@aria-framework/ai 0.4.0 → 0.7.0

package/health.js

@@ -12,9 +12,19 @@
  * Reporting that as "healthy" hides the only fact worth knowing. So `reachable` and `modelPresent`
  * are separate, and the UI shows them separately.
  *
- * `modelPresent` is NULL, not false, where residency has no meaning — a hosted API does not load
- * models on demand, and flagging one as "not resident" would be a false alarm on a working
- * provider. Null means "not applicable", false means "we asked and it is not there".
+ * `modelPresent` is NULL, not false, whenever the answer is not KNOWN — a hosted API does not load
+ * models on demand, and a server that cannot report load state must not be guessed at. Null means
+ * "cannot tell", false means "we asked something that knows, and it said no".
+ *
+ * ── /v1/models CANNOT ANSWER THIS, WHICH TOOK REAL HARDWARE TO FIND ─────────────────────────────
+ * Measured against a live LM Studio: `/v1/models` returned FIVE models with keys `id, object,
+ * owned_by`, while only TWO were actually resident. It lists what is DOWNLOADED, not what is
+ * LOADED — so residency judged from it reports "loaded" for every model on the disk, and the
+ * evicted-model warning this module exists for could never fire at all.
+ *
+ * LM Studio's native `/api/v0/models` carries a `state` field (`loaded` / `not-loaded`), which is
+ * the real signal. So the probe asks that FIRST for kinds that load on demand, and falls back to
+ * `/v1/models` purely for reachability — reporting residency as null rather than inventing it.
  *
  * ── THE CIRCUIT BREAKER IS WHY THIS IS NOT JUST A PING ──────────────────────────────────────────
  * Without one, every call to a dead provider pays the full timeout before failing over. Annoying at
@@ -35,8 +45,19 @@ const DEFAULT_TIMEOUT_MS = 8000;
 /** Kinds that load models on demand, where residency is a real question. */
 const RESIDENT_KINDS = new Set(['lmstudio', 'openai-compatible']);
 
-/** Strip a trailing slash so `${base}/models` never becomes `//models`. */
-const apiRoot = (baseUrl) => String(baseUrl || '').replace(/\/+$/, '');
+/**
+ * THE ADAPTER'S root, not a second one.
+ *
+ * This file originally rolled its own — strip the trailing slash and append `/models` — which is
+ * wrong for the base URL people actually type. `http://host:1234` has no path, so the adapter adds
+ * `/v1`; the probe did not, and asked for `/models`. LM Studio answers that with
+ * "Unexpected endpoint or method. Returning 200 anyway" — a 200 with no model list — so the probe
+ * saw an empty listing and reported a perfectly loaded model as EVICTED, on a server that was
+ * answering completions the whole time.
+ *
+ * Two roots that disagree about where the API lives is a bug waiting to happen twice. There is one.
+ */
+const { apiRoot } = require('./providers/openai-compatible');
 
