feat: route codex patch repair through responses api

packages/casan-harness/scripts/bash/model-call.py

@@ -239,6 +239,58 @@ def call_openai(model_name, prompt, role):
     }
 
 
+def parse_responses_payload(payload):
+    try:
+        text = str(payload.get("output_text") or "").strip()
+        if not text:
+            text = "".join(
+                str(part.get("text") or "")
+                for item in payload.get("output", []) if isinstance(item, dict)
+                for part in item.get("content", []) if isinstance(part, dict) and part.get("type") == "output_text"
+            ).strip()
+        usage = payload["usage"]
+        input_tokens = int(usage["input_tokens"])
+        output_tokens = int(usage["output_tokens"])
+        if not text:
+            raise ValueError("output text missing")
+    except (KeyError, TypeError, ValueError) as exc:
+        fail(f"provider_usage_invalid openai-responses {type(exc).__name__}: {str(exc)[:80]}")
+    return text, input_tokens, output_tokens
+
+
+def call_openai_responses(model_name, prompt, role):
+    """Use Responses API for Codex models that are not Chat Completions routes."""
+    host = "api.openai.com"
+    if host not in ALLOWED_CLOUD:
+        fail(f"endpoint_not_allowed openai host={host}")
+    key = os.environ["OPENAI_API_KEY"]
+    body = {
+        "model": model_name,
+        "input": prompt,
+        "max_output_tokens": generation_max_tokens(role),
+        "store": False,
+    }
+    data = json.dumps(body).encode()
+    req = urllib.request.Request(
+        f"https://{host}/v1/responses", data=data,
+        headers={"Content-Type": "application/json", "Authorization": f"Bearer {key}"},
+    )
+    t0 = time.time()
+    try:
+        with urllib.request.urlopen(req, timeout=REQUEST_TIMEOUT) as resp:
+            payload = json.loads(resp.read().decode())
+    except Exception as exc:
+        fail(f"backend_unreachable {type(exc).__name__}: {str(exc)[:120]}")
+    latency_ms = int((time.time() - t0) * 1000)
+    text, input_tokens, output_tokens = parse_responses_payload(payload)
+    return {
+        "text": text,
+        "input_tokens": input_tokens,
+        "output_tokens": output_tokens,
+        "latency_ms": latency_ms,
+    }
+
+
 def openai_compatible_url():
     """Return a vetted OpenAI-compatible chat-completions endpoint.
 
@@ -412,7 +464,7 @@ def main():
     if backend == "ollama":
         result = call_ollama(model_name, prompt, args.role)
     elif backend == "openai":
-        result = call_openai(model_name, prompt, args.role)
+        result = call_openai_responses(model_name, prompt, args.role) if model_name.endswith("-codex") else call_openai(model_name, prompt, args.role)
     elif backend == "openai-compatible":
         result = call_openai_compatible(model_name, prompt, args.role)
     else:  # anthropic