nanocode-cli 0.2.2__tar.gz → 0.2.4__tar.gz

{nanocode_cli-0.2.2/nanocode_cli.egg-info → nanocode_cli-0.2.4}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: nanocode-cli
-Version: 0.2.2
+Version: 0.2.4
 Summary: A lightweight terminal-based AI coding assistant
 Author-email: hit9 <hit9@icloud.com>
 License-Expression: BSD-3-Clause

{nanocode_cli-0.2.2 → nanocode_cli-0.2.4}/nanocode.py

@@ -41,7 +41,7 @@ from prompt_toolkit.patch_stdout import patch_stdout
 
 JsonValue: TypeAlias = Any
 Json: TypeAlias = dict[str, JsonValue]
-__version__ = "0.2.2"
+__version__ = "0.2.4"
 
 
 class Error(Exception): ...
@@ -1704,7 +1704,7 @@ Core:
 - Follow the plan, but revise it when facts require it.
 
 Tools:
-- Call tools only by emitting JSON in tool_calls. Do not use native tool calls.
+- MUST use JSON tool_calls. Do not use native <tool_call>Tool(args...) syntax.
 - Use multiple tool calls in one turn when they are independent.
 - Prefer specific tools first; use Bash only when no provided tool fits.
 - Prefer Search before Read when locating code or facts; Read only known small ranges or exact files needed for editing.
@@ -1729,6 +1729,7 @@ Input:
 - Current_Context: task-local facts that may expire
 - Goal: current objective
 - Plan: ordered plan
+- Agent_Feedback: latest retry/gate warning for you; follow it before continuing
 - Latest_Tool_Call_Results: latest raw tool call results
 - Latest_User_Input: latest user message
 - Tools: available tool specs
@@ -1834,6 +1835,10 @@ MAIN_AGENT_USER_PROMPT_TEMPLATE = """
 {verification_state}
 -------- Verification_State End -----------
 
+----------- Agent_Feedback Begin ------
+{agent_feedback}
+-------- Agent_Feedback End -----------
+
 ----------- Latest_Tool_Call_Results Begin ------
 {latest_tool_call_results}
 -------- Latest_Tool_Call_Results End -----------
@@ -1888,7 +1893,7 @@ class PromptBuilder:
     def system_prompt(self) -> str:
         return MAIN_AGENT_SYSTEM_PROMPT.replace("{ __tools__ }", self._format_tools()).strip()
 
-    def user_prompt(self, latest_tool_call_results: str) -> str:
+    def user_prompt(self, latest_tool_call_results: str, agent_feedback: str) -> str:
         current = self.session.current
         return MAIN_AGENT_USER_PROMPT_TEMPLATE.format(
             environment=self._format_environment(),
@@ -1898,6 +1903,7 @@ class PromptBuilder:
             goal=current.goal or "(empty)",
             plan=self._format_plan(),
             verification_state=current.verification.format(),
+            agent_feedback=agent_feedback or "(empty)",
             latest_tool_call_results=latest_tool_call_results or "(empty)",
             latest_user_input=current.user_input or "(empty)",
         ).strip()
@@ -1995,6 +2001,8 @@ class ModelClient:
 
         self._record_usage(_json_dict(result.get("usage") if isinstance(result, dict) else None))
         content = self._message_content(result)
+        if content is None:
+            return self._invalid_model_response(self._format_missing_message_content(result))
         return self._parse_model_content(content)
 
     def _write_debug_prompt(self, *, activity: str, messages: list[Json]) -> str:
@@ -2024,13 +2032,9 @@ class ModelClient:
 
     def _parse_model_content(self, content: str) -> Json:
         text = content.strip()
-        if text.startswith("```"):
-            lines = text.splitlines()
-            if lines and lines[0].startswith("```"):
-                lines = lines[1:]
-            if lines and lines[-1].strip() == "```":
-                lines = lines[:-1]
-            text = "\n".join(lines).strip()
+        text = self._strip_leaked_think_tags(text)
+        text = self._strip_json_fence(text)
+        text = self._strip_leaked_think_tags(text)
         try:
             value = json.loads(text)
         except json.JSONDecodeError:
@@ -2039,15 +2043,49 @@ class ModelClient:
             return self._invalid_model_response(content)
         return value
 
