@ebony-vinyl/dsh-our-free-model 2.8.3 → 2.8.4

package/index.js

@@ -43,7 +43,7 @@ import { clearEacUser, readEacUser, writeEacUser } from './src/eac-user.js'
 import { createEacLoginPoller } from './src/eac-login.js'
 import { unlockSealedLane } from './src/vault.js'
 import { DEFAULT_LEVEL, budgetLadder, hasDeclaredEffortMenu } from './src/effort.js'
-import { modelAvailability } from './src/model-test.js'
+import { readPerf, readSeries, siteRootFor, modelAvailability } from './src/model-test.js'
 import { AnnouncementFeed } from './src/feed.js'
 import { createPickStats } from './src/picks.js'
 import { PluginUpdater, installationIsManaged } from './src/updater.js'
@@ -170,6 +170,19 @@ export function apply(ctx, config) {
   if (generation > 1) {
     settings.update({ reloadedAt: Date.now(), reloadCount: generation - 1 })
     settings.flush()
+  } else {
+    // Before issue #761 the only "already seen" marker lived in the browser's
+    // origin-scoped localStorage. A settings file from that build can carry an
+    // old reloadedAt with no acknowledgement beside it; adopting it here keeps
+    // a plain process restart (a new desktop port is a new origin) from
+    // replaying a stale notice. A genuinely new reload writes a new timestamp.
+    const stored = settings.get()
+    const at = Number(stored.reloadedAt ?? 0)
+    const seen = Number(stored.reloadNoticeSeen ?? 0)
+    if (Number.isSafeInteger(at) && at > 0 && seen === 0) {
+      settings.update({ reloadNoticeSeen: at })
+      settings.flush()
+    }
   }
 
   /** `managed` stands down everything that would rewrite the installed bytes. */
@@ -1152,6 +1165,10 @@ export function apply(ctx, config) {
     },
     subscriptions,
     onSubscriptionsChanged: syncAggregateCatalog,
+    // The sealed credential names the lane's base (`…/eac/v1`); the perf route
+    // derives the site root from it so the metrics endpoint is found without a
+    // second hard-coded address.
+    sealedCredential: sealedCredentialOf,
     // ── the aggregate lane's settings surface ────────────────────────────────
     // The registry is static data; the configured half is the settings file with
     // every key masked. Writing one provider rebuilds the catalog and re-emits
@@ -1404,6 +1421,7 @@ export function apply(ctx, config) {
       generation,
       reloadedAt: settings.get().reloadedAt ?? 0,
       reloadCount: settings.get().reloadCount ?? 0,
+      reloadNoticeSeen: settings.get().reloadNoticeSeen ?? 0,
       dataDir,
       autoReloadWatch: settings.get().autoReloadWatch === true,
       lastReload: globalThis[Symbol.for('our-free-model.last-reload')] ?? undefined,
@@ -1816,6 +1834,29 @@ function imageResolver(ctx, logger) {
 /** The fixed failure classes the pool proxy may surface to the client. */
 const POOL_REASONS = new Set(['no-lane', 'gateway-status', 'malformed', 'unreachable'])
 
