feat read files

.gitignore

@@ -3,6 +3,7 @@ __pycache__/
 *.egg-info/
 *.gguf
 *.log
+*.exe
 log.txt
 .venv/
 .env

chatbot.py

@@ -49,7 +49,9 @@ SYSTEM_PROMPT = (
     "identifiers, filenames, and commands mentioned in prose. Use the "
     "web_search tool when a question depends on current information you "
     "may not know (recent releases, changelogs, current best practices, "
-    "error messages you don't recognize)."
+    "error messages you don't recognize). If the user provides a local "
+    "file path or asks to inspect a specific file, prefer the read_file "
+    "tool instead of web_search."
 )
 
 # Number of prior user/assistant exchanges to keep in the prompt. Bounds the
@@ -75,6 +77,29 @@ WEB_SEARCH_TOOL = {
     },
 }
 
+FILE_READ_TOOL = {
+    "type": "function",
+    "function": {
+        "name": "read_file",
+        "description": "Read a local file from disk and return its contents.",
+        "parameters": {
+            "type": "object",
+            "properties": {
+                "path": {"type": "string", "description": "Absolute or relative path to the file."},
+                "start_line": {
+                    "type": "integer",
+                    "description": "Optional 1-based line number to start reading from.",
+                },
+                "end_line": {
+                    "type": "integer",
+                    "description": "Optional 1-based line number to stop reading at.",
+                },
+            },
+            "required": ["path"],
+        },
+    },
+}
+
 # Matches the <tool_call><function=name>...</function></tool_call> blocks
 # Qwen3.5's chat template trains it to emit when it wants to invoke a tool -
 # an XML-style syntax, not Qwen2.5's JSON-in-<tool_call> format.
@@ -109,7 +134,35 @@ def web_search(query: str, max_results: int = 5) -> str:
     return "\n\n".join(f"{r['title']}\n{r['href']}\n{r['body']}" for r in results)
 
 
-TOOLS = {"web_search": web_search}
+def read_file(path: str, start_line: int | None = None, end_line: int | None = None) -> str:
+    """Read a local file from disk. If start_line/end_line are provided, read
+    only that line range (1-based, inclusive). Returns a helpful error string if
+    the file cannot be read."""
+    file_path = Path(path)
+    if not file_path.exists():
+        return f"File not found: {path}"
+    if not file_path.is_file():
+        return f"Not a file: {path}"
+
+    try:
+        lines = file_path.read_text(encoding="utf-8").splitlines()
+    except UnicodeDecodeError:
+        try:
+            lines = file_path.read_text(encoding="utf-8-sig").splitlines()
+        except Exception as exc:  # pragma: no cover - defensive fallback
+            return f"Could not decode file: {path} ({exc})"
+
+    if start_line is None and end_line is None:
+        return "\n".join(lines) + ("\n" if lines and file_path.read_bytes() else "")
+
+    start = 1 if start_line is None else max(1, int(start_line))
+    end = len(lines) if end_line is None else min(len(lines), int(end_line))
+    if start > end:
+        return ""
+    return "\n".join(lines[start - 1 : end]) + ("\n" if start <= end and end <= len(lines) else "")
+
+
+TOOLS = {"web_search": web_search, "read_file": read_file}
 
 
 def extract_tool_calls(text: str) -> list[dict]:
@@ -140,6 +193,9 @@ class ChatSession:
     created_at: str
     updated_at: str
     history: list[tuple[str, str]] = field(default_factory=list)
+    # Tokens occupied in this chat's own context after its last model call.
+    # None until that chat has been evaluated in some session of the app.
+    context_tokens: int | None = None
 
 
 def new_session() -> ChatSession:
@@ -167,6 +223,7 @@ def save_session(session: ChatSession) -> None:
         "created_at": session.created_at,
         "updated_at": session.updated_at,
         "history": [[question, answer] for question, answer in session.history],
+        "context_tokens": session.context_tokens,
     }
     _session_path(session.id).write_text(json.dumps(payload, indent=2), encoding="utf-8")
 
@@ -191,6 +248,7 @@ def list_sessions() -> list[ChatSession]:
                     created_at=data["created_at"],
                     updated_at=data["updated_at"],
                     history=[tuple(turn) for turn in data["history"]],
+                    context_tokens=data.get("context_tokens"),
                 )
             )
         except (json.JSONDecodeError, OSError, KeyError, TypeError):
@@ -367,7 +425,13 @@ def ask(
     invoked the same way but for text inside the model's <think> block,
     letting a caller render its reasoning trace separately from the answer."""
     for _ in range(MAX_TOOL_ROUNDS):
-        content = _stream_content(llm, messages, [WEB_SEARCH_TOOL], on_token, on_thinking_token)
+        content = _stream_content(
+            llm,
+            messages,
+            [WEB_SEARCH_TOOL, FILE_READ_TOOL],
+            on_token,
+            on_thinking_token,
+        )
         if on_usage is not None:
             on_usage({"total_tokens": llm.n_tokens})
         calls = extract_tool_calls(content)
@@ -385,7 +449,7 @@ def ask(
                 result = f"Unknown tool: {name}"
             else:
                 if on_tool_call is not None:
-                    on_tool_call(args.get("query", ""))
+                    on_tool_call(name, args.get("query") or args.get("path", ""))
                 result = func(**args)
             messages.append({"role": "tool", "content": result})
 

