Fix MCP tool call events for Codex MCP

REPORT-codex-mcp-test-hang.md

@@ -0,0 +1,23 @@
+# Codex MCP test hang investigation
+
+## Test that appeared stuck
+- `CodexMcpAgentClient (MCP integration) > maps thread/item events for file changes, MCP tools, web search, and todo lists`
+- Wait loop: `packages/server/src/server/agent/providers/codex-mcp-agent.test.ts:577` (`for await (const event of session.stream(prompt))` until `turn_completed`/`turn_failed`)
+
+## What was actually happening
+- The test was slow (40s+) because the Codex MCP server performed multiple tool call cycles before completing the turn.
+- The run ultimately failed (not a true deadlock) because the MCP tool call timeline item never appeared, so `rawItemTypes.has("mcp_tool_call")` stayed false.
+
+## Root cause
+- Codex MCP server emits `mcp_tool_call_begin` and `mcp_tool_call_end` events with tool invocation + result payloads.
+- `codex-mcp-agent.ts` handled raw_response_item tool calls but ignored `mcp_tool_call_*` events, so MCP tool calls were never mapped to timeline items when those events were the only reliable signal.
+
+## Evidence
+- Observed event shapes from Codex MCP:
+  - `mcp_tool_call_begin` with `call_id` and `invocation` (server/tool/arguments)
+  - `mcp_tool_call_end` with `result.Ok.structuredContent`
+- When mapping those events to tool_call timeline items, the test passed.
+
+## Fix
+- Add schemas + handlers for `mcp_tool_call_begin`/`mcp_tool_call_end` in `packages/server/src/server/agent/providers/codex-mcp-agent.ts`.
+- Emit `tool_call` timeline items for running/completed MCP tool calls and surface structured tool output.

packages/server/src/server/agent/providers/codex-mcp-agent.ts

@@ -735,6 +735,46 @@ const PatchApplyEndEventSchema = z
     stderr: data.stderr,
   }));
 
+const McpToolCallInvocationSchema = z
+  .object({
+    server: z.string(),
+    tool: z.string(),
+    arguments: z.unknown().optional(),
+  })
+  .passthrough();
+
+const McpToolCallBeginEventSchema = z
+  .object({
+    type: z.literal("mcp_tool_call_begin"),
+    call_id: CallIdSchema,
+    invocation: McpToolCallInvocationSchema,
+  })
+  .passthrough()
+  .transform((data) => ({
+    type: data.type,
+    callId: data.call_id,
+    server: data.invocation.server,
+    tool: data.invocation.tool,
+    input: data.invocation.arguments,
+  }));
+
+const McpToolCallEndEventSchema = z
+  .object({
+    type: z.literal("mcp_tool_call_end"),
+    call_id: CallIdSchema,
+    invocation: McpToolCallInvocationSchema,
+    result: z.unknown().optional(),
+  })
+  .passthrough()
+  .transform((data) => ({
+    type: data.type,
+    callId: data.call_id,
+    server: data.invocation.server,
+    tool: data.invocation.tool,
+    input: data.invocation.arguments,
+    result: data.result,
+  }));
+
 const AgentMessageEventSchema = z
   .object({
     type: z.literal("agent_message"),
@@ -1667,6 +1707,8 @@ const CodexEventSchema = z.union([
   ExecCommandEndEventSchema,
   PatchApplyBeginEventSchema,
   PatchApplyEndEventSchema,
+  McpToolCallBeginEventSchema,
+  McpToolCallEndEventSchema,
   ThreadStartedEventSchema,
   TurnStartedEventSchema,
   TurnCompletedEventSchema,
@@ -1876,6 +1918,10 @@ function normalizeStructuredPayload(value: unknown): unknown {
   }
 }
 
+function isKeyedObject(value: unknown): value is { [key: string]: unknown } {
+  return typeof value === "object" && value !== null;
+}
+
 function normalizeToolName(toolName: string): string {
   if (!toolName.startsWith("mcp__")) {
     return toolName;
@@ -1889,6 +1935,44 @@ function normalizeToolName(toolName: string): string {
   return `${serverName}.${toolParts.join("__")}`;
 }
 