+// ── per-model performance figures ─────────────────────────────────────────────
+// The roster asks for these on every mount, and they are hourly samples: a
+// minute of staleness costs nothing, while an outbound hop per page load is
+// pure load on the site. The in-flight promise is shared as well, so a burst of
+// mounts costs one request rather than one each.
+const PERF_FRESH_MS = 60_000
+let perfCache = { at: 0, data: null }
+let perfPending = null
+function perfSnapshot(root, fetchImpl) {
+  if (root === null || root === undefined) return Promise.resolve({ models: {} })
+  if (perfCache.data !== null && Date.now() - perfCache.at < PERF_FRESH_MS) return Promise.resolve(perfCache.data)
+  if (perfPending !== null) return perfPending
+  perfPending = readPerf({ root, fetch: fetchImpl })
+    .then(data => {
+      // Only a non-empty answer is worth keeping: caching "the site did not
+      // answer" would hide a recovery for a full minute.
+      if (Object.keys(data.models).length > 0) perfCache = { at: Date.now(), data }
+      return data
+    })
+    .finally(() => { perfPending = null })
+  return perfPending
+}
+
 export function createApiRoutes(deps) {
   return async function handler(req, res) {
     const url = new URL(req.url ?? '/', 'http://localhost')
@@ -1876,6 +1917,30 @@ export function createApiRoutes(deps) {
       if (method === 'GET' && routePath === '/stats') {
         return send(200, buildStats(deps.stats.get(), deps.catalog()))
       }
+      // Per-model performance figures for the roster's cards. The site's own
+      // metrics endpoint is public but sends no CORS headers, so the page
+      // cannot call it directly — this is the server-side hop. One request
+      // answers every model, and the map holds only models whose figure is
+      // trustworthy (see src/model-test.js); an absent model draws no strip.
+      //
+      // Cached for a minute and de-duplicated while in flight: the roster asks
+      // on every mount, and the figures are hourly samples — a fresh outbound
+      // request per page load buys nothing but load on the site.
+      if (method === 'GET' && routePath === '/perf') {
+        const credential = deps.sealedCredential?.() ?? null
+        const root = siteRootFor(credential?.base)
+        return send(200, await perfSnapshot(root, deps.perfFetch ?? directFetch))
+      }
+      // One model's hourly trend series, for the charts behind 详情. Separate
+      // from `/perf` because it is per model: the roster would otherwise pay
+      // twelve requests to draw twelve strips nobody has expanded yet.
+      if (method === 'GET' && routePath === '/perf/series') {
+        const modelId = url.searchParams.get('model') ?? ''
+        const credential = deps.sealedCredential?.() ?? null
+        const root = siteRootFor(credential?.base)
+        const series = await readSeries({ root, modelId, fetch: deps.perfFetch ?? directFetch })
+        return send(200, series === null ? { series: null } : series)
+      }
       if (method === 'GET' && routePath === '/outlet') {
         return send(200, await deps.outletStatus())
       }
@@ -1967,6 +2032,21 @@ export function createApiRoutes(deps) {
       if (method === 'GET' && routePath === '/meta') {
         return send(200, { ...deps.meta(), feed: { fetchedAt: deps.announcements.view().fetchedAt, source: deps.announcements.view().source, error: deps.announcements.view().error }, update: deps.update.status() })
       }
+      if (method === 'POST' && routePath === '/reload-notice/ack') {
+        const body = await readJson(req)
+        const at = body?.at
+        const current = Number(deps.settings.get().reloadedAt ?? 0)
+        // Only the timestamp this process announced may be acknowledged: an
+        // older page or a hand-written request must not pre-mark a future
+        // reload as seen.
+        if (typeof at !== 'number' || !Number.isSafeInteger(at) || at <= 0 || at !== current) {
+          return send(400, { error: 'at must match the current reloadedAt' })
+        }
+        deps.settings.update({ reloadNoticeSeen: at })
+        deps.settings.flush()
+        if (deps.settings.writeFailed === true) return send(500, { error: 'could not persist settings' })
+        return send(200, { ok: true, reloadNoticeSeen: at })
+      }
       if (method === 'POST' && routePath === '/eac/login/start') {
         const started = await deps.eacAuth.start()
         return send(started.error === undefined ? 200 : 404, started.error === undefined ? started : { error: started.error })

