Agentic Batch Changes can now merge your changesets when you ask it to.

Nothing merges until you click Allow on the card the agent shows in the conversation. The card lists the changesets with problems first, such as merge conflicts or failed checks. The merges then run in the background. When they finish, the agent picks the conversation back up on its own and reports which changesets merged and which failed, so you don't have to keep checking.

The commits page now helps you narrow down a repository's history and look at changes right next to the list. Filter by message, author, and date, step through diffs with the keyboard, and compare any two commits in place.

A diff only shows a few lines around each change, which often isn't enough to understand it. You can now reveal the hidden lines right in the diff on commit, compare, and changeset pages, or show the whole file with the changes highlighted inline.

When Sourcegraph indexes a repository, it skips some files, for example because they are too large or binary. You can't search the contents of those files, so it's good to know which ones they are. Until now, you had to open each repository's settings to find out.

Agents need more than code to get the full picture. Issue trackers and internal documentation live elsewhere, and many of those services already run MCP servers. With connectors, you can connect MCP servers, so Agentic Batch Changes and the coding agents it runs can interact with those tools on your behalf.

This will let you build more end-to-end workflows right in Agentic Batch Changes, with more workflow extensions to come soon.

You can now keep talking to Agentic Batch Changes while the workspaces it started are running. Ask how a migration is going or what has finished so far.

Before, the agent kept its turn open until the workspace runs finished, which blocked follow-up messages. It now starts the runs and ends its turn, leaving the conversation available while execution continues.

Sourcegraph now supports precise code navigation in .svelte files. Hover over a symbol to see its type and documentation, jump to its definition, or find its references—including symbols used in component templates, powered by SCIP, for compiler-grade results.

Long conversations in Agentic Batch Changes and Deep Search are now easier to get around. Hover the timeline rail next to the conversation and a card opens with your messages and the agent's questions, so you can find the part you need and jump straight to it.

Before, each strip on the rail had its own popover, so you had to hover strips one by one to find a specific question. The card shows the whole thread at once and highlights the part you're currently reading.

Agentic Batch Changes now honors an administrator-defined exclusion list. A repository or file matched by a context filter is invisible to its agents.

The filters apply at every point where the agent touches code. The coordination agent cannot search or read excluded content, workspaces in excluded repositories are neither resolved nor executed, and publishing from an excluded repository is refused. The configuration validator reports invalid rules that do not compile, and if an invalid rule reaches the agent through file-based configuration, every repository is blocked until the rule is fixed.

Our Compare page now supports code intelligence while viewing diffs. Access documentation, look at definitions, or navigate to references across your entire codebase, right from the diff without friction. This of course uses our leading in class, compiler accurate precise code intelligence.

When a changeset fails to publish in Agentic Batch Changes, the Resolve view now tells you why and how to fix it. Before, every failure looked the same: a generic Failed badge and a raw error from the code host.

Clicking Execute in Agentic Batch Changes now queues the selected workspaces right away, even while the agent is in the middle of a turn. A toast shows what was queued, what was skipped, and why.

Before, the button sent the agent a chat message asking it to run them. You waited on a model turn for something the server can do on its own, and the request showed up in the conversation as if you had typed it. A button should do what it says, and the agent should be told, not asked.

Administrators can now connect code hosts through guided wizards instead of editing JSON. Each wizard walks through connection details and authentication, helps you find repositories with search and query previews, and offers optional advanced settings.

Signed-commit rules blocked the merge of pull requests that Batch Changes opened for you. You can now generate an SSH commit signing key for your existing GitHub or GitLab credential in Sourcegraph, without help from an administrator.

Go to Settings, Batch Changes, and Code Hosts. Click Generate signing key, and add the public key to your code host. Commits that Sourcegraph pushes for you from then on carry a Verified badge and pass the rule. Existing commits are not updated.

Typing speed is the bottleneck to express yourself. Talking lets you have your creative ideas flow and produce better prompts, and get better results.

Codebases are growing and changing faster than ever, but the problem of maintaining systems - old and new - is not going away for codebase owners. More of the code is written by agents now, and fewer people can hold the whole codebase in their head. The changes that touch everything (migrations, upgrades, security fixes) are the ones that keep getting deferred.

Code Finder can now search a specified revision, including a branch, tag, or commit.

Use it to investigate regressions at exact commits, find code on non-default release branches, or explore older release tags.

Large diffs are hard to review because there's no natural order to follow. Guided diff tour groups related files into sections, orders them into a suggested reading path, and adds short explanations so you can review quicker and with more confidence.