tamfis-code 0.6.1__tar.gz → 0.6.2__tar.gz

{tamfis_code-0.6.1/tamfis_code.egg-info → tamfis_code-0.6.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: tamfis-code
-Version: 0.6.1
+Version: 0.6.2
 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-0.6.1 → tamfis_code-0.6.2}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "tamfis-code"
-version = "0.6.1"
+version = "0.6.2"
 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-0.6.1 → tamfis_code-0.6.2}/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__ = "0.6.1"
+__version__ = "0.6.2"
 
 # Bumped whenever a CLI release requires a minimum backend Remote API
 # contract version. Only meaningful for --remote; the standalone path has no

{tamfis_code-0.6.1 → tamfis_code-0.6.2}/tamfis_code/model_registry.py

@@ -59,6 +59,24 @@ MODELS: dict[str, ModelRecord] = {
         ModelCapabilities(vision=True, long_context=True), 1000000,
         ("long_context_review", "repository_search", "vision_assisted_coding"), "high", "medium",
     ),
+    "qwen/qwen3-coder": ModelRecord(
+        "qwen/qwen3-coder", "openrouter",
+        ModelCapabilities(parallel_tool_calls=True, long_context=True), 256000,
+        ("repository_audit", "multi_file_edit", "debugging", "planning", "tool_heavy_execution"),
+        "frontier", "medium",
+    ),
+    "deepseek/deepseek-chat-v3-0324": ModelRecord(
+        "deepseek/deepseek-chat-v3-0324", "openrouter",
+        ModelCapabilities(parallel_tool_calls=True, long_context=True), 128000,
+        ("repository_search", "multi_file_edit", "debugging", "planning"),
+        "high", "low",
+    ),
+    "nvidia/nemotron-3-nano-omni-30b-a3b-reasoning": ModelRecord(
+        "nvidia/nemotron-3-nano-omni-30b-a3b-reasoning", "nvidia",
+        ModelCapabilities(parallel_tool_calls=True, long_context=True), 128000,
+        ("repository_audit", "multi_file_edit", "debugging", "planning", "tool_heavy_execution"),
+        "frontier", "medium",
+    ),
     "Qwen/Qwen2.5-7B-Instruct": ModelRecord(
         "Qwen/Qwen2.5-7B-Instruct", "hf",
         ModelCapabilities(long_context=False), 32768,

{tamfis_code-0.6.1 → tamfis_code-0.6.2}/tamfis_code/providers.py

@@ -226,22 +226,14 @@ class ProviderManager:
             name="NVIDIA NIM",
             base_url="https://integrate.api.nvidia.com/v1",
             api_key_env="NVIDIA_API_KEY",
-            # meta/llama-3.1-70b-instruct is the stable coding/tool default.
-            # Nemotron remains selectable, but live streams have produced
-            # corrupted lexical loops that pass an ordinary finish-reason
-            # check. moonshotai/kimi-k2.6 was tried as the default before
-            # this -- confirmed live (real HTTP calls, real account) that
-            # it's a hard, permanent 404 ("Function ... Not found for
-            # account"), not a transient outage: it's listed in NVIDIA's
-            # general /v1/models catalog but this account (and, going by
-            # the entitlement-shaped error, likely most ordinary NVIDIA API
-            # keys -- third-party marketplace models commonly need separate
-            # enablement NVIDIA's own first-party models don't) has no
-            # actual deployment access to it. A default that 404s outright
-            # for typical accounts is strictly worse than nemotron's
-            # occasional loop -- confirmed the replacement actually works
-            # live before switching to it.
-            default_model="meta/llama-3.1-70b-instruct",
+            # The plain Llama instruct route can answer fluently but has
+            # repeatedly narrated/fabricated local tool results in this CLI.
+            # This NVIDIA reasoning model was verified to return genuine
+            # tool_calls and to accept reasoning parameters without the
+            # indefinite stream stall seen on the plain instruct models.
+            # Keep the older models selectable for accounts where this route
+            # is not enabled, but do not make them the automatic coding path.
+            default_model="nvidia/nemotron-3-nano-omni-30b-a3b-reasoning",
             models=[
                 "nvidia/nemotron-3-super-120b-a12b",
                 "nvidia/nemotron-3-ultra-550b-a55b",
@@ -303,7 +295,10 @@ class ProviderManager:
             name="OpenRouter",
             base_url="https://openrouter.ai/api/v1",
             api_key_env="OPENROUTER_API_KEY",
-            default_model="google/gemini-2.5-flash",
+            # Paid coding/repair work gets a coder-tuned route. Routine work
+            # still uses the free model below, so this does not turn every
+            # inspection into a paid request.
+            default_model="qwen/qwen3-coder",
             # Used for plain chat/Q&A/inspection -- anything that isn't
             # deep research or genuinely demanding coding/analysis work --
             # so routine turns don't spend paid OpenRouter credits at all.

