diff --git a/README.md b/README.md
index b8f1028..acdc163 100644
--- a/README.md
+++ b/README.md
@@ -34,6 +34,15 @@ step guarded by `if: failure()`. On a red run you get an annotation like
 `python-test-failure (confidence 0.89) — guide: getpatchrail.com/fix/...` plus a
 job summary block.
 
+When no rule matches the log, the class stays `unknown` — PatchRail does not
+guess — but the annotation hands back the line the runner itself flagged for the
+failing step, so you still land on the error instead of on "no signal found":
+
+```
+unknown (confidence 0.15) — runner reported: "github-token" length must be less
+than or equal to 100 characters long — guide: getpatchrail.com/fix
+```
+
 ### Capturing the log correctly
 
 The capture step above is deliberate in two ways, and both matter:
diff --git a/action.yml b/action.yml
index 82ae815..8802605 100644
--- a/action.yml
+++ b/action.yml
@@ -58,10 +58,13 @@ runs:
           # `latest-patchrail` job in .github/workflows/test.yml runs this same
           # path against the newest release, so the range moves deliberately.
           #
-          # Floor is 0.5.0: every release below it reads a tool the job merely
-          # named — a `GRADLE_HOME=` line, a `Collecting mypy` from pip — as the
-          # cause of death, so it annotates the PR with a confident wrong answer.
-          python -m pip install --quiet "patchrail>=0.5.0,<0.6.0"
+          # Floor is 0.6.0. Below it, a release reads a tool the job merely named —
+          # a `GRADLE_HOME=` line, a `Collecting mypy` from pip, a tool named inside
+          # a filename — as the cause of death, and annotates the PR with a confident
+          # wrong answer. 0.6.0 also reports the runner's own `##[error]` line on an
+          # `unknown` verdict, which is the only thing that makes an unclassified run
+          # worth annotating at all; `scripts/annotate.py` surfaces it.
+          python -m pip install --quiet "patchrail>=0.6.0,<0.7.0"
         fi
 
     - name: Explain failure and annotate
diff --git a/scripts/annotate.py b/scripts/annotate.py
index dba82dd..d16ec0f 100644
--- a/scripts/annotate.py
+++ b/scripts/annotate.py
@@ -20,6 +20,12 @@
 # red. Refuse a schema we do not know instead of inventing a classification.
 RESULT_SCHEMA = "patchrail.ci_result.v1"
 
+# What one `unknown` verdict may drag into the annotation. patchrail already
+# de-duplicates and caps `runner_errors`, but the content is log text: a matrix
+# build emits one annotation per leg, and a single line can be a whole stack trace.
+RUNNER_ERROR_LIMIT = 3
+RUNNER_ERROR_MAX_CHARS = 200
+
 # Shown whenever there is no log to classify. Both halves matter: without
 # `2>&1` a tool that reports only on stderr leaves an empty log, and without
 # pipefail (`shell: bash`) a failing command piped into `tee` exits 0, so the
@@ -78,6 +84,49 @@ def guide_url(failure_class: str) -> str:
     return FIX_GUIDE_BASE
 
