v0.9.67

• Changes panel: see what the agent changed, read the diff, commit the files you pick, undo with a one-hour Restore. Nothing is pushed.
• Chat rows show age, branch and lines changed.
• Claude Code and Codex sessions: continue the chats you started in those apps, or use them as a model, on their own sign-in. Off until you turn it on in Settings.
• Open chats as tabs; a clearer composer bar; long turns fold to "Worked for ...".

v0.9.66

Grok's non-thinking models work again. A chat with a model that refuses any thinking level (xAI's "non-reasoning" Grok models) used to stop with a red 400 error. Radiant now drops the level, asks again in the same message, and remembers the model, so later messages go straight through. Models that do think still get the level you chose.

v0.9.65

Graphs show what each round cost. A graph run now shows the tokens each step used and what each round spent, with a dollar estimate at list price for Claude and OpenAI models. A round that found nothing and still cost money is plain to see while it runs.

A spending limit for repeating graphs. With "Keep going until nothing new turns up" ticked, "Stop after … thousand tokens" stops the graph after the round that reaches it.

A loop step can be told it cannot be done. A loop's checker can now answer "blocked" with a reason, when no further attempt could pass. The loop stops at once and says why, instead of spending the attempts it has left.

v0.9.64

Thinking models keep their train of thought. When a model reasons before acting, that reasoning now carries to its next step within the same reply. Claude with a thinking level set no longer fails after its first tool call. With a ChatGPT sign-in, a 30-task coding test solved every task either way, and it cost about 10% less with this on.

A fence around the agent's commands (optional). Settings → Automation → Sandbox commands. "Project folder only" lets commands change files only in the chat's project folder. "Project folder, no internet" also blocks the web, while local servers keep working. Off unless you turn it on.

Lighter file reads. Files the agent reads carry fewer line numbers, which leaves more room in a long chat.

v0.9.63

When a chat grows past what its model can take, Radiant now first sets aside earlier exchanges about something else, keeping them in the chat but not sending them, and only summarizes if that is not enough.

v0.9.62

As a chat grows, Radiant now asks Jev which earlier tool results the task no longer needs and sets those aside, along with any result a later identical call replaced, so what the task depends on keeps its place.

v0.9.61

Radiant now refuses a second copy of the same message, so a retried connection or a repeated voice request never runs the same work twice. When a send fails, it checks whether your message actually got through before telling you it did not send.

v0.9.60

Tasks is now a real kanban board. Click a card for a two-column view: the description and one feed of comments, changes and every run with its worker log on the left; status, priority, labels, due date and linked tasks on the right. Tasks can wait on other tasks, and agents can use the board themselves to split work into subtasks and leave comments. A stopped task no longer stays in Working.

v0.9.59

Jev Router is now the first model in the OpenRouter list. Choose it and each message goes to whichever model Jev judges best for it, and every reply is labeled with the model that actually answered.

v0.9.58

Every chat now has a bin to the right of the Archive button, so a chat can be deleted in one step instead of going through the archive first. Press and hold it while a red ring fills; let go early to cancel.