diff --git a/src/leanci/llm.py b/src/leanci/llm.py
index 72588b1..0e14aec 100644
--- a/src/leanci/llm.py
+++ b/src/leanci/llm.py
@@ -152,7 +152,7 @@ def _post_chat(self, body: dict[str, Any]) -> dict[str, Any]:
                 return _load_json_object(raw)
             except HTTPError as exc:
                 last_error = exc
-                if exc.code not in {429, 500, 502, 503, 504} or attempt >= attempts - 1:
+                if not _is_retriable_http(exc) or attempt >= attempts - 1:
                     detail = _http_error_detail(exc)
                     raise LLMError(
                         f"LLM HTTP {exc.code} from {url}: {exc.reason}{detail}"
@@ -167,13 +167,49 @@ def _post_chat(self, body: dict[str, Any]) -> dict[str, Any]:
         raise LLMError(f"LLM request to {url} failed: {last_error}")
 
 
+def _is_retriable_http(exc: HTTPError) -> bool:
+    """Transient upstream failures worth retrying (incl. Groq tool_use_failed)."""
+    if exc.code in {429, 500, 502, 503, 504}:
+        return True
+    if exc.code == 400:
+        # Groq sometimes returns 400 when the model emits malformed tool XML.
+        return "tool_use_failed" in _peek_http_body(exc)
+    return False
+
+
 def _retry_delay_s(exc: HTTPError, attempt: int) -> float:
     """Backoff for retriable HTTP errors; honor Retry-After / body hints on 429."""
     if exc.code == 429:
         hinted = _retry_after_seconds(exc)
         floor = _RATE_LIMIT_BACKOFF_S[min(attempt, len(_RATE_LIMIT_BACKOFF_S) - 1)]
         return max(hinted or 0.0, floor)
-    return float(_RETRY_BACKOFF_S[attempt])
+    return float(_RETRY_BACKOFF_S[min(attempt, len(_RETRY_BACKOFF_S) - 1)])
+
+
+def _peek_http_body(exc: HTTPError) -> str:
+    """Read and stash the HTTP error body once for retry checks + error text."""
+    raw = getattr(exc, "_leanci_body", None)
+    if raw is None:
+        try:
+            raw = exc.read()
+        except Exception:
+            return ""
+        # urllib may hand back a file-like; normalize to bytes.
+        if hasattr(raw, "read") and not isinstance(raw, (bytes, bytearray)):
+            try:
+                raw = raw.read()
+            except Exception:
+                return ""
+        if isinstance(raw, str):
+            raw = raw.encode("utf-8", errors="replace")
+        if not isinstance(raw, (bytes, bytearray)):
+            return ""
+        if raw:
+            setattr(exc, "_leanci_body", bytes(raw))
+            raw = bytes(raw)
+    if not raw:
+        return ""
+    return raw.decode("utf-8", errors="replace")
 
 
 def _retry_after_seconds(exc: HTTPError) -> float | None:
@@ -186,16 +222,7 @@ def _retry_after_seconds(exc: HTTPError) -> float | None:
             except ValueError:
                 pass
     # Gemini often embeds "Please retry in 48.55s" in the JSON body.
-    try:
-        body = exc.read()
-    except Exception:
-        return None
-    if not body:
-        return None
-    text = body.decode("utf-8", errors="replace")
-    # Stash for final error formatting if this was the last attempt — body already
-    # consumed, so attach a copy for _http_error_detail via a private attr.
-    setattr(exc, "_leanci_body", body)
+    text = _peek_http_body(exc)
     match = re.search(r"retry in\s+([0-9]+(?:\.[0-9]+)?)\s*s", text, flags=re.I)
     if match:
         return float(match.group(1))
@@ -204,15 +231,9 @@ def _retry_after_seconds(exc: HTTPError) -> float | None:
 
 def _http_error_detail(exc: HTTPError) -> str:
     """Best-effort provider error body for CI/debug (truncate to keep logs readable)."""
-    raw = getattr(exc, "_leanci_body", None)
-    if raw is None:
-        try:
-            raw = exc.read()
-        except Exception:
-            return ""
-    if not raw:
+    text = _peek_http_body(exc).strip()
+    if not text:
         return ""
-    text = raw.decode("utf-8", errors="replace").strip()
     if len(text) > 800:
         text = text[:800] + "…"
     return f" — {text}"
diff --git a/tests/test_llm.py b/tests/test_llm.py
index 64100b7..eb42107 100644
--- a/tests/test_llm.py
+++ b/tests/test_llm.py
@@ -248,6 +248,30 @@ def test_chat_retries_429_honors_retry_after_header() -> None:
     sleep.assert_called_once_with(90.0)
 
 
+def test_chat_retries_groq_tool_use_failed_400() -> None:
+    import io
+
+    client = LLMClient(api_key="sk", model="llama-3.3-70b-versatile")
+    body = b'{"error":{"code":"tool_use_failed","message":"Failed to call a function"}}'
+    err = HTTPError(
+        url="http://x",
+        code=400,
+        msg="Bad Request",
+        hdrs=None,  # type: ignore[arg-type]
+        fp=io.BytesIO(body),
+    )
+    ok = _http_response(_assistant_payload(content="ok"))
+
+    with (
+        patch("leanci.llm.urlopen", side_effect=[err, ok]),
+        patch("leanci.llm.time.sleep") as sleep,
+    ):
+        result = client.chat([{"role": "user", "content": "hi"}])
+
+    assert result.content == "ok"
+    sleep.assert_called_once_with(2)
+
+
 def test_chat_retries_twice_on_5xx_then_raises() -> None:
     client = LLMClient(api_key="sk", model="gpt-4.1-mini")
     err = HTTPError(