See what's happening in your repository at a glance.
AI agents are changing everything.
Code is written in "factories", partially by humans, partially automated. The pace of change is relentless, and we've got to keep up with it.
We're releasing a new repository overview: our first iteration into making Sourcegraph a platform that empowers you to stay plugged in to the changes across your codebase.
Give it a whirl, and let us know what you think.
Deep Search can now cross-reference and aggregate larger result sets, and write its findings to downloadable CSV, JSON, and SVG files.
Ask Deep Search for codebase wide inventories, audits, or migration reports that would be too large to present cleanly in an answer. Deep Search summarizes the key findings in chat and attaches the full structured output for further analysis or tracking.
Deep Search can also read generated CSV and JSON files in follow-up investigations. This allows it to filter, join, or enrich previously collected results without repeating the original codebase searches.
Generated files appear directly in the Deep Search answer, where you can preview or download them.
Try prompts like:
profile-service through getProfileFeatures. Create a CSV with the consuming service, repository, source URL, field mask, feature group, and request type.old/package but does not yet import new/package.Learn more in the Deep Search documentation.
Give Deep Search durable instructions and task-specific skills tailored to your organization.
Deep Search can now follow your organization’s conventions on every run. Add durable AGENTS.md-style guidance for standards and preferences, plus task-specific skills that help the agent handle specialized workflows with the right instructions.
Configure customization globally, for an organization, or in your personal settings. Guidance and skills can be entered directly or loaded from an explicitly configured Sourcegraph repository location, with an optional path and branch or revision. Repository sources can contain AGENTS.md and skills at skills/<name>/SKILL.md or .agents/skills/<name>/SKILL.md; they are not automatically discovered from every repository involved in a search.
Guidance is included in agent runs, while skills are advertised by name and description and loaded only when relevant, preserving context for the task. Settings are applied from global to organization to personal scope, and later skills override earlier skills with the same name, case-insensitively.
Agentic Batch Changes supports the same customization model for its coordinating agent. Guidance and skills do not flow into delegated coding agents such as Claude Code or Codex.
Run exhaustive, codebase-wide searches from your own integrations with Search Jobs in the Sourcegraph API.
Search Jobs have landed in the Sourcegraph API. A Search Job runs your query across every repository, branch, and revision in your codebase. Use it for audits, security reviews, and large scale migrations.
Search Jobs are served from the versioned /api/ endpoints and authenticated with access tokens scoped to externalapi:read and externalapi:write. Jobs run asynchronously, and results are available as JSONL once a job completes. Each token only sees and manages its own jobs.
You can explore the available operations and download the OpenAPI schema from the searchjobs service section of the interactive API reference at /api-reference (e.g. https://sourcegraph.example.com/api-reference).
The Search Jobs operations in the Sourcegraph GraphQL API, now the debug API, remain available and unchanged. For long-term support and a stable integration experience, we recommend building on the new Sourcegraph API instead.
If you have any feedback, please reach out to support@sourcegraph.com.
Our new MCP tool finds relevant code faster with lower token usage.
Code Finder, a fast and efficient agent optimized specifically for code search, is now available in Beta through the Sourcegraph MCP server.
Code Finder is a tool that helps coding agents quickly locate relevant implementations, configuration, tests, or documentation within a single repository. It returns a concise summary with links and line ranges, giving your agent focused results it can immediately use.
To show the strengths of using an optimized search agent, we assembled a benchmark suite with a variety of search tasks across several repositories, from small to large. The results show clear improvements in speed and cost. In particular, Code Finder is more than twice as fast compared to a coding agent searching with local tools and up to 40% cheaper than a coding agent using our MCP toolset, achieving comparable result quality.
| Variant | Duration | Cost |
|---|---|---|
| Code Finder | 1.00× | 1.00× |
| MCP tools | 1.14× | 1.63× |
| Local search | 2.19× | 1.32× |
Code Finder already powers file discovery within Deep Search, letting Deep Search spend less of its token budget locating files and more on reasoning about the code and producing a thorough final response.
Code Finder is free while in Beta. The feature will become billable with credits once it becomes generally available. For more information on our MCP server and Code Finder usage, see the documentation.
See exactly what changed, at any scale.