{tamfis_code-0.6.1 → tamfis_code-0.6.2}/tamfis_code/render.py

@@ -432,6 +432,32 @@ class StreamRenderer:
         if self._live is not None:
             self._live.update(self._build_status())
 
+    def _print_plan_snapshot(
+        self, items: list[dict[str, Any]], *, title: str,
+        assumptions: Optional[list[Any]] = None, risks: Optional[list[Any]] = None,
+    ) -> None:
+        """Print an authoritative plan snapshot to scrollback.
+
+        Rich Live is intentionally transient for the spinner, so it cannot
+        be the source of truth for a plan users need to follow. Every plan
+        creation and every status transition gets a durable snapshot here.
+        """
+        table = Table.grid(padding=(0, 1))
+        table.add_column(width=4, justify="right", style="cyan")
+        table.add_column(ratio=1)
+        for index, item in enumerate(items, start=1):
+            status = str(item.get("status") or "pending")
+            marker = "✓" if status == "completed" else (
+                "✗" if status == "failed" else ("▶" if status == "in_progress" else "○")
+            )
+            table.add_row(f"{index}. {marker}", f"{item.get('step') or ''} · {status}")
+        body: list[Any] = [table]
+        if assumptions:
+            body.extend([Text("Assumptions", style="bold"), Text(" • " + "\n • ".join(map(str, assumptions)))])
+        if risks:
+            body.extend([Text("Risks", style="bold yellow"), Text(" • " + "\n • ".join(map(str, risks)))])
+        self.console.print(Panel(Group(*body), title=title, border_style="cyan", expand=False))
+
     def _stop_live(self) -> None:
         """End the progress display before ordinary streamed output begins."""
         if self._live is not None:
@@ -574,14 +600,15 @@ class StreamRenderer:
             return
 
         if event_type == "plan_step_progress":
-            # In-place step-status refresh (see engine.py's
-            # _sync_plan_progress) -- unlike plan_created, this never
-            # reprints the plan banner or touches the spinner phase; it
-            # only updates the live step markers, silently, every round.
+            # Update both the live view and durable scrollback. The live view
+            # may already have been stopped when assistant output began, so
+            # refreshing it alone makes progress appear to disappear.
+            self._close_assistant()
             items = payload.get("items") if isinstance(payload.get("items"), list) else []
             if items:
                 self._plan_steps = [item for item in items if isinstance(item, dict) and item.get("status") != "context"]
                 self._refresh_live()
+                self._print_plan_snapshot(self._plan_steps, title=payload.get("title") or "Plan progress")
             return
 
         if event_type == "plan_created":
@@ -595,21 +622,12 @@ class StreamRenderer:
                 # Rich's TTY Live region is transient and is stopped when
                 # assistant output begins. Always print a durable snapshot;
                 # otherwise the plan disappears at execution start.
-                table = Table.grid(padding=(0, 1))
-                table.add_column(width=4, justify="right", style="cyan")
-                table.add_column(ratio=1)
-                for index, item in enumerate(self._plan_steps, start=1):
-                    status = item.get("status")
-                    marker = "✓" if status == "completed" else ("▶" if status == "in_progress" else "○")
-                    table.add_row(f"{index}. {marker}", str(item.get("step") or ""))
                 assumptions = payload.get("assumptions") or []
                 risks = payload.get("risks") or []
-                body: list[Any] = [table]
-                if assumptions:
-                    body.extend([Text("Assumptions", style="bold"), Text(" • " + "\n • ".join(map(str, assumptions)))])
-                if risks:
-                    body.extend([Text("Risks", style="bold yellow"), Text(" • " + "\n • ".join(map(str, risks)))])
-                self.console.print(Panel(Group(*body), title=payload.get("title") or "Execution plan", border_style="cyan", expand=False))
+                self._print_plan_snapshot(
+                    self._plan_steps, title=payload.get("title") or "Execution plan",
+                    assumptions=assumptions, risks=risks,
+                )
                 return
             if stage == "tool_execution":
                 match = _TOOL_ANNOUNCE_RE.search(content)

