fix: make h2 patch repair failures diagnosable

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

@@ -67,6 +67,37 @@ def generation_max_tokens(role):
     return min(max(configured, 64), 8192)
 
 
+def decode_provider_json(response, provider):
+    """Decode a provider response without mislabelling an empty 2xx body.
+
+    Gateways occasionally terminate an upstream stream after accepting a
+    request and return an empty (or HTML) 2xx response.  Treating that as a
+    generic JSONDecodeError makes it look like the network failed and obscures
+    the provider that needs attention.  The diagnostic deliberately contains
+    only response shape metadata: never response content or credentials.
+    """
+    raw = response.read()
+    content_type = ""
+    headers = getattr(response, "headers", None)
+    if headers is not None:
+        try:
+            content_type = str(headers.get("Content-Type", "")).split(";", 1)[0].lower()
+        except (AttributeError, TypeError):
+            content_type = ""
+    if not raw:
+        fail(f"provider_response_empty {provider} content_type={content_type or 'unknown'}")
+    try:
+        payload = json.loads(raw.decode("utf-8"))
+    except (UnicodeDecodeError, json.JSONDecodeError) as exc:
+        fail(
+            f"provider_response_invalid {provider} {type(exc).__name__} "
+            f"bytes={len(raw)} content_type={content_type or 'unknown'}"
+        )
+    if not isinstance(payload, dict):
+        fail(f"provider_response_invalid {provider} payload_type={type(payload).__name__}")
+    return payload
+
+
 def enforce_call_budget():
     """Fail closed when a run exceeds CASAN_MODEL_MAX_CALLS. The count is tracked in
     CASAN_MODEL_CALL_COUNTER_FILE so it spans the many per-step model-call processes
@@ -192,7 +223,7 @@ def call_ollama(model_name, prompt, role):
     t0 = time.time()
     try:
         with urllib.request.urlopen(req, timeout=REQUEST_TIMEOUT) as resp:
-            payload = json.loads(resp.read().decode())
+            payload = decode_provider_json(resp, "ollama")
     except Exception as exc:  # backend/model failure -> honest non-zero, no fake success
         fail(f"backend_unreachable {type(exc).__name__}: {str(exc)[:120]}")
     latency_ms = int((time.time() - t0) * 1000)
@@ -226,7 +257,7 @@ def call_openai(model_name, prompt, role):
     t0 = time.time()
     try:
         with urllib.request.urlopen(req, timeout=REQUEST_TIMEOUT) as resp:
-            payload = json.loads(resp.read().decode())
+            payload = decode_provider_json(resp, "openai-chat")
     except Exception as exc:  # honest non-zero, no fake success
         fail(f"backend_unreachable {type(exc).__name__}: {str(exc)[:120]}")
     latency_ms = int((time.time() - t0) * 1000)
@@ -278,7 +309,7 @@ def call_openai_responses(model_name, prompt, role):
     t0 = time.time()
     try:
         with urllib.request.urlopen(req, timeout=REQUEST_TIMEOUT) as resp:
-            payload = json.loads(resp.read().decode())
+            payload = decode_provider_json(resp, "openai-responses")
     except Exception as exc:
         fail(f"backend_unreachable {type(exc).__name__}: {str(exc)[:120]}")
     latency_ms = int((time.time() - t0) * 1000)
@@ -345,7 +376,7 @@ def call_openai_compatible(model_name, prompt, role):
     t0 = time.time()
     try:
         with urllib.request.urlopen(req, timeout=REQUEST_TIMEOUT) as resp:
-            payload = json.loads(resp.read().decode())
+            payload = decode_provider_json(resp, "openai-compatible")
     except Exception as exc:
         fail(f"backend_unreachable {type(exc).__name__}: {str(exc)[:120]}")
     latency_ms = int((time.time() - t0) * 1000)
@@ -413,7 +444,7 @@ def call_anthropic(model_name, prompt, role):
     t0 = time.time()
     try:
         with urllib.request.urlopen(req, timeout=REQUEST_TIMEOUT) as resp:
-            payload = json.loads(resp.read().decode())
+            payload = decode_provider_json(resp, "anthropic")
     except Exception as exc:  # honest non-zero, no fake success
         fail(f"backend_unreachable {type(exc).__name__}: {str(exc)[:120]}")
     latency_ms = int((time.time() - t0) * 1000)