Execution Flow from Coordinator#26
Conversation
| | Provider | Models | Notes | | ||
| |----------|--------|-------| | ||
| | **Gemini** | `gemini-2.0-flash-exp` (default)<br>`gemini-1.5-pro`<br>`gemini-1.5-flash` | Remote, requires API key | | ||
| | **Ollama** | `llama3.2` (default)<br>`llama3.1`<br>`codellama`<br>`mistral` | Local, free | |
There was a problem hiding this comment.
I changed this bc Ruff linter was complaining about it
| console.print(f"[dim]Using Gemini model: {model}[/dim]") | ||
|
|
||
|
|
||
| def _print_tools_info(tools_description: str, verbose: bool = False) -> None: |
There was a problem hiding this comment.
I moved the print functions to a util file to clean up the logic in this file
| coord = CoordinationService( | ||
| registry_path=registry_path, | ||
| openai_model=openai_model, # for Coordinator to chain subprompts | ||
| step_executor=detected_model, # "ollama" | "gemini" | "openai" |
There was a problem hiding this comment.
This is so the executor could use a variation of LLM Providers, but right now implementation is to require OpenAi Responses for chaining logic.
| typer.echo(plan_obj.model_dump_json(indent=2)) | ||
| # Print trace of tool calls | ||
| trace = result.get("trace") or [] | ||
| if trace: |
There was a problem hiding this comment.
I should prob move this out of this file this is just printing tracing
| """ | ||
| Produce JSON args for one tool call. | ||
|
|
||
| Make compatible with other LLM Providers (needs testing) |
There was a problem hiding this comment.
note comment here - this logic is implemented but I did not test it
| res = self.stdio.request("initialize", params, timeout=timeout) | ||
|
|
||
| # Initialize before sending request | ||
| self.notify("notifications/initialized") |
There was a problem hiding this comment.
need this handshake
| tools_txt = "\n".join(f"- {name}: {desc}" for name, desc in tool_summaries) | ||
|
|
||
| instructions = PLANNING_PROMPT_FULL | ||
| # This is using OpenAI *only* for planning - can be updated to use other LLM provider |
There was a problem hiding this comment.
Note this comment - we should have access to other providers but only using openai right now.
| _DOLLAR_KEY_RE = re.compile( | ||
| r"\$([A-Za-z][A-Za-z0-9_]*)\$?" | ||
| ) # tolerated: $KEY or $KEY$ (could happen in LLM provider mistakes) | ||
|
|
There was a problem hiding this comment.
I added more logic and therefore expanded definition of 'sanitization' here
these methods are for data validation and normalization. Some of the normalization is specific to the flows I have been testing
| ] | ||
| } | ||
| }, | ||
| "email-operations": { |
There was a problem hiding this comment.
adding this back in as default just for simplicity's sake when a user whats to try out the 'hello world' of ShardGuard when copy/pasting the prompt. Without this the user has to 'add-mcp' before trying out the sample prompt
| "args": [ | ||
| "-m", | ||
| "shardguard.mcp_servers.web_server" | ||
| [ |
There was a problem hiding this comment.
fixing issue reading this
1 participant
No description provided.