No usage recorded yet in this window.
) : null} + {unreportedTokens.length > 0 ? ( ++ Token counts were not reported for: {unreportedTokens.join(", ")}. + Those turns do not contribute a fabricated cost or token total. +
+ ) : null} + {missingRates.length > 0 ? (
Rates unknown for: {missingRates.join(", ")}. Those turns do not
diff --git a/packages/insights/src/client.test.ts b/packages/insights/src/client.test.ts
index 425e91a62..c0671bca8 100644
--- a/packages/insights/src/client.test.ts
+++ b/packages/insights/src/client.test.ts
@@ -7,6 +7,8 @@ import {
formatCount,
formatUsd,
INSIGHTS_WINDOW_DAYS,
+ modelsWithMissingRates,
+ modelsWithUnreportedTokens,
tokensLabel,
topModelsByCost,
usageChromeLabel,
@@ -165,3 +167,55 @@ describe("CL-6877 usage chrome labels", () => {
expect(usageChromeLabel(EMPTY_OVERALL_USAGE)).not.toContain("0 tok");
});
});
+
+describe("CL-6659 unreported tokens vs missing rates", () => {
+ const zeroTokens = {
+ input: 0,
+ cacheRead: 0,
+ cacheWrite: 0,
+ output: 0,
+ thinking: 0,
+ total: 0,
+ };
+
+ test("modelsWithUnreportedTokens names models that recorded turns but no token counts", () => {
+ expect(
+ modelsWithUnreportedTokens({
+ turns: 2,
+ tokens: zeroTokens,
+ costUsd: null,
+ byModel: [
+ {
+ model: "qwen3:latest",
+ turns: 2,
+ tokens: zeroTokens,
+ costUsd: null,
+ },
+ ],
+ }),
+ ).toEqual(["qwen3:latest"]);
+ expect(modelsWithUnreportedTokens(EMPTY_OVERALL_USAGE)).toEqual([]);
+ });
+
+ test("modelsWithMissingRates still requires tokens so unreported turns are not labeled as unknown rates", () => {
+ expect(
+ modelsWithMissingRates({
+ turns: 1,
+ tokens: zeroTokens,
+ costUsd: null,
+ byModel: [
+ {
+ model: "qwen3:latest",
+ turns: 1,
+ tokens: zeroTokens,
+ costUsd: null,
+ },
+ ],
+ }),
+ ).toEqual([]);
+ });
+
+ test("usageChromeLabel is an em-dash when cost is unknown and tokens were not reported", () => {
+ expect(usageChromeLabel({ costUsd: null, tokens: zeroTokens })).toBe("—");
+ });
+});
diff --git a/packages/insights/src/client.ts b/packages/insights/src/client.ts
index d830c10c2..1da3eeacb 100644
--- a/packages/insights/src/client.ts
+++ b/packages/insights/src/client.ts
@@ -221,6 +221,7 @@ export function usageChromeLabel(usage: {
}): string {
const cost = formatUsd(usage.costUsd);
const tok = tokensLabel(usage.tokens);
+ if (usage.costUsd === null && tok === undefined) return "—";
return tok === undefined ? cost : `${cost} · ${tok}`;
}
@@ -230,3 +231,12 @@ export function modelsWithMissingRates(usage: OverallUsage): readonly string[] {
.filter((m) => m.costUsd === null && m.tokens.total > 0)
.map((m) => m.model);
}
+
+/** Models that recorded turns but no token counts — not the same as $0 spend. */
+export function modelsWithUnreportedTokens(
+ usage: OverallUsage,
+): readonly string[] {
+ return usage.byModel
+ .filter((m) => m.turns > 0 && m.tokens.total === 0)
+ .map((m) => m.model);
+}
diff --git a/packages/insights/src/index.ts b/packages/insights/src/index.ts
index 852dee5d1..c136cb708 100644
--- a/packages/insights/src/index.ts
+++ b/packages/insights/src/index.ts
@@ -54,6 +54,7 @@ export {
summarizeLatency,
summarizeUsage,
summarizeUsageByTenant,
+ teamSpaceWorkbenchRows,
type DayActivity,
type LatencyStageStat,
type LatencySummary,
diff --git a/packages/insights/src/queries.test.ts b/packages/insights/src/queries.test.ts
index e8016aa42..4ee8ed719 100644
--- a/packages/insights/src/queries.test.ts
