dsh-plugin-discover 0.0.0-stage → 1.0.1

package/lib/vector.js

@@ -0,0 +1,592 @@
+/**
+ * Optional vector rerank over the BM25 top-50.
+ *
+ * The model lives in a child process (the optional model package); this module
+ * owns the fp16 caches, README excerpts, cosine similarity, and the RRF merge
+ * with the lexical order.
+ *
+ * The ranking fuses up to three ranked lists with weighted RRF; weights are
+ * mode-specific (RRF_WEIGHTS): fields mode BM25 1 : fields 0.35, readme mode
+ * BM25 1 : fields 0.3 : README 0.4. `fields` (balance) has only the first two
+ * lists; `readme` (high) computes both vector lists. A list a record is
+ * missing contributes nothing, so a record without a README vector still
+ * ranks through BM25 and its field vector.
+ *
+ * Modes:
+ *   off     — no vectors, no model
+ *   fields  — name/title/topics/capability/description vectors
+ *   readme  — field vectors plus README excerpt vectors
+ */
+import { ModelClient, ModelUnavailableError, resolveModelServer } from './model-client.js'
+import { VectorStore, decodeVector, hashText } from './vector-store.js'
+import { cleanReadmeText, README_CLEAN_VERSION } from './readme-clean.js'
+import {
+  README_CONCURRENCY,
+  README_RATE_PER_SECOND,
+  README_RATE_BURST,
+  README_REQUEST_TIMEOUT_MS,
+  README_TOTAL_BUDGET_MS,
+  NEGATIVE_CACHE_TTL_MS,
+  RateLimiter,
+  githubGuard,
+  githubNotice,
+  readmeTargets,
+  fetchGithubReadme,
+  fetchNpmReadme,
+  resolveNpmVersion,
+} from './readme-source.js'
+
+const RRF_K = 60
+const CACHE_TTL_MS = 90 * 86_400_000
+const README_TOKEN_LIMIT = 384
+const BATCH_SIZE = 64
+
+/** Identifies the ranking algorithm in code, reports, and diagnostics. */
+export const ALGORITHM_VERSION = 'v1.1'
+
+/** gate2 default: max query-to-candidate cosine must exceed this to count as semantic evidence. */
+export const GATE2_MAX_SIMILARITY = 0.85
+
+/**
+ * Weighted-RRF weights by mode (retuned 2026-10-25 on the real-usage set):
+ * the field list is weak on near-duplicate clusters, the README list carries
+ * high mode. A zero weight disables that list's contribution.
+ */
+export const RRF_WEIGHTS = Object.freeze({
+  fields: Object.freeze({ bm25: 1, fields: 0.35, readme: 0 }),
+  readme: Object.freeze({ bm25: 1, fields: 0.3, readme: 0.4 }),
+})
+
+let sharedClient = null
+
+function getSharedClient(serverPath) {
+  if (sharedClient && sharedClient.serverPath === serverPath && !sharedClient.closed) return sharedClient
+  sharedClient = new ModelClient(serverPath)
+  return sharedClient
+}
+
+/** Stop the idle model child (plugin unload or process shutdown). */
+export function disposeVectorRuntime() {
+  try { sharedClient?.dispose() } catch {}
+  sharedClient = null
+}
+
+export function vectorRuntimeAvailable(indexDir) {
+  return typeof indexDir === 'string' && resolveModelServer(indexDir) !== null
+}
+
+function fieldsText(record) {
+  return [
+    record.name,
+    record.title,
+    (record.topics ?? []).join(' '),
+    (record.capability ?? []).join(' '),
+    record.description,
+  ].filter(value => typeof value === 'string' && value.trim() !== '').join('\n')
+}
+
+function fieldKey(record) {
+  const contentHash = typeof record.contentHash === 'string' && record.contentHash !== ''
+    ? record.contentHash
+    : hashText(fieldsText(record))
+  return `${record.id}:${contentHash}`
+}
+
+function fresh(entry, now, ttl = CACHE_TTL_MS) {
+  return Boolean(entry && typeof entry.at === 'number' && now - entry.at < ttl)
+}
+
+function dot(a, b) {
+  const length = Math.min(a.length, b.length)
+  let sum = 0
+  for (let index = 0; index < length; index += 1) sum += a[index] * b[index]
+  return sum
+}
+
+/**
+ * Weighted reciprocal-rank fusion. Ranks are 1-based; a missing list
+ * contributes nothing. This keeps a record without a README vector eligible
+ * through its BM25 and field-vector ranks instead of dropping it.
