antechamber 0.2.25 → 0.2.26

package/CHANGELOG.md

@@ -7,6 +7,17 @@ and this project uses [Semantic Versioning](https://semver.org/).
 
 ## [Unreleased]
 
+## [0.2.26] - 2026-10-17
+
+### Changed
+
+- Name a subagent card for an `agent()` call made inside eval after its agent type (for example "Explore Agent"),
+  and describe it with the call's label, the task's summary, or the first line of its prompt, instead of a bare "agent".
+- Show subagents an eval cell starts with `tool.task(...)` as subagent cards that open their session, live and after
+  reloading history, including ones started in the background.
+- Show an `agent()` call that OmO refused (for example one with no agent type) as a failed subagent card. OmO
+  records it as a success but starts no task.
+
 ## [0.2.25] - 2026-10-17
 
 - feat(tools): show subagents and monitors started inside eval (#30)

package/dist/subagents.d.ts

@@ -1,5 +1,5 @@
 import { type SessionStore } from "./session-store.js";
-import { type ToolResult } from "./tool-map.js";
+import { type EvalAgent, type ToolResult } from "./tool-map.js";
 import type { ManagedSession } from "./sessions.js";
 import type { OCEvent, OCMessageWithParts, OCToolPart } from "./types.js";
 export declare function stripTaskSuffix(text: string): string;
@@ -17,9 +17,10 @@ export declare function findTaskBySession(sessionId: string): {
 /**
  * The session of a subagent an eval cell started with `agent()`. The recorded call keeps only its
  * arguments, so the task record is found by parent, prompt and label: the oldest one created since
- * the cell started and not yet claimed by an earlier call of the same cell.
+ * the cell started and not yet claimed by an earlier call of the same cell. `null` means the parent
+ * was searched and no record matched; `undefined` means it could not be searched.
  */
-export declare function evalAgentSession(parentPiId: string | undefined, args: Record<string, unknown>, claimed: Set<string>, after: number): string | undefined;
+export declare function evalAgentSession(parentPiId: string | undefined, args: Record<string, unknown>, claimed: Set<string>, after: number): EvalAgent | null | undefined;
 export declare function findTask(id: string, parentPiId: string): {
     path: string;
     data: Record<string, unknown>;

package/dist/subagents.js

@@ -193,10 +193,23 @@ export function findTaskBySession(sessionId) {
     }
     return undefined;
 }
+/**
+ * Whether a recorded argument names a task record's value. Codemode caps recorded strings at 512 code
+ * points and ends a cut one with "…" (call-capture.ts), and a schema agent() call sends its prompt with
+ * JSON instructions appended (agent-bridge.ts toTaskParams), so both match as prefixes.
+ */
+function recordedMatch(recorded, stored) {
+    if (typeof stored !== "string")
+        return false;
+    if (stored === recorded)
+        return true;
+    return recorded.endsWith("…") ? stored.startsWith(recorded.slice(0, -1)) : stored.startsWith(`${recorded}\n\n`);
+}
 /**
  * The session of a subagent an eval cell started with `agent()`. The recorded call keeps only its
  * arguments, so the task record is found by parent, prompt and label: the oldest one created since
- * the cell started and not yet claimed by an earlier call of the same cell.
+ * the cell started and not yet claimed by an earlier call of the same cell. `null` means the parent
+ * was searched and no record matched; `undefined` means it could not be searched.
  */
 export function evalAgentSession(parentPiId, args, claimed, after) {
     if (parentPiId === undefined || typeof args.prompt !== "string")
@@ -209,21 +222,31 @@ export function evalAgentSession(parentPiId, args, claimed, after) {
         if (data?.parent_session_id !== parentPiId)
             continue;
         const spec = record(data.spawn_spec) ? data.spawn_spec : {};
-        if (spec.prompt !== args.prompt)
+        if (!recordedMatch(args.prompt, spec.prompt))
             continue;
-        if (typeof args.label === "string" && data.name !== args.label)
+        if (typeof args.label === "string" && !recordedMatch(args.label, data.name))
             continue;
         const created = typeof data.created_at === "string" ? Date.parse(data.created_at) : Number.NaN;
         // The cell's start is taken when its card opens, a moment after OmO starts it.
         if (!Number.isFinite(created) || created < after - 5_000)
             continue;
         if (best === undefined || created < best.created)
-            best = { id, created };
+            best = { id, created, data };
     }
     if (best === undefined)
-        return undefined;
+        return null;
     claimed.add(best.id);
-    return ocIdFor(best.id);
+    const text = (value) => (typeof value === "string" && value !== "" ? value : undefined);
+    // OmO names an unlabelled task after its id, which says nothing.
+    const name = text(best.data.name);
+    const summary = text(best.data.task_summary);
+    const agentType = text(best.data.agent_type);
+    return {
+        sessionID: ocIdFor(best.id),
+        ...(name !== undefined && name !== best.id ? { name } : {}),
+        ...(summary !== undefined ? { summary } : {}),
+        ...(agentType !== undefined ? { agentType } : {}),
+    };
 }
 export function findTask(id, parentPiId) {
     if (!taskId(id))

