refactor(model-routing): centralize explicit foreground fallback policy (#6020)

* refactor(model-routing): centralize explicit foreground fallback policy Make foreground fallback an explicit per-user, availability-only policy shared by streaming Chat, non-stream Chat, and Agent runs. Preserve strict defaults, owner/model and credential boundaries, pinned Agent routes, and truthful per-round provenance/accounting. Carry provider-reported model identifiers through native streaming adapters, non-stream responses, and caches, and keep legacy default_model_fallbacks as tombstoned raw storage that generic settings APIs and agent tools cannot expose or mutate. * fix(agent-loop): restore rebase-dropped qwen routing, workspace prompt, and temperature clamp * fix(model-routing): thread selected endpoint identity, fix cost classification and fallback eligibility * fix(chat): restore stream helpers and harden run stop lifecycle * fix(model-routing): let numeric provider codes win over symbolic rate-limit statuses * fix(agent-loop): apply qwen temperature and notes-tool clamps per fallback candidate * fix(chat): honor queued stop across resend and reload canonical terminal on EOF * fix(chat): track stop queue and cleanup ownership by per-send generation * fix(agent-loop): preserve requested temperature for non-qwen fallback candidates * fix(chat): reserve send ownership before any await and scope stop to the current send * fix(chat): clear the previous run identity at send reservation --------- Co-authored-by: RaresKeY <158580472+RaresKeY@users.noreply.github.com> Co-authored-by: StressTestor <212606152+StressTestor@users.noreply.github.com>
2026-10-19 13:22:20 +02:00 · 2026-08-14 08:10:30 +01:00
parent b52296471b
commit c4369305f0
46 changed files with 11418 additions and 1006 deletions
@@ -615,10 +615,36 @@ export function sameModelName(left, right) {
    || shortModel(a).toLowerCase() === shortModel(b).toLowerCase();
 }

-export function modelRouteLabel(requestedModel, actualModel) {
+function shortEndpointLabel(label) {
+  const value = modelValue(label);
+  if (!value) return '';
+  return value.length > 18 ? value.slice(0, 17) + '…' : value;
+}
+
+export function modelRouteLabel(
+  requestedModel,
+  actualModel,
+  requestedEndpointLabel = '',
+  actualEndpointLabel = '',
+  requestedEndpointId = '',
+  actualEndpointId = '',
+) {
  const requested = modelValue(requestedModel);
  const actual = modelValue(actualModel) || requested;
-  if (!requested || sameModelName(requested, actual)) return shortModel(actual || requested);
+  const requestedRoute = modelValue(requestedEndpointId || requestedEndpointLabel);
+  const actualRoute = modelValue(actualEndpointId || actualEndpointLabel);
+  const routeChanged = Boolean(
+    actualRoute
+    && requestedRoute
+    && actualRoute !== requestedRoute
+  );
+  if (!requested || sameModelName(requested, actual)) {
+    const model = shortModel(actual || requested);
+    if (!routeChanged) return model;
+    const from = shortEndpointLabel(requestedEndpointLabel || 'Selected route');
+    const to = shortEndpointLabel(actualEndpointLabel || actualEndpointId);
+    return model + ' (' + from + ' -> ' + to + ')';
+  }
  return shortModel(requested) + ' -> ' + shortModel(actual);
 }

@@ -629,10 +655,24 @@ export function replyModelPair(modelName, metadata) {
  if (actualFromMeta || requestedFromMeta) {
    const actual = actualFromMeta || requestedFromMeta || modelValue(modelName);
    const requested = requestedFromMeta || actual;
-    return { requestedModel: requested, actualModel: actual };
+    return {
+      requestedModel: requested,
+      actualModel: actual,
+      requestedEndpointId: meta.requested_endpoint_id || null,
+      requestedEndpointLabel: meta.requested_endpoint_label || 'Selected route',
+      actualEndpointId: meta.endpoint_id || null,
+      actualEndpointLabel: meta.endpoint_label || meta.requested_endpoint_label || 'Selected route',
+    };
  }
  const fallback = modelValue(modelName);
-  return { requestedModel: fallback, actualModel: fallback };
+  return {
+    requestedModel: fallback,
+    actualModel: fallback,
+    requestedEndpointId: null,
+    requestedEndpointLabel: 'Selected route',
+    actualEndpointId: null,
+    actualEndpointLabel: 'Selected route',
+  };
 }