+++ b/packages/insights/src/queries.test.ts
@@ -6,6 +6,7 @@ import {
emptyOverallUsageSummary,
summarizeUsage,
summarizeUsageByTenant,
+ teamSpaceWorkbenchRows,
} from "./queries";
describe("summarizeUsage", () => {
@@ -432,3 +433,128 @@ describe("summarizeUsageByTenant", () => {
expect(summedCost).toBe(aggregate.costUsd ?? 0);
});
});
+
+describe("CL-6659 unreported tokens and team-space breakdowns", () => {
+ const zeroTokens = {
+ input: 0,
+ cacheRead: 0,
+ cacheWrite: 0,
+ output: 0,
+ thinking: 0,
+ };
+
+ test("a recorded turn with no token counts is null cost, not $0.00", async () => {
+ const store = createMemoryUsageStore();
+ await store.insertUsage({
+ id: "u1",
+ tenantId: "tenant-acme",
+ sessionId: "s1",
+ turnId: "t1",
+ provider: "ollama",
+ model: "qwen3:latest",
+ tokens: zeroTokens,
+ });
+
+ const summary = await summarizeUsage(store, ["tenant-acme"]);
+ expect(summary.turns).toBe(1);
+ expect(summary.tokens.total).toBe(0);
+ expect(summary.costUsd).toBeNull();
+ expect(summary.byModel[0]?.costUsd).toBeNull();
+ });
+
+ test("ollama turns with real tokens and a FREE catalog row cost $0, not null", async () => {
+ const store = createMemoryUsageStore();
+ await store.insertUsage({
+ id: "u1",
+ tenantId: "tenant-acme",
+ sessionId: "s1",
+ turnId: "t1",
+ provider: "ollama",
+ model: "qwen3.8:27b",
+ tokens: {
+ input: 100,
+ cacheRead: 0,
+ cacheWrite: 0,
+ output: 50,
+ thinking: 0,
+ },
+ });
+
+ const summary = await summarizeUsage(store, ["tenant-acme"]);
+ expect(summary.turns).toBe(1);
+ expect(summary.tokens.total).toBe(150);
+ expect(summary.costUsd).toBe(0);
+ });
+
+ test("team-space parent with recorded turns stays in the workbench breakdown", () => {
+ const rows = [
+ {
+ tenantId: "team-root",
+ turns: 3,
+ tokens: {
+ input: 0,
+ cacheRead: 0,
+ cacheWrite: 0,
+ output: 0,
+ thinking: 0,
+ total: 0,
+ },
+ costUsd: null,
+ },
+ {
+ tenantId: "workbench-a",
+ turns: 0,
+ tokens: {
+ input: 0,
+ cacheRead: 0,
+ cacheWrite: 0,
+ output: 0,
+ thinking: 0,
+ total: 0,
+ },
+ costUsd: 0,
+ },
+ ];
+ const kept = teamSpaceWorkbenchRows(rows, {
+ tenantId: "team-root",
+ isTeamSpace: true,
+ });
+ expect(kept.map((r) => r.tenantId)).toEqual(["team-root", "workbench-a"]);
+ });
+
+ test("empty team-space parent is still dropped as a duplicate of All workbenches", () => {
+ const rows = [
+ {
+ tenantId: "team-root",
+ turns: 0,
+ tokens: {
+ input: 0,
+ cacheRead: 0,
+ cacheWrite: 0,
+ output: 0,
+ thinking: 0,
+ total: 0,
+ },
+ costUsd: 0,
+ },
+ {
+ tenantId: "workbench-a",
+ turns: 2,
+ tokens: {
+ input: 10,
+ cacheRead: 0,
+ cacheWrite: 0,
+ output: 5,
+ thinking: 0,
+ total: 15,
+ },
+ costUsd: null,
+ },
+ ];
+ const kept = teamSpaceWorkbenchRows(rows, {
+ tenantId: "team-root",
+ isTeamSpace: true,
+ });
+ expect(kept.map((r) => r.tenantId)).toEqual(["workbench-a"]);
+ });
+});
diff --git a/packages/insights/src/queries.ts b/packages/insights/src/queries.ts
index 13bf1cd0b..fc0a7fcbe 100644
--- a/packages/insights/src/queries.ts
+++ b/packages/insights/src/queries.ts
@@ -93,6 +93,10 @@ export function emptyOverallUsageSummary(): OverallUsageSummary {
function turnCostUsd(row: UsageTurnRecord): number | null {
if (row.reportedCostUsd !== null) return row.reportedCostUsd;
+ // A recorded turn with no token counts is not $0 spend — we did not
+ // observe usage, so cost is unknown (CL-6659). Catalog FREE rates still
+ // apply once tokens are present.
