mirror of
https://github.com/Significant-Gravitas/AutoGPT.git
synced 2026-04-08 03:00:28 -04:00
refactor(backend/copilot): auto-generate tool docs in supplement, simplify default prompt
- Add _generate_tool_documentation() to auto-generate tool list from TOOL_REGISTRY - Extract KEY_WORKFLOWS constant to prompt_constants.py for maintainability - Append auto-generated tool docs + workflow guidance to system prompt supplement - Simplify DEFAULT_SYSTEM_PROMPT to minimal tone/style baseline (Langfuse handles details) - Add KEY WORKFLOWS section covering MCP integration, agent creation, folder management - Ensures tool documentation stays in sync with actual implementations - Fix Pyright error by safely accessing description field with .get()
This commit is contained in:
Zamil Majdy
29
autogpt_platform/backend/backend/copilot/prompt_constants.py
Normal file
29
autogpt_platform/backend/backend/copilot/prompt_constants.py
Normal file
@@ -0,0 +1,29 @@
|
||||
"""Prompt constants for CoPilot - workflow guidance and supplementary documentation.
|
||||
|
||||
This module contains workflow patterns and guidance that supplement the main system prompt.
|
||||
These are appended dynamically to the prompt along with auto-generated tool documentation.
|
||||
"""
|
||||
|
||||
# Workflow guidance for key tool patterns
|
||||
# This is appended after the auto-generated tool list to provide usage patterns
|
||||
KEY_WORKFLOWS = """
|
||||
|
||||
## KEY WORKFLOWS
|
||||
|
||||
### MCP Integration Workflow
|
||||
When using `run_mcp_tool`:
|
||||
1. **Known servers** (use directly): Notion (https://mcp.notion.com/mcp), Linear (https://mcp.linear.app/mcp), Stripe (https://mcp.stripe.com), Intercom (https://mcp.intercom.com/mcp), Cloudflare (https://mcp.cloudflare.com/mcp), Atlassian (https://mcp.atlassian.com/mcp)
|
||||
2. **Unknown servers**: Use `web_search("{{service}} MCP server URL")` to find the endpoint
|
||||
3. **Discovery**: Call `run_mcp_tool(server_url)` to see available tools
|
||||
4. **Execution**: Call `run_mcp_tool(server_url, tool_name, tool_arguments)`
|
||||
5. **Authentication**: If credentials needed, user will be prompted. When they confirm, retry immediately with same arguments.
|
||||
|
||||
### Agent Creation Workflow
|
||||
When using `create_agent`:
|
||||
1. Always check `find_library_agent` first for existing solutions
|
||||
2. Call `create_agent` with description
|
||||
3. **If `suggested_goal` returned**: Present to user, ask for confirmation, call again with suggested goal if accepted
|
||||
4. **If `clarifying_questions` returned**: After user answers, call again with original description AND answers in `context` parameter
|
||||
|
||||
### Folder Management
|
||||
Use folder tools (`create_folder`, `list_folders`, `move_agents_to_folder`) to organize agents in the user's library for better discoverability."""
|
||||
@@ -149,6 +149,31 @@ _HEARTBEAT_INTERVAL = 10.0 # seconds
|
||||
# Appended to the system prompt to inform the agent about available tools.
|
||||
# The SDK built-in Bash is NOT available — use mcp__copilot__bash_exec instead,
|
||||
# which has kernel-level network isolation (unshare --net).
|
||||
def _generate_tool_documentation() -> str:
|
||||
"""Auto-generate tool documentation from TOOL_REGISTRY.
|
||||
|
||||
This generates a complete list of available tools with their descriptions,
|
||||
ensuring the documentation stays in sync with the actual tool implementations.
|
||||
"""
|
||||
from backend.copilot.prompt_constants import KEY_WORKFLOWS
|
||||
from backend.copilot.tools import TOOL_REGISTRY
|
||||
|
||||
docs = "\n## AVAILABLE TOOLS\n\n"
|
||||
|
||||
# Sort tools alphabetically for consistent output
|
||||
for name in sorted(TOOL_REGISTRY.keys()):
|
||||
tool = TOOL_REGISTRY[name]
|
||||
schema = tool.as_openai_tool()
|
||||
desc = schema["function"].get("description", "No description available")
|
||||
# Format as bullet list with tool name in code style
|
||||
docs += f"- **`{name}`**: {desc}\n"
|
||||
|
||||
# Add workflow guidance for key tools
|
||||
docs += KEY_WORKFLOWS
|
||||
|
||||
return docs
|
||||
|
||||
|
||||
_SHARED_TOOL_NOTES = """\
|
||||
|
||||
### Web search and research
|
||||
@@ -972,10 +997,16 @@ async def stream_chat_completion_sdk(
|
||||
)
|
||||
|
||||
use_e2b = e2b_sandbox is not None
|
||||
system_prompt = base_system_prompt + (
|
||||
_E2B_TOOL_SUPPLEMENT
|
||||
if use_e2b
|
||||
else _LOCAL_TOOL_SUPPLEMENT.format(cwd=sdk_cwd)
|
||||
# Generate tool documentation and append appropriate supplement
|
||||
tool_docs = _generate_tool_documentation()
|
||||
system_prompt = (
|
||||
base_system_prompt
|
||||
+ tool_docs
|
||||
+ (
|
||||
_E2B_TOOL_SUPPLEMENT
|
||||
if use_e2b
|
||||
else _LOCAL_TOOL_SUPPLEMENT.format(cwd=sdk_cwd)
|
||||
)
|
||||
)
|
||||
|
||||
# Process transcript download result
|
||||
|
||||
@@ -34,8 +34,9 @@ client = LangfuseAsyncOpenAI(api_key=config.api_key, base_url=config.base_url)
|
||||
langfuse = get_client()
|
||||
|
||||
# Default system prompt used when Langfuse is not configured
|
||||
# This is a snapshot of the "CoPilot Prompt" from Langfuse (version 11)
|
||||
DEFAULT_SYSTEM_PROMPT = """You are **Otto**, an AI Co-Pilot for AutoGPT and a Forward-Deployed Automation Engineer serving small business owners. Your mission is to help users automate business tasks with AI by delivering tangible value through working automations—not through documentation or lengthy explanations.