 /**
@@ -824,12 +864,50 @@ export function isCostTrackedEndpoint(url) {
 }

 /** Cost for the current turn, returning null for non-billable endpoints. */
-function _billableCost(model, inputTokens, outputTokens) {
-  const url = _currentEndpointUrl();
-  if (!isCostTrackedEndpoint(url)) return null;
+function _billableCost(model, inputTokens, outputTokens, endpointCostTracked, selectedEndpointUrl) {
+  // Foreground fallback can answer on a different endpoint than the session's
+  // selected route. Prefer the backend's non-secret actual-route
+  // classification; retain the selected-endpoint check for older history.
+  if (endpointCostTracked === false) return null;
+  const selectedUrl = selectedEndpointUrl === undefined
+    ? _currentEndpointUrl()
+    : selectedEndpointUrl;
+  if (endpointCostTracked !== true && !isCostTrackedEndpoint(selectedUrl)) {
+    return null;
+  }
  return getModelCost(model, inputTokens, outputTokens);
 }

+/** Sum cost using the route/model that produced each Agent round. */
+function _metricsBillableCost(metrics, model, inputTokens, outputTokens, selectedEndpointUrl) {
+  const buckets = Array.isArray(metrics.usage_buckets) ? metrics.usage_buckets : [];
+  if (!buckets.length) {
+    return _billableCost(
+      model,
+      inputTokens,
+      outputTokens,
+      metrics.endpoint_cost_tracked,
+      selectedEndpointUrl,
+    );
+  }
+  let total = 0;
+  let hasPricedUsage = false;
+  for (const bucket of buckets) {
+    if (!bucket || typeof bucket !== 'object') continue;
+    const bucketCost = _billableCost(
+      bucket.model || model,
+      Number(bucket.input_tokens) || 0,
+      Number(bucket.output_tokens) || 0,
+      bucket.endpoint_cost_tracked,
+      selectedEndpointUrl,
+    );
+    if (bucketCost === null) continue;
+    total += bucketCost;
+    hasPricedUsage = true;
+  }
+  return hasPricedUsage ? total : null;
+}
+
 export function getImageCost(model, quality, size) {
  if (!model) return null;
  const m = model.toLowerCase();
@@ -844,6 +922,9 @@ export function getImageCost(model, quality, size) {

 /* ── Session cost helpers ─────────────────────────────────────────── */
 const _COST_KEY = 'ody-session-cost';
+const _COST_RUNS_KEY = 'ody-session-cost-runs';
+const _MAX_COST_RUNS_PER_SESSION = 256;
+const _COST_LEDGER_LOCK = 'odysseus-session-cost-ledger';

 /** Return the accumulated cost for the current (or given) session. */
 export function getSessionCost(sessionId) {
@@ -851,7 +932,14 @@ export function getSessionCost(sessionId) {
  if (!sid) return 0;
  try {
    const costs = JSON.parse(localStorage.getItem(_COST_KEY) || '{}');
-    return costs[sid] || 0;
+    const runCosts = JSON.parse(localStorage.getItem(_COST_RUNS_KEY) || '{}');
+    const recordedRuns = runCosts[sid] && typeof runCosts[sid] === 'object'
+      ? Object.values(runCosts[sid])
+      : [];
+    return (costs[sid] || 0) + recordedRuns.reduce(
+      (total, value) => total + (Number(value) || 0),
+      0,
+    );
  } catch (_e) { return 0; }
 }

