feat(workers): Local LLM First-Class Tool Calling Adapter for Ollama / vLLM

[sdageltc]

Summary & Motivation

letitloop already features a zero-dependency REST client (orchestrator/llm.py) that connects to any OpenAI-compatible endpoint (such as Ollama or vLLM). However, advanced coding models (like Qwen 2.5 Coder 32B, Llama 3.3 70B, and DeepSeek Coder) provide structured function/tool calling that can execute multi-turn file edits and bash commands directly.

We want to build a dedicated Local Tool-Calling Worker Adapter that empowers users to run complete, self-contained coding loops on consumer GPUs (or Apple Silicon Macs) without third-party agent CLIs or cloud API keys.

---

Proposed Solution

1. LocalToolWorkerAdapter (orchestrator/worker_adapters.py):
   • Connects directly to local Ollama / vLLM /v1/chat/completions API.
   • Defines a standard local tool schema: read_file, write_file, replace_lines, execute_command.
   • Executes multi-turn tool loops within the contract's allowed workspace scope until the objective is achieved or max iterations are reached.
2. Deterministic Output & Evidence Recording:
   • Records every executed tool call and terminal output into the contract's task execution journal.

---

Acceptance Criteria

• Works seamlessly with local Ollama (ollama run qwen2.5-coder) and vLLM instances.
• Restricts local tool execution strictly to contract.workspace_scope.
• Handles tool call parse errors and JSON decoding failures gracefully with automatic prompt repair.
• Unit tests with mocked local LLM tool-calling responses.