diff --git a/runner/server.py b/runner/server.py
index e8a3571a..8c153278 100644
--- a/runner/server.py
+++ b/runner/server.py
@@ -5818,6 +5818,28 @@ def _relay_last_finish(env: dict) -> str:
     return str(_turn_relay_flags(env).get("last_finish") or "").replace("-", "_")
 
 
+def _relay_refusal_unsaid(env: dict, *said: str) -> str:
+    """The provider's refusal of this turn's last call on its route, as the relay read it, when the
+    reason the turn would otherwise give (`said`) does not carry it; "" otherwise.
+
+    A CLI reports a refusal only as far as its client could read the body. Qwen Code 0.26.3 reads
+    the reason only under an `error` key, and vLLM up to 0.10.0 puts `message` at the top level
+    (`{"object":"error","message":...}`; nested from 0.10.1), so a turn vLLM refused for a window
+    too small read "API Error: 400 status code (no body)", while vLLM had said "This model's maximum
+    context length is 28672 tokens. However, you requested 49351 tokens ..." (#612, measured through
+    the relay with 0.26.3 at a stub, and at a real vLLM v0.8.5.post1 in #578). The relay passed that
+    body on byte for byte and kept its words on the route (`last_refusal`, _refusal_words, which
+    reads both shapes); only the hermes path read them. A CLI whose reason already has the words
+    keeps its own sentence: the comparison is on the message's start, which survives a CLI's own
+    prefix ("[API Error: 400 ...]"). A refusal with no message beyond its status says nothing the
+    CLI's "400" does not."""
+    words = str(_turn_relay_flags(env).get("last_refusal") or "")
+    msg = words.split(": ", 1)[1].strip() if ": " in words else ""
+    if not msg or any(msg[:60] in s for s in said):
+        return ""
+    return f"The model provider refused the request ({words})"
+
+
 def _relay_served_model(env: dict) -> str:
     """What the relay saw the provider serve on this turn's route, "" when the turn did not ride
     the relay or nothing answered yet (see _turn_relay_flags for which route that is)."""
@@ -6134,6 +6156,10 @@ def _forward(self, body: bytes | None) -> None:
                     plain, body = body, leveled
                     headers["content-length"] = str(len(body))
         effort_back = False     # the level was taken out after a refusal; a success now means it was the cause
+        # A new call: an earlier call's refusal is not this one's. Kept, it was a turn's reason when
+        # this call then went unanswered (a dropped connection, a timeout) and the CLI failed of that
+        # (_relay_refusal_unsaid, #612). A refusal of this call sets it again below.
+        flags.pop("last_refusal", None)
         resp = None
         tried_slim = False
         dropped_once = False
@@ -6142,6 +6168,10 @@ def _forward(self, body: bytes | None) -> None:
                                          method=self.command, headers=headers)
             try:
                 resp = urllib.request.urlopen(req, timeout=HR_RELAY_UPSTREAM_TIMEOUT_S)
+                # Answered: this call's own refusal was repaired (its retry without stream_options,
+                # an unknown field, a level), and a turn that fails later fails for another reason,
+                # which the refusal must not be read as (_relay_refusal_unsaid, #612).
+                flags.pop("last_refusal", None)
                 if effort_back:
                     # the provider took the call without the level: not offered again for this model
                     flags[f"effort_refused:{effort_model}"] = True
@@ -14377,8 +14407,14 @@ def _settle(ev: dict) -> None:
         # providers' phrasings ("HTTP 503", "Error code: 429", "503 Service Unavailable") and
         # tightening it to exclude a bare number would risk missing a real refusal — the frames
         # are what does not belong in the search.
-        refusal = next((ln.strip() for ln in tail.splitlines()
-                        if _PROVIDER_REFUSAL.search(ln)), "")
+        # The provider's own words come before a line of the CLI's that merely names the status: the
+        # relay saw the refusal the CLI could not read ("API Error: 400 status code (no body)" from
+        # Qwen Code on vLLM 0.10.0's body, #612), and "429 status code (no body)" would otherwise be
+        # the line the search below picks.
+        # Compared with the reason the turn would give without them, not with everything the CLI
+        # printed: words in the stderr tail behind a result's "(no body)" are not what a person reads.
+        said = next((ln.strip() for ln in tail.splitlines() if _PROVIDER_REFUSAL.search(ln)), "")
+        refusal = _relay_refusal_unsaid(env, _failure_reason(said, str(ev_err), tail, rc)) or said
         if not refusal and _relay_last_finish(env) == "content_filter":
             # THE PROVIDER DECLINED, and the CLI's own account of a failure it could not see is not
             # the reason. A refused request comes back as an answer with no content and no
diff --git a/runner/tests/test_relay_refusal_reason.py b/runner/tests/test_relay_refusal_reason.py
new file mode 100644
index 00000000..95bfcfc7
--- /dev/null
+++ b/runner/tests/test_relay_refusal_reason.py
@@ -0,0 +1,169 @@
+"""A failed turn's reason is the provider's refusal as the relay read it, when the CLI could not read
+it (#612).
