fix(acp): close open tool rows with the turn and register the Hoplite icon

The ACP shim only terminalized its own tool calls when a turn ended, so a
running row published by a transformer stayed running after a cancel or
failure. Track transformer tool rows and close them with the turn.

A canceled turn also marked open tool calls failed with a null error,
which the wire schema rejects, making the agent's timeline unfetchable.
Cancel those rows instead, and give failed rows a non-null error.

Add `hoplite` to the known ACP provider icon names so the catalog icon
resolves without a snapshot SVG, and stop listing Hoplite under the
promise that local MCP servers stay intact.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

packages/plugin/src/server/acp-internal/connection.ts

@@ -358,6 +358,8 @@ class AcpRuntime {
   private emit: (event: ProviderEvent) => void;
   private readonly messages = new Map<string, string>();
   private readonly toolCalls = new Map<string, AcpToolCallSnapshot>();
+  // Running tool rows published by transformers; the turn's terminal closes any left open.
+  private readonly vendorToolCalls = new Map<string, ToolCallTimelineItem>();
   private readonly pendingCompactions = new Set<string>();
   private readonly permissions = new Map<
     string,
@@ -1026,9 +1028,17 @@ class AcpRuntime {
       this.emit({
         type: "timeline.item",
         sessionId: this.options.boundarySessionId,
-        item: toolTimelineItem(completed),
+        item: terminalToolItem(toolTimelineItem(snapshot), state),
       });
     }
+    for (const item of this.vendorToolCalls.values()) {
+      this.emit({
+        type: "timeline.item",
+        sessionId: this.options.boundarySessionId,
+        item: terminalToolItem(item, state),
+      });
+    }
+    this.vendorToolCalls.clear();
     for (const id of this.pendingCompactions) {
       this.emit({
         type: "timeline.item",
@@ -1075,6 +1085,10 @@ class AcpRuntime {
       if (this.configTransaction) this.stagedTransformerConfig = update.config;
       else this.emitConfigSnapshot(update.config);
     } else if (update.type === "timeline") {
+      if (update.item.type === "tool_call") {
+        if (update.item.status === "running") this.vendorToolCalls.set(update.item.id, update.item);
+        else this.vendorToolCalls.delete(update.item.id);
+      }
       this.emit({
         type: "timeline.item",
         sessionId: this.options.boundarySessionId,
@@ -1382,7 +1396,9 @@ function configValues(
   return Object.fromEntries(options.map((option) => [option.id, option.currentValue]));
 }
 
-function toolTimelineItem(snapshot: AcpToolCallSnapshot): ProviderTimelineItem {
+type ToolCallTimelineItem = Extract<ProviderTimelineItem, { type: "tool_call" }>;
+
+function toolTimelineItem(snapshot: AcpToolCallSnapshot): ToolCallTimelineItem {
   const input = jsonRecord(snapshot.input);
   const isEdit = snapshot.kind === "edit" || snapshot.name?.toLowerCase().includes("edit") === true;
   const detail = isEdit
@@ -1410,6 +1426,18 @@ function toolTimelineItem(snapshot: AcpToolCallSnapshot): ProviderTimelineItem {
   return { ...base, status, error: null };
 }
 
+// A row the turn left open ends with the turn. The wire contract requires a non-null error on a
+// failed tool call, so a canceled turn cancels its rows instead of failing them.
+function terminalToolItem(
+  item: ToolCallTimelineItem,
+  state: "completed" | "failed" | "canceled",
+): ToolCallTimelineItem {
+  if (state === "failed") {
+    return { ...item, status: "failed", error: item.error ?? { message: "Turn failed" } };
+  }
+  return { ...item, status: state, error: null };
+}
+
 function jsonValue(value: unknown) {
   return z.json().parse(value);
 }

packages/plugin/src/server/acp.test.ts

@@ -619,6 +619,146 @@ lines.on("line", (line) => {
     await connection.close();
   });
 
