Forward provider output caps in bg-review + cron agents; chunk OpenViking batch sync

- agent/background_review.py, cron/scheduler.py: forward custom_providers
  max_output_tokens so qwen-family models don't fall back to the 65,536
  family cap and 400 on long prompts against a 131k local vLLM.
- plugins/memory/openviking: chunk /messages/batch posts at 100 items so
  long sessions keep structured sync instead of the 4k text fallback.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

agent/background_review.py

@@ -87,12 +87,18 @@ def _resolve_review_runtime(agent: Any) -> Dict[str, Any]:
             explicit_api_key=task_api_key,
             explicit_base_url=task_base_url,
         )
+        # Forward the per-provider output cap (custom_providers
+        # max_output_tokens) so the routed fork doesn't fall back to the
+        # model-family output limit (e.g. qwen3 -> 65,536), which can exceed
+        # the local server's real context budget. Mirrors gateway/run.py.
+        _rt_max_tokens = rp.get("max_output_tokens")
         return {
             "provider": rp.get("provider") or task_provider,
             "model": task_model,
             "api_key": rp.get("api_key"),
             "base_url": rp.get("base_url"),
             "api_mode": rp.get("api_mode"),
+            "max_tokens": _rt_max_tokens if isinstance(_rt_max_tokens, int) and _rt_max_tokens > 0 else None,
             "routed": True,
         }
     except Exception as e:
@@ -647,6 +653,7 @@ def _run_review_in_thread(
             # (The runtime whitelist below still restricts dispatch.)
             review_agent = AIAgent(
                 model=_rt.get("model") or agent.model,
+                max_tokens=_rt.get("max_tokens"),
                 max_iterations=16,
                 quiet_mode=True,
                 platform=agent.platform,

cron/scheduler.py

@@ -2751,6 +2751,7 @@ def run_job(job: dict) -> tuple[bool, str, str, Optional[str]]:
             # example DeepSeek) for cron jobs that do not pin provider/model.
             runtime_kwargs = {
                 "requested": job.get("provider"),
+                "target_model": model or None,
             }
             if job.get("base_url"):
                 runtime_kwargs["explicit_base_url"] = job.get("base_url")
@@ -2871,8 +2872,15 @@ def run_job(job: dict) -> tuple[bool, str, str, Optional[str]]:
                 job_id, _mcp_exc,
             )
 
+        # Forward the per-provider output cap (custom_providers
+        # max_output_tokens) so cron agents don't fall back to the
+        # model-family output limit (e.g. qwen3 -> 65,536), which can
+        # exceed the local server's real context budget. Mirrors
+        # agent/background_review.py.
+        _rt_max_tokens = runtime.get("max_output_tokens")
         agent = AIAgent(
             model=model,
+            max_tokens=_rt_max_tokens if isinstance(_rt_max_tokens, int) and _rt_max_tokens > 0 else None,
             api_key=runtime.get("api_key"),
             base_url=runtime.get("base_url"),
             provider=runtime.get("provider"),

plugins/memory/openviking/__init__.py

@@ -3116,7 +3116,14 @@ class OpenVikingMemoryProvider(MemoryProvider):
                             json.dumps(payload, ensure_ascii=False),
                         )
                     try:
-                        client.post(f"/api/v1/sessions/{sid}/messages/batch", payload)
+                        # The OpenViking server caps batches at 100 messages;
+                        # chunk so long sessions keep structured sync instead
+                        # of degrading to the 4k-char text fallback.
+                        for _start in range(0, len(batch_messages), 100):
+                            client.post(
+                                f"/api/v1/sessions/{sid}/messages/batch",
+                                {"messages": batch_messages[_start:_start + 100]},
+                            )
                         return
                     except Exception as batch_error:
                         logger.warning(