feat(agent-sessions): grade the list's Errors cell as the Overview grades its checks

apps/ai/src/mcp/tools/__tests__/agent-sessions.test.ts

@@ -61,6 +61,23 @@ const listRow = {
 	errorAgentSpans: 1,
 	toolErrors: 1,
 	turnErrors: 0,
+	failures: [
+		[
+			"span-tool-1",
+			"span-agent-1",
+			1,
+			0,
+			"TimeoutError",
+			"search_traces",
+			"eve",
+			"search timed out",
+			"",
+			"",
+			"trace-1",
+			1755599605825,
+		],
+	],
+	terminalSpanId: "",
 	serviceNames: ["agent-runner"],
 	models: ["gpt-5"],
 	agentNames: ["maple"],
@@ -291,7 +308,8 @@ describe("list_agent_sessions rendering", () => {
 		expect(cells?.[1]).toBe("maple")
 		expect(cells?.[2]).toBe("eve")
 		// Errors are agent/tool/turn, in that order.
-		expect(cells?.[7]).toBe("1/1/0")
+		// The breakdown by label, not the raw counts, once the index classified one.
+		expect(cells?.[7]).toBe("tool_timeout · search_traces")
 		// The hint is the row's bounds PADDED the way the page pads them: a row's
 		// bounds are its agent spans' extent, and both read levels bound on
 		// `Timestamp`, so handing them over verbatim would drop the app spans

apps/ai/src/mcp/tools/list-agent-sessions.ts

@@ -14,6 +14,7 @@ import { formatNextSteps } from "../lib/next-steps"
 import { windowHint } from "../lib/agent-sessions"
 import { Effect, Schema } from "effect"
 import {
+	type AiSessionFailureSummary,
 	AI_SESSION_SEARCH_MAX_CHARS,
 	AiSessionSortDir,
 	AiSessionSortKey,
@@ -150,7 +151,8 @@ export function registerListAgentSessionsTool(server: McpToolRegistrar) {
 				formatDurationFromMs(session.durationMs),
 				formatNumber(session.llmCalls),
 				formatNumber(session.toolCalls),
-				`${session.errorSpanCount}/${session.toolErrorCount}/${session.turnErrorCount}`,
+				failuresCell(session.failures) ??
+					`${session.errorSpanCount}/${session.toolErrorCount}/${session.turnErrorCount}`,
 				formatNumber(session.totalTokens),
 				session.cost > 0 ? formatCost(session.cost) : "—",
 				truncate(session.models.join(", "), 40),
@@ -160,7 +162,7 @@ export function registerListAgentSessionsTool(server: McpToolRegistrar) {
 			const lines: string[] = [
 				`## AI agent sessions (showing ${offset + 1}–${offset + sessions.length})`,
 				`Time range: ${st} — ${et}`,
-				`Every figure is over the session's AGENT spans; the app's own spans in the same traces are not counted. Errors are agent/tool/turn.`,
+				`Every figure is over the session's AGENT spans; the app's own spans in the same traces are not counted. Failures are by label, ×count; "!" marks one that needs a fix (the run died on it, or its kind always does), the rest were survived. A bare a/b/c is errored agent/tool/turn spans the index could not classify.`,
 				``,
 				formatTable(
 					[
@@ -171,7 +173,7 @@ export function registerListAgentSessionsTool(server: McpToolRegistrar) {
 						"Duration",
 						"LLM calls",
 						"Tool calls",
-						"Errors",
+						"Failures",
 						"Tokens",
 						"Cost",
 						"Models",
@@ -180,7 +182,10 @@ export function registerListAgentSessionsTool(server: McpToolRegistrar) {
 					rows,
 				),
 				...(sessions.length === limit
-					? [``, `The page is full; more sessions may match — call again with offset=${offset + sessions.length}.`]
+					? [
+							``,
+							`The page is full; more sessions may match — call again with offset=${offset + sessions.length}.`,
+						]
 					: []),
 				formatNextSteps(
 					sessions.slice(0, 3).map(
@@ -197,3 +202,24 @@ export function registerListAgentSessionsTool(server: McpToolRegistrar) {
 		}),
 	)
 }
+
+/** `!context_length_exceeded, tool_error · run_tests ×2` — the row's failures
+ *  by label, a `!` on each one that needs a fix, whole labels only and
+ *  `+N more` past the cell's width. `undefined` when the index classified
+ *  none, and the raw counts say what it saw. */
+function failuresCell(failures: ReadonlyArray<AiSessionFailureSummary>): string | undefined {
+	if (failures.length === 0) return undefined
+	const labels = failures.map(
+		(failure) =>
+			`${failure.severity === "failure" ? "!" : ""}${failure.label}${failure.count > 1 ? ` ×${failure.count}` : ""}`,
+	)
+	const shown: string[] = []
+	for (const label of labels) {
+		const next = [...shown, label].join(", ")
+		if (shown.length > 0 && next.length > FAILURES_CELL_MAX) break
+		shown.push(label)
+	}
+	const more = labels.length - shown.length
+	return shown.join(", ") + (more > 0 ? ` +${more} more` : "")
+}
+const FAILURES_CELL_MAX = 80

