orionfold-relay 0.22.0 → 0.22.1

package/dist/cli.js

@@ -6910,6 +6910,34 @@ var init_output_scanner = __esm({
   }
 });
 
+// src/lib/agents/runtime/model-preference.ts
+var model_preference_exports = {};
+__export(model_preference_exports, {
+  resolvePreferredModel: () => resolvePreferredModel
+});
+async function resolvePreferredModel(runtimeId, options) {
+  if (options?.pinnedModelId) {
+    return { modelId: options.pinnedModelId, source: "pin" };
+  }
+  const models = getRuntimeCatalogEntry(runtimeId).models;
+  const preference = await getModelPreference();
+  const tierModel = preference === "balanced" ? models.tiers?.balanced : preference === "cost" ? models.tiers?.fast : preference === "quality" ? models.tiers?.quality : void 0;
+  if (tierModel) {
+    return { modelId: tierModel, source: "preference" };
+  }
+  return {
+    modelId: models.tiers?.quality ?? models.default,
+    source: "default"
+  };
+}
+var init_model_preference = __esm({
+  "src/lib/agents/runtime/model-preference.ts"() {
+    "use strict";
+    init_catalog2();
+    init_helpers();
+  }
+});
+
 // src/lib/agents/runtime/claude-sdk.ts
 function buildClaudeSdkEnv(authEnv) {
   const { CLAUDECODE, ANTHROPIC_API_KEY, ...cleanEnv } = process.env;
@@ -7679,6 +7707,9 @@ async function deriveUsageCostMicros(input) {
   if (!input.modelId) {
     return { costMicros: null, pricingVersion: null };
   }
+  if (input.providerId === "ollama") {
+    return { costMicros: 0, pricingVersion: "local-free" };
+  }
   if (input.providerId !== "anthropic" && input.providerId !== "openai") {
     return { costMicros: null, pricingVersion: null };
   }
@@ -18074,13 +18105,17 @@ ${learnedCtx}
 You are operating inside a git worktree (branch: ${ws.gitBranch ?? "unknown"}). All file operations MUST use paths relative to the working directory: ${cwd}. Do NOT navigate to or create files in the main repository directory.` : "";
   const systemInstructions = [worktreeNote, profileInstructions, learnedCtxBlock, docContext, tableContext, outputInstructions].filter(Boolean).join("\n\n");
   const maxTurns = profile?.maxTurns ?? DEFAULT_MAX_TURNS;
+  const { modelId } = await resolvePreferredModel("claude-code", {
+    pinnedModelId: profile?.capabilityOverrides?.["claude-code"]?.modelId
+  });
   return {
     userPrompt: basePrompt,
     systemInstructions,
     cwd,
     payload,
     maxTurns,
-    canUseToolPolicy: payload?.canUseToolPolicy
+    canUseToolPolicy: payload?.canUseToolPolicy,
+    modelId
   };
 }
 async function executeClaudeTask(taskId) {
@@ -18126,6 +18161,10 @@ async function executeClaudeTask(taskId) {
       prompt: ctx.userPrompt,
       options: {
         abortController,
+        // Explicit model: profile pin > onboarding preference > quality
+        // default. Omitting this let the SDK pick ITS default (Opus) and
+        // silently bill the wrong tier.
+        model: ctx.modelId,
         includePartialMessages: true,
         cwd: ctx.cwd,
         env: buildClaudeSdkEnv(authEnv),
@@ -18248,6 +18287,9 @@ async function resumeClaudeTask(taskId) {
       options: {
         resume: task.sessionId,
         abortController,
+        // Same model resolution as the original run — a resume must not
+        // silently hop tiers (profile pin > preference > quality default).
+        model: ctx.modelId,
         includePartialMessages: true,
         cwd: ctx.cwd,
         env: buildClaudeSdkEnv(authEnv),
@@ -18366,6 +18408,7 @@ var init_claude_agent = __esm({
     init_context_builder2();
     init_output_scanner();
     init_registry2();
+    init_model_preference();
     init_compatibility();
     init_claude_sdk();
     init_types2();
@@ -18384,9 +18427,8 @@ var init_claude_agent = __esm({
 // src/lib/agents/runtime/claude.ts
 import { query as query2 } from "@anthropic-ai/claude-agent-sdk";
 import { eq as eq34 } from "drizzle-orm";
-function claudeCodeModelAlias() {
-  const models = getRuntimeCatalogEntry("claude-code").models;
-  return models.tiers?.quality ?? models.default;
+async function claudeCodeModelAlias() {
+  return (await resolvePreferredModel("claude-code")).modelId;
 }
 function buildTaskAssistSystemPrompt(profileIds) {
   const profileList = profileIds.length > 0 ? `Available agent profiles: ${profileIds.join(", ")}
@@ -18480,7 +18522,7 @@ Provide a brief analysis (2-3 paragraphs max). Include specific terminology rele
       prompt,
       options: {
         abortController,
-        model: claudeCodeModelAlias(),
+        model: await claudeCodeModelAlias(),
         includePartialMessages: true,
         env: buildClaudeSdkEnv(authEnv),
         allowedTools: []
@@ -18621,7 +18663,7 @@ async function runMetaCompletion(input) {
       prompt: input.prompt,
       options: {
