v0.5.0

The coding agent now writes real code, and the web UI can run the whole loop — from a new idea to an applied change — without the CLI. 7 breaking changes, with Was/Now rows in CHANGELOG.md.

The coding agent works

v0.4.0 shipped a Hermes executor that could not produce a file: its command line was rejected by the pinned image, the hardened container could not create the agent's home, and setup reported the CLI missing from the image it had just downloaded. All three are fixed, and idea executors setup --kind hermes --model <provider/model> --smoke --yes passes for OpenAI, Anthropic, Gemini, DeepSeek, OpenRouter and Ollama, each producing a real file.

The agent is pinned to the provider you configured — base URL and key variable — rather than Hermes's own provider names, which sent an OpenAI key's work to OpenRouter and did not know Ollama at all.

The agent now works on a copy of your project. Its working directory used to start empty, so asked to change server.js it wrote a new one from nothing, and applying it deleted the code that was there. Each run now starts from a fresh copy (git's tracked and untracked, non-ignored files). .env files and anything that looks like a credential are never copied, because the agent has network access; what was left out is listed on the change set. Applying refuses — and writes nothing — if a file was edited after the agent ran, and keeps file permissions.

The web UI runs the loop

• Build has a "Run TASK-n" button, disabled with the reason when a run can't start. A failed run shows the agent's own error. Diffs show only the changed lines; every finding shows on its change set before you approve; a conflict explains which files changed and offers "Reject and re-run".
• Settings shows whether the agent's image is pulled and its key is set, and rejects a model it can't use with the reason.
• Define and Plan show each finding and decision once, on the page where it's decided — including a dependency cycle that blocked plan approval and had shown on no page.
• Demo mode is labelled everywhere; nothing demo-made is called "SHIPPED".
• Works at phone width and in dark mode.

Current models

Defaults are models that answered a real call on 2026-09-17: claude-sonnet-5, gpt-5.6-luna, gemini-3.6-flash, deepseek-v4-pro, anthropic/claude-sonnet-5 on OpenRouter, and qwen3.8 on Ollama. Six ids that now return 404 were removed from the catalog. Every catalog model has a price, so max_cost_usd enforces at real rates — current models had been budgeted at a fraction of theirs.

Read this before upgrading

• Ollama without a default_model now asks for qwen3.8. Run ollama pull qwen3.8, or set default_model: ollama/llama3.1 to keep the old one.
• Tasks run one at a time. POST …/execute is refused while another task's change set in the same initiative is still open, and while the PRD or plan is waiting for review. Apply or reject each change set before running the next task.
• PUT …/executors rejects a model the agent can't use with 400.
• GET …/decisions lists each question once (?all=true for every row), and answering a decision answers its open duplicates.
• One daemon per state directory, even on different addresses.
• idea daemon stop waits for running work to be cancelled (up to 30 s) and prints the pid.

Installing

Download the archive for your platform and put idea on your PATH. The web UI is embedded — no Go or Node toolchain needed.

shasum -a 256 -c checksums.txt

linux/arm64 is built and shipped, but its SQLite driver is compiled per-architecture and is not exercised in CI; see the install section of the README.

Trying it

idea init && idea new "<your idea>" runs the whole loop on a built-in demo provider with no API key, and idea init prints the web UI's address. For real code, set up the coding agent — docs/agents.md covers the providers and the security tradeoffs.

Pre-1.0: the CLI surface, HTTP API and config schema may change in any release.