+ */
+export function weightedRrf(ranks, weights) {
+  let score = 0
+  for (const list of ['bm25', 'fields', 'readme']) {
+    const rank = ranks?.[list]
+    const weight = weights?.[list] ?? 0
+    if (!Number.isFinite(rank) || rank < 1 || weight <= 0) continue
+    score += weight / (RRF_K + rank)
+  }
+  return score
+}
+
+function topTwoAverage(similarities) {
+  if (similarities.length === 0) return 0
+  const sorted = [...similarities].sort((a, b) => b - a)
+  return sorted.length >= 2 ? (sorted[0] + sorted[1]) / 2 : sorted[0]
+}
+
+async function mapLimit(items, limit, worker) {
+  const results = new Array(items.length)
+  let next = 0
+  async function run(workerIndex) {
+    while (next < items.length) {
+      const index = next
+      next += 1
+      results[index] = await worker(items[index], index, workerIndex)
+    }
+  }
+  const workers = Array.from({ length: Math.min(limit, items.length) }, (_, index) => run(index))
+  await Promise.all(workers)
+  return results
+}
+
+function excerptSource(entry, key) {
+  if (typeof entry?.source === 'string' && entry.source !== '') return entry.source
+  return key.startsWith('npm:') ? 'npm-cdn' : 'github-raw'
+}
+
+async function finishExcerpt(fetched, key, client, store, now) {
+  try {
+    const cleaned = cleanReadmeText(fetched.text)
+    if (cleaned.trim() === '') return null
+    const truncated = await client.truncate(cleaned, README_TOKEN_LIMIT)
+    if (typeof truncated.text !== 'string' || truncated.text.trim() === '') return null
+    const excerpt = {
+      text: truncated.text,
+      hash: hashText(truncated.text),
+      etag: fetched.etag ?? null,
+      fetchedAt: now,
+      source: fetched.source,
+      cleanVersion: README_CLEAN_VERSION,
+      ...(fetched.provider ? { provider: fetched.provider } : {}),
+    }
+    store.setExcerpt(key, excerpt)
+    return { ...excerpt, at: Date.now() }
+  } catch {
+    return null
+  }
+}
+
+async function readmeExcerpt(record, client, store, now, {
+  fetchImpl,
+  limiter,
+  githubDeadline,
+  requestTimeoutMs = README_REQUEST_TIMEOUT_MS,
+}) {
+  const { github, npm } = readmeTargets(record)
+  const githubKey = github?.key ?? ''
+  const githubCached = githubKey ? store.getExcerpt(githubKey) : null
+  const githubCurrent = Boolean(
+    githubCached && !githubCached.noReadme
+    && githubCached.cleanVersion === README_CLEAN_VERSION
+    && fresh(githubCached, now),
+  )
+  if (githubCurrent) {
+    return { ...githubCached, source: excerptSource(githubCached, githubKey) }
+  }
+  const githubNegative = Boolean(
+    githubCached?.noReadme && fresh(githubCached, now, NEGATIVE_CACHE_TTL_MS),
+  )
+  let stale = githubCached && !githubCached.noReadme
+    && githubCached.cleanVersion === README_CLEAN_VERSION
+    ? githubCached
+    : null
+
+  let npmVersion = npm?.version ?? null
+  let npmKey = npmVersion ? `npm:${npm.name}@${npmVersion}` : ''
+  let npmCached = npmKey ? store.getExcerpt(npmKey) : null
+  const npmNegative = Boolean(
+    npmCached?.noReadme && fresh(npmCached, now, NEGATIVE_CACHE_TTL_MS),
+  )
+
+  if (github && !githubNegative && Date.now() < githubDeadline && githubGuard.canUse()) {
+    const revalidate = githubCached && !githubCached.noReadme
+      && githubCached.cleanVersion === README_CLEAN_VERSION
