From ade5ea1210fcf0fa51222290f47f6f0819ef16d9 Mon Sep 17 00:00:00 2001
From: wuyang <5700876+banisherwy@user.noreply.gitee.com>
Date: Sun, 26 Jul 2026 17:08:27 +0800
Subject: [PATCH] fix: make invalid source retries actionable

---
 agent_platform/runtime/loop.py | 27 +++++++++++++++++++-
 tests/test_agent_loop.py       | 46 ++++++++++++++++++++++++++++++++++
 2 files changed, 72 insertions(+), 1 deletion(-)

diff --git a/agent_platform/runtime/loop.py b/agent_platform/runtime/loop.py
index c908ff8..5f09f65 100644
--- a/agent_platform/runtime/loop.py
+++ b/agent_platform/runtime/loop.py
@@ -409,7 +409,15 @@ def python_source_error(arguments: dict[str, Any]) -> str | None:
         ast.parse(content, filename=path)
     except SyntaxError as exc:
         location = f"line {exc.lineno}" if exc.lineno else "unknown line"
-        return f"{exc.msg} ({location})"
+        source_lines = content.splitlines()
+        snippet = ""
+        if exc.lineno:
+            first = max(1, exc.lineno - 1)
+            last = min(len(source_lines), exc.lineno + 1)
+            snippet = "\n" + "\n".join(
+                f"{line_number}: {source_lines[line_number - 1][:240]}" for line_number in range(first, last + 1)
+            )
+        return f"{exc.msg} ({location}). Source around the error:{snippet}"
     return None
 
 
@@ -580,6 +588,7 @@ class AgentLoop:
         unresolved_execution_failure: str | None = None
         failed_execution_revisions: dict[str, int] = {}
         successful_write_fingerprints: set[str] = set()
+        failed_write_fingerprints: set[str] = set()
         consecutive_checkpoint_rejections = 0
         for iteration in range(max_iterations):
             await recorder.emit(
@@ -697,6 +706,9 @@ class AgentLoop:
                 ):
                     blocked_call_reasons[call_id] = "duplicate_write"
                     continue
+                if fingerprint in failed_write_fingerprints:
+                    blocked_call_reasons[call_id] = "failed_write_retry"
+                    continue
                 if fingerprint:
                     batch_write_fingerprints.add(fingerprint)
                 key = execution_call_key(name, arguments) if isinstance(arguments, dict) else None
@@ -760,6 +772,10 @@ class AgentLoop:
                     key = execution_call_key(name, arguments) if isinstance(arguments, dict) else None
                     if key:
                         failed_execution_revisions[key] = mutation_count
+                if not result.get("ok", False):
+                    fingerprint = write_call_fingerprint(name, arguments) if isinstance(arguments, dict) else None
+                    if fingerprint and not result.get("policy_duplicate", False):
+                        failed_write_fingerprints.add(fingerprint)
                 messages.append(
                     {
                         "role": "tool",
@@ -913,6 +929,15 @@ class AgentLoop:
                     "already requested it without an intervening repair."
                 ),
             }
+        elif blocked_reason == "failed_write_retry":
+            result = {
+                "ok": False,
+                "policy_duplicate": True,
+                "error": (
+                    "Execution policy skipped this unchanged write because identical content already failed "
+                    "validation in this run. Inspect the prior validation error and make a real source change."
+                ),
+            }
         elif blocked_reason == "failed_retry":
             result = {
                 "ok": False,
diff --git a/tests/test_agent_loop.py b/tests/test_agent_loop.py
index 3fe8109..0cd2909 100644
--- a/tests/test_agent_loop.py
+++ b/tests/test_agent_loop.py
@@ -59,6 +59,7 @@ def test_invalid_complete_python_source_is_detected_before_write() -> None:
     error = python_source_error({"path": "benchmark.py", "content": "# Report\n- O(n²)\n"})
     assert error is not None
     assert "invalid character" in error
+    assert "2: - O(n²)" in error
     assert python_source_error({"path": "benchmark.py", "content": "print('ok')\n"}) is None
 
 
@@ -900,6 +901,51 @@ async def test_bare_interactive_exec_is_blocked_before_registry(settings) -> Non
         await store.close()
 
 
+async def test_unchanged_invalid_python_write_is_not_revalidated(settings) -> None:
+    invalid_call = {
+        "choices": [
+            {
+                "message": {
+                    "role": "assistant",
+                    "tool_calls": [tool_call("1", "write_file", {"path": "broken.py", "content": "print('x'\n"})],
+                },
+                "finish_reason": "tool_calls",
+            }
+        ]
+    }
+    repeated_call = copy.deepcopy(invalid_call)
+    repeated_call["choices"][0]["message"]["tool_calls"][0]["id"] = "2"
+    final = {"choices": [{"message": {"role": "assistant", "content": "无法完成"}, "finish_reason": "stop"}]}
+    provider = ScriptedProvider([invalid_call, repeated_call, final])
+    registry = RecordingRegistry()
+    store = RuntimeStore(settings.database_url)
+    await store.initialize()
+
+    async def callback(event_type, payload):
+        return None
+
+    try:
+        answer = await AgentLoop(provider, registry, store, max_tool_output_chars=10_000).run(
+            spec=get_model_spec("work-light"),
+            messages=[{"role": "user", "content": "你好"}],
+            identity=UserIdentity("u1", "", "", "user"),
+            raw_user_jwt="jwt",
+            chat_id="invalid-repeat",
+            callback=callback,
+        )
+    finally:
+        await store.close()
+
+    assert answer == "无法完成"
+    assert registry.calls == []
+    repeated_result = next(
+        message["content"]
+        for message in provider.requests[2]["messages"]
+        if message.get("role") == "tool" and "unchanged write" in message.get("content", "")
+    )
+    assert "already failed validation" in repeated_result
+
+
 def test_tool_events_render_one_friendly_completed_card() -> None:
     assert tool_event_details("tool.started", {"name": "exec"}) is None
     rendered = tool_event_details(