package/dist/tool-map.d.ts

@@ -32,8 +32,15 @@ export declare function mapTool(omoName: string, args: unknown, result?: ToolRes
 export interface NestedSessions {
     /** The monitor child session for a recorded `monitor` call. */
     monitor?: (args: Record<string, unknown>) => string | undefined;
-    /** The subagent session for a recorded `agent()` call; `claimed` keeps repeated calls apart. */
-    agent?: (args: Record<string, unknown>, claimed: Set<string>, after: number) => string | undefined;
+    /** The subagent a recorded `agent()` call started; `claimed` keeps repeated calls apart. */
+    agent?: (args: Record<string, unknown>, claimed: Set<string>, after: number) => EvalAgent | null | undefined;
+}
+/** The task record behind an `agent()` call made inside eval. */
+export interface EvalAgent {
+    sessionID: string;
+    name?: string;
+    summary?: string;
+    agentType?: string;
 }
 /**
  * Finished eval calls, from the raw RPC capture of installed omo-ai 5.1.33 / senpi runtime, 2026-10-17.

package/dist/tool-map.js

@@ -1,3 +1,4 @@
+import { ocIdFor } from "./session-store.js";
 function record(value) {
     return value !== null && typeof value === "object" && !Array.isArray(value);
 }
@@ -134,13 +135,41 @@ export function mapTool(omoName, args, result) {
         ...(Object.keys(metadata).length > 0 ? { metadata } : {}),
     };
 }
-/** A subagent-style card, as OpenChamber draws `task`: it opens `sessionID` when there is one. */
-function nestedSubagent(description, prompt, type, sessionID) {
+/** "Explore Agent" for an agent type or category, "Subagent" when there is none. */
+function agentTitle(type) {
+    return type === "subagent" ? "Subagent" : `${type.charAt(0).toUpperCase()}${type.slice(1)} Agent`;
+}
+/**
+ * The task id OmO puts in a `task` result: "[task_id: st_x - " when it waited, "(st_x, running" in the
+ * background. The preview is capped at 160 code points, so only an id followed by its delimiter is whole.
+ */
+const TASK_ID = /\[task_id: (st_[A-Za-z0-9_-]+) - |\((st_[A-Za-z0-9_-]+), running\b/;
+/** The prompt's first line, cut at a word boundary near 60 characters. */
+function shortPrompt(prompt) {
+    const line = prompt
+        .split("\n")
+        .find((text) => text.trim() !== "")
+        ?.trim();
+    if (line === undefined || line.length <= 60)
+        return line;
+    const cut = line.slice(0, 60);
+    const space = cut.lastIndexOf(" ");
+    return `${(space > 30 ? cut.slice(0, space) : cut).trimEnd()}…`;
+}
+/**
+ * A subagent-style card, as OpenChamber draws `task`: it opens `sessionID` when there is one.
+ * `agentTitle`/`agentSubtitle` feed the OmO Compatibility extension's `task` header rule.
+ */
+function nestedSubagent(heading, input, sessionID) {
     return {
         tool: "task",
-        input: { description, prompt, subagent_type: type },
-        title: description.slice(0, 60),
-        metadata: sessionID !== undefined ? { sessionID } : {},
+        input: { description: input.description, prompt: input.prompt, subagent_type: input.type },
+        title: input.description.slice(0, 60),
+        metadata: {
+            agentTitle: heading.title,
+            agentSubtitle: heading.subtitle,
+            ...(sessionID !== undefined ? { sessionID } : {}),
+        },
     };
 }
 /**
@@ -154,8 +183,9 @@ export function nestedToolParts(parent, details, sessions = {}) {
     const start = parent.state.status === "pending" ? 0 : parent.state.time.start;
     const claimed = new Set();
     details.toolCalls.forEach((entry, index) => {
+        let failure;
         if (!record(entry) ||
-            (entry.name !== "monitor" && entry.name !== "__agent__") ||
+            (entry.name !== "monitor" && entry.name !== "__agent__" && entry.name !== "task") ||
             typeof entry.ok !== "boolean" ||
             typeof entry.durationMs !== "number" ||
             !Number.isFinite(entry.durationMs) ||
@@ -169,14 +199,44 @@ export function nestedToolParts(parent, details, sessions = {}) {
                 return; // nothing to open; the cell's code already shows the call
             const command = typeof args.command === "string" ? args.command : "";
             const label = typeof args.description === "string" && args.description !== "" ? args.description : command;
-            mapped = nestedSubagent(`Monitor: ${label}`, command, "monitor", sessionID);