@@ -863,6 +951,9 @@ export function resetSessionCost(sessionId) {
    const costs = JSON.parse(localStorage.getItem(_COST_KEY) || '{}');
    delete costs[sid];
    localStorage.setItem(_COST_KEY, JSON.stringify(costs));
+    const runCosts = JSON.parse(localStorage.getItem(_COST_RUNS_KEY) || '{}');
+    delete runCosts[sid];
+    localStorage.setItem(_COST_RUNS_KEY, JSON.stringify(runCosts));
  } catch (_e) { /* ignore */ }
  updateSessionCostUI();
 }
@@ -871,21 +962,8 @@ export function resetSessionCost(sessionId) {
 export function updateSessionCostUI() {
  const el = document.getElementById('session-cost-display');
  if (!el) return;
-  // Non-billable endpoint? Hide the badge and clear stale cost that a previous
-  // cloud-rate calculation may have left in localStorage for this session.
-  const _url = _currentEndpointUrl();
-  if (!isCostTrackedEndpoint(_url)) {
-    const sid = window.sessionModule && window.sessionModule.getCurrentSessionId();
-    if (sid && getSessionCost(sid) > 0) {
-      try {
-        const costs = JSON.parse(localStorage.getItem(_COST_KEY) || '{}');
-        delete costs[sid];
-        localStorage.setItem(_COST_KEY, JSON.stringify(costs));
-      } catch (_e) { /* ignore */ }
-    }
-    el.style.display = 'none';
-    return;
-  }
+  // The ledger records billable work already performed in this session. A
+  // selected local endpoint does not erase cost from a paid fallback route.
  const cost = getSessionCost();
  if (cost > 0) {
    el.textContent = '$' + (cost < 0.01 ? cost.toFixed(4) : cost < 1 ? cost.toFixed(3) : cost.toFixed(2));
@@ -895,6 +973,94 @@ export function updateSessionCostUI() {
  }
 }

+/** Record one metrics payload in a session ledger at most once. */
+export function recordSessionMetricsCost(metrics, sessionId, selectedEndpointUrl) {
+  if (!metrics || typeof metrics !== 'object') return null;
+  const cost = _metricsBillableCost(
+    metrics,
+    metrics.model || 'Unknown',
+    metrics.input_tokens || 0,
+    metrics.output_tokens || 0,
+    selectedEndpointUrl,
+  );
+  if (metrics._fromHistory) return cost;
+  const sid = sessionId || (
+    window.sessionModule && window.sessionModule.getCurrentSessionId()
+  );
+  if (!sid || cost === null) return cost;
+  const runId = typeof metrics._costRecordId === 'string'
+    ? metrics._costRecordId.trim()
+    : '';
+  if ((metrics._costRecorded || metrics._costRecordPending) && !runId) return cost;
+  // Recorded is only set once the write actually runs; pending covers the
+  // window while the write waits on the cross-tab lock, so a replay in that
+  // window cannot double-add and a tab closed mid-queue never claims recorded.
+  metrics._costRecordPending = true;
+  const writeCost = () => {
+    if (runId) {
+      try {
+        const runCosts = JSON.parse(localStorage.getItem(_COST_RUNS_KEY) || '{}');
+        const sessionRuns = runCosts[sid] && typeof runCosts[sid] === 'object'
+          ? runCosts[sid]
+          : {};
+        // Assigning by detached-run identity is replay-idempotent even when a
+        // refresh produces a fresh metrics object. The Web Lock around this
+        // read/modify/write also keeps distinct runs from two tabs from
+        // overwriting one another's stale snapshot.
+        sessionRuns[runId] = cost;
+        const entries = Object.entries(sessionRuns);
+        if (entries.length > _MAX_COST_RUNS_PER_SESSION) {
+          const overflow = entries.slice(0, entries.length - _MAX_COST_RUNS_PER_SESSION);
+          const costs = JSON.parse(localStorage.getItem(_COST_KEY) || '{}');
+          costs[sid] = (costs[sid] || 0) + overflow.reduce(
+            (total, entry) => total + (Number(entry[1]) || 0),
+            0,
+          );
+          overflow.forEach(([oldRunId]) => delete sessionRuns[oldRunId]);
+          localStorage.setItem(_COST_KEY, JSON.stringify(costs));
+        }
+        runCosts[sid] = sessionRuns;
+        localStorage.setItem(_COST_RUNS_KEY, JSON.stringify(runCosts));
+      } catch (_e) { /* ignore */ }
+    } else {
+      try {
+        const costs = JSON.parse(localStorage.getItem(_COST_KEY) || '{}');
+        costs[sid] = (costs[sid] || 0) + cost;
+        localStorage.setItem(_COST_KEY, JSON.stringify(costs));
+      } catch (_e) { /* ignore */ }
+    }
+    metrics._costRecorded = true;
+    metrics._costRecordPending = false;
+    const currentSid = window.sessionModule && window.sessionModule.getCurrentSessionId();
+    if (currentSid === sid) updateSessionCostUI();
+  };
+
+  let writeStarted = false;
+  const guardedWrite = () => {
+    writeStarted = true;
+    writeCost();
+  };
+  try {
+    if (
+      typeof navigator !== 'undefined'
+      && navigator.locks
+      && typeof navigator.locks.request === 'function'
+    ) {
+      const pendingWrite = navigator.locks.request(_COST_LEDGER_LOCK, guardedWrite);
+      if (pendingWrite && typeof pendingWrite.catch === 'function') {
+        pendingWrite.catch(() => {
+          if (!writeStarted) guardedWrite();
+        });
+      }
+    } else {
+      guardedWrite();
+    }
+  } catch (_e) {
+    if (!writeStarted) guardedWrite();
+  }
+  return cost;
+}
+
 /** Create a timestamp span for role labels.
 * Pass an ISO string / Date / epoch-ms to render the message's own time
 * (used when replaying history). Falls back to "now" when no value is given. */
@@ -1874,23 +2040,19 @@ export function displayMetrics(messageElement, metrics) {
  const isReal = metrics.usage_source === 'real';
  const ctxPct = metrics.context_percent;
  const model = metrics.model || 'Unknown';
-  const cost = _billableCost(model, inputTokens, outputTokens);
+  const cost = _metricsBillableCost(
+    metrics,
+    model,
+    inputTokens,
+    outputTokens,
+  );

