Surface per-step budget in agent observations

Add a global observation_template in configs/mini/default.yaml with a
[Step n/limit used - k left] banner, and thread it onto the model via a
ClashAgentConfig field pushed in ClashAgent.__init__. Applies to every model
backend, tournament type, and arena; models otherwise learn the step limit only
once at round start with no running counter.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Author: john-b-yang

codeclash/agents/minisweagent.py

@@ -18,6 +18,16 @@
 )
 
 
+class ClashAgentConfig(AgentConfig):
+    """`AgentConfig` plus an optional global observation-format override.
+
+    `observation_template` is a model-side field in mini-swe-agent, so putting it in the shared
+    agent config (`configs/mini/default.yaml`) doesn't reach the model on its own. Declaring it
+    here lets the one global config drive it; `ClashAgent` pushes it onto the model at init."""
+
+    observation_template: str | None = None
+
+
 class ClashAgent(DefaultAgent):
     """`DefaultAgent` from mini-SWE-agent (https://github.com/SWE-agent/mini-swe-agent)
     with per-player debug logging."""
@@ -28,11 +38,15 @@ def __init__(
         env: DockerEnvironment,
         *,
         logger: logging.Logger,
-        config_class: Callable = AgentConfig,
+        config_class: Callable = ClashAgentConfig,
         **kwargs,
     ):
         super().__init__(model, env, config_class=config_class, **kwargs)
         self.logger = logger
+        # Honor the global observation_template (agent config) by pushing it onto the model,
+        # which is what actually renders observations. Applies to every backend uniformly.
+        if getattr(self.config, "observation_template", None):
+            self.model.config.observation_template = self.config.observation_template
 
     def add_messages(self, *messages: dict) -> list[dict]:
         result = super().add_messages(*messages)

configs/mini/default.yaml

@@ -102,3 +102,7 @@ instance_template: |
   You cannot continue working in any way on this task after submitting.
 step_limit: 30
 cost_limit: 0
+# Global observation format for every model/tournament/arena. ClashAgent pushes this onto the
+# model config (mini-swe-agent renders observations model-side). The leading banner surfaces the
+# per-round step budget so the model can pace itself; the rest is mini's default observation body.
+observation_template: "[Step {{n_model_calls}}/{{step_limit}} used - {{step_limit - n_model_calls}} left]\n{% if output.exception_info %}<exception>{{output.exception_info}}</exception>\n{% endif %}<returncode>{{output.returncode}}</returncode>\n<output>\n{{output.output}}</output>"