORE-BACKED through the SHARED {@link readEntity} — the same reader
+ * `graph inspect` uses — so a resource read and a `graph inspect` of the same
+ * URI return identical content (the mirror contract).
+ *
+ * The listing is CURATED, not paged. MCP has a `cursor`/`nextCursor` contract
+ * for `resources/list`, but the SDK's high-level `McpServer` list handler
+ * ignores `request.params.cursor` and never returns a `nextCursor` (verified on
+ * the wire against 1.27.1), so "list everything, the client will page it" means
+ * "send everything, every connect". Listing all 712 indexed entities cost an
+ * agent ~155 KB of context before it had asked a single question. So a module
+ * DECLARES the slices of the index it contributes ({@link McpListable}) and the
+ * listing is their union: the collections, plus the schema that describes them.
+ * The individuals stay fully reachable — by the `{+uri}` template, whose
+ * autocomplete still offers every one of them.
  *
  * Degradation (Risk4): on a missing or legacy (pre-v2) index the listing returns
  * a single "run `pragma sources update`" hint — never a live re-index of the
- * store on the resource path.
+ * store on the resource path. That hint is itself READABLE (it is the only
+ * resource a cold server advertises, and an advertised resource that refuses to
+ * be read is worse than no resource at all).
  *
  * The MCP SDK's `ResourceTemplate` is dynamic-imported inside `register` so this
  * module — reachable on the capabilities import graph — never pulls the SDK onto
  * the `--help`/`__complete` fast path.
  */
 
-import { readPackIndex } from "../../kernel/completion/entitySource.js";
+import { BIN_NAME } from "../../constants.js";
+import {
+  matchesType,
+  readPackIndex,
+} from "../../kernel/completion/entitySource.js";
 import { asPragmaError } from "../../kernel/error/fromTaskError.js";
-import { PragmaError } from "../../kernel/error/PragmaError.js";
+import { PragmaError } from "../../kernel/error/index.js";
 import { mcpErrorFrom } from "../../kernel/project/mcp/mcpError.js";
-import type { PackIndex } from "../../kernel/runtime/graphpack/types.js";
+import { toTurtle } from "../../kernel/render/turtle.js";
+import type {
+  PackIndex,
+  PackIndexEntity,
+} from "../../kernel/runtime/graphpack/types.js";
 import { readEntity } from "../../kernel/runtime/readEntity.js";
-import type { McpResourceProvider } from "../../kernel/spec/types.js";
+import { resolveSources } from "../../kernel/runtime/resolveSources.js";
+import type {
+  CapabilityModule,
+  McpListable,
+  McpListableRef,
+  McpResourceProvider,
+} from "../../kernel/spec/types.js";
+
+/**
+ * The weight an unweighted slice carries — "as important as any other".
+ *
+ * The ONLY editorial number left in this module. The judgements that used to
+ * live here as `CLASS_PRIORITY = 0.9` / `INDIVIDUAL_PRIORITY = 0.3` are now
+ * declared data (`McpListableRef.weight`, derived from `PackLookup.weights`):
+ * which of a distribution's classes deserves an agent's attention first is a
+ * fact about that distribution's graph, not about the kernel that serves it.
+ */
+const DEFAULT_WEIGHT = 1;
 
-/** Max autocomplete suggestions returned for a partial URI. */
-const COMPLETION_LIMIT = 50;
+/** The reserved listing entry that explains an unusable store. */
+const SOURCES_URI = "pragma:sources";
+/** The `{+uri}` variable value that addresses {@link SOURCES_URI}. */
+const SOURCES_VARIABLE = "sources";
 
-/** MCP annotation priority for a schema (TBox) entry — surfaced above individuals. */
-const CLASS_PRIORITY = 0.9;
-/** MCP annotation priority for an individual (ABox) entry. */
-const INDIVIDUAL_PRIORITY = 0.3;
+/**
+ * The media type an entity read RETURNS — and therefore the one the listing
+ * ADVERTISES for every entry it mints.
+ *
+ * One constant, three readers (the template registration, the read body, and
+ * each listed entry). The listing promising `application/json` while the read
+ * answers Turtle is the same shape of failure as the `recipes:{+uri}` template
+ * over `pragma:` URIs: a surface describing content it does not serve.
+ */
+const ENTITY_MIME_TYPE = "text/turtle";
+/** The recovery hint is prose about a broken store, not graph content. */
+const RECOVERY_MIME_TYPE = "text/plain";
 
