Compare commits

..
Author SHA1 Message Date
Mohamed Boudra 87f5472901 fix(server): keep short strings whole and typed arrays valid when limiting items
Strings up to 1 KiB are never cut, so a tool call keeps its callId and an
item limits to the same result twice. Arrays drop trailing elements
without inserting a marker element, which typed arrays such as todo items
would reject.
2026-10-30 09:40:57 +01:00
Mohamed Boudra 40ba2cee69 fix(server): bound every timeline item so one fetch cannot close the socket
A timeline page is one WebSocket message. Only three tool-call fields were
capped, so MCP results, file contents, errors and merged assistant rows could
make a 20-entry or even a 1-entry fetch pass the outbound high-water mark.

limitAgentTimelineItemContent is now one budgeted walk over any item: 64 KiB
for tool calls, 256 KiB for other items. Over budget, the largest strings are
cut to a shared length with a visible marker and oversized data URLs become a
marker. Fetches also apply it to merged assistant and reasoning rows.
2026-10-30 08:03:21 +01:00
57 changed files with 1443 additions and 691 deletions
+10 -4
View File
@@ -21,10 +21,16 @@ The invariants are:
> agent establishes the daemon's current tail in one complete-turn request, with older history reachable
> through backward pagination.
Tool output is bounded before it enters either delivery path. Canonical shell tool output is sliced
to 64 KiB, and the same bounded item is used for runtime timeline rows and live stream events.
Provider history hydration applies the same rule so reopening an agent cannot restore an oversized
tool payload.
Every item is bounded before it enters either delivery path, because a timeline page is one
WebSocket message and the daemon closes a socket whose outbound buffer passes 64 MiB. Tool calls get
64 KiB and other items 256 KiB. Over budget, the largest strings are cut to a shared length with a
`…[truncated 12.3 MB]` marker, and an oversized `data:` URL is replaced by the marker. Strings up
to 1 KiB stay whole, so ids and paths survive. Arrays drop their trailing elements without a marker
element, so typed arrays still match the wire schema. Limiting an item twice gives the same item. The
budget does not bound wide objects or multi-byte text. The same
bounded item is used for runtime timeline rows and live stream events. Provider history hydration
applies the same rule so reopening an agent cannot restore an oversized payload. Fetches apply it
again to assistant and reasoning rows the projection merges from many chunks.
## Presence is not delivery
@@ -64,7 +64,7 @@ function resolveShouldFill(
): boolean {
if (!fillAvailableHeight) return false;
const t = detail?.type;
return t === "shell" || t === "edit" || t === "write" || t === "read";
return t === "shell" || t === "edit" || t === "write" || t === "read" || t === "sub_agent";
}
function useDetailStyles(
@@ -186,6 +186,184 @@ function ShellDetailSection({ command, output, ds }: ShellDetailProps) {
);
}
function resolveSubAgentFallbackHeader(
subAgentType: string | null | undefined,
description: string | null | undefined,
fallbackText: string,
): string {
if (subAgentType && description) {
return `${subAgentType}: ${description}`;
}
return subAgentType ?? description ?? fallbackText;
}
interface SubAgentDetailProps {
log: string;
childSessionId: string | null | undefined;
subAgentType: string | null | undefined;
description: string | null | undefined;
ds: DetailStyles;
}
interface SubAgentActivityRow {
index: number;
toolName: string;
summary?: string;
}
interface ParsedSubAgentLog {
actions: SubAgentActivityRow[];
remainingLog: string;
}
function parseBracketedSubAgentLine(line: string, index: number): SubAgentActivityRow | null {
const match = line.match(/^\[([^\]]+)\](?:\s+(.*))?$/);
if (!match) {
return null;
}
const toolName = match[1]?.trim();
if (!toolName) {
return null;
}
const summary = match[2]?.trim();
return {
index,
toolName,
...(summary ? { summary } : {}),
};
}
function parseSubAgentLog(log: string): ParsedSubAgentLog {
const actions: SubAgentActivityRow[] = [];
const remainingLines: string[] = [];
for (const line of log.replace(/^\n+/, "").split("\n")) {
const normalizedLine = line.trim();
if (!normalizedLine) {
continue;
}
const parsedAction = parseBracketedSubAgentLine(normalizedLine, actions.length + 1);
if (parsedAction) {
actions.push(parsedAction);
} else {
remainingLines.push(line);
}
}
return {
actions,
remainingLog: remainingLines.join("\n").replace(/^\n+/, ""),
};
}
function SubAgentActionRow({ action }: { action: SubAgentActivityRow }) {
return (
<View style={styles.subAgentActionRow}>
<Text selectable style={styles.subAgentActionTool}>
{formatSubAgentToolName(action.toolName)}
</Text>
{action.summary ? (
<Text selectable style={styles.subAgentActionSummary}>
{action.summary}
</Text>
) : null}
</View>
);
}
function formatSubAgentToolName(toolName: string): string {
const trimmed = toolName.trim();
if (!trimmed) {
return toolName;
}
return trimmed
.replace(/[._-]+/g, " ")
.split(" ")
.filter((segment) => segment.length > 0)
.map((segment) => `${segment[0]?.toUpperCase() ?? ""}${segment.slice(1)}`)
.join(" ");
}
function SubAgentLogText({
activityLog,
fallbackHeader,
hasActions,
}: {
activityLog: string;
fallbackHeader: string;
hasActions: boolean;
}) {
if (activityLog.length > 0) {
return (
<Text selectable style={styles.scrollText}>
{activityLog}
</Text>
);
}
if (!hasActions) {
return (
<Text selectable style={styles.scrollText}>
{fallbackHeader}
</Text>
);
}
return null;
}
function SubAgentDetailSection({
log,
childSessionId,
subAgentType,
description,
ds,
}: SubAgentDetailProps) {
const { t } = useTranslation();
const { actions, remainingLog } = useMemo(() => parseSubAgentLog(log), [log]);
const fallbackHeader = resolveSubAgentFallbackHeader(
subAgentType,
description,
t("toolCallDetails.subAgentActivity"),
);
const hasActions = actions.length > 0;
return (
<View style={ds.sectionFillStyle}>
<View style={ds.codeBlockFillStyle}>
<ScrollView
style={ds.codeVerticalScrollStyle}
contentContainerStyle={styles.codeVerticalContent}
nestedScrollEnabled
showsVerticalScrollIndicator
>
<ScrollView
horizontal
nestedScrollEnabled
showsHorizontalScrollIndicator
contentContainerStyle={styles.codeHorizontalContent}
>
<View style={styles.codeLine} dataSet={CODE_SURFACE_DATASET}>
{childSessionId ? (
<Text selectable style={styles.subAgentSessionText}>
session {childSessionId}
</Text>
) : null}
{hasActions ? (
<View style={styles.subAgentActions}>
{actions.map((action) => (
<SubAgentActionRow key={action.index} action={action} />
))}
</View>
) : null}
<SubAgentLogText
activityLog={remainingLog}
fallbackHeader={fallbackHeader}
hasActions={hasActions}
/>
</View>
</ScrollView>
</ScrollView>
</View>
</View>
);
}
interface EditDetailProps {
diffLines: DiffLine[] | undefined;
ds: DetailStyles;
@@ -458,7 +636,19 @@ function buildDetailSections(
<ShellDetailSection key="shell" command={detail.command} output={detail.output} ds={ds} />,
];
}
if (detail.type === "worktree_setup" || detail.type === "sub_agent") return [];
if (detail.type === "worktree_setup") return [];
if (detail.type === "sub_agent") {
return [
<SubAgentDetailSection
key="sub-agent"
log={detail.log}
childSessionId={detail.childSessionId}
subAgentType={detail.subAgentType}
description={detail.description}
ds={ds}
/>,
];
}
if (detail.type === "edit") {
return [<EditDetailSection key="edit" diffLines={diffLines} ds={ds} />];
}
@@ -710,6 +900,34 @@ const styles = StyleSheet.create((theme) => {
shellPrompt: {
color: theme.colors.foregroundMuted,
},
subAgentSessionText: {
fontFamily: theme.fontFamily.mono,
fontSize: theme.fontSize.code,
color: theme.colors.foregroundMuted,
lineHeight: 18,
marginBottom: theme.spacing[2],
},
subAgentActions: {
gap: theme.spacing[1],
marginBottom: theme.spacing[2],
},
subAgentActionRow: {
flexDirection: "row",
alignItems: "center",
gap: theme.spacing[2],
},
subAgentActionTool: {
fontFamily: theme.fontFamily.mono,
fontSize: theme.fontSize.code,
color: theme.colors.foregroundMuted,
lineHeight: 18,
},
subAgentActionSummary: {
fontFamily: theme.fontFamily.mono,
fontSize: theme.fontSize.code,
color: theme.colors.foreground,
lineHeight: 18,
},
jsonScroll: {
borderWidth: theme.borderWidth[1],
borderColor: theme.colors.border,
+1
View File
@@ -832,6 +832,7 @@ describe("translation resources", () => {
expect(en.common.placeholders.search).toBe("Search...");
expect(en.common.empty.noResults).toBe("No results found");
expect(en.common.empty.noOptionsMatchSearch).toBe("No options match your search.");
expect(en.toolCallDetails.subAgentActivity).toBe("Sub-agent activity");
expect(en.panels.file.failedToLoadPreview).toBe("Failed to load file preview");
});
+1
View File
@@ -2013,6 +2013,7 @@ export const ar: TranslationResources = {
toolCallDetails: {
error: "خطأ",
empty: "لا توجد تفاصيل إضافية متاحة",
subAgentActivity: "نشاط الوكيل الفرعي",
input: "مدخل",
output: "الإخراج",
},
+1
View File
@@ -2038,6 +2038,7 @@ export const en = {
toolCallDetails: {
error: "Error",
empty: "No additional details available",
subAgentActivity: "Sub-agent activity",
input: "Input",
output: "Output",
},
+1
View File
@@ -2061,6 +2061,7 @@ export const es: TranslationResources = {
toolCallDetails: {
error: "Error",
empty: "No hay detalles adicionales disponibles",
subAgentActivity: "Actividad de subagente",
input: "Aporte",
output: "Producción",
},
+1
View File
@@ -2060,6 +2060,7 @@ export const fr: TranslationResources = {
toolCallDetails: {
error: "Erreur",
empty: "Aucun détail supplémentaire disponible",
subAgentActivity: "Activité du sous-agent",
input: "Entrée",
output: "Sortie",
},
+1
View File
@@ -2033,6 +2033,7 @@ export const ja: TranslationResources = {
toolCallDetails: {
error: "エラー",
empty: "追加の詳細はありません",
subAgentActivity: "サブエージェントアクティビティ",
input: "入力",
output: "出力",
},
+1
View File
@@ -2023,6 +2023,7 @@ export const ko: TranslationResources = {
toolCallDetails: {
error: "오류",
empty: "추가 세부 정보가 없습니다",
subAgentActivity: "서브에이전트 활동",
input: "입력",
output: "출력",
},
+1
View File
@@ -2046,6 +2046,7 @@ export const ptBR: TranslationResources = {
toolCallDetails: {
error: "Erro",
empty: "Nenhum detalhe adicional disponível",
subAgentActivity: "Atividade do subagente",
input: "Entrada",
output: "Saída",
},
+1
View File
@@ -2046,6 +2046,7 @@ export const ru: TranslationResources = {
toolCallDetails: {
error: "Ошибка",
empty: "Дополнительные сведения отсутствуют",
subAgentActivity: "Активность субагента",
input: "Входные данные",
output: "Выходные данные",
},
+1
View File
@@ -1991,6 +1991,7 @@ export const zhCN: TranslationResources = {
toolCallDetails: {
error: "错误",
empty: "没有可用的更多详情",
subAgentActivity: "Sub-agent 活动",
input: "输入",
output: "输出",
},
@@ -40,6 +40,12 @@ function hasMeaningfulFetchDetail(detail: Extract<ToolCallDetail, { type: "fetch
return Boolean(detail.url || detail.result || detail.codeText);
}
function hasMeaningfulSubAgentDetail(
detail: Extract<ToolCallDetail, { type: "sub_agent" }>,
): boolean {
return Boolean(detail.subAgentType || detail.description || detail.log);
}
function hasMeaningfulPlainTextDetail(
detail: Extract<ToolCallDetail, { type: "plain_text" }>,
): boolean {
@@ -69,9 +75,9 @@ export function hasMeaningfulToolCallDetail(detail: ToolCallDetail | undefined):
case "fetch":
return hasMeaningfulFetchDetail(detail);
case "worktree_setup":
// A sub-agent card is a launch marker; the child's activity lives in the sub-agent panel.
case "sub_agent":
return false;
case "sub_agent":
return hasMeaningfulSubAgentDetail(detail);
case "plain_text":
return hasMeaningfulPlainTextDetail(detail);
case "plan":
@@ -86,7 +92,7 @@ export function isPendingToolCallDetail(params: {
status: "executing" | "running" | "completed" | "failed" | "canceled";
error: unknown;
}): boolean {
if (params.detail?.type === "worktree_setup" || params.detail?.type === "sub_agent") return false;
if (params.detail?.type === "worktree_setup") return false;
const isRunning = params.status === "running" || params.status === "executing";
return isRunning && params.error == null && !hasMeaningfulToolCallDetail(params.detail);
}
+1 -1
View File
@@ -288,7 +288,7 @@ export type ToolCallDetail =
subAgentType?: string;
description?: string;
childSessionId?: string;
log?: string;
log: string;
actions?: Array<{
index: number;
toolName: string;
+1 -3
View File
@@ -638,9 +638,7 @@ const ToolCallDetailPayloadSchema: z.ZodType<ToolCallDetail, unknown> = z.discri
subAgentType: z.string().optional(),
description: z.string().optional(),
childSessionId: z.string().optional(),
// COMPAT(subagent-log): added in v0.14.1, remove after 2027-04-30. A sub-agent card is a
// launch marker; apps before v0.14.1 required `log`, so daemons still send it empty.
log: z.string().optional(),
log: z.string(),
// Compat cruft for clients <= 0.1.65-beta.3 that required this field. Producers still
// emit `[]`; nothing reads it. Drop the field (and the `[]` emissions) once those
// clients are no longer in the field.