+      ? githubCached
+      : null
+    const result = await fetchGithubReadme(github, {
+      fetchImpl,
+      limiter,
+      timeoutMs: requestTimeoutMs,
+      etag: revalidate?.etag ?? null,
+    })
+    if (result.notModified && revalidate) {
+      githubGuard.recordSuccess()
+      store.setExcerpt(githubKey, { ...revalidate, fetchedAt: now, source: 'github-raw' })
+      return { ...revalidate, at: Date.now(), fetchedAt: now, source: 'github-raw' }
+    }
+    if (result.status === 200 && typeof result.text === 'string') {
+      githubGuard.recordSuccess()
+      const excerpt = await finishExcerpt(result, githubKey, client, store, now)
+      if (excerpt) return excerpt
+    } else if (result.status === 404 || result.empty) {
+      githubGuard.recordSuccess()
+      store.setExcerpt(githubKey, { noReadme: true, source: 'github-raw' })
+    } else if (result.rateLimited) {
+      githubGuard.recordSuccess()
+      githubGuard.recordRateLimit(Date.now(), result.retryAfterMs ?? 0)
+    } else if (result.network || result.transport) {
+      githubGuard.recordTransportFailure(result.timeout ? 'timeout' : 'transport')
+    }
+  }
+
+  if (npm && !npmNegative) {
+    if (!npmVersion) {
+      npmVersion = await resolveNpmVersion(npm.name, {
+        fetchImpl,
+        limiter,
+        timeoutMs: requestTimeoutMs,
+        now: () => Date.now(),
+      })
+      if (npmVersion) {
+        npmKey = `npm:${npm.name}@${npmVersion}`
+        npmCached = store.getExcerpt(npmKey)
+      }
+    }
+    if (npmVersion) {
+      if (npmCached && !npmCached.noReadme && npmCached.cleanVersion === README_CLEAN_VERSION) {
+        if (fresh(npmCached, now)) {
+          return { ...npmCached, source: excerptSource(npmCached, npmKey) }
+        }
+        stale ??= npmCached
+      }
+      if (npmCached?.noReadme && fresh(npmCached, now, NEGATIVE_CACHE_TTL_MS)) return stale
+      const result = await fetchNpmReadme(npm.name, npmVersion, {
+        fetchImpl,
+        limiter,
+        timeoutMs: requestTimeoutMs,
+        etag: npmCached && npmCached.cleanVersion === README_CLEAN_VERSION ? npmCached.etag : null,
+      })
+      if (result.notModified && npmCached && npmCached.cleanVersion === README_CLEAN_VERSION) {
+        store.setExcerpt(npmKey, { ...npmCached, fetchedAt: now, source: 'npm-cdn' })
+        return { ...npmCached, at: Date.now(), fetchedAt: now, source: 'npm-cdn' }
+      }
+      if (result.status === 200 && typeof result.text === 'string') {
+        const excerpt = await finishExcerpt(result, npmKey, client, store, now)
+        if (excerpt) return excerpt
+      } else if (result.status === 404 || result.empty) {
+        store.setExcerpt(npmKey, { noReadme: true, source: 'npm-cdn' })
+      }
+    }
+  }
+  return stale
+}
+
+async function fillFields({ records, candidates, client, store, vectors, now }) {
+  const missing = []
+  for (const entry of candidates) {
+    const record = records[entry.doc]
+    const key = fieldKey(record)
+    const cached = store.getField(key)
+    if (fresh(cached, now)) {
+      vectors.set(entry.doc, decodeVector(cached.v))
+      continue
+    }
+    missing.push({ entry, record, key, text: fieldsText(record) })
+  }
+  for (let offset = 0; offset < missing.length; offset += BATCH_SIZE) {