  // Nothing useful to show — bail out (only if ALL metrics are missing)
  if (!responseTime && !inputTokens && !outputTokens && tps == null && !ctxPct) return;

-  // Accumulate session cost (only on fresh metrics, not history reload)
-  if (!metrics._fromHistory) {
-    const _sid = window.sessionModule && window.sessionModule.getCurrentSessionId();
-    if (_sid && cost !== null) {
-      try {
-        const _costs = JSON.parse(localStorage.getItem(_COST_KEY) || '{}');
-        _costs[_sid] = (_costs[_sid] || 0) + cost;
-        localStorage.setItem(_COST_KEY, JSON.stringify(_costs));
-      } catch (_e) { /* ignore */ }
-      updateSessionCostUI();
-    }
-  }
+  // Rendering can occur when metrics arrive and again after [DONE]. The
+  // ledger mutation is idempotent for that shared payload.
+  recordSessionMetricsCost(metrics);

  // Keep token counts in the Message Stats popup; the footer should stay slim.
  const costStr0 = cost !== null ? `$${cost < 0.01 ? cost.toFixed(4) : cost.toFixed(3)}` : null;
@@ -2307,9 +2469,19 @@ export function addMessage(role, content, modelName, metadata) {
    const textRaw = Array.isArray(content) ? markdownModule.renderContent(content) : content;

    // --- Agent multi-bubble reconstruction from saved metadata ---
-    if (role === 'assistant' && metadata && metadata.tool_events && metadata.tool_events.length > 0) {
+    if (
+      role === 'assistant'
+      && metadata
+      && (
+        (Array.isArray(metadata.tool_events) && metadata.tool_events.length > 0)
+        || (Array.isArray(metadata.round_texts) && metadata.round_texts.length > 1)
+      )
+    ) {
      const roundTexts = metadata.round_texts || [];
-      const toolEvents = metadata.tool_events;
+      const roundModels = metadata.round_models || [];
+      const roundEndpointIds = metadata.round_endpoint_ids || [];
+      const roundEndpointLabels = metadata.round_endpoint_labels || [];
+      const toolEvents = metadata.tool_events || [];
      let pendingAskUser = null;
      let lastWrap = null;
      let firstMsgAi = null;
@@ -2322,7 +2494,8 @@ export function addMessage(role, content, modelName, metadata) {
        toolsByRound[r].push(ev);
      }

