Render one assistant message per turn, and make the request timeout a setting

Two setup failures found while calibrating the new benchmark tasks,
both of which would also fail real milestones.

Devstral's chat template (Mistral) rejected any turn in which the model
wrote some text before a tool call: the OpenAI renderer made one message
per event, so the turn became a text message and a tool-call message in
a row, and llama.cpp answered 500. The renderer now folds consecutive
user or assistant messages into one, which is the OpenAI format for a
turn anyway. Checked against the live server: the captured failing
request no longer occurs, over six real runs.

LiteLLMClient gave up on any request after a fixed 120 s. A dense 27B
model writing a whole module in one turn takes longer, so the executor
escalated work that was going well. llm.request_timeout_s now sets it
for both clients, default 600.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

nisse-dev/src/nisse_dev/agent.py

@@ -144,7 +144,10 @@ def make_llm_clients(
 
     normalized = budgeted(
         LiteLLMClient(
-            config.llm_base_url, config.llm_token, max_tokens=config.max_output_tokens
+            config.llm_base_url,
+            config.llm_token,
+            max_tokens=config.max_output_tokens,
+            timeout_s=config.request_timeout_s,
         )
     )
     # Delegations prefer the vendor-native route; fidelity is cheap
@@ -152,7 +155,10 @@ def make_llm_clients(
     # select_client, not by where the caller happens to sit.
     passthrough = budgeted(
         AnthropicPassthroughClient(
-            config.llm_base_url, config.llm_token, max_tokens=config.max_output_tokens
+            config.llm_base_url,
+            config.llm_token,
+            max_tokens=config.max_output_tokens,
+            timeout_s=config.request_timeout_s,
         )
     )
     return normalized, passthrough

nisse-dev/src/nisse_dev/config.py

@@ -23,7 +23,7 @@ from typing import Any
 
 import yaml
 
-from nisse_dev.llm import DEFAULT_MAX_OUTPUT_TOKENS, RoleMap
+from nisse_dev.llm import DEFAULT_MAX_OUTPUT_TOKENS, DEFAULT_REQUEST_TIMEOUT_S, RoleMap
 from nisse_dev.matrix import MatrixConfig
 
 DEFAULT_CONFIG_FILE = "nisse-dev.yaml"
@@ -48,7 +48,7 @@ REQUIRED_KEYS = (
     "gitea.repo",
     "gitea.operators",
 )
-OPTIONAL_KEYS = ("llm.max_output_tokens", "gitea.base_branch")
+OPTIONAL_KEYS = ("llm.max_output_tokens", "llm.request_timeout_s", "gitea.base_branch")
 
 _REFERENCE = re.compile(r"\$\{([A-Za-z_][A-Za-z0-9_]*)\}")
 
@@ -221,6 +221,7 @@ class AgentConfig:
     gitea_operators: frozenset[str]
     base_branch: str = "main"
     max_output_tokens: int = DEFAULT_MAX_OUTPUT_TOKENS
+    request_timeout_s: float = DEFAULT_REQUEST_TIMEOUT_S
     tuning: Tuning = field(default_factory=Tuning)
     config_file: Path | None = None
 
@@ -291,6 +292,14 @@ class AgentConfig:
                 raise ValueError
         except (TypeError, ValueError):
             errors.append("llm.max_output_tokens: must be a positive integer")
+
+        request_timeout = values.get("llm.request_timeout_s", DEFAULT_REQUEST_TIMEOUT_S)
+        try:
+            request_timeout = float(request_timeout)
+            if request_timeout <= 0:
+                raise ValueError
+        except (TypeError, ValueError):
+            errors.append("llm.request_timeout_s: must be a positive number of seconds")
         if errors:
             raise ConfigError("cannot start: " + "; ".join(errors))
 
@@ -319,6 +328,7 @@ class AgentConfig:
             gitea_operators=_people(values["gitea.operators"]),
             base_branch=str(values.get("gitea.base_branch") or "main"),
             max_output_tokens=max_output,
+            request_timeout_s=request_timeout,
             tuning=tuning,
             config_file=config_file,
         )

nisse-dev/src/nisse_dev/llm.py

@@ -46,6 +46,12 @@ RATE_LIMIT_RETRY_S = 20.0
 # was observed at 4,064 against a 4,096 limit.
 DEFAULT_MAX_OUTPUT_TOKENS = 16384
 
+# How long one request may take before it is abandoned (llm.request_timeout_s).
+# A safety net for a hung connection, not a budget: a dense local model
+# writing a whole module in one turn was measured taking over 120 s, and
+# timing it out failed work that was going well.
+DEFAULT_REQUEST_TIMEOUT_S = 600.0
+
 
 class UpstreamUnavailable(LLMError):
     """The proxy or provider could not be reached, or refused for reasons
@@ -124,7 +130,7 @@ class LiteLLMClient:
         base_url: str,
         token: str,
         *,
-        timeout_s: float = 120.0,
+        timeout_s: float = DEFAULT_REQUEST_TIMEOUT_S,
         transport: httpx.BaseTransport | None = None,
         sleep: Any = time.sleep,
         max_tokens: int | None = None,
@@ -250,7 +256,7 @@ class AnthropicPassthroughClient:
         base_url: str,
         token: str,
         *,
-        timeout_s: float = 300.0,
+        timeout_s: float = DEFAULT_REQUEST_TIMEOUT_S,
         max_tokens: int | None = None,
         transport: httpx.BaseTransport | None = None,
         sleep: Any = time.sleep,

nisse-dev/src/nisse_dev/render.py

@@ -60,7 +60,7 @@ class OpenAIChatRenderer:
         kept = effective_events(events)
         answered = {e.event.call_id for e in kept if isinstance(e.event, ToolResult)}
         for stored in kept:
-            messages.append(self._message_for(stored.event))
+            _append_merged(messages, self._message_for(stored.event))
             # Providers reject a tool call with no result following it. One
             # can legitimately be missing: a crash between the call and its
             # result (A3), or a turn-ending gate tool. Render a placeholder so
@@ -128,3 +128,28 @@ class OpenAIChatRenderer:
                 }
             case _:
                 raise ValueError(f"unrenderable event: {event!r}")
+
+
+def _append_merged(messages: list[dict[str, Any]], message: dict[str, Any]) -> None:
+    """Append, folding into the previous message when both are the same user
+    or assistant role.
+
+    One model turn is often several events (some text, then one or more tool
+    calls), and the OpenAI format carries a turn as one assistant message
+    with a tool_calls list. Strict chat templates enforce that: Devstral's
+    (Mistral) rejected a turn rendered as a text message then a tool-call
+    message with a 500, failing runs whenever the model spoke before acting."""
+    previous = messages[-1] if messages else None
+    role = message["role"]
+    if (
+        previous is None
+        or previous["role"] != role
+        or role not in ("user", "assistant")
+    ):
+        messages.append(message)
+        return
+    texts = [t for t in (previous.get("content"), message.get("content")) if t]
+    previous["content"] = "\n\n".join(texts) if texts else None
+    calls = previous.get("tool_calls", []) + message.get("tool_calls", [])
+    if calls:
+        previous["tool_calls"] = calls