Normalize Codex paseo_voice.speak MCP calls and wrap spoken input

Broaden the Codex speak tool name matcher to recognize paseo_voice.speak
MCP calls alongside the existing paseo.speak convention. Wrap voice
transcription input in <spoken-input> tags so agents can distinguish
spoken from typed messages.

packages/server/src/server/agent/providers/codex/tool-call-mapper.test.ts

@@ -539,6 +539,26 @@ describe("codex tool-call mapper", () => {
     });
   });
 
+  it("normalizes codex paseo_voice.speak mcp calls and extracts spoken text", () => {
+    const item = mapCodexToolCallFromThreadItem({
+      type: "mcpToolCall",
+      id: "codex-speak-thread-2",
+      status: "completed",
+      server: "paseo_voice",
+      tool: "speak",
+      arguments: { text: "Voice response from Codex via paseo_voice." },
+      result: { ok: true },
+    });
+
+    expect(item).toBeTruthy();
+    expect(item?.name).toBe("speak");
+    expect(item?.detail).toEqual({
+      type: "unknown",
+      input: "Voice response from Codex via paseo_voice.",
+      output: null,
+    });
+  });
+
   it("normalizes codex paseo speak rollout names and extracts spoken text", () => {
     const item = expectMapped(
       mapCodexRolloutToolCall({

packages/server/src/server/agent/providers/codex/tool-call-mapper.ts

@@ -3,6 +3,7 @@ import { z } from "zod";
 import type { ToolCallTimelineItem } from "../../agent-sdk-types.js";
 import { extractCodexShellOutput, truncateDiffText } from "../tool-call-mapper-utils.js";
 import { deriveCodexToolDetail, normalizeCodexFilePath } from "./tool-call-detail-parser.js";
+import { isSpeakToolName } from "../../tool-name-normalization.js";
 
 type CodexMapperOptions = { cwd?: string | null };
 
@@ -67,7 +68,7 @@ const CodexEditToolNameSchema = z.union([
   z.literal("apply_diff"),
 ]);
 const CodexSearchToolNameSchema = z.union([z.literal("search"), z.literal("web_search")]);
-const CodexSpeakToolNameSchema = z.literal("paseo.speak");
+const CodexSpeakToolNameSchema = z.string().min(1).refine((name) => isSpeakToolName(name.trim()));
 
 const CodexToolKindSchema = z.enum([
   "shell",

packages/server/src/server/session.ts

@@ -133,6 +133,7 @@ import {
   buildVoiceAgentMcpServerConfig,
   buildVoiceModeSystemPrompt,
   stripVoiceModeSystemPrompt,
+  wrapSpokenInput,
 } from "./voice-config.js";
 import { isVoicePermissionAllowed } from "./voice-permission-policy.js";
 import {
@@ -2834,6 +2835,7 @@ export class Session {
     messageId?: string,
     images?: Array<{ data: string; mimeType: string }>,
     runOptions?: AgentRunOptions,
+    options?: { spokenInput?: boolean },
   ): Promise<void> {
     this.sessionLogger.info(
       { agentId, textPreview: text.substring(0, 50), imageCount: images?.length ?? 0 },
@@ -2861,7 +2863,8 @@ export class Session {
       );
     }
 
-    const prompt = this.buildAgentPrompt(text, images);
+    const promptText = options?.spokenInput ? wrapSpokenInput(text) : text;
+    const prompt = this.buildAgentPrompt(promptText, images);
 
     this.startAgentStream(agentId, prompt, runOptions);
   }
@@ -7019,9 +7022,14 @@ export class Session {
       return;
     }
 
-    // Route voice utterances through the same send path as regular text input:
-    // interrupt-if-running, record message, then start a new stream.
-    await this.handleSendAgentMessage(agentId, result.text);
+    await this.handleSendAgentMessage(
+      agentId,
+      result.text,
+      undefined,
+      undefined,
+      undefined,
+      { spokenInput: true },
+    );
     await this.flushPendingAudioSegments("transcription complete");
   }
 

packages/server/src/server/voice-config.ts

@@ -53,6 +53,10 @@ export function buildVoiceModeSystemPrompt(existing: string | undefined, enabled
     .join("\n\n");
 }
 
+export function wrapSpokenInput(text: string): string {
+  return `<spoken-input>\n${text}\n</spoken-input>`;
+}
+
 export function buildVoiceAgentMcpServerConfig(params: {
   command: string;
   baseArgs: string[];