apps/api/src/routes/internal/ai-sessions.http.test.ts

@@ -454,6 +454,41 @@ describe("POST /internal/ai-sessions/list", () => {
 		errorAgentSpans: "1",
 		toolErrors: 1,
 		turnErrors: 0,
+		// The failed spans, as the wire renders the tuple: a rate-limited model
+		// call — no tool, so the summary carries no `tool` key at all — and a
+		// tool that timed out, both in a trace that is not the session's last,
+		// so both warnings.
+		failures: [
+			[
+				"span-llm-1",
+				"span-agent-1",
+				0,
+				1,
+				"",
+				"",
+				"eve",
+				"429 Too Many Requests",
+				"",
+				"",
+				"trace-1",
+				"1755599604825",
+			],
+			[
+				"span-tool-1",
+				"span-agent-1",
+				1,
+				0,
+				"TimeoutError",
+				"search_traces",
+				"eve",
+				"search timed out",
+				"",
+				"",
+				"trace-1",
+				"1755599605825",
+			],
+		],
+		terminalSpanId: "",
 		totalTokens: 18_400,
 		inputTokens: 12_000,
 		cacheReadTokens: 4_000,
@@ -523,6 +558,24 @@ describe("POST /internal/ai-sessions/list", () => {
 				errorSpanCount: 1,
 				toolErrorCount: 1,
 				turnErrorCount: 0,
+				// Classified off the shipped tuples by the detail page's own rule.
+				failures: [
+					{
+						kind: "rateLimited",
+						label: "rate_limit",
+						count: 1,
+						severity: "anomaly",
+						terminal: false,
+					},
+					{
+						kind: "toolTimeout",
+						label: "tool_timeout · search_traces",
+						tool: "search_traces",
+						count: 1,
+						severity: "anomaly",
+						terminal: false,
+					},
+				],
 				models: ["claude-sonnet-5"],
 				agentNames: ["web-fetcher", "slack-agent"],
 				firstAgentName: "slack-agent",

apps/web/src/components/agent-sessions/agent-sessions-list.test.tsx

@@ -50,6 +50,7 @@ const session: AgentSessionRow = {
 	errorSpanCount: 0,
 	toolErrorCount: 0,
 	turnErrorCount: 0,
+	failures: [],
 	serviceNames: ["maple-slack-agent"],
 	models: ["claude-sonnet-5"],
 	agentNames: ["web-fetcher", "slack-agent"],
@@ -114,7 +115,13 @@ describe("AgentSessionsList", () => {
 	it("asks for the next page when the sentinel comes into view, but not while one is in flight", () => {
 		const onReachEnd = vi.fn()
 		const view = renderList(
-			<AgentSessionsList {...sort} sessions={[session]} hasMore onReachEnd={onReachEnd} loadingMore={false} />,
+			<AgentSessionsList
+				{...sort}
+				sessions={[session]}
+				hasMore
+				onReachEnd={onReachEnd}
+				loadingMore={false}
+			/>,
 		)
 
 		const first = MockIntersectionObserver.instances[0]!
@@ -141,7 +148,7 @@ describe("AgentSessionsList", () => {
 		expect(MockIntersectionObserver.instances).toHaveLength(0)
 	})
 
-	it("names the framework by its mark alone, and splits the failures by kind", () => {
+	it("names the framework by its mark alone, and splits the failures by severity", () => {
 		const view = renderList(
 			<AgentSessionsList
 				{...sort}
@@ -151,6 +158,23 @@ describe("AgentSessionsList", () => {
 						errorSpanCount: 5,
 						toolErrorCount: 2,
 						turnErrorCount: 1,
+						failures: [
+							{
+								kind: "contextExceeded",
+								label: "context_length_exceeded",
+								count: 1,
+								severity: "failure",
+								terminal: true,
+							},
+							{
+								kind: "error",
+								label: "tool_error · run_tests",
+								tool: "run_tests",
+								count: 2,
+								severity: "anomaly",
+								terminal: false,
+							},
+						],
 					},
 				]}
 			/>,
@@ -164,8 +188,26 @@ describe("AgentSessionsList", () => {
 		// count and noun into fixed-width slots, so match on the whole chip's text.
 		const chip = (label: string) => (_: string, element: Element | null) =>
 			element?.classList.contains("rounded-full") === true && element.textContent === label
-		expect(view.getAllByText(chip("2 tools"))).toHaveLength(1)
-		expect(view.getAllByText(chip("1 turn"))).toHaveLength(1)
+		expect(view.getAllByText(chip("1 failure"))).toHaveLength(1)
+		expect(view.getAllByText(chip("2 warnings"))).toHaveLength(1)
+		// The breakdown is the chip's accessible name: the labels exist nowhere
+		// else until the tooltip opens.
+		expect(view.getByLabelText("context_length_exceeded (ended the run)")).toBeTruthy()
+		expect(view.getByLabelText("tool_error · run_tests ×2")).toBeTruthy()
+	})
+
+	it("counts the errored spans when the index classified none of them", () => {
+		const view = renderList(
+			<AgentSessionsList
+				{...sort}
+				sessions={[
+					{ ...session, errorSpanCount: 3, toolErrorCount: 0, turnErrorCount: 0, failures: [] },
+				]}
+			/>,
+		)
+		const chip = (label: string) => (_: string, element: Element | null) =>
+			element?.classList.contains("rounded-full") === true && element.textContent === label
+		expect(view.getAllByText(chip("3 spans"))).toHaveLength(1)
 		expect(view.getByText("18.4k")).toBeTruthy()
 		expect(view.getByText("maple-slack-agent")).toBeTruthy()
 	})
@@ -199,7 +241,12 @@ describe("AgentSessionsList", () => {
 	it("sorts through the column headers, marking the one the rows are in", () => {
 		const onSortChange = vi.fn()
 		const view = renderList(
-			<AgentSessionsList sessions={[session]} sortBy="cost" sortDir="desc" onSortChange={onSortChange} />,
+			<AgentSessionsList
+				sessions={[session]}
+				sortBy="cost"
+				sortDir="desc"
+				onSortChange={onSortChange}
+			/>,
 		)
 		const cost = view.getByRole("columnheader", { name: "Cost" })
 		expect(cost.getAttribute("aria-sort")).toBe("descending")