akm-cli 0.9.0-beta.5 → 0.9.0-beta.50

package/dist/llm/memory-infer.js

@@ -18,20 +18,16 @@
  * the connection via `resolveIndexPassLLM("memory", config)` and pass it
  * straight through.
  */
+import memoryInferSystemPrompt from "../assets/prompts/memory-infer-system.md" with { type: "text" };
+import memoryInferUserPrompt from "../assets/prompts/memory-infer-user.md" with { type: "text" };
 import { toErrorMessage } from "../core/common.js";
 import { warn } from "../core/warn.js";
-import { chatCompletion, LlmCallError, parseEmbeddedJsonResponse } from "./client.js";
-import { tryLlmFeature } from "./feature-gate.js";
+import { parseEmbeddedJsonResponse } from "./client.js";
+import { callStructured } from "./structured-call.js";
 /** Hard cap on body chars sent to the model — pragmatic and matches `runLlmEnrich`. */
 const MAX_BODY_CHARS = 4000;
-const SYSTEM_PROMPT = "You compress a developer memory into one high-signal derived memory for later retrieval. " +
-    "Return only valid JSON. No prose outside the JSON object. No markdown fences.";
-const USER_PROMPT_PREFIX = `Compress the memory below into one derived memory. Output ONLY JSON:
-{"title":"short title string","description":"one sentence summary string","tags":["tag1","tag2"],"searchHints":["search phrase 1","search phrase 2"],"content":"2-3 sentence compressed body preserving key facts verbatim"}
-Rules: be specific, no vague generalizations, preserve key facts (names/versions/paths/config keys verbatim), merge related points, 3-8 tags, 3-6 searchHints. The content field must be a plain string with 2-3 sentences.
-
-Memory:
-`;
+const SYSTEM_PROMPT = memoryInferSystemPrompt;
+const USER_PROMPT_PREFIX = memoryInferUserPrompt;
 /**
  * Strict JSON Schema for the derived-memory payload. Sent to providers that
  * opt in via `LlmConnectionConfig.supportsJsonSchema = true`; the client
@@ -63,26 +59,39 @@ const DERIVED_MEMORY_JSON_SCHEMA = {
  * Errors are logged via `warn()` but never thrown — a failed split for one memory
  * must not abort the rest of the index pass.
  *
- * Routes through `tryLlmFeature("memory_inference", ...)` so the feature gate
- * and onFallback hook are honoured uniformly (Fix C5).
+ * Routes through `callStructured({ feature: "memory_inference", ... })` so the
+ * feature gate, error classification, and onFallback hook are honoured uniformly
+ * (Fix C5).
  */
 export async function compressMemoryToDerivedMemory(llmConfig, body, signal, akmConfig, onFallback, telemetry, onRetryAttempt) {
     const trimmedBody = body.trim();
     if (!trimmedBody)
         return undefined;
     const userPrompt = `${USER_PROMPT_PREFIX}${trimmedBody.slice(0, MAX_BODY_CHARS)}`;
-    return tryLlmFeature("memory_inference", akmConfig, async () => {
-        try {
-            const raw = await chatCompletion(llmConfig, [
-                { role: "system", content: SYSTEM_PROMPT },
-                { role: "user", content: userPrompt },
-            ], {
-                temperature: 0.1,
-                timeoutMs: llmConfig.timeoutMs,
-                signal,
-                responseSchema: DERIVED_MEMORY_JSON_SCHEMA,
-                onRetryAttempt,
-            });
+    // Memory-inference is ALWAYS gated: no `akmConfig` ⇒ gate closed (no chat,
+    // `disabled` fallback), never the seam's ungated/propagate path (which is for
+    // direct callers like `enhanceMetadata`). This is the gate-closed branch
+    // `tryLlmFeature(_, undefined, _)` took before the migration.
+    if (!akmConfig) {
+        onFallback?.({ feature: "memory_inference", reason: "disabled" });
+        return undefined;
+    }
+    return callStructured({
+        feature: "memory_inference",
+        akmConfig,
+        config: llmConfig,
+        messages: [
+            { role: "system", content: SYSTEM_PROMPT },
+            { role: "user", content: userPrompt },
+        ],
+        request: {
+            temperature: 0.1,
+            timeoutMs: llmConfig.timeoutMs,
+            signal,
+            responseSchema: DERIVED_MEMORY_JSON_SCHEMA,
+            onRetryAttempt,
+        },
+        parse: (raw) => {
             if (!raw)
                 return undefined;
             const parsed = parseEmbeddedJsonResponse(raw);
@@ -112,9 +121,9 @@ export async function compressMemoryToDerivedMemory(llmConfig, body, signal, akm
                 return undefined;
             }
             return { title, description, tags, searchHints, content };
-        }
-        catch (err) {
-            if (err instanceof LlmCallError && err.code === "provider_html_error") {
+        },
+        onError: (cls, err) => {
+            if (cls === "html") {
                 if (telemetry)
                     telemetry.htmlErrorCount = (telemetry.htmlErrorCount ?? 0) + 1;
                 warn(`memory inference: provider returned HTML instead of JSON; skipping memory: ${toErrorMessage(err)}`);
@@ -122,9 +131,8 @@ export async function compressMemoryToDerivedMemory(llmConfig, body, signal, akm
             }
             warn(`memory inference failed: ${toErrorMessage(err)}`);
             return undefined;
-        }
-    }, undefined, {
-        timeoutMs: llmConfig.timeoutMs,
+        },
+        fallback: undefined,
         onFallback,
     });
 }