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@ebony-vinyl/dsh-our-free-model",
-  "version": "2.8.3",
+  "version": "2.8.4",
   "type": "module",
   "description": "Our Free Model — 在 dsh 里装上这个插件即可，无需登录、注册或填 API Key，就能使用包括 DeepSeek V4.1 Flash、Kimi K3 在内的前沿模型，桌面端另有 EAC 协付渠道；2.0 起接入 LobsterAI、Qoder、TRAE、Cline、Loomy、Raccoon、MiniMax、ZCode、Gemini 十个账号渠道（原有匿名免费通道保留）（登录、账号池、积分领取、本地 OpenAI 网关与局域网转发、数据看板与运行日志），六页毛玻璃界面统一呈现。",
   "repository": {

package/src/catalog.js

@@ -296,17 +296,33 @@ export const EAC_CHANNEL = 'eac'
 export const EAC_TAG = 'EAC'
 
 const EAC_DISPLAY_NAMES = {
-  // The OpenCode free lane is relayed server-side; the org prefix is routing
-  // detail, so the picker shows the model's own name with the lane's mark.
-  'opencode/deepseek-v4.1-flash-free': 'DeepSeek V4.1 Flash Highspeed',
-  'opencode/glm-5.3-flash-free': 'GLM 5.3 Flash Highspeed',
+  // These two ids are the lane's routing names; the models behind them are
+  // Dahl Inference's (`inference.dahl.global`), whose own docs name them
+  // "DeepSeek V4 Flash" and "GLM 5.3 Flash". The card shows that name, not the
+  // relay's internal spelling — the org prefix and the `-free` suffix are both
+  // routing detail.
+  'opencode/deepseek-v4.1-flash-free': 'DeepSeek V4 Flash',
+  'opencode/glm-5.3-flash-free': 'GLM 5.3 Flash',
   'deepseek-ai/deepseek-v4.1-flash': 'DeepSeek V4.1 Flash',
-  'moonshotai/kimi-k2.6': 'Kimi K2.6',
+  // Kimi K2.6 was removed from the NVIDIA channel on 2026-11-14 (its listing
+  // there never served this lane), so it is no longer named here either — an
+  // entry the catalog cannot receive is a name the picker never shows.
   'moonshotai/kimi-k3': 'Kimi K3',
   'openai/gpt-oss-20b': 'GPT-OSS 20B',
   'z-ai/glm-5.3': 'GLM 5.3',
   'z-ai/glm-5.3-flash': 'GLM 5.3 Flash',
   'mistralai/mistral-large-4-1t-a52b': 'Mistral Large 4 1T A52B',
+  // AMD Radeon (developer.amd.com.cn). The lane namespaces its ids `amd/…`;
+  // the card drops the prefix the same way it drops `org/` elsewhere.
+  'amd/DeepSeek-V4.1-Flash': 'DeepSeek V4.1 Flash (AMD)',
+  'amd/DeepSeek-V4-Flash': 'DeepSeek V4 Flash (AMD)',
+  'amd/Gemma-4-31B-IT': 'Gemma 4 31B IT (AMD)',
+  'amd/GLM-5.3-Flash': 'GLM 5.3 Flash (AMD)',
+  'amd/MiMo-V2.6-Flash': 'MiMo V2.6 Flash (AMD)',
+  'amd/MiniCPM5-2B': 'MiniCPM5 2B (AMD)',
+  'amd/MiniCPM-V-4.7-1B': 'MiniCPM-V 4.7 1B (AMD)',
+  'amd/MiniCPM-V-4.7-35B-A3B': 'MiniCPM-V 4.7 35B A3B (AMD)',
+  'amd/Qwen3.8-27B': 'Qwen3.8 27B (AMD)',
 }
 
