diff --git a/quantprobe/auto.py b/quantprobe/auto.py
index 91d0d2a..dd14d13 100644
--- a/quantprobe/auto.py
+++ b/quantprobe/auto.py
@@ -550,6 +550,11 @@ def run(a):
             dry_run=False,
             imatrix=None if getattr(a, "no_imatrix", False) else "auto",
             imatrix_chunks=100,
+            # `auto --yes` is answered BEFORE the source download; probe asks its
+            # time-commitment question after it. Omitting this re-asked a consenting user
+            # (and aborted with "re-run with --yes", which is what they did). Forwarded as
+            # given - without the flag the probe still confirms.
+            yes=bool(getattr(a, "yes", False)),
         )
         probemod.run(pa)
         print("\n[quantprobe auto --custom] your personalized model:")
diff --git a/tests/smoke.py b/tests/smoke.py
index fce0397..b296669 100644
--- a/tests/smoke.py
+++ b/tests/smoke.py
@@ -1229,6 +1229,23 @@ def t_auto_custom_dry():
         return
     assert "source:" in out and "fragile band" in out and "nothing downloaded" in out, out[-300:]
 
+def t_auto_custom_forwards_the_confirmation_it_was_given():
+    """`auto --custom --yes` must reach probe.run with the consent, and without it must not.
+
+    Lives in tests/test_auto_confirmation.py as unittest cases (hermetic: HF listing, download,
+    eval fetch and the probe itself are all replaced); this wrapper is what keeps it in the smoke
+    suite, which is the gate CONTRIBUTING names."""
+    import unittest
+    import test_auto_confirmation as mod
+
+    r = unittest.TextTestRunner(stream=io.StringIO(), verbosity=0).run(
+        unittest.defaultTestLoader.loadTestsFromModule(mod)
+    )
+    assert r.testsRun >= 4, f"only {r.testsRun} confirmation-forwarding cases ran"
+    assert not (r.failures or r.errors), "\n".join(
+        f"{t}: {why.strip().splitlines()[-1]}" for t, why in (r.failures + r.errors)
+    )
+
 def t_quantize_missing_file_graceful():
     # quantize on a missing GGUF must give a CLEAN error, never a traceback
     rc, out = cli("quantize", "--gguf", "nope.gguf", "--out", "o.gguf", "--protect-late", "12", "--dry")
diff --git a/tests/test_auto_confirmation.py b/tests/test_auto_confirmation.py
new file mode 100644
index 0000000..dc15985
--- /dev/null
+++ b/tests/test_auto_confirmation.py
@@ -0,0 +1,198 @@
+"""`auto --custom --yes` must carry the consent it was given into the probe.
+
+`auto` accepts `--yes` ("skip time-commitment confirmations") and `probe.run` is the step that
+asks for one: above ~2h estimated it prompts, and with no tty it exits. `auto._custom` builds the
+probe's argument Namespace by hand, so a flag it forgets to copy is simply not there - and the
+user who answered the question up front is asked it again AFTER the multi-GB source download,
+under an error that tells them to "re-run with --yes", which is what they did.
+
+Hermetic: the HF listing, the download, the eval-corpus fetch and the probe itself are the four
+external boundaries, and all four are replaced. No GGUF is read, nothing is fetched, no llama.cpp
+runs. What is exercised is the real `cli.main` -> `auto.run` custom pipeline, including the real
+optimizer and planner, on a preset machine.
+"""
+
+from __future__ import annotations
+
+import io
+import os
+import shutil
+import sys
+import tempfile
+import unittest
+from contextlib import redirect_stdout
+from unittest import mock
+
+sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
+
+from quantprobe import auto as automod
+
+PRESET = "qwen3-30b"
+TOTAL_B = automod.MODEL_REPOS[PRESET][1]
+REPO = automod.MODEL_REPOS[PRESET][0]
+SOURCE = "Qwen3-30B-A3B-Q8_0.gguf"
+
+
+def _bytes_at(bits):
+    """Size a synthetic listing entry so `auto`'s size->bits arithmetic lands on `bits`."""
+    return int(bits * TOTAL_B * 1e9 / 8)
+
+
+# One high-precision source (what --custom requantizes) plus two ordinary quants, so the
+# pre-custom ranking loop has something real to score.
+LISTING = [
+    ("Qwen3-30B-A3B-Q2_K.gguf", _bytes_at(2.7)),
+    ("Qwen3-30B-A3B-Q4_K_M.gguf", _bytes_at(4.6)),
+    (SOURCE, _bytes_at(8.2)),
+]
+
+
+class _Recorder:
+    """The four external boundaries, recorded in call order."""
+
+    def __init__(self, dest):
+        self.dest = dest
+        self.fetched = []
+        self.eval_dirs = []
+        self.probe_args = []
+        self.order = []
+
+    def fetch(self, repo, dest, fname, tok, tries=100, force=False):
+        self.fetched.append((repo, dest, fname))
+        self.order.append("fetch")
+        return True
+
+    def ensure_eval(self, dest_dir):
+        self.eval_dirs.append(dest_dir)
+        self.order.append("ensure_eval")
+        path = os.path.join(dest_dir, "wiki.test.raw")
+        with open(path, "w") as f:
+            f.write("synthetic held-out text\n")
+        return path
+
+    def probe_run(self, a):
+        self.probe_args.append(a)
+        self.order.append("probe.run")
+
+
+def run_auto(*extra):
+    """Drive the real CLI so the Namespace `auto` reads is the one argparse actually builds."""
