Fix inflated live turn counter in hackbot Reporter

agents/autowebcompat-repro/hackbot_agents/autowebcompat_repro/agent.py

@@ -131,7 +131,7 @@ async def run_autowebcompat_repro(
     user_prompt = build_user_prompt(bug_data, bug_id)
 
     result_msg: ResultMessage | None = None
-    with Reporter(verbose=verbose, log_path=log) as reporter:
+    with Reporter(verbose=verbose, log_path=log, max_turns=max_turns) as reporter:
         reporter.header(subject)
         async with ClaudeSDKClient(options=options) as client:
             await client.query(user_prompt)

agents/bug-fix/hackbot_agents/bug_fix/agent.py

@@ -161,7 +161,7 @@ async def run_bug_fix(
         )
 
     result_msg: ResultMessage | None = None
-    with Reporter(verbose=verbose, log_path=log) as reporter:
+    with Reporter(verbose=verbose, log_path=log, max_turns=max_turns) as reporter:
         reporter.header(f"bug {bug}")
         async with ClaudeSDKClient(options=options) as client:
             await client.query(user_prompt)

agents/build-repair/hackbot_agents/build_repair/agent.py

@@ -204,7 +204,7 @@ async def run_build_repair(
     captured: dict[str, dict] = {}
     tracked = {BUILD_TOOL, *([TRY_PUSH_TOOL] if run_try_push else [])}
 
-    with Reporter(verbose=verbose, log_path=log) as reporter:
+    with Reporter(verbose=verbose, log_path=log, max_turns=max_turns) as reporter:
         # Stage 1: analysis (high effort, no source edits yet).
         reporter.header(f"{label}: analysis")
         analysis_opts = _build_options(

agents/frontend-triage/hackbot_agents/frontend_triage/agent.py

@@ -231,7 +231,7 @@ async def run_frontend_triage(
         )
 
     result_msg: ResultMessage | None = None
-    with Reporter(verbose=verbose, log_path=log) as reporter:
+    with Reporter(verbose=verbose, log_path=log, max_turns=max_turns) as reporter:
         reporter.header(f"bug {bug}")
         async with ClaudeSDKClient(options=options) as client:
             await client.query(user_prompt)

libs/hackbot-runtime/hackbot_runtime/claude.py

@@ -32,10 +32,14 @@ def _truncate(s: str, n: int = 500) -> str:
 class Reporter:
     """Routes streamed claude-agent-sdk messages to stdout and/or a log file."""
 
-    def __init__(self, verbose: bool, log_path: Path | None):
+    def __init__(
+        self, verbose: bool, log_path: Path | None, max_turns: int | None = None
+    ):
         self.verbose = verbose
         self._log = log_path.open("w", encoding="utf-8") if log_path else None
+        self.max_turns = max_turns
         self._turn = 0
+        self._last_msg_id: str | None = None
 
     def __enter__(self):
         return self
@@ -47,6 +51,7 @@ def __exit__(self, *exc):
     def header(self, title: str) -> None:
         """Emit a section header (e.g. ``"bug 12345"``) and reset the turn count."""
         self._turn = 0
+        self._last_msg_id = None
         banner = f"\n{'#' * 60}\n# {title}\n{'#' * 60}"
         self._emit(banner, always=True)
 
@@ -62,8 +67,19 @@ def message(self, msg) -> None:
             is_main = msg.parent_tool_use_id is None
             label = "agent" if is_main else "subagent"
             if is_main:
-                self._turn += 1
-                self._emit(f"\n--- turn {self._turn} ---")
+                msg_id = msg.message_id
+                # The CLI streams one logical model response as several
+                # AssistantMessages (thinking / text / tool_use), all sharing
+                # one message_id. Count a turn only when a new id appears so the
+                # live marker matches ResultMessage.num_turns. If message_id is
+                # unavailable (older CLI), fall back to counting every message.
+                if msg_id is None or msg_id != self._last_msg_id:
+                    self._turn += 1
+                    self._last_msg_id = msg_id
+                    if self.max_turns:
+                        self._emit(f"\n--- turn {self._turn}/{self.max_turns} ---")
+                    else:
+                        self._emit(f"\n--- turn {self._turn} ---")
             for block in msg.content:
                 if isinstance(block, TextBlock):
                     self._emit(f"\n[{label}] {block.text}", always=is_main)