-/** The `pragma:` URI scheme + the single reserved-expansion template variable. */
+/**
+ * The `pragma:` URI scheme + the single reserved-expansion template variable.
+ *
+ * A LITERAL, and it must stay one until the wire identifiers move together.
+ * `surface/surface.v2.json` freezes `pragma:{+uri}` as protocol identity, and
+ * `buildResourceList` mints `pragma:<prefixed>` for every entry it lists —
+ * three writings of one decision. Deriving only this one made a fork advertise
+ * `recipes:{+uri}` over a list of 653 `pragma:` URIs: every resource the server
+ * offered became unreadable, and the readable form was never advertised.
+ * `resources.test.ts` pins the pair through {@link resourceProvider}'s declared
+ * surface; `identity.test.ts` masks this template out of its leak scan for the
+ * same reason.
+ */
 const URI_TEMPLATE = "pragma:{+uri}";
 
 const TEMPLATE_DESCRIPTION =
-  "Knowledge-graph entities from the local pack. Read any entry by its prefixed " +
-  "URI (e.g. pragma:ds:button); autocomplete matches URI or label substrings. " +
-  "Content mirrors `pragma graph inspect <uri>`.";
+  "Knowledge-graph entities from the local pack, as Turtle. Read any entry by " +
+  "its prefixed URI (e.g. pragma:ds:global.component.button); autocomplete " +
+  "matches URI or label substrings. A read returns the entity's neighbourhood — " +
+  "its own triples, the edges pointing AT it, and its blank nodes inlined. " +
+  `Content mirrors \`${BIN_NAME} graph inspect <uri> --format llm\`.`;
 
 /** Agent-navigability annotations mirrored onto a listed resource. */
 export interface ResourceAnnotations {
-  /** MCP audience roles (the SDK's literal union) — always the assistant here. */
+  /** MCP audience roles (the SDK's literal union). */
   readonly audience: ("user" | "assistant")[];
+  /** The declared weight of the slice that contributed this entry (0–1). */
   readonly priority: number;
 }
 
-/** The audience every listed resource carries (assistant-facing). */
-const ASSISTANT_AUDIENCE: ("user" | "assistant")[] = ["assistant"];
+/**
+ * The audience every listed resource carries.
+ *
+ * BOTH roles. `["assistant"]` tells a client that honours the field to hide the
+ * resource from its human picker, and browsing a design system in a picker is
+ * exactly a human use case — the same entries serve an agent orienting itself
+ * and a person looking for the button spec.
+ */
+const LISTING_AUDIENCE: ("user" | "assistant")[] = ["user", "assistant"];
 
 /** One resource entry in the listing. */
 export interface ListedResource {
   readonly uri: string;
+  /** The STABLE prefixed URI (`ds:Component`) — an identifier, not a label. */
   readonly name: string;
+  /** The human label (MCP `title`), when the index carries one. */
+  readonly title?: string;
   readonly description?: string;
-  readonly mimeType: "application/json";
-  /** Audience + priority (schema surfaced above individuals). */
+  readonly mimeType: typeof ENTITY_MIME_TYPE | typeof RECOVERY_MIME_TYPE;
+  /** Audience + the declared weight of the slice that listed it. */
   readonly annotations?: ResourceAnnotations;
-  /** Taxonomy metadata: `pragma/box` (tbox|abox) + optional instance count. */
+  /** Taxonomy metadata: `pragma/box`, `pragma/type`, optional instance count. */
   readonly _meta?: Record<string, unknown>;
 }
 
