Skip to content

💎 CodeGem: The harness dedicated for Google's Gemma and Gemini models

Choose a tag to compare

View all tags
@github-actions github-actions released this 13 Oct 17:40
· 58 commits to main since this release
v2.20.0
cce1417

This CodeGem release brings together the Gemini and Gemma coding harness
introduced in 2.19.0 with terminal integration, Gemini 4 Argon support, and
local Gemma improvements.

Added

  • CodeGem, a native Python harness dedicated to Gemini and Gemma coding, with
    six bounded tools: read, write, edit, search, Python, and Bash.
  • Gemini-native conversation handling, resumable sessions, and Gemma tool-call
    capability detection with a validated JSON fallback.
  • CodeGem in the terminal connection menu alongside Core, RLM, PiPy, Workbench
    and SystemOne, with focused Gemini and local Gemma model choices.
  • Gemini 4 Argon in the Google model catalog and terminal model lists.
  • Optional local EmbeddingGemma semantic code search, plus reproducible setup,
    cross-runtime validation, and benchmark guides for Ollama, LM Studio, and
    llama.cpp.
  • Installed superqode-codegem-index and superqode-codegem-download commands
    for local retrieval and verified model downloads without a source checkout.
  • CodeGem quick starts, model/runtime configuration guidance, transparent
    evaluation notes, and catalog metadata for first-time users.
  • Multi-file coding validation with API compatibility and input validation checks.

Changed

  • Use the cross-platform Ollama gemma4:12b model in the CodeGem example;
    Apple Silicon users can select the MLX-specific tag when it is available.

Fixed

  • Cancel CodeGem requests using the active terminal session ID.
  • Preserve tool-result IDs when compatible servers receive results in a different order.
  • Route terminal image attachments to CodeGem's Gemini and Ollama providers.
  • Keep local context, output and compaction budgets bounded for imported Gemma models.
  • Run local EmbeddingGemma inference outside the terminal event loop.
  • Support semantic search scoped to one file and ignore stale or replaced source paths.
  • Recognize the native RLM Python tools and Monty boundary in harness diagnostics.
  • Include Gemini thinking tokens in estimated API costs rather than displaying zero.
  • Preserve source-view scroll position across file reloads, theme refreshes,
    and workspace resize events.
Assets 4
Loading