Cicada 0.4.0

Cicada 0.4.0

Frictionless first run and a self-healing pipeline. This release focuses on what happens when things go wrong — especially with a small local model — and makes recovery automatic or one click away.

Zero-setup first launch

• The default model now downloads itself. On first run, if no .gguf is present, Cicada fetches Qwen2.5-Coder 3B Instruct (~1.9 GB) from Hugging Face with live progress in the status pill — on top of the existing llama.cpp auto-install. Fresh machine → working agentic IDE with no manual steps.
• Interrupted downloads never leave a truncated model behind (atomic .part → rename, with a sanity check against error pages).

Self-healing resilience

• Transient model failures are retried. Connection drops, 5xx and 429 responses in any pipeline stage, chat turn, fix loop, or inpaint heal are retried automatically, polling llama-server's /health between attempts so a reloading server gets a real chance. Bad requests (4xx) and user cancels are never retried.
• A crashed llama-server restarts itself — up to 3 times in a rolling 10-minute window, narrated in the status pill ("Model crashed — restarting it…"). Persistent crashes surface a real, actionable error instead of a dead pill.
• No-code answers are re-asked. When a code-producing stage (Generate, Apply Changes, Edit Selection) returns prose with no usable fenced block — the most common small-model failure — the pipeline automatically asks once more with a firmer instruction before reporting an error. Token-budget truncation still gets the targeted "raise context size" message.

Actionable errors, better feedback

• Failed agent runs now raise an error toast with one-click Retry (create, refine, and edit-selection runs all remember their last request).
• The status pill is clickable when the model is stopped or errored: one click re-runs the full recovery path — find/download the binary, find/download the model, restart the server. Full error detail lives in the tooltip.
• Starting a pipeline while the model is down now kicks recovery in the background and says so, instead of a bare "Model is not ready yet."
• Setup failures (binary or model download) show a Retry toast.

Fixes & polish

• Windows path display fixes in the console header and Render captions (no more full C:\… paths).
• Windows-safe relative paths in multi-file compile error output.
• Dock panes get a subtle fade/rise transition (disabled under reduced-motion).
• Guarded an unhandled-rejection path in background recovery.

Verification

• All 15 model-free test suites pass (llm, inpaint, memory, env, datasets, repo, llama-installer, snapshots, experiments, github, doctor, stop, sysmon, insights, syntax check).
• New retry layer covered by an integration check against a mock failing server (503×2 → success).

Downloads: Cicada-Setup-0.4.0-x64.exe (installer) · Cicada-Portable-0.4.0-x64.exe (portable, no install)

Binaries are unsigned; SmartScreen may ask to "Run anyway" on first launch.

Cicada 0.2.0

Windows build of Cicada — an agentic Python IDE powered by a local GGUF model.

Default model: Qwen2.5-Coder-3B-Instruct (Q4_K_M). On first run the app auto-downloads llama.cpp for the best backend it detects — CUDA (NVIDIA), Metal (Apple Silicon), or CPU — and the model into your Downloads folder.

Downloads

• Cicada-Setup-0.2.0-x64.exe — installer
• Cicada-Portable-0.2.0-x64.exe — portable, no install

Builds are unsigned; Windows SmartScreen will prompt "More info → Run anyway" on first launch.