 
+def runner_errors(result: dict) -> list[str]:
+    """The lines the runner itself flagged, as patchrail 0.6.0 reports them.
+
+    Only present on an `unknown` result: when no rule matches, patchrail hands back
+    the runner's own annotation for the failing step (`##[error]…`), so the job can
+    say *where* it died even when PatchRail cannot say *why*. Without this, `unknown`
+    annotates a red run with `inspect CI log and run the failing job locally` — the
+    user learns nothing they would not have learned by never running the action.
+
+    Absent on a classified result (its `signals` already explain it) and on any
+    patchrail below 0.6.0, which the `patchrail-version` input still allows: an
+    older release simply has no such key and this returns nothing.
+    """
+    reported = result.get("runner_errors")
+    if not isinstance(reported, list):
+        return []
+    lines = []
+    for entry in reported[:RUNNER_ERROR_LIMIT]:
+        # Log text, so it arrives with whatever shape the failing build gave it.
+        line = " ".join(str(entry).split())
+        if not line:
+            continue
+        if len(line) > RUNNER_ERROR_MAX_CHARS:
+            line = line[: RUNNER_ERROR_MAX_CHARS - 1].rstrip() + "…"
+        lines.append(line)
+    return lines
+
+
+def annotation_safe(text: str) -> str:
+    """Escape log text for a workflow command, `%` first.
+
+    GitHub decodes `%0A` in a `::warning::` message back into a newline, and this
+    text is a line from a build any PR author can write. Left raw, a log containing
+    `%0A::error::<anything>` would close our annotation and forge a second one.
+    """
+    return text.replace("%", "%25").replace("\r", "%0D").replace("\n", "%0A")
+
+
+def summary_safe(text: str) -> str:
+    """Keep log text inside its code span in the job summary, instead of formatting it."""
+    return text.replace("`", "'")
+
+
 def write_kv(path_env: str, lines: list[str]) -> None:
     path = os.environ.get(path_env)
     if not path:
@@ -162,13 +211,15 @@ def main() -> int:
     subsystem = result.get("likely_subsystem") or "unknown"
     repro = result.get("reproduction_command") or ""
     strategy = result.get("minimal_repair_strategy") or ""
+    reported = runner_errors(result)
     url = guide_url(failure_class)
 
-    # GitHub annotation (shows up inline on the run).
-    print(
-        f"::warning title=PatchRail CI Triage::{failure_class} "
-        f"(confidence {confidence}) — guide: {url}"
-    )
+    # GitHub annotation (shows up inline on the run). One line, always: GitHub reads
+    # one annotation per line, so the runner's line is escaped, never appended raw.
+    headline = f"{failure_class} (confidence {confidence})"
+    if reported:
+        headline += f" — runner reported: {annotation_safe(reported[0])}"
+    print(f"::warning title=PatchRail CI Triage::{headline} — guide: {url}")
 
