diff --git a/.fallowrc.json b/.fallowrc.json
index 5fdc941a5..ff1a43b76 100644
--- a/.fallowrc.json
+++ b/.fallowrc.json
@@ -6,6 +6,9 @@
     "**/*.{test,spec}.{ts,tsx}",
     "**/*.eval.ts"
   ],
+  "ignorePatterns": [
+    ".claude/**"
+  ],
   "dynamicallyLoaded": [
     "src/deploy/runtime.ts",
     "src/deploy/runtime.worker.ts",
diff --git a/.gitignore b/.gitignore
index 29a283088..05d40d7ae 100644
--- a/.gitignore
+++ b/.gitignore
@@ -52,3 +52,6 @@ docs/api/
 
 # fallow cache/snapshots
 .fallow/
+
+# Local agent worktrees and session state
+.claude/
diff --git a/src/execution/AgentExecutor.ts b/src/execution/AgentExecutor.ts
index 51f81c8ec..71afe4d68 100644
--- a/src/execution/AgentExecutor.ts
+++ b/src/execution/AgentExecutor.ts
@@ -494,11 +494,14 @@ export class AgentExecutor {
 
       // A failed tool carries its message as `{error}` (the same shape
       // resume.ts uses) - `result` is null then, so the model would
-      // otherwise see a bare "null" and never learn the call failed.
+      // otherwise see a bare "null" and never learn the call failed. A
+      // failure that already has a structured result (argument validation)
+      // keeps it, so the model gets the per-issue detail.
       const failed = toolResult.error !== undefined;
+      const failurePayload = toolResult.result ?? { error: toolResult.error };
       state.messages.push({
         role: 'tool',
-        content: JSON.stringify(failed ? { error: toolResult.error } : toolResult.result),
+        content: JSON.stringify(failed ? failurePayload : toolResult.result),
         name: toolCall.function.name,
         toolCallId: toolCall.id,
         toolName: toolCall.function.name,
diff --git a/vitest.config.ts b/vitest.config.ts
index d34b929c4..7b4a68470 100644
--- a/vitest.config.ts
+++ b/vitest.config.ts
@@ -35,6 +35,9 @@ export default defineConfig({
       // DOM globals (File, Blob, HTMLAnchorElement, ...) those tests need.
       // Each app is tested via its own `npm run test --workspace=apps/*`.
       'apps/**',
+      // Agent worktrees (.claude/worktrees/*) are full checkouts of this repo;
+      // without this the main checkout would run every copy's tests too.
+      '.claude/**',
     ],
     coverage: {
       provider: 'v8',
@@ -49,7 +52,7 @@ export default defineConfig({
       // LOU-P's growth (new server/chat + ChatPanel/ApprovalCard code, all
       // covered by apps/agent-forge's own 49/49 + 74/74 suites) tipped the
       // denominator enough to trip the global threshold below.
-      exclude: ['**/*.test.ts', '**/*.spec.ts', 'dist/**', 'node_modules/**', 'apps/**'],
+      exclude: ['**/*.test.ts', '**/*.spec.ts', 'dist/**', 'node_modules/**', 'apps/**', '.claude/**'],
       // Floor set 1-2 points below the measured baseline (LOU-B4). Measured with
       // `npm run test:coverage`, excluding the 3 provider test suites that fail
       // to even load in this repo because @ai-sdk/openai and ollama-ai-provider