+    from quantprobe import cli
+    from quantprobe import fetch as fetchmod
+    from quantprobe import probe as probemod
+
+    dest = tempfile.mkdtemp(prefix="qp-auto-confirm-")
+    rec = _Recorder(dest)
+    argv = [
+        "quantprobe",
+        "auto",
+        PRESET,
+        "--machine",
+        "2016-xmp",
+        "--custom",
+        "--dir",
+        dest,
+        *extra,
+    ]
+    buf = io.StringIO()
+    with (
+        mock.patch.object(automod, "list_ggufs", lambda _repo: list(LISTING)),
+        mock.patch.object(automod, "ensure_eval", rec.ensure_eval),
+        mock.patch.object(fetchmod, "fetch", rec.fetch),
+        mock.patch.object(fetchmod, "token", lambda: None),
+        mock.patch.object(probemod, "run", rec.probe_run),
+        mock.patch.object(sys, "argv", argv),
+        redirect_stdout(buf),
+    ):
+        cli.main()
+    rec.out = buf.getvalue()
+    return rec
+
+
+class AutoCustomConfirmationTest(unittest.TestCase):
+    def run_auto(self, *extra):
+        rec = run_auto(*extra)
+        self.addCleanup(shutil.rmtree, rec.dest, ignore_errors=True)
+        return rec
+
+    def assert_took_custom_path(self, rec):
+        # A machine that declined the surgery would skip the custom branch entirely and leave
+        # every assertion below vacuously true, so every test checks the branch actually ran.
+        self.assertIn(
+            "[quantprobe auto --custom] source: " + SOURCE,
+            rec.out,
+            "the custom branch did not run - the machine gate declined the surgery",
+        )
+
+    def test_yes_reaches_probe_run(self):
+        rec = self.run_auto("--yes")
+        self.assert_took_custom_path(rec)
+        self.assertEqual(len(rec.probe_args), 1, "probe.run must be called exactly once")
+        pa = rec.probe_args[0]
+        self.assertIs(
+            getattr(pa, "yes", None),
+            True,
+            "auto --custom --yes must forward the consent to probe.run, or the probe re-asks "
+            "for it after the source download and aborts telling the user to pass --yes",
+        )
+
+    def test_without_yes_the_probe_still_confirms(self):
+        rec = self.run_auto()
+        self.assert_took_custom_path(rec)
+        pa = rec.probe_args[0]
+        self.assertFalse(
+            getattr(pa, "yes", False),
+            "no --yes must leave probe.run's confirmation in place - this fix forwards a flag, "
+            "it does not grant approval nobody asked for",
+        )
+
+    def test_the_flag_changes_nothing_but_the_flag(self):
+        """The download steps, the probe's build parameters and the ordering stay identical."""
+        plain, consented = self.run_auto(), self.run_auto("--yes")
+        for rec in (plain, consented):
+            self.assert_took_custom_path(rec)
+
+        self.assertEqual(
+            [(r, f) for r, _, f in plain.fetched],
+            [(REPO, SOURCE)],
+            "the source download must still be the one fetch --custom does",
+        )
+        self.assertEqual(
+            [(r, f) for r, _, f in plain.fetched],
+            [(r, f) for r, _, f in consented.fetched],
+            "--yes must not change what is downloaded or from where",
+        )
+        self.assertEqual(
+            plain.order,
+            ["fetch", "ensure_eval", "probe.run"],
+            "--custom must download the source and the eval corpus before probing",
+        )
+        self.assertEqual(plain.order, consented.order, "--yes must not reorder the pipeline")
+        self.assertEqual(
+            len(plain.eval_dirs), 1, "the eval corpus is resolved once, in the download directory"
+        )
+
+        a_plain, a_yes = vars(plain.probe_args[0]).copy(), vars(consented.probe_args[0]).copy()
+        # the two runs use separate temp dirs, so path-valued fields differ by construction
+        for k in ("gguf", "workdir", "out", "eval"):
+            a_plain.pop(k, None)
+            a_yes.pop(k, None)
+        differing = {k for k in set(a_plain) | set(a_yes) if a_plain.get(k) != a_yes.get(k)}
+        self.assertEqual(
+            differing,
+            {"yes"},
+            "--yes must alter the confirmation flag and nothing else in the probe's arguments",
+        )
+        self.assertTrue(a_yes["apply"], "the custom build still applies the depth-aware recipe")
+        self.assertFalse(a_yes["dry_run"], "a real --custom run is not a dry run")
+
+    def test_dry_still_executes_nothing(self):
+        rec = self.run_auto("--yes", "--dry")
+        self.assert_took_custom_path(rec)
+        self.assertIn("(--dry: nothing downloaded)", rec.out)
+        self.assertEqual(rec.order, [], "--dry must not fetch, resolve an eval corpus, or probe")
+        self.assertEqual(os.listdir(rec.dest), [], "--dry must leave the download directory empty")
+
+
+if __name__ == "__main__":
+    unittest.main()