         abortController,
-        model: claudeCodeModelAlias(),
+        model: await claudeCodeModelAlias(),
         includePartialMessages: true,
         cwd: getLaunchCwd(),
         env: buildClaudeSdkEnv(authEnv),
@@ -18801,7 +18843,7 @@ ${userMessage}`;
       prompt,
       options: {
         abortController,
-        model: claudeCodeModelAlias(),
+        model: await claudeCodeModelAlias(),
         includePartialMessages: true,
         cwd: getLaunchCwd(),
         env: buildClaudeSdkEnv(authEnv),
@@ -18876,7 +18918,7 @@ ${userMessage}`;
       prompt,
       options: {
         abortController,
-        model: claudeCodeModelAlias(),
+        model: await claudeCodeModelAlias(),
         includePartialMessages: true,
         cwd: getLaunchCwd(),
         env: buildClaudeSdkEnv(authEnv),
@@ -18941,7 +18983,7 @@ async function testClaudeConnection() {
       prompt: "Reply with exactly: OK",
       options: {
         abortController,
-        model: claudeCodeModelAlias(),
+        model: await claudeCodeModelAlias(),
         maxTurns: 1,
         includePartialMessages: false,
         cwd: getLaunchCwd(),
@@ -18988,6 +19030,7 @@ var init_claude = __esm({
     init_compatibility();
     init_claude_agent();
     init_catalog2();
+    init_model_preference();
     init_claude_sdk();
     init_workspace_context();
     init_helpers();
@@ -20882,7 +20925,8 @@ ${outputInstructions}`;
       initialMessages = [{ role: "user", content: ctx.userPrompt }];
     }
     const { getSetting: getSetting2 } = await Promise.resolve().then(() => (init_helpers(), helpers_exports));
-    const modelId = await getSetting2("anthropic_direct_model") ?? getRuntimeCatalogEntry("anthropic-direct").models.default;
+    const { resolvePreferredModel: resolvePreferredModel2 } = await Promise.resolve().then(() => (init_model_preference(), model_preference_exports));
+    const modelId = await getSetting2("anthropic_direct_model") ?? (await resolvePreferredModel2("anthropic-direct")).modelId;
     const maxTurns = ctx.maxTurns ?? DEFAULT_MAX_TURNS;
     await db.insert(agentLogs).values({
       id: crypto.randomUUID(),
@@ -21353,7 +21397,8 @@ ${outputInstructions}`;
     );
     const pluginMcpTools = mcpServersToOpenAiTools(mergedMcpServers);
     const { getSetting: getSetting2 } = await Promise.resolve().then(() => (init_helpers(), helpers_exports));
-    const modelId = await getSetting2("openai_direct_model") ?? getRuntimeCatalogEntry("openai-direct").models.default;
+    const { resolvePreferredModel: resolvePreferredModel2 } = await Promise.resolve().then(() => (init_model_preference(), model_preference_exports));
+    const modelId = await getSetting2("openai_direct_model") ?? (await resolvePreferredModel2("openai-direct")).modelId;
     const maxTurns = ctx.maxTurns ?? DEFAULT_MAX_TURNS;
     let previousResponseId = null;
     if (isResume) {
@@ -22193,6 +22238,17 @@ async function getUsageAggregates(policy, runtimeStates, now = /* @__PURE__ */ n
       runtimes[runtimeId].daily.totalTokens += row.totalTokens ?? 0;
     }
   });
+  const metered = {
+    daily: { costMicros: 0, totalTokens: 0 },
+    monthly: { costMicros: 0, totalTokens: 0 }
+  };
+  for (const runtimeId of SUPPORTED_AGENT_RUNTIMES) {
+    metered.daily.costMicros += runtimes[runtimeId].daily.costMicros;
+    metered.daily.totalTokens += runtimes[runtimeId].daily.totalTokens;
+    metered.monthly.costMicros += runtimes[runtimeId].monthly.costMicros;
+    metered.monthly.totalTokens += runtimes[runtimeId].monthly.totalTokens;
+  }
+  let planPricedMonthlyMicros = null;
   if (runtimeStates["claude-code"].billingMode === "subscription") {
     const planPriceUsd = await getClaudeOAuthPlanPrice(
       policy.runtimes["claude-code"].claudeOAuthPlan
@@ -22201,6 +22257,7 @@ async function getUsageAggregates(policy, runtimeStates, now = /* @__PURE__ */ n
     const dailyMicros = Math.round(monthlyMicros / daysInMonth(now));
     runtimes["claude-code"].monthly.costMicros = monthlyMicros;
     runtimes["claude-code"].daily.costMicros = dailyMicros;
+    planPricedMonthlyMicros = monthlyMicros;
   }
   const overall = {
     daily: { costMicros: 0, totalTokens: 0 },
@@ -22218,6 +22275,8 @@ async function getUsageAggregates(policy, runtimeStates, now = /* @__PURE__ */ n
   return {
     overall,
     runtimes,
+    metered,
+    planPricedMonthlyMicros,
     ...getBudgetWindowBounds(now)
   };
 }
@@ -25681,8 +25740,8 @@ import { execFileSync as execFileSync3 } from "child_process";
 import yaml12 from "js-yaml";
 import semver from "semver";
 function relayCoreVersion() {
-  if (semver.valid("0.22.0")) {
-    return "0.22.0";
+  if (semver.valid("0.22.1")) {
+    return "0.22.1";
   }
   try {
     const root = getAppRoot(import.meta.dirname, 3);

