fix: fixed prompt chaining, added test (fixes #106)

gptme/cli.py

@@ -45,11 +45,12 @@
 
 
 docstring = f"""
-GPTMe, a chat-CLI for LLMs, enabling them to execute commands and code.
+gptme is a chat-CLI for LLMs, empowering them with tools to run shell commands, execute code, read and manipulate files, and more.
 
 If PROMPTS are provided, a new conversation will be started with it.
 
-If one of the PROMPTS is '{MULTIPROMPT_SEPARATOR}', following prompts will run after the assistant is done answering the first one.
+If one of the PROMPTS is '{MULTIPROMPT_SEPARATOR}', the PROMPTS will form a chain,
+where following prompts will be submitted after the assistant is done answering the previous one.
 
 The interface provides user commands that can be used to interact with the system.
 
@@ -176,7 +177,7 @@ def main(
     if resume:
         name = "resume"  # magic string to load last conversation
 
-    # join prompts, grouped by `-` if present, since that's the separator for multiple-round prompts
+    # join prompts, grouped by `-` if present, since that's the separator for "chained"/multiple-round prompts
     sep = "\n\n" + MULTIPROMPT_SEPARATOR
     prompts = [p.strip() for p in "\n\n".join(prompts).split(sep) if p]
     prompt_msgs = [Message("user", p) for p in prompts]
@@ -260,34 +261,49 @@ def chat(
 
     # main loop
     while True:
-        # if prompt_msgs given, insert next prompt into log
+        # if prompt_msgs given, process each prompt fully before moving to the next
         if prompt_msgs:
-            msg = prompt_msgs.pop(0)
-            if not msg.content.startswith("/"):
-                msg = _include_paths(msg)
-            log.append(msg)
-            # if prompt is a user-command, execute it
-            if execute_cmd(msg, log):
-                continue
+            while prompt_msgs:
+                msg = prompt_msgs.pop(0)
+                if not msg.content.startswith("/"):
+                    msg = _include_paths(msg)
+                log.append(msg)
+                # if prompt is a user-command, execute it
+                if execute_cmd(msg, log):
+                    continue
+
+                # Generate and execute response for this prompt
+                while True:
+                    response_msgs = list(step(log, no_confirm, stream=stream))
+                    for response_msg in response_msgs:
+                        log.append(response_msg)
+                        # run any user-commands, if msg is from user
+                        if response_msg.role == "user" and execute_cmd(
+                            response_msg, log
+                        ):
+                            break
+
+                    # Check if there are any runnable tools left
+                    last_content = next(
+                        (m.content for m in reversed(log) if m.role == "assistant"), ""
+                    )
+                    if not any(
+                        tooluse.is_runnable
+                        for tooluse in ToolUse.iter_from_content(last_content)
+                    ):
+                        break
+
+            # All prompts processed, continue to next iteration
+            continue
+
         # if:
         #  - prompts exhausted
         #  - non-interactive
         #  - no executable block in last assistant message
         # then exit
         elif not interactive:
-            # noreorder
-
-            # continue if we can run tools on the last message
-            last_content = next(
-                (m.content for m in reversed(log) if m.role == "assistant"), ""
-            )
-            tooluses = list(ToolUse.iter_from_content(last_content))
-            runnable = False
-            if tooluses:
-                runnable = any(tooluse.is_runnable for tooluse in tooluses)
-            if not runnable:
-                logger.debug("Non-interactive and exhausted prompts, exiting")
-                break
+            logger.debug("Non-interactive and exhausted prompts, exiting")
+            break
 
         # ask for input if no prompt, generate reply, and run tools
         for msg in step(log, no_confirm, stream=stream):  # pragma: no cover

tests/test_cli.py

@@ -243,6 +243,29 @@ def test_stdin(args: list[str], runner: CliRunner):
     assert result.exit_code == 0
 
 
+@pytest.mark.slow
+def test_chain(args: list[str], runner: CliRunner):
+    """tests that the "-" argument works to chain commands, executing after the agent has exhausted the previous command"""
+    # first command needs to be something requiring two tools, so we can check both are ran before the next chained command
+    args.append("write a test.txt file, then patch it")
+    args.append("-")
+    args.append("read the contents")
+    result = runner.invoke(gptme.cli.main, args)
+    print(result.output)
+    # check that outputs came in expected order
+    user1_loc = result.output.index("User:")
+    user2_loc = result.output.index("User:", user1_loc + 1)
+    save_loc = result.output.index("```save")
+    patch_loc = result.output.index("```patch")
+    print_loc = result.output.rindex("cat test.txt")
+    print(f"{user1_loc=} {save_loc=} {patch_loc=} {user2_loc=} {print_loc=}")
+    assert user1_loc < user2_loc
+    assert save_loc < patch_loc
+    assert patch_loc < user2_loc
+    assert user2_loc < print_loc
+    assert result.exit_code == 0
+
+
 # TODO: move elsewhere
 @pytest.mark.slow
 def test_tmux(args: list[str], runner: CliRunner):