@@ -200,8 +200,8 @@ describe("wire schema compatibility", () => {
});
});
test("a sub-agent launch marker still parses against the v0.1.65-beta.3 schema", () => {
const card = {
test("sub_agent tool-call payload still parses against the v0.1.65-beta.3 schema", () => {
const parsed = LegacySubAgentToolCallSchema.parse({
type: "tool_call",
callId: "call-sub-agent-1",
name: "Task",
@@ -212,26 +212,12 @@ describe("wire schema compatibility", () => {
subAgentType: "Explore",
description: "Inspect repository structure",
childSessionId: "child-session-1",
log: "",
log: "[Read] README.md",
actions: [],
},
} as const;
const parsed = LegacySubAgentToolCallSchema.parse(JSON.parse(JSON.stringify(card)));
expect(parsed.detail).toMatchObject({ log: "", actions: [] });
const { log: _log, ...withoutLog } = card.detail;
expect(LegacySubAgentToolCallSchema.safeParse({ ...card, detail: withoutLog }).success).toBe(
false,
);
expect(AgentTimelineItemPayloadSchema.parse({ ...card, detail: withoutLog })).toMatchObject({
detail: { type: "sub_agent", description: "Inspect repository structure" },
});
expect(
AgentTimelineItemPayloadSchema.parse({
...card,
detail: { ...card.detail, log: "[Read] README.md" },
}),
).toMatchObject({ detail: { log: "[Read] README.md" } });
expect(parsed.detail.actions).toEqual([]);
});
test("old clients parse agent snapshots with rewind capabilities", () => {
@@ -247,7 +247,7 @@ second line'`,
]);
});
it("lists each sub-agent as its launch marker", () => {
it("keeps nested sub-agent logs beside the sub-agent that produced them", () => {
const timeline: AgentTimelineItem[] = [
{ type: "assistant_message", text: "Before first child." },
toolCallItem({
@@ -257,7 +257,7 @@ second line'`,
type: "sub_agent",
subAgentType: "Child one",
description: "First investigation",
log: "",
log: "[Assistant] First child result.",
},
}),
{ type: "assistant_message", text: "Between children." },
@@ -268,7 +268,7 @@ second line'`,
type: "sub_agent",
subAgentType: "Child two",
description: "Second investigation",
log: "",
log: "[Assistant] Second child result.",
},
}),
{ type: "assistant_message", text: "After second child." },
@@ -279,8 +279,10 @@ second line'`,
expect(result.split("\n")).toEqual([
"[Assistant] Before first child.",
"[Child one] First investigation",
"[Assistant] First child result.",
"[Assistant] Between children.",
"[Child two] Second investigation",
"[Assistant] Second child result.",
"[Assistant] After second child.",
]);
});
@@ -164,6 +164,9 @@ function curateProjectedActivityEntries(
break;
}
entries.push(formatToolCallEntry(item, options));
if (item.detail.type === "sub_agent" && item.detail.log.trim()) {
entries.push(activityEntry(item.detail.log.trim()));
}
break;
}
case "todo":
@@ -6,6 +6,7 @@ import { afterEach, describe, expect, test, vi } from "vitest";
import { createTestLogger } from "../../test-utils/test-logger.js";
import { AgentManager, type AgentManagerEvent } from "./agent-manager.js";
import { AGENT_STREAM_COALESCE_DEFAULT_WINDOW_MS } from "./agent-stream-coalescer.js";
import { limitAgentTimelineItemContent } from "./agent-timeline-content.js";
import type { AgentTimelineRow } from "./agent-timeline-store-types.js";
import { projectTimelineRows } from "./timeline-projection.js";
import type {
@@ -31,7 +32,6 @@ import type {
const COALESCE_WINDOW_MS = AGENT_STREAM_COALESCE_DEFAULT_WINDOW_MS;
const BEFORE_COALESCE_WINDOW_MS = Math.max(COALESCE_WINDOW_MS - 1, 0);
const TOOL_CALL_CONTENT_MAX_LENGTH = 64 * 1024;
const TEST_CAPABILITIES: AgentCapabilityFlags = {
supportsStreaming: false,
@@ -400,10 +400,7 @@ describe("target coalesced behavior", () => {
try {
const { agentId, session } = await createManagedSession(harness);
const output = `${"a".repeat(512 * 1024)}${"z".repeat(512 * 1024)}`;
const expectedItem = toolCall({
status: "completed",
output: "a".repeat(TOOL_CALL_CONTENT_MAX_LENGTH),
});
const expectedItem = limitAgentTimelineItemContent(toolCall({ status: "completed", output }));
session.pushEvent(timelineEvent(toolCall({ status: "completed", output })));
await waitForSessionEventQueue();
@@ -425,10 +422,7 @@ describe("target coalesced behavior", () => {
try {
const { agentId } = await createManagedSession(harness);
const output = `${"a".repeat(512 * 1024)}${"z".repeat(512 * 1024)}`;
const expectedItem = toolCall({
status: "completed",
output: "a".repeat(TOOL_CALL_CONTENT_MAX_LENGTH),
});
const expectedItem = limitAgentTimelineItemContent(toolCall({ status: "completed", output }));
await harness.manager.appendTimelineItem(agentId, toolCall({ status: "completed", output }));
@@ -449,10 +443,7 @@ describe("target coalesced behavior", () => {
try {
const { agentId, session } = await createManagedSession(harness);
const output = `${"a".repeat(512 * 1024)}${"z".repeat(512 * 1024)}`;
const expectedItem = toolCall({
status: "completed",
output: "a".repeat(TOOL_CALL_CONTENT_MAX_LENGTH),
});
const expectedItem = limitAgentTimelineItemContent(toolCall({ status: "completed", output }));
session.setHistory([timelineEvent(toolCall({ status: "completed", output }))]);
await harness.manager.hydrateTimelineFromProvider(agentId, { force: true });
@@ -470,10 +461,7 @@ describe("target coalesced behavior", () => {
try {
const { agentId } = await createManagedSession(harness);
const output = `${"a".repeat(512 * 1024)}${"z".repeat(512 * 1024)}`;
const expectedItem = toolCall({
status: "completed",
output: "a".repeat(TOOL_CALL_CONTENT_MAX_LENGTH),
});
const expectedItem = limitAgentTimelineItemContent(toolCall({ status: "completed", output }));
await harness.manager.emitLiveTimelineItem(
agentId,
@@ -495,10 +483,9 @@ describe("target coalesced behavior", () => {
try {
const { agentId, session } = await createManagedSession(harness);
const content = `${"a".repeat(512 * 1024)}${"z".repeat(512 * 1024)}`;
const expectedItem = toolCall({
status: "failed",
error: { content: "a".repeat(TOOL_CALL_CONTENT_MAX_LENGTH) },
});
const expectedItem = limitAgentTimelineItemContent(
toolCall({ status: "failed", error: { content } }),
);
session.pushEvent(timelineEvent(toolCall({ status: "failed", error: { content } })));
await waitForSessionEventQueue();
@@ -21,6 +21,7 @@ import {
} from "./agent-manager.js";
import { AgentStorage } from "./agent-storage.js";
import { InMemoryAgentTimelineStore } from "./agent-timeline-store.js";
import { limitAgentTimelineItemContent } from "./agent-timeline-content.js";
import { toAgentPayload } from "./agent-projections.js";
import { projectTimelineRows } from "./timeline-projection.js";
import { getOpenAgentTabLabel, PARENT_AGENT_ID_LABEL } from "@getpaseo/protocol/agent-labels";
@@ -4592,7 +4593,7 @@ test("importProviderSession imports the selected session without listing and pub
expect(manager.getTimeline(imported.id)).toEqual([
{ type: "user_message", text: "Trace provider imports" },
{ type: "assistant_message", text: "Done" },
{
limitAgentTimelineItemContent({
type: "tool_call",
callId: "large-shell-result",
name: "shell",
@@ -4601,10 +4602,10 @@ test("importProviderSession imports the selected session without listing and pub
detail: {
type: "shell",
command: "print output",
output: "x".repeat(64 * 1024),
output: "x".repeat(1024 * 1024),
exitCode: 0,
},
},
}),
]);
expect(manager.listProviderSubagents(imported.id)).toEqual([
expect.objectContaining({ id: "thread-child", title: "Imported child", status: "completed" }),
@@ -11957,10 +11958,7 @@ test("commits startup notices once on create and after restored history", async
exitCode: 0,
},
};
const limitedToolOutput: AgentTimelineItem = {
...toolOutput,
detail: { type: "shell", command: "print output", output: "x".repeat(64 * 1024), exitCode: 0 },
};
const limitedToolOutput = limitAgentTimelineItemContent(toolOutput);
class NotifyingSession extends TestAgentSession {
readonly initialTimeline = [{ item: notice, timestamp: "2026-09-22T00:00:00.000Z" }];
override async *streamHistory(): AsyncGenerator<AgentStreamEvent> {
@@ -313,10 +313,7 @@ export type ToolCallDetail =
subAgentType?: string;
description?: string;
childSessionId?: string;
// COMPAT(subagent-log): added in v0.14.1, remove after 2027-04-30. Apps before v0.14.1
// require `log`; every producer sends it empty and carries the same tag. Optional here because
// protocol timeline items, where `log` is optional, flow into this type.
log?: string;
log: string;
actions?: Array<{
index: number;
toolName: string;
@@ -1,12 +1,30 @@
import { AgentTimelineItemPayloadSchema } from "@getpaseo/protocol/messages";
import { describe, expect, test } from "vitest";
import type { AgentTimelineItem } from "./agent-sdk-types.js";
import { limitAgentTimelineItemContent } from "./agent-timeline-content.js";
function serializedBytes(item: AgentTimelineItem): number {
return Buffer.byteLength(JSON.stringify(item), "utf8");
}
describe("agent timeline content", () => {
test("returns an item within budget unchanged", () => {
const item: AgentTimelineItem = {
type: "tool_call",
callId: "small",
name: "Bash",
status: "completed",
error: null,
detail: { type: "shell", command: "ls", output: "x".repeat(60 * 1024) },
};
expect(limitAgentTimelineItemContent(item)).toBe(item);
});
test.each([false, true])(
"preserves the exact UTF-16 prefix for oversized shell output (failed=%s)",
"cuts oversized shell output and keeps short fields whole (failed=%s)",
(failed) => {
const text = "x".repeat(64 * 1024 - 1) + "🙂";
const text = "x".repeat(1024 * 1024) + "🙂";
const item = limitAgentTimelineItemContent({
type: "tool_call",
callId: "large-shell",
@@ -15,45 +33,116 @@ describe("agent timeline content", () => {
detail: { type: "shell", command: "echo large", output: text },
error: failed ? { content: text } : null,
});
expect(item).toEqual({
type: "tool_call",
expect(item).toMatchObject({
callId: "large-shell",
name: "Bash",
status: failed ? "failed" : "completed",
detail: { type: "shell", command: "echo large", output: text.slice(0, 64 * 1024) },
error: failed ? { content: text.slice(0, 64 * 1024) } : null,
detail: { type: "shell", command: "echo large" },
});
if (item.type !== "tool_call" || item.detail.type !== "shell") throw new Error("shape");
expect(item.detail.output).toMatch(/^x+…\[truncated \d{3}\.\d KB\]$/);
expect(serializedBytes(item)).toBeLessThanOrEqual(64 * 1024);
},
);
test("marks truncated plain-text tool content within the content budget", () => {
const oversizedInput = "x".repeat(64 * 1024 + 1);
test("replaces an oversized data URL in an unknown tool result with a marker", () => {
const url = `data:image/jpeg;base64,${"A".repeat(14 * 1024 * 1024)}`;
const item = limitAgentTimelineItemContent({
type: "tool_call",
callId: "terminal-session-4242",
name: "terminal",
callId: "screenshot",
name: "node_repl.js",
status: "completed",
error: null,
detail: {
type: "plain_text",
text: oversizedInput,
icon: "square_terminal",
type: "unknown",
input: { index: 1 },
output: { content: [{ type: "image", url }], note: "captured" },
},
});
expect(limitAgentTimelineItemContent(item)).toEqual(item);
expect(item).toEqual({
type: "tool_call",
callId: "terminal-session-4242",
name: "terminal",
callId: "screenshot",
name: "node_repl.js",
status: "completed",
error: null,
detail: {
type: "plain_text",
text: "x".repeat(64 * 1024 - "\n\n[Content truncated]".length) + "\n\n[Content truncated]",
icon: "square_terminal",
type: "unknown",
input: { index: 1 },
output: { content: [{ type: "image", url: "…[truncated 14.0 MB]" }], note: "captured" },
},
});
});
test("gives non-tool items a 256 KiB budget", () => {
const text = "a".repeat(10 * 1024 * 1024);
const item = limitAgentTimelineItemContent({ type: "assistant_message", text });
if (item.type !== "assistant_message") throw new Error("shape");
expect(item.text).toMatch(/^a+…\[truncated 9\.\d MB\]$/);
expect(item.text.length).toBeGreaterThan(200 * 1024);
expect(serializedBytes(item)).toBeLessThanOrEqual(256 * 1024);
});
test("bounds an item whose structure alone exceeds the budget", () => {
const item = limitAgentTimelineItemContent({
type: "tool_call",
callId: "many",
name: "query",
status: "completed",
error: null,
detail: { type: "unknown", input: null, output: Array.from({ length: 1_000_000 }, () => 1) },
});
if (item.type !== "tool_call" || item.detail.type !== "unknown") throw new Error("shape");
const output = item.detail.output as unknown[];
expect(output.length).toBeLessThan(1_000_000);
expect(output.every((value) => value === 1)).toBe(true);
expect(serializedBytes(item)).toBeLessThanOrEqual(64 * 1024);
});
test("keeps short strings whole and limits idempotently when an item has many strings", () => {
const callId = "3f2b8c1e-9a4d-4e6f-8b2a-1c3d5e7f9a0b";
const item = limitAgentTimelineItemContent({
type: "tool_call",
callId,
name: "mcp__search__files",
status: "completed",
error: null,
detail: {
type: "unknown",
input: { query: "useWorkspaceSomethingVeryDescriptive" },
output: Array.from(
{ length: 10_000 },
(_, index) => `${String(index).padStart(5, "0")}${"p".repeat(28)}`,
),
},
});
if (item.type !== "tool_call" || item.detail.type !== "unknown") throw new Error("shape");
expect(item.callId).toBe(callId);
expect(item.detail.input).toEqual({ query: "useWorkspaceSomethingVeryDescriptive" });
const output = item.detail.output as string[];
expect(output.length).toBeLessThan(10_000);
expect(output.every((path) => path.length === 33)).toBe(true);
expect(serializedBytes(item)).toBeLessThanOrEqual(64 * 1024);
expect(limitAgentTimelineItemContent(item)).toBe(item);
});
test("keeps a long typed array valid against the wire schema", () => {
const item = limitAgentTimelineItemContent({
type: "todo",
items: Array.from({ length: 20_000 }, (_, index) => ({
text: `Task ${index}`,
completed: false,
})),
});
if (item.type !== "todo") throw new Error("shape");
expect(item.items.length).toBeLessThan(20_000);
expect(AgentTimelineItemPayloadSchema.safeParse(item).success).toBe(true);
expect(serializedBytes(item)).toBeLessThanOrEqual(256 * 1024);
});
test("is stable on an already-limited item", () => {
const item = limitAgentTimelineItemContent({
type: "reasoning",
text: "r".repeat(1024 * 1024),
});
expect(limitAgentTimelineItemContent(item)).toBe(item);
});
});
@@ -1,9 +1,19 @@
import type { AgentTimelineItem } from "./agent-sdk-types.js";
import type { JsonValue } from "@getpaseo/protocol/agent-types";
const TOOL_CALL_CONTENT_MAX_LENGTH = 64 * 1024;
export const PLUGIN_TIMELINE_DATA_MAX_BYTES = 64 * 1024;
// One timeline fetch sends its entries in one WebSocket message, so every item must stay small.