{tamfis_code-0.6.1 → tamfis_code-0.6.2}/tamfis_code/runner.py

@@ -546,7 +546,7 @@ async def _stream_task(
                 continue
 
             phase_by_event = {
-                "plan_created": "plan", "tool_call_requested": "execute",
+                "plan_created": "plan", "plan_step_progress": "execute", "tool_call_requested": "execute",
                 "command_started": "execute", "file_mutation": "execute",
                 "approval_required": "waiting_for_approval", "task_diagnostics": "validate",
                 "ai_task_completed": "report", "ai_task_failed": "report",
@@ -566,6 +566,13 @@ async def _stream_task(
                 if state.active_plan_id:
                     local_state.update_plan_steps(session_id, state.active_plan_id, items)
 
+            if event_type == "plan_step_progress":
+                items = payload.get("items") if isinstance(payload.get("items"), list) else []
+                last_plan_items = items
+                state = local_state.get_session_state(session_id)
+                if state.active_plan_id and items:
+                    local_state.update_plan_steps(session_id, state.active_plan_id, items)
+
             if event_type == "file_mutation":
                 state = local_state.get_session_state(session_id)
                 mutation_id = str(payload.get("mutation_id") or "")

{tamfis_code-0.6.1 → tamfis_code-0.6.2}/tamfis_code/runner_local.py

@@ -96,6 +96,18 @@ NARRATED_TOOL_CORRECTION = (
     "Only provide a final answer after the requested inspection or action actually ran."
 )
 
+# Keep the user-facing completion readable even when a provider streams a
+# long audit. Plans and live progress already have their own durable renderer
+# panels; the model must not recreate those as a second, drifting bullet list.
+FINAL_RESPONSE_FORMAT_INSTRUCTION = (
+    "When you finish, write a concise evidence-backed response using exactly these "
+    "sections when applicable: Summary, Changes, Verification, and Remaining issues. "
+    "Use one short bullet per concrete fact under a section; do not emit nested or "
+    "unrelated bullet lists, do not repeat the execution plan, and do not claim a "
+    "tool ran unless a real tool result appears in the conversation. If no files "
+    "changed, say so plainly under Summary."
+)
+
 # Same one-chance-then-fallback shape as narrated tool intent, for the
 # distinct failure of giving up outright instead of narrating.
 MAX_CAPITULATION_RETRIES_PER_PROVIDER = 1
@@ -1947,7 +1959,7 @@ def _looks_like_capitulation(text: str) -> bool:
 # same way: refuse the round, ask for a real tool call, and fall back
 # across providers the same one-chance-then-switch way as narrated intent.
 _FABRICATED_TOOL_RESULT_RE = re.compile(
-    r"the\s+\w+\s+tool\s+(?:has\s+)?(?:found|returned|shows?|revealed|indicates?)|"
+    r"the\s+\w+\s+tool\s+(?:has\s+)?(?:found|returned|executed|ran|shows?|revealed|indicates?)|"
     r"(?:the\s+)?results?\s+(?:suggest|indicate)s?\b|"
     r"i\s+encountered\s+an?\s+(?:access|permission)\s+issue|"
     r"(?:here\s+are\s+the\s+)?allowed\s+directories\s+(?:are|include|where\s+i\s+can)|"
@@ -2929,6 +2941,10 @@ async def run_local_agent_turn(
     # it survives later compaction and remains authoritative in every round.
     insert_at = 1 if working_messages and working_messages[0].get("role") == "system" else 0
     working_messages.insert(insert_at, scope_message)
+    working_messages.insert(insert_at + 1, {
+        "role": "system",
+        "content": FINAL_RESPONSE_FORMAT_INSTRUCTION,
+    })
     if resumed_from_checkpoint or resumed_from_legacy or _requests_autonomous_execution(incoming_objective):
         working_messages.insert(insert_at + 1, {"role": "system", "content": RESUME_EXECUTION_INSTRUCTION})
     checkpoint_mode = str(prior_checkpoint.get("mode") or ("read_only" if read_only else "execute"))

{tamfis_code-0.6.1 → tamfis_code-0.6.2/tamfis_code.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: tamfis-code
-Version: 0.6.1
+Version: 0.6.2
 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-0.6.1 → tamfis_code-0.6.2}/tests/test_fake_tool_detection.py

@@ -25,6 +25,7 @@ from tamfis_code.runner_local import (
     _looks_like_narrated_tool_intent,
     _requests_autonomous_execution,
     _requests_no_confirmation,
+    _looks_like_fabricated_tool_result,
 )
 
 
@@ -55,6 +56,16 @@ class FakeToolCallDetectionTests(unittest.TestCase):
     def test_unregistered_function_looking_text_is_not_flagged(self):
         self.assertFalse(_looks_like_fake_tool_call("def calculate_total(items): return sum(items)"))
 