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "orionfold-relay",
-  "version": "0.22.0",
+  "version": "0.22.1",
   "description": "Orionfold Relay — a local-first, multi-agent orchestration runtime and builder scaffold for AI-native work.",
   "keywords": [
     "ai",

package/src/app/api/telemetry/route.ts

@@ -175,9 +175,6 @@ export async function GET() {
       getFailuresByDay(7),
     ]);
 
-    const overallDaily = budget.statuses.find(
-      (s) => s.scopeId === "overall" && s.window === "daily",
-    );
     const overallMonthly = budget.statuses.find(
       (s) => s.scopeId === "overall" && s.window === "monthly",
     );
@@ -196,8 +193,12 @@ export async function GET() {
       activeProjects: activeProjects?.count ?? 0,
       activeWorkflows: activeWorkflows?.count ?? 0,
       reviewPending,
-      costTodayMicros: overallDaily?.currentValue ?? 0,
-      costToDateMicros: overallMonthly?.currentValue ?? 0,
+      // Metered ledger sums only — the guardrail statuses' plan-priced budget
+      // basis must never render as "cost" (fix-dashboard-budget-vs-cost-labeling).
+      costTodayMicros: budget.meteredSpend.dailyMicros,
+      costToDateMicros: budget.meteredSpend.monthlyMicros,
+      budgetMonthlyCapMicros: overallMonthly?.limitValue ?? null,
+      planPricedMonthlyMicros: budget.planPricedMonthlyMicros,
       runtimeLabel,
       providerId,
       runtimeSdkVersion,

package/src/components/shell/telemetry-rail.tsx

@@ -20,16 +20,19 @@ import {
 } from "lucide-react";
 import { RailCell, formatMicros } from "./rail-cell";
 import { useTelemetry } from "./use-telemetry";
+import type { TelemetrySnapshot } from "./telemetry-types";
 
 // The standing instrument cluster: a single dense horizontal row beneath the app
 // bar (mirrors `.hp-rail`). A cockpit for a multi-agent harness — eight real
 // cells: HOST (folder · cpu/mem) · RUNTIME (label · sdk version) · TASKS
 // (running + 24h activity spark) · THROUGHPUT (completed today + 7d spark) ·
-// FAILURES (failed + 7d spark, red) · REVIEW (pending) · COST TODAY · COST TO
+// FAILURES (failed + 7d spark, red) · REVIEW (pending) · SPEND TODAY · SPEND TO
 // DATE — plus a live/error status foot. No fabricated data: while loading, cells
 // show "—"; on a poll error the last good snapshot stays visible and the foot
 // flips to an explicit error pip. Static identity (cwd/runtime) is compressed
-// into sub-lines so the live throughput signal owns the foreground.
+// into sub-lines so the live throughput signal owns the foreground. The SPEND
+// cells render real metered ledger sums; the budget cap and any flat plan price
+// live in the sub-line, named as what they are — never presented as spend.
 