+/**
+ * The human name for an entity, best-available-first.
+ *
+ * `altNames` is the distribution's declared alternative-name property — the
+ * name a bespoke lookup MATCHES on ("Button", "FileTree"). 383 of the embedded
+ * pack's 714 entities carry one and no `rdfs:label`, so before this the listing
+ * showed them as a bare URI with no description while the readable name sat
+ * unused in the index one field away.
+ */
+function humanName(entity: PackIndexEntity): string | undefined {
+  return entity.label ?? entity.altNames?.[0] ?? undefined;
+}
+
+/** Render one index entity as a listed resource at a given weight. */
+function listEntity(
+  entity: PackIndexEntity,
+  weight: number,
+  instanceCount?: number,
+): ListedResource {
+  const title = humanName(entity);
+  // `pragma/box` + `pragma/instanceCount` are FROZEN protocol identity (see
+  // `surface.v2.json` and docs/mcp-integration.md). `pragma/type` is an
+  // ADDITION: all 714 entities carry a type, and publishing it lets an agent
+  // narrow the listing to a family without spending a read to find out what
+  // each entry is.
+  const meta: Record<string, unknown> = {
+    "pragma/box": entity.box ?? "abox",
+    "pragma/type": entity.type,
+  };
+  if (typeof instanceCount === "number") {
+    meta["pragma/instanceCount"] = instanceCount;
+  }
+  return {
+    uri: `pragma:${entity.prefixed ?? entity.name}`,
+    name: entity.prefixed ?? entity.name,
+    ...(title ? { title } : {}),
+    ...(entity.description ? { description: entity.description } : {}),
+    mimeType: ENTITY_MIME_TYPE,
+    annotations: { audience: LISTING_AUDIENCE, priority: weight },
+    _meta: meta,
+  };
+}
+
+/** The recovery entry a missing/legacy index degrades to (and its read body). */
+const SOURCES_RECOVERY = `No enriched entity index. Run \`${BIN_NAME} sources update\` to build it.`;
+
+/** The single entry a server with no usable index advertises. */
+function sourcesEntry(): ListedResource {
+  return {
+    uri: SOURCES_URI,
+    name: SOURCES_VARIABLE,
+    title: "Store not indexed",
+    description: SOURCES_RECOVERY,
+    mimeType: RECOVERY_MIME_TYPE,
+  };
+}
+
+/**
+ * Resolve one declared slice into listed resources.
+ *
+ * `as: "collection"` yields at most ONE entry — the class entry for the slice's
+ * type, carrying `pragma/instanceCount`. A type the index does not know
+ * contributes NOTHING rather than an empty collection: `ds:Token` is declared by
+ * a read story but SPARQL-served, never indexed, and advertising an empty
+ * collection for it would promise a browse that has nothing to show.
+ */
+function listSlice(
+  index: PackIndex,
+  byPrefixed: ReadonlyMap<string, PackIndexEntity>,
+  ref: McpListableRef,
+): ListedResource[] {
+  const weight = ref.weight ?? DEFAULT_WEIGHT;
+  if (ref.as === "collection") {
+    const type = ref.type;
+    if (!type) return [];
+    const entity = byPrefixed.get(type);
+    if (!entity) return [];
+    // A class with no instances has NO key in `instanceCountByType`, so reading
+    // it straight through reports `undefined` — and an empty collection then
+    // looks identical to an entry that is not a collection at all. A declared
+    // collection whose class the index knows always states a count, and for an
+    // empty one that count is zero.
+    const count = entity.uri ? (index.instanceCountByType[entity.uri] ?? 0) : 0;
+    return [listEntity(entity, weight, count)];
+  }
+  const listed: ListedResource[] = [];
+  for (const entity of index.entities) {
+    if (ref.box && (entity.box ?? "abox") !== ref.box) continue;
+    if (!matchesType(entity, ref.type ?? "")) continue;
+    const count =
+      entity.box === "tbox" && entity.uri
+        ? index.instanceCountByType[entity.uri]
+        : undefined;
+    listed.push(listEntity(entity, weight, count));
+  }
+  return listed;
+}
+
 /**
  * Build the resource listing from the storeless pack index.
  *
  * @param index - The active pack index, or `undefined` when none is reachable.
- * @returns One resource per indexed entity, schema (TBox) first; a single
+ * @param listing - The union of the modules' declared slices.
+ * @returns One resource per contributed entry, heaviest slice first; a single
  *   recovery entry when the index is missing or pre-v2 (no enrichment).