package/dist/llm/metadata-enhance.js

@@ -1,9 +1,17 @@
 // This Source Code Form is subject to the terms of the Mozilla Public
 // License, v. 2.0. If a copy of the MPL was not distributed with this
 // file, You can obtain one at https://mozilla.org/MPL/2.0/.
-import { chatCompletion, parseJsonResponse } from "./client.js";
-import { tryLlmFeature } from "./feature-gate.js";
-const SYSTEM_PROMPT = `You are a metadata generator for a developer asset registry. Given a script/skill/command/agent entry, generate improved metadata. Respond with ONLY valid JSON, no markdown fencing.`;
+/**
+ * LLM-driven metadata enhancement for stash entries.
+ *
+ * Split out of `llm.ts` so the higher-level workflow (prompting the LLM to
+ * improve descriptions/tags/searchHints) lives separately from the low-level
+ * transport client in `client.ts`.
+ */
+import metadataEnhanceSystemPrompt from "../assets/prompts/metadata-enhance-system.md" with { type: "text" };
+import { parseJsonResponse } from "./client.js";
+import { callStructured } from "./structured-call.js";
+const SYSTEM_PROMPT = metadataEnhanceSystemPrompt;
 /**
  * Use an LLM to enhance a stash entry's metadata: improve description,
  * generate searchHints, and suggest tags.
@@ -33,34 +41,39 @@ Generate improved metadata for this ${entry.type}. Return JSON with these fields
 - "tags": an array of 3-8 relevant keyword tags
 
