v0.0.4

Added

• LSP document synchronization: didOpen, didChange, didSave, didClose notifications
• LSP completions: textDocument/completion request builder with full round-trip support
• CompletionParams construction with URI, position, and context
• Completion response parsing for CompletionList and CompletionItem[]
• Buffer::apply_completion() with textEdit range handling (delete range before insert)
• Buffer::apply_text_edit() for proper LSP range replacement
• Completion ranking: sort by sort_text (server intent), then label length, deterministic deduplication via BTreeMap
• completion.rs module for bridging LSP results to buffer
• rust-analyzer integration test for real completion requests
• Document sync test with Python mock server (gated by python3 availability)

Fixed

• textEdit range bug: Completions now properly delete the specified range before inserting, preventing doubled text (e.g., "fn fo" + "fn foo" → "fn foo")
• Dead code shadow in update_cursor (removed variable re-declaration)
• Cursor positioning: clamp to last character index (line_len - 1), accounting for trailing newlines

Changed

• Moved rank_completions from LspClient to completion.rs module (better organization)
• Completion ranking now respects sort_text field per LSP spec
• Updated Claude workflow to use claude-code-action@v1 with custom prompt and model settings
• Expanded CLAUDE.md with full SDLC coverage, SemVer guidance, and deployment context

v0.0.3 — LSP Client Lifecycle

Highlights

• LSP client lifecycle: spawn, shutdown (graceful + force-kill timeout), restart with backoff, stderr capture.
• JSON-RPC transport over stdio: Content-Length framing, pending response tracking, notification broadcast; round-trip and malformed-message tests.
• Initialize handshake helper: client capabilities, stored InitializeResult; rust-analyzer integration test gated by RUN_RA_INT=1.
• WSL2 path handling: Windows ↔ /mnt/ URI normalization, round-trip tests; rust-analyzer path integration scaffold gated by RUN_WSL_RA_TEST=1.
• CI: installs rust-analyzer component; initialize→shutdown sequence test.

Technical notes

• rust-analyzer integration test gate: set RUN_RA_INT=1.
• WSL integration test gate: set RUN_WSL_RA_TEST=1 (requires /mnt/c and rust-analyzer).

v0.0.2

Idep v0.0.2 — AI Works End-to-End

Released: 2026-03-12

The AI layer is now fully operational. Completions flow end-to-end from keypress through
the completion engine to Ollama and back. The core buffer primitives are in place.
This is the foundation everything else builds on.

---

What's New

AI Completions

• CompletionEngine wired to llm-ls LSP bridge
• FIM (fill-in-the-middle) support for three model families:
  • DeepSeek Coder — <｜fim▁begin｜> / <｜fim▁hole｜> / <｜fim▁end｜>
  • StarCoder2 — <fim_prefix> / <fim_suffix> / <fim_middle>
  • CodeLlama — <PRE> / <SUF> / <MID>
• Model-specific stop sequences — completions halt at function boundaries
• Post-processing truncation via truncate_on_stop() as fallback
• Debounce logic — configurable, default 300ms
• Ollama backend: raw: true + temperature: 0.0 for all completion requests

AI Chat

• Streaming token callback restored to ChatSession::send()
• Debounce wired through to chat context

Core Buffer

• Buffer::insert(pos, text)
• Buffer::delete(range)
• Buffer::lines() -> impl Iterator
• Buffer::to_string()
• Cursor position tracking
• Workspace::open_file(path) -> Buffer
• Workspace::save_file(path, buffer)
• File watcher (notify crate) → triggers Indexer::reindex_file on save
• Full unit test coverage for all buffer operations

---

Validated

Tested against deepseek-coder:1.3b running locally via Ollama on WSL2:

fn add(a: i32, b: i32) -> i32 {
    // ← model fills this
}

Model output: a + b — correct, idiomatic, stopped at }\n. 7 tokens, ~550ms.

See [ollama-smoke-test.md](docs/testing/ollama-smoke-test.md) for the full validation procedure.

---

What's Next — v0.0.3

LSP client goes in next. Target: Week 5.

• idep-lsp client lifecycle: initialize → initialized → shutdown
• Spawn rust-analyzer and typescript-language-server
• textDocument/completion bridged to CompletionEngine
• llm-ls wired as virtual LSP for AI completions
• tree-sitter AST chunking for the indexer (replaces naive line chunking)
• fastembed-rs local embeddings + usearch in-process vector search

---

Notes

• No binary release — editor UI does not exist yet. Binaries ship at v0.1.0-alpha.
• GitHub Sponsors is live (quiet launch, no announcement).
• The ollama_backend integration test currently uses a mock server for CI speed.
  Run cargo test -- --ignored locally to hit a real Ollama instance.

---

Built in Bali 🌴 — [idep.dev](https://idep.dev) · [GitHub](https://github.com/idep-editor/idep)

v0.0.1

v0.0.1 — Repo Foundation

Think in code. Own your tools.

This is the foundation release. No binary, no editor UI — but a clean, tested base to build on in public.

---

What's in this release

Cargo workspace — five crates scaffolded: idep-core, idep-ai, idep-lsp, idep-plugin, idep-index.

AI backend layer — all four backends implemented, tested, and hardened:

• OllamaBackend — local, no API key
• AnthropicBackend
• HuggingFaceBackend
• OpenAiCompatBackend — Groq, Together, LM Studio, anything OpenAI-compatible

All backends include retry with exponential backoff and rate-limit handling.

BackendInfo / info() trait method — runtime diagnostics: which backend is active, whether it's cloud-dependent, whether auth is required.

Config schema — ~/.config/idep/config.toml with XDG resolution. Switch backends by changing one line. Full reference in config.example.toml.

Project docs — SECURITY.md, SUSTAINABILITY.md, CONTRIBUTING.md, CHANGELOG.md, CI, pre-commit hooks.

---

What's not here yet

The editor UI does not exist. Completions are not wired end-to-end. This is a source-only release.

See TODO.md for what's next.

---

Next: v0.0.2

Ollama completions working end-to-end — the first thing you'll actually be able to run.

---

Built in Bali 🌴 — idep.dev