+ if (totalTokens(row.tokens) === 0) return null;
return costUsd({
provider: row.provider ?? "",
model: row.model,
@@ -209,10 +213,27 @@ export async function summarizeUsageByTenant(
});
}
+/**
+ * Team-space `/workbenches` rows: drop an empty parent (it is a duplicate
+ * of the "All workbenches" landing, CL-6368). Keep the parent when it
+ * recorded turns of its own — otherwise those turns vanish from the
+ * breakdown (CL-6659).
+ */
+export function teamSpaceWorkbenchRows<
+ T extends { readonly tenantId: string; readonly turns: number },
+>(
+ rows: readonly T[],
+ opts: { readonly tenantId: string; readonly isTeamSpace: boolean },
+): readonly T[] {
+ if (!opts.isTeamSpace) return rows;
+ return rows.filter((row) => row.tenantId !== opts.tenantId || row.turns > 0);
+}
+
/**
* Activity histogram by UTC day across a tenant scope (see summarizeUsage
- * for what `tenantIds` means). Token totals are always known for recorded
- * turns; pre-sink history simply does not appear.
+ * for what `tenantIds` means). Token totals are known when the adapter
+ * reported them; a recorded turn with every class at zero is an honest
+ * absence, not $0.00. Pre-sink history simply does not appear.
*/
export async function activityByDay(
store: UsageStore,
diff --git a/packages/insights/src/routes.ts b/packages/insights/src/routes.ts
index a5b86c265..08cd71efe 100644
--- a/packages/insights/src/routes.ts
+++ b/packages/insights/src/routes.ts
@@ -23,6 +23,7 @@ import {
summarizeLatency,
summarizeUsage,
summarizeUsageByTenant,
+ teamSpaceWorkbenchRows,
type RunTraceReader,
type ToolCallReader,
} from "./queries";
@@ -210,12 +211,11 @@ export function createInsightsRoutes(
* seeing the scope's sum. Calling it for a leaf workbench (no
* descendants) returns that one workbench's own row.
*
- * A `parentId === null` requested tenant is never itself a workbench —
- * it is the account root, the container real workbenches (each its own
- * child tenant, CL-6089) live under. Its own row would just be a
- * zero-usage rollup-of-itself duplicate of the "All workbenches"
- * landing this chart already sits on, so it is dropped rather than
- * listed alongside its children (CL-6368).
+ * A `parentId === null` requested tenant is the account root, the
+ * container real workbenches live under (CL-6089). An empty parent row
+ * is a duplicate of the "All workbenches" landing and is dropped
+ * (CL-6368). A parent that recorded turns of its own stays in the
+ * breakdown so those turns are not silently excluded (CL-6659).