// Budgets count UTF-16 units plus a fixed cost per node; neither is exact JSON bytes. They bound
// long strings and long arrays, not wide objects (every entry is kept) or multi-byte text.
const TOOL_CALL_BUDGET = 64 * 1024;
const ITEM_BUDGET = 256 * 1024;
const NODE_COST = 8;
const MARKER_COST = 32;
// Ids, names, paths and queries stay whole, so an item keeps its identity across passes.
const SHORT_STRING = 1024;
const MAX_DEPTH = 64;
export function assertPluginTimelineDataSize(data: JsonValue): void {
const serializedBytes = Buffer.byteLength(JSON.stringify(data), "utf8");
if (serializedBytes > PLUGIN_TIMELINE_DATA_MAX_BYTES) {
@@ -17,66 +27,119 @@ function retainTextPrefix(text: string, length: number): string {
return Buffer.from(text.slice(0, length), "utf16le").toString("utf16le");
}
function limitFailedShellError(item: AgentTimelineItem): AgentTimelineItem {
if (
item.type !== "tool_call" ||
item.detail.type !== "shell" ||
item.status !== "failed" ||
typeof item.error !== "object" ||
item.error === null ||
!("content" in item.error) ||
typeof item.error.content !== "string" ||
item.error.content.length <= TOOL_CALL_CONTENT_MAX_LENGTH
) {
return item;
}
return {
...item,
error: {
...item.error,
content: retainTextPrefix(item.error.content, TOOL_CALL_CONTENT_MAX_LENGTH),
},
};
function formatSize(units: number): string {
if (units < 1024) return `${units} B`;
if (units < 1024 * 1024) return `${(units / 1024).toFixed(1)} KB`;
return `${(units / (1024 * 1024)).toFixed(1)} MB`;
}
function limitPlainText(item: AgentTimelineItem): AgentTimelineItem {
if (
item.type !== "tool_call" ||
item.detail.type !== "plain_text" ||
typeof item.detail.text !== "string" ||
item.detail.text.length <= TOOL_CALL_CONTENT_MAX_LENGTH
) {
return item;
}
const notice = "\n\n[Content truncated]";
return {
...item,
detail: {
...item.detail,
text:
retainTextPrefix(item.detail.text, TOOL_CALL_CONTENT_MAX_LENGTH - notice.length) + notice,
},
};
function truncatedMarker(units: number): string {
return `…[truncated ${formatSize(units)}]`;
}
interface ItemSize {
nodeCost: number;
stringLengths: number[];
tooDeep: boolean;
}
function measure(value: unknown, depth: number, size: ItemSize): void {
size.nodeCost += NODE_COST;
if (depth > MAX_DEPTH) {
size.tooDeep = true;
} else if (typeof value === "string") {
size.stringLengths.push(value.length);
} else if (Array.isArray(value)) {
for (const element of value) measure(element, depth + 1, size);
} else if (typeof value === "object" && value !== null) {
for (const [key, child] of Object.entries(value)) {
size.nodeCost += key.length;
measure(child, depth + 1, size);
}
}
}
/** The largest per-string length that keeps all strings, once cut, within `available`. */
function findStringCap(lengths: readonly number[], available: number): number {
let low = 0;
let high = 0;
for (const length of lengths) high = Math.max(high, length);
while (low < high) {
const cap = Math.ceil((low + high) / 2);
let total = 0;
for (const length of lengths) total += Math.min(length, cap);
if (total <= available) low = cap;
else high = cap - 1;
}
return low;
}
interface CutState {
stringCap: number;
remaining: number;
}
function cutString(text: string, cap: number): string {
if (text.length <= cap) return text;
// A partial data URL is useless to every reader, so drop it whole.
if (text.startsWith("data:")) return truncatedMarker(text.length);
return retainTextPrefix(text, cap) + truncatedMarker(text.length - cap);
}
function cut(value: unknown, depth: number, state: CutState): unknown {
state.remaining -= NODE_COST;
if (depth > MAX_DEPTH) return "…[truncated]";
if (typeof value === "string") {
const text = cutString(value, state.stringCap);
state.remaining -= text.length;
return text;
}
if (Array.isArray(value)) {
let changed = false;
const output: unknown[] = [];
for (const element of value) {
const remainingBefore = state.remaining;
const next = cut(element, depth + 1, state);
if (state.remaining < 0) {
// Drop this and later elements. A marker element would break typed arrays.
state.remaining = remainingBefore;
changed = true;
break;
}
changed ||= next !== element;
output.push(next);
}
return changed ? output : value;
}
if (typeof value === "object" && value !== null) {
let changed = false;
const output: Record<string, unknown> = {};
for (const [key, child] of Object.entries(value)) {
state.remaining -= key.length;
const next = cut(child, depth + 1, state);
changed ||= next !== child;
output[key] = next;
}
return changed ? output : value;
}
return value;
}
/** Returns the item itself when it fits its budget; otherwise a copy with its largest strings cut. */
export function limitAgentTimelineItemContent(item: AgentTimelineItem): AgentTimelineItem {
item = limitFailedShellError(item);
item = limitPlainText(item);
if (
item.type !== "tool_call" ||
item.detail.type !== "shell" ||
typeof item.detail.output !== "string"
) {
return item;
}
if (item.detail.output.length <= TOOL_CALL_CONTENT_MAX_LENGTH) {
return item;
}
return {
...item,
detail: {
...item.detail,
output: retainTextPrefix(item.detail.output, TOOL_CALL_CONTENT_MAX_LENGTH),
},
};
const budget = item.type === "tool_call" ? TOOL_CALL_BUDGET : ITEM_BUDGET;
const size: ItemSize = { nodeCost: 0, stringLengths: [], tooDeep: false };
measure(item, 0, size);
let stringTotal = 0;
for (const length of size.stringLengths) stringTotal += length;
if (!size.tooDeep && size.nodeCost + stringTotal <= budget) return item;
// Strings get what structure leaves, but at least half the budget, less a margin for JSON
// escaping. Arrays drop the elements that do not fit, which bounds items whose short strings
// or structure alone are too big.
const markerCost =
MARKER_COST * size.stringLengths.filter((length) => length > SHORT_STRING).length;
const available = (budget * 15) / 16 - Math.min(size.nodeCost, budget / 2) - markerCost;
const stringCap = Math.max(SHORT_STRING, findStringCap(size.stringLengths, available));
return cut(item, 0, { stringCap, remaining: budget }) as AgentTimelineItem;
}
@@ -181,6 +181,20 @@ describe("projected timeline retention", () => {
},
]);
});
it("serves a merged assistant message within the item budget", () => {
const store = new InMemoryAgentTimelineStore();
store.initialize("agent", { epoch: "e" });
const chunk = "a".repeat(200 * 1024);
for (let index = 0; index < 10; index++) {
store.append("agent", { type: "assistant_message", messageId: "m", text: chunk });
}
const [row] = store.fetch("agent", { limit: 1 }).rows;
expect(row).toMatchObject({ seqStart: 1, seqEnd: 10 });
expect(row.item).toMatchObject({ type: "assistant_message", messageId: "m" });
expect(Buffer.byteLength(JSON.stringify(row.item))).toBeLessThanOrEqual(256 * 1024);
expect(store.getLastAssistantMessage("agent")).toHaveLength(10 * chunk.length);
});
});
describe("projected sequence ownership", () => {
@@ -1,5 +1,6 @@
import { randomUUID } from "node:crypto";
import { TimelineProjection, type ProjectedTimelineRow } from "./timeline-projection.js";
import { limitAgentTimelineItemContent } from "./agent-timeline-content.js";
import type { AgentTimelineItem } from "./agent-sdk-types.js";
import type {
AgentTimelineFetchOptions,
@@ -120,7 +121,12 @@ export class InMemoryAgentTimelineStore {
hasNewer: page.hasNewer,
startSeq: page.startSeq,
endSeq: page.endSeq,
rows: page.entries.map((entry) => Object.assign({ seq: entry.seqEnd }, entry)),
// Stored rows are already within budget; this bounds assistant and reasoning merges.
rows: page.entries.map((entry) =>
Object.assign({ seq: entry.seqEnd }, entry, {
item: limitAgentTimelineItemContent(entry.item),
}),
),
};
}
@@ -80,7 +80,7 @@ describe("ProviderSubagentStore", () => {
const [row] = subagents.fetchTimeline("parent-a", "child-1").rows;
expect(row?.item).toMatchObject({
type: "tool_call",
detail: { type: "shell", output: "x".repeat(64 * 1024) },
detail: { type: "shell", output: expect.stringMatching(/^x+…\[truncated \d+\.\d KB\]$/) },
});
});
@@ -466,13 +466,13 @@ describe("background Claude subagents", () => {
type: "sub_agent",
subAgentType: "general-purpose",
description: "Count files here",
log: "",
log: "[Bash] ls -1",
},
});
});
test.each(["before", "after"])(
"keeps one Task card when child tools start %s the parent finishes",
"keeps one logged Task card when child tools start %s the parent finishes",
async (timing) => {
// A child can stream across the main session's turn boundary. Claude opens the next
// main-session turn with system/init; bare child frames do not open a turn.
@@ -605,7 +605,7 @@ describe("background Claude subagents", () => {
type: "sub_agent",
subAgentType: "general-purpose",
description: "Count files here",
log: "",
log: "[Bash] ls -1",
},
});
},
@@ -737,7 +737,7 @@ describe("ClaudeAgentSession persisted subagent replay", () => {
);
}, 120_000);
test("limits retained child shell output to the timeline's existing content limit", async () => {
test("limits retained child shell output to the timeline item budget", async () => {
const output = "x".repeat(128 * 1024);
writeSession({
parentLines: [taskToolUse(), taskToolResult()],
@@ -782,7 +782,10 @@ describe("ClaudeAgentSession persisted subagent replay", () => {
type: "timeline",
id: TOOL_USE_ID,
item: expect.objectContaining({
detail: expect.objectContaining({ type: "shell", output: output.slice(0, 64 * 1024) }),
detail: expect.objectContaining({
type: "shell",
output: expect.stringMatching(/^x+…\[truncated \d+\.\d KB\]$/),
}),
}),
}),
]);
@@ -273,7 +273,7 @@ describe("ClaudeAgentSession sub-agent sidechain updates", () => {
queryFactory.mockReset();
});
test("labels the Task card from its input and preserves callId lifecycle collapse", async () => {
test("accumulates lightweight sub_agent detail and preserves callId lifecycle collapse", async () => {
const session = await new ClaudeAgentClient({
logger,
queryFactory,
@@ -307,7 +307,9 @@ describe("ClaudeAgentSession sub-agent sidechain updates", () => {
expect(latest.detail.subAgentType).toBe("Explore");
expect(latest.detail.description).toBe("Inspect repository structure");
expect(latest.detail.log).toBe("");
expect(latest.detail.log).toContain("[Read] README.md");
expect(latest.detail.log).toContain("[Edit] src/index.ts");
expect(latest.detail.log).not.toContain("VERY_LARGE_OLD_STRING");
const rows: AgentTimelineRow[] = timelineToolCalls.map((item, index) => ({
seq: index + 1,
@@ -394,7 +396,10 @@ describe("ClaudeAgentSession sub-agent sidechain updates", () => {
.map((event) => event.item)
.at(-1);
expect(latestSubAgentUpdate?.detail).toMatchObject({ type: "sub_agent", log: "" });
expect(latestSubAgentUpdate?.detail).toMatchObject({
type: "sub_agent",
log: expect.stringContaining("[Read] README.md"),
});
});
test("routes a resumed SendMessage sidechain to the original child without a second Task card", async () => {
@@ -548,7 +553,7 @@ describe("ClaudeAgentSession sub-agent sidechain updates", () => {
).toMatchObject({ type: "upsert", id: "task-tail-1", status: "failed" });
});
test("keeps the Task card the same size while the child runs", async () => {
test("tails sub-agent actions instead of dropping latest entries at cap", async () => {
queryFactory.mockImplementation(() => buildQueryMock(buildTailScenarioEvents(205)));
const session = await new ClaudeAgentClient({
@@ -563,27 +568,24 @@ describe("ClaudeAgentSession sub-agent sidechain updates", () => {
const events = await collectUntilTerminal(streamSession(session, "delegate work"));
await session.close();
const runningCards = events
.flatMap((event) =>
event.type === "timeline" &&
event.item.type === "tool_call" &&
event.item.callId === "task-tail-1" &&
event.item.status === "running" &&
event.item.detail.type === "sub_agent"
? [event.item]
: [],
const timelineToolCalls = events
.filter(
(event): event is Extract<AgentStreamEvent, { type: "timeline" }> =>
event.type === "timeline" && event.item.type === "tool_call",
)
.map((item) => JSON.stringify(item));
expect(runningCards).toHaveLength(1);
.map((event) => event.item)
.filter((item) => item.callId === "task-tail-1");
const subAgentUpdates = timelineToolCalls.filter((item) => item.detail.type === "sub_agent");
const latest = subAgentUpdates[subAgentUpdates.length - 1];
expect(latest).toBeDefined();
if (!latest || latest.detail.type !== "sub_agent") {
throw new Error("expected sub_agent detail");
}
const childToolCalls = events.filter(
(event) =>
event.type === "provider_subagent" &&
event.event.type === "timeline" &&
event.event.id === "task-tail-1" &&
event.event.item.type === "tool_call",
);
expect(childToolCalls).toHaveLength(205);
expect(latest.detail.log).not.toContain("[Read] file-1.md");
expect(latest.detail.log).not.toContain("[Read] file-5.md");
expect(latest.detail.log).toContain("[Read] file-6.md");
expect(latest.detail.log).toContain("[Read] file-205.md");
});
test("keeps a long main-thread tool call out of the subagents track", async () => {
@@ -4667,7 +4667,7 @@ class ClaudeAgentSession implements AgentSession {
/**
* The Task card in the parent transcript, built from the declaration.
*
* The sidechain tracker also emits this card when it first sees the child, and the two
* The sidechain tracker also emits this card, enriched with the child's action log, and the two
* collapse on the shared tool-call id. Emitting it at declaration matters because a
* backgrounded subagent produces no sidechain frames at all — verified on the wire — so the
* tracker never runs for one and its card would otherwise stay an unlabeled "Task".
