drydock-cli 1.1.6__tar.gz → 1.2.0__tar.gz

{drydock_cli-1.1.6 → drydock_cli-1.2.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: drydock-cli
-Version: 1.1.6
+Version: 1.2.0
 Summary: Drydock — nautical CLI coding agent
 Project-URL: Homepage, https://github.com/fbobe321/drydock
 Project-URL: Repository, https://github.com/fbobe321/drydock

{drydock_cli-1.1.6 → drydock_cli-1.2.0}/drydock/core/prompts/cli.md

@@ -17,9 +17,27 @@ Investigate: user wants understanding, explanation, audit, review, or diagnosis
 Change: user wants code created, modified, or fixed → proceed to Plan then Execute.
 If unclear, default to investigate. It is better to explain what you would do than to make an unwanted change.
 
+Multi-Agent Delegation (IMPORTANT):
+You have subagents that run in their own context. USE THEM:
+- For codebase exploration: `task(task="Explore the project structure and report what you find", agent="explore")`
+- For investigating bugs: `task(task="Find the root cause of the X error in module Y", agent="explore")`
+- For code review: `invoke_skill(skill_name="review")`
+- For security audits or architecture review: delegate to a subagent so your main context stays clean
+
+WHEN to delegate:
+- The project has 3+ files to examine → delegate exploration to a subagent
+- You need to understand a codebase before making changes → subagent explores, you fix
+- You're doing a review or audit → use /review skill or delegate
+- A task has multiple independent parts → run subagents in parallel
+
+WHEN NOT to delegate:
+- Simple single-file fixes
+- Quick questions that need 1-2 tool calls
+- Tasks where you already know exactly what to do
+
 Explore. Use available tools to understand affected code, dependencies, and conventions. Never edit a file you haven't read in this session.
 Identify constraints: language, framework, test setup, and any user restrictions on scope.
-When given multiple file paths or a complex task: Do not start reading files immediately. First, summarize your understanding of the task and propose a short plan. Wait for the user to confirm before exploring any files. This prevents wasted effort on the wrong path.
+When given multiple file paths or a complex task: delegate exploration to a subagent with `task(task="...", agent="explore")` first. The subagent explores and reports back. You then fix based on its findings. This keeps your main context clean.
 
 Phase 2 — Plan (Change tasks only)
 State your plan before writing code:

{drydock_cli-1.1.6 → drydock_cli-1.2.0}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "drydock-cli"
-version = "1.1.6"
+version = "1.2.0"
 description = "Drydock — nautical CLI coding agent"
 readme = "README.md"
 requires-python = ">=3.12"

drydock_cli-1.2.0/tests/test_multi_agent.py

@@ -0,0 +1,112 @@
+"""Tests that DryDock actually uses multi-agent delegation.
+
+These MUST run against the real vLLM backend.
+They should FAIL if the model does everything in a single context
+without using the task tool for subagent delegation.
+"""
+
+from __future__ import annotations
+
+import asyncio
+import os
+from pathlib import Path
+
+import httpx
+import pytest
+
+from drydock.core.config.harness_files import init_harness_files_manager
+try:
+    init_harness_files_manager("user", "project")
+except RuntimeError:
+    pass
+
+from drydock.core.agent_loop import AgentLoop
+from drydock.core.agents.models import BuiltinAgentName
+from drydock.core.config import Backend, ModelConfig, ProviderConfig, VibeConfig
+from drydock.core.types import AssistantEvent, BaseEvent, ToolCallEvent, ToolResultEvent
+
+
+def _vllm_ok():
+    try:
+        return httpx.get("http://localhost:8000/v1/models", timeout=3).status_code == 200
+    except Exception:
+        return False
+
+pytestmark = pytest.mark.skipif(not _vllm_ok(), reason="vLLM not running")
+
+
+def _config(tmp_path):
+    return VibeConfig(
+        active_model="devstral", auto_approve=True, enable_telemetry=False,
+        include_project_context=False, system_prompt_id="cli",
+        providers=[ProviderConfig(name="local", api_base="http://localhost:8000/v1", api_key_env_var="", backend=Backend.GENERIC)],
+        models=[ModelConfig(name="devstral", provider="local", input_price=0, output_price=0)],
+        session_logging={"enabled": False, "save_dir": str(tmp_path / "logs")},
+    )
+
+def _agent(tmp_path, max_turns=12):
+    return AgentLoop(config=_config(tmp_path), agent_name=BuiltinAgentName.AUTO_APPROVE, max_turns=max_turns)
+
+async def _run(agent, prompt, max_events=80):
+    events = []
+    async for ev in agent.act(prompt):
+        events.append(ev)
+        if len(events) >= max_events:
+            break
+    return events
+
+
+@pytest.mark.asyncio
+async def test_agent_uses_subagent_naturally(tmp_path):
+    """Agent should use subagents WITHOUT being told to.
+
+    The system prompt must instruct the model to delegate complex tasks
+    to subagents. If the model does everything alone, this test FAILS.
+
+    The prompt does NOT mention subagents, task tool, or delegation.
+    """
+    # Create a multi-file project with enough complexity
+    proj = tmp_path / "project"
+    proj.mkdir()
+    (proj / "main.py").write_text(
+        "from auth import login\nfrom db import get_user\n\n"
+        "def handle_request(username, password):\n"
+        "    user = get_user(username)\n"
+        "    if login(user, password):\n"
+        "        return {'status': 'ok'}\n"
+        "    return {'status': 'denied'}\n"
+    )
+    (proj / "auth.py").write_text(
+        "import hashlib\n\ndef login(user, password):\n"
+        "    hashed = hashlib.md5(password.encode()).hexdigest()\n"
+        "    return user['password_hash'] == hashed\n"
+    )
+    (proj / "db.py").write_text(
+        "USERS = {'admin': {'password_hash': '21232f297a57a5a743894a0e4a801fc3', 'role': 'admin'}}\n\n"
+        "def get_user(username):\n    return USERS.get(username)\n"
+    )
+    (proj / "config.py").write_text("DB_HOST = 'localhost'\nDB_PORT = 5432\nSECRET = 'changeme'\n")
+    (proj / "README.md").write_text("# My App\nA simple auth system.\n")
+
+    agent = _agent(tmp_path)
+
+    # Natural prompt — NO mention of subagents or task tool
+    events = await _run(agent,
+        f"Review the project at {proj}. It has multiple files. "
+        f"Find any security issues and explain the architecture."
+    )
+
+    tool_calls = {}
+    for ev in events:
+        if isinstance(ev, ToolCallEvent):
+            tool_calls[ev.tool_name] = tool_calls.get(ev.tool_name, 0) + 1
+
+    used_task = "task" in tool_calls
+    used_skill = "invoke_skill" in tool_calls
+
+    assert used_task or used_skill, (
+        f"FAIL: Agent did NOT use subagents for a multi-file review task. "
+        f"Tool calls: {tool_calls}. "
+        f"The system prompt must tell the model to use the task tool "
+        f"for codebase exploration and complex multi-file analysis."
+    )