fix(eval): import paths, patch extraction, model configs

- Fix imports from eval.agents/eval.environments to relative (agents/environments)
- Add hatch wheel config for correct package discovery
- Extract git diff patch from container for SWE-bench submission
- Use sys.executable instead of hardcoded "python" for venv compat
- Upgrade claude-haiku config to 4.5
- Add minimax-m2.1 model config

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

eval/analysis/analyze_results.py

@@ -19,6 +19,7 @@ import json
 import logging
 import os
 import subprocess
+import sys
 from pathlib import Path
 from typing import Any
 
@@ -164,7 +165,7 @@ def run_swebench_evaluation(results_dir: Path, run_id: str, subset: str = "lite"
     try:
         eval_output = results_dir / run_id / "swebench_eval"
         cmd = [
-            "python", "-m", "swebench.harness.run_evaluation",
+            sys.executable, "-m", "swebench.harness.run_evaluation",
             "--dataset_name", dataset_mapping.get(subset, subset),
             "--predictions_path", str(preds_path),
             "--max_workers", "4",

eval/configs/models/claude-haiku.yaml

@@ -1,8 +1,7 @@
-# Claude 3.5 Haiku — fast, cheap, good baseline
+# Claude Haiku 4.5 — fast, cheap, good baseline
 # Via OpenRouter (set OPENROUTER_API_KEY in .env)
-# To use Anthropic directly, change to: anthropic/claude-3-5-haiku-20241022
 model:
-  model_name: "openrouter/anthropic/claude-3.5-haiku"
+  model_name: "openrouter/anthropic/claude-haiku-4.5"
   cost_tracking: "ignore_errors"
   model_kwargs:
     max_tokens: 8192

eval/configs/models/minimax-m2.1.yaml

@@ -0,0 +1,11 @@
+# MiniMax M2.5 — via OpenRouter (set OPENROUTER_API_KEY in .env)
+# Uses text-based model class because MiniMax doesn't support tool_calls natively.
+# The action_regex tells mini-swe-agent to parse ```bash blocks from responses.
+model:
+  model_class: litellm_textbased
+  model_name: "openrouter/minimax/minimax-m2.5"
+  action_regex: "```(?:bash|mswea_bash_command)\\s*\\n(.*?)\\n```"
+  cost_tracking: "ignore_errors"
+  model_kwargs:
+    max_tokens: 8192
+    temperature: 0

eval/pyproject.toml

@@ -30,6 +30,10 @@ gitnexus-eval-analyze = "analysis.analyze_results:app"
 requires = ["hatchling"]
 build-backend = "hatchling.build"
 
+[tool.hatch.build.targets.wheel]
+packages = ["agents", "environments", "analysis", "bridge"]
+extra-files = ["run_eval.py"]
+
 [tool.ruff]
 line-length = 120
 target-version = "py311"

eval/run_eval.py

@@ -178,7 +178,7 @@ def process_instance(
         env_class_name = env_config.pop("environment_class", "docker")
 
         if env_class_name == "eval.environments.gitnexus_docker.GitNexusDockerEnvironment":
-            from eval.environments.gitnexus_docker import GitNexusDockerEnvironment
+            from environments.gitnexus_docker import GitNexusDockerEnvironment
             env_config["image"] = get_swebench_docker_image(instance)
             env = GitNexusDockerEnvironment(**env_config)
         else:
@@ -189,7 +189,7 @@ def process_instance(
         agent_config = dict(config.get("agent", {}))
         agent_class_name = agent_config.pop("agent_class", "eval.agents.gitnexus_agent.GitNexusAgent")
 
-        from eval.agents.gitnexus_agent import GitNexusAgent
+        from agents.gitnexus_agent import GitNexusAgent
         traj_path = instance_dir / f"{instance_id}.traj.json"
         agent_config["output_path"] = traj_path
         agent = GitNexusAgent(model, env, **agent_config)
@@ -199,11 +199,18 @@ def process_instance(
         info = agent.run(instance["problem_statement"])
 
         result["exit_status"] = info.get("exit_status")
-        result["submission"] = info.get("submission", "")
         result["cost"] = agent.cost
         result["n_calls"] = agent.n_calls
         result["gitnexus_metrics"] = agent.gitnexus_metrics.to_dict()
 
+        # Extract git diff patch from the container (SWE-bench needs the model_patch)
+        try:
+            patch_output = env.execute({"command": "cd /testbed && git diff"})
+            result["submission"] = patch_output.get("output", "").strip()
+        except Exception as patch_err:
+            logger.warning(f"[{run_id}] Failed to extract patch: {patch_err}")
+            result["submission"] = info.get("submission", "")
+
     except Exception as e:
         logger.error(f"[{run_id}] Error on {instance_id}: {e}")
         result["exit_status"] = type(e).__name__