@@ -6276,7 +6276,6 @@ function buildClaudeSubagentCardDetail(
type: "sub_agent",
...(facts.title ? { subAgentType: facts.title } : {}),
...(facts.description ? { description: facts.description } : {}),
// COMPAT(subagent-log): remove after 2027-04-30.
log: "",
actions: [],
};
@@ -4,7 +4,7 @@ import type { SDKMessage } from "@anthropic-ai/claude-agent-sdk";
import { ClaudeSidechainTracker } from "./sidechain-tracker.js";
describe("ClaudeSidechainTracker", () => {
it("keeps a background child's card across the parent finishing", () => {
it("retains a background child's actions across the parent finishing", () => {
const tracker = new ClaudeSidechainTracker({
getToolInput: () => ({ subagent_type: "general-purpose", description: "List files" }),
isDescriptorOwnedElsewhere: () => true,
@@ -15,67 +15,27 @@ describe("ClaudeSidechainTracker", () => {
parent_tool_use_id: "task-1",
message: { content: [{ type: "tool_use", id, name: "Bash", input: { command } }] },
}) as unknown as SDKMessage;
const parentCards = (events: ReturnType<typeof tracker.handleMessage>) =>
events.filter((event) => event.type === "timeline");
const childCallIds = (events: ReturnType<typeof tracker.handleMessage>) =>
events.flatMap((event) =>
event.type === "provider_subagent" &&
event.event.type === "timeline" &&
event.event.item.type === "tool_call"
? [event.event.item.callId]
: [],
);
expect(parentCards(tracker.handleMessage(action("ls", "ls -1"), "task-1"))).toHaveLength(1);
tracker.handleMessage(action("ls", "ls -1"), "task-1");
expect(tracker.finishAll("completed")).toEqual([]);
const afterFinish = tracker.handleMessage(action("pwd", "pwd"), "task-1");
expect(parentCards(afterFinish)).toEqual([]);
expect(childCallIds(afterFinish)).toEqual(["pwd"]);
const events = tracker.handleMessage(action("pwd", "pwd"), "task-1");
expect(events).toContainEqual(
expect.objectContaining({
type: "timeline",
item: expect.objectContaining({
detail: expect.objectContaining({ log: "[Bash] ls -1\n[Bash] pwd" }),
}),
}),
);
tracker.delete("task-1");
const next = tracker.handleMessage(action("next", "ls"), "task-1");
expect(parentCards(next)).toHaveLength(1);
expect(childCallIds(next)).toEqual(["next"]);
});
it("shows a streamed child tool call's input once the assistant message brings it", () => {
const tracker = new ClaudeSidechainTracker({ getToolInput: () => null });
const frames = [
{
type: "stream_event",
parent_tool_use_id: "task-1",
event: {
type: "content_block_start",
index: 0,
content_block: { type: "tool_use", id: "bash-1", name: "Bash", input: {} },
},
},
{
type: "assistant",
parent_tool_use_id: "task-1",
message: {
content: [
{ type: "tool_use", id: "bash-1", name: "Bash", input: { command: "echo child" } },
],
},
},
] as unknown as SDKMessage[];
const childCards = frames
.flatMap((frame) => tracker.handleMessage(frame, "task-1"))
.flatMap((event) =>
event.type === "provider_subagent" &&
event.event.type === "timeline" &&
event.event.item.type === "tool_call"
? [event.event.item]
: [],
);
expect(childCards.at(-1)).toMatchObject({
callId: "bash-1",
status: "running",
detail: { type: "shell", command: "echo child" },
});
expect(next).toContainEqual(
expect.objectContaining({
type: "timeline",
item: expect.objectContaining({
detail: expect.objectContaining({ log: "[Bash] ls" }),
}),
}),
);
});
it("uses Claude's native agent name for the provider subagent title", () => {
@@ -5,6 +5,7 @@ import {
mapClaudeFailedToolCall,
mapClaudeRunningToolCall,
} from "./tool-call-mapper.js";
import { buildToolCallDisplayModel } from "@getpaseo/protocol/tool-call-display";
import {
isClaudeSubagentHandbackToolName,
readClaudeSubagentHandback,
@@ -19,17 +20,21 @@ interface ClaudeContentChunk {
type ClaudeToolCallItem = Extract<AgentTimelineItem, { type: "tool_call" }>;
interface SubAgentChildCall {
interface SubAgentActionEntry {
index: number;
toolName: string;
input: unknown;
summary?: string;
}
interface SubAgentActivityState {
name?: string;
subAgentType?: string;
description?: string;
/** The child's tool calls awaiting results, which carry no input of their own. */
runningCalls: Map<string, SubAgentChildCall>;
actions: SubAgentActionEntry[];
actionKeys: string[];
nextActionIndex: number;
actionIndexByKey: Map<string, number>;
completedActionKeys: Set<string>;
}
@@ -39,6 +44,7 @@ interface SubAgentActionCandidate {
input: unknown;
}
const MAX_SUB_AGENT_LOG_ENTRIES = 200;
const MAX_SUB_AGENT_SUMMARY_CHARS = 160;
function readTrimmedString(value: unknown): string | undefined {
@@ -73,7 +79,8 @@ export class ClaudeSidechainTracker {
needsSyntheticParentToolCard?: (toolUseId: string) => boolean;
/**
* The parent's tool call once Claude has answered it. A background child streams its frames
* after that answer, so its card keeps the settled status and output.
* after that answer, so its card keeps the settled status and output and gains the child's
* action log.
*/
getSettledParentToolCall?: (toolUseId: string) => ClaudeToolCallItem | undefined;
}) {
@@ -84,27 +91,28 @@ export class ClaudeSidechainTracker {
}
handleMessage(message: SDKMessage, parentToolUseId: string): AgentStreamEvent[] {
const existingState = this.activeSidechains.get(parentToolUseId);
const state =
existingState ??
this.activeSidechains.get(parentToolUseId) ??
({
runningCalls: new Map<string, SubAgentChildCall>(),
actions: [],
actionKeys: [],
nextActionIndex: 1,
actionIndexByKey: new Map<string, number>(),
completedActionKeys: new Set<string>(),
} satisfies SubAgentActivityState);
this.activeSidechains.set(parentToolUseId, state);
// The parent card is a launch marker: it changes only when the child is first seen or the
// task input names it differently, never with the child's activity.
const cardUpdated =
this.updateSubAgentContextFromTaskInput(state, parentToolUseId) || !existingState;
const contextUpdated = this.updateSubAgentContextFromTaskInput(state, parentToolUseId);
const actionCandidates = this.extractSubAgentActionCandidates(message);
const childTimelineItems = [
...this.extractSubAgentTimelineItems(message, state),
...this.extractSubAgentToolResults(message, state),
];
let actionUpdated = false;
for (const action of actionCandidates) {
if (state.completedActionKeys.has(action.key)) continue;
if (this.trackChildCall(state, action)) {
if (this.appendSubAgentAction(state, action)) {
actionUpdated = true;
const toolCall = mapClaudeRunningToolCall({
name: action.toolName,
callId: action.key,
@@ -117,7 +125,7 @@ export class ClaudeSidechainTracker {
}
}
if (!cardUpdated && childTimelineItems.length === 0) {
if (!contextUpdated && !actionUpdated && childTimelineItems.length === 0) {
return [];
}
@@ -137,8 +145,11 @@ export class ClaudeSidechainTracker {
type: "sub_agent",
...(state.subAgentType ? { subAgentType: state.subAgentType } : {}),
...(state.description ? { description: state.description } : {}),
// COMPAT(subagent-log): remove after 2027-04-30.
log: "",
log: state.actions
.map((action) =>
action.summary ? `[${action.toolName}] ${action.summary}` : `[${action.toolName}]`,
)
.join("\n"),
actions: [],
};
@@ -159,19 +170,18 @@ export class ClaudeSidechainTracker {
},
];
const parentCard: AgentStreamEvent[] =
cardUpdated && this.needsSyntheticParentToolCard(parentToolUseId)
? [
{
type: "timeline",
item: {
...toolCall,
detail,
},
provider: "claude",
const parentCard: AgentStreamEvent[] = this.needsSyntheticParentToolCard(parentToolUseId)
? [
{
type: "timeline",
item: {
...toolCall,
detail,
},
]
: [];
provider: "claude",
},
]
: [];
return [
...descriptorEvents,
@@ -293,7 +303,8 @@ export class ClaudeSidechainTracker {
if (!isClaudeContentChunk(block) || !block.type.endsWith("tool_result")) continue;
const callId = readTrimmedString(block.tool_use_id);
if (!callId || state.completedActionKeys.has(callId)) continue;
const action = state.runningCalls.get(callId);
const actionIndex = state.actionIndexByKey.get(callId);
const action = actionIndex === undefined ? undefined : state.actions[actionIndex];
const toolName = action?.toolName ?? readTrimmedString(block.tool_name);
if (!toolName) continue;
const params = {
@@ -306,7 +317,6 @@ export class ClaudeSidechainTracker {
? mapClaudeFailedToolCall({ ...params, error: block })
: mapClaudeCompletedToolCall(params);
if (toolCall) {
state.runningCalls.delete(callId);
state.completedActionKeys.add(callId);
items.push(toolCall);
}
@@ -402,12 +412,13 @@ export class ClaudeSidechainTracker {
const key =
readTrimmedString(block.id) ??
`stream:${block.name}:${typeof event.index === "number" ? event.index : 0}`;
// A streamed call starts with `{}` until its input arrives, so an empty input is not final.
const input =
block.input && typeof block.input === "object" && Object.keys(block.input).length > 0
? block.input
: null;
return [{ key, toolName: block.name, input }];
return [
{
key,
toolName: block.name,
input: block.input ?? null,
},
];
}
private extractSubAgentActionCandidates(message: SDKMessage): SubAgentActionCandidate[] {
@@ -437,20 +448,84 @@ export class ClaudeSidechainTracker {
return [];
}
/** Returns whether the call is new or has gained its input, so its running card changes. */
private trackChildCall(
private appendSubAgentAction(
state: SubAgentActivityState,
candidate: SubAgentActionCandidate,
): boolean {
const toolName = readTrimmedString(candidate.toolName);
if (!toolName) {
const normalizedToolName = readTrimmedString(candidate.toolName);
if (!normalizedToolName) {
return false;
}
const existing = state.runningCalls.get(candidate.key);
if (existing && (existing.input !== null || candidate.input === null)) {
return false;
const summary = this.deriveSubAgentActionSummary(normalizedToolName, candidate.input);
const existingIndex = state.actionIndexByKey.get(candidate.key);
if (existingIndex !== undefined) {
const existing = state.actions[existingIndex];
if (!existing) {
return false;
}
const nextSummary = existing.summary ?? summary;
if (existing.toolName === normalizedToolName && existing.summary === nextSummary) {
return false;
}
state.actions[existingIndex] = {
...existing,
toolName: normalizedToolName,
input: existing.input ?? candidate.input,
...(nextSummary ? { summary: nextSummary } : {}),
};
return true;
}
state.runningCalls.set(candidate.key, { toolName, input: candidate.input });
state.actions.push({
index: state.nextActionIndex,
toolName: normalizedToolName,
input: candidate.input,
...(summary ? { summary } : {}),
});
state.nextActionIndex += 1;
state.actionKeys.push(candidate.key);
this.trimSubAgentTail(state);
this.rebuildSubAgentActionIndex(state);
return true;
}
private trimSubAgentTail(state: SubAgentActivityState): void {
while (state.actions.length > MAX_SUB_AGENT_LOG_ENTRIES) {
state.actions.shift();
const removedKey = state.actionKeys.shift();
if (removedKey) state.completedActionKeys.delete(removedKey);
}
}
private rebuildSubAgentActionIndex(state: SubAgentActivityState): void {
state.actionIndexByKey.clear();
for (let index = 0; index < state.actionKeys.length; index += 1) {
const key = state.actionKeys[index];
if (key) {
state.actionIndexByKey.set(key, index);
}
}
}
private deriveSubAgentActionSummary(toolName: string, input: unknown): string | undefined {
const runningToolCall = mapClaudeRunningToolCall({
name: toolName,
callId: `sub-agent-summary-${toolName}`,
input,
output: null,
});
if (!runningToolCall) {
return undefined;
}
const display = buildToolCallDisplayModel({
name: runningToolCall.name,
status: runningToolCall.status,
error: runningToolCall.error,
detail: runningToolCall.detail,
metadata: runningToolCall.metadata,
});
return this.normalizeSubAgentText(display.summary);
}
}
@@ -77,19 +77,12 @@ CHILD_SENTINEL in your own response.
const completedChildIndex = events.findIndex(
(event) =>
event.type === "provider_subagent" &&
event.event.type === "upsert" &&
event.event.status === "completed",
event.type === "timeline" &&
event.item.type === "tool_call" &&
event.item.status === "completed" &&
event.item.detail.type === "sub_agent" &&
event.item.detail.log.includes("CHILD_SENTINEL"),
);
const childAssistantText = events
.flatMap((event) =>
event.type === "provider_subagent" &&
event.event.type === "timeline" &&
event.event.item.type === "assistant_message"
? [event.event.item.text]
: [],
)
.join("");
const rootMessageIndexes = events.flatMap((event, index) =>
event.type === "timeline" && event.item.type === "assistant_message" ? [index] : [],
);
@@ -112,7 +105,6 @@ CHILD_SENTINEL in your own response.