 /**
@@ -364,24 +380,126 @@ const EAC_CAPABILITIES = [
     effortPatch: { reasoning_effort: '$effort' },
     effortOffPatch: { reasoning_effort: 'none' },
   },
-  // The OpenCode lane's two ids, relayed through the highspeed NVIDIA Build
-  // channel. The model is the same one the NVIDIA entries above describe, so
-  // its window and its menu are theirs: the shared `reasoning_effort` enum
-  // NVIDIA Build validates, with `none` as the off rung. `maxOutput` stays the
-  // free relay's own re-verified ceiling (384000 / 131072) — the relay enforces
-  // its own limit on the wire, which is tighter than NVIDIA's published one.
-  // The upstream id keeps its `-free` suffix (routing detail); the card's label
-  // says highspeed.
+  // The OpenCode lane's two ids, served by Dahl Inference
+  // (`inference.dahl.global`, docs read 2026-11-14). Both are the broker's own
+  // ids: DeepSeek-V4-Flash-0731 and GLM-5.3-Flash.
+  //
+  // - Window: Dahl's `/docs/models/` publishes 400K for both.
+  // - Vision: Dahl states plainly that vision "is not offered on current
+  //   models", and `image_url` content is rejected with a 400 — so `vision` is
+  //   false, not the NVIDIA Build `true` its namesakes carry.
+  // - Menu: Dahl validates `reasoning_effort` against
+  //   none|minimal|low|medium|high|xhigh|max and rejects `disabled` by name
+  //   ("unsupported value: got \"disabled\""), so the off rung is `none` — the
+  //   `effortOffPatch` below, which is what the plugin sends for `disabled`.
+  // - Output: DeepSeek's ceiling is measured, not guessed — a 5M `max_tokens`
+  //   request came back `finish_reason: length` at 8192 completion tokens, so
+  //   the broker caps output there regardless of what the caller asks for. GLM
+  //   shares that broker cap; Dahl publishes no per-model figure.
   {
     match: /^opencode\/deepseek-v4/,
-    vision: true, reasoning: true, contextWindow: 1048576, maxOutput: 384000,
+    vision: false, reasoning: true, contextWindow: 409600, maxOutput: 8192,
     efforts: ['disabled', 'minimal', 'low', 'medium', 'high', 'xhigh', 'max'], effortDefault: 'high',
     effortPatch: { reasoning_effort: '$effort' },
     effortOffPatch: { reasoning_effort: 'none' },
   },
   {
+    // GLM answers with reasoning even at `none` (measured: `none` still spent
+    // 57 reasoning tokens), so its menu has no off rung and the card shows the
+    // shared-budget tag rather than promising a switch it cannot honour.
     match: /^opencode\/glm-5\.3-flash/,
-    vision: true, reasoning: true, contextWindow: 1048576, maxOutput: 131072,
+    vision: false, reasoning: true, contextWindow: 409600, maxOutput: 8192,
+    efforts: ['minimal', 'low', 'medium', 'high', 'xhigh', 'max'], effortDefault: 'high',
+    effortPatch: { reasoning_effort: '$effort' },
+  },
+  // ── AMD Radeon (developer.amd.com.cn) ───────────────────────────────────────
+  //
+  // Two sources, each used for what it actually publishes:
+  //
+  // - The platform's own listing (`GET /radeon/api/v1/models`, read 2026-11-14)
+  //   carries `context_length`, `architecture.input_modalities`, and a
+  //   per-provider block with `vision`, `reasoning` and `tools`.
+  //   `supported_parameters` lists `reasoning_effort`, so the level is the real
+  //   control on the wire.
+  // - The output ceiling comes from **each model's own vendor docs**, because
+  //   the platform publishes none. Where a vendor publishes one it is used
+  //   verbatim; where a vendor publishes none either, `maxOutput` falls back to
+  //   the platform's window and the comment says so — a guess is worse than a
+  //   generous ceiling the vendor may clamp.
+  //
+  // Per-model vendor sources:
+  // - DeepSeek V4.1 Flash / V4 Flash: api-docs.deepseek.com model table —
+  //   1M context, 384K max output (`deepseek-flash`, "DeepSeek-V4.1-Flash-1017";
+  //   the older v4-flash name is served by the same model).
+  // - GLM-5.3-Flash: docs.z.ai/guides/llm/glm-5.3-flash — 1M context, 128K max
+  //   output. (The platform's listing reports 262144 for this id; the vendor's
+  //   own figure is the larger one, so the window below is the platform's, which
+  //   is the one requests actually hit.)
+  // - Gemma 4 31B IT: ai.google.dev model card — 256K context; Google publishes
+  //   no output ceiling, so it inherits the window.
+  // - MiMo V2.6 Flash / Qwen3.8 27B / MiniCPM-V 4.7 / MiniCPM5 2B: neither the
+  //   platform nor the vendor publishes an output ceiling (checked 2026-11-14),