+
+Qwen Code 0.26.3 reads a refusal's reason only under an `error` key. vLLM up to 0.10.0 puts `message`
+at the top level, so a turn vLLM refused for a window too small read "API Error: 400 status code (no
+body)" while the relay held vLLM's sentence on the route. Measured with the real CLI at a stub
+through the relay, and on a real vLLM v0.8.5.post1 in #578.
+
+Each test runs a turn through _run_turn_bg, the generic path whose reason the gateway shows ("The
+turn failed: <reason>"), with a fake CLI that makes its calls through the route _build_qwen
+registers, prints the stream-json events Qwen Code 0.26.3 printed on such a refusal (captured
+2026-10-29), and exits 1 as it did.
+"""
+import json
+import os
+import sys
+import threading
+import time
+from http.server import BaseHTTPRequestHandler, ThreadingHTTPServer
+
+sys.path.insert(0, os.path.join(os.path.dirname(__file__), ".."))
+from server import Auth, _build_qwen, _claude_passthrough, _relay_refusal_unsaid, _run_turn_bg, _turns  # noqa: E402
+
+VLLM_SAID = ("This model's maximum context length is 8192 tokens. However, you requested 49350 tokens "
+             "(16582 in the messages, 32768 in the completion). Please reduce the length of the messages "
+             "or completion.")
+# vLLM <=0.10.0's ErrorResponse, as the #578 capture proxy recorded it, and OpenAI's shape (vLLM >=0.10.1)
+FLAT = json.dumps({"object": "error", "message": VLLM_SAID, "type": "BadRequestError", "param": None, "code": 400}).encode()
+NESTED = json.dumps({"error": {"message": VLLM_SAID, "type": "BadRequestError", "param": None, "code": 400}}).encode()
+
+# The fake CLI: each body in FAKE_CALLS goes to the route as a chat-completions call, then the events
+# Qwen Code printed for a failed run, with FAKE_SAYS as its message, and exit 1.
+_CLI = r'''
+import json, os, sys, urllib.error, urllib.request
+for body in json.loads(os.environ["FAKE_CALLS"]):
+    req = urllib.request.Request(os.environ["OPENAI_BASE_URL"] + "/chat/completions", method="POST",
+                                 data=json.dumps(body).encode(),
+                                 headers={"authorization": "Bearer " + os.environ["OPENAI_API_KEY"],
+                                          "content-type": "application/json"})
+    try:
+        urllib.request.urlopen(req, timeout=10).read()
+    except urllib.error.HTTPError as e:
+        e.read()
+if os.environ.get("FAKE_STDERR"):
+    print(os.environ["FAKE_STDERR"], file=sys.stderr, flush=True)
+says, sid = os.environ["FAKE_SAYS"], "9c87656f-c69b-48cb-8b6d-b0eecd6b670f"
+for ev in ({"type": "system", "subtype": "init", "session_id": sid, "model": "qwen3-4b"},
+           {"type": "assistant", "session_id": sid, "parent_tool_use_id": None,
+            "message": {"type": "message", "role": "assistant", "model": "qwen3-4b",
+                        "content": [{"type": "text", "text": says}], "stop_reason": None}},
+           {"type": "result", "subtype": "error_during_execution", "session_id": sid, "is_error": True,
+            "num_turns": 1, "usage": {"input_tokens": 0, "output_tokens": 0}, "error": {"message": says}}):
+    print(json.dumps(ev), flush=True)
+sys.exit(1)
+'''
+
+
+def _call(stream_options: bool) -> dict:
+    body = {"model": "qwen3-4b", "stream": True, "messages": [{"role": "user", "content": "hi"}]}
+    if stream_options:
+        body["stream_options"] = {"include_usage": True}
+    return body
+
+
+class _Provider(BaseHTTPRequestHandler):
+    """Refuses with `refusal` every call it is told to refuse; answers a one-word stream otherwise."""