.map((item) => item.text)
.join("");
expect(topLevelAssistantText).not.toContain("CHILD_SENTINEL");
expect(childAssistantText).toContain("CHILD_SENTINEL");
expect(completedChildIndex).toBeGreaterThanOrEqual(0);
expect(rootMessageIndexes[0]).toBeGreaterThan(completedChildIndex);
expect(terminalIndexes).toHaveLength(1);
@@ -3524,7 +3524,7 @@ describe("Codex app-server provider", () => {
]);
});
test("keeps child-thread Codex activity out of the parent sub-agent card", async () => {
test("folds child-thread Codex activity into the parent sub-agent tool call", async () => {
const session = createSession();
const events: AgentStreamEvent[] = [];
session.subscribe((event) => events.push(event));
@@ -3562,6 +3562,7 @@ describe("Codex app-server provider", () => {
});
const timelineEvents = events.filter((event) => event.type === "timeline");
expect(timelineEvents).toHaveLength(4);
expect(timelineEvents.every((event) => event.item.type === "tool_call")).toBe(true);
const finalItem = timelineEvents.at(-1)?.item;
expect(finalItem).toMatchObject({
@@ -3573,7 +3574,7 @@ describe("Codex app-server provider", () => {
type: "sub_agent",
subAgentType: "Sub-agent",
description: "Report findings.",
log: "",
log: "[Assistant] Found the path.",
actions: [],
},
});
@@ -3650,9 +3651,13 @@ describe("Codex app-server provider", () => {
id: "settled-child-thread",
status: "completed",
});
expect(events.findLast((event) => event.type === "timeline")).toMatchObject({
expect(events.at(-1)).toMatchObject({
type: "timeline",
item: { callId: "call-settled-child", status: "completed" },
item: {
callId: "call-settled-child",
status: "completed",
detail: { type: "sub_agent", log: "[Assistant] Late trailing output." },
},
});
appServer.completesSubAgentActivity({
@@ -3957,8 +3962,6 @@ describe("Codex app-server provider", () => {
createTestLogger(),
async () => appServer.child,
);
const events: AgentStreamEvent[] = [];
session.subscribe((event) => events.push(event));
try {
const resultPromise = session.run("Delegate the investigation, then report the result.");
@@ -4002,18 +4005,11 @@ describe("Codex app-server provider", () => {
type: "tool_call",
callId: "spawn-child-early",
status: "completed",
detail: { type: "sub_agent", log: "" },
detail: {
type: "sub_agent",
log: "[Assistant] Early child findings.",
},
});
expect(events).toContainEqual(
expect.objectContaining({
type: "provider_subagent",
event: expect.objectContaining({
type: "timeline",
id: "child-thread-early",
item: expect.objectContaining({ type: "assistant_message" }),
}),
}),
);
appServer.assertNoErrors();
} finally {
await session.close();
@@ -4084,7 +4080,10 @@ describe("Codex app-server provider", () => {
expect(toolCalls.at(-1)).toMatchObject({
callId: "spawn-child-stable",
status: "canceled",
detail: { type: "sub_agent", log: "" },
detail: {
type: "sub_agent",
log: "[Assistant] Still on the same card.",
},
});
appServer.assertNoErrors();
} finally {
@@ -4164,10 +4163,11 @@ describe("Codex app-server provider", () => {
const result = await resultPromise;
const toolCalls = result.timeline.filter((item) => item.type === "tool_call");
expect(toolCalls.map((item) => item.status)).toEqual(["running", "completed"]);
expect(toolCalls.map((item) => item.status)).toEqual(["running", "completed", "completed"]);
expect(toolCalls.at(-1)).toMatchObject({
callId: "spawn-child-late-compaction",
status: "completed",
detail: { type: "sub_agent", log: "[Compacted]" },
});
appServer.assertNoErrors();
} finally {
@@ -4175,7 +4175,7 @@ describe("Codex app-server provider", () => {
}
});
test("routes legacy child tools to the child timeline", async () => {
test("projects legacy child tools into one stable sub-agent log", async () => {
const appServer = createFakeCodexAppServer();
const session = new CodexAppServerAgentSession(
createConfig({ cwd: "/workspace/project" }),
@@ -4183,8 +4183,6 @@ describe("Codex app-server provider", () => {
createTestLogger(),
async () => appServer.child,
);
const events: AgentStreamEvent[] = [];
session.subscribe((event) => events.push(event));
try {
const resultPromise = session.run("Delegate the implementation.");
@@ -4227,26 +4225,19 @@ describe("Codex app-server provider", () => {
expect(finalToolCall).toMatchObject({
callId: "spawn-legacy-tool-child",
status: "completed",
detail: { type: "sub_agent", log: "" },
detail: { type: "sub_agent" },
});
const childToolCalls = events.flatMap((event) =>
event.type === "provider_subagent" &&
event.event.type === "timeline" &&
event.event.id === "legacy-tool-child" &&
event.event.item.type === "tool_call"
? [event.event.item]
: [],
);
expect(new Set(childToolCalls.map((item) => item.detail.type))).toEqual(
new Set(["shell", "edit"]),
);
if (finalToolCall?.detail.type === "sub_agent") {
expect(finalToolCall.detail.log.match(/\[Shell\]/g)).toHaveLength(1);
expect(finalToolCall.detail.log).toContain("[Edit]");
}
appServer.assertNoErrors();
} finally {
await session.close();
}
});
test("keeps nested MultiAgentV2 output out of the root sub-agent card", async () => {
test("keeps nested MultiAgentV2 output inside the root sub-agent card", async () => {
const session = createSession();
const events: AgentStreamEvent[] = [];
session.subscribe((event) => events.push(event));
@@ -4306,22 +4297,8 @@ describe("Codex app-server provider", () => {
expect(beforeParentCompletes.at(-1)).toMatchObject({
callId: "spawn-child-root",
status: "running",
detail: { type: "sub_agent", log: "" },
detail: { type: "sub_agent", log: expect.stringContaining("Grandchild findings.") },
});
expect(events).toContainEqual(
expect.objectContaining({
type: "provider_subagent",
event: {
type: "timeline",
id: "grandchild-thread",
item: {
type: "assistant_message",
messageId: "grandchild-message",
text: "Grandchild findings.",
},
},
}),
);
await asInternals(session).handleNotification("turn/completed", {
threadId: "child-thread-root",
@@ -4388,20 +4365,13 @@ describe("Codex app-server provider", () => {
"child-thread-root",
"child-thread-root",
]);
expect(events).toContainEqual(
expect.objectContaining({
type: "provider_subagent",
event: {
type: "timeline",
id: "grandchild-thread",
item: {
type: "assistant_message",
messageId: "grandchild-after-wait",
text: "Still nested.",
},
},
}),
);
expect(events.at(-1)).toMatchObject({
type: "timeline",
item: {
callId: "spawn-child-root",
detail: { type: "sub_agent", log: expect.stringContaining("Still nested.") },
},
});
});
test("never treats an unmapped foreign terminal as the root terminal", async () => {
@@ -4821,7 +4791,7 @@ describe("Codex app-server provider", () => {
expect(events).toHaveLength(eventCountAfterSpawn);
});
test("keeps child context compaction on the child timeline", async () => {
test("keeps child context compaction inside the child card", async () => {
const session = createSession();
const events: AgentStreamEvent[] = [];
session.subscribe((event) => events.push(event));
@@ -4854,70 +4824,10 @@ describe("Codex app-server provider", () => {
(item) => item.type === "tool_call" && item.callId === "spawn-child-compaction",
),
).toBe(true);
expect(
events.flatMap((event) =>
event.type === "provider_subagent" && event.event.type === "timeline" ? [event.event] : [],
),
).toEqual([
{
type: "timeline",
id: "child-thread-compaction",
item: { type: "compaction", status: "loading" },
},
{
type: "timeline",
id: "child-thread-compaction",
item: { type: "compaction", status: "completed" },
},
]);
});
test("settles a child compaction once when the item completion is missing or duplicated", async () => {
const session = createSession();
const events: AgentStreamEvent[] = [];
session.subscribe((event) => events.push(event));
for (const child of ["compacted-child", "turn-ended-child"]) {
await asInternals(session).handleNotification("item/completed", {
threadId: "test-thread",
item: {
type: "subAgentActivity",
id: `spawn-${child}`,
kind: "started",
agentThreadId: child,
agentPath: `/root/${child}`,
},
});
await asInternals(session).handleNotification("item/started", {
threadId: child,
item: { type: "contextCompaction", id: `${child}-compaction` },
});
}
await asInternals(session).handleNotification("thread/compacted", {
threadId: "compacted-child",
expect(timelineItems.at(-1)).toMatchObject({
type: "tool_call",
detail: { type: "sub_agent", log: "[Compacted]" },
});
await asInternals(session).handleNotification("item/completed", {
threadId: "compacted-child",
item: { type: "contextCompaction", id: "compacted-child-compaction" },
});
await asInternals(session).handleNotification("turn/completed", {
threadId: "turn-ended-child",
turn: { status: "completed" },
});
const compactions = events.flatMap((event) =>
event.type === "provider_subagent" && event.event.type === "timeline"
? [
`${event.event.id} ${event.event.item.type === "compaction" ? event.event.item.status : ""}`,
]
: [],
);
expect(compactions).toEqual([
"compacted-child loading",
"turn-ended-child loading",
"compacted-child completed",
"turn-ended-child completed",
]);
});
test("keeps the parent sub-agent running when a child command fails during the child turn", async () => {
@@ -4952,7 +4862,7 @@ describe("Codex app-server provider", () => {
},
});
expect(events.findLast((event) => event.type === "timeline")?.item).toMatchObject({
expect(events.at(-1)?.item).toMatchObject({
type: "tool_call",
callId: "call-sub-agent-child-command-failure",
name: "Sub-agent",
@@ -5247,7 +5157,7 @@ describe("Codex app-server provider", () => {
);
expect(liveToolCalls.at(-1)).toMatchObject({
status: "running",
detail: { type: "sub_agent", log: "" },
detail: { type: "sub_agent", log: "[Assistant] More findings after resume." },
});
liveEvents.length = 0;
@@ -5260,12 +5170,12 @@ describe("Codex app-server provider", () => {
itemId: "legacy-child-message-after-resume",
delta: "Legacy findings after resume.",
});
expect(liveEvents.findLast((event) => event.type === "timeline")).toMatchObject({
expect(liveEvents.at(-1)).toMatchObject({
type: "timeline",
item: {
callId: "legacy-spawn-history",
status: "running",
detail: { type: "sub_agent", log: "" },
detail: { type: "sub_agent", log: "[Assistant] Legacy findings after resume." },
},
});
});
@@ -49,10 +49,10 @@ import { Dirent } from "node:fs";
import fs from "node:fs/promises";
import os from "node:os";
import path from "node:path";
import { isDeepStrictEqual } from "node:util";
import { z } from "zod";
import { renderPromptAttachmentAsText } from "../prompt-attachments.js";
import { composeSystemPromptParts } from "../system-prompt.js";
import { curateAgentActivity } from "../activity-curator.js";
import { CodexAsyncQuestions, codexAsyncQuestionToTimeline } from "./codex/async-questions.js";
import {
mapCodexToolCallEnvelope,
@@ -3384,6 +3384,8 @@ interface CodexSubAgentCallState {
activityItemIds: Set<string>;
pendingCommandOutputDeltas: Map<string, string[]>;
pendingFileChangeOutputDeltas: Map<string, string[]>;
childItemOrder: string[];
childItems: Map<string, AgentTimelineItem>;
childThreadIds: Set<string>;
}
@@ -3485,9 +3487,6 @@ export class CodexAppServerAgentSession implements AgentSession {
private compactionTriggerByItemId = new Map<string, "auto" | "manual">();
private pendingRootCompactionItemIds = new Set<string>();
private pendingAnonymousRootCompactions = 0;
// Child threads with a compaction shown as loading. Some Codex builds settle it only through
// `thread/compacted` or the end of the child's turn.
private pendingChildCompactionThreadIds = new Set<string>();
// Codex can report one completed compaction through both channels:
// `thread/compacted` and a completed `contextCompaction` item.
private unpairedCompactionNotificationCompletions = 0;
@@ -5517,9 +5516,6 @@ export class CodexAppServerAgentSession implements AgentSession {
case "patch_apply_completed":
this.handlePatchApplyCompletedNotification(parsed, callId);
return;
case "context_compacted":
this.settleChildCompaction(parsed.threadId);
return;
default:
// Aggregate child telemetry is redundant and must not leak into the
// root timeline. Concrete legacy tools are projected above for Codex
@@ -5696,10 +5692,20 @@ export class CodexAppServerAgentSession implements AgentSession {
activityItemIds: new Set<string>(),
pendingCommandOutputDeltas: new Map<string, string[]>(),
pendingFileChangeOutputDeltas: new Map<string, string[]>(),
childItemOrder: [],
childItems: new Map<string, AgentTimelineItem>(),
childThreadIds: new Set<string>(),
} satisfies CodexSubAgentCallState);
state.toolCall = timelineItem;
state.toolCall = {
...timelineItem,
detail: {
...timelineItem.detail,
log:
timelineItem.detail.log ||
(state.toolCall.detail.type === "sub_agent" ? state.toolCall.detail.log : ""),
},
};
state.parentCallId ??= parentCallId;
state.parentSubagentId ??= parentSubagentId;
const activity = readCodexSubAgentActivity(rawItem);
@@ -5752,17 +5758,25 @@ export class CodexAppServerAgentSession implements AgentSession {
const activityToolCall = mapCodexToolCallFromThreadItem(rawItem, {
cwd: this.config.cwd ?? null,
});
const subAgentType =
activityToolCall?.detail.type === "sub_agent"
? activityToolCall.detail.subAgentType
: undefined;
if (
activityToolCall?.detail.type === "sub_agent" &&
state.toolCall.detail.type === "sub_agent"
) {
state.toolCall = {
...state.toolCall,
detail: {
...state.toolCall.detail,
subAgentType: activityToolCall.detail.subAgentType,
},
};
}
let nextStatus: ToolCallTimelineItem["status"] | undefined = "running";
if (activity.kind === "interrupted") {
nextStatus = "canceled";
} else if (isTerminalSubAgentStatus(state.toolCall.status)) {
nextStatus = undefined;
}
this.emitSubAgentActivityUpdate(callId, nextStatus, { subAgentType });
this.emitSubAgentActivityUpdate(callId, nextStatus);
return true;
}
@@ -5803,13 +5817,7 @@ export class CodexAppServerAgentSession implements AgentSession {
): boolean {
if (
childSubAgentCallId &&
parsed.threadId &&
this.handleSubAgentContextCompactionItem(
childSubAgentCallId,
parsed.threadId,
parsed.item,
"completed",
)
this.handleSubAgentContextCompactionItem(childSubAgentCallId, parsed.item, "completed")
) {
return true;
}
@@ -5819,6 +5827,17 @@ export class CodexAppServerAgentSession implements AgentSession {
);
}
private upsertSubAgentChildItem(callId: string, itemId: string, item: AgentTimelineItem): void {
const state = this.subAgentCallsByCallId.get(callId);
if (!state) {
return;
}
if (!state.childItems.has(itemId)) {
state.childItemOrder.push(itemId);
}
state.childItems.set(itemId, item);
}
private emitCodexToolTimelineItem(
timelineItem: ToolCallTimelineItem,
subAgentCallId: string | null,
@@ -5828,6 +5847,7 @@ export class CodexAppServerAgentSession implements AgentSession {
this.emitEvent({ type: "timeline", provider: CODEX_PROVIDER, item: timelineItem });
return;
}
this.upsertSubAgentChildItem(subAgentCallId, timelineItem.callId, timelineItem);
const state = this.subAgentCallsByCallId.get(subAgentCallId);
if (state) {
const targetThreadIds = childThreadId ? [childThreadId] : state.childThreadIds;
@@ -5841,15 +5861,26 @@ export class CodexAppServerAgentSession implements AgentSession {
);
}
private getSubAgentChildTimeline(state: CodexSubAgentCallState): AgentTimelineItem[] {
return state.childItemOrder
.map((itemId) => state.childItems.get(itemId))
.filter((item): item is AgentTimelineItem => Boolean(item));
}
private emitSubAgentActivityUpdate(
callId: string,
status?: ToolCallTimelineItem["status"],
options?: { reopen?: boolean; subAgentType?: string },
options?: { reopen?: boolean },
): void {
const state = this.subAgentCallsByCallId.get(callId);
if (!state || state.toolCall.detail.type !== "sub_agent") {
return;
}
const childTimeline = this.getSubAgentChildTimeline(state);
const log =
childTimeline.length > 0
? curateAgentActivity(childTimeline, { labelAssistantMessages: true })
: "";
let resolvedStatus = status ?? state.toolCall.status;
if (
status === "running" &&
@@ -5858,12 +5889,16 @@ export class CodexAppServerAgentSession implements AgentSession {
) {
resolvedStatus = state.toolCall.status;
}
const baseToolCall = options?.subAgentType
? {
...state.toolCall,
detail: { ...state.toolCall.detail, subAgentType: options.subAgentType },
}
: state.toolCall;
for (const childThreadId of state.childThreadIds) {
this.emitProviderSubagentUpsert(childThreadId, state, resolvedStatus);
}
const baseToolCall = {
...state.toolCall,
detail: {
...state.toolCall.detail,
log,
},
};
const nextToolCall: ToolCallTimelineItem =
resolvedStatus === "failed"
? {
@@ -5876,14 +5911,10 @@ export class CodexAppServerAgentSession implements AgentSession {
status: resolvedStatus,
error: null,
};
// The card is a launch marker: the child's activity streams to its own timeline, so the card
// is re-sent only when its status or label changes.