 // Compose the HOST sub-line from whatever live metrics the platform reports;
 // falls back to git branch so the cell is never empty.
@@ -45,6 +48,20 @@ function hostSub(
   return branch ? `git:${branch}` : "no git";
 }
 
+// Sub-line for SPEND TO DATE: name the flat plan price when billing is
+// subscription (it sits on top of metered spend, it is not spend), else show
+// the monthly budget cap as "budget", else the plain window label.
+function spendToDateSub(data: TelemetrySnapshot | null): string {
+  if (!data) return "—";
+  if (data.planPricedMonthlyMicros != null) {
+    return `+ plan ${formatMicros(data.planPricedMonthlyMicros)}/mo`;
+  }
+  if (data.budgetMonthlyCapMicros != null) {
+    return `of ${formatMicros(data.budgetMonthlyCapMicros)} budget`;
+  }
+  return "monthly";
+}
+
 export function TelemetryRail() {
   const telemetry = useTelemetry();
   const data = telemetry.data;
@@ -71,9 +88,13 @@ export function TelemetryRail() {
         loading={loading}
         value={data?.runtimeLabel ?? "—"}
         sub={
-          data?.runtimeSdkVersion
-            ? `sdk ${data.runtimeSdkVersion}`
-            : data?.providerId ?? "not configured"
+          // Never fabricate "not configured" while the snapshot is still
+          // loading — only claim it once data has actually said so.
+          data
+            ? data.runtimeSdkVersion
+              ? `sdk ${data.runtimeSdkVersion}`
+              : data.providerId ?? "not configured"
+            : "—"
         }
       />
       <RailCell
@@ -129,18 +150,18 @@ export function TelemetryRail() {
         sub="active"
       />
       <RailCell
-        label="Cost Today"
+        label="Spend Today"
         icon={<Coins aria-hidden />}
         loading={loading}
         value={data ? formatMicros(data.costTodayMicros) : "—"}
-        sub="daily"
+        sub="metered"
       />
       <RailCell
-        label="Cost To Date"
+        label="Spend To Date"
         icon={<Wallet aria-hidden />}
         loading={loading}
         value={data ? formatMicros(data.costToDateMicros) : "—"}
-        sub="monthly"
+        sub={spendToDateSub(data)}
       />
       <div className="ml-auto flex items-center gap-2 px-4 font-mono text-xs text-muted-foreground/60">
         {errored ? (

package/src/components/shell/telemetry-types.ts

@@ -16,10 +16,14 @@ export interface TelemetrySnapshot {
   activeWorkflows: number;
   /** Unread `permission_required` notifications awaiting a human decision. */
   reviewPending: number;
-  /** Overall daily spend in micros (USD * 1e6). */
+  /** Real metered spend today in micros (usage_ledger sum, USD * 1e6) — never a budget/plan figure. */
   costTodayMicros: number;
-  /** Overall monthly (to-date) spend in micros (USD * 1e6). */
+  /** Real metered spend this month in micros (usage_ledger sum, USD * 1e6) — never a budget/plan figure. */
   costToDateMicros: number;
+  /** Overall monthly budget cap in micros, or null when unlimited. */
+  budgetMonthlyCapMicros: number | null;
+  /** Flat subscription plan price in micros when billing is subscription (not metered spend), else null. */
+  planPricedMonthlyMicros: number | null;
   /** Display label of the active runtime (e.g. "Claude Code"), or null if none configured. */
   runtimeLabel: string | null;
   /** Provider behind the active runtime ("anthropic" | "openai" | "ollama"), or null. */