*/
app.get(
"/workbenches",
@@ -229,9 +229,10 @@ export function createInsightsRoutes(
summarizeUsageByTenant(deps.store, scope, range),
tenantNames(deps.db, tenant.id, tenant.name, scope),
]);
- const isTeamSpace = tenant.parentId === null;
- const items = rows
- .filter((row) => !isTeamSpace || row.tenantId !== tenant.id)
+ const items = teamSpaceWorkbenchRows(rows, {
+ tenantId: tenant.id,
+ isTeamSpace: tenant.parentId === null,
+ })
.map((row) => ({
tenantId: row.tenantId,
name: names.get(row.tenantId) ?? row.tenantId,
diff --git a/packages/ollama-adapter/src/adapter.test.ts b/packages/ollama-adapter/src/adapter.test.ts
index 45ca88d53..b25dc73c5 100644
--- a/packages/ollama-adapter/src/adapter.test.ts
+++ b/packages/ollama-adapter/src/adapter.test.ts
@@ -74,7 +74,9 @@ describe("createOllamaAdapter", () => {
const inner = createOpenAIAdapter(source);
const wrappedBuilt = wrapped.buildRequest(messages, "gpt-oss:20b", options);
const innerBuilt = inner.buildRequest(messages, "gpt-oss:20b", options);
- expect(bodyOf(wrappedBuilt)).toEqual(bodyOf(innerBuilt));
+ const wrappedBody = bodyOf(wrappedBuilt);
+ const { stream_options: _streamOptions, ...rest } = wrappedBody;
+ expect(rest).toEqual(bodyOf(innerBuilt));
expect(wrappedBuilt.url).toBe(innerBuilt.url);
expect(wrappedBuilt.headers).toEqual(innerBuilt.headers);
});
@@ -180,4 +182,57 @@ describe("createOllamaAdapter", () => {
query: "this person",
});
});
+
+ test("every request asks the OpenAI-compat endpoint for a terminal usage chunk", () => {
+ const wrapped = createOllamaAdapter(source, undefined);
+ const body = bodyOf(wrapped.buildRequest(messages, "gpt-oss:20b", options));
+ expect(body["stream_options"]).toEqual({ include_usage: true });
+ });
+
+ test("parseResponse maps native prompt_eval_count/eval_count onto inference.usage", () => {
+ const wrapped = createOllamaAdapter(source, undefined);
+ const events = wrapped.parseResponse(
+ JSON.stringify({
+ choices: [{ delta: {}, finish_reason: "stop" }],
+ prompt_eval_count: 42,
+ eval_count: 17,
+ }),
+ );
+ const usage = events.filter((event) => event.type === "inference.usage");
+ expect(usage).toHaveLength(1);
+ expect(usage[0]?.data).toEqual({
+ usage: {
+ input: 42,
+ output: 17,
+ cacheRead: 0,
+ cacheWrite: 0,
+ thinking: 0,
+ },
+ source,
+ });
+ });
+
+ test("parseResponse does not double-emit usage when the OpenAI-shaped object is already present", () => {
+ const wrapped = createOllamaAdapter(source, undefined);
+ const events = wrapped.parseResponse(
+ JSON.stringify({
+ choices: [{ delta: {}, finish_reason: "stop" }],
+ usage: { prompt_tokens: 10, completion_tokens: 4 },
+ prompt_eval_count: 99,
+ eval_count: 99,
+ }),
+ );
+ const usage = events.filter((event) => event.type === "inference.usage");
+ expect(usage).toHaveLength(1);
+ expect(usage[0]?.data).toEqual({
+ usage: {
+ input: 10,
+ output: 4,
+ cacheRead: 0,
+ cacheWrite: 0,
+ thinking: 0,
+ },
+ source,
+ });
+ });
});
diff --git a/packages/ollama-adapter/src/adapter.ts b/packages/ollama-adapter/src/adapter.ts
index 4eb2520dc..bead8d88d 100644
--- a/packages/ollama-adapter/src/adapter.ts
+++ b/packages/ollama-adapter/src/adapter.ts
@@ -18,7 +18,11 @@ import type {
BuiltRequest,
ProviderAdapter,
} from "@intx/inference";
-import type { LastCycleSource } from "@intx/types/runtime";
+import type {
+ InferenceEvent,
+ LastCycleSource,
+ TokenUsage,
+} from "@intx/types/runtime";
import {
parseOllamaAdapterConfig,
@@ -38,20 +42,18 @@ type OllamaChatBody = {
max_tokens?: number;
max_completion_tokens?: number;
reasoning_effort?: string;
+ stream_options?: { include_usage: boolean };
};
function applyOverride(
built: BuiltRequest,
override: OllamaAdapterOverride,
): BuiltRequest {
- if (
- override.numCtx === undefined &&
- override.maxOutputTokens === undefined &&
- override.reasoningEffort === undefined
- ) {
- return built;
- }
const body = JSON.parse(built.body) as OllamaChatBody;
+ // Without include_usage, Ollama's OpenAI-compat stream often ends with no
+ // usage object; the harness then synthesizes zero token counts that Insights
+ // used to display as Cost $0.00 / 0/0 (CL-6659).
+ body.stream_options = { include_usage: true };
if (override.numCtx !== undefined) {
body.options = { ...body.options, num_ctx: override.numCtx };
}
@@ -71,6 +73,81 @@ function applyOverride(
return { ...built, body: JSON.stringify(body) };
}
+function isPlainObject(value: unknown): value is Record