const cardChanged = !isDeepStrictEqual(nextToolCall, state.toolCall);
state.toolCall = nextToolCall;
for (const childThreadId of state.childThreadIds) {
this.emitProviderSubagentUpsert(childThreadId, state, resolvedStatus);
}
if (!cardChanged || (state.parentCallId && state.parentCallId !== callId)) {
if (state.parentCallId && state.parentCallId !== callId) {
this.upsertSubAgentChildItem(state.parentCallId, state.callId, nextToolCall);
this.emitSubAgentActivityUpdate(state.parentCallId);
return;
}
this.emitEvent({ type: "timeline", provider: CODEX_PROVIDER, item: nextToolCall });
@@ -5991,6 +6022,7 @@ export class CodexAppServerAgentSession implements AgentSession {
): void {
this.applyBufferedDeltaTextToTimelineItem(timelineItem, itemId);
if (itemId) {
this.upsertSubAgentChildItem(callId, itemId, timelineItem);
this.pendingAgentMessages.delete(itemId);
this.pendingReasoning.delete(itemId);
this.pendingCommandOutputDeltas.delete(itemId);
@@ -6001,29 +6033,19 @@ export class CodexAppServerAgentSession implements AgentSession {
private handleSubAgentContextCompactionItem(
callId: string,
childThreadId: string,
item: { id?: string; type?: string; [key: string]: unknown },
status: "loading" | "completed",
): boolean {
if (!this.isContextCompactionItem(item)) {
return false;
}
if (status === "loading") {
this.pendingChildCompactionThreadIds.add(childThreadId);
this.emitProviderSubagentTimeline(childThreadId, { type: "compaction", status });
} else {
this.settleChildCompaction(childThreadId);
if (item.id) {
this.upsertSubAgentChildItem(callId, item.id, { type: "compaction", status });
}
this.emitSubAgentActivityUpdate(callId);
return true;
}
private settleChildCompaction(childThreadId: string): void {
if (this.pendingChildCompactionThreadIds.delete(childThreadId)) {
this.emitProviderSubagentTimeline(childThreadId, { type: "compaction", status: "completed" });
}
}
private shouldSkipCompletedThreadItem(
timelineItem: AgentTimelineItem,
normalizedItemType: string | undefined,
@@ -6054,6 +6076,11 @@ export class CodexAppServerAgentSession implements AgentSession {
text: parsed.delta,
});
}
this.upsertSubAgentChildItem(subAgentCallId, parsed.itemId, {
type: "assistant_message",
messageId: parsed.itemId,
text,
});
this.emitSubAgentActivityUpdate(subAgentCallId, "running");
return;
}
@@ -6087,6 +6114,10 @@ export class CodexAppServerAgentSession implements AgentSession {
text: parsed.delta,
});
}
this.upsertSubAgentChildItem(subAgentCallId, parsed.itemId, {
type: "reasoning",
text: prev.join(""),
});
this.emitSubAgentActivityUpdate(subAgentCallId, "running");
return;
}
@@ -6154,7 +6185,6 @@ export class CodexAppServerAgentSession implements AgentSession {
): void {
const subAgentCallId = this.getSubAgentCallIdForThread(parsed.threadId);
if (subAgentCallId) {
if (parsed.threadId) this.settleChildCompaction(parsed.threadId);
let status: ToolCallTimelineItem["status"] = "completed";
if (parsed.status === "failed") {
status = "failed";
@@ -6741,13 +6771,7 @@ export class CodexAppServerAgentSession implements AgentSession {
const childSubAgentCallId = this.getSubAgentCallIdForThread(parsed.threadId);
if (
childSubAgentCallId &&
parsed.threadId &&
this.handleSubAgentContextCompactionItem(
childSubAgentCallId,
parsed.threadId,
parsed.item,
"loading",
)
this.handleSubAgentContextCompactionItem(childSubAgentCallId, parsed.item, "loading")
) {
return;
}
@@ -6778,6 +6802,9 @@ export class CodexAppServerAgentSession implements AgentSession {
});
if (childSubAgentCallId) {
this.emitStartedProviderSubagentItem(parsed.threadId, timelineItem);
if (parsed.item.id) {
this.upsertSubAgentChildItem(childSubAgentCallId, parsed.item.id, timelineItem);
}
this.emitSubAgentActivityUpdate(childSubAgentCallId, "running");
await this.replayPendingSubAgentNotifications(registeredChildThreadIds);
return;
@@ -6831,6 +6858,9 @@ export class CodexAppServerAgentSession implements AgentSession {
if (parsed.threadId) {
this.emitProviderSubagentTimeline(parsed.threadId, timelineItem);
}
if (itemId) {
this.upsertSubAgentChildItem(childSubAgentCallId, itemId, timelineItem);
}
this.emitSubAgentActivityUpdate(childSubAgentCallId, "running");
return;
}
@@ -950,7 +950,6 @@ function mapCollabAgentToolCallItem(
type: "sub_agent",
subAgentType: "Sub-agent",
...(item.prompt ? { description: item.prompt } : {}),
// COMPAT(subagent-log): remove after 2027-04-30.
log: "",
actions: [],
};
@@ -1007,7 +1006,6 @@ function mapSubAgentActivityItem(
type: "sub_agent",
subAgentType: nativeName,
description,
// COMPAT(subagent-log): remove after 2027-04-30.
log: "",
actions: [],
},
@@ -69,7 +69,7 @@ import {
type OmpRuntimeOptions,
} from "./provider-config.js";
export { formatOmpVersionSupport, resolveOmpDiagnosticPaths } from "./provider-config.js";
import { OmpSubagentCardTracker } from "./subagent-card-tracker.js";
import { OmpSubagentCardTracker, type OmpSubagentCardScheduler } from "./subagent-card-tracker.js";
import { ompSkillPromptUserText, shouldDisplayOmpCustomMessage } from "./custom-message.js";
import { getUserMessageText } from "./message-history.js";
import { mapOmpSystemNoticeToNotification } from "./system-notice.js";
@@ -130,6 +130,7 @@ export interface OmpAgentClientOptions {
logger: Logger;
runtimeSettings?: ProviderRuntimeSettings;
runtime?: OmpRuntime;
subagentCardScheduler?: OmpSubagentCardScheduler;
providerIdleScheduler?: OmpProviderIdleScheduler;
noTurnScheduler?: OmpNoTurnScheduler;
usagePollScheduler?: OmpUsagePollScheduler;
@@ -204,6 +205,7 @@ interface OmpAgentSessionOptions {
initialState: OmpSessionState;
currentModeId?: string | null;
logger: Logger;
subagentCardScheduler?: OmpSubagentCardScheduler;
providerIdleScheduler?: OmpProviderIdleScheduler;
noTurnScheduler?: OmpNoTurnScheduler;
usagePollScheduler?: OmpUsagePollScheduler;
@@ -771,7 +773,9 @@ export class OmpAgentSession implements AgentSession {
this.logger.debug({ err: error }, "OMP context usage poll failed");
},
});
this.subagentCardTracker = new OmpSubagentCardTracker();
this.subagentCardTracker = new OmpSubagentCardTracker({
scheduler: options.subagentCardScheduler,
});
this.lastKnownThinkingOptionId =
normalizeOmpThinkingOption(options.config.thinkingOptionId) ??
this.state.thinkingLevel ??
@@ -2151,7 +2155,8 @@ export class OmpAgentSession implements AgentSession {
return mapOmpToolDetail(toolCall, result, {
toolCallId,
bridgedTool: this.hostTools?.bridge?.tool(toolCall.toolName),
mapSubagentDetail: (detail) => this.subagentCardTracker.detailFor(toolCallId, detail),
mapSubagentDetail: (detail) =>
this.subagentCardTracker.detailFor(toolCallId, detail) ?? detail,
});
}
@@ -2241,6 +2246,7 @@ export class OmpAgentClient implements AgentClient {
private readonly logger: Logger;
private readonly runtimeSettings?: ProviderRuntimeSettings;
private readonly subagentCardScheduler?: OmpSubagentCardScheduler;
private readonly providerIdleScheduler?: OmpProviderIdleScheduler;
private readonly noTurnScheduler?: OmpNoTurnScheduler;
private readonly usagePollScheduler?: OmpUsagePollScheduler;
@@ -2259,6 +2265,7 @@ export class OmpAgentClient implements AgentClient {
);
this.logger = options.logger;
this.runtimeSettings = runtimeSettings;
this.subagentCardScheduler = options.subagentCardScheduler;
this.providerIdleScheduler = options.providerIdleScheduler;
this.noTurnScheduler = options.noTurnScheduler;
this.usagePollScheduler = options.usagePollScheduler;
@@ -2342,6 +2349,7 @@ export class OmpAgentClient implements AgentClient {
initialState,
currentModeId: launchMode.modeId,
logger: this.logger,
subagentCardScheduler: this.subagentCardScheduler,
providerIdleScheduler: this.providerIdleScheduler,
noTurnScheduler: this.noTurnScheduler,
usagePollScheduler: this.usagePollScheduler,
@@ -2413,6 +2421,7 @@ export class OmpAgentClient implements AgentClient {
initialState,
currentModeId: launchMode.modeId,
logger: this.logger,
subagentCardScheduler: this.subagentCardScheduler,
providerIdleScheduler: this.providerIdleScheduler,
noTurnScheduler: this.noTurnScheduler,
usagePollScheduler: this.usagePollScheduler,
@@ -637,7 +637,7 @@ describe("OMP history mapper", () => {
subAgentType: "explore",
description: "Inspect files",
childSessionId: "/tmp/omp-task/Explore.jsonl",
log: "",
log: "done\ntranscript: /tmp/omp-task/Explore.jsonl",
},
error: null,
},
@@ -1,6 +1,6 @@
import { describe, expect, test } from "vitest";
import { OmpSubagentCardTracker } from "./subagent-card-tracker.js";
import { OmpSubagentCardTracker, type OmpSubagentCardScheduler } from "./subagent-card-tracker.js";
import type { OmpSubagentLifecyclePayload, OmpSubagentProgressPayload } from "./rpc-types.js";
const PARENT_TOOL_CALL_ID = "task-1";
@@ -59,84 +59,184 @@ const PROGRESS: OmpSubagentProgressPayload[] = [
},
];
class ManualScheduler implements OmpSubagentCardScheduler {
private currentMs = 0;
private nextId = 1;
private readonly timers = new Map<number, { dueMs: number; callback: () => void }>();
now(): number {
return this.currentMs;
}
setTimeout(callback: () => void, delayMs: number) {
const id = this.nextId;
this.nextId += 1;
this.timers.set(id, { dueMs: this.currentMs + delayMs, callback });
return { token: id };
}
clearTimeout(timer: { token: unknown }): void {
if (typeof timer.token === "number") {
this.timers.delete(timer.token);
}
}
advance(ms: number): void {
this.currentMs += ms;
const dueTimers = [...this.timers.entries()]
.filter(([, timer]) => timer.dueMs <= this.currentMs)
.sort((left, right) => left[1].dueMs - right[1].dueMs);
for (const [id, timer] of dueTimers) {
if (this.timers.delete(id)) {
timer.callback();
}
}
}
}
describe("OmpSubagentCardTracker", () => {
test("keeps the card a launch marker while the child runs", () => {
test("folds lifecycle and progress into one throttled sub-agent detail", () => {
const scheduler = new ManualScheduler();
const emitted: string[] = [];
const emit = (toolCallId: string) => {
emitted.push(toolCallId);
return true;
};
const tracker = new OmpSubagentCardTracker();
tracker.handleLifecycle(LIFECYCLE, emit);
const tracker = new OmpSubagentCardTracker({
scheduler,
emitToolCall: (toolCallId) => {
emitted.push(toolCallId);
return true;
},
});
tracker.handleLifecycle(LIFECYCLE);
for (const payload of PROGRESS) {
tracker.handleProgress(payload, emit);
tracker.handleProgress(payload);
}
expect(emitted).toEqual([PARENT_TOOL_CALL_ID]);
expect(
tracker.detailFor(PARENT_TOOL_CALL_ID, {
type: "sub_agent",
subAgentType: "task",
description: "Task arg description wins",
log: "",
}),
).toEqual({
const detailBeforeTrailing = tracker.detailFor(PARENT_TOOL_CALL_ID, {
type: "sub_agent",
subAgentType: "task",
description: "Task arg description wins",
log: "",
});
expect(detailBeforeTrailing).toEqual({
type: "sub_agent",
subAgentType: "task",
description: "Task arg description wins",
childSessionId: SESSION_FILE,
log: "",
log: [
"EchoSubagent started",
"[bash] echo subagent-hi",
"[yield]",
"EchoSubagent completed",
].join("\n"),
});
});
test("re-emits the card when a later event names the child", () => {
const emitted: string[] = [];
const emit = (toolCallId: string) => {
emitted.push(toolCallId);
return true;
};
const tracker = new OmpSubagentCardTracker();
tracker.handleLifecycle({ ...LIFECYCLE, description: undefined, sessionFile: undefined }, emit);
tracker.handleProgress(PROGRESS[0], emit);
scheduler.advance(500);
expect(emitted).toEqual([PARENT_TOOL_CALL_ID, PARENT_TOOL_CALL_ID]);
expect(tracker.detailFor(PARENT_TOOL_CALL_ID, { type: "sub_agent", log: "" })).toEqual({
type: "sub_agent",
childSessionId: SESSION_FILE,
log: "",
});
test("aggregates batch progress streams into one index-prefixed log", () => {
const scheduler = new ManualScheduler();
const emitted: string[] = [];
const tracker = new OmpSubagentCardTracker({
scheduler,
emitToolCall: (toolCallId) => {
emitted.push(toolCallId);
return true;
},
});
tracker.delete(PARENT_TOOL_CALL_ID);
expect(tracker.detailFor(PARENT_TOOL_CALL_ID, { type: "sub_agent", log: "" })).toEqual({
tracker.handleLifecycle({
id: "Explore",
agent: "task",
description: "Inspect files",
status: "started",
sessionFile: "/tmp/one.jsonl",
parentToolCallId: "task.batch-1",
index: 0,
});
scheduler.advance(600);
tracker.handleProgress({
index: 5,
agent: "task",
task: "Run tests",
parentToolCallId: "task.batch-1",
progress: {
id: "Test",
status: "running",
description: "Run tests",
recentTools: [{ tool: "bash", args: "npm test", endMs: 10 }],
},
sessionFile: "/tmp/two.jsonl",
});
expect(emitted).toEqual(["task.batch-1", "task.batch-1"]);
expect(
tracker.detailFor("task.batch-1", {
type: "sub_agent",
subAgentType: "batch",
actions: [],
log: "",
}),
).toEqual({
type: "sub_agent",
log: "",
subAgentType: "batch",
description: "Inspect files",
childSessionId: "/tmp/one.jsonl",
log: "[1/6] Explore started\n[6/6] [bash] npm test",
actions: [],
});
});
test("leaves the card alone when a later child of the same task starts", () => {
const emitted: string[] = [];
const emit = (toolCallId: string) => {
emitted.push(toolCallId);
return true;
};
const tracker = new OmpSubagentCardTracker();
tracker.handleLifecycle(LIFECYCLE, emit);
tracker.handleLifecycle(
{
...LIFECYCLE,
id: "Second",
index: 1,
description: "Second child",
sessionFile: "/s.jsonl",
},
emit,
);
expect(emitted).toEqual([PARENT_TOOL_CALL_ID]);
expect(tracker.detailFor(PARENT_TOOL_CALL_ID, { type: "sub_agent", log: "" })).toEqual({
type: "sub_agent",
description: "Run echo in subagent",
childSessionId: SESSION_FILE,
log: "",
test("keeps dirty trailing state available for final detail and cancels it on cleanup", () => {
const scheduler = new ManualScheduler();
const emitted: string[] = [];
const tracker = new OmpSubagentCardTracker({
scheduler,
emitToolCall: (toolCallId) => {
emitted.push(toolCallId);
return true;
},
});
tracker.handleLifecycle({
id: "Explore",
agent: "task",
description: "Inspect files",
status: "started",
parentToolCallId: "task-1",
index: 0,
});
scheduler.advance(100);
tracker.handleProgress({
index: 0,
agent: "task",
task: "Inspect files",
parentToolCallId: "task-1",
progress: {
id: "Explore",
status: "running",
recentOutput: ["found target file"],
},
});
expect(emitted).toEqual(["task-1"]);
expect(
tracker.detailFor("task-1", {
type: "sub_agent",
log: "",
}).log,
).toBe("Explore started\nfound target file");
tracker.delete("task-1");
scheduler.advance(500);
expect(emitted).toEqual(["task-1"]);
expect(
tracker.detailFor("task-1", {
type: "sub_agent",
log: "static",
}).log,
).toBe("static");
});
});
@@ -1,106 +1,361 @@
import type { ToolCallDetail } from "../../agent-sdk-types.js";
import type { OmpSubagentLifecyclePayload, OmpSubagentProgressPayload } from "./rpc-types.js";
interface OmpSubagentCardChild {
export interface OmpSubagentCardTimer {
readonly token: unknown;
}
export interface OmpSubagentCardScheduler {
now(): number;
setTimeout(callback: () => void, delayMs: number): OmpSubagentCardTimer;
clearTimeout(timer: OmpSubagentCardTimer): void;
}
export interface OmpSubagentCardTrackerOptions {
scheduler?: OmpSubagentCardScheduler;
emitToolCall?: (toolCallId: string) => boolean;
}
interface OmpSubagentLogLine {
key: string;
text: string;
}
interface OmpSubagentCardItem {
index: number;
description?: string;
agent?: string;
status?: "pending" | "running" | "completed" | "failed" | "aborted";
childSessionId?: string;
lines: OmpSubagentLogLine[];
lineKeys: Set<string>;
}
interface OmpSubagentCardState {
items: Map<number, OmpSubagentCardItem>;
emitToolCall: (toolCallId: string) => boolean;
lastEmitMs: number | null;
trailingTimer: OmpSubagentCardTimer | null;
dirty: boolean;
}
const MAX_LOG_LINES_PER_ITEM = 200;
const MAX_LOG_LINE_CHARS = 240;
const THROTTLE_INTERVAL_MS = 500;
export function createOmpSubagentCardScheduler(): OmpSubagentCardScheduler {
return {
now: () => Date.now(),
setTimeout: (callback, delayMs) => ({ token: setTimeout(callback, delayMs) }),
clearTimeout: (timer) => {
clearTimeout(timer.token as ReturnType<typeof setTimeout>);
},
};
}
/**
* Fills a task card's description and child session from OMP's sub-agent events. The card is a
* launch marker that names the task's first child; the children's activity belongs to their own
* timelines.