+  it.each([
+    ["end_turn", "completed", "completed"],
+    ["cancelled", "canceled", "canceled"],
+  ] as const)(
+    "closes a transformer's running tool row when the turn ends with %s",
+    async (stopReason, turnState, rowStatus) => {
+      const harness = connectorHarness({
+        handleMessage(instance, message) {
+          if (!("method" in message) || !("id" in message)) return false;
+          if (message.method !== "session/prompt") return false;
+          instance.notify("vendor/phase", { id: "phase-1", label: "Preparing" });
+          instance.respond(message as AcpRequestMessage, { stopReason });
+          return true;
+        },
+      });
+      const registration = runAcpProvider({
+        id: "vendor-row-acp",
+        label: "Vendor row ACP",
+        connector: harness.connector,
+        transformers: [
+          {
+            notification(notification) {
+              if (notification.method !== "vendor/phase") return null;
+              return {
+                type: "timeline",
+                item: {
+                  type: "tool_call",
+                  id: "phase-1",
+                  callId: "phase-1",
+                  name: "Vendor phase",
+                  detail: { type: "plain_text", label: "Preparing" },
+                  status: "running",
+                  error: null,
+                },
+              };
+            },
+          },
+        ],
+      });
+      const connection = await registration.connect({
+        versions: [1],
+        capabilities: ["prompt.message"],
+      });
+      const events: ProviderEvent[] = [];
+      connection.onEvent((event) => events.push(event));
+      await connection.send(openInput());
+      await waitForEvent(events, (event) => event.type === "session.ready");
+
+      await connection.send({
+        type: "session.prompt",
+        sessionId: "session-1",
+        prompt: {
+          clientMessageId: "first",
+          delivery: "auto",
+          input: { type: "message", content: [{ type: "text", text: "first" }] },
+        },
+      });
+      await waitForEvent(
+        events,
+        (event) => event.type === "session.turn" && event.state === turnState,
+      );
+
+      const rows = events.flatMap((event) =>
+        event.type === "timeline.item" && event.item.id === "phase-1" ? [event.item] : [],
+      );
+      expect(rows.map((row) => row.type === "tool_call" && row.status)).toEqual([
+        "running",
+        rowStatus,
+      ]);
+      const closed = events.findIndex(
+        (event) =>
+          event.type === "timeline.item" &&
+          event.item.id === "phase-1" &&
+          event.item.type === "tool_call" &&
+          event.item.status === rowStatus,
+      );
+      const terminal = events.findIndex(
+        (event) => event.type === "session.turn" && event.state === turnState,
+      );
+      expect(closed).toBeLessThan(terminal);
+      // The wire contract only accepts a failed tool call with a non-null error.
+      expect(rows.at(-1)).toMatchObject({ error: null });
+      await connection.close();
+    },
+  );
+
+  it("cancels an in-progress ACP tool call with the canceled turn instead of failing it without an error", async () => {
+    const harness = connectorHarness({
+      handleMessage(instance, message) {
+        if (!("method" in message) || !("id" in message)) return false;
+        if (message.method !== "session/prompt") return false;
+        instance.notify("session/update", {
+          sessionId: "connector-session",
+          update: {
+            sessionUpdate: "tool_call",
+            toolCallId: "tool-1",
+            title: "Run tests",
+            kind: "execute",
+            status: "in_progress",
+          },
+        });
+        instance.respond(message as AcpRequestMessage, { stopReason: "cancelled" });
+        return true;
+      },
+    });
+    const registration = runAcpProvider({
+      id: "cancel-tool-acp",
+      label: "Cancel tool ACP",
+      connector: harness.connector,
+    });
+    const connection = await registration.connect({
+      versions: [1],
+      capabilities: ["prompt.message"],
+    });
+    const events: ProviderEvent[] = [];
+    connection.onEvent((event) => events.push(event));
+    await connection.send(openInput());
+    await waitForEvent(events, (event) => event.type === "session.ready");
+
+    await connection.send({
+      type: "session.prompt",
+      sessionId: "session-1",
+      prompt: {
+        clientMessageId: "first",
+        delivery: "auto",
+        input: { type: "message", content: [{ type: "text", text: "first" }] },
+      },
+    });
+    await waitForEvent(
+      events,
+      (event) => event.type === "session.turn" && event.state === "canceled",
+    );
+
+    const rows = events.flatMap((event) =>
+      event.type === "timeline.item" && event.item.id === "tool-1" ? [event.item] : [],
+    );
+    expect(rows.at(-1)).toMatchObject({ type: "tool_call", status: "canceled", error: null });
+    await connection.close();
+  });
+
   it("serializes configuration mutations", async () => {
     let releaseMutation: (() => void) | null = null;
     const configOptions = [

packages/protocol/src/provider-icon-names.ts

@@ -31,6 +31,7 @@ export const ACP_PROVIDER_ICON_NAMES = [
   "glm-acp-agent",
   "goose",
   "grok",
+  "hoplite",
   "junie",
   "kilo",
   "kimi",

public-docs/providers.md

@@ -8,7 +8,7 @@ category: Providers
 
 # Providers
 
-Paseo doesn't ship its own coding agent. It launches and supervises **existing CLIs you've already installed and authenticated**, Claude Code, Codex, OpenCode, Antigravity, Muse Code, Hoplite, Cursor, Gemini, and the rest. Your subscriptions, your config, your skills, your MCP servers all stay intact. Paseo just gives you a UI, a CLI, a relay, and orchestration on top.
+Paseo doesn't ship its own coding agent. It launches and supervises **existing CLIs you've already installed and authenticated**, Claude Code, Codex, OpenCode, Antigravity, Muse Code, Cursor, Gemini, and the rest. Your subscriptions, your config, your skills, your MCP servers all stay intact. Paseo just gives you a UI, a CLI, a relay, and orchestration on top.
 
 ## Mental model
 
@@ -16,7 +16,7 @@ A provider is the contract between Paseo and one external agent CLI: how to laun
 
 ## Two tiers
 
-- **Native support**, Paseo ships a bundled adapter for the major agents (Claude Code, Codex, OpenCode, Pi, [Antigravity](/docs/supported-providers#antigravity), [Muse Code](/docs/muse-code), [Hoplite](/docs/supported-providers#hoplite)). Auto-discovered when the underlying CLI is installed, with mode metadata and voice support where applicable.
+- **Native support**, Paseo ships a bundled adapter for the major agents (Claude Code, Codex, OpenCode, Pi, [Antigravity](/docs/supported-providers#antigravity), [Muse Code](/docs/muse-code), [Hoplite](/docs/supported-providers#hoplite)). Auto-discovered when the underlying CLI is installed, with mode metadata and voice support where applicable. Hoplite's CLI runs on your machine, but its agent works in a Hoplite cloud workspace, so your local MCP servers and Paseo tools are not available there.
 - **ACP catalog**, any agent speaking the [Agent Client Protocol](https://agentclientprotocol.com) is supported through a generic adapter. Paseo ships a curated catalog of one-click installs (Cursor, Gemini, GitHub Copilot, Hermes, Kimi, Qwen Code, and 25+ more), and you can add any other ACP agent yourself.
 
 Either way, **you install the underlying CLI**. Paseo runs it.