v0.44.0 — the app is an editor now

The desktop app stopped being a console for agents and became a place to write code. Seven phases of
work land together: an editor with tabs, diagnostics that agree with CI, inline completion from a
local model, search across the workspace, retrieval over the repository, a command runner, and the
ability to hand a coding turn to Claude Code or Gemini CLI.

What you get

An editor at #/edit. CodeMirror with tabs, per-file undo and cursor, and a warning when the
agent changes a file you have open and unsaved. Not on the chat screen — asking a question should not
begin with navigating a tree.

Diagnostics from ruff server. The same rules the build enforces, so a squiggle means something.
The buffer travels with the request, so what you are looking at is what gets checked.

Inline completion from a local model. Grey text ahead of the cursor, Tab takes it, Escape
refuses. Fill-in-the-middle, so the model sees the text on both sides of the caret.

A command runner beside the file. History per project, across reloads, on the arrow keys. Stop
kills the process tree, and closing the panel does the same.

Search across the workspace (ripgrep where it exists, a bounded walk where it does not — the
answer says which), what the machine is spending (CPU, memory, NVIDIA VRAM), and retrieval over
the repository with a pre-registered A/B in bench/rag/.

Somebody else's agent, driving. A coding turn can go to Claude Code or Gemini CLI over ACP and
comes back through the same events, verifier, snapshot and revert as a native turn.

Setting it up

Nothing is required beyond a provider key. Each extra capability is optional and the app tells you
what it needs — run chimera doctor for the whole list, present or absent, with the command that
fixes absent:

To get	Install
Editor diagnostics	pip install ruff (already there if you installed the dev extra)
Inline completion	Ollama, then ollama pull qwen2.5-coder:1.5b-base — a base tag; an instruct model ignores the text after the cursor. Set it in Settings → Models.
Your own Claude Code	Node 22+ and npx. The app uses your existing Claude authentication.
Gemini CLI	npm i -g @google/gemini-cli — its ACP mode is flagged experimental upstream.
Another agent	Point a custom command at whatever adapter you have.

Honest caveats

• The installers are unsigned. Windows SmartScreen and macOS Gatekeeper will warn on first run.
• An external agent has its own file and shell tools. When you hand a turn to Claude Code, our
  write guards are an offer it can decline — the guarantee is checkpoint and revert, not prevention.
  The app says so on screen rather than in a footnote.
• No claim about how good the completions are. The acceptance rate is instrumented and not yet
  measured; GET /api/complete/stats reports it as null until suggestions have been answered.
• Retrieval is not wired into the agent yet. The keyword baseline is measured and published in
  bench/rag/; the semantic half needs an embedder and a number before it ships.
• The runner is not a terminal. No PTY: each command is a fresh process, so cd and export do
  not carry over and nothing interactive gets input. The panel says so where you read it first.

Full detail, including the two bugs found by asking questions the tests had not, is in the
CHANGELOG.