feat: tools refactor, started work on subagent

Makefile

@@ -50,3 +50,14 @@ clean-test:
 	echo $$HOME/.local/share/gptme/logs/*test-*-test_*
 	rm -I $$HOME/.local/share/gptme/logs/*test-*-test_*/*.jsonl || true
 	rm --dir $$HOME/.local/share/gptme/logs/*test-*-test_*/ || true
+
+cloc: cloc-core cloc-tools
+
+cloc-core:
+	cloc gptme/*.py
+
+cloc-tools:
+	cloc gptme/tools
+
+cloc-tests:
+	cloc tests/*.py

eval/filestore.py

@@ -28,7 +28,7 @@ def download(self) -> Files:
             if path.is_file():
                 key = str(path.relative_to(self.working_dir))
                 try:
-                    with open(path, "r") as f:
+                    with open(path) as f:
                         files[key] = f.read()
                 except UnicodeDecodeError:
                     # file is binary

eval/types.py

@@ -1,7 +1,8 @@
 from dataclasses import dataclass
-from typing import Callable, Dict, TypedDict
+from typing import TypedDict
+from collections.abc import Callable
 
-Files = Dict[str, str | bytes]
+Files = dict[str, str | bytes]
 
 
 @dataclass

gptme/cli.py

@@ -192,6 +192,9 @@ def chat(
     """
     Run the chat loop.
 
+    prompt_msgs: list of messages to execute in sequence.
+    initial_msgs: list of history messages.
+
     Callable from other modules.
     """
     # init
@@ -233,7 +236,7 @@ def chat(
             codeblock = log.get_last_code_block("assistant", history=1, content=False)
             if not (codeblock and is_supported_codeblock(codeblock)):
                 logger.info("Non-interactive and exhausted prompts, exiting")
-                exit(0)
+                break
 
         # ask for input if no prompt, generate reply, and run tools
         for msg in step(log, no_confirm, model, stream=stream):  # pragma: no cover

gptme/message.py

@@ -9,6 +9,7 @@
 from rich import print
 from rich.console import Console
 from rich.syntax import Syntax
+from tomlkit._utils import escape_string
 from typing_extensions import Self
 
 from .constants import ROLE_COLOR
@@ -80,7 +81,10 @@ def to_toml(self) -> str:
         flags_toml = "\n".join(f"{flag} = true" for flag in flags)
 
         # doublequotes need to be escaped
-        content = self.content.replace('"', '\\"')
+        # content = self.content.replace('"', '\\"')
+        content = escape_string(self.content)
+        content = content.replace("\\n", "\n")
+
         return f'''[message]
 role = "{self.role}"
 content = """

gptme/models.py

@@ -34,13 +34,7 @@ class _ModelDictMeta(TypedDict):
 # TODO: can we get this from the API?
 MODELS: dict[str, dict[str, _ModelDictMeta]] = {
     "openai": {
-        "gpt-4": {
-            "context": 8193,
-            # 0.03 USD per 1k input tokens
-            # 0.06 USD per 1k output tokens
-            "price_input": 0.03,
-            "price_output": 0.06,
-        },
+        # gpt-3.5
         "gpt-3.5-turbo": {
             "context": 4097,
             # 0.001 USD per 1k input tokens
@@ -51,6 +45,17 @@ class _ModelDictMeta(TypedDict):
         "gpt-3.5-turbo-16k": {
             "context": 16385,
         },
+        "gpt-3.5-turbo-1106": {
+            "context": 16385,
+        },
+        # gpt-4 (non-turbo, small context)
+        "gpt-4": {
+            "context": 8193,
+            # 0.03 USD per 1k input tokens
+            # 0.06 USD per 1k output tokens
+            "price_input": 0.03,
+            "price_output": 0.06,
+        },
         # gpt-4-turbo
         # https://openai.com/blog/new-models-and-developer-products-announced-at-devday
         "gpt-4-1106-preview": {
@@ -59,8 +64,8 @@ class _ModelDictMeta(TypedDict):
         "gpt-4-vision-preview": {
             "context": 128_000,
         },
-        "gpt-3.5-turbo-1106": {
-            "context": 16385,
+        "gpt-4-turbo": {
+            "context": 128_000,
         },
     }
 }

gptme/prompts.py

@@ -9,7 +9,15 @@
 
 from .config import get_config
 from .message import Message
-from .tools import browser, patch
+from .tools import (
+    browser,
+    init_tools,
+    loaded_tools,
+    patch,
+    python,
+    save,
+    shell,
+)
 
 PromptType = Literal["full", "short"]
 
@@ -41,6 +49,8 @@ def prompt_full() -> Generator[Message, None, None]:
     yield from prompt_gptme()
 
     yield from prompt_tools()
+    # Useful in debugging
+    #yield from prompt_tools_from_spec()
     yield from prompt_examples()
     yield from prompt_gh()
 
@@ -133,6 +143,24 @@ def prompt_code_interpreter() -> Generator[Message, None, None]:  # pragma: no c
     )
 
 
+def prompt_tools_from_spec() -> Generator[Message, None, None]:
+    # TODO: this should be moved to tools.py
+    # tools must have been initialized by now
+    init_tools()
+    prompt = ""
+    assert loaded_tools, "No tools loaded"
+    for tool in loaded_tools:
+        prompt += (
+            f"""## {tool.name}
+
+{tool.desc.strip()}
+
+{tool.instructions.strip()}""".strip()
+            + "\n\n"
+        )
+    yield Message("system", prompt.strip())
+
+
 def prompt_tools() -> Generator[Message, None, None]:
     python_libraries = get_installed_python_libraries()
     python_libraries_str = "\n".join(f"- {lib}" for lib in python_libraries)
@@ -147,24 +175,24 @@ def prompt_tools() -> Generator[Message, None, None]:
 
 ## python
 
-When you send a message containing Python code (and is not a file block), it will be executed in a stateful environment.
-Python will respond with the output of the execution.
+{python.instructions}
 
 The following libraries are available:
 {python_libraries_str}
 
+The following functions are available in the REPL:
+{python.get_functions_prompt()}
+
 ## bash
 
-When you send a message containing bash code, it will be executed in a stateful bash shell.
-The shell will respond with the output of the execution.
+{shell.instructions}
 
 These programs are available, among others:
 {shell_programs_str}
 
 ## saving files
 
-To save a file, output a code block with a filename on the first line, like "```src/example.py" (a "file block").
-It is very important that such blocks begin with a filename, otherwise the code will be executed instead of saved.
+{save.instructions}
 
 ## patching files
 
@@ -191,29 +219,15 @@ def prompt_examples() -> Generator[Message, None, None]:
 
 ## bash
 
-> User: learn about the project
-```bash
-git ls-files
-```
-> stdout: `README.md`
-```bash
-cat README.md
-```
+{shell.examples}
 
 ## Python
 
-> User: print hello world
-```python
-print("Hello world")
-```
+{python.examples}
 
 ## Save files
 
-> User: write a Hello world script to hello.py
-```hello.py
-print("Hello world")
-```
-Saved to `hello.py`.
+{save.examples}
 
 ## Read files