 Return ONLY the JSON object, no explanation.`;
-    const runLlm = async () => {
-        const raw = await chatCompletion(config, [
+    // `parse` owns the raw response: the `!raw`/unparseable case ⇒ `{}`, plus the
+    // description/searchHints/tags shaping. `enhanceMetadata` never warns and
+    // never bumps telemetry, so `onError` (gated path only) just swallows to `{}`
+    // — identical to the surrounding control flow's pre-migration behaviour. The
+    // ungated path (akmConfig === undefined) propagates errors via callStructured.
+    return callStructured({
+        feature: "metadata_enhance",
+        akmConfig,
+        config,
+        messages: [
             { role: "system", content: SYSTEM_PROMPT },
             { role: "user", content: userPrompt },
-        ], { signal });
-        const parsed = parseJsonResponse(raw);
-        if (!parsed)
-            return {};
-        const result = {};
-        if (typeof parsed.description === "string" && parsed.description) {
-            result.description = parsed.description;
-        }
-        if (Array.isArray(parsed.searchHints)) {
-            result.searchHints = parsed.searchHints
-                .filter((s) => typeof s === "string" && s.trim().length > 0)
-                .slice(0, 8);
-        }
-        if (Array.isArray(parsed.tags)) {
-            result.tags = parsed.tags.filter((s) => typeof s === "string" && s.trim().length > 0).slice(0, 10);
-        }
-        return result;
-    };
-    // When no akmConfig is provided, bypass the feature gate entirely: run the
-    // LLM call directly and let errors propagate to the caller (pre-gate
-    // behaviour). When akmConfig is present, honour the feature flag and swallow
-    // errors to {} via tryLlmFeature.
-    if (akmConfig === undefined) {
-        return runLlm();
-    }
-    return tryLlmFeature("metadata_enhance", akmConfig, runLlm, {}, { timeoutMs: config.timeoutMs });
+        ],
+        request: { signal, timeoutMs: config.timeoutMs },
+        parse: (raw) => {
+            const parsed = raw ? parseJsonResponse(raw) : undefined;
+            if (!parsed)
+                return {};
+            const result = {};
+            if (typeof parsed.description === "string" && parsed.description) {
+                result.description = parsed.description;
+            }
+            if (Array.isArray(parsed.searchHints)) {
+                result.searchHints = parsed.searchHints
+                    .filter((s) => typeof s === "string" && s.trim().length > 0)
+                    .slice(0, 8);
+            }
+            if (Array.isArray(parsed.tags)) {
+                result.tags = parsed.tags.filter((s) => typeof s === "string" && s.trim().length > 0).slice(0, 10);
+            }
+            return result;
+        },
+        onError: () => ({}),
+        fallback: {},
+    });
 }

package/dist/llm/structured-call.js

@@ -0,0 +1,49 @@
+// This Source Code Form is subject to the terms of the Mozilla Public
+// License, v. 2.0. If a copy of the MPL was not distributed with this
+// file, You can obtain one at https://mozilla.org/MPL/2.0/.
+import { chatCompletion, isContextSizeError, LlmCallError } from "./client.js";
+import { tryLlmFeature } from "./feature-gate.js";
+/**
+ * Classify a thrown LLM error into one of the three buckets. This is the single
+ * home for the `isContextSizeError -> html -> other` ladder that was previously
+ * inlined at every call site.
+ */
+export function classifyLlmError(err) {
+    const message = err instanceof Error ? err.message : String(err);
+    if (isContextSizeError(message))
+        return "context_limit";
+    if (err instanceof LlmCallError && err.code === "provider_html_error")
+        return "html";
+    return "other";
+}
+export async function callStructured(opts) {
+    const { feature, akmConfig, config, messages, request, parse, onError, fallback, onFallback } = opts;
+    const chat = request?.chat ?? chatCompletion;
+    const chatOptions = {
+        temperature: request?.temperature,
+        timeoutMs: request?.timeoutMs,
+        signal: request?.signal,
+        responseSchema: request?.responseSchema,
+        onRetryAttempt: request?.onRetryAttempt,
+    };
+    // UNGATED: run the chat+parse directly. Errors propagate — no `onError`
+    // funnel — matching the pre-gate behaviour of direct callers.
+    if (akmConfig === undefined) {
+        const raw = await chat(config, messages, chatOptions);
+        return parse(raw);
+    }
+    // GATED: run through `tryLlmFeature`. A throw inside is classified ONCE and
+    // routed to `onError`; `tryLlmFeature` returns `fallback` on disablement/timeout.
+    return tryLlmFeature(feature, akmConfig, async () => {
+        try {
+            const raw = await chat(config, messages, chatOptions);
+            return parse(raw);
+        }
+        catch (err) {
+            return onError(classifyLlmError(err), err);
+        }
+    }, fallback, {
+        timeoutMs: request?.timeoutMs,
+        onFallback,
+    });
+}