+    const batch = missing.slice(offset, offset + BATCH_SIZE)
+    const started = Date.now()
+    const result = await client.embed(batch.map(item => item.text), 'passage')
+    if (process.env.DSH_VECTOR_DEBUG) {
+      const chars = batch.reduce((sum, item) => sum + item.text.length, 0)
+      console.error(`[vector] fields batch ${batch.length} texts, ${chars} chars, ${Date.now() - started} ms`)
+    }
+    for (let index = 0; index < batch.length; index += 1) {
+      const vector = result[index]
+      if (!vector) continue
+      vectors.set(batch[index].entry.doc, vector)
+      store.setField(batch[index].key, vector, batch[index].record.contentHash ?? null)
+    }
+  }
+}
+
+async function fillReadme({
+  records,
+  candidates,
+  client,
+  store,
+  vectors,
+  now,
+  fetchImpl,
+  readmeOptions = null,
+  readmeTotalBudgetMs = README_TOTAL_BUDGET_MS,
+}) {
+  githubGuard.beginRound(Date.now())
+  const concurrency = Math.max(1, Math.round(
+    Number(readmeOptions?.concurrency) || README_CONCURRENCY,
+  ))
+  const requestTimeoutMs = Math.max(1000,
+    Number(readmeOptions?.requestTimeoutMs) || README_REQUEST_TIMEOUT_MS)
+  const totalBudgetMs = readmeOptions?.totalBudgetMs !== undefined && readmeOptions?.totalBudgetMs !== null
+    ? Math.max(0, Number(readmeOptions.totalBudgetMs))
+    : readmeTotalBudgetMs
+  const githubDeadline = Date.now() + totalBudgetMs
+  const limiters = Array.from({ length: concurrency }, () => new RateLimiter({
+    rate: README_RATE_PER_SECOND,
+    burst: README_RATE_BURST,
+  }))
+  const tasks = candidates.map(entry => async workerIndex => {
+    const record = records[entry.doc]
+    const excerpt = await readmeExcerpt(record, client, store, now, {
+      fetchImpl,
+      limiter: limiters[workerIndex],
+      githubDeadline,
+      requestTimeoutMs,
+    })
+    if (!excerpt) return null
+    return { entry, record, excerpt, key: `${record.id}:${excerpt.hash}` }
+  })
+  const resolved = (await mapLimit(
+    tasks,
+    concurrency,
+    (task, _index, workerIndex) => task(workerIndex),
+  )).filter(Boolean)
+  const readmeSources = { github: 0, npm: 0, fields: 0 }
+  for (const item of resolved) {
+    if (item.excerpt.source === 'npm-cdn') readmeSources.npm += 1
+    else readmeSources.github += 1
+  }
+  const missing = []
+  let readmeCovered = 0
+  for (const item of resolved) {
+    const cached = store.getReadme(item.key)
+    if (fresh(cached, now)) {
+      vectors.set(item.entry.doc, decodeVector(cached.v))
+      readmeCovered += 1
+    } else {
+      missing.push(item)
+    }
+  }
+  for (let offset = 0; offset < missing.length; offset += BATCH_SIZE) {
+    const batch = missing.slice(offset, offset + BATCH_SIZE)
+    const result = await client.embed(batch.map(item => item.excerpt.text), 'passage')
+    for (let index = 0; index < batch.length; index += 1) {
+      const vector = result[index]
+      if (!vector) continue
+      vectors.set(batch[index].entry.doc, vector)
+      store.setReadme(batch[index].key, vector, batch[index].excerpt.hash)
+      readmeCovered += 1
+    }
+  }
+
+  // No field fallback here: the field list stays separate, so a record
+  // without a README vector still contributes through its field vector. The
+  // counter is coverage diagnostics only.