*/
export class OmpSubagentCardTracker {
private readonly firstChildren = new Map<string, OmpSubagentCardChild>();
private readonly scheduler: OmpSubagentCardScheduler;
private readonly emitToolCall: (toolCallId: string) => boolean;
private readonly states = new Map<string, OmpSubagentCardState>();
constructor(options: OmpSubagentCardTrackerOptions) {
this.scheduler = options.scheduler ?? createOmpSubagentCardScheduler();
this.emitToolCall = options.emitToolCall ?? (() => false);
}
handleLifecycle(
payload: OmpSubagentLifecyclePayload,
emitToolCall: (toolCallId: string) => boolean,
emitToolCall: (toolCallId: string) => boolean = this.emitToolCall,
): void {
this.update(
payload.parentToolCallId,
{
index: payload.index,
description: payload.description,
childSessionId: payload.sessionFile,
},
emitToolCall,
const parentToolCallId = readTrimmedString(payload.parentToolCallId);
if (!parentToolCallId) {
return;
}
this.stateFor(parentToolCallId).emitToolCall = emitToolCall;
const item = this.upsertItem(parentToolCallId, {
index: payload.index,
agent: payload.agent,
description: payload.description,
status: payload.status === "started" ? "running" : payload.status,
childSessionId: payload.sessionFile,
});
this.appendLine(
item,
`lifecycle:${payload.status}:${payload.id}`,
`${payload.id} ${payload.status}`,
);
this.requestEmit(parentToolCallId);
}
handleProgress(
payload: OmpSubagentProgressPayload,
emitToolCall: (toolCallId: string) => boolean,
emitToolCall: (toolCallId: string) => boolean = this.emitToolCall,
): void {
this.update(
payload.parentToolCallId,
{
index: payload.index,
description: payload.progress.description,
childSessionId: payload.sessionFile,
},
emitToolCall,
);
const parentToolCallId = readTrimmedString(payload.parentToolCallId);
if (!parentToolCallId) {
return;
}
this.stateFor(parentToolCallId).emitToolCall = emitToolCall;
const item = this.upsertItem(parentToolCallId, {
index: payload.index,
agent: payload.agent,
description: payload.progress.description,
status: payload.progress.status,
childSessionId: payload.sessionFile,
});
const currentToolLine = summarizeTool(readRecord(payload.progress.currentTool));
if (currentToolLine) {
this.appendLine(item, `current:${currentToolLine.key}`, currentToolLine.text);
}
for (const tool of readRecordArray(payload.progress.recentTools)) {
const line = summarizeTool(tool);
if (line) {
this.appendLine(item, `tool:${line.key}`, line.text);
}
}
for (const output of readOutputLines(payload.progress.recentOutput)) {
this.appendLine(item, `output:${output}`, output);
}
if (payload.progress.status !== "running") {
this.appendLine(
item,
`status:${payload.progress.status}:${payload.progress.id}`,
`${payload.progress.id} ${payload.progress.status}`,
);
}
this.requestEmit(parentToolCallId);
}
detailFor(toolCallId: string, baseDetail: ToolCallDetail): ToolCallDetail {
if (baseDetail.type !== "sub_agent") {
return baseDetail;
}
const first = this.firstChildren.get(toolCallId);
if (!first) {
const state = this.states.get(toolCallId);
if (!state) {
return baseDetail;
}
const description = baseDetail.description ?? first.description;
const childSessionId = first.childSessionId ?? baseDetail.childSessionId;
return {
...baseDetail,
...(description ? { description } : {}),
...(childSessionId ? { childSessionId } : {}),
const items = [...state.items.values()].sort((left, right) => left.index - right.index);
const firstItem = items[0];
if (!firstItem) {
return baseDetail;
}
const detail: ToolCallDetail = {
type: "sub_agent",
...(baseDetail.subAgentType ? { subAgentType: baseDetail.subAgentType } : {}),
...((baseDetail.description ?? firstItem.description)
? { description: baseDetail.description ?? firstItem.description }
: {}),
...((firstItem.childSessionId ?? baseDetail.childSessionId)
? { childSessionId: firstItem.childSessionId ?? baseDetail.childSessionId }
: {}),
log: buildLog(items, baseDetail.log),
};
return baseDetail.actions ? { ...detail, actions: baseDetail.actions } : detail;
}
flush(toolCallId: string): void {
const state = this.states.get(toolCallId);
if (!state || !state.dirty) {
return;
}
this.emitNow(toolCallId, state);
}
delete(toolCallId: string): void {
this.firstChildren.delete(toolCallId);
const state = this.states.get(toolCallId);
if (state?.trailingTimer) {
this.scheduler.clearTimeout(state.trailingTimer);
}
this.states.delete(toolCallId);
}
clear(): void {
this.firstChildren.clear();
for (const toolCallId of this.states.keys()) {
this.delete(toolCallId);
}
}
private update(
parentToolCallId: string | undefined,
input: { index: number; description?: string; childSessionId?: string },
emitToolCall: (toolCallId: string) => boolean,
): void {
const toolCallId = readTrimmedString(parentToolCallId);
if (!toolCallId) {
return;
private upsertItem(
parentToolCallId: string,
input: {
index: number;
agent?: string;
status?: "pending" | "running" | "completed" | "failed" | "aborted";
description?: string;
childSessionId?: string;
},
): OmpSubagentCardItem {
const state = this.stateFor(parentToolCallId);
const existing = state.items.get(input.index);
if (existing) {
existing.agent = readTrimmedString(input.agent) ?? existing.agent;
existing.status = input.status ?? existing.status;
existing.description = readTrimmedString(input.description) ?? existing.description;
existing.childSessionId = readTrimmedString(input.childSessionId) ?? existing.childSessionId;
return existing;
}
const first = this.firstChildren.get(toolCallId);
if (first && input.index > first.index) {
return;
}
const existing = first?.index === input.index ? first : undefined;
const next: OmpSubagentCardChild = {
const item: OmpSubagentCardItem = {
index: input.index,
description: readTrimmedString(input.description) ?? existing?.description,
childSessionId: readTrimmedString(input.childSessionId) ?? existing?.childSessionId,
lines: [],
lineKeys: new Set<string>(),
...(readTrimmedString(input.agent) ? { agent: readTrimmedString(input.agent) } : {}),
...(input.status ? { status: input.status } : {}),
...(readTrimmedString(input.description)
? { description: readTrimmedString(input.description) }
: {}),
...(readTrimmedString(input.childSessionId)
? { childSessionId: readTrimmedString(input.childSessionId) }
: {}),
};
if (
existing &&
existing.description === next.description &&
existing.childSessionId === next.childSessionId
) {
state.items.set(input.index, item);
return item;
}
private stateFor(parentToolCallId: string): OmpSubagentCardState {
const existing = this.states.get(parentToolCallId);
if (existing) {
return existing;
}
const state: OmpSubagentCardState = {
items: new Map(),
emitToolCall: this.emitToolCall,
lastEmitMs: null,
trailingTimer: null,
dirty: false,
};
this.states.set(parentToolCallId, state);
return state;
}
private appendLine(item: OmpSubagentCardItem, key: string, text: string): void {
const normalizedText = normalizeLine(text);
if (!normalizedText || item.lineKeys.has(key)) {
return;
}
this.firstChildren.set(toolCallId, next);
emitToolCall(toolCallId);
item.lines.push({ key, text: normalizedText });
item.lineKeys.add(key);
while (item.lines.length > MAX_LOG_LINES_PER_ITEM) {
const removed = item.lines.shift();
if (removed) {
item.lineKeys.delete(removed.key);
}
}
}
private requestEmit(toolCallId: string): void {
const state = this.stateFor(toolCallId);
const now = this.scheduler.now();
const elapsedMs = state.lastEmitMs === null ? THROTTLE_INTERVAL_MS : now - state.lastEmitMs;
if (elapsedMs >= THROTTLE_INTERVAL_MS) {
this.emitNow(toolCallId, state);
return;
}
state.dirty = true;
if (state.trailingTimer) {
return;
}
state.trailingTimer = this.scheduler.setTimeout(() => {
state.trailingTimer = null;
if (state.dirty) {
this.emitNow(toolCallId, state);
}
}, THROTTLE_INTERVAL_MS - elapsedMs);
}
private emitNow(toolCallId: string, state: OmpSubagentCardState): void {
if (state.trailingTimer) {
this.scheduler.clearTimeout(state.trailingTimer);
state.trailingTimer = null;
}
state.dirty = false;
state.lastEmitMs = this.scheduler.now();
state.emitToolCall(toolCallId);
}
}
function buildLog(items: OmpSubagentCardItem[], fallback: string): string {
const itemCount = Math.max(...items.map((item) => item.index)) + 1;
const lines = items.flatMap((item) => {
const prefix = itemCount > 1 ? `[${item.index + 1}/${itemCount}] ` : "";
return item.lines.map((line) => `${prefix}${line.text}`);
});
return lines.length > 0 ? lines.join("\n") : fallback;
}
function summarizeTool(tool: Record<string, unknown> | null): OmpSubagentLogLine | null {
if (!tool) {
return null;
}
const toolName = readTrimmedString(tool.tool) ?? readTrimmedString(tool.name);
if (!toolName) {
return null;
}
const args = readTrimmedString(tool.args) ?? summarizeUnknown(tool.input);
const endMs = typeof tool.endMs === "number" ? `:${tool.endMs}` : "";
return {
key: `${toolName}:${args ?? ""}${endMs}`,
text: args ? `[${toolName}] ${args}` : `[${toolName}]`,
};
}
function readOutputLines(value: unknown): string[] {
if (!Array.isArray(value)) {
return [];
}
return value.flatMap((item) => {
const text =
readTrimmedString(item) ??
(isRecord(item)
? (readTrimmedString(item.text) ??
readTrimmedString(item.output) ??
readTrimmedString(item.content))
: undefined);
return text ? [text] : [];
});
}
function readRecord(value: unknown): Record<string, unknown> | null {
return isRecord(value) ? value : null;
}
function readRecordArray(value: unknown): Record<string, unknown>[] {
return Array.isArray(value) ? value.filter(isRecord) : [];
}
function summarizeUnknown(value: unknown): string | undefined {
if (value === undefined || value === null) {
return undefined;
}
if (typeof value === "string") {
return readTrimmedString(value);
}
const json = JSON.stringify(value);
return typeof json === "string" ? normalizeLine(json) : undefined;
}
function normalizeLine(value: string): string | undefined {
const normalized = value.trim().replace(/\s+/g, " ");
if (!normalized) {
return undefined;
}
if (normalized.length <= MAX_LOG_LINE_CHARS) {
return normalized;
}
return `${normalized.slice(0, MAX_LOG_LINE_CHARS)}...`;
}
function readTrimmedString(value: unknown): string | undefined {
return typeof value === "string" && value.trim().length > 0 ? value.trim() : undefined;
}
function isRecord(value: unknown): value is Record<string, unknown> {
return typeof value === "object" && value !== null && !Array.isArray(value);
}
@@ -85,7 +85,7 @@ describe("OMP tool call mapper", () => {
expect(mapOmpToolDetail(parseToolArgs("todo", { op: "view" }), null)).toBeNull();
});
test("takes the child session from the task transcript path", () => {
test("uses task result text and transcript path as the best static replay detail", () => {
expect(
mapOmpToolDetail(
parseToolArgs("task", {
@@ -106,7 +106,7 @@ describe("OMP tool call mapper", () => {
subAgentType: "explore",
description: "Inspect the target files",
childSessionId: "/tmp/omp-task-static/Explore.jsonl",
log: "",
log: "done\ntranscript: /tmp/omp-task-static/Explore.jsonl",
});
});
@@ -179,6 +179,7 @@ const OPAQUE_TEXT_TOOLS = new Set([
function mapOmpTaskDetail(args: unknown, result: OmpToolResult): ToolCallDetail {
const argRecord = isRecord(args) ? args : {};
const firstTask = Array.isArray(argRecord.tasks) ? argRecord.tasks.find(isRecord) : undefined;
const resultText = extractTextFromToolResult(result);
const childSessionId = readChildSessionId(result);
return {
type: "sub_agent",
@@ -199,8 +200,7 @@ function mapOmpTaskDetail(args: unknown, result: OmpToolResult): ToolCallDetail
argRecord.assignment,
),
...(childSessionId ? { childSessionId } : {}),
// COMPAT(subagent-log): remove after 2027-04-30.