gui.py

@@ -439,7 +439,7 @@ class ModelLoaderThread(QThread):
 
 
 class AskThread(QThread):
-    tool_call = pyqtSignal(str)
+    tool_call = pyqtSignal(str, str)
     usage = pyqtSignal(dict)
     token = pyqtSignal(str)
     thinking_token = pyqtSignal(str)
@@ -479,6 +479,11 @@ class ChatWindow(QMainWindow):
             self._sessions[0] if self._sessions else chatbot.new_session()
         )
         self._history: list[tuple[str, str]] = list(self._current_session.history)
+        # The loaded model has a single KV cache. Each chat keeps its own
+        # snapshot of that cache so switching chats restores that chat's
+        # context instead of continuing whichever chat ran last.
+        self._context_states: dict[str, object] = {}
+        self._context_owner: str | None = None
         self._pending_question = ""
         self._ask_thread: AskThread | None = None
         # Position in chat_view's document where the in-progress streamed
@@ -718,6 +723,7 @@ class ChatWindow(QMainWindow):
             return  # already on an empty, unsaved chat
         self._current_session = chatbot.new_session()
         self._history = []
+        self._activate_session_context(self._current_session)
         self.chat_view.clear()
         self._append_system("Dev assist chatbot ready. Ask a development question.")
         self.session_list.clearSelection()
@@ -734,6 +740,7 @@ class ChatWindow(QMainWindow):
             return
         self._current_session = session
         self._history = list(session.history)
+        self._activate_session_context(session)
         self._render_session_history(session)
         self.statusBar().showMessage("Ready.")
 
@@ -759,6 +766,9 @@ class ChatWindow(QMainWindow):
         if reply != QMessageBox.StandardButton.Yes:
             return
         chatbot.delete_session(session_id)
+        self._context_states.pop(session_id, None)
+        if self._context_owner == session_id:
+            self._context_owner = None
         self._sessions = [s for s in self._sessions if s.id != session_id]
         if self._current_session.id == session_id:
             if self._sessions:
@@ -766,6 +776,7 @@ class ChatWindow(QMainWindow):
             else:
                 self._current_session = chatbot.new_session()
             self._history = list(self._current_session.history)
+            self._activate_session_context(self._current_session)
             self.chat_view.clear()
             if self._current_session.history:
                 self._render_session_history(self._current_session)
@@ -776,7 +787,7 @@ class ChatWindow(QMainWindow):
     def _on_model_loaded(self, llm) -> None:
         self._llm = llm
         self._n_ctx = llm.n_ctx()
-        self.context_status_label.setText(f"Context: 0 / {self._n_ctx}")
+        self._activate_session_context(self._current_session)
         self.statusBar().showMessage("Ready.")
         self._append_system("Dev assist chatbot ready. Ask a development question.")
         self._set_input_enabled(True)
@@ -799,6 +810,7 @@ class ChatWindow(QMainWindow):
         self._answer_started = False
         self._stream_thinking_id = None
 
+        self._activate_session_context(self._current_session)
         messages = chatbot.build_messages(self._history, question)
         self._pending_question = question
 
@@ -936,16 +948,47 @@ class ChatWindow(QMainWindow):
             return
         self._render_thinking(block_id, expanded=not block["expanded"])
 
-    def _on_tool_call(self, query: str) -> None:
+    def _on_tool_call(self, tool_name: str, target: str) -> None:
         self._collapse_open_thinking()
         self._discard_stream_card()
-        self.statusBar().showMessage(f"Searching: {query}")
-        self._append_system(f"Searching the web for: {query}")
+        if tool_name == "read_file":
+            self.statusBar().showMessage(f"Reading file: {target}")
+            self._append_system(f"Reading file: {target}")
+        else:
+            self.statusBar().showMessage(f"Searching: {target}")
+            self._append_system(f"Searching the web for: {target}")
+
+    def _show_context_usage(self, total: int | None) -> None:
+        if total is None or not self._n_ctx:
+            shown = "--" if total is None else total
+            limit = self._n_ctx or "--"
+            self.context_status_label.setText(f"Context: {shown} / {limit}")
+            return
+        pct = 100 * total / self._n_ctx
+        self.context_status_label.setText(f"Context: {total} / {self._n_ctx} ({pct:.0f}%)")
+
+    def _activate_session_context(self, session: "chatbot.ChatSession") -> None:
+        """Make the loaded model's KV cache match this chat.
+
+        Leaving a chat stores its cache; entering one restores it, or clears
+        the cache when that chat has no snapshot yet (first visit this run).
+        """
+        self._show_context_usage(session.context_tokens)
+        if self._llm is None or self._context_owner == session.id:
+            return
+        if self._context_owner is not None:
+            self._context_states[self._context_owner] = self._llm.save_state()
+        state = self._context_states.get(session.id)
+        if state is None:
+            self._llm.reset()
+        else:
+            self._llm.load_state(state)
+        self._context_owner = session.id
 