package/src/lib/agents/claude-agent.ts

@@ -13,6 +13,7 @@ import {
   scanTaskOutputDocuments,
 } from "@/lib/documents/output-scanner";
 import { getProfile } from "./profiles/registry";
+import { resolvePreferredModel } from "./runtime/model-preference";
 import { resolveProfileRuntimePayload, type ResolvedProfileRuntimePayload } from "./profiles/compatibility";
 import type { CanUseToolPolicy } from "./profiles/types";
 import {
@@ -483,6 +484,13 @@ export interface TaskQueryContext {
   maxTurns: number;
   /** Profile's canUseToolPolicy */
   canUseToolPolicy?: CanUseToolPolicy;
+  /**
+   * Concrete model to pass to `query()`: profile pin > onboarding model
+   * preference tier > quality default. Without an explicit model the SDK
+   * silently falls back to ITS default — which billed Opus to users who chose
+   * "Balanced" (fix-workflow-model-preference-propagation).
+   */
+  modelId: string;
 }
 
 export async function buildTaskQueryContext(
@@ -533,6 +541,10 @@ export async function buildTaskQueryContext(
   // F9: Use profile maxTurns or fall back to default
   const maxTurns = profile?.maxTurns ?? DEFAULT_MAX_TURNS;
 
+  const { modelId } = await resolvePreferredModel("claude-code", {
+    pinnedModelId: profile?.capabilityOverrides?.["claude-code"]?.modelId,
+  });
+
   return {
     userPrompt: basePrompt,
     systemInstructions,
@@ -540,6 +552,7 @@ export async function buildTaskQueryContext(
     payload,
     maxTurns,
     canUseToolPolicy: payload?.canUseToolPolicy,
+    modelId,
   };
 }
 
@@ -612,6 +625,10 @@ export async function executeClaudeTask(taskId: string): Promise<void> {
       prompt: ctx.userPrompt,
       options: {
         abortController,
+        // Explicit model: profile pin > onboarding preference > quality
+        // default. Omitting this let the SDK pick ITS default (Opus) and
+        // silently bill the wrong tier.
+        model: ctx.modelId,
         includePartialMessages: true,
         cwd: ctx.cwd,
         env: buildClaudeSdkEnv(authEnv),
@@ -771,6 +788,9 @@ export async function resumeClaudeTask(taskId: string): Promise<void> {
       options: {
         resume: task.sessionId,
         abortController,
+        // Same model resolution as the original run — a resume must not
+        // silently hop tiers (profile pin > preference > quality default).
+        model: ctx.modelId,
         includePartialMessages: true,
         cwd: ctx.cwd,
         env: buildClaudeSdkEnv(authEnv),

package/src/lib/agents/runtime/anthropic-direct.ts

@@ -376,9 +376,14 @@ async function executeAnthropicDirectTask(taskId: string, isResume = false): Pro
       initialMessages = [{ role: "user", content: ctx.userPrompt }];
     }
 
-    // Resolve model from settings
+    // Resolve model: explicit runtime setting > onboarding model preference
+    // tier > catalog default ("Balanced" means Sonnet on every runtime, not
+    // just chat — fix-workflow-model-preference-propagation).
     const { getSetting } = await import("@/lib/settings/helpers");
-    const modelId = (await getSetting("anthropic_direct_model")) ?? getRuntimeCatalogEntry("anthropic-direct").models.default;
+    const { resolvePreferredModel } = await import("./model-preference");
+    const modelId =
+      (await getSetting("anthropic_direct_model")) ??
+      (await resolvePreferredModel("anthropic-direct")).modelId;
 
     const maxTurns = ctx.maxTurns ?? DEFAULT_MAX_TURNS;
 

package/src/lib/agents/runtime/claude.ts

@@ -8,6 +8,7 @@ import { getProfile, listProfiles } from "@/lib/agents/profiles/registry";
 import { resolveProfileRuntimePayload } from "@/lib/agents/profiles/compatibility";
 import { executeClaudeTask, resumeClaudeTask } from "@/lib/agents/claude-agent";
 import { getRuntimeCapabilities, getRuntimeCatalogEntry } from "./catalog";
+import { resolvePreferredModel } from "./model-preference";
 import { buildClaudeSdkEnv } from "./claude-sdk";
 import { getLaunchCwd } from "@/lib/environment/workspace-context";
 import { getSetting } from "@/lib/settings/helpers";
@@ -34,10 +35,13 @@ import {
  * model in that family instead of pinning to a string that silently ages out.
  * Without this, `query()` omits `model` entirely and the SDK falls back to its
  * own default — which is not necessarily the family the chat picker selected.
+ *
+ * Honors the user's onboarding model preference ("Balanced" means Sonnet
+ * everywhere, not just chat); falls back to the quality tier when no
+ * preference is recorded.
  */
-function claudeCodeModelAlias(): string {
-  const models = getRuntimeCatalogEntry("claude-code").models;
-  return models.tiers?.quality ?? models.default;
+async function claudeCodeModelAlias(): Promise<string> {
+  return (await resolvePreferredModel("claude-code")).modelId;
 }
 