 /**
  * The default probe: ask the endpoint what models it has.
@@ -45,6 +66,39 @@ const apiRoot = (baseUrl) => String(baseUrl || '').replace(/\/+$/, '');
  * for a dead host and [] for a live host with nothing loaded, which is exactly the difference this
  * module exists to report.
  */
+/**
+ * LM Studio's native model list, which reports LOAD STATE.
+ *
+ * Returns null when the endpoint is not there (any other server), so the caller falls back rather
+ * than treating its absence as a failure.
+ */
+async function nativeResidency(cfg, fetchImpl, signalMs) {
+  // The native API sits beside the OpenAI-compatible one, not under /v1.
+  let base;
+  try {
+    const u = new URL(String(cfg.baseUrl || ''));
+    base = `${u.protocol}//${u.host}`;
+  } catch (_) { return null; }
+
+  try {
+    const res = await fetchImpl(`${base}/api/v0/models`, {
+      headers: cfg.apiKey ? { Authorization: `Bearer ${cfg.apiKey}` } : {},
+      signal: AbortSignal.timeout(signalMs)
+    });
+    if (!res.ok) return null;
+    const payload = await res.json();
+    if (!Array.isArray(payload && payload.data)) return null;
+    const loaded = payload.data
+      .filter((m) => m && m.state === 'loaded')
+      .map((m) => m.id || m.key || '')
+      .filter(Boolean);
+    // An empty `data` is a real answer (nothing loaded); a missing one was handled above.
+    return { loaded, all: payload.data.map((m) => (m && m.id) || '').filter(Boolean) };
+  } catch (_) {
+    return null;
+  }
+}
+
 async function defaultProbe(cfg, { fetchImpl = fetch, timeoutMs } = {}) {
   const url = `${apiRoot(cfg.baseUrl)}/models`;
   const started = Date.now();
@@ -62,25 +116,47 @@ async function defaultProbe(cfg, { fetchImpl = fetch, timeoutMs } = {}) {
     // lives in `e.cause.code`, so lead with that and keep the message as context.
     const cause = (e && e.cause) || {};
     const detail = cause.code || (e && e.message) || 'unreachable';
+    // NAME THE URL. Nothing normalises a base URL any more, so a wrong one is a real possibility
+    // and the operator's first question is "what did it actually call". `ECONNREFUSED` alone sends
+    // them to check the server; `ECONNREFUSED at http://host:1234/models` shows them the typo.
+    const reason = timedOut
+      ? `no response within ${timeoutMs || cfg.timeoutMs || DEFAULT_TIMEOUT_MS}ms`
+      : (cause.code ? `${cause.code}${cause.message ? ' — ' + cause.message : ''}` : detail);
     return {
-      reachable: false, models: [], ms: Date.now() - started,
-      error: timedOut
-        ? `no response within ${timeoutMs || cfg.timeoutMs || DEFAULT_TIMEOUT_MS}ms`
-        : (cause.code ? `${cause.code}${cause.message ? ' — ' + cause.message : ''}` : detail)
+      reachable: false, listed: false, models: [], loaded: null, ms: Date.now() - started,
+      url, error: `${reason} (${url})`
     };
   }
   const ms = Date.now() - started;
   if (!res.ok) {
     // It ANSWERED, so the host is reachable — the failure is auth, quota or a bad path, and saying
     // "unreachable" would send an operator to check the network instead of the key.
-    return { reachable: true, models: [], ms, error: `HTTP ${res.status}` };
+    return { reachable: true, listed: false, models: [], loaded: null, ms, url, error: `HTTP ${res.status} (${url})` };
   }
   let payload = null;
   try { payload = await res.json(); } catch (_) { payload = null; }
-  const models = Array.isArray(payload && payload.data)
+
+  // "A 200 THAT IS NOT A MODEL LISTING" IS NOT "NO MODELS LOADED".
+  //
+  // Servers answer 200 to things they do not implement — LM Studio logs
+  // "Unexpected endpoint or method. Returning 200 anyway" and returns a body with no `data`.
+  // Treating that as an empty list makes a loaded model look evicted, which is the exact false
+  // alarm this module exists to avoid. So `listed` says whether a real listing came back, and
+  // residency is only judged when it did.
+  const listed = Array.isArray(payload && payload.data);
+  const models = listed
     ? payload.data.map((m) => (m && (m.id || m.name)) || '').filter(Boolean)
     : [];
-  return { reachable: true, models, ms, error: null };
+
+  // Now ask something that actually knows about residency. Only for kinds that load on demand,
+  // and only as an addition — a server without it still reports reachable, with residency null.
+  let loaded = null;
+  if (RESIDENT_KINDS.has(cfg.provider)) {
+    const native = await nativeResidency(cfg, fetchImpl, timeoutMs || cfg.timeoutMs || DEFAULT_TIMEOUT_MS);
+    if (native) loaded = native.loaded;
+  }
+
+  return { reachable: true, listed, models, loaded, ms, url, error: null };
 }
 
