fix(agent): auto-recover from persistent 400 format errors by reloading session

When the LLM API returns a 400 "tool_call_id is not found" error, the corrupted
in-memory message state persists, blocking all subsequent messages until restart.
This adds a recovery handler that reloads sanitized messages from disk and retries,
similar to the existing context overflow recovery pattern.

Also adds `default` export condition to core package.json to fix tsx watch
resolution for subpath exports (gateway startup failure).

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

packages/core/package.json

@@ -5,35 +5,43 @@
   "exports": {
     ".": {
       "types": "./dist/index.d.ts",
-      "import": "./dist/index.js"
+      "import": "./dist/index.js",
+      "default": "./dist/index.js"
     },
     "./agent": {
       "types": "./dist/agent/index.d.ts",
-      "import": "./dist/agent/index.js"
+      "import": "./dist/agent/index.js",
+      "default": "./dist/agent/index.js"
     },
     "./hub": {
       "types": "./dist/hub/index.d.ts",
-      "import": "./dist/hub/index.js"
+      "import": "./dist/hub/index.js",
+      "default": "./dist/hub/index.js"
     },
     "./channels": {
       "types": "./dist/channels/index.d.ts",
-      "import": "./dist/channels/index.js"
+      "import": "./dist/channels/index.js",
+      "default": "./dist/channels/index.js"
     },
     "./cron": {
       "types": "./dist/cron/index.d.ts",
-      "import": "./dist/cron/index.js"
+      "import": "./dist/cron/index.js",
+      "default": "./dist/cron/index.js"
     },
     "./heartbeat": {
       "types": "./dist/heartbeat/index.d.ts",
-      "import": "./dist/heartbeat/index.js"
+      "import": "./dist/heartbeat/index.js",
+      "default": "./dist/heartbeat/index.js"
     },
     "./media": {
       "types": "./dist/media/index.d.ts",
-      "import": "./dist/media/index.js"
+      "import": "./dist/media/index.js",
+      "default": "./dist/media/index.js"
     },
     "./client": {
       "types": "./dist/client/index.d.ts",
-      "import": "./dist/client/index.js"
+      "import": "./dist/client/index.js",
+      "default": "./dist/client/index.js"
     }
   },
   "main": "./dist/index.js",

packages/core/src/agent/auth-profiles/error-classification.test.ts

@@ -20,6 +20,13 @@ describe("classifyError", () => {
     expect(classifyError(new Error("Schema validation failed"))).toBe("format");
   });
 
+  it("classifies tool_call_id 400 errors as format (recoverable via transcript repair)", () => {
+    expect(
+      classifyError(new Error("400 tool_call_id  is not found in the list of tool calls")),
+    ).toBe("format");
+    expect(classifyError(new Error("400 Bad Request: tool_call_id not found"))).toBe("format");
+  });
+
   it("classifies 429/rate limit as rate_limit", () => {
     expect(classifyError(new Error("429 Too Many Requests"))).toBe("rate_limit");
     expect(classifyError(new Error("Rate limit exceeded"))).toBe("rate_limit");

packages/core/src/agent/runner.ts

@@ -794,6 +794,8 @@ export class Agent {
 
       const MAX_OVERFLOW_COMPACTION_ATTEMPTS = 2;
       let overflowAttempts = 0;
+      const MAX_FORMAT_REPAIR_ATTEMPTS = 1;
+      let formatRepairAttempts = 0;
 
       // Loop to exhaust all candidate profiles on rotatable errors
       while (true) {
@@ -860,6 +862,31 @@ export class Agent {
             reason,
             rotatable: isRotatableError(reason),
           });
+
+          // Format error recovery: reload sanitized messages from disk and retry.
+          // This handles corrupted in-memory state (e.g. orphaned tool_call_id references)
+          // that causes persistent 400 errors until process restart.
+          if (reason === "format" && formatRepairAttempts < MAX_FORMAT_REPAIR_ATTEMPTS) {
+            formatRepairAttempts++;
+            this.stderr.write(
+              `[format-repair] Format error detected (attempt ${formatRepairAttempts}/${MAX_FORMAT_REPAIR_ATTEMPTS}), reloading messages from disk...\n`,
+            );
+            this.runLog.log("format_repair", {
+              attempt: formatRepairAttempts,
+              error: errorMsg.slice(0, 200),
+              messages_before: this.agent.state.messages.length,
+            });
+            const repairedMessages = this.session.loadMessages();
+            if (repairedMessages.length > 0) {
+              this.runLog.log("format_repair_reloaded", {
+                messages_after: repairedMessages.length,
+              });
+              this.agent.replaceMessages(repairedMessages);
+              this.output.state.lastAssistantText = "";
+              continue; // retry with sanitized messages
+            }
+          }
+
           if (this.currentProfileId && isRotatableError(reason)) {
             markAuthProfileFailure(this.currentProfileId, reason);
           }