+  const readmeFallback = Math.max(0, candidates.length - readmeCovered)
+  readmeSources.fields = readmeFallback
+  const circuit = githubGuard.status()
+  return {
+    readmeCovered,
+    readmeFallback,
+    readmeSources,
+    ...(circuit ? { circuit, notice: githubNotice(circuit.reason) } : {}),
+  }
+}
+
+function degraded(mode, reason, candidates, error) {
+  const notice = degradationNotice(mode, reason, error)
+  return {
+    mode,
+    algorithm: ALGORITHM_VERSION,
+    covered: 0,
+    degraded: reason,
+    ...(error ? { error } : {}),
+    ...(notice ? { notice } : {}),
+    candidates,
+  }
+}
+
+/** English notice so the session LLM can warn the user about the fallback. */
+function degradationNotice(mode, reason, error) {
+  if (reason === 'runtime-missing') {
+    return `Vector rerank (mode "${mode}") was skipped: the local embedding model could not be started${error ? ` (${error})` : ''}. Results are lexical (BM25) only. Warn the user that vector enhancement degraded; suggest checking that the optional model package (dsh-plugin-discover-model) is installed and runnable, or setting vector.mode to "off".`
+  }
+  if (reason === 'embedding-failed') {
+    return `Vector rerank (mode "${mode}") failed while embedding: ${error ?? 'unknown error'}. Results are lexical (BM25) only. Warn the user that vector enhancement degraded for this query.`
+  }
+  return null
+}
+
+const MODEL_ATTEMPTS = 3
+const MODEL_RETRY_DELAYS_MS = [0, 800, 2000]
+
+function sleep(ms) {
+  return new Promise(resolve => setTimeout(resolve, ms))
+}
+
+/**
+ * Rerank the grouped BM25 top-50 with weighted RRF over the BM25, field-vector,
+ * and — in readme mode — README-vector lists (weights: RRF_WEIGHTS by mode).
+ * @param {object[]} candidates grouped candidates in lexical order
+ * @param {object[]} routes prepared query routes (original + rewrites)
+ * @param {object} options
+ */
+export async function vectorRerank(candidates, routes, {
+  mode = 'off',
+  gate2Threshold = GATE2_MAX_SIMILARITY,
+  indexDir = null,
+  records = [],
+  now = Date.now(),
+  clientFactory = null,
+  fetchImpl = null,
+  serverPath = null,
+  disposeClient = true,
+  readmeOptions = null,
+  readmeTotalBudgetMs = README_TOTAL_BUDGET_MS,
+} = {}) {
+  if (mode !== 'fields' && mode !== 'readme') return null
+  const selectedGate2Threshold = Number.isFinite(Number(gate2Threshold))
+    ? Number(gate2Threshold)
+    : GATE2_MAX_SIMILARITY
+  if (typeof indexDir !== 'string' || indexDir === '') return degraded(mode, 'index-dir-missing', candidates, null)
+  const resolvedServer = serverPath ?? resolveModelServer(indexDir)
+  if (!resolvedServer) return degraded(mode, 'runtime-missing', candidates, 'model package is not installed')
+  const store = new VectorStore(indexDir)
+  const fetchReader = fetchImpl ?? globalThis.fetch
+
+  async function attempt(client) {
+    const texts = routes
+      .map(route => (typeof route.text === 'string' && route.text !== '' ? route.text : (route.terms ?? []).join(' ')))
+      .filter(text => text !== '')
+    if (texts.length === 0) return degraded(mode, 'no-query-vectors', candidates, null)
+    const queryVectors = await client.embed(texts, 'query')
+    if (!queryVectors.length) return degraded(mode, 'no-query-vectors', candidates, null)
+
+    const fieldVectors = new Map()
+    const readmeVectors = new Map()
+    let readmeState = null
+    // Field vectors are cheap and always computed; the README list is an
+    // additional, independent list in `readme` mode.
