feat(agentgym): EvoSwarm → AgentGym-RL → HuggingFace publisher pipeline

.claude/context/nats-subjects.md

@@ -646,6 +646,73 @@ nats server report connections
 - `nats_server_messages_in` - Messages received
 - `nats_server_messages_out` - Messages sent
 
+## AgentGym RL Training Subjects
+
+### Training Lifecycle
+
+**`agentgym.train.started.v1`**
+- **Direction:** Published by EvoSwarm (evo-controller) → Consumed by AgentGym-RL coordinator, monitoring
+- **Purpose:** Notify that RL training has been triggered
+- **Payload:**
+  ```json
+  {
+    "training_run_id": "run-abc123",
+    "environment": "pmoves-hirag",
+    "trigger_reason": "fitness_plateau|new_constellation|scheduled|fitness_degradation",
+    "population_id": "pop-5",
+    "algorithm": "ppo|grpo|rloo|reinforce++",
+    "horizon": 10,
+    "num_epochs": 25,
+    "learning_rate": 1e-6,
+    "geometry_config": {
+      "cgp_fitness_weight": 0.2,
+      "retrieval_quality_weight": 0.3,
+      "task_success_weight": 0.4
+    },
+    "timestamp": "2026-03-14T12:00:00Z"
+  }
+  ```
+- **Subscribers:** AgentGym-RL coordinator, observability dashboards
+
+**`agentgym.train.completed.v1`**
+- **Direction:** Published by EvoSwarm (evo-controller) → Consumed by AgentGym-RL coordinator
+- **Purpose:** Training run finished — triggers auto-publish to HuggingFace Hub
+- **Payload:**
+  ```json
+  {
+    "training_run_id": "run-abc123",
+    "trajectory_ids": ["traj-1", "traj-2"],
+    "model_id": "Qwen3-8B-Instruct",
+    "population_id": "pop-5",
+    "fitness_metrics": {
+      "avg_reward": 0.82,
+      "task_success_rate": 0.91,
+      "retrieval_quality": 0.78
+    },
+    "epoch": 50,
+    "generation": 5,
+    "timestamp": "2026-03-14T14:00:00Z"
+  }
+  ```
+- **Subscribers:** AgentGym-RL coordinator (auto-publishes to HF), monitoring
+- **Triggers:** `agentgym.model.published.v1`, `skills.pipeline.model-benchmark-viz.v1`
+
+**`agentgym.model.published.v1`**
+- **Direction:** Published by AgentGym-RL coordinator → Consumed by monitoring, Agent Zero
+- **Purpose:** Model/dataset published to HuggingFace Hub
+- **Payload:**
+  ```json
+  {
+    "training_run_id": "run-abc123",
+    "model_id": "Qwen3-8B-Instruct",
+    "dataset_id": "pmoves/agentgym-run-abc123",
+    "repo_url": "https://huggingface.co/datasets/pmoves/agentgym-run-abc123",
+    "trajectory_count": 15,
+    "source": "agentgym-rl-coordinator"
+  }
+  ```
+- **Subscribers:** Agent Zero, Discord Publisher, observability dashboards
+
 ## BoTZ MCP GitHub Subjects
 
 ### GitHub Tool Execution

pmoves/env.agentgym.example

@@ -12,7 +12,11 @@ AGENTGYM_ENABLE=true
 AGENTGYM_COORDINATOR_URL=http://agentgym-rl-coordinator:8114
 
 # Base model for agent training (HuggingFace model ID or local path)
-AGENTGYM_BASE_MODEL=Qwen2.5-7B-Instruct
+AGENTGYM_BASE_MODEL=Qwen/Qwen3-8B
+
+# HuggingFace Hub credentials (required for dataset publishing)
+HF_TOKEN=  # Your HuggingFace API token
+HF_ORG=pmoves  # HuggingFace organization for published datasets
 