     # Job summary.
     summary = [
@@ -178,6 +229,11 @@ def main() -> int:
         f"- **Confidence:** `{confidence}`",
         f"- **Subsystem:** {subsystem}",
     ]
+    # On an `unknown` verdict this is the only line in the summary worth reading, so
+    # it goes above the generic "reproduce it locally" advice, not below it.
+    if reported:
+        summary.append("- **Errors the runner reported:**")
+        summary += [f"  - `{summary_safe(line)}`" for line in reported]
     if repro:
         summary.append(f"- **Reproduce:** `{repro}`")
     if strategy:
diff --git a/tests/test_runner_errors.py b/tests/test_runner_errors.py
new file mode 100644
index 0000000..16d96d7
--- /dev/null
+++ b/tests/test_runner_errors.py
@@ -0,0 +1,175 @@
+"""An `unknown` verdict must hand back the line the runner flagged, and nothing else.
+
+Before this, an unclassified run got the worst annotation the action can produce:
+`unknown (confidence 0.15)`, subsystem `unknown`, and the advice to "inspect CI log
+and run the failing job locally" — on a red run, from a tool the user installed
+precisely so they would not have to. patchrail 0.6.0 reports the runner's own
+`##[error]` line for the failing step in `runner_errors`, so the action can at least
+say *where* the job died when it cannot say *why*. It was already in the JSON the
+action reads; the action was dropping it on the floor.
+
+The lines are log text from a build any PR author can write, so they are untrusted:
+they reach a `::warning` workflow command, where GitHub decodes `%0A` back into a
+newline, and a job summary, which is Markdown. Escaped on the way out, both times.
+"""
+from __future__ import annotations
+
+import json
+import subprocess
+import sys
+from pathlib import Path
+
+import pytest
+
+ANNOTATE = Path(__file__).resolve().parent.parent / "scripts" / "annotate.py"
+GUIDE_INDEX = "https://getpatchrail.com/fix"
+
+# What patchrail 0.6.0 emits for a log no rule matches. The runner error is the real
+# one from psf/requests run 29295524780, the run that motivated the patchrail fix.
+UNKNOWN_RESULT = {
+    "schema_version": "patchrail.ci_result.v1",
+    "failure_class": "unknown",
+    "confidence": 0.15,
+    "likely_subsystem": "unknown",
+    "reproduction_command": "inspect CI log and run the failing job locally",
+    "minimal_repair_strategy": "Do not auto-repair until the failing subsystem is identified.",
+    "runner_errors": ['"github-token" length must be less than or equal to 100 characters long'],
+}
+
+# A log that classifies: its `signals` explain it, and patchrail reports no
+# `runner_errors` at all. Also what every patchrail below 0.6.0 emits.
+CLASSIFIED_RESULT = {
+    "schema_version": "patchrail.ci_result.v1",
+    "failure_class": "python_test_failure",
+    "confidence": 0.95,
+    "likely_subsystem": "Python tests",
+    "reproduction_command": "python -m pytest -q",
+    "minimal_repair_strategy": "Reproduce the failing test and patch the drift.",
+}
+
+
+@pytest.fixture()
+def annotate(tmp_path):
+    def run(result: dict) -> tuple[subprocess.CompletedProcess, str, str]:
+        result_path = tmp_path / "patchrail-ci-result.json"
+        result_path.write_text(json.dumps(result), encoding="utf-8")
+        output = tmp_path / "output"
+        summary = tmp_path / "summary"
+        output.touch()
+        summary.touch()
+        proc = subprocess.run(
+            [sys.executable, str(ANNOTATE), str(result_path)],
+            capture_output=True,
+            text=True,
+            env={
+                "PATH": "/usr/bin:/bin",
+                "GITHUB_OUTPUT": str(output),
+                "GITHUB_STEP_SUMMARY": str(summary),
+            },
+        )
+        return proc, output.read_text(), summary.read_text()
+
+    return run
+
+
+def with_errors(*errors: object) -> dict:
+    return {**UNKNOWN_RESULT, "runner_errors": list(errors)}
+
+
+def test_the_runner_line_reaches_the_annotation(annotate) -> None:
+    """The whole point: the red run now names the line it died on."""
+    proc, _, _ = annotate(UNKNOWN_RESULT)
+    assert proc.returncode == 0, proc.stderr
+    assert '"github-token" length must be less than or equal to 100' in proc.stdout
+
+
+def test_the_runner_line_reaches_the_job_summary(annotate) -> None:
+    _, _, summary = annotate(UNKNOWN_RESULT)
+    assert "**Errors the runner reported:**" in summary
+    assert '"github-token" length must be less than or equal to 100' in summary
+
+
+def test_the_runner_line_outranks_the_generic_advice(annotate) -> None:
+    """On `unknown` it is the only line worth reading, so it must not sit under
+    "inspect CI log and run the failing job locally"."""
