llm chat --tool and --functions (#1062)

* Tool support for llm chat, closes #1004

docs/help.md

@@ -173,6 +173,13 @@ Options:
   -d, --database FILE          Path to log database
   --no-stream                  Do not stream output
   --key TEXT                   API key to use
+  -T, --tool TEXT              Name of a tool to make available to the model
+  --functions TEXT             Python code block or file path defining functions
+                               to register as tools
+  --td, --tools-debug          Show full details of tool executions
+  --ta, --tools-approve        Manually approve every tool execution
+  --cl, --chain-limit INTEGER  How many chained tool responses to allow, default
+                               5, set 0 for unlimited
   --help                       Show this message and exit.
 ```
 

docs/usage.md

@@ -162,6 +162,17 @@ Approve tool call? [y/N]:
 ```
 The `--functions` option can be passed more than once, and can also point to the filename of a `.py` file containing one or more functions.
 
+If you have any tools that have been made available via plugins you can add them to the prompt using `--tool/-T` option. For example, using [llm-tools-simpleeval](https://github.com/simonw/llm-tools-simpleeval) like this:
+
+```bash
+llm install llm-tools-simpleeval
+llm --tool simple_eval "4444 * 233423" --td
+```
+Run this command to see a list of available tools from plugins:
+```bash
+llm tools
+```
+
 (usage-extract-fenced-code)=
 ### Extracting fenced code blocks
 
