From 2e5fd9f2b45b1c61c9e205e0de6328f2f37e35c4 Mon Sep 17 00:00:00 2001
From: HiranoMasaaki <lambda.groove@gmail.com>
Date: Thu, 12 Mar 2026 04:29:50 +0900
Subject: [PATCH] fix: persist startRun and stopRunByError events for failed
 delegations

When a delegate's subprocess failed to start (e.g., MCP connection
failure, executable not found), no events were persisted because the
state machine never started. The error was silently converted to a
tool result text and returned to the parent expert.

Now the catch block in DelegationExecutor.executeSingleDelegation()
emits startRun + stopRunByError events via the parent's storeEvent
and eventListener callbacks, ensuring failed delegations are recorded
in the event store and visible in log viewers.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---
 .changeset/persist-failed-delegation.md       |  5 ++
 .../src/orchestration/delegation-executor.ts  | 56 ++++++++++++++++++-
 2 files changed, 58 insertions(+), 3 deletions(-)
 create mode 100644 .changeset/persist-failed-delegation.md

diff --git a/.changeset/persist-failed-delegation.md b/.changeset/persist-failed-delegation.md
new file mode 100644
index 00000000..347f2a68
--- /dev/null
+++ b/.changeset/persist-failed-delegation.md
@@ -0,0 +1,5 @@
+---
+"@perstack/runtime": patch
+---
+
+Persist startRun and stopRunByError events for failed delegations
diff --git a/packages/runtime/src/orchestration/delegation-executor.ts b/packages/runtime/src/orchestration/delegation-executor.ts
index 318d112e..ad7d3fa1 100644
--- a/packages/runtime/src/orchestration/delegation-executor.ts
+++ b/packages/runtime/src/orchestration/delegation-executor.ts
@@ -13,7 +13,12 @@ import type {
   ToolResult,
   Usage,
 } from "@perstack/core"
-import { resolveModelTier, resolveModelTierWithFallback } from "@perstack/core"
+import {
+  resolveModelTier,
+  resolveModelTierWithFallback,
+  startRun,
+  stopRunByError,
+} from "@perstack/core"
 
 /** Reference to the parent Expert that delegated */
 type DelegatedBy = NonNullable<Checkpoint["delegatedBy"]>
@@ -247,12 +252,57 @@ export class DelegationExecutor {
         { ...parentOptions, returnOnDelegationComplete: true },
       )
     } catch (error) {
-      // Child run crashed (e.g., MCP connection failure) - return error to parent
+      // Child run crashed before state machine started (e.g., MCP connection failure).
+      // Emit startRun + stopRunByError events so the failure is persisted in the event store.
+      const errorMessage = error instanceof Error ? error.message : String(error)
+      const errorName = error instanceof Error ? error.name : "Error"
+
+      const startRunEvent = startRun(delegateSetting, delegateCheckpoint, {
+        initialCheckpoint: delegateCheckpoint,
+        inputMessages: [],
+        model: delegateSetting.model ?? "",
+      })
+
+      const errorCheckpoint: Checkpoint = {
+        ...delegateCheckpoint,
+        status: "stoppedByError",
+        error: {
+          name: errorName,
+          message: errorMessage,
+          isRetryable: false,
+        },
+      }
+
+      const stopRunByErrorEvent = stopRunByError(delegateSetting, errorCheckpoint, {
+        checkpoint: errorCheckpoint,
+        step: {
+          stepNumber: 1,
+          newMessages: [],
+          usage: createEmptyUsage(),
+          startedAt: Date.now(),
+        },
+        error: {
+          name: errorName,
+          message: errorMessage,
+          isRetryable: false,
+        },
+      })
+
+      // Persist events via parent's callbacks
+      if (parentOptions?.storeEvent) {
+        await parentOptions.storeEvent(startRunEvent).catch(() => {})
+        await parentOptions.storeEvent(stopRunByErrorEvent).catch(() => {})
+      }
+      if (parentOptions?.eventListener) {
+        parentOptions.eventListener(startRunEvent)
+        parentOptions.eventListener(stopRunByErrorEvent)
+      }
+
       return {
         toolCallId,
         toolName,
         expertKey: expert.key,
-        text: `Delegation to ${expert.key} failed: ${error instanceof Error ? error.message : String(error)}`,
+        text: `Delegation to ${expert.key} failed: ${errorMessage}`,
         stepNumber: parentContext.stepNumber,
         deltaUsage: createEmptyUsage(),
       }