 /**
@@ -166,9 +242,13 @@ function createHealthChecker(opts = {}) {
 
       const r = await probe(cfg);
       // Residency only means something where models are loaded on demand.
+      // RESIDENCY COMES FROM `loaded`, NEVER FROM THE MODEL LIST. `/v1/models` reports what is
+      // downloaded; judging residency from it says "loaded" for every model on the disk. Null
+      // whenever nothing authoritative answered — an unknown is honest, a guess is not.
       let modelPresent = null;
-      if (RESIDENT_KINDS.has(cfg.provider) && r.reachable && !r.error) {
-        modelPresent = cfg.model ? r.models.includes(cfg.model) : null;
+      if (RESIDENT_KINDS.has(cfg.provider) && r.reachable && !r.error
+          && Array.isArray(r.loaded) && cfg.model) {
+        modelPresent = r.loaded.includes(cfg.model);
       }
 
       // REACHABLE BUT NOT LOADED IS NOT A FAILURE. The call will still succeed; it will just be

package/index.js

@@ -177,6 +177,13 @@ module.exports = {
   get createProviderStore() { return require('./providerStore').createProviderStore; },
   // No database behind health, so it loads eagerly like the rest of the seam.
   ...require('./health'),
+  /**
+   * Where this package's EJS partials live, for the consumer's view-roots list.
+   *
+   * Same contract as backup/server/notify/uploads: the package knows its own layout, the app
+   * puts its own views FIRST so a local file of the same name wins.
+   */
+  viewsDir: require('path').join(__dirname, 'views'),
   get providerSchemaFor() { return require('./providerStore').schemaFor; },
   get usageSchemaFor() { return require('./usageStore').schemaFor; },
   createAiClient,

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "@aria-framework/ai",
   "description": "Aria App Framework — AI module. A dependency-injected model seam (createAiClient) over several providers (LM Studio / OpenAI-compatible / Anthropic), with a fact-preservation guard, generic Polish and Generate writing engines, and a browser polish widget. Prompts and config stay in the consuming app.",
-  "version": "0.4.0",
+  "version": "0.7.0",
   "license": "UNLICENSED",
   "private": false,
   "publishConfig": {
@@ -16,7 +16,7 @@
     "generate.js",
     "providers/openai-compatible.js",
     "providers/anthropic.js",
-    "browser/ai-polish.js", "usageStore.js", "providerStore.js", "health.js"
+    "browser/ai-polish.js", "usageStore.js", "providerStore.js", "health.js", "views/"
   ],
   "peerDependencies": {
     "@aria-framework/db-worker": ">=0.7.0"
@@ -25,6 +25,6 @@
     "@aria-framework/db-worker": { "optional": true }
   },
   "scripts": {
-    "test": "node test/smoke.js && node test/usageStore.js && node test/providerStore.js && node test/health.js"
+    "test": "node test/smoke.js && node test/usageStore.js && node test/providerStore.js && node test/health.js && node test/views.js"
   }
 }

package/providers/openai-compatible.js

@@ -251,15 +251,29 @@ async function httpError(res, label, apiKey) {
  * left alone, because other servers in this family mount elsewhere (LiteLLM behind a prefix, for
  * one) and second-guessing an explicit path would break them.
  */
+/**
+ * The configured base URL, with a trailing slash trimmed. Nothing else.
+ *
+ * IT USED TO APPEND `/v1` WHEN THE URL HAD NO PATH, and that was wrong for reasons that only got
+ * clearer the more providers there were:
+ *
+ *   - the stored value and the called value differed, so the form did not show what went on the
+ *     wire;
+ *   - it cannot generalise. One provider wants `/v1`, another `/api/v1`, another nothing at all.
+ *     A blanket rule is wrong for every provider it was not written for, and there is no way to
+ *     know which is which;
+ *   - it turned a typo into a SILENT wrong answer that surfaced three layers away. A base URL
+ *     missing `/v1` still completed chats — because this quietly fixed it — while the health probe
+ *     asked a path LM Studio answers with "Unexpected endpoint, returning 200 anyway", and a
+ *     perfectly loaded model was reported as evicted.
+ *
+ * So: call exactly what was configured. A wrong URL now fails immediately and says which URL it
+ * called, which is a configuration error an operator can act on rather than a mystery.
+ *
+ * Trailing-slash trimming stays because it changes no meaning — it only prevents `//models`.
+ */
 function apiRoot(baseUrl) {
-  const trimmed = String(baseUrl || '').replace(/\/+$/, '');
-  let path;
-  try {
-    path = new URL(trimmed).pathname.replace(/\/+$/, '');
-  } catch (err) {
-    return trimmed;   // not a URL we can parse; leave it exactly as typed
-  }
-  return path === '' ? trimmed + '/v1' : trimmed;
+  return String(baseUrl || '').replace(/\/+$/, '');
 }
 
