Repository navigation
Releases
v1.1.0
DevAssistant 1.1.0
A rebuilt inside: the same features on a cleaner, layered architecture, with an agent that
never freezes the window and fixes for problems the rework brought to light.
Download DevAssistant-v1.1.0-windows-x64.zip below, unpack it and run
DevAssistantQt.exe. You also need a model server: Ollama
(for example ollama pull qwen2.5-coder:7b) or any OpenAI-compatible server.
What changed
- The window stays responsive. Model requests are asynchronous; commands that wait on a
program or the network (shell, Python, git, web and document search) run in a worker
thread. "Stop" works at any moment, and answers that arrive for a stopped run are dropped. - The model sees the files it opened.
{{OPEN_FILES}}was always empty before; it now
holds the opened files and lines with their numbers, program output and search results. - Context usage in the status bar shows the size of the last prompt against the model's
context window (it showed 0 before). - No modal pop-ups at the end of a run: the outcome goes to the status bar and the log.
- Agent mode uses the model selected in the chat, as the setting says.
- Model servers that need a key (OpenAI-compatible): the key from the settings is sent.
- File descriptions are generated in the background with a progress window that can be
cancelled.
Architecture
src/coreis a static library without QtWidgets: the session state, an asynchronous model
client (LlmClient,ModelReply), the agent engine (AgentEngine,CommandRunner,
LoopGuard,WorkflowLibrary), the commands, the context, the project, RAG.src/uionly builds the widgets and connects them; the main window went from 2300 lines
and ten "service" interfaces to a few hundred lines.- 12 test suites, among them the agent engine end to end with a scripted model on a
temporary project, and the model client against a fake HTTP server.
See docs/ARCHITECTURE.md.
Requirements
Windows 10 or 11, 64-bit. Voice features need Python 3.10+ and
pip install -r scripts\requirements.txt.
v1.0.0
DevAssistant 1.0.0
The first public release: a desktop AI assistant for programmers that works on local models
and changes the project only through commands you can see and undo.
Download DevAssistant-v1.0.0-windows-x64.zip below, unpack it and run
DevAssistantQt.exe. You also need a model server: Ollama
(for example ollama pull qwen2.5-coder:7b) or any OpenAI-compatible server.
What is in it
- Commands instead of free-form edits. The model answers with JSON commands (open a file
region, edit a line range, apply a diff, search the project, run a program, ask the user);
36 commands, described in docs/COMMANDS.md. - Safe by default.
- Every path must stay inside the project folder.
- Each workflow step lists the commands it may use.
- Read-only commands run at once; changes, programs and deletions are shown first with
the exact command line, paths and line ranges. - Changes are applied all-or-nothing and can be undone, up to 50 steps back.
- Questions about your documents (RAG). A persistent SQLite index next to the project,
incremental updates, answers with the excerpts they rely on; the agent can search it with
SEARCH_DOCS. - LightRAG client. Upload documents, query in graph or vector modes, explore the graph.
- Code insight. An editor with highlighting, the tree-sitter syntax tree, and a project
graph of classes, methods and calls (C/C++, Python, Dart). - Workflows. Multi-step prompt sequences grouped into solution templates. An optional
agent mode plans each step with a separate model call. - Voice (optional, off by default). Dictation with faster-whisper, Vosk, whisper.cpp or
a wav2vec2 model, and answers read aloud. - English and Russian interface. It follows the system language by default; you can
change it in Settings.
Requirements
Windows 10 or 11, 64-bit. Voice features need Python 3.10+ and
pip install -r scripts\requirements.txt.