  */
 export function buildResourceList(
   index: PackIndex | undefined,
+  listing: McpListable,
 ): ListedResource[] {
   // readPackIndex returns raw JSON.parse output (no zod), so guard structurally:
   // a malformed index missing `version` must degrade to the recovery hint, not
   // fall through to `[...index.entities]` (a TypeError inside the MCP handler).
   if (!index || index.version !== 2 || !Array.isArray(index.entities)) {
-    return [
-      {
-        uri: "pragma:sources",
-        name: "Store not indexed",
-        description:
-          "No enriched entity index. Run `pragma sources update` to build it.",
-        mimeType: "application/json",
-      },
-    ];
+    return [sourcesEntry()];
+  }
+  const byPrefixed = new Map<string, PackIndexEntity>();
+  for (const entity of index.entities) {
+    const key = entity.prefixed ?? entity.name;
+    // A punned subject is indexed as two facets sharing one key; the schema
+    // facet is the one a collection entry means.
+    if (!byPrefixed.has(key) || entity.box === "tbox")
+      byPrefixed.set(key, entity);
   }
-  const ordered = [...index.entities].sort((a, b) => {
-    const boxRank = (e: typeof a) => (e.box === "tbox" ? 0 : 1);
-    return boxRank(a) - boxRank(b) || a.name.localeCompare(b.name);
+  const candidates = listing.sources.flatMap((ref) =>
+    listSlice(index, byPrefixed, ref),
+  );
+  candidates.sort((a, b) => {
+    const priority = (r: ListedResource) => r.annotations?.priority ?? 0;
+    const boxRank = (r: ListedResource) =>
+      r._meta?.["pragma/box"] === "tbox" ? 0 : 1;
+    return (
+      priority(b) - priority(a) ||
+      boxRank(a) - boxRank(b) ||
+      a.uri.localeCompare(b.uri)
+    );
   });
-  // Dedup by resource URI: an OWL-punned subject (a class/property IRI ALSO
-  // asserted as a domain individual) is indexed as TWO entities — a tbox and an
-  // abox facet — that compact to the SAME `pragma:<uri>`. Mapping both would
-  // list one URI twice (A8). The sort above puts the tbox facet first, so
-  // keeping the first occurrence per URI surfaces the schema facet (its higher
-  // priority + instance count) and drops the punned duplicate.
+  // Dedup by resource URI. Two slices may contribute the same entry (a story's
+  // collection IS one of the ontology's schema entries), and an OWL-punned
+  // subject (a class/property IRI ALSO asserted as a domain individual) is
+  // indexed as TWO entities — a tbox and an abox facet — that compact to the
+  // SAME `pragma:<uri>` (A8). The sort above puts the heavier slice, then the
+  // tbox facet, first, so keeping the first occurrence per URI surfaces the
+  // richer entry (its instance count and higher priority) and drops the rest.
   const seen = new Set<string>();
   const resources: ListedResource[] = [];
-  for (const entity of ordered) {
-    const uri = `pragma:${entity.prefixed ?? entity.name}`;
-    if (seen.has(uri)) continue;
-    seen.add(uri);
-    const isTbox = entity.box === "tbox";
-    // Port of the old shell's `_meta` taxonomy: `pragma/box` + a priority that
-    // ranks schema (classes/properties) above individuals, so an agent browsing
-    // the resource list sees the schema first. `pragma/instanceCount` is the
-    // per-type count carried in the index for a TBox class entry.
-    // `instanceCountByType` is keyed by the full type IRI; a TBox class entry's
-    // own IRI is that key, so it carries the count of its individuals.
-    const instanceCount =
-      isTbox && entity.uri ? index.instanceCountByType[entity.uri] : undefined;
-    const meta: Record<string, unknown> = {
-      "pragma/box": entity.box ?? "abox",
-    };
-    if (typeof instanceCount === "number") {
-      meta["pragma/instanceCount"] = instanceCount;
-    }
-    resources.push({
-      uri,
-      name: entity.label || entity.name,
-      ...(entity.description ? { description: entity.description } : {}),
-      mimeType: "application/json" as const,
-      annotations: {
-        audience: ASSISTANT_AUDIENCE,
-        priority: isTbox ? CLASS_PRIORITY : INDIVIDUAL_PRIORITY,
-      },
-      _meta: meta,
-    });
+  for (const candidate of candidates) {
+    if (seen.has(candidate.uri)) continue;
+    seen.add(candidate.uri);
+    resources.push(candidate);
   }
   return resources;
 }
 