 /** Both providers report tokens; they name the fields differently. One shape reaches the caller. */

package/views/ai/provider-card.ejs

@@ -0,0 +1,138 @@
+<%#
+  One provider: what it is, whether it works, and what it has cost.
+
+  LOCALS
+    p        a row from createProviderStore.all()
+    health   a status object from createHealthChecker.status(p.id) — may be undefined
+    usage    { calls, total_tokens } for the window, or undefined
+    canEdit  boolean — hides controls rather than disabling them, because a greyed-out
+             button still advertises a feature and invites a support ticket
+    routes   [{ id, position }] this provider appears in, for the shared-fallback warning
+
+  WHY "reachable" AND "model loaded" ARE SHOWN SEPARATELY: a local server answers /models
+  perfectly while the model itself has been evicted, and the next call then pays a
+  multi-second reload. One green dot would hide the only fact worth knowing.
+-%>
+<%
+  const h = typeof health !== 'undefined' && health ? health : { status: 'unknown', failures: 0 };
+  const local = p.kind === 'lmstudio' || p.kind === 'openai-compatible';
+  const dot = !p.enabled ? 'secondary'
+    : h.status === 'down' ? 'danger'
+    : h.status === 'up' ? (h.modelPresent === false ? 'warning' : 'success')
+    : 'secondary';
+  const ago = (ts) => {
+    if (!ts) return null;
+    const s = Math.max(0, Math.round((Date.now() - ts) / 1000));
+    if (s < 60) return s + 's ago';
+    if (s < 3600) return Math.round(s / 60) + 'm ago';
+    if (s < 86400) return Math.round(s / 3600) + 'h ago';
+    return Math.round(s / 86400) + 'd ago';
+  };
+  const shared = (typeof routes !== 'undefined' && routes) ? routes : [];
+-%>
+<article class="card mb-3">
+  <div class="card-body d-flex flex-wrap gap-3 justify-content-between align-items-start">
+    <div class="flex-grow-1" style="min-width:18rem">
+
+      <div class="d-flex align-items-center gap-2 mb-1">
+        <span class="badge rounded-pill bg-<%= dot %>" style="width:.6rem;height:.6rem;padding:0"
+              aria-hidden="true"></span>
+        <strong class="fs-6"><%= p.id %></strong>
+        <span class="badge text-bg-light border"><%= p.kind %></span>
+        <% if (local) { %><span class="badge text-bg-light border">local</span><% } %>
+        <% if (!p.enabled) { %><span class="badge text-bg-secondary">disabled</span><% } %>
+        <% if (h.status === 'down') { %>
+          <span class="badge text-bg-danger">
+            down<%= h.cooldownRemainingMs ? ' · retrying in ' + Math.ceil(h.cooldownRemainingMs / 1000) + 's' : '' %>
+          </span>
+        <% } else if (h.modelPresent === false) { %>
+          <span class="badge text-bg-warning">model not loaded</span>
+        <% } %>
+      </div>
+
+      <div class="text-body-secondary small font-monospace mb-2">
+        <%= p.model || '(no chat model)' %><% if (p.embedding_model) { %> · <%= p.embedding_model %><% } %>
+        <% if (p.base_url) { %> · <%= p.base_url %><% } %>
+      </div>
+
+      <div class="d-flex flex-wrap gap-4 small">
+        <% if (p.context_tokens) { %>
+          <div><div class="text-body-secondary text-uppercase" style="font-size:.68rem;letter-spacing:.06em">Context</div>
+            <span class="font-monospace"><%= Number(p.context_tokens).toLocaleString() %></span></div>
+        <% } %>
+        <% if (h.lastMs != null) { %>
+          <div><div class="text-body-secondary text-uppercase" style="font-size:.68rem;letter-spacing:.06em">Latency</div>
+            <span class="font-monospace"><%= h.lastMs %> ms</span></div>
+        <% } %>
+        <% if (typeof usage !== 'undefined' && usage) { %>
+          <div><div class="text-body-secondary text-uppercase" style="font-size:.68rem;letter-spacing:.06em">Calls</div>