+    def test_transcript_style_list_directory_claim_is_fabricated_result(self):
+        self.assertTrue(_looks_like_fabricated_tool_result(
+            "The list_directory tool has found the following files and directories."
+        ))
+
+    def test_transcript_style_execute_command_claim_is_fabricated_result(self):
+        self.assertTrue(_looks_like_fabricated_tool_result(
+            "The execute_command tool has executed the npm test command."
+        ))
+
 
 class NarratedToolIntentDetectionTests(unittest.TestCase):
     def test_let_me_check_is_detected(self):

{tamfis_code-0.6.1 → tamfis_code-0.6.2}/tests/test_routing.py

@@ -96,17 +96,13 @@ def test_openrouter_default_is_not_openai_family():
     assert all(not model.startswith("openai/") for model in cfg.models)
 
 
-def test_nvidia_default_model_is_not_loop_prone_nemotron_or_unentitled_kimi():
-    # Live-reported regression: moonshotai/kimi-k2.6 was tried as the
-    # default to avoid nemotron's occasional lexical-loop corruption, but
-    # confirmed live (real HTTP calls against a real account) to be a hard,
-    # permanent 404 -- listed in NVIDIA's general /v1/models catalog but
-    # with no actual deployment access on an ordinary API key. A default
-    # that 404s outright is strictly worse than nemotron's occasional loop.
+def test_nvidia_default_model_is_tool_capable_and_not_unentitled_kimi():
+    # The plain Llama route is fluent but has been observed fabricating local
+    # tool results. Use the verified NVIDIA reasoning/tool route instead;
+    # never use the account-unentitled Kimi route as the default.
     default_model = ProviderManager.PROVIDERS[ProviderType.NVIDIA].default_model
-    assert default_model != "nvidia/nemotron-3-super-120b-a12b"
     assert default_model != "moonshotai/kimi-k2.6"
-    assert default_model == "meta/llama-3.1-70b-instruct"
+    assert default_model == "nvidia/nemotron-3-nano-omni-30b-a3b-reasoning"
 
 
 def test_kimi_k2_6_is_still_selectable_on_openrouter_and_hf():
@@ -126,6 +122,12 @@ def test_kimi_k2_6_is_still_selectable_on_openrouter_and_hf():
     assert ProviderManager.PROVIDERS[ProviderType.HF].default_model != "moonshotai/Kimi-K2.6"
 
 
+def test_openrouter_paid_coding_default_is_qwen_coder():
+    config = ProviderManager.PROVIDERS[ProviderType.OPENROUTER]
+    assert config.default_model == "qwen/qwen3-coder"
+    assert config.default_model in config.models
+
+
 def test_remote_fallback_candidates_stay_in_policy_order():
     manager = _manager_with(ProviderType.NVIDIA, ProviderType.OPENROUTER, ProviderType.HF)
     # Use a tool-requiring task that does not require long context, so HF is

{tamfis_code-0.6.1 → tamfis_code-0.6.2}/tests/test_tamfis_code_render.py

@@ -396,6 +396,32 @@ class StreamRendererTests(unittest.TestCase):
             renderer.finish()
         self.assertIn("inspect mission pipeline", console.file.getvalue())
 
+    def test_plan_progress_is_durable_after_live_panel_stops(self):
+        console = Console(file=StringIO(), no_color=True, width=200, force_terminal=True)
+        renderer = StreamRenderer(console)
+        try:
+            renderer.handle_event({
+                "event_type": "plan_created",
+                "payload": {"title": "Plan", "items": [
+                    {"step": "Inspect", "status": "pending"},
+                    {"step": "Fix", "status": "pending"},
+                ]},
+            })
+            renderer.handle_event({"event_type": "assistant_delta", "payload": {"content": "Working."}})
+            renderer.handle_event({
+                "event_type": "plan_step_progress",
+                "payload": {"items": [
+                    {"step": "Inspect", "status": "completed"},
+                    {"step": "Fix", "status": "in_progress"},
+                ]},
+            })
+        finally:
+            renderer.finish()
+        output = console.file.getvalue()
+        self.assertIn("Plan progress", output)
+        self.assertIn("Inspect · completed", output)
+        self.assertIn("Fix · in_progress", output)
+
     def test_approval_gate_suspends_and_resumes_the_live_status_line(self):
         # Regression guard: Rich's Live redraws on its own timer, independent
         # of a blocking console.input() approval prompt -- without suspending