 function buildTaskAssistSystemPrompt(profileIds: string[]): string {
@@ -148,7 +152,7 @@ export async function runSingleProfileTest(
       prompt,
       options: {
         abortController,
-        model: claudeCodeModelAlias(),
+        model: await claudeCodeModelAlias(),
         includePartialMessages: true,
         env: buildClaudeSdkEnv(authEnv),
         allowedTools: [],
@@ -306,7 +310,7 @@ export async function runMetaCompletion(input: {
       prompt: input.prompt,
       options: {
         abortController,
-        model: claudeCodeModelAlias(),
+        model: await claudeCodeModelAlias(),
         includePartialMessages: true,
         cwd: getLaunchCwd(),
         env: buildClaudeSdkEnv(authEnv),
@@ -497,7 +501,7 @@ async function runClaudeProfileAssist(
       prompt,
       options: {
         abortController,
-        model: claudeCodeModelAlias(),
+        model: await claudeCodeModelAlias(),
         includePartialMessages: true,
         cwd: getLaunchCwd(),
         env: buildClaudeSdkEnv(authEnv),
@@ -584,7 +588,7 @@ async function runClaudeTaskAssist(
       prompt,
       options: {
         abortController,
-        model: claudeCodeModelAlias(),
+        model: await claudeCodeModelAlias(),
         includePartialMessages: true,
         cwd: getLaunchCwd(),
         env: buildClaudeSdkEnv(authEnv),
@@ -655,7 +659,7 @@ async function testClaudeConnection(): Promise<RuntimeConnectionResult> {
       prompt: "Reply with exactly: OK",
       options: {
         abortController,
-        model: claudeCodeModelAlias(),
+        model: await claudeCodeModelAlias(),
         maxTurns: 1,
         includePartialMessages: false,
         cwd: getLaunchCwd(),

package/src/lib/agents/runtime/model-preference.ts

@@ -0,0 +1,54 @@
+import { getRuntimeCatalogEntry, type AgentRuntimeId } from "./catalog";
+import { getModelPreference } from "@/lib/settings/helpers";
+
+/**
+ * Resolve the concrete model a claude-code (or any catalog) execution should
+ * run on, honoring the user's onboarding model preference
+ * (`chat.modelPreference`) outside chat. Order:
+ *
+ *   1. an explicit profile/step pin (`pinnedModelId`) — the user said exactly
+ *      what to run, the preference never overrides it;
+ *   2. the preference mapped to the runtime's tier ("balanced" → Sonnet,
+ *      "cost" → fast/Haiku, "quality" → Opus);
+ *   3. the runtime's quality tier / catalog default — the pre-existing
+ *      behavior when no preference is recorded.
+ *
+ * "privacy" is a runtime-level preference (route to Ollama), not a model tier
+ * within a cloud runtime — it resolves as "no within-runtime opinion".
+ *
+ * The `source` field keeps the routing inspectable: surfaces can say WHY a
+ * model was chosen instead of silently swapping tiers.
+ */
+export interface ResolvedPreferredModel {
+  modelId: string;
+  source: "pin" | "preference" | "default";
+}
+
+export async function resolvePreferredModel(
+  runtimeId: AgentRuntimeId,
+  options?: { pinnedModelId?: string | null },
+): Promise<ResolvedPreferredModel> {
+  if (options?.pinnedModelId) {
+    return { modelId: options.pinnedModelId, source: "pin" };
+  }
+
+  const models = getRuntimeCatalogEntry(runtimeId).models;
+  const preference = await getModelPreference();
+  const tierModel =
+    preference === "balanced"
+      ? models.tiers?.balanced
+      : preference === "cost"
+        ? models.tiers?.fast
+        : preference === "quality"
+          ? models.tiers?.quality
+          : undefined;
+
+  if (tierModel) {
+    return { modelId: tierModel, source: "preference" };
+  }
+
+  return {
+    modelId: models.tiers?.quality ?? models.default,
+    source: "default",
+  };
+}

package/src/lib/agents/runtime/openai-direct.ts

@@ -319,9 +319,14 @@ async function executeOpenAIDirectTask(taskId: string, isResume = false): Promis
     );
     const pluginMcpTools = mcpServersToOpenAiTools(mergedMcpServers);
 