+            <span class="font-monospace"><%= Number(usage.calls || 0).toLocaleString() %></span></div>
+          <div><div class="text-body-secondary text-uppercase" style="font-size:.68rem;letter-spacing:.06em">Tokens</div>
+            <span class="font-monospace"><%= Number(usage.total_tokens || 0).toLocaleString() %></span></div>
+        <% } %>
+        <div>
+          <%# LAST GOOD, not just a dot. "last good 3h ago" tells a story a green light cannot. %>
+          <div class="text-body-secondary text-uppercase" style="font-size:.68rem;letter-spacing:.06em">Last good</div>
+          <span class="font-monospace"><%= ago(h.lastGoodAt) || 'never' %></span>
+        </div>
+      </div>
+
+      <% if (p.daily_token_cap) { %>
+        <%
+          const spent = (typeof usage !== 'undefined' && usage) ? Number(usage.total_tokens || 0) : 0;
+          const pct = Math.min(100, Math.round((spent / Number(p.daily_token_cap)) * 100));
+        -%>
+        <div class="mt-3" style="max-width:26rem">
+          <div class="d-flex justify-content-between small text-body-secondary">
+            <span>Daily token cap</span>
+            <span class="font-monospace"><%= spent.toLocaleString() %> / <%= Number(p.daily_token_cap).toLocaleString() %></span>
+          </div>
+          <div class="progress" style="height:.4rem" role="progressbar" aria-valuenow="<%= pct %>"
+               aria-valuemin="0" aria-valuemax="100">
+            <div class="progress-bar bg-<%= pct >= 100 ? 'danger' : pct >= 80 ? 'warning' : 'success' %>"
+                 style="width:<%= pct %>%"></div>
+          </div>
+        </div>
+      <% } %>
+
+      <% if (h.status === 'down' && h.lastError) { %>
+        <div class="alert alert-danger d-flex gap-2 mt-3 mb-0 py-2">
+          <i class="bi bi-exclamation-triangle mt-1"></i>
+          <div class="small">
+            <strong>Marked down after <%= h.failures %> consecutive failure<%= h.failures === 1 ? '' : 's' %>.</strong>
+            Last error: <code><%= h.lastError %></code>. Calls skip this provider entirely rather than
+            paying its timeout, and it returns to service on the first success.
+          </div>
+        </div>
+      <% } else if (h.modelPresent === false) { %>
+        <div class="alert alert-warning d-flex gap-2 mt-3 mb-0 py-2">
+          <i class="bi bi-hourglass-split mt-1"></i>
+          <div class="small">
+            <strong>Reachable, but <code><%= p.model %></code> is not loaded.</strong>
+            The next call pays a reload before it does anything. This is not a failure and nothing
+            fails over — but if it keeps happening, turn off automatic model unloading on the server.
+          </div>
+        </div>
+      <% } %>
+
+      <% if (shared.length > 1) { %>
+        <div class="alert alert-warning d-flex gap-2 mt-3 mb-0 py-2">
+          <i class="bi bi-diagram-3 mt-1"></i>
+          <div class="small">
+            <strong>Serves <%= shared.length %> routes.</strong>
+            <%= shared.map((r) => r.id + ' (' + (r.position === 0 ? 'primary' : 'fallback') + ')').join(', ') %>.
+            If it goes down, one route fails over <em>and</em> another loses its safety net at the
+            same moment.
+          </div>
+        </div>
+      <% } %>
+    </div>
+
+    <div class="d-flex flex-column gap-2 align-items-end">
+      <% if (typeof canEdit !== 'undefined' && canEdit) { %>
+        <button class="btn btn-sm btn-outline-secondary" name="test_provider" value="<%= p.id %>">Test</button>
+        <a class="btn btn-sm btn-outline-secondary" href="?edit=<%= encodeURIComponent(p.id) %>">Edit</a>
+      <% } %>
+    </div>
+  </div>
+</article>

