fix: keep the reviewer's objection, not its praise

src/agent_harness/executor.py

@@ -1350,6 +1350,34 @@ def is_disk_exhaustion(detail: str) -> bool:
 """
 
 
+def review_reason(verdict_text: str, limit: int = 1200) -> str:
+    """The reviewer's verdict, kept from the end rather than the beginning.
+
+    `last_error` is what a retry is told and what an operator reads, and it
+    used to be the first 500 characters of the verdict. The rubric asks for
+    "what I verified" first and "why" last, so those 500 characters were the
+    list of things the reviewer was *happy* with, and the objection — the only
+    part anyone can act on — was cut off.
+
+    Measured: two items were rejected twice for the same fault after being
+    told, both times, only the preamble praising the parts that were fine.
+
+    So the tail wins when something has to go. The head is a summary of the
+    diff, which the reader already has; the tail is the decision.
+    """
+    text = verdict_text.strip()
+    if len(text) <= limit:
+        return text
+    # Prefer starting at a section boundary, so a retry is not handed half a
+    # sentence about something it cannot see the start of.
+    tail = text[-limit:]
+    for marker in ("\n3. ", "\n2. ", "\n\n"):
+        cut = tail.find(marker)
+        if 0 <= cut < limit // 2:
+            return "…" + tail[cut:].strip()
+    return "…" + tail.strip()
+
+
 def review_context(repo: Path, diff: str, budget: int) -> str:
     """The touched files as they now stand, for the reviewer.
 
@@ -2413,7 +2441,7 @@ def _review_stage(
         if outcome.pr_url and self.github is not None:
             self._record_verdict(record, outcome.pr_url, verdict, verdict_text)
         if verdict != APPROVED:
-            outcome.reason = f"review rejected: {verdict_text.strip()[:500]}"
+            outcome.reason = f"review rejected: {review_reason(verdict_text)}"
             outcome.stop = Stop(REFUSED, REVIEW_REJECTED, detail=outcome.reason)
             return outcome
 

src/agent_harness/session_executor.py

@@ -139,6 +139,7 @@
 #: items were rejected for exactly the artefacts those fixes had already
 #: retired — in the other file (#167).
 from .executor import REVIEW_PROMPT as REVIEW_PROMPT  # noqa: E402
+from .executor import review_reason  # noqa: E402
 
 
 @dataclass
@@ -647,7 +648,7 @@ def _execute(self, record: WorkRecord) -> Outcome:
                 self._record_verdict(record, outcome.pr_url, verdict, verdict_text)
 
             if verdict != APPROVED:
-                outcome.reason = f"review rejected: {verdict_text.strip()[:500]}"
+                outcome.reason = f"review rejected: {review_reason(verdict_text)}"
                 outcome.stop = Stop(REFUSED, REVIEW_REJECTED, detail=outcome.reason)
                 return outcome
 

tests/test_reviewer.py

@@ -240,3 +240,37 @@ def call(self, _role: str, messages: Any, **_: Any) -> Any:
 
     assert "hello harness" in seen["prompt"], "the reviewer was not shown the committed work"
     assert "-hello world" in seen["prompt"], "nor what it replaced"
+
+
+def test_the_kept_reason_is_the_objection_not_the_praise() -> None:
+    """`last_error` is what a retry is told, and it kept the wrong end.
+
+    The rubric asks for "what I verified" first and "why" last. Keeping the
+    first 500 characters therefore kept the list of things the reviewer was
+    happy with and cut off the objection — the only part anyone can act on.
+
+    Measured: two items were rejected twice for the same fault, each retry
+    having been told only the preamble praising the parts that were fine.
+    """
+    from agent_harness.executor import review_reason
+
+    verdict = (
+        "REJECTED\n\n"
+        "1. **What I verified** — " + "the call site is right, " * 60 + "\n\n"
+        "2. **What I could not verify** — " + "nothing much, " * 20 + "\n\n"
+        "3. **Why**\nIt widens the repository trait's return type."
+    )
+
+    kept = review_reason(verdict)
+
+    assert "widens the repository trait" in kept, "the objection must survive"
+    assert len(kept) <= 1300
+    # And the old behaviour must not come back: the head is a summary of a
+    # diff the reader already has.
+    assert not kept.startswith("REJECTED\n\n1. **What I verified**")
+
+
+def test_a_short_verdict_is_kept_whole() -> None:
+    from agent_harness.executor import review_reason
+
+    assert review_reason("REJECTED\nToo narrow.") == "REJECTED\nToo narrow."