-      const maxRound = Math.max(...Object.keys(toolsByRound).map(Number), roundTexts.length);
+      const toolRounds = Object.keys(toolsByRound).map(Number);
+      const maxRound = Math.max(toolRounds.length ? Math.max(...toolRounds) : 0, roundTexts.length);

      for (let r = 0; r < maxRound; r++) {
        const roundNum = r + 1;
@@ -2334,10 +2507,31 @@ export function addMessage(role, content, modelName, metadata) {
          const roleEl = document.createElement('div');
          roleEl.className = 'role';
          const pair = replyModelPair(modelName, metadata);
-          const contModel = pair.actualModel || pair.requestedModel;
-          roleEl.textContent = modelRouteLabel(pair.requestedModel, contModel);
-          if (pair.requestedModel && contModel && !sameModelName(pair.requestedModel, contModel)) {
-            roleEl.title = pair.requestedModel + ' -> ' + contModel;
+          const contModel = roundModels[r] || pair.actualModel || pair.requestedModel;
+          const contEndpointId = r < roundEndpointIds.length
+            ? roundEndpointIds[r]
+            : pair.actualEndpointId;
+          const contEndpointLabel = r < roundEndpointLabels.length
+            ? roundEndpointLabels[r]
+            : pair.actualEndpointLabel;
+          roleEl.textContent = modelRouteLabel(
+            pair.requestedModel,
+            contModel,
+            pair.requestedEndpointLabel,
+            contEndpointLabel,
+            pair.requestedEndpointId,
+            contEndpointId,
+          );
+          if (
+            pair.requestedModel
+            && contModel
+            && (
+              !sameModelName(pair.requestedModel, contModel)
+              || (pair.requestedEndpointId && contEndpointId && pair.requestedEndpointId !== contEndpointId)
+            )
+          ) {
+            roleEl.title = pair.requestedModel + ' -> ' + contModel
+              + ' (' + pair.requestedEndpointLabel + ' -> ' + contEndpointLabel + ')';
          }
          applyModelColor(roleEl, contModel);
          if (r === 0) roleEl.appendChild(roleTimestamp(metadata?.timestamp));
@@ -2492,7 +2686,14 @@ export function addMessage(role, content, modelName, metadata) {
    const isCompacted = metadata?.compacted;
    const replyModels = replyModelPair(modelName, metadata);
    const resolvedModel = replyModels.actualModel || replyModels.requestedModel;
-    var _roleText = role === 'user' ? 'You' : (isSlash || isCompacted) ? 'Odysseus' : modelRouteLabel(replyModels.requestedModel, resolvedModel);
+    var _roleText = role === 'user' ? 'You' : (isSlash || isCompacted) ? 'Odysseus' : modelRouteLabel(
+      replyModels.requestedModel,
+      resolvedModel,
+      replyModels.requestedEndpointLabel,
+      replyModels.actualEndpointLabel,
+      replyModels.requestedEndpointId,
+      replyModels.actualEndpointId,
+    );
    if (role === 'assistant' && (metadata?.research || metadata?.research_clarification)) {
      _roleText += ' (Research)';
    }
@@ -2503,8 +2704,14 @@ export function addMessage(role, content, modelName, metadata) {
    }
    r.textContent = _roleText;
    if (role !== 'user') {
-      if (!isSlash && !isCompacted && replyModels.requestedModel && resolvedModel && !sameModelName(replyModels.requestedModel, resolvedModel)) {
-        r.title = replyModels.requestedModel + ' -> ' + resolvedModel;
+      const endpointChanged = Boolean(
+        replyModels.requestedEndpointId
+        && replyModels.actualEndpointId
+        && replyModels.requestedEndpointId !== replyModels.actualEndpointId
+      );
+      if (!isSlash && !isCompacted && replyModels.requestedModel && resolvedModel && (!sameModelName(replyModels.requestedModel, resolvedModel) || endpointChanged)) {
+        r.title = replyModels.requestedModel + ' -> ' + resolvedModel
+          + ' (' + replyModels.requestedEndpointLabel + ' -> ' + replyModels.actualEndpointLabel + ')';
      }
      if (!isSlash && !isCompacted) applyModelColor(r, resolvedModel);
      r.appendChild(roleTimestamp(metadata?.timestamp));
@@ -2788,6 +2995,7 @@ const chatRenderer = {
  getSessionCost,
  resetSessionCost,
  updateSessionCostUI,
+  recordSessionMetricsCost,
  roleTimestamp,
  stripToolBlocks,
  copyMessageText,