feat(coding-agent): add standalone streaming codegen CLI and docs

[chrisnk1]

Title

feat(coding-agent): add standalone streaming codegen CLI and docs

Summary

Introduces a self-contained coding agent example that streams code generation from local Ollama models via LangChain. This lives outside the RAG pipeline and is runnable with a single command.

Motivation

• Decouple coding-agent experimentation from the RAG pipeline.
• Provide a minimal, fast-to-run baseline that works with local models (e.g., gemma3, mistral).
• Establish a foundation for upcoming multi-agent (LangGraph) work.

What’s included

• coding_agent_example/run.py: simple streaming code generator using ChatOllama and a minimal prompt.
• coding_agent_example/start.sh: one-liner wrapper with sensible defaults.
• coding_agent_example/tools.py and coding_agent_example/agent.py: scaffolding for future tool/agent work (not used by the simple CLI).
• coding_agent_example/requirements.txt: minimal deps for the example.
• coding_agent_example/README.md: quickstart, advanced usage, and next-steps plan.

How to run

• Simplest:

ollama pull gemma3:4b

./coding_agent_example/start.sh

• Advanced:

uv run -p 3.11 --with coding_agent_example/requirements.txt \
  python coding_agent_example/run.py --repl --model gemma3:latest --base-url http://localhost:11434

• Single-shot to file:

uv run -p 3.11 --with coding_agent_example/requirements.txt \
  python coding_agent_example/run.py --model gemma3:latest --language python --output scratch/example.py

Notes

• Streams tokens for responsive UX; logs show model selection, connectivity, and completion stats.
• Autodetects a usable Ollama model if --model isn’t provided.
• Keeps the RAG stack untouched.

Next steps (planned follow-up PRs)

• Implement a multi-agent architecture using LangGraph, following the Open SWE pattern:
  • Manager: orchestrates workflow and user interaction.
  • Planner: researches the codebase/requirements and proposes a plan (pauses for human approval).
  • Programmer: executes the approved plan (write/modify repo code).
  • Reviewer: validates via tests/linters and verifies the issue is resolved.
• Add human-in-the-loop checkpoints, telemetry, and error recovery.
• Gradually introduce safe file ops and repo-wide context ingestion.