log: "",
log: resultText?.trim() ?? "",
};
}
@@ -4846,6 +4846,15 @@ class OpenCodeAgentSession implements AgentSession {
...item,
status: "failed",
error,
detail:
item.detail.type === "sub_agent"
? {
...item.detail,
log: [item.detail.log, error.message]
.filter((entry) => entry.trim().length > 0)
.join("\n"),
}
: item.detail,
},
},
turnId,
@@ -153,6 +153,21 @@ const OpencodeGrepOutputSchema = z
])
.nullable();
function formatLogEntry(value: unknown): string | undefined {
const outputText = readOutputText(value);
if (outputText) {
return outputText;
}
if (value === null || value === undefined) {
return undefined;
}
try {
return JSON.stringify(value, null, 2);
} catch {
return String(value);
}
}
function extractOpenCodeTaskSessionId(value: unknown): string | undefined {
const text = readOutputText(value);
if (text) {
@@ -180,6 +195,7 @@ function extractOpenCodeTaskSessionId(value: unknown): string | undefined {
function deriveOpencodeTaskDetail(
input: unknown,
output: unknown,
error: unknown,
metadata: Record<string, unknown> | undefined,
): ToolCallDetail | null {
if (!isRecord(input)) {
@@ -192,6 +208,7 @@ function deriveOpencodeTaskDetail(
return null;
}
const log = [formatLogEntry(output), formatLogEntry(error)].filter((entry) => entry).join("\n");
const childSessionId =
normalizeSubAgentText(metadata?.sessionId) ?? extractOpenCodeTaskSessionId(output);
return {
@@ -199,8 +216,7 @@ function deriveOpencodeTaskDetail(
...(subAgentType ? { subAgentType } : {}),
...(description ? { description } : {}),
...(childSessionId ? { childSessionId } : {}),
// COMPAT(subagent-log): remove after 2027-04-30.
log: "",
log,
actions: [],
};
}
@@ -357,10 +373,11 @@ export function deriveOpencodeToolDetail(
toolName: string,
input: unknown,
output: unknown,
error: unknown = null,
metadata?: Record<string, unknown>,
): ToolCallDetail {
if (toolName.trim().toLowerCase() === "task") {
const taskDetail = deriveOpencodeTaskDetail(input, output, metadata);
const taskDetail = deriveOpencodeTaskDetail(input, output, error, metadata);
if (taskDetail) {
return taskDetail;
}
@@ -460,7 +460,7 @@ describe("opencode tool-call mapper", () => {
});
});
it("maps completed task calls to sub_agent detail", () => {
it("maps completed task calls with final output log to sub_agent detail", () => {
const item = expectMapped(
mapOpencodeToolCall({
toolName: "task",
@@ -480,7 +480,7 @@ describe("opencode tool-call mapper", () => {
type: "sub_agent",
subAgentType: "explore",
description: "Explore agent-tools codebase",
log: "",
log: "Found the CLI entrypoint and provider registry.",
actions: [],
});
});
@@ -505,12 +505,12 @@ describe("opencode tool-call mapper", () => {
subAgentType: "explore",
description: "Explore current directory",
childSessionId: "ses_2268db431ffe299vL1bbot8R7Z",
log: "",
log: '<task id="ses_2268db431ffe299vL1bbot8R7Z">done</task>',
actions: [],
});
});
it("maps aborted task calls to sub_agent detail", () => {
it("maps aborted task calls with preserved error log to sub_agent detail", () => {
const item = expectMapped(
mapOpencodeToolCall({
toolName: "task",
@@ -531,7 +531,7 @@ describe("opencode tool-call mapper", () => {
type: "sub_agent",
subAgentType: "explore",
description: "Explore agent-tools codebase",
log: "",
log: "Tool execution aborted",
actions: [],
});
});
@@ -43,7 +43,7 @@ export function mapOpencodeToolCall(params: OpencodeToolCallParams): ToolCallTim
const error = raw.error ?? null;
const rawStatus = typeof raw.status === "string" ? raw.status : undefined;
const status = normalizeToolCallStatus(rawStatus, error, output);
const detail = deriveOpencodeToolDetail(name, input, output, raw.metadata);
const detail = deriveOpencodeToolDetail(name, input, output, error, raw.metadata);
if (status === "failed") {
return {
@@ -1,4 +1,5 @@
import { z } from "zod";
import { extractTextFromToolResult } from "../../tool-call-mapper.js";
import type { PiChildSessionFile, PiExtension } from "../contract.js";
import type { ProviderSubagentInputEvent } from "../../../../provider-subagents/store.js";
@@ -98,8 +99,7 @@ export const edxethPiSubagents: PiExtension = {
detail: {
type: "sub_agent",
description,
// COMPAT(subagent-log): remove after 2027-04-30.
log: "",
log: extractTextFromToolResult(call.result)?.trim() ?? "",
},
subagents,
childSessions,
@@ -130,18 +130,6 @@ describe("@gotgenes/pi-subagents adapter", () => {
.filter((event) => event.event.type === "upsert")
.map((event) => (event.event.type === "upsert" ? event.event.status : null)),
).toEqual(["running", "completed"]);
expect(
events.flatMap((event) => (event.event.type === "timeline" ? [event.event] : [])),
).toEqual([
{
type: "timeline",
id: "call_502323",
item: {
type: "assistant_message",
text: "Agent completed in 4.7s (1 tool uses, 9.0k token).\nAgent ID: 848583f5-6749-4a1\n\nThe workspace is empty.",
},
},
]);
});
test("correlates a background notification and result collection", async () => {
const source = readSubagentFixture(new URL("./fixtures/background.json", import.meta.url));
@@ -64,8 +64,7 @@ export const gotgenesPiSubagents: PiExtension = {
type: "sub_agent" as const,
subAgentType: title,
description,
// COMPAT(subagent-log): remove after 2027-04-30.
log: "",
log: extractTextFromToolResult(call.result)?.trim() ?? "",
};
if (call.status === "running")
return {
@@ -93,10 +92,6 @@ export const gotgenesPiSubagents: PiExtension = {
outputFileFromToolResult(call.result) ?? pendingFilesByAgent.get(details.data.agentId);
pendingFilesByAgent.delete(details.data.agentId);
const childSessions = childSession(call.callId, file);
const childStatus = status(details.data.status);
// Without a session file, as on replay, the child's answer is the tool result.
const answer =
!file && childStatus !== "running" ? extractTextFromToolResult(call.result)?.trim() : "";
return {
detail,
subagents: [
@@ -106,17 +101,8 @@ export const gotgenesPiSubagents: PiExtension = {
title,
description,
toolCallId: call.callId,
status: childStatus,
status: status(details.data.status),
},
...(answer
? [
{
type: "timeline" as const,
id: call.callId,
item: { type: "assistant_message" as const, text: answer },
},
]
: []),
],
childSessions,
};
@@ -132,8 +118,7 @@ export const gotgenesPiSubagents: PiExtension = {
const detail = {
type: "sub_agent" as const,
description: details.success ? details.data.description : undefined,
// COMPAT(subagent-log): remove after 2027-04-30.
log: "",
log: extractTextFromToolResult(call.result)?.trim() ?? "",
};
if (call.status === "running" || !details.success) return { detail };
const file = details.data.transcriptPath;
@@ -210,7 +210,7 @@ describe("Pi extension host", () => {
type: "sub_agent",
subAgentType: "scout",
description: "Inspect files",
log: "",
log: "Found two files",
},
}),
);
@@ -137,7 +137,7 @@ describe("pi-subagents adapter", () => {
type: "sub_agent",
subAgentType: "reviewer",
description: "Review the Pi mapper change",
log: "",
log: "The mapper change preserves provider status.",
});
});
@@ -131,8 +131,7 @@ export const piSubagents: PiExtension = {
type: "sub_agent" as const,
subAgentType: args.data.agent,
description: args.data.task,
// COMPAT(subagent-log): remove after 2027-04-30.
log: "",
log: extractTextFromToolResult(call.result)?.trim() ?? "",
};
const base = {
type: "upsert" as const,
@@ -54,6 +54,8 @@ export async function verifySubagentFixture(
try {
await session.startTurn("Delegate work");
for (const event of fixture.events) pi.latestSession().emit(event);
const immediate = live.filter((event) => event.type === "provider_subagent");
expect(immediate.some((event) => event.event.type === "timeline")).toBe(false);
pi.latestSession().finishTurn();
const replay: AgentStreamEvent[] = [];
for await (const event of streamPiHistory("pi", fixture.messages)) replay.push(event);
@@ -36,17 +36,12 @@ describe("@tintinweb/pi-subagents adapter", () => {
});
expect(mapping?.subagents).toEqual([
expect.objectContaining({ id: "call-1", status: "completed" }),
{
type: "timeline",
id: "call-1",
item: { type: "assistant_message", text: "Wrapped up (turn limit)" },
},
]);
expect(mapping?.detail).toEqual({
type: "sub_agent",
subAgentType: "Explore",
description: "Inspect",
log: "",
log: "Wrapped up (turn limit)",
});
});
test("maps captured foreground lifecycle live and on replay", async () => {
@@ -58,18 +53,6 @@ describe("@tintinweb/pi-subagents adapter", () => {
.filter((event) => event.event.type === "upsert")
.map((event) => (event.event.type === "upsert" ? event.event.status : null)),
).toEqual(["running", "completed"]);
expect(
events.flatMap((event) => (event.event.type === "timeline" ? [event.event] : [])),
).toEqual([
{
type: "timeline",
id: "call_458926",
item: {
type: "assistant_message",
text: "Agent completed in 6.6s (1 tool uses, 18.8k token).\n\nThe workspace directory `/tmp/paseo-pi-extension-phase3/workspace` is empty (no files or hidden dotfiles).",
},
},
]);
});
test("uses a structured notification to complete a background child", async () => {
const source = readSubagentFixture(new URL("./fixtures/background.json", import.meta.url));
@@ -164,7 +147,7 @@ describe("@tintinweb/pi-subagents adapter", () => {
expect(mapping?.detail).toEqual({
type: "sub_agent",
description: undefined,
log: "",
log: "Wrapped up (turn limit)",
});
},
);
@@ -57,8 +57,7 @@ export const tintinwebPiSubagents: PiExtension = {
type: "sub_agent" as const,
subAgentType: title,
description,
// COMPAT(subagent-log): remove after 2027-04-30.
log: "",
log: extractTextFromToolResult(call.result)?.trim() ?? "",
};
if (call.status === "running")
return {
@@ -85,10 +84,6 @@ export const tintinwebPiSubagents: PiExtension = {
const file = outputFileFromToolResult(call.result);
const childSessions = file && !readSessions.has(file) ? [{ id: call.callId, file }] : [];
if (file) readSessions.add(file);
const childStatus = status(details.data.status);
// A foreground child has no session file; its answer is the tool result.
const answer =
!file && childStatus !== "running" ? extractTextFromToolResult(call.result)?.trim() : "";
return {
detail,
subagents: [
@@ -98,17 +93,8 @@ export const tintinwebPiSubagents: PiExtension = {
title,
description,
toolCallId: call.callId,
status: childStatus,
status: status(details.data.status),
},
...(answer
? [
{
type: "timeline" as const,
id: call.callId,
item: { type: "assistant_message" as const, text: answer },
},
]
: []),
],
childSessions,
};
@@ -124,8 +110,7 @@ export const tintinwebPiSubagents: PiExtension = {
const detail = {
type: "sub_agent" as const,
description: details.success ? details.data.description : undefined,
// COMPAT(subagent-log): remove after 2027-04-30.
log: "",
log: extractTextFromToolResult(call.result)?.trim() ?? "",
};
if (call.status === "running" || !details.success) return { detail };
return {
@@ -1,3 +1,6 @@
import { execFile } from "node:child_process";
import { promisify } from "node:util";
import { fileURLToPath } from "node:url";
import { mkdtempSync, rmSync } from "node:fs";
import { tmpdir } from "node:os";
import path from "node:path";
@@ -7,7 +10,7 @@ import { ClaudeAgentClient } from "../agent/providers/claude/agent.js";
import { createTestPaseoDaemon } from "../test-utils/paseo-daemon.js";
import { DaemonClient } from "../test-utils/daemon-client.js";
test("background children keep their actions after the parent finishes (#6253)", async () => {
test("background cards retain child actions after the parent finishes (#6253)", async () => {
const cwd = mkdtempSync(path.join(tmpdir(), "paseo-background-card-"));
const logger = pino({ level: "silent" });
const daemon = await createTestPaseoDaemon({
@@ -59,6 +62,22 @@ test("background children keep their actions after the parent finishes (#6253)",
limit: 0,
projection: "projected",
});
const { stdout } = await promisify(execFile)(
process.execPath,
[
"--conditions=source",
"--import",
"tsx",
fileURLToPath(new URL("../../../../cli/src/index.ts", import.meta.url)),
"--host",
`127.0.0.1:${daemon.port}`,
"logs",
agent.id,
],
{ env: { ...process.env, PASEO_HOME: daemon.paseoHome }, timeout: 30_000 },
);
expect(stdout).toContain("[Bash] ls -1");
expect(stdout).toContain("[Bash] pwd");
expect(
atFinish.entries.filter(
(row) => row.item.type === "tool_call" && row.item.detail?.type === "sub_agent",
@@ -75,6 +94,9 @@ test("background children keep their actions after the parent finishes (#6253)",
),
).toBe(true);
}
for (const card of cards) {
expect(card.detail?.type === "sub_agent" ? card.detail.log : "").toContain("[Bash]");
}
} finally {
await client.close();
await daemon.close();
@@ -0,0 +1,46 @@
import { mkdtempSync, rmSync } from "node:fs";
import { tmpdir } from "node:os";
import path from "node:path";
import { expect, test } from "vitest";
import { createDaemonTestContext } from "../test-utils/index.js";
// #2220: a timeline page is one WebSocket message, so a few huge entries used to close the socket.
test.each([40, 0])(
"fetching image-heavy history with limit %i keeps the direct connection usable",
async (limit) => {
const ctx = await createDaemonTestContext();
const cwd = mkdtempSync(path.join(tmpdir(), "oversized-timeline-"));
try {
const agent = await ctx.client.createAgent({ provider: "codex", cwd });
const imageUrl = `data:image/jpeg;base64,${"A".repeat(14 * 1024 * 1024)}`;
for (let index = 0; index < 5; index += 1) {
await ctx.daemon.daemon.agentManager.appendTimelineItem(agent.id, {
type: "tool_call",
callId: `screenshot-${index}`,
name: "node_repl.js",
status: "completed",
error: null,
detail: {
type: "unknown",
input: { index },
output: { _meta: { "codex/toolSurface": { screenshot: { url: imageUrl } } } },
},
});
}
const page = await ctx.client.fetchAgentTimeline(agent.id, { direction: "tail", limit });
const screenshots = page.entries.filter((entry) => entry.item.type === "tool_call");
expect(screenshots).toHaveLength(5);
for (const entry of screenshots) {
expect(Buffer.byteLength(JSON.stringify(entry.item))).toBeLessThanOrEqual(64 * 1024);
}
const next = await ctx.client.fetchAgentTimeline(agent.id, { direction: "tail", limit: 1 });
expect(next.entries).toHaveLength(1);
} finally {
await ctx.cleanup();
rmSync(cwd, { recursive: true, force: true });
}
},
60_000,
);