llmclient: add agentic client for code reviews

[academo]

this PR adds a basic agent client that can use openai/anthropic/google apis to answer questions about code.

The agent is barebones and it is not yet used anywhere. The idea is introduce the client here and then move code-diff to use it directly instead of the gemini-cli

AgenticClient

The agentic client runs a tool-calling loop where the LLM can explore code using read-only tools, then submits structured answers via submit_answer.

Tools: list_directory, read_file, grep, git (allowlisted subcommands), submit_answer

Providers: Google (Gemini), Anthropic (Claude), OpenAI

sequenceDiagram
    participant Caller
    participant AgenticClient
    participant LLM
    participant Tools

    Caller->>AgenticClient: CallLLM(prompt, repoPath, opts)
    AgenticClient->>LLM: system prompt + user prompt + tool definitions

    loop until submit_answer or 100 tool calls
        LLM-->>AgenticClient: tool call(s)
        alt submit_answer
            AgenticClient->>AgenticClient: collect answer
            AgenticClient-->>LLM: "Answer recorded"
        else read_file / list_directory / grep / git
            AgenticClient->>Tools: execute tool (sandboxed to repoPath)
            Tools-->>AgenticClient: result
            AgenticClient-->>LLM: tool result
        end
    end

    Note over LLM,AgenticClient: LLM sends message with no tool calls → done
    AgenticClient-->>Caller: []AnswerSchema

Loading

[academo]

please keep an eye for "testdata" files. like this one. all this file content is fake an irrelevant.

[academo]

this is a real file

[xnyo]

Awesome work! I have some comments but I really like this idea, much much better than the Gemini CLI approach we currently have 🥇 🙌

[xnyo]

LGTM, awesome work! 🙌 🚀