+    await fillFields({ records, candidates, client, store, vectors: fieldVectors, now })
+    if (mode === 'readme') {
+      readmeState = await fillReadme({
+        records, candidates, client, store, vectors: readmeVectors, now, fetchImpl: fetchReader,
+        readmeOptions,
+        readmeTotalBudgetMs,
+      })
+    }
+
+    const scored = candidates.map((entry, bm25Rank) => {
+      const fieldVector = fieldVectors.get(entry.doc) ?? null
+      const readmeVector = readmeVectors.get(entry.doc) ?? null
+      const fieldSimilarities = fieldVector
+        ? queryVectors.map(query => dot(fieldVector, query))
+        : []
+      const readmeSimilarities = readmeVector
+        ? queryVectors.map(query => dot(readmeVector, query))
+        : []
+      const fieldSimilarity = fieldSimilarities.length ? topTwoAverage(fieldSimilarities) : null
+      const readmeSimilarity = readmeSimilarities.length ? topTwoAverage(readmeSimilarities) : null
+      const fieldMax = fieldSimilarities.length ? Math.max(...fieldSimilarities) : null
+      const readmeMax = readmeSimilarities.length ? Math.max(...readmeSimilarities) : null
+      const gate2MaxSimilarity = mode === 'readme'
+        ? Math.max(fieldMax ?? -Infinity, readmeMax ?? -Infinity)
+        : (fieldMax ?? -Infinity)
+      return {
+        entry,
+        bm25Rank,
+        fieldVector,
+        readmeVector,
+        fieldSimilarity,
+        readmeSimilarity,
+        fieldMaxSimilarity: fieldMax,
+        readmeMaxSimilarity: readmeMax,
+        gate2MaxSimilarity,
+      }
+    })
+    const fieldOrder = scored
+      .filter(item => item.fieldVector)
+      .sort((a, b) => b.fieldSimilarity - a.fieldSimilarity
+        || a.bm25Rank - b.bm25Rank
+        || a.entry.doc - b.entry.doc)
+    const fieldRank = new Map(fieldOrder.map((item, index) => [item.entry.doc, index + 1]))
+    const readmeOrder = scored
+      .filter(item => item.readmeVector)
+      .sort((a, b) => b.readmeSimilarity - a.readmeSimilarity
+        || a.bm25Rank - b.bm25Rank
+        || a.entry.doc - b.entry.doc)
+    const readmeRank = new Map(readmeOrder.map((item, index) => [item.entry.doc, index + 1]))
+
+    const weights = RRF_WEIGHTS[mode]
+    for (const item of scored) {
+      const fieldRankValue = fieldRank.get(item.entry.doc)
+      const readmeRankValue = readmeRank.get(item.entry.doc)
+      const gate2Max = Number.isFinite(item.gate2MaxSimilarity) ? item.gate2MaxSimilarity : null
+      item.entry.gate2MaxSimilarity = gate2Max
+      item.entry.gate2Pass = gate2Max !== null && gate2Max > selectedGate2Threshold
+      item.entry.fieldMaxSimilarity = item.fieldMaxSimilarity
+      item.entry.readmeMaxSimilarity = item.readmeMaxSimilarity
+      // Ranks are 1-based; a missing list contributes nothing.
+      item.entry.score = weightedRrf({
+        bm25: item.bm25Rank + 1,
+        fields: fieldRankValue,
+        readme: readmeRankValue,
+      }, weights)
+      item.entry.vectorSimilarity = mode === 'readme'
+        ? (item.readmeSimilarity ?? item.fieldSimilarity)
+        : item.fieldSimilarity
+      item.entry.fieldSimilarity = item.fieldSimilarity
+      item.entry.readmeSimilarity = item.readmeSimilarity
+      item.entry.fieldRank = fieldRankValue ?? null
+      item.entry.readmeRank = readmeRankValue ?? null
+    }
+    const ordered = scored
+      .slice()
+      .sort((a, b) => b.entry.score - a.entry.score
+        || a.bm25Rank - b.bm25Rank
+        || a.entry.doc - b.entry.doc)
+      .map(item => item.entry)
+    const covered = scored.filter(item => item.fieldVector || item.readmeVector).length
+    if (covered === 0) return degraded(mode, 'no-vectors', candidates, null)
+    const gate2Values = scored
+      .map(item => item.gate2MaxSimilarity)
+      .filter(value => Number.isFinite(value))
+    const gate2MaxSimilarity = gate2Values.length ? Math.max(...gate2Values) : null
+    const gate2 = {
+      threshold: selectedGate2Threshold,
+      maxSimilarity: gate2MaxSimilarity,
+      passed: gate2MaxSimilarity !== null && gate2MaxSimilarity > selectedGate2Threshold,
+    }
+    return {
+      mode,
+      algorithm: ALGORITHM_VERSION,
+      covered,
+      fieldCovered: fieldOrder.length,
+      gate2,
+      candidates: ordered,
+      ...(mode === 'readme' ? readmeState : {}),
+    }
+  }
+
+  // Retry with a fresh model process so a transient spawn failure does not
+  // silently degrade to lexical-only results; caches make retries cheap.