package/dist/output/context.js

@@ -12,10 +12,10 @@
  * Initialized from `cli.ts` before `runMain`.
  */
 import { UsageError } from "../core/errors.js";
-export const OUTPUT_FORMATS = ["json", "yaml", "text", "jsonl", "md", "html"];
-export const DETAIL_LEVELS = ["brief", "normal", "full"];
-export const SHAPE_MODES = ["human", "agent", "summary"];
-export function parseOutputFormat(value) {
+const OUTPUT_FORMATS = ["json", "yaml", "text", "jsonl", "md", "html"];
+const DETAIL_LEVELS = ["brief", "normal", "full"];
+const SHAPE_MODES = ["human", "agent", "summary"];
+function parseOutputFormat(value) {
     if (!value)
         return undefined;
     if (OUTPUT_FORMATS.includes(value))
@@ -29,7 +29,7 @@ export function parseDetailLevel(value) {
         return value;
     throw new UsageError(`Invalid value for --detail: ${value}. Expected one of: ${DETAIL_LEVELS.join("|")}`, "INVALID_DETAIL_VALUE");
 }
-export function parseShapeMode(value) {
+function parseShapeMode(value) {
     if (!value)
         return undefined;
     if (SHAPE_MODES.includes(value))

package/dist/output/renderers.js

@@ -468,6 +468,44 @@ const sessionMdRenderer = {
         };
     },
 };
+// ── 9. fact-md ───────────────────────────────────────────────────────────────
+/**
+ * Renderer for the `fact` asset type. A fact is durable stash-level semantic
+ * knowledge (personal/team/project details, coding conventions, stash-meta).
+ * It carries `category` (personal|team|project|convention|meta) and an
+ * optional `pinned` flag marking it as part of the always-injected core. The
+ * renderer surfaces a one-liner (category + pinned marker) so an agent can tell
+ * at a glance what kind of fact it is and whether it is core context.
+ */
+const factMdRenderer = {
+    name: "fact-md",
+    buildShowResponse(ctx) {
+        const name = deriveName(ctx);
+        const parsed = parseFrontmatter(ctx.content());
+        const fm = parsed.data;
+        const category = asNonEmptyString(fm.category);
+        const description = asNonEmptyString(fm.description);
+        const pinned = fm.pinned === true;
+        const headerParts = [
+            category ? `category: ${category}` : undefined,
+            pinned ? "pinned (core context)" : undefined,
+        ].filter((p) => !!p);
+        const action = [
+            "Durable stash fact — apply it as background context.",
+            headerParts.length > 0 ? headerParts.join("  ") : undefined,
+        ]
+            .filter((p) => !!p)
+            .join("\n");
+        return {
+            type: "fact",
+            name,
+            path: ctx.absPath,
+            action,
+            description,
+            content: parsed.content,
+        };
+    },
+};
 function applySessionMetadata(entry, ctx) {
     try {
         const fm = applyFrontmatterDescriptionAndTags(entry, ctx);
@@ -501,6 +539,34 @@ function applyTocMetadata(entry, ctx) {
         // Non-fatal: skip TOC if file can't be read
     }
 }
+/**
+ * Fact metadata: surface `category` and the `pinned` core marker as tags +
+ * search hints (no dedicated DB columns — same encoding pattern as session /
+ * task). `pinned` is mirrored to both a `pinned` tag and a `pinned` search
+ * hint so the ranking contributor can detect it and queries can target it.
+ */
+function applyFactMetadata(entry, ctx) {
+    try {
+        const fm = applyFrontmatterDescriptionAndTags(entry, ctx);
+        const tags = new Set([...(entry.tags ?? []), "fact"]);
+        const hints = new Set(entry.searchHints ?? []);
+        const category = asNonEmptyString(fm.category);
+        if (category) {
+            tags.add(category);
+            hints.add(`category:${category}`);
+        }
+        if (fm.pinned === true) {
+            tags.add("pinned");
+            hints.add("pinned");
+        }
+        entry.tags = Array.from(tags).filter(Boolean);
+        if (hints.size > 0)
+            entry.searchHints = Array.from(hints).filter(Boolean);
+    }
+    catch {
+        // Non-fatal: skip metadata extraction on parse error
+    }
+}
 /**
  * Parse frontmatter, apply description (if not already set) and merge tags
  * into `entry`. Returns the raw frontmatter data object so callers can access
@@ -660,6 +726,11 @@ registerMetadataContributor({
     appliesTo: ({ rendererName }) => rendererName === "session-md",
     contribute: (entry, ctx) => applySessionMetadata(entry, ctx.renderContext),
 });
+registerMetadataContributor({
+    name: "fact-md-metadata",
+    appliesTo: ({ rendererName }) => rendererName === "fact-md",
+    contribute: (entry, ctx) => applyFactMetadata(entry, ctx.renderContext),
+});
 // ── Registration ─────────────────────────────────────────────────────────────
 /** All built-in renderers. */
 const builtinRenderers = [
@@ -676,6 +747,7 @@ const builtinRenderers = [
     secretFileRenderer,
     taskMdRenderer,
     sessionMdRenderer,
+    factMdRenderer,
 ];
 /**
  * Register all built-in renderers with the file-context registry.
@@ -687,4 +759,4 @@ export function registerBuiltinRenderers() {
     }
 }
 // ── Named exports for testing ────────────────────────────────────────────────
-export { agentMdRenderer, commandMdRenderer, envFileRenderer, INTERPRETER_MAP, knowledgeMdRenderer, lessonMdRenderer, memoryMdRenderer, SETUP_SIGNALS, scriptSourceRenderer, secretFileRenderer, skillMdRenderer, wikiMdRenderer, workflowMdRenderer, };
+export { agentMdRenderer, commandMdRenderer, envFileRenderer, factMdRenderer, INTERPRETER_MAP, knowledgeMdRenderer, lessonMdRenderer, memoryMdRenderer, SETUP_SIGNALS, scriptSourceRenderer, secretFileRenderer, skillMdRenderer, wikiMdRenderer, workflowMdRenderer, };