package/views/ai/usage-table.ejs

@@ -0,0 +1,58 @@
+<%#
+  Per-provider usage over a window.
+
+  LOCALS
+    rows    from createUsageStore.byProvider(sinceDay) — [{provider, model, calls, ...}]
+    prices  optional { [model]: costPerMillionTokens } supplied BY THE APP. Prices change, and
+            a stale hardcoded price silently produces confidently wrong cost reports, so the
+            package never carries one.
+    kinds   optional { [providerId]: kind } so a local endpoint can be shown as having no cost
+            rather than a cost of zero — different facts.
+-%>
+<%
+  const priceFor = (m) => (typeof prices !== 'undefined' && prices && prices[m] != null) ? Number(prices[m]) : null;
+  const kindOf = (id) => (typeof kinds !== 'undefined' && kinds) ? kinds[id] : undefined;
+  const isLocal = (id) => ['lmstudio', 'openai-compatible'].includes(kindOf(id));
+  const n = (v) => Number(v || 0).toLocaleString();
+-%>
+<div class="table-responsive">
+  <table class="table table-sm align-middle mb-0">
+    <thead>
+      <tr class="small text-body-secondary text-uppercase">
+        <th>Provider</th><th>Model</th>
+        <th class="text-end">Calls</th><th class="text-end">Prompt</th>
+        <th class="text-end">Completion</th><th class="text-end">Total</th><th class="text-end">Cost</th>
+      </tr>
+    </thead>
+    <tbody>
+      <% if (!rows || !rows.length) { %>
+        <tr><td colspan="7" class="text-body-secondary py-3">Nothing recorded in this window.</td></tr>
+      <% } %>
+      <% (rows || []).forEach(function (r) {
+           const price = priceFor(r.model);
+           const cost = price == null ? null : (Number(r.total_tokens || 0) / 1e6) * price;
+      -%>
+        <tr>
+          <td><%= r.provider %></td>
+          <td class="font-monospace small"><%= r.model %></td>
+          <td class="text-end font-monospace"><%= n(r.calls) %></td>
+          <td class="text-end font-monospace"><%= n(r.prompt_tokens) %></td>
+          <td class="text-end font-monospace"><%= n(r.completion_tokens) %></td>
+          <td class="text-end font-monospace"><%= n(r.total_tokens) %></td>
+          <td class="text-end font-monospace">
+            <%# A local GPU has NO cost, which is a different fact from a cost of zero. Showing
+                "0.00" would invite someone to compare it against a cloud row as if it were cheap. %>
+            <% if (isLocal(r.provider)) { %><span class="text-body-secondary">&mdash;</span>
+            <% } else if (cost == null) { %><span class="text-body-secondary" title="No price configured for this model">?</span>
+            <% } else { %><%= cost.toFixed(2) %><% } %>
+          </td>
+        </tr>
+      <% }); -%>
+    </tbody>
+  </table>
+</div>
+<p class="small text-body-secondary mt-2 mb-0">
+  Tokens are recorded for every call, including local models. Cost is derived from tokens and the
+  price table this app supplies &mdash; never stored &mdash; so a local endpoint still reports real
+  numbers, just different ones: throughput and utilisation rather than spend.
+</p>