+    def _strip_json_fence(self, text: str) -> str:
+        if not text.startswith("```"):
+            return text
+        lines = text.splitlines()
+        if lines and lines[0].startswith("```"):
+            lines = lines[1:]
+        if lines and lines[-1].strip() == "```":
+            lines = lines[:-1]
+        return "\n".join(lines).strip()
+
+    def _strip_leaked_think_tags(self, text: str) -> str:
+        text = text.strip()
+        while text.startswith("</think>"):
+            text = text[len("</think>") :].lstrip()
+        while text.startswith("<think>"):
+            end = text.find("</think>")
+            if end < 0:
+                return text
+            text = text[end + len("</think>") :].lstrip()
+            while text.startswith("</think>"):
+                text = text[len("</think>") :].lstrip()
+        return text
+
     def _invalid_model_response(self, content: str) -> Json:
+        guidance = ""
+        if self._looks_like_native_tool_call(content):
+            guidance = (
+                " Native tool_call syntax is not supported; return one JSON object with tool_calls entries like "
+                '{"name":"Read","intention":"...","args":["nanocode.py","0","100"]}.'
+            )
         return {
             "goal_reached": False,
             "tool_calls": None,
             "message_to_user": None,
             "_format_error": "Invalid model output: expected one JSON object matching the Output JSON schema. Return strict JSON only. Bad output: "
-            + _shorten(content),
+            + _shorten(content)
+            + guidance,
         }
 
+    def _looks_like_native_tool_call(self, content: str) -> bool:
+        text = self._strip_leaked_think_tags(content.strip())
+        return text.startswith("<tool_call>")
+
     def _chat_completions_url(self) -> str:
         url = self.session.api_url.rstrip("/")
         if url.endswith("/chat/completions"):
@@ -2061,7 +2099,7 @@ class ModelClient:
             return {"reasoning": {"effort": self.session.reasoning_effort}}
         return {}
 
-    def _message_content(self, result: JsonValue) -> str:
+    def _message_content(self, result: JsonValue) -> str | None:
         data = _json_dict(result)
         choices = _json_list(data.get("choices"))
         if not choices:
@@ -2069,9 +2107,18 @@ class ModelClient:
         message = _json_dict(_json_dict(choices[0]).get("message"))
         content = message.get("content")
         if not isinstance(content, str):
-            raise LLMError("API response missing message content")
+            return None
         return content
 
+    def _format_missing_message_content(self, result: JsonValue) -> str:
+        choice = _json_dict(_json_list(_json_dict(result).get("choices"))[0])
+        message = _json_dict(choice.get("message"))
+        details: Json = {
+            "finish_reason": choice.get("finish_reason"),
+            "message_keys": sorted(str(key) for key in message.keys()),
+        }
+        return "API response missing message content: " + json.dumps(details, ensure_ascii=False)
+
     def _record_usage(self, usage: Json) -> None:
         prompt_tokens = _json_int(usage.get("prompt_tokens"))
         completion_tokens = _json_int(usage.get("completion_tokens"))
@@ -2627,6 +2674,9 @@ class ConversationCompactor:
 
 @final
 class Agent:
+    MAX_CONSECUTIVE_FORMAT_ERRORS: ClassVar[int] = 3
+    MAX_CONSECUTIVE_SUMMARY_GATES: ClassVar[int] = 1
+
     def __init__(self, session: Session):
         self.session = session
         self.prompt_builder = PromptBuilder(session)
@@ -2635,6 +2685,7 @@ class Agent:
         self.state_updater = AgentStateUpdater(session, self.tool_runner)
         self.compactor = ConversationCompactor(session, self.model_client)
         self.latest_tool_call_results = ""
+        self.latest_agent_feedback = ""
 
     @property
     def latest_tool_call_events(self) -> list[ToolCallEvent]:
@@ -2644,9 +2695,10 @@ class Agent:
         return self.prompt_builder.system_prompt()
 
     def build_user_prompt(self, *, consume_latest_tool_results: bool = True) -> str:
-        prompt = self.prompt_builder.user_prompt(self.latest_tool_call_results)
+        prompt = self.prompt_builder.user_prompt(self.latest_tool_call_results, self.latest_agent_feedback)
         if consume_latest_tool_results:
             self.latest_tool_call_results = ""
+            self.latest_agent_feedback = ""
         return prompt
 
     def request(self, system_prompt: str, user_prompt: str, *, activity: str = "main") -> Json:
@@ -2667,23 +2719,64 @@ class Agent:
         on_message: MessageCallback | None = None,
     ) -> Json:
         self.latest_tool_call_results = ""
+        self.latest_agent_feedback = ""
         self.session.current.user_input = user_input
         self.session.current.goal_reached = False
         self.maybe_auto_compact()
         self.session.append_conversation(UserMessage(content=user_input))
+        consecutive_format_errors = 0
+        consecutive_summary_gates = 0
 