@@ -416,6 +427,8 @@ Type '!edit' to open your default editor and modify the prompt.
 > !edit
 ```
 
+`llm chat` takes the same `--tool/-T` and `--functions` options as `llm prompt`. You can use this to start a chat with the specified {ref}`tools <usage-tools>` enabled.
+
 ## Listing available models
 
 The `llm models` command lists every model that can be used with LLM, along with their aliases. This includes models that have been installed using {ref}`plugins <plugins>`.

llm/cli.py

@@ -909,6 +909,42 @@ def prompt(
 )
 @click.option("--no-stream", is_flag=True, help="Do not stream output")
 @click.option("--key", help="API key to use")
+@click.option(
+    "tools",
+    "-T",
+    "--tool",
+    multiple=True,
+    help="Name of a tool to make available to the model",
+)
+@click.option(
+    "python_tools",
+    "--functions",
+    help="Python code block or file path defining functions to register as tools",
+    multiple=True,
+)
+@click.option(
+    "tools_debug",
+    "--td",
+    "--tools-debug",
+    is_flag=True,
+    help="Show full details of tool executions",
+    envvar="LLM_TOOLS_DEBUG",
+)
+@click.option(
+    "tools_approve",
+    "--ta",
+    "--tools-approve",
+    is_flag=True,
+    help="Manually approve every tool execution",
+)
+@click.option(
+    "chain_limit",
+    "--cl",
+    "--chain-limit",
+    type=int,
+    default=5,
+    help="How many chained tool responses to allow, default 5, set 0 for unlimited",
+)
 def chat(
     system,
     model_id,
@@ -920,6 +956,11 @@ def chat(
     no_stream,
     key,
     database,
+    tools,
+    python_tools,
+    tools_debug,
+    tools_approve,
+    chain_limit,
 ):
     """
     Hold an ongoing chat with a model.
@@ -987,7 +1028,18 @@ def chat(
             raise click.ClickException(render_errors(ex.errors()))
 
     kwargs = {}
-    kwargs.update(validated_options)
+    if validated_options:
+        kwargs["options"] = validated_options
+
+    tool_functions = _gather_tools(tools, python_tools)
+
+    if tool_functions:
+        kwargs["chain_limit"] = chain_limit
+        if tools_debug:
+            kwargs["after_call"] = _debug_tool_call
+        if tools_approve:
+            kwargs["before_call"] = _approve_tool_call
+        kwargs["tools"] = tool_functions
 
     should_stream = model.can_stream and not no_stream
     if not should_stream:
@@ -1042,7 +1094,7 @@ def chat(
                 prompt = new_prompt
         if prompt.strip() in ("exit", "quit"):
             break
-        response = conversation.prompt(prompt, system=system, **kwargs)
+        response = conversation.chain(prompt, system=system, **kwargs)
         # System prompt only sent for the first message:
         system = None
         for chunk in response:

llm/models.py

@@ -332,7 +332,7 @@ class Conversation(_BaseConversation):
         details: bool = False,
         key: Optional[str] = None,
         options: Optional[dict] = None,
-    ):
+    ) -> "ChainResponse":
         self.model._validate_attachments(attachments)
         return ChainResponse(
             Prompt(

pyproject.toml

@@ -67,6 +67,7 @@ test = [
     "types-click",
     "types-PyYAML",
     "types-setuptools",
+    "llm-echo==0.3a1",
 ]
 
 [build-system]

tests/conftest.py

@@ -2,6 +2,7 @@ import pytest
 import sqlite_utils
 import json
 import llm
+import llm_echo
 from llm.plugins import pm
 from pydantic import Field
 from pytest_httpx import IteratorStream
@@ -233,19 +234,20 @@ def register_embed_demo_model(embed_demo, mock_model, async_mock_model):
 
 
 @pytest.fixture(autouse=True)
-def register_echo_model():
-    class SimpleEchoModelPlugin:
-        __name__ = "SimpleEchoModelPlugin"
+def register_echo_models():
+    class EchoModelsPlugin:
+        __name__ = "EchoModelsPlugin"
 
         @llm.hookimpl
         def register_models(self, register):
             register(SimpleEchoModel())
+            register(llm_echo.Echo(), llm_echo.EchoAsync())
 
-    pm.register(SimpleEchoModelPlugin(), name="undo-SimpleEchoModelPlugin")
+    pm.register(EchoModelsPlugin(), name="undo-EchoModelsPlugin")
     try:
         yield
     finally:
-        pm.unregister(name="undo-SimpleEchoModelPlugin")
+        pm.unregister(name="undo-EchoModelsPlugin")
 
 
 @pytest.fixture

tests/test_chat.py

@@ -1,9 +1,11 @@
 from click.testing import CliRunner
+import json
 import llm.cli
 from unittest.mock import ANY
 import pytest
 import sys
 import sqlite_utils
+import textwrap
 
 
 @pytest.mark.xfail(sys.platform == "win32", reason="Expected to fail on Windows")
@@ -265,3 +267,69 @@ def test_llm_chat_creates_log_database(tmpdir, monkeypatch, custom_database_path
         assert (user_path / "logs.db").exists()
         db_path = str(user_path / "logs.db")
     assert sqlite_utils.Database(db_path)["responses"].count == 2
+
+
+@pytest.mark.xfail(sys.platform == "win32", reason="Expected to fail on Windows")
+def test_chat_tools(logs_db):
+    runner = CliRunner()
+    functions = textwrap.dedent(
+        """
+    def upper(text: str) -> str:
+        "Convert text to upper case"
+        return text.upper()                         
+    """
+    )
+    result = runner.invoke(
+        llm.cli.cli,
+        ["chat", "-m", "echo", "--functions", functions],
+        input="\n".join(
+            [
+                json.dumps(
+                    {
+                        "prompt": "Convert hello to uppercase",
+                        "tool_calls": [
+                            {"name": "upper", "arguments": {"text": "hello"}}
+                        ],
+                    }
+                ),
+                "quit",
+            ]
+        ),
+        catch_exceptions=False,
+    )
+    assert result.exit_code == 0
+    assert result.output == (
+        "Chatting with echo\n"
+        "Type 'exit' or 'quit' to exit\n"
+        "Type '!multi' to enter multiple lines, then '!end' to finish\n"
+        "Type '!edit' to open your default editor and modify the prompt\n"
+        '> {"prompt": "Convert hello to uppercase", "tool_calls": [{"name": "upper", '
+        '"arguments": {"text": "hello"}}]}\n'
+        "{\n"
+        '  "prompt": "Convert hello to uppercase",\n'
+        '  "system": "",\n'
+        '  "attachments": [],\n'
+        '  "stream": true,\n'
+        '  "previous": []\n'
+        "}{\n"
+        '  "prompt": "",\n'
+        '  "system": "",\n'
+        '  "attachments": [],\n'
+        '  "stream": true,\n'
+        '  "previous": [\n'
+        "    {\n"
+        '      "prompt": "{\\"prompt\\": \\"Convert hello to uppercase\\", '
+        '\\"tool_calls\\": [{\\"name\\": \\"upper\\", \\"arguments\\": {\\"text\\": '
+        '\\"hello\\"}}]}"\n'
+        "    }\n"
+        "  ],\n"
+        '  "tool_results": [\n'
+        "    {\n"
+        '      "name": "upper",\n'
+        '      "output": "HELLO",\n'
+        '      "tool_call_id": null\n'
+        "    }\n"
+        "  ]\n"
+        "}\n"
+        "> quit\n"
+    )

tests/test_plugins.py

@@ -297,11 +297,11 @@ def test_plugins_command():
     result = runner.invoke(cli.cli, ["plugins"])
     assert result.exit_code == 0
     expected = [
+        {"name": "EchoModelsPlugin", "hooks": ["register_models"]},
         {
             "name": "MockModelsPlugin",
             "hooks": ["register_embedding_models", "register_models"],
         },
-        {"name": "SimpleEchoModelPlugin", "hooks": ["register_models"]},
     ]
     actual = json.loads(result.output)
     actual.sort(key=lambda p: p["name"])