libs/hackbot-runtime/tests/test_runtime.py

@@ -215,3 +215,107 @@ def test_publish_file_copies_locally_without_uploader(tmp_path):
     assert key == "logs/agent.log"
     written = tmp_path / "artifacts" / "local-test" / "logs" / "agent.log"
     assert written.read_text() == "hello log"
+
+
+# --- Reporter turn counting (requires the claude-sdk extra) -----------------
+
+
+def _asst(*, message_id, parent=None):
+    """A main- or sub-agent AssistantMessage carrying a single text block.
+
+    The block content is irrelevant to turn counting; only ``message_id`` and
+    ``parent_tool_use_id`` (main vs subagent) matter here.
+    """
+    from claude_agent_sdk import AssistantMessage, TextBlock
+
+    return AssistantMessage(
+        content=[TextBlock(text="x")],
+        model="claude-test",
+        parent_tool_use_id=parent,
+        message_id=message_id,
+    )
+
+
+def _turn_lines(text):
+    return [ln for ln in text.splitlines() if ln.startswith("--- turn ")]
+
+
+def test_reporter_counts_one_turn_per_message_id(capsys):
+    """Stream fragments sharing a message_id collapse into a single turn."""
+    pytest.importorskip("claude_agent_sdk")
+    from hackbot_runtime.claude import Reporter
+
+    with Reporter(verbose=True, log_path=None) as reporter:
+        # One logical response streamed as three fragments (id "m1"), then a
+        # response with two parallel tool fragments (id "m2"), then "m3".
+        for mid in ["m1", "m1", "m1", "m2", "m2", "m3"]:
+            reporter.message(_asst(message_id=mid))
+
+    assert _turn_lines(capsys.readouterr().out) == [
+        "--- turn 1 ---",
+        "--- turn 2 ---",
+        "--- turn 3 ---",
+    ]
+
+
+def test_reporter_ignores_subagent_messages(capsys):
+    pytest.importorskip("claude_agent_sdk")
+    from hackbot_runtime.claude import Reporter
+
+    with Reporter(verbose=True, log_path=None) as reporter:
+        reporter.message(_asst(message_id="m1"))
+        # Subagent (parent_tool_use_id set) must not bump the turn counter,
+        # even with a fresh message_id.
+        reporter.message(_asst(message_id="s1", parent="tool_1"))
+        reporter.message(_asst(message_id="m2"))
+
+    assert _turn_lines(capsys.readouterr().out) == [
+        "--- turn 1 ---",
+        "--- turn 2 ---",
+    ]
+
+
+def test_reporter_falls_back_when_message_id_missing(capsys):
+    """Older CLI without message_id: every main message counts as a turn."""
+    pytest.importorskip("claude_agent_sdk")
+    from hackbot_runtime.claude import Reporter
+
+    with Reporter(verbose=True, log_path=None) as reporter:
+        for _ in range(3):
+            reporter.message(_asst(message_id=None))
+
+    assert _turn_lines(capsys.readouterr().out) == [
+        "--- turn 1 ---",
+        "--- turn 2 ---",
+        "--- turn 3 ---",
+    ]
+
+
+def test_reporter_header_resets_turn_and_id(capsys):
+    pytest.importorskip("claude_agent_sdk")
+    from hackbot_runtime.claude import Reporter
+
+    with Reporter(verbose=True, log_path=None) as reporter:
+        reporter.message(_asst(message_id="m1"))
+        reporter.header("bug 2")
+        # Same id as before the reset must still open a fresh turn 1.
+        reporter.message(_asst(message_id="m1"))
+
+    assert _turn_lines(capsys.readouterr().out) == [
+        "--- turn 1 ---",
+        "--- turn 1 ---",
+    ]
+
+
+def test_reporter_shows_max_turns_budget(capsys):
+    pytest.importorskip("claude_agent_sdk")
+    from hackbot_runtime.claude import Reporter
+
+    with Reporter(verbose=True, log_path=None, max_turns=30) as reporter:
+        reporter.message(_asst(message_id="m1"))
+        reporter.message(_asst(message_id="m2"))
+
+    assert _turn_lines(capsys.readouterr().out) == [
+        "--- turn 1/30 ---",
+        "--- turn 2/30 ---",
+    ]