v0.2.1

Fix server lifecycle handling while inference providers are loading. FIMLoom now reuses responding endpoints, including llama.cpp HTTP 503 responses, without launching duplicate servers.

FIMLoom.nvim v0.2.0

Added

• Matching-prefix consumption for visible inline suggestions, including
  character-by-character typing, multi-character insertion, newlines, and
  indentation, with strict editor-state validation and no duplicate acceptance.
• completion.preserve_matching_prefix, enabled by default.
• Configurable FIMLoomSuggestion colors through ui.suggestion, optional
  highlight linking through ui.suggestion_hl, and colorscheme-safe restoration.

Fixed

• Restore configured suggestion highlight links after a colorscheme explicitly
  clears FIMLoomSuggestion, while continuing to preserve user and colorscheme
  definitions.
• Remove trailing whitespace-only output lines before both rendering and
  acceptance through the default-enabled completion.trim_trailing_blank_lines.
• Preserve internal blank lines and indentation while sanitizing multiline
  completions.
• Avoid replacement inference while a matching suggestion prefix is being
  consumed; restart normal debounced inference after divergence.

FIMLoom.nvim v0.1.1

Highlights

• Fixed <Tab> acceptance by scheduling ghost-text insertion outside Neovim’s expression-mapping text lock.
• Added Codeium-style accept-key precedence with literal Tab fallback when no suggestion is visible.
• Preserved conflict protection for dismiss and manual-completion mappings.
• Documented sharing a detached llama.cpp server across multiple Neovim sessions.
• Documented --sleep-idle-seconds for unloading inactive model weights.

Verification

• 80 automated tests passing.
• Native Neovim help validated.
• Tab acceptance tested in a real Neovim configuration with Mellum 4B and llama.cpp.

Full changelog: v0.1.0...v0.1.1

FIMLoom.nvim v0.1.0

FIMLoom.nvim v0.1.0 is the first public release of local-first inline code completion for Neovim.

Highlights

• Copilot-style single-line and multiline ghost-text suggestions.
• Native fill-in-the-middle support for Ollama and llama.cpp.
• Configurable local endpoints, ports, models, generation settings, and context limits.
• Debounced asynchronous requests with cancellation and stale-response rejection.
• Optional managed inference-server startup and ownership-aware shutdown.
• :FIMLoom controls and :checkhealth fimloom diagnostics.
• Bounded, redacted diagnostics with optional JSONL persistence.
• Tested end-to-end with JetBrains Mellum 4B through both managed providers.
• Documented lazy.nvim, Ollama, and llama.cpp setup workflows.
• 79 automated tests across Neovim 0.10.4 and stable CI targets.

Requirements

• Neovim 0.10 or newer
• curl
• Ollama or llama-server
• A genuine fill-in-the-middle-capable code model

See the README for installation and tested Mellum setup instructions.

Important model guidance

A suffix-aware prompt template does not make a general chat model FIM-capable. Use model weights whose documentation explicitly reports fill-in-the-middle training or evaluation.

Known limitations

This release does not include streaming, chat or agent modes, cross-file context, multiple suggestions, Windows lifecycle guarantees, or shared ownership across multiple Neovim instances.

Full details are available in the changelog.