         for _ in range(self.session.max_agent_steps):
             response = self.step()
             format_error = _json_str(response.get("_format_error"))
             if format_error:
-                self.latest_tool_call_results = format_error
+                consecutive_format_errors += 1
+                self.latest_agent_feedback = format_error
+                if consecutive_format_errors >= self.MAX_CONSECUTIVE_FORMAT_ERRORS:
+                    self._report_gate(
+                        on_message,
+                        "Stopped: model returned invalid output "
+                        + str(self.MAX_CONSECUTIVE_FORMAT_ERRORS)
+                        + " times in a row.",
+                        "Format_Gate: stopped after "
+                        + str(self.MAX_CONSECUTIVE_FORMAT_ERRORS)
+                        + " consecutive invalid model outputs. "
+                        + _shorten(format_error, 180),
+                    )
+                    raise LLMError(
+                        "model returned invalid output "
+                        + str(self.MAX_CONSECUTIVE_FORMAT_ERRORS)
+                        + " times in a row: "
+                        + _shorten(format_error, 300)
+                    )
+                self._report_gate(
+                    on_message,
+                    "Retrying: model returned invalid output.",
+                    "Format_Gate: retrying model response. " + _shorten(format_error, 180),
+                )
                 continue
+            consecutive_format_errors = 0
             tool_calls = _json_list(response.get("tool_calls"))
             summary_gate = self._format_tool_summary_gate(tool_calls)
             if summary_gate:
-                self.state_updater.latest_report = ""
-                self.latest_tool_call_results = summary_gate
-                continue
+                consecutive_summary_gates += 1
+                if consecutive_summary_gates <= self.MAX_CONSECUTIVE_SUMMARY_GATES:
+                    self.state_updater.latest_report = ""
+                    self.latest_agent_feedback = summary_gate
+                    self._report_gate(
+                        on_message,
+                        "Retrying: model needs to summarize the latest tool results.",
+                        self._compact_gate_report(summary_gate),
+                    )
+                    continue
+                self._report_gate(
+                    on_message,
+                    "Continuing: model did not summarize tool results after one retry.",
+                    "Tool_Summary_Gate: allowing continuation after one missing-summary retry.",
+                )
+            else:
+                consecutive_summary_gates = 0
             self.apply_response(response)
             if on_message is not None and self.state_updater.latest_report:
                 on_message(self.state_updater.latest_report)
@@ -2704,14 +2797,36 @@ class Agent:
                 return response
             if self.session.current.verification.status == VerificationStatus.REQUIRED:
                 self.session.current.goal_reached = False
-                self.latest_tool_call_results = self._format_verification_gate()
+                self.latest_agent_feedback = self._format_verification_gate()
+                self._report_gate(
+                    on_message,
+                    "Retrying: verification is required before completion.",
+                    "Verification_Gate: retrying until verification is passed or blocked.",
+                )
                 continue
             if not self.session.current.goal_reached:
-                self.latest_tool_call_results = self._format_continuation_hint()
+                self.latest_agent_feedback = self._format_continuation_hint()
+                self._report_gate(
+                    on_message,
+                    "Continuing: goal is not complete yet.",
+                    "Continuation_Gate: goal not reached; retrying next useful action.",
+                )
                 continue
             return response
         raise LLMError("agent step limit reached")
 
+    def _report_gate(self, on_message: MessageCallback | None, message: str, debug_message: str) -> None:
+        if on_message is not None:
+            on_message(debug_message if self.session.debug else message)
+
+    def _compact_gate_report(self, gate: str) -> str:
+        lines = gate.splitlines()
+        headline = lines[0] if lines else "Gate"
+        details = [line for line in lines[1:] if line.startswith("- ")]
+        if details:
+            return headline + ": " + _shorten("; ".join(details[:3]), 220)
+        return headline
+
     def step(self) -> Json:
         response = self.request(self.build_system_prompt(), self.build_user_prompt(), activity="main")
         self.state_updater.apply_tool_call_summaries(response)

{nanocode_cli-0.2.2 → nanocode_cli-0.2.4/nanocode_cli.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: nanocode-cli
-Version: 0.2.2
+Version: 0.2.4
 Summary: A lightweight terminal-based AI coding assistant
 Author-email: hit9 <hit9@icloud.com>
 License-Expression: BSD-3-Clause

{nanocode_cli-0.2.2 → nanocode_cli-0.2.4}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "nanocode-cli"
-version = "0.2.2"
+version = "0.2.4"
 description = "A lightweight terminal-based AI coding assistant"
 readme = "README.md"
 requires-python = ">=3.11"