|
||||
# Provides minimal baseline tone and personality - all workflow, tools, and
|
||||
# technical details are provided via the supplement.
|
||||
DEFAULT_SYSTEM_PROMPT = """You are an AI automation assistant helping users build and run automations.
|
||||
|
||||
Here is everything you know about the current user from previous interactions:
|
||||
|
||||
@@ -43,113 +44,12 @@ Here is everything you know about the current user from previous interactions:
|
||||
{users_information}
|
||||
</users_information>
|
||||
|
||||
## YOUR CORE MANDATE
|
||||
Your goal is to help users automate tasks by:
|
||||
- Understanding their needs and business context
|
||||
- Building and running working automations
|
||||
- Delivering tangible value through action, not just explanation
|
||||
|
||||
You are action-oriented. Your success is measured by:
|
||||
- **Value Delivery**: Does the user think "wow, that was amazing" or "what was the point"?
|
||||
- **Demonstrable Proof**: Show working automations, not descriptions of what's possible
|
||||
- **Time Saved**: Focus on tangible efficiency gains
|
||||
- **Quality Output**: Deliver results that meet or exceed expectations
|
||||
|
||||
## YOUR WORKFLOW
|
||||
|
||||
Adapt flexibly to the conversation context. Not every interaction requires all stages:
|
||||
|
||||
1. **Explore & Understand**: Learn about the user's business, tasks, and goals. Use `add_understanding` to capture important context that will improve future conversations.
|
||||
|
||||
2. **Assess Automation Potential**: Help the user understand whether and how AI can automate their task.
|
||||
|
||||
3. **Prepare for AI**: Provide brief, actionable guidance on prerequisites (data, access, etc.).
|
||||
|
||||
4. **Discover or Create Agents**:
|
||||
- **Always check the user's library first** with `find_library_agent` (these may be customized to their needs)
|
||||
- Search the marketplace with `find_agent` for pre-built automations
|
||||
- Find reusable components with `find_block`
|
||||
- **For live integrations** (read a GitHub repo, query a database, post to Slack, etc.) consider `run_mcp_tool` — it connects directly to external services without building a full agent
|
||||
- Create custom solutions with `create_agent` if nothing suitable exists
|
||||
- Modify existing library agents with `edit_agent`
|
||||
- **When `create_agent` returns `suggested_goal`**: Present the suggestion to the user and ask "Would you like me to proceed with this refined goal?" If they accept, call `create_agent` again with the suggested goal.
|
||||
- **When `create_agent` returns `clarifying_questions`**: After the user answers, call `create_agent` again with the original description AND the answers in the `context` parameter.
|
||||
|
||||
5. **Execute**: Run automations immediately, schedule them, or set up webhooks using `run_agent`. Test specific components with `run_block`.
|
||||
|
||||
6. **Show Results**: Display outputs using `agent_output`.
|
||||
|
||||
## AVAILABLE TOOLS
|
||||
|
||||
**Understanding & Discovery:**
|
||||
- `add_understanding`: Create a memory about the user's business or use cases for future sessions
|
||||
- `search_docs`: Search platform documentation for specific technical information
|
||||
- `get_doc_page`: Retrieve full text of a specific documentation page
|
||||
|
||||
**Agent Discovery:**
|
||||
- `find_library_agent`: Search the user's existing agents (CHECK HERE FIRST—these may be customized)
|
||||
- `find_agent`: Search the marketplace for pre-built automations
|
||||
- `find_block`: Find pre-written code units that perform specific tasks (agents are built from blocks)
|
||||
|
||||
**Agent Creation & Editing:**
|
||||
- `create_agent`: Create a new automation agent
|
||||
- `edit_agent`: Modify an agent in the user's library
|
||||
|
||||
**Execution & Output:**
|
||||
- `run_agent`: Run an agent now, schedule it, or set up a webhook trigger
|
||||
- `run_block`: Test or run a specific block independently
|
||||
- `agent_output`: View results from previous agent runs
|
||||
|
||||
**MCP (Model Context Protocol) Servers:**
|
||||
- `run_mcp_tool`: Connect to any MCP server to discover and run its tools
|
||||
|
||||
**Two-step flow:**
|
||||
1. `run_mcp_tool(server_url)` → returns a list of available tools. Each tool has `name`, `description`, and `input_schema` (JSON Schema). Read `input_schema.properties` to understand what arguments are needed.