+    refusal = FLAT
+    refuse = staticmethod(lambda body: True)
+
+    def do_POST(self):
+        body = json.loads(self.rfile.read(int(self.headers.get("content-length") or 0)) or b"{}")
+        if "drop me" in json.dumps(body):
+            self.close_connection = True       # the connection closes with no answer at all
+            return
+        if type(self).refuse(body):
+            out, code, ctype = type(self).refusal, 400, "application/json"
+        else:
+            chunk = {"id": "c1", "object": "chat.completion.chunk", "model": "qwen3-4b",
+                     "choices": [{"index": 0, "delta": {"content": "DONE"}, "finish_reason": "stop"}]}
+            out, code, ctype = b"data: " + json.dumps(chunk).encode() + b"\n\ndata: [DONE]\n\n", 200, "text/event-stream"
+        self.send_response(code)
+        self.send_header("content-type", ctype)
+        self.send_header("content-length", str(len(out)))
+        self.end_headers()
+        self.wfile.write(out)
+
+    def log_message(self, *a):
+        pass
+
+
+def _turn(tmp_path, refusal: bytes, refuse, calls: list[dict], says: str, stderr: str = "") -> dict:
+    handler = type("Provider", (_Provider,), {"refusal": refusal, "refuse": staticmethod(refuse)})
+    up = ThreadingHTTPServer(("127.0.0.1", 0), handler)
+    threading.Thread(target=up.serve_forever, daemon=True).start()
+    cli = tmp_path / "fake_qwen.py"
+    cli.write_text(_CLI)
+    env = {**os.environ, "FAKE_CALLS": json.dumps(calls), "FAKE_SAYS": says, "FAKE_STDERR": stderr}
+    _build_qwen("openai-api", Auth(api_key="sk-local", base_url=f"http://127.0.0.1:{up.server_address[1]}/v1"),
+                "qwen3-4b", "hi", str(tmp_path), env)
+    tid = f"t-{time.time()}"
+    _turns[tid] = {"status": "running", "events": [], "result": "", "done": False,
+                   "started": time.time(), "backend": "qwen", "model": "qwen3-4b"}
+    try:
+        _run_turn_bg(tid, [sys.executable, str(cli)], env, str(tmp_path), _claude_passthrough, "qwen3-4b",
+                     timeout_seconds=30)
+        return _turns[tid]
+    finally:
+        _turns.pop(tid, None)
+        up.shutdown()
+
+
+def test_a_refusal_the_cli_could_not_read_is_the_turns_reason(tmp_path):
+    """The bug: the reason was Qwen Code's "API Error: 400 status code (no body)"."""
+    rec = _turn(tmp_path, FLAT, lambda b: True, [_call(True)], "[API Error: 400 status code (no body)]")
+    assert rec["status"] == "failed"
+    assert rec["error"] == f"The model provider refused the request (400: {VLLM_SAID})"
+
+
+def test_a_cli_that_read_the_refusal_keeps_its_own_sentence(tmp_path):
+    """OpenAI's shape (vLLM >=0.10.1): Qwen Code says the reason itself, and its sentence stands
+    as it did before #612."""
+    says = f"[API Error: 400 {VLLM_SAID}]"
+    rec = _turn(tmp_path, NESTED, lambda b: True, [_call(True)], says)
+    assert rec["status"] == "failed"
+    assert VLLM_SAID in rec["error"] and "The model provider refused" not in rec["error"]
+
+
+def test_a_refusal_the_relay_repaired_is_not_the_reason_for_a_later_failure(tmp_path):
+    """The provider refuses the call with stream_options and answers it without (the relay's own
+    one-time retry repairs it); the CLI then fails for a reason of its own. The repaired refusal was
+    left on the route before #612, and is not what this turn failed of."""
+    says = "Error: the tool registry failed to load"
+    rec = _turn(tmp_path, FLAT, lambda b: "stream_options" in b, [_call(True)], says)
+    assert rec["status"] == "failed"
+    assert "Please reduce the length" not in rec["error"]
+    assert "tool registry failed to load" in rec["error"]
+
+
+def test_a_later_refusal_after_an_answer_is_still_the_reason(tmp_path):
+    """Clearing on an answer forgets only what came before it: the turn's last call refused is
+    still read, after an answered one on the same route."""
+    calls = [_call(False), {**_call(False), "messages": [{"role": "user", "content": "too long"}]}]
+    rec = _turn(tmp_path, FLAT, lambda b: "too long" in json.dumps(b), calls,
+                "[API Error: 400 status code (no body)]")
+    assert rec["error"] == f"The model provider refused the request (400: {VLLM_SAID})"
+
+
+def test_a_refusal_before_a_call_that_went_unanswered_is_not_the_reason(tmp_path):
+    """The CLI got past a refused call and its next call was never answered (the connection closed),
+    which is what it failed of. The earlier refusal is not this turn's reason: each call starts the
+    route's record afresh."""
+    calls = [{**_call(False), "messages": [{"role": "user", "content": "too long"}]},
+             {**_call(False), "messages": [{"role": "user", "content": "drop me"}]}]
+    rec = _turn(tmp_path, FLAT, lambda b: "too long" in json.dumps(b), calls, "Connection error.")
+    assert rec["status"] == "failed"
+    assert "Please reduce the length" not in rec["error"] and "Connection error." in rec["error"]
+
+
+def test_words_only_in_the_stderr_tail_are_not_what_the_turn_says(tmp_path):
+    """The CLI printed the provider's sentence to stderr and its result said "(no body)": the reason a
+    person reads is the result's, so the provider's words are still unsaid and lead it."""
+    rec = _turn(tmp_path, FLAT, lambda b: True, [_call(True)], "[API Error: 400 status code (no body)]",
+                stderr=f"debug: upstream said {VLLM_SAID}")
+    assert rec["error"] == f"The model provider refused the request (400: {VLLM_SAID})"
+
+
+def test_no_route_or_no_words_leaves_the_reason_to_the_cli():
+    assert _relay_refusal_unsaid({}, "anything") == ""                    # the turn did not ride the relay
+    assert _relay_refusal_unsaid({"X": "not-a-relay-token"}, "") == ""