-    // Resolve model
+    // Resolve model: explicit runtime setting > onboarding model preference
+    // tier > catalog default ("Balanced" means the balanced tier on every
+    // runtime, not just chat — fix-workflow-model-preference-propagation).
     const { getSetting } = await import("@/lib/settings/helpers");
-    const modelId = (await getSetting("openai_direct_model")) ?? getRuntimeCatalogEntry("openai-direct").models.default;
+    const { resolvePreferredModel } = await import("./model-preference");
+    const modelId =
+      (await getSetting("openai_direct_model")) ??
+      (await resolvePreferredModel("openai-direct")).modelId;
     const maxTurns = ctx.maxTurns ?? DEFAULT_MAX_TURNS;
 
     // For resume: load previous response ID

package/src/lib/chat/ollama-engine.ts

@@ -18,6 +18,7 @@ import {
 } from "@/lib/data/chat";
 import { buildChatContext } from "./context-builder";
 import { getWorkspaceContext } from "@/lib/environment/workspace-context";
+import { recordUsageLedgerEntry } from "@/lib/usage/ledger";
 import type { ChatStreamEvent } from "./types";
 
 /**
@@ -109,6 +110,47 @@ export async function* sendOllamaMessage(
 
   // Stream from Ollama
   let accumulated = "";
+
+  // Meter the turn like every other chat path (main engine writes a
+  // chat_turn row on success, degrade, and error). Ollama's final chunk
+  // reports prompt_eval_count / eval_count; local runs are recorded at $0 —
+  // those rows are what proves blended-cost savings on /costs.
+  const startedAt = new Date();
+  let inputTokens: number | null = null;
+  let outputTokens: number | null = null;
+  let ledgerRecorded = false;
+  const recordTurn = async (status: "completed" | "failed" | "cancelled") => {
+    if (ledgerRecorded) return;
+    ledgerRecorded = true;
+    await recordUsageLedgerEntry({
+      projectId: conversation.projectId ?? null,
+      activityType: "chat_turn",
+      runtimeId: "ollama",
+      providerId: "ollama",
+      modelId,
+      inputTokens,
+      outputTokens,
+      totalTokens:
+        inputTokens != null && outputTokens != null
+          ? inputTokens + outputTokens
+          : null,
+      status,
+      startedAt,
+      finishedAt: new Date(),
+    });
+  };
+  const captureTokenCounts = (parsed: {
+    prompt_eval_count?: unknown;
+    eval_count?: unknown;
+  }) => {
+    if (typeof parsed.prompt_eval_count === "number") {
+      inputTokens = parsed.prompt_eval_count;
+    }
+    if (typeof parsed.eval_count === "number") {
+      outputTokens = parsed.eval_count;
+    }
+  };
+
   try {
     const response = await fetch(`${baseUrl}/api/chat`, {
       method: "POST",
@@ -125,6 +167,7 @@ export async function* sendOllamaMessage(
       const errorText = await response.text().catch(() => "Unknown error");
       yield { type: "error", message: `Ollama error (${response.status}): ${errorText}` };
       await updateMessageStatus(assistantMsg.id, "complete");
+      await recordTurn("failed");
       return;
     }
 
@@ -132,6 +175,7 @@ export async function* sendOllamaMessage(
     if (!reader) {
       yield { type: "error", message: "No response stream from Ollama" };
       await updateMessageStatus(assistantMsg.id, "complete");
+      await recordTurn("failed");
       return;
     }
 