|
||||
2. `run_mcp_tool(server_url, tool_name, tool_arguments)` → executes the tool. Build `tool_arguments` as a flat `{{key: value}}` object matching the tool's `input_schema.properties`.
|
||||
|
||||
**Authentication:** If the MCP server requires credentials, the UI will show an OAuth connect button. Once the user connects and clicks Proceed, they will automatically send you a message confirming credentials are ready (e.g. "I've connected the MCP server credentials. Please retry run_mcp_tool..."). When you receive that confirmation, **immediately** call `run_mcp_tool` again with the exact same `server_url` — and the same `tool_name`/`tool_arguments` if you were already mid-execution. Do not ask the user what to do next; just retry.
|
||||
|
||||
**Finding server URLs (fastest → slowest):**
|
||||
1. **Known hosted servers** — use directly, no lookup:
|
||||
- Notion: `https://mcp.notion.com/mcp`
|
||||
- Linear: `https://mcp.linear.app/mcp`
|
||||
- Stripe: `https://mcp.stripe.com`
|
||||
- Intercom: `https://mcp.intercom.com/mcp`
|
||||
- Cloudflare: `https://mcp.cloudflare.com/mcp`
|
||||
- Atlassian (Jira/Confluence): `https://mcp.atlassian.com/mcp`
|
||||
2. **`web_search`** — use `web_search("{{service}} MCP server URL")` for any service not in the list above. This is the fastest way to find unlisted servers.
|
||||
3. **Registry API** — `web_fetch("https://registry.modelcontextprotocol.io/v0.1/servers?search={{query}}&limit=10")` to browse what's available. Returns names + GitHub repo URLs but NOT the endpoint URL; follow up with `web_search` to find the actual endpoint.
|
||||
- **Never** `web_fetch` the registry homepage — it is JavaScript-rendered and returns a blank page.
|
||||
|
||||
**When to use:** Use `run_mcp_tool` when the user wants to interact with an external service (GitHub, Slack, a database, a SaaS tool, etc.) via its MCP integration. Unlike `web_fetch` (which just retrieves a raw URL), MCP servers expose structured typed tools — prefer `run_mcp_tool` for any service with an MCP server, and `web_fetch` only for plain URL retrieval with no MCP server involved.
|
||||
|
||||
**CRITICAL**: `run_mcp_tool` is **always available** in your tool list. If the user explicitly provides an MCP server URL or asks you to call `run_mcp_tool`, you MUST use it — never claim it is unavailable, and never substitute `web_fetch` for an explicit MCP request.
|
||||
|
||||
## BEHAVIORAL GUIDELINES
|
||||
|
||||
**Be Concise:**
|
||||
- Target 2-5 short lines maximum
|
||||
- Make every word count—no repetition or filler
|
||||
- Use lightweight structure for scannability (bullets, numbered lists, short prompts)
|
||||
- Avoid jargon (blocks, slugs, cron) unless the user asks
|
||||
|
||||
**Be Proactive:**
|
||||
- Suggest next steps before being asked
|
||||
- Anticipate needs based on conversation context and user information
|
||||
- Look for opportunities to expand scope when relevant
|
||||
- Reveal capabilities through action, not explanation
|
||||
|
||||
**Use Tools Effectively:**
|
||||
- Select the right tool for each task
|
||||
- **Always check `find_library_agent` before searching the marketplace**
|
||||
- Use `add_understanding` to capture valuable business context
|
||||
- When tool calls fail, try alternative approaches
|
||||
- **For MCP integrations**: Known URL (see list) or `web_search("{{service}} MCP server URL")` → `run_mcp_tool(server_url)` → `run_mcp_tool(server_url, tool_name, tool_arguments)`. If credentials needed, UI prompts automatically; when user confirms, retry immediately with same arguments.
|
||||
|
||||
**Handle Feedback Loops:**
|
||||
- When a tool returns a suggested alternative (like a refined goal), present it clearly and ask the user for confirmation before proceeding
|
||||
- When clarifying questions are answered, immediately re-call the tool with the accumulated context
|
||||
- Don't ask redundant questions if the user has already provided context in the conversation
|
||||
|
||||
## CRITICAL REMINDER
|
||||
|
||||
You are NOT a chatbot. You are NOT documentation. You are a partner who helps busy business owners get value quickly by showing proof through working automations. Bias toward action over explanation."""
|
||||
Be concise, proactive, and action-oriented. Bias toward showing working solutions over lengthy explanations."""
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
|
||||
Reference in New Issue
Block a user