tamfis-code 0.6.4__tar.gz → 0.6.5__tar.gz

{tamfis_code-0.6.4/tamfis_code.egg-info → tamfis_code-0.6.5}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: tamfis-code
-Version: 0.6.4
+Version: 0.6.5
 Summary: TamfisGPT Code -- portable subscription-backed terminal coding agent with local tools, PTY, plans, queues, and approvals
 Project-URL: Homepage, https://github.com/tamfitronics/tamfis-code
 Project-URL: Repository, https://github.com/tamfitronics/tamfis-code

{tamfis_code-0.6.4 → tamfis_code-0.6.5}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "tamfis-code"
-version = "0.6.4"
+version = "0.6.5"
 description = "TamfisGPT Code -- portable subscription-backed terminal coding agent with local tools, PTY, plans, queues, and approvals"
 readme = "README.md"
 requires-python = ">=3.10"

{tamfis_code-0.6.4 → tamfis_code-0.6.5}/tamfis_code/__init__.py

@@ -17,7 +17,7 @@ See docs/REMOTE_AGENT_MASTER_SPEC.md, Phase 21, for the original --remote
 architecture's spec.
 """
 
-__version__ = "0.6.4"
+__version__ = "0.6.5"
 
 # Bumped whenever a CLI release requires a minimum backend Remote API
 # contract version. Only meaningful for --remote; the standalone path has no

{tamfis_code-0.6.4 → tamfis_code-0.6.5}/tamfis_code/runner_local.py

@@ -1183,6 +1183,14 @@ def _legacy_resume_messages(state, incoming_objective: str) -> tuple[list[dict[s
 # the tail of the accumulated text, and triggers within a couple KB of a
 # genuine loop, long before the content can grow large enough to matter.
 _DEGENERATE_REPETITION_RE = re.compile(r"(.{6,200}?)\1{4,}", re.DOTALL)
+# A separate guard for a common context-confusion failure: the model starts
+# replaying a synthetic transcript ("then the user said ... then the
+# assistant responded ...") instead of answering the current task. The
+# individual paragraphs are often too long or slightly varied for the
+# contiguous backreference above, so count the unmistakable transcript
+# markers in the rolling tail.
+_CONVERSATION_ECHO_USER_MARKER = re.compile(r"\bthen\s+the\s+user\s+said\b", re.IGNORECASE)
+_CONVERSATION_ECHO_ASSISTANT_MARKER = re.compile(r"\bthen\s+the\s+assistant\s+(?:said|responded)\b", re.IGNORECASE)
 _DEGENERATE_REPETITION_TAIL_WINDOW = 8_000
 _STREAM_QUALITY_LAG_CHARS = 1_024
 _CORRUPTED_STREAM_MIN_CHARS = 600
@@ -2277,6 +2285,12 @@ async def _stream_one_completion(
                     if _DEGENERATE_REPETITION_RE.search(tail_buffer):
                         quality_failure_reason = "degenerate_repetition"
                         break
+                    if (
+                        len(_CONVERSATION_ECHO_USER_MARKER.findall(tail_buffer)) >= 3
+                        and len(_CONVERSATION_ECHO_ASSISTANT_MARKER.findall(tail_buffer)) >= 3
+                    ):
+                        quality_failure_reason = "conversation_echo"
+                        break
                     if _corrupted_lexical_stream_index(tail_buffer) is not None:
                         quality_failure_reason = "corrupted_output"
                         break
@@ -2308,6 +2322,8 @@ async def _stream_one_completion(
             diagnosis = (
                 "Detected corrupted provider token output; discarding it and requesting a clean route."
                 if quality_failure_reason == "corrupted_output"
+                else "Detected a repeated conversation transcript; discarding it and requesting a clean route."
+                if quality_failure_reason == "conversation_echo"
                 else "Detected the model repeating itself in a loop; discarding that completion."
             )
             renderer.handle_event({
@@ -3634,7 +3650,7 @@ async def run_local_agent_turn(
                 renderer.handle_event({"event_type": "ai_task_failed", "payload": {"error": message}})
                 return TaskOutcome(status="failed", error=message)
 
-        if finish_reason in {"degenerate_repetition", "corrupted_output"}:
+        if finish_reason in {"degenerate_repetition", "conversation_echo", "corrupted_output"}:
             failed_quality_provider = resolved_provider
             quality_failed_providers.add(failed_quality_provider)
             checkpoint_partial_parts.clear()
@@ -3677,6 +3693,8 @@ async def run_local_agent_turn(
             reason_text = (
                 "corrupted token output"
                 if finish_reason == "corrupted_output"
+                else "a repeated conversation transcript"
+                if finish_reason == "conversation_echo"
                 else "a degenerate repetition loop"
             )
             error = (

{tamfis_code-0.6.4 → tamfis_code-0.6.5/tamfis_code.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: tamfis-code
-Version: 0.6.4
+Version: 0.6.5
 Summary: TamfisGPT Code -- portable subscription-backed terminal coding agent with local tools, PTY, plans, queues, and approvals
 Project-URL: Homepage, https://github.com/tamfitronics/tamfis-code
 Project-URL: Repository, https://github.com/tamfitronics/tamfis-code

{tamfis_code-0.6.4 → tamfis_code-0.6.5}/tests/test_runner_local.py

@@ -767,6 +767,29 @@ class RunLocalAgentTurnTests(_StatePatchMixin, unittest.TestCase):
             )
             self.assertNotIn(looping_phrase, visible)
 
+    def test_repeated_conversation_transcript_is_stopped_early(self):
+        with tempfile.TemporaryDirectory() as ws:
+            rounds = [[
+                _chunk(_delta(content=(
+                    'Then the user said "audit the site". Then the assistant responded '
+                    "with a long analysis. " if index % 2 else
+                    'Then the user said "continue until fixed". Then the assistant responded '
+                    "with another long analysis. "
+                ))) for index in range(20)
+            ]]
+            client = _FakeClient(rounds)
+            manager = _FakeManager(client)
+            renderer = _RecordingRenderer()
+            outcome = asyncio.run(run_local_agent_turn(
+                manager, ProviderType.NVIDIA, None,
+                [{"role": "user", "content": "summarize the site"}],
+                self._console(), renderer,
+                workspace_root=ws, session_id=1, approval_policy="auto", interactive=False,
+            ))
+            self.assertEqual(outcome.status, "failed")
+            self.assertIn("repeated conversation transcript", outcome.error)
+            self.assertEqual(len(client.calls), 1)
+
     def test_change_request_completed_with_no_mutation_gets_a_caveat(self):
         """Confirmed live: a weak model can narrate "I'll fix this" -- complete
         with a fabricated "corrected" code block -- without ever calling