@@ -157,7 +201,10 @@ export async function* sendOllamaMessage(
             accumulated += delta;
             yield { type: "delta", content: delta };
           }
-          if (parsed.done) break;
+          if (parsed.done) {
+            captureTokenCounts(parsed);
+            break;
+          }
         } catch {
           // Skip malformed lines
         }
@@ -173,6 +220,7 @@ export async function* sendOllamaMessage(
           accumulated += delta;
           yield { type: "delta", content: delta };
         }
+        if (parsed.done) captureTokenCounts(parsed);
       } catch {
         // ignore
       }
@@ -181,6 +229,7 @@ export async function* sendOllamaMessage(
     // Persist the complete response
     await updateMessageContent(assistantMsg.id, accumulated);
     await updateMessageStatus(assistantMsg.id, "complete");
+    await recordTurn("completed");
 
     yield { type: "done", messageId: assistantMsg.id, quickAccess: [] };
   } catch (err) {
@@ -194,5 +243,6 @@ export async function* sendOllamaMessage(
       await updateMessageContent(assistantMsg.id, accumulated);
     }
     await updateMessageStatus(assistantMsg.id, "complete");
+    await recordTurn(signal?.aborted ? "cancelled" : "failed");
   }
 }

package/src/lib/settings/budget-guardrails.ts

@@ -68,6 +68,10 @@ export interface BudgetSnapshot {
   monthlyResetAtIso: string;
   runtimeStates: Record<AgentRuntimeId, RuntimeSetupState>;
   pricing: PricingRegistrySnapshot;
+  /** Real metered spend (usage_ledger sums) — never the plan-priced budget basis. */
+  meteredSpend: { dailyMicros: number; monthlyMicros: number };
+  /** Flat subscription price counted as the budget basis, when billing is subscription. */
+  planPricedMonthlyMicros: number | null;
 }
 
 interface BudgetGuardInput {
@@ -358,6 +362,23 @@ async function getUsageAggregates(
     }
   });
 
+  // Real metered spend: the plain usage_ledger sums across every runtime,
+  // captured BEFORE the subscription plan-price substitution below. Guardrail
+  // statuses budget against the plan price (a flat subscription is the real
+  // monthly outlay), but display surfaces must never present that basis as
+  // spend — they read this instead.
+  const metered = {
+    daily: { costMicros: 0, totalTokens: 0 },
+    monthly: { costMicros: 0, totalTokens: 0 },
+  };
+  for (const runtimeId of SUPPORTED_AGENT_RUNTIMES) {
+    metered.daily.costMicros += runtimes[runtimeId].daily.costMicros;
+    metered.daily.totalTokens += runtimes[runtimeId].daily.totalTokens;
+    metered.monthly.costMicros += runtimes[runtimeId].monthly.costMicros;
+    metered.monthly.totalTokens += runtimes[runtimeId].monthly.totalTokens;
+  }
+
+  let planPricedMonthlyMicros: number | null = null;
   if (runtimeStates["claude-code"].billingMode === "subscription") {
     const planPriceUsd = await getClaudeOAuthPlanPrice(
       policy.runtimes["claude-code"].claudeOAuthPlan
@@ -366,6 +387,7 @@ async function getUsageAggregates(
     const dailyMicros = Math.round(monthlyMicros / daysInMonth(now));
     runtimes["claude-code"].monthly.costMicros = monthlyMicros;
     runtimes["claude-code"].daily.costMicros = dailyMicros;
+    planPricedMonthlyMicros = monthlyMicros;
   }
 
   const overall = {
@@ -387,6 +409,8 @@ async function getUsageAggregates(
   return {
     overall,
     runtimes,
+    metered,
+    planPricedMonthlyMicros,
     ...getBudgetWindowBounds(now),
   };
 }
@@ -714,5 +738,10 @@ export async function getBudgetGuardrailSnapshot(): Promise<BudgetSnapshot> {
     monthlyResetAtIso: aggregates.monthlyEnd.toISOString(),
     runtimeStates,
     pricing,
+    meteredSpend: {
+      dailyMicros: aggregates.metered.daily.costMicros,
+      monthlyMicros: aggregates.metered.monthly.costMicros,
+    },
+    planPricedMonthlyMicros: aggregates.planPricedMonthlyMicros,
   };
 }

package/src/lib/usage/pricing.ts

@@ -15,6 +15,13 @@ export async function deriveUsageCostMicros(input: {
     return { costMicros: null, pricingVersion: null };
   }
 
+  // Local inference has no per-token billing: a known-free $0, not unknown
+  // pricing. The $0 rows are the evidence for the blended-cost savings story —
+  // null here would demote every local run to "unknown_pricing" and hide them.
+  if (input.providerId === "ollama") {
+    return { costMicros: 0, pricingVersion: "local-free" };
+  }
+
   if (input.providerId !== "anthropic" && input.providerId !== "openai") {
     return { costMicros: null, pricingVersion: null };
   }