fix(agent-classifier): recognize 'Agent session error (subtype=…)' wrapper

The error classifier keyed on `agent_status=error_max_turns` (pipeline.py's
wrapper) but missed runner.py:515's `Agent session error (subtype='error_max_turns')`.
A real max-turns failure fell through to UNKNOWN → 'Unexpected error'
(live-caught ABCA-483: a task hit the 100-turn cap but the Linear reply read
'Unexpected error'). The max_turns / max_budget / error_during_execution
patterns now match BOTH the `agent_status=` and `subtype=` wrappers.

Platform-agnostic, MAIN-BOUND (cherry-pick clean: single file +
colocated test, no Linear/orchestration coupling).

Author: isadeks

cdk/src/handlers/shared/error-classifier.ts

@@ -210,11 +210,17 @@ const PATTERNS: readonly ErrorPattern[] = [
   // ``Task did not succeed.*agent_status=`` catch-all so the concrete
   // cap / runtime-error signals surface to users rather than the
   // opaque "Agent task did not succeed" title. Each matches the
-  // ``agent_status`` literals emitted by ``agent/src/pipeline.py``
-  // (see ``_resolve_overall_task_status``) and
-  // ``agent/src/runner.py``.
+  // status literal under BOTH wrappers the agent emits:
+  //   - ``agent_status=error_max_turns`` — ``agent/src/pipeline.py``
+  //     (``_resolve_overall_task_status``); and
+  //   - ``Agent session error (subtype='error_max_turns')`` —
+  //     ``agent/src/runner.py:515`` (the terminal-error path).
+  // Keying on only ``agent_status=`` missed the ``subtype=`` wrapper, so a
+  // real max-turns failure fell through to UNKNOWN → "Unexpected error"
+  // (live-caught on ABCA-483: a task hit the 100-turn cap but the reply
+  // said "Unexpected error"). Match either ``agent_status=``/``subtype=``.
   {
-    pattern: /agent_status=['"]?error_max_turns['"]?/i,
+    pattern: /(?:agent_status|subtype)=['"]?error_max_turns['"]?/i,
     classification: {
       category: ErrorCategory.TIMEOUT,
       title: 'Exceeded max turns',
@@ -224,7 +230,7 @@ const PATTERNS: readonly ErrorPattern[] = [
     },
   },
   {
-    pattern: /agent_status=['"]?error_max_budget_usd['"]?/i,
+    pattern: /(?:agent_status|subtype)=['"]?error_max_budget_usd['"]?/i,
     classification: {
       category: ErrorCategory.TIMEOUT,
       title: 'Exceeded max budget',
@@ -234,7 +240,7 @@ const PATTERNS: readonly ErrorPattern[] = [
     },
   },
   {
-    pattern: /agent_status=['"]?error_during_execution['"]?/i,
+    pattern: /(?:agent_status|subtype)=['"]?error_during_execution['"]?/i,
     classification: {
       category: ErrorCategory.AGENT,
       title: 'Agent errored during execution',

cdk/test/handlers/shared/error-classifier.test.ts

@@ -256,6 +256,23 @@ describe('classifyError', () => {
       expect(result!.retryable).toBe(true);
     });
 
+    test('classifies the runner.py "Agent session error (subtype=...)" wrapper, not just agent_status= (K5, live-caught ABCA-483)', () => {
+      // runner.py:515 emits ``Agent session error (subtype='error_max_turns')``
+      // — a DIFFERENT wrapper from pipeline.py's ``agent_status=``. Pre-K5 this
+      // fell through to UNKNOWN → "Unexpected error" even though the task hit the
+      // 100-turn cap (live: a 1-line README task burned 101 turns, reply said
+      // "Unexpected error"). The pattern must match the subtype= wrapper too.
+      const turns = classifyError("Agent session error (subtype='error_max_turns')");
+      expect(turns!.title).toBe('Exceeded max turns');
+      expect(turns!.category).toBe(ErrorCategory.TIMEOUT);
+
+      const budget = classifyError("Agent session error (subtype='error_max_budget_usd')");
+      expect(budget!.title).toBe('Exceeded max budget');
+
+      const exec = classifyError("Agent session error (subtype='error_during_execution')");
+      expect(exec!.title).toBe('Agent errored during execution');
+    });
+
     test('matches agent_status with or without quotes around the literal', () => {
       // Defensive: the agent writer currently emits single-quoted
       // repr values (``agent_status='error_max_turns'``) but a future