+    _, _, summary = annotate(UNKNOWN_RESULT)
+    assert summary.index("Errors the runner reported") < summary.index("Reproduce")
+
+
+def test_an_unknown_verdict_is_still_unknown(annotate) -> None:
+    """An annotation says where the job died, not why. Reporting it must not be
+    allowed to look like a classification: patchrail does not know this log."""
+    proc, output, _ = annotate(UNKNOWN_RESULT)
+    assert "failure-class=unknown" in output
+    assert "confidence=0.15" in output
+    assert f"guide-url={GUIDE_INDEX}" in output.strip().splitlines()  # index, not a 404
+
+
+def test_a_classified_result_is_untouched(annotate) -> None:
+    """patchrail sends no `runner_errors` when a rule matched, and neither does any
+    patchrail below 0.6.0, which `patchrail-version` still lets a user pin."""
+    proc, output, summary = annotate(CLASSIFIED_RESULT)
+    assert proc.returncode == 0, proc.stderr
+    assert "runner reported" not in proc.stdout
+    assert "Errors the runner reported" not in summary
+    assert "python_test_failure (confidence 0.95)" in proc.stdout
+    assert f"guide-url={GUIDE_INDEX}/python-test-failure" in output
+
+
+def test_a_log_line_cannot_forge_a_second_annotation(annotate) -> None:
+    """GitHub decodes `%0A` in a workflow command back into a newline. Raw, this log
+    line would end our warning and open an `::error::` of the author's choosing."""
+    proc, _, _ = annotate(with_errors("boom%0A::error::forged by the log"))
+    commands = [ln for ln in proc.stdout.splitlines() if ln.startswith("::")]
+    assert len(commands) == 1
+    assert commands[0].startswith("::warning title=PatchRail CI Triage::")
+    # The `%` is escaped, so GitHub renders `%0A` as those two literal characters
+    # instead of decoding it into the newline that would start the forged command.
+    assert "%250A" in commands[0]
+    assert "%0A::error::" not in commands[0]
+
+
+def test_the_annotation_stays_on_one_line(annotate) -> None:
+    """GitHub reads one annotation per line; a wrapped message would be truncated."""
+    proc, _, _ = annotate(with_errors("first line\nsecond line\r\nthird"))
+    warnings = [ln for ln in proc.stdout.splitlines() if ln.startswith("::warning")]
+    assert len(warnings) == 1
+    assert "first line second line third" in warnings[0]
+
+
+def test_a_log_line_cannot_break_out_of_its_code_span(annotate) -> None:
+    """The job summary is Markdown, and this is text from someone else's build."""
+    _, _, summary = annotate(with_errors("`# not a heading, and no <img> either"))
+    reported = next(ln for ln in summary.splitlines() if "not a heading" in ln)
+    assert reported.startswith("  - `")
+    assert reported.endswith("`")
+    assert reported.count("`") == 2
+
+
+def test_a_matrix_build_cannot_flood_the_annotation(annotate) -> None:
+    """One failing leg per matrix entry, each with its own annotation."""
+    _, _, summary = annotate(with_errors(*[f"leg {i} died" for i in range(9)]))
+    assert summary.count("leg ") == 3
+
+
+def test_a_stack_trace_on_one_line_is_truncated(annotate) -> None:
+    _, _, summary = annotate(with_errors("x" * 5000))
+    reported = next(ln for ln in summary.splitlines() if "xxx" in ln)
+    assert len(reported) < 260
+    assert reported.endswith("…`")
+
+
+def test_junk_in_the_field_is_ignored_rather_than_annotated(annotate) -> None:
+    """`.get()` never raises, so a future patchrail that reshapes this field would
+    otherwise annotate a red run with a stringified dict."""
+    for junk in ["a bare string", {"error": "a dict"}, 17, None]:
+        proc, _, summary = annotate(with_errors() | {"runner_errors": junk})
+        assert proc.returncode == 0, proc.stderr
+        assert "runner reported" not in proc.stdout
+        assert "Errors the runner reported" not in summary
+
+
+def test_empty_and_blank_entries_never_become_a_bullet(annotate) -> None:
+    proc, _, summary = annotate(with_errors("", "   ", "\n"))
+    assert "runner reported" not in proc.stdout
+    assert "Errors the runner reported" not in summary
diff --git a/tests/test_schema_guard.py b/tests/test_schema_guard.py
index 5fcd607..cf9995a 100644
--- a/tests/test_schema_guard.py
+++ b/tests/test_schema_guard.py
@@ -26,7 +26,7 @@
 # test agree with any future edit to it, which is the drift it exists to catch.
 SUPPORTED_SCHEMA = "patchrail.ci_result.v1"
 
-# What patchrail emits today (verified against 0.3.1, 0.4.0 and 0.5.0).
+# What patchrail emits today (verified against 0.3.1, 0.4.0, 0.5.0 and 0.6.0).
 V1_RESULT = {
     "schema_version": SUPPORTED_SCHEMA,
     "failure_class": "python_test_failure",