feat read files

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})