+  //   so each inherits its window and the card shows the window as the ceiling.
+  {
+    match: /^amd\/DeepSeek-V4\.1-Flash/,
+    vision: true, reasoning: true, contextWindow: 1048576, maxOutput: 393216,
+    efforts: ['disabled', 'minimal', 'low', 'medium', 'high', 'xhigh', 'max'], effortDefault: 'high',
+    effortPatch: { reasoning_effort: '$effort' },
+    effortOffPatch: { reasoning_effort: 'none' },
+  },
+  {
+    match: /^amd\/DeepSeek-V4-Flash/,
+    vision: false, reasoning: true, contextWindow: 1048576, maxOutput: 393216,
+    efforts: ['disabled', 'minimal', 'low', 'medium', 'high', 'xhigh', 'max'], effortDefault: 'high',
+    effortPatch: { reasoning_effort: '$effort' },
+    effortOffPatch: { reasoning_effort: 'none' },
+  },
+  {
+    match: /^amd\/GLM-5\.3-Flash/,
+    vision: false, reasoning: true, contextWindow: 262144, maxOutput: 131072,
+    efforts: ['disabled', 'minimal', 'low', 'medium', 'high', 'xhigh', 'max'], effortDefault: 'high',
+    effortPatch: { reasoning_effort: '$effort' },
+    effortOffPatch: { reasoning_effort: 'none' },
+  },
+  {
+    // Xiaomi publishes no output ceiling for MiMo V2.6 Flash, so it inherits
+    // the window rather than a guessed 32K.
+    match: /^amd\/MiMo-V2\.6-Flash/,
+    vision: true, reasoning: true, contextWindow: 1048576, maxOutput: 1048576,
+    efforts: ['disabled', 'minimal', 'low', 'medium', 'high', 'xhigh', 'max'], effortDefault: 'high',
+    effortPatch: { reasoning_effort: '$effort' },
+    effortOffPatch: { reasoning_effort: 'none' },
+  },
+  {
+    // Google's Gemma 4 model card states 256K of context and no output ceiling.
+    match: /^amd\/Gemma-4-31B-IT/,
+    vision: true, reasoning: true, contextWindow: 262144, maxOutput: 262144,
+    efforts: ['disabled', 'minimal', 'low', 'medium', 'high', 'xhigh', 'max'], effortDefault: 'high',
+    effortPatch: { reasoning_effort: '$effort' },
+    effortOffPatch: { reasoning_effort: 'none' },
+  },
+  {
+    // Alibaba publishes no output ceiling for Qwen3.8-27B on the model-studio
+    // pages; it inherits the window.
+    match: /^amd\/Qwen3\.8-27B/,
+    vision: true, reasoning: true, contextWindow: 262144, maxOutput: 262144,
+    efforts: ['disabled', 'minimal', 'low', 'medium', 'high', 'xhigh', 'max'], effortDefault: 'high',
+    effortPatch: { reasoning_effort: '$effort' },
+    effortOffPatch: { reasoning_effort: 'none' },
+  },
+  {
+    // OpenBMB publishes no output ceiling for the MiniCPM-V 4.7 pair either.
+    match: /^amd\/MiniCPM-V/,
+    vision: true, reasoning: true, contextWindow: 262144, maxOutput: 262144,
+    efforts: ['disabled', 'minimal', 'low', 'medium', 'high', 'xhigh', 'max'], effortDefault: 'high',
+    effortPatch: { reasoning_effort: '$effort' },
+    effortOffPatch: { reasoning_effort: 'none' },
+  },
+  {
+    // MiniCPM5-2B: the project README states "Context Length 131,072" and no
+    // output ceiling.
+    match: /^amd\/MiniCPM5-2B/,
+    vision: false, reasoning: true, contextWindow: 131072, maxOutput: 131072,
     efforts: ['disabled', 'minimal', 'low', 'medium', 'high', 'xhigh', 'max'], effortDefault: 'high',
     effortPatch: { reasoning_effort: '$effort' },
     effortOffPatch: { reasoning_effort: 'none' },
@@ -431,25 +549,25 @@ const EAC_FEATURED = new Set([
 /**
  * Display-only ids for the card's small line.
  *
- * The OpenCode lane serves its ids with a `-free` suffix that is upstream
- * routing detail; this build relays those two models through the highspeed
- * NVIDIA Build channel, so the line shows the id the channel actually names
- * (`deepseek-v4.1-flash-free 1017` → `deepseek-v4.1-flash-highspeed 1017`).
- * Only the label moves: the catalog `id` is what the adapter matches, and it
- * keeps the upstream spelling.
+ * The lane serves its ids under its own routing names, with a `-free` suffix
+ * that is upstream detail. Behind them are Dahl Inference's models, and the
+ * line shows the id Dahl itself publishes (`GET /v1/models`, read 2026-11-14)
+ * — that is the string a reader can paste into a Dahl-compatible client. Only
+ * the label moves: the catalog `id` is what the adapter matches, and it keeps
+ * the lane's spelling.
  */
 const EAC_LABELS = {
-  'opencode/deepseek-v4.1-flash-free': 'deepseek-v4.1-flash-highspeed 1017',
-  'opencode/glm-5.3-flash-free': 'glm-5.3-flash-highspeed',
+  'opencode/deepseek-v4.1-flash-free': 'deepseek-ai/DeepSeek-V4-Flash-0731',
+  'opencode/glm-5.3-flash-free': 'zai-org/GLM-5.3-Flash',
 }
 