package/dist/output/shapes/curate.js

@@ -5,7 +5,7 @@ import { capDescription, NORMAL_DESCRIPTION_LIMIT, pickFields } from "./helpers.
 // Curation is a small, high-signal top-N. Even at `brief` we keep `followUp`
 // (the actionable `akm show <ref>` command) and `reason` (why this asset was
 // selected) — these are the point of curate, unlike a bulk search listing.
-const BRIEF_FIELDS = ["source", "type", "name", "ref", "id", "followUp", "reason"];
+const BRIEF_FIELDS = ["source", "type", "name", "ref", "id", "supportRefs", "followUp", "reason"];
 const NORMAL_FIELDS = [
     "source",
     "type",
@@ -17,12 +17,24 @@ const NORMAL_FIELDS = [
     "keys",
     "parameters",
     "run",
+    "supportRefs",
     "followUp",
     "reason",
     "score",
 ];
 // Agent shape: the minimal field set an LLM needs to decide and act.
-const AGENT_FIELDS = ["source", "type", "name", "ref", "id", "description", "followUp", "reason", "score"];
+const AGENT_FIELDS = [
+    "source",
+    "type",
+    "name",
+    "ref",
+    "id",
+    "description",
+    "supportRefs",
+    "followUp",
+    "reason",
+    "score",
+];
 function shapeCurateItem(item, detail, shape) {
     if (shape === "agent") {
         return capDescription(pickFields(item, AGENT_FIELDS), NORMAL_DESCRIPTION_LIMIT);

package/dist/output/shapes/passthrough.js

@@ -23,7 +23,6 @@ const PASSTHROUGH_COMMANDS = [
     "agent-result",
     "clone",
     "config",
-    "db-backups",
     "disable",
     "enable",
     "env-create",

package/dist/output/text/helpers.js

@@ -1005,6 +1005,15 @@ export function formatCuratePlain(r, detail) {
             lines.push(`  run: ${String(item.run)}`);
         if (item.followUp)
             lines.push(`  show: ${String(item.followUp)}`);
+        if (Array.isArray(item.supportRefs) && item.supportRefs.length > 0) {
+            for (const support of item.supportRefs) {
+                if (!support.ref)
+                    continue;
+                const label = typeof support.type === "string" ? `[${support.type}] ` : "";
+                const why = typeof support.reason === "string" ? ` — ${support.reason}` : "";
+                lines.push(`  support: ${label}${String(support.ref)}${why}`);
+            }
+        }
         if (detail !== "brief" && item.reason)
             lines.push(`  why: ${String(item.reason)}`);
     }
@@ -1026,8 +1035,14 @@ export function formatCuratePlain(r, detail) {
 }
 export function formatInitPlain(r) {
     let out = `Stash initialized at ${r.stashDir ?? "unknown"}`;
-    if (r.configPath)
+    // When --dir scaffolded a secondary stash but the default was deliberately
+    // left untouched, tell the user instead of silently repointing their default.
+    if (r.defaultStashUpdated === false && typeof r.previousStashDir === "string" && r.previousStashDir) {
+        out += `\nYour default stash is unchanged (${r.previousStashDir}). Re-run with --set-default to make ${r.stashDir} the default.`;
+    }
+    else if (r.configPath) {
         out += `\nConfig saved to ${r.configPath}`;
+    }
     return out;
 }
 export function formatIndexPlain(r) {

