Here's a summary of all user-facing changes on Superconductor from September 6-12:
Major Changes
New model launches
Google's newest Flash model is here. You can now pick Gemini 3.8 Flash from the agent selector on any ticket or chat when you want a fast, inexpensive agent for smaller changes — it's offered through both OpenCode and Cursor.
Pi also picked up DeepSeek V4.1 Flash, at High reasoning effort, alongside the DeepSeek V4 Pro and V4 Flash profiles already there — another quicker, cheaper option when you want to put several agents on the same ticket and compare.
Other Changes
Unblock a domain for just one chat
When the network sandbox blocks an agent in a chat, you can now click Allow for Chat on the banner instead of opening the domain up for the whole project. Tickets have had this since network sandboxing shipped; chats now work the same way.
Work logs say what the agent is doing, not just what it typed
When a Claude Code agent runs a terminal command, the work log now leads with the agent's own one-line description of what it's up to, i.e. "Run the secrets model tests" instead of a wall of shell commands, so you can skim a long run at a glance. The exact command is still one click away: expand a step and you'll see it above the output. Steps with a description can now be expanded while they're still running too, which means you can check what a long-running command is actually executing instead of waiting for it to finish.
Long conversations open faster
Opening a ticket or chat with hundreds of messages used to mean waiting on the whole thread before you could do anything. Superconductor now renders the most recent messages right away and fills in the history behind them, without moving what you're already reading.
The work log tells you when Claude is retrying
If the Claude API times out or returns an error mid-run, the work log now shows it: a step explaining that the request didn't complete, how long the agent waited, and which automatic retry it's on. Runs that stall on a flaky provider no longer look like an agent that quietly stopped thinking.
Benchmarks stop penalizing agents for work you never asked for
When you add a ground truth PR, Superconductor turns it into a spec that agents are graded against. Anything the original developer added while iterating that was never part of the original ticket is now tracked separately, listed on the benchmark ticket under "Ground truth behaviors not in the spec", and left out of scoring.
Meetings follow everyone who joins
Meeting Assistant [Beta] now follows a meeting for every workspace member who joins the call, as soon as they join, instead of waiting for them to say something the transcript picks up. Sit quietly through a call and it still shows up in your Recent Activity sidebar with any tickets the assistant created.
QA Checks stop failing on minor edge cases
QA Checks [Beta] now reserve functionality failures for things that actually matter: unmet ticket requirements, broken primary workflows, regressions, and security or data integrity defects. An unhandled edge case on the margins of a new feature shows up as a suggested follow-up rather than a failing check.
More room for chats in the iOS app
Opening a project chat on iOS now hides the native tab bar and the duplicate title, the way implementation pages already did, so the conversation gets the full screen.
Fixed Issues
Connecting a ChatGPT plan no longer reports false success
Pasting Codex credentials could show a connected ChatGPT plan even when the credentials were expired or invalid, leaving you to find out later when a run failed. Superconductor now validates and refreshes the credentials before saving them, and shows an error in the modal if they don't work.
Finished QA checks mark the conversation unread
A completed QA Check [Beta] posts its report card into the implementation chat, but the conversation wasn't getting the blue unread dot for followers, so the result was easy to miss. It now appears like any other new activity.
Empty diffs no longer contradict the agent
While an implementation had no changes yet, the Full Diff and Uncommitted views could tell you "The agent made no code changes" with the agent still working, or keep showing a waiting state after it had finished. Both now follow the implementation's own status, so what the review pane says matches what the agent is doing.
Have feedback or ideas for things you'd like to see on Superconductor? Let us know!