 # Model storage path (Docker volume mount)
 AGENTGYM_MODEL_PATH=/models

pmoves/services/agentgym-rl-coordinator/app.py

@@ -4,6 +4,7 @@
 import re
 import asyncio
 from contextlib import asynccontextmanager
+from datetime import datetime, timezone
 from typing import Optional
 from uuid import UUID
 from fastapi import FastAPI, HTTPException
@@ -142,7 +143,7 @@ async def hf_model_handler(msg):
         async def training_completed_handler(msg):
             """
             Handle a training completion event by optionally publishing associated trajectories to HuggingFace, emitting related NATS events, and recording the completion in storage.
-            
+
             Parameters:
                 msg: NATS message whose `data` is a JSON-encoded payload containing at minimum a `training_run_id` and optionally `trajectory_ids` and `model_id`.
             """
@@ -156,11 +157,25 @@ async def training_completed_handler(msg):
                     logger.warning("agentgym.train.completed.v1 missing training_run_id")
                     return
 
+                if not DATASET_NAME_PATTERN.match(training_run_id):
+                    logger.warning("Invalid training_run_id format: %s", training_run_id[:100])
+                    return
+
                 logger.info(
                     "Training completed: run=%s, trajectories=%d, model=%s",
                     training_run_id, len(trajectory_ids), model_id,
                 )
 
+                # Validate trajectory IDs as UUIDs before publishing
+                valid_trajectory_ids = []
+                for tid in trajectory_ids:
+                    try:
+                        UUID(tid)
+                        valid_trajectory_ids.append(tid)
+                    except (ValueError, AttributeError):
+                        logger.warning("Skipping invalid trajectory_id: %s", str(tid)[:100])
+                trajectory_ids = valid_trajectory_ids
+
                 # Auto-publish to HuggingFace if publisher is available
                 if hf_publisher and trajectory_ids:
                     dataset_name = f"agentgym-{training_run_id}"
@@ -184,6 +199,7 @@ async def training_completed_handler(msg):
                                 "repo_url": result.get("repo_url"),
                                 "trajectory_count": len(trajectory_ids),
                                 "source": "agentgym-rl-coordinator",
+                                "timestamp": datetime.now(timezone.utc).strftime("%Y-%m-%dT%H:%M:%SZ"),
                             })
                             await nc.publish(
                                 "agentgym.model.published.v1",
@@ -537,11 +553,11 @@ async def publish_dataset(
         for tid in trajectory_ids:
             try:
                 UUID(tid)
-            except ValueError:
+            except ValueError as e:
                 raise HTTPException(
                     status_code=400,
-                    detail=f"Invalid trajectory_id format: {tid}. Must be a valid UUID."
-                )
+                    detail=f"Invalid trajectory_id format: {tid}. Must be a valid UUID.",
+                ) from e
 
     if not hf_publisher:
         raise HTTPException(status_code=503, detail="HuggingFace publisher not available")

pmoves/services/evo-controller/agentgym_integration.py

@@ -298,7 +298,7 @@ async def launch_agentgym_training(
         if not decision.get("should_train"):
             return None
 
-        base_model = os.getenv("AGENTGYM_BASE_MODEL", "Qwen3-8B-Instruct")
+        base_model = os.getenv("AGENTGYM_BASE_MODEL", "Qwen/Qwen3-8B")
         env_namespace = os.getenv("AGENTGYM_ENV_NAMESPACE", "pmoves.consciousness")
 
         # Build training request
@@ -360,16 +360,16 @@ async def on_training_completed(
     ) -> None:
         """
         Publish a training completion event to the agent-zero event bus.
-        
+
         Sends an event with training metadata so downstream services (AgentGym-RL coordinator, publishing workflows) can process the completed run. The emitted payload includes training_run_id, trajectory_ids, model_id (resolved from the provided value or environment), population_id, fitness_metrics, epoch, generation, and an ISO-8601 UTC timestamp. On success the method logs the publication; on failure it logs a warning.
-        
+
         Parameters:
             training_run_id (str): Identifier of the completed training run.
             trajectory_ids (List[str]): List of trajectory identifiers produced by the run.
             model_id (Optional[str]): Model identifier used for training; if omitted, the base model is resolved from environment variables or a default.
             fitness_metrics (Optional[Dict[str, float]]): Final fitness metrics collected from the training; an empty dict is sent if omitted.
         """
-        base_model = model_id or os.getenv("AGENTGYM_BASE_MODEL", "Qwen3-8B-Instruct")
+        base_model = model_id or os.getenv("AGENTGYM_BASE_MODEL", "Qwen/Qwen3-8B")
         base = os.getenv("AGENT_ZERO_BASE_URL") or os.getenv("AGENTZERO_BASE_URL") or "http://agent-zero:8080"
         url = f"{base.rstrip('/')}/events/publish"