package/dist/registry/providers/skills-sh.js

@@ -2,46 +2,12 @@
 // License, v. 2.0. If a copy of the MPL was not distributed with this
 // file, You can obtain one at https://mozilla.org/MPL/2.0/.
 import { fetchWithRetry } from "../../core/common.js";
-import { rethrowIfTestIsolationError } from "../../core/errors.js";
-import { closeDatabase, getRegistryIndexCache, openDatabase, upsertRegistryIndexCache } from "../../indexer/db/db.js";
 import { md5Hex } from "../../runtime.js";
+import { fetchCachedJson } from "../../storage/repositories/registry-cache.js";
 import { registerProvider } from "../factory.js";
 // ── Constants ───────────────────────────────────────────────────────────────
 /** Per-query cache TTL in milliseconds (15 minutes). */
 const QUERY_CACHE_TTL_MS = 15 * 60 * 1000;
-// ── Cache DB lifecycle ────────────────────────────────────────────────────────
-/**
- * RAII-style lifecycle helper for the registry cache DB. Opens the DB (treating
- * a failed open exactly like the legacy fall-through: the bun-test isolation
- * guard is re-thrown, any other failure yields `db = undefined`), runs `fn`,
- * and guarantees the DB is closed in a `finally` after `fn` has fully settled
- * (the await is required: the callbacks are async, and closing before they
- * settle would tear the DB down mid-write).
- */
-async function withRegistryCacheDb(fn) {
-    let db;
-    try {
-        db = openDatabase();
-    }
-    catch (err) {
-        // Never mask the bun-test isolation guard as "DB unavailable".
-        rethrowIfTestIsolationError(err);
-        db = undefined;
-    }
-    try {
-        return await fn(db);
-    }
-    finally {
-        if (db) {
-            try {
-                closeDatabase(db);
-            }
-            catch {
-                /* ignore */
-            }
-        }
-    }
-}
 // ── Provider class ──────────────────────────────────────────────────────────
 class SkillsShProvider {
     type = "skills-sh";
@@ -66,131 +32,39 @@ class SkillsShProvider {
             return { hits: [], warnings: [`Registry ${label}: ${message}`] };
         }
     }
-    // ── v1-spec §3.1 surface ────────────────────────────────────────────────
-    async searchKits(q) {
-        const result = await this.search({
-            query: q.text,
-            limit: q.limit ?? 20,
-            includeAssets: false,
-        });
-        return result.hits.map((hit) => ({
-            id: hit.id,
-            title: hit.title,
-            summary: hit.description,
-            installRef: hit.installRef,
-            score: hit.score,
-        }));
-    }
-    async searchAssets(q) {
-        const result = await this.search({
-            query: q.text,
-            limit: q.limit ?? 20,
-            includeAssets: true,
-        });
-        return (result.assetHits ?? []).map((hit) => ({
-            kitId: hit.stash.id,
-            type: hit.assetType,
-            name: hit.assetName,
-            summary: hit.description,
-            cloneRef: hit.action.replace(/^akm add\s+/, ""),
-        }));
-    }
-    /**
-     * skills.sh has no `getKit` API — every entry corresponds to a GitHub
-     * repository whose metadata we already include in the search result. We
-     * synthesize a manifest from the search hit when the caller knows the stash
-     * id; if not present in the most recent results, return null.
-     */
-    async getKit(id) {
-        if (!id.startsWith("skills-sh:"))
-            return null;
-        const slug = id.slice("skills-sh:".length);
-        // Best-effort: the API gives us search-by-name; extract the leaf segment.
-        const segments = slug.split("/").filter(Boolean);
-        const leaf = segments[segments.length - 1] ?? slug;
-        const result = await this.search({ query: leaf, limit: 50, includeAssets: false });
-        const match = result.hits.find((hit) => hit.id === id);
-        if (!match)
-            return null;
-        return { id: match.id, installRef: match.installRef };
-    }
-    /**
-     * skills.sh entries are always GitHub repositories. Claim only refs whose
-     * parsed source is `github`; defer everything else (npm tarballs, local
-     * paths, raw git URLs) to other registries.
-     */
-    canHandle(ref) {