 function eacLabel(modelId) {
   const known = EAC_LABELS[modelId]
   if (known !== undefined) return known
   const bare = modelId.includes('/') ? modelId.slice(modelId.indexOf('/') + 1) : modelId
-  // Any remaining `-free` suffix is the same upstream detail; it is shown as
-  // the channel's highspeed mark rather than leaked into the card.
-  return bare.replace(/(?:^|[-_])free(?:$|[-_.])/, match => match.replace('free', 'highspeed'))
+  // Any remaining `-free` suffix is the lane's own routing detail; the line
+  // drops it rather than leaking it into the card.
+  return bare.replace(/(?:^|[-_])free(?:$|[-_.])/, '')
 }
 
 /** `org/model` → the model's own name under the channel tag. */
@@ -491,11 +609,10 @@ export function buildEacCatalog(ids) {
       // The card's small line under the name. The `id` keeps its `org/` prefix —
       // that is what keeps this row distinct from the free lane's same-named
       // model during the merge — but the prefix is routing detail, so the line
-      // shows the model's own id, with a display-only version marker where needed.
-      // The OpenCode lane's two ids are shown as their highspeed ids: the lane
-      // relays them through the highspeed NVIDIA Build channel, and the raw
-      // upstream id keeps its `-free` suffix, which is routing detail the line
-      // must not show. The catalog id is untouched, so routing still matches.
+      // shows the id the serving platform itself publishes: for the lane's two
+      // models that is Dahl's own `deepseek-ai/DeepSeek-V4-Flash-0731` /
+      // `zai-org/GLM-5.3-Flash`, which a reader can paste into any Dahl client.
+      // The catalog id is untouched, so routing still matches.
       label: eacLabel(id),
       channel: EAC_CHANNEL,
       wire: 'chat',