fix(launcher): address PR #1176 review feedback

core.py: fix SSH user precedence — explicit user arg now wins over
slurm_config.user, which wins over getpass.getuser()

query.py:
- re-raise connection errors so datasets.map() halts the shard
  instead of silently swallowing them in worker processes
- map developer-role messages to system per OpenAI schema conventions

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Signed-off-by: chenhany <chenhany@nvidia.com>

Author: ChenhanYu

tools/launcher/common/query.py

@@ -81,6 +81,7 @@ def generate(self, messages, verbose=False, **chat_template_kwargs):
 
             if "Connection error" in str(e):
                 early_termination = True
+                raise  # propagate so datasets.map() halts the shard
 
             new_message = None
 
@@ -159,12 +160,15 @@ def synthesize(data):
                 current_messages.append(stripped)
             else:
                 current_messages.append(new_message)
+        elif role == "developer":
+            # Map developer-role messages to system per OpenAI schema conventions.
+            current_messages.append({"role": "system", "content": msg["content"]})
         elif role == "assistant":
             # Original assistant messages are not used — the model generates fresh responses.
             pass
         else:
-            # Skip unknown roles (e.g. tool) rather than raising; tool turns are
-            # part of agentic datasets but are not sent to the generation model.
+            # Skip unknown roles (e.g. tool) — agentic datasets include tool turns
+            # that are not sent to the generation model.
             pass
 
     # Restore the full reasoning trace for the last generated assistant turn.

tools/launcher/core.py

@@ -254,7 +254,7 @@ def build_slurm_executor(
 
     tunnel = run.SSHTunnel(
         host=slurm_config.host,
-        user=getattr(slurm_config, "user", None) or (getpass.getuser() if user is None else user),
+        user=user or getattr(slurm_config, "user", None) or getpass.getuser(),
         port=slurm_config.port,
         job_dir=job_dir,
         identity=identity,