 /**
- * Rank URI autocomplete candidates against a partial query (case-insensitive over
- * both the prefixed URI and the label; exact > prefix > substring), capped.
+ * The per-type ranking weights a declared listing implies.
+ *
+ * A slice that names a `type` is an opinion about that type's importance, and
+ * the SAME opinion should decide both what the listing surfaces first and what
+ * autocomplete offers first — one declaration, two readers. A slice that names
+ * only a `box` carries no opinion about a candidate URI, so it contributes none.
+ *
+ * @param listing - The union of the modules' declared slices.
+ * @returns Prefixed type → weight (heaviest declaration wins on a repeat).
  */
+export function completionWeights(
+  listing: McpListable,
+): Record<string, number> {
+  const weights: Record<string, number> = {};
+  for (const ref of listing.sources) {
+    if (!ref.type) continue;
+    const weight = ref.weight ?? DEFAULT_WEIGHT;
+    weights[ref.type] = Math.max(weights[ref.type] ?? 0, weight);
+  }
+  return weights;
+}
+
+/**
+ * Rank URI autocomplete candidates against a partial query (case-insensitive
+ * over both the prefixed URI and the label; exact > prefix > substring).
+ *
+ * UNCAPPED on purpose. The SDK derives the wire's `total`/`hasMore` from the
+ * array this returns — it slices to 100, reports `total: values.length` and
+ * `hasMore: values.length > 100` — so pre-truncating here made a truncated
+ * result describe itself as complete: `ds:` answered
+ * `{ values: 50, total: 50, hasMore: false }` while 499 entities matched.
+ * Completion keeps offering EVERY entity: it is the discovery path for the
+ * individuals the curated listing no longer enumerates.
+ *
+ * @param entities - The index entities to rank.
+ * @param query - The partial the agent typed.
+ * @param weights - Prefixed type → weight, breaking ties between equal matches.
+ * @returns Every matching prefixed URI, best first.
+ */
+/**
+ * The weight of an entity across EVERY type it is declared under.
+ *
+ * `PackIndexEntity.type` is one lexically chosen primary among `types`, and the
+ * listing's own `matchesType` already honours both. Weighing only the primary
+ * therefore ignored a weight declared for a secondary type, and let ranking
+ * shift when an unrelated, lexically earlier type was added to an entity.
+ *
+ * The LOWEST declared weight wins: a weight below the default is a demotion, and
+ * a demotion that any membership asks for should not be cancelled by another
+ * membership that simply never asked for anything. Types with no declared weight
+ * do not vote, so an entity nobody weighted keeps the default.
+ */
+function effectiveWeight(
+  entity: PackIndexEntity,
+  weights: Readonly<Record<string, number>>,
+): number {
+  const memberships = entity.types ?? [entity.type];
+  const declared = memberships
+    .map((type) => weights[type])
+    .filter((weight): weight is number => weight !== undefined);
+  return declared.length > 0 ? Math.min(...declared) : DEFAULT_WEIGHT;
+}
+
 export function rankUriCompletions(
-  entities: readonly PackIndex["entities"][number][],
+  entities: readonly PackIndexEntity[],
   query: string,
-  limit: number,
+  weights: Readonly<Record<string, number>> = {},
 ): string[] {
   const needle = query.toLowerCase();
   const score = (field: string): number =>
@@ -150,18 +369,28 @@ export function rankUriCompletions(
         : field.includes(needle)
           ? 1
           : 0;
-  const scored: { prefixed: string; score: number }[] = [];
+  const scored: { prefixed: string; score: number; weight: number }[] = [];
   for (const entity of entities) {
     const prefixed = entity.prefixed ?? entity.name;
     const uriScore = score(prefixed.toLowerCase());
     const labelScore = entity.label ? score(entity.label.toLowerCase()) : 0;
     const best = Math.max(uriScore, labelScore);
-    if (best > 0) scored.push({ prefixed, score: best });
+    if (best > 0) {
+      scored.push({
+        prefixed,
+        score: best,
+        weight: effectiveWeight(entity, weights),
+      });
+    }
   }
-  scored.sort((a, b) =>
-    b.score !== a.score
-      ? b.score - a.score
-      : a.prefixed.localeCompare(b.prefixed),
+  // Match quality first, then the DECLARED weight, then the alphabet. The
+  // weight step is what stops 86 `…-close_button` subcomponents from burying
+  // `ds:global.component.button` on the strength of their spelling alone.
+  scored.sort(
+    (a, b) =>
+      b.score - a.score ||
+      b.weight - a.weight ||
+      a.prefixed.localeCompare(b.prefixed),
   );
   const seen = new Set<string>();
   const ranked: string[] = [];
@@ -169,11 +398,24 @@ export function rankUriCompletions(
     if (seen.has(prefixed)) continue;
     seen.add(prefixed);
     ranked.push(prefixed);
-    if (ranked.length >= limit) break;
   }
   return ranked;
 }
 