+function extractMcpToolResultPayload(result: unknown): {
+  output: unknown;
+  success: boolean;
+} {
+  let success = true;
+  let payload = result;
+
+  if (isKeyedObject(result)) {
+    if ("Ok" in result) {
+      payload = result.Ok;
+      success = true;
+    } else if ("ok" in result) {
+      payload = result.ok;
+      success = true;
+    } else if ("Err" in result) {
+      payload = result.Err;
+      success = false;
+    } else if ("err" in result) {
+      payload = result.err;
+      success = false;
+    }
+  }
+
+  if (isKeyedObject(payload)) {
+    if ("structuredContent" in payload && payload.structuredContent !== undefined) {
+      return { output: payload.structuredContent, success };
+    }
+    if ("structured_content" in payload && payload.structured_content !== undefined) {
+      return { output: payload.structured_content, success };
+    }
+    if ("content" in payload && payload.content !== undefined) {
+      return { output: payload.content, success };
+    }
+  }
+
+  return { output: payload, success };
+}
+
 function extractWebSearchQuery(
   input: unknown,
   output: unknown
@@ -3190,6 +3274,46 @@ class CodexMcpAgentSession implements AgentSession {
         }
         return;
       }
+      case "mcp_tool_call_begin": {
+        const input = normalizeStructuredPayload(parsedEvent.input);
+        this.emitEvent({
+          type: "timeline",
+          provider: CODEX_PROVIDER,
+          item: createToolCallTimelineItem({
+            server: parsedEvent.server,
+            tool: parsedEvent.tool,
+            status: "running",
+            callId: parsedEvent.callId,
+            displayName: `${parsedEvent.server}.${parsedEvent.tool}`,
+            kind: "tool",
+            input,
+          }),
+        });
+        return;
+      }
+      case "mcp_tool_call_end": {
+        const input = normalizeStructuredPayload(parsedEvent.input);
+        const { output, success } = extractMcpToolResultPayload(parsedEvent.result);
+        const normalizedOutput = normalizeStructuredPayload(output);
+        if (!success) {
+          this.turnState && (this.turnState.sawError = true);
+        }
+        this.emitEvent({
+          type: "timeline",
+          provider: CODEX_PROVIDER,
+          item: createToolCallTimelineItem({
+            server: parsedEvent.server,
+            tool: parsedEvent.tool,
+            status: success ? "completed" : "failed",
+            callId: parsedEvent.callId,
+            displayName: `${parsedEvent.server}.${parsedEvent.tool}`,
+            kind: "tool",
+            input,
+            output: normalizedOutput,
+          }),
+        });
+        return;
+      }
     }
   }
 

plan.md

@@ -624,6 +624,28 @@ Build a new Codex MCP provider side‑by‑side with the existing Codex SDK prov
 
 - [⏳] **Fix**: Codex MCP thread/item mapping failure in `codex-mcp-agent.test.ts` (file change, MCP tool, web search, todo list assertions).
 
+  **BREAK IT DOWN - test each event type individually:**
+
+  1. Run a minimal test for JUST `file_change`:
+     ```bash
+     npm run test --workspace=@paseo/server -- codex-mcp-agent.test.ts -t "file_change"
+     ```
+     Does Codex emit file_change events? Log raw events to see.
+
+  2. Run a minimal test for JUST `mcp_tool_call`:
+     - Does Codex emit `function_call` or `mcp_tool_call`?
+     - What's the EXACT shape of the event?
+     - Log: `console.log("RAW EVENT:", JSON.stringify(event))`
+
+  3. Run a minimal test for JUST `web_search`:
+     - What event type does Codex actually send?
+
+  4. For each: capture the EXACT raw event JSON, then fix the parser.
+
+  **Don't try to fix everything at once. Fix ONE event type, verify it works, then move to the next.**
+
+  **IMMEDIATE ACTION:** Add `console.log("RAW MCP EVENT:", JSON.stringify(msg))` in `handleMcpEvent` and run the test. Post the output.
+
 - [x] **Fix**: Typecheck error in `agent-projections.ts:198` - {} not assignable to JsonValue.
 
   - Error: TS2322 at `src/server/agent/agent-projections.ts:198:3`
@@ -648,7 +670,7 @@ Build a new Codex MCP provider side‑by‑side with the existing Codex SDK prov
   - Run `npm run typecheck --workspace=@paseo/server` → zero errors
   - **Done (2025-12-25 12:38)**: WHAT: tightened thread item schema transforms with literal `as const` types and added thread item type guards/read_file narrowing in `packages/server/src/server/agent/providers/codex-mcp-agent.ts:872`, `packages/server/src/server/agent/providers/codex-mcp-agent.ts:1453`, `packages/server/src/server/agent/providers/codex-mcp-agent.ts:3247`, `packages/server/src/server/agent/providers/codex-mcp-agent.ts:3366`. RESULT: TS2339 discriminated-union errors resolved and server typecheck completes with no errors. EVIDENCE: `npm run typecheck --workspace=@paseo/server`.
 
-- [ ] **Fix**: Test hang in `codex-mcp-agent.test.ts` - stuck at 3/13 tests.
+- [x] **Fix**: Test hang in `codex-mcp-agent.test.ts` - stuck at 3/13 tests.
 
   **DO NOT just report "it stalled" - INVESTIGATE:**
 
@@ -685,6 +707,7 @@ Build a new Codex MCP provider side‑by‑side with the existing Codex SDK prov
   - "It stalled" is NOT an acceptable answer
   - "Codex MCP server issue" is NOT an acceptable answer
   - This is OUR bug. Find it. Fix it.
+  - **Done (2025-12-25 13:03)**: WHAT: added MCP tool call begin/end event schemas + handlers and MCP result parsing in `packages/server/src/server/agent/providers/codex-mcp-agent.ts:746`, `packages/server/src/server/agent/providers/codex-mcp-agent.ts:1938`, `packages/server/src/server/agent/providers/codex-mcp-agent.ts:3277` to emit `tool_call` timeline items for MCP tools; documented investigation in `REPORT-codex-mcp-test-hang.md:1`. RESULT: the previously “stuck at 3/13” test now completes and passes with MCP tool calls recorded. EVIDENCE: `npm run test --workspace=@paseo/server -- codex-mcp-agent.test.ts -t "maps thread/item events for file changes, MCP tools, web search, and todo lists"` (1 passed, 12 skipped).
 
 - [ ] **Verify**: Rerun typecheck and full test suite after fixes.