     def _on_usage(self, usage: dict) -> None:
         total = usage["total_tokens"]
-        pct = 100 * total / self._n_ctx if self._n_ctx else 0
-        self.context_status_label.setText(f"Context: {total} / {self._n_ctx} ({pct:.0f}%)")
+        self._current_session.context_tokens = total
+        self._show_context_usage(total)
 
     def _on_answer(self, answer: str) -> None:
         # Normally already collapsed by the first answer token (_on_token),

packaging/build.spec

@@ -69,6 +69,49 @@ a = Analysis(  # noqa: F821
     excludes=[],
     noarchive=False,
 )
+
+# PyInstaller's own dependency walker separately discovers cublas64_13.dll
+# and cublasLt64_13.dll (ggml-cuda.dll imports them directly) and stages its
+# own copy at the bundle root, on top of the copy the `binaries` list above
+# already places in llama_cpp/lib/ (which is where they need to be, next to
+# ggml-cuda.dll, for ITS dependency lookup to find them - see the comment
+# above). The root copy is pure duplication: nothing loads cublas from
+# there. Drop it - it's ~547MB.
+_DUPLICATED_AT_ROOT = {"cublas64_13.dll", "cublasLt64_13.dll"}
+a.binaries = [
+    entry
+    for entry in a.binaries
+    if not (Path(entry[0]).name in _DUPLICATED_AT_ROOT and Path(entry[0]).parent == Path("."))
+]
+
+# Qt/PyQt6 features this app doesn't use, bundled defensively by PyQt6's
+# PyInstaller hook. Confirmed safe to drop: Qt6Widgets/Qt6Gui/Qt6Core don't
+# link any of these at the DLL level (only load them on demand), no
+# PyQt6.QtNetwork/QtSvg/QtPdf Python module is bundled or imported anywhere
+# in this app, no QTranslator is ever installed, and app.setStyle("Fusion")
+# in gui.py forces a style that's built into Qt6Widgets itself rather than
+# loaded from the styles/ plugin.
+_UNUSED_QT_BASENAMES = {
+    "opengl32sw.dll",  # software OpenGL fallback; no QOpenGLWidget/3D used
+    "qt6pdf.dll",
+    "qt6network.dll",
+    "qt6svg.dll",
+    "qsvg.dll",  # imageformats SVG plugin
+    "qsvgicon.dll",  # iconengines SVG plugin
+    "qpdf.dll",  # imageformats PDF plugin (depends on the Qt6Pdf.dll dropped above)
+    "qmodernwindowsstyle.dll",  # styles plugin; app always forces Fusion
+}
+
+
+def _is_unused_qt_binary(dest: str) -> bool:
+    return Path(dest).name.lower() in _UNUSED_QT_BASENAMES
+
+
+a.binaries = [entry for entry in a.binaries if not _is_unused_qt_binary(entry[0])]
+# Qt's .qm translation files ship as data, not binaries - this app never
+# installs a QTranslator, so none of the 96 bundled languages are ever used.
+a.datas = [entry for entry in a.datas if not str(entry[0]).lower().endswith(".qm")]
+
 pyz = PYZ(a.pure)  # noqa: F821
 
 exe = EXE(  # noqa: F821

test_chatbot_tools.py

@@ -0,0 +1,57 @@
+import chatbot
+
+
+def test_read_file_tool_returns_file_contents(tmp_path):
+    sample = tmp_path / "note.txt"
+    sample.write_text("hello\nworld\n", encoding="utf-8")
+
+    assert "read_file" in chatbot.TOOLS
+    assert chatbot.read_file(str(sample)) == "hello\nworld\n"
+
+
+def test_ask_uses_tool_name_in_callback(tmp_path):
+    sample = tmp_path / "note.txt"
+    sample.write_text("hello\nworld\n", encoding="utf-8")
+
+    class DummyLLM:
+        n_tokens = 0
+
+        def __init__(self):
+            self.calls = 0
+
+        def create_chat_completion(self, *args, **kwargs):
+            self.calls += 1
+            if self.calls == 1:
+                yield {
+                    "choices": [{
+                        "delta": {
+                            "content": (
+                                f"<tool_call><function=read_file>"
+                                f"<parameter=path>{sample}</parameter></function></tool_call>"
+                            )
+                        }
+                    }]
+                }
+                return
+            yield {"choices": [{"delta": {"content": "Final answer"}}]}
+
+    seen = []
+    llm = DummyLLM()
+    result = chatbot.ask(
+        llm,
+        [{"role": "user", "content": "read this file"}],
+        on_tool_call=lambda tool_name, target: seen.append((tool_name, target)),
+    )
+
+    assert result == "Final answer"
+    assert seen == [("read_file", str(sample))]
+
+
+def test_gui_tool_signal_accepts_name_and_target():
+    import gui
+    from PyQt6.QtWidgets import QApplication
+
+    app = QApplication.instance() or QApplication([])
+    thread = gui.AskThread(None, [])
+    thread.tool_call.emit("read_file", "C:/temp/example.py")
+    assert app is not None