-        return ref.source === "github";
-    }
     async fetchSkills(query, limit) {
         // Build a stable DB cache key for this query
         const dbCacheKey = this.queryDbCacheKey(query, limit);
-        return withRegistryCacheDb(async (db) => {
-            // ── Step 1: Try DB cache (index.db) ───────────────────────────────────
-            let dbCacheResult;
-            try {
-                if (db) {
-                    dbCacheResult = getRegistryIndexCache(db, dbCacheKey, QUERY_CACHE_TTL_MS);
-                }
-            }
-            catch (err) {
-                // Never mask the bun-test isolation guard as "DB unavailable" — see
-                // rethrowIfTestIsolationError in src/core/errors.ts. Without this,
-                // a leaky test silently gets a cold cache + fresh fetch instead of
-                // the loud TEST_ISOLATION_MISSING failure the guard intends.
-                rethrowIfTestIsolationError(err);
-                // index.db not available yet (pre-migration install or test env) — fall through
-            }
-            if (dbCacheResult) {
+        const baseUrl = this.config.url.replace(/\/+$/, "");
+        const url = `${baseUrl}/api/search?q=${encodeURIComponent(query)}&limit=${limit}`;
+        return fetchCachedJson({
+            cacheKey: dbCacheKey,
+            ttlMs: QUERY_CACHE_TTL_MS,
+            // A fresh hit returns even an empty array; a stale fallback only when
+            // non-empty. Corrupt cache JSON is swallowed and treated as a miss.
+            parseCache: (json, { stale }) => {
                 try {
-                    const parsed = JSON.parse(dbCacheResult.indexJson);
-                    if (Array.isArray(parsed)) {
-                        const entries = parsed.filter(isValidSkillsEntry);
-                        return entries;
-                    }
+                    const parsed = JSON.parse(json);
+                    if (!Array.isArray(parsed))
+                        return undefined;
+                    const entries = parsed.filter(isValidSkillsEntry);
+                    if (stale && entries.length === 0)
+                        return undefined;
+                    return entries;
                 }
                 catch {
-                    /* corrupt DB entry — fall through */
+                    return undefined;
                 }
-            }
-            // ── Step 2: Fetch from API ─────────────────────────────────────────────
-            const baseUrl = this.config.url.replace(/\/+$/, "");
-            const url = `${baseUrl}/api/search?q=${encodeURIComponent(query)}&limit=${limit}`;
-            try {
+            },
+            fetchFresh: async () => {
                 const response = await fetchWithRetry(url, undefined, { timeout: 10_000, retries: 1 });
                 if (!response.ok) {
                     throw new Error(`HTTP ${response.status}`);
                 }
                 const data = (await response.json());
                 const entries = parseSkillsResponse(data);
-                // Write to DB cache (primary)
-                if (db) {
-                    try {
-                        upsertRegistryIndexCache(db, dbCacheKey, JSON.stringify(entries));
-                    }
-                    catch {
-                        /* best-effort */
-                    }
-                }
-                return entries;
-            }
-            catch (err) {
-                // Fetch failed — use stale DB cache if available
-                if (dbCacheResult) {
-                    try {
-                        const parsed = JSON.parse(dbCacheResult.indexJson);
-                        if (Array.isArray(parsed)) {
-                            const entries = parsed.filter(isValidSkillsEntry);
-                            if (entries.length > 0)
-                                return entries;
-                        }
-                    }
-                    catch {
-                        /* ignore */
-                    }
-                }
-                throw err;
-            }
+                return { value: entries, cacheJson: JSON.stringify(entries) };
+            },
         });
     }
     mapToHits(entries) {