+/**
+ * The union of every module's declared listing.
+ *
+ * @param modules - The effective capability modules.
+ * @returns One listing holding every declared slice, in module order.
+ */
+export function collectListable(
+  modules: readonly CapabilityModule[],
+): McpListable {
+  return {
+    sources: modules.flatMap((module) => module.mcpListable?.sources ?? []),
+  };
+}
+
 /** Read the single `uri` template variable, tolerating array expansion. */
 function readUriVariable(value: string | string[] | undefined): string {
   const uri = Array.isArray(value) ? value.at(0) : value;
@@ -190,23 +432,35 @@ export const resourceProvider: McpResourceProvider = {
   // The single authoring point for the template id — the surface emitter reads
   // this so `mcpSurface.resources` cannot drift from what `register` installs.
   surface: { templates: [URI_TEMPLATE] },
-  async register(server, runtime) {
+  async register(server, runtime, modules) {
     const { ResourceTemplate } = await import(
       "@modelcontextprotocol/sdk/server/mcp.js"
     );
     const { McpError, ErrorCode } = await import(
       "@modelcontextprotocol/sdk/types.js"
     );
+    // The index of the pack the boot decision names — `undefined` when the
+    // store is unavailable, which `buildResourceList` renders as the recovery
+    // entry rather than a catalogue of unreadable URIs.
+    const activeIndex = async (): Promise<PackIndex | undefined> =>
+      readPackIndex(resolveSources(await runtime.loadConfig(), runtime.cwd));
+    // What the SERVER lists is the union of what its modules declare — not what
+    // this one module knows. Collected once, at registration.
+    const listing = collectListable(modules);
+    const weights = completionWeights(listing);
+
     const template = new ResourceTemplate(URI_TEMPLATE, {
       list: async () => ({
-        resources: buildResourceList(readPackIndex(runtime.cwd)),
+        resources: buildResourceList(await activeIndex(), listing),
       }),
       complete: {
+        // Uncapped: the SDK slices to 100 and derives `total`/`hasMore` from
+        // the full array, so returning it is what makes those numbers honest.
         uri: async (value: string) =>
           rankUriCompletions(
-            readPackIndex(runtime.cwd)?.entities ?? [],
+            (await activeIndex())?.entities ?? [],
             value,
-            COMPLETION_LIMIT,
+            weights,
           ),
       },
     });
@@ -214,20 +468,42 @@ export const resourceProvider: McpResourceProvider = {
     server.registerResource(
       "graph-entity",
       template,
-      { description: TEMPLATE_DESCRIPTION, mimeType: "application/json" },
+      // `text/turtle` because that is what the body IS. The surface used to
+      // claim `application/json` for a graph, and paid ~95 bytes of wrapper per
+      // named node to do it — one button read carried ~6.1 KB of pure structure.
+      // Turtle spends the prefixes once, writes each term as itself, and says
+      // IRI-versus-literal in syntax rather than in a `termType` field.
+      { description: TEMPLATE_DESCRIPTION, mimeType: ENTITY_MIME_TYPE },
       async (
         url: URL,
         variables: Record<string, string | string[] | undefined>,
       ) => {
         try {
           const uri = readUriVariable(variables.uri);
+          // The recovery entry is a real resource, not a label. It is the ONLY
+          // thing a server with no usable index advertises, and reading it used
+          // to fail `INVALID_INPUT — Invalid uri "sources"`: the one resource on
+          // offer refused to be read, so the hint that explains the cold store
+          // was unreachable from the surface that showed it.
+          if (uri === SOURCES_VARIABLE) {
+            return {
+              contents: [
+                {
+                  uri: url.href,
+                  mimeType: RECOVERY_MIME_TYPE,
+                  text: SOURCES_RECOVERY,
+                },
+              ],
+            };
+          }
           const entity = await readEntity(runtime, uri);
+          const { prefixes } = await runtime.store.get();
           return {
             contents: [
               {
                 uri: url.href,
-                mimeType: "application/json",
-                text: JSON.stringify(entity, null, 2),
+                mimeType: ENTITY_MIME_TYPE,
+                text: toTurtle(entity, prefixes),
               },
             ],
           };