+            mapped = nestedSubagent({ title: "Monitor", subtitle: label }, { description: `Monitor: ${label}`, prompt: command, type: "monitor" }, sessionID);
+        }
+        else if (entry.name === "task") {
+            const text = (value) => (typeof value === "string" && value !== "" ? value : undefined);
+            const prompt = text(args.prompt) ?? "";
+            const match = typeof entry.resultPreview === "string" ? TASK_ID.exec(entry.resultPreview) : null;
+            const taskID = match?.[1] ?? match?.[2];
+            if (taskID !== undefined)
+                claimed.add(taskID);
+            // A long reply pushes the id out of the preview; the task record still names the session.
+            const sessionID = !entry.ok
+                ? undefined
+                : taskID !== undefined
+                    ? ocIdFor(taskID)
+                    : (sessions.agent?.({ prompt, ...(text(args.name) ? { label: args.name } : {}) }, claimed, start) ??
+                        undefined)?.sessionID;
+            const description = text(args.task_summary) ?? text(args.description) ?? shortPrompt(prompt) ?? "subagent";
+            const type = text(args.subagent_type) ?? text(args.category) ?? "subagent";
+            mapped = nestedSubagent({ title: agentTitle(type), subtitle: description }, { description, prompt, type }, sessionID);
         }
         else {
             const prompt = typeof args.prompt === "string" ? args.prompt : "";
-            const label = [args.label, args.description, prompt.split("\n")[0]].find((value) => typeof value === "string" && value !== "");
-            const sessionID = entry.ok ? sessions.agent?.(args, claimed, start) : undefined;
-            const type = typeof args.agent === "string" && args.agent !== "" ? args.agent : "subagent";
-            mapped = nestedSubagent(label ?? "subagent", prompt, type, sessionID);
+            const found = entry.ok ? sessions.agent?.(args, claimed, start) : undefined;
+            const text = (value) => (typeof value === "string" && value !== "" ? value : undefined);
+            const description = text(args.label) ??
+                text(args.description) ??
+                found?.summary ??
+                found?.name ??
+                shortPrompt(prompt) ??
+                "subagent";
+            const type = text(args.agent) ?? found?.agentType ?? "subagent";
+            mapped = nestedSubagent({
+                title: agentTitle(type),
+                subtitle: description,
+            }, { description, prompt, type }, found?.sessionID);
+            // OmO writes the task record before agent() returns, but records a refused spawn as ok with no record.
+            if (entry.ok && found === null)
+                failure = "OmO started no subagent for this call (agent() needs an agent type or a category)";
         }
         const key = typeof entry.callId === "string" ? `id:${encodeURIComponent(entry.callId)}` : `index:${index}`;
         const callID = `${encodeURIComponent(parent.callID)}:${key}`;
@@ -187,19 +247,22 @@ export function nestedToolParts(parent, details, sessions = {}) {
             metadata: { ...mapped.metadata, nested: true, parentCallID: parent.callID, durationMs: entry.durationMs },
             time: { start, end: start + entry.durationMs },
         };
+        const failed = !entry.ok || failure !== undefined;
         parts.push({
             id: `prt_${parent.sessionID}_nested_${callID}`,
             sessionID: parent.sessionID,
             messageID: parent.messageID,
             type: "tool",
             callID: `nested:${callID}`,
-            tool: mapped.tool,
-            state: entry.ok
+            // OpenChamber 1.24 never styles a task card as failed or shows its error (ToolPart.tsx isTaskTool),
+            // and a failed spawn has no session to open, so it goes out as a plain card that does both.
+            tool: failed && mapped.tool === "task" ? "agent" : mapped.tool,
+            state: !failed
                 ? { ...state, status: "completed" }
                 : {
                     ...state,
                     status: "error",
-                    error: typeof entry.error === "string" && entry.error !== "" ? entry.error : "failed",
+                    error: failure ?? (typeof entry.error === "string" && entry.error !== "" ? entry.error : "failed"),
                 },
         });
     });

