feat(python): implemented host tool event normalization and error checks

- Added `_normalize_host_tool_event` to `RpcClient` to remap transport tool events to executed host tools.
- Updated worker tool-end handling to verify `is_error` is false before marking terminal actions complete.
- Added test coverage in `test_client.py` and `test_worker.py` for host tool event normalization and errored review submission behavior.

python/robomp/src/github_client.py

@@ -265,7 +265,13 @@ class GitHubClient:
                 last_exc = exc
                 log.warning(
                     "transient github 5xx, retrying",
-                    extra={"method": method, "path": path, "attempt": attempt + 1, "delay": delay, "status": exc.status},
+                    extra={
+                        "method": method,
+                        "path": path,
+                        "attempt": attempt + 1,
+                        "delay": delay,
+                        "status": exc.status,
+                    },
                 )
                 time.sleep(delay)
         raise last_exc  # type: ignore[misc]
@@ -294,7 +300,13 @@ class GitHubClient:
                 last_exc = exc
                 log.warning(
                     "transient github 5xx, retrying",
-                    extra={"method": method, "path": path, "attempt": attempt + 1, "delay": delay, "status": exc.status},
+                    extra={
+                        "method": method,
+                        "path": path,
+                        "attempt": attempt + 1,
+                        "delay": delay,
+                        "status": exc.status,
+                    },
                 )
                 await asyncio.sleep(delay)
         raise last_exc  # type: ignore[misc]

python/robomp/src/worker.py

@@ -500,14 +500,21 @@ def _run_rpc_blocking(
 
     def _on_tool_end(event: ToolExecutionEndEvent) -> None:
         tool_name = event.tool_name
-        if event.result is not None:
+        # `tool_name` is transport-normalized by omp_rpc: an xd:// device
+        # dispatch (`write xd://submit_pr_review`) reports the host tool that
+        # ran, so terminal-action detection can match on host-tool names. A
+        # failed execution (`is_error`) does not count as reaching the
+        # terminal action — a rejected submit must still trigger the
+        # completion reminder.
+        ok = event.result is not None and not event.is_error
+        if ok:
             tools_called.add(tool_name)
         log.info(
             "tool_end",
             extra={
                 "issue": bindings.issue_key,
                 "tool": tool_name,
-                "ok": event.result is not None,
+                "ok": ok,
             },
         )
 

python/robomp/tests/test_github_events.py

@@ -975,5 +975,3 @@ def test_route_non_directive_comment_carries_no_pragmas() -> None:
     )
     assert decision.directive is False
     assert decision.directive_pragmas == ()
-
-

python/robomp/tests/test_worker.py

@@ -650,7 +650,7 @@ async def test_run_rpc_stops_reminding_after_terminal_tool(tmp_path: Path, setti
             # driver registers the callback before prompt_and_wait, so we
             # replay it here.
             for cb in client._tool_end_callbacks:
-                cb(SimpleNamespace(tool_name="gh_open_pr", result={}))
+                cb(SimpleNamespace(tool_name="gh_open_pr", result={}, is_error=None))
 
     # Capture the registered tool_end callback on the fake.
     original_on_tool_end = _FakeRpcClient.on_tool_execution_end
@@ -759,7 +759,7 @@ async def test_run_rpc_review_pr_stops_after_submit_without_dirty_probe(
 
     def _on_prompt(client: _FakeRpcClient, _prompt: str) -> None:
         for cb in client._tool_end_callbacks:
-            cb(SimpleNamespace(tool_name="submit_pr_review", result={}))
+            cb(SimpleNamespace(tool_name="submit_pr_review", result={}, is_error=None))
 
     _FakeRpcClient.on_tool_execution_end = _record_tool_end  # type: ignore[assignment]
     try:
@@ -783,6 +783,55 @@ async def test_run_rpc_review_pr_stops_after_submit_without_dirty_probe(
     assert fake.prompts == ["kickoff"]
 
 
+@pytest.mark.asyncio
+async def test_run_rpc_review_pr_still_reminds_when_submit_fails(tmp_path: Path, settings: Settings) -> None:
+    """An errored terminal-tool end event does not count as the terminal action.
+
+    omp_rpc normalizes xd:// device dispatches to the host-tool name, so a
+    rejected `submit_pr_review` surfaces as an end event with `is_error=True`
+    under its real name. Counting it would end the review task silently with
+    no review submitted — the completion reminder must still fire.
+    """
+    inputs, bindings = _make_inputs(tmp_path, settings, session_has_jsonl=False)
+    original_on_tool_end = _FakeRpcClient.on_tool_execution_end
+
+    def _record_tool_end(self, cb) -> None:
+        self._tool_end_callbacks = getattr(self, "_tool_end_callbacks", [])
+        self._tool_end_callbacks.append(cb)
+
+    def _on_prompt(client: _FakeRpcClient, _prompt: str) -> None:
+        for cb in client._tool_end_callbacks:
+            cb(
+                SimpleNamespace(
+                    tool_name="submit_pr_review",
+                    result={"content": [{"type": "text", "text": "GitHub rejected PR review: 422"}]},
+                    is_error=True,
+                )
+            )
+
+    _FakeRpcClient.on_tool_execution_end = _record_tool_end  # type: ignore[assignment]
+    try:
+        _FakeRpcClient.on_prompt = staticmethod(_on_prompt)  # type: ignore[attr-defined]
+        loop = asyncio.new_event_loop()
+        try:
+            worker._run_rpc_blocking(
+                inputs,
+                task_kind="review_pr",
+                prompt="kickoff",
+                loop=loop,
+                bindings=bindings,  # type: ignore[arg-type]
+            )
+        finally:
+            loop.close()
+    finally:
+        _FakeRpcClient.on_tool_execution_end = original_on_tool_end  # type: ignore[assignment]
+        delattr(_FakeRpcClient, "on_prompt")
+
+    fake = _FakeRpcClient.instances[0]
+    assert len(fake.prompts) == 1 + settings.task_completion_max_reminders
+    assert all("submit_pr_review" in p for p in fake.prompts[1:])
+
+
 # ---------------------------------------------------------------------------
 # Dirty-state watchdog
 # ---------------------------------------------------------------------------