+  let lastError = null
+  try {
+    for (let round = 0; round < MODEL_ATTEMPTS; round += 1) {
+      if (MODEL_RETRY_DELAYS_MS[round] > 0) await sleep(MODEL_RETRY_DELAYS_MS[round])
+      const ephemeral = Boolean(clientFactory) || disposeClient
+      const client = clientFactory
+        ? clientFactory(resolvedServer)
+        : (ephemeral ? new ModelClient(resolvedServer) : getSharedClient(resolvedServer))
+      try {
+        return await attempt(client)
+      } catch (error) {
+        if (!(error instanceof ModelUnavailableError)) {
+          return degraded(mode, 'embedding-failed', candidates, error instanceof Error ? error.message : String(error))
+        }
+        lastError = error
+        try { client.dispose?.() } catch {}
+        if (!ephemeral && sharedClient === client) sharedClient = null
+      } finally {
+        if (ephemeral) {
+          try { client.dispose?.() } catch {}
+        }
+      }
+    }
+    const message = lastError instanceof Error ? lastError.message : String(lastError ?? 'model unavailable')
+    return degraded(mode, 'runtime-missing', candidates, message)
+  } finally {
+    try { store.flush() } catch {}
+  }
+}

package/locale/en.json

@@ -0,0 +1,6 @@
+{
+  "meta": {
+    "title": "Plugin Discover",
+    "description": "Search DSH plugins locally with Chinese natural language: deterministic ranking, quality filters, install targets. No server, no model calls, never installs."
+  }
+}

package/locale/zh.json

@@ -0,0 +1,6 @@
+{
+  "meta": {
+    "title": "插件检索",
+    "description": "用中文口语在本地检索 DSH 插件：确定性排序、质量过滤、给出安装目标。不连服务端、不调模型、从不安装。"
+  }
+}

package/package.json

@@ -1,6 +1,67 @@
-{
-  "name": "dsh-plugin-discover",
-  "version": "0.0.0-stage",
-  "stub": true,
-  "description": "Temporary package placeholder for staged publishing"
-}
+{
+  "name": "dsh-plugin-discover",
+  "version": "1.0.1",
+  "description": "Find DeepSeek Harness plugins with a local, model-free index: Chinese natural-language search, quality filters, deterministic ranking, and install targets — never installs anything.",
+  "type": "module",
+  "main": "index.js",
+  "exports": {
+    ".": "./index.js",
+    "./client": "./client.js",
+    "./package.json": "./package.json",
+    "./locale/*.json": "./locale/*.json",
+    "./icon": "./icon.svg"
+  },
+  "bin": {
+    "dsh-plugin-discover": "lib/cli.js"
+  },
+  "files": [
+    "index.js",
+    "client.js",
+    "lib",
+    "cordis.patch.yml",
+    "locale/*.json",
+    "icon.svg",
+    "README.md",
+    "README.zh.md",
+    "CHANGELOG.md",
+    "LICENSE"
+  ],
+  "scripts": {
+    "test": "node --test --test-isolation=none test/*.test.js"
+  },
+  "keywords": [
+    "deepseek-harness",
+    "dsh",
+    "dsh-plugin",
+    "plugin-discovery",
+    "search",
+    "bm25",
+    "chinese",
+    "local-first",
+    "offline"
+  ],
+  "license": "MIT",
+  "engines": {
+    "node": ">=22.19.0"
+  },
+  "dsh": {
+    "bundle": {
+      "patch": "./cordis.patch.yml"
+    },
+    "client": {
+      "platform": "web",
+      "immediately": true,
+      "inject": [
+        "@deepseek-ai/dsh-client-locale",
+        "@deepseek-ai/dsh-client-ui-plugin-manager"
+      ]
+    }
+  },
+  "peerDependencies": {
+    "@deepseek-ai/cordis": "^4.0.5",
+    "@deepseek-ai/dsh-tools": "^0.2.1-rc.2"
+  },
+  "dependencies": {
+    "@deepseek-ai/schemastery": "^3.18.1"
+  }
+}