package/extensions/omo-compat/README.md

@@ -6,6 +6,10 @@ extension declare. It has no panel, page, service or permissions.
 
 What it does today:
 
+- **Subagent card headers.** A subagent an eval cell started with `agent()` is shown as a `task` card. Its header
+  reads the agent type and a short description (the call's label, the task's summary, or the prompt's first line)
+  instead of the generic "Agent Task".
+
 - **Eval cell headers.** OpenChamber 1.24 streams live output only for shell
   cards, so antechamber shows OmO eval cells as shell cards. With this extension
   their header reads `JS Eval` (or `Python Eval`) and the cell's summary instead

package/extensions/omo-compat/package.json

@@ -1,6 +1,6 @@
 {
   "name": "antechamber-omo-compat",
-  "version": "0.2.25",
+  "version": "0.2.26",
   "description": "OpenChamber extension: displays OmO tool calls the way OmO means them when OpenChamber runs on antechamber",
   "private": true,
   "type": "module",
@@ -21,6 +21,16 @@
           "match": "bash",
           "title": "{metadata.evalTitle}",
           "subtitle": "{metadata.evalSubtitle}"
+        },
+        {
+          "match": "task",
+          "title": "{metadata.agentTitle}",
+          "subtitle": "{metadata.agentSubtitle}"
+        },
+        {
+          "match": "agent",
+          "title": "{metadata.agentTitle}",
+          "subtitle": "{metadata.agentSubtitle}"
         }
       ]
     }

package/extensions/omo-dag/package.json

@@ -1,6 +1,6 @@
 {
   "name": "antechamber-omo-graph",
-  "version": "0.2.25",
+  "version": "0.2.26",
   "description": "OpenChamber extension: live, interactive graph of the OmO sub-agents behind the current session",
   "private": true,
   "type": "module",

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "antechamber",
-  "version": "0.2.25",
+  "version": "0.2.26",
   "description": "OpenCode-compatible HTTP/SSE server that lets OpenChamber drive the pi coding agent (and OmO) over RPC",
   "type": "module",
   "license": "MIT",