Following its Beta period, the redesigned compare page is now generally available. It's a ground-up rebuild of how you view changes between two revisions — fast, complete, and built to stay useful even on very large diffs.
What you get:
There are two ways in: open any repository and click Compare, or go to a repository's Branches tab and compare branches directly.
This is the diff view we're adopting across Sourcegraph: it already powers the commit page and Agentic Batch Changes, and it'll soon roll out to more places where you review code.
Making it easier for humans to follow the progress of coding agents.
In the Agentic Batch Changes Beta, steps utilizing inner loop coding agents like Claude Code and Codex now stream thinking and progress in real time.
Early on, when building out Agentic Batch Changes, we decided to give the outer-loop agent that orchestrates your Batch Change effectively two tools to make code changes:
At first, we weren't sure if our coordination agent would see the benefit of using deterministic scripts over the simple option of using a coding agent. Coding agents can do everything now!
After letting the outer-loop agent rip through some of the most complicated problems we could come up with on the spot, though, we realized that the agent was making much smarter decisions than we could've expected, using both tools:
It would write scripts for the parts that were easy to automate because it was much easier for the outer-loop agent to predict the exact outcome. That gave the outer-loop agent more control and made things much more effective. Instead of having 500 inner-loop coding agents each solve the same dependency-installation problem in 500 different repositories, the outer-loop agent would use the knowledge it gained from studying the problem and the codebase to automate as much as possible with a script.
Then, if, during execution, it learned that there were edge cases that were hard or even impossible to script against, it would start involving coding agents. And when one coding agent goes off track, the orchestration agent would be able to steer it and apply that steering and learning to the coding agents across all other workspaces as well.
With those patterns, it was often able to make 500 changes with one outer-loop agent thread and zero inner-loop coding agents involved. And then, when they would truly make a difference, it could pull them in.
When it deployed inner-loop coding agents, following along was difficult in our first beta. The agent CLI tools we support emit beautifully structured JSON, but that output is not easy for humans to grasp.
So, for this week's update, we have:
The Agentic Batch Changes Beta is now available upon request: talk to your account team or schedule a demo to get started.
Deep Search can cross-reference and aggregate large result sets for codebase-wide inventories, audits, migration planning, and other structured investigations
Migrated Deep Search guidance setting from experimental features to stable configuration path.
Recent Deep Search contexts are now reconciled against live data when the picker opens, preventing stale or deleted contexts from appearing in the recents list.
Added source.read_file module to the evaluator, enabling file search, reading exact repo/path/revision matches, and source context inspection within a single evaluator execution.
Added an alert for SCIP uploads that exceed the configured processing time limit.
Added a Grafana dashboard for monitoring Batch Changes server-side background workers, queue stores, and stalled-job resetters.
Added indicator highlighting suggested Batch Changes code host credential method
Batch Changes now uses required GitHub checks when computing changeset CI status, preventing optional check failures from incorrectly marking changesets as failed.
The coordinator agent in Agentic Batch Changes can now cancel workspace execution.
The Agentic Batch Changes coordinator agent can now read files in repositories directly, cutting down on planning time.
Bundled coding-agent CLIs (Codex and Claude Code) with checksum-pinned versions at build time, eliminating runtime downloads from public installers. Supported across v1 and v2 executor runtimes with executor version gating.
Migrated Agentic Batch Change guidance setting from experimental features to stable configuration path.
The Duration column in the batch-change-agents workspace list now updates live for in-flight workspaces (queued/processing/canceling), showing elapsed time that ticks up in real time.
Agentic Batch Changes' coordinator agent now uses Claude Opus 5.
Agentic batch change agents now wake automatically when a changeset's CI passes, even when no onCIFailure hook fired. The resume is labelled "Resumed automatically: changeset CI passed" in the thread timeline.
Read files generated from Deep Search in Agentic Batch Changes
The managed coding-agent base image now includes Python 3, ripgrep, and jq by default, avoiding the need to install these tools at execution time.
Agentic Batch Changes links that you can't access now show a placeholder indicating how to request access
Polished the batch change agent detail page UI: improved table layout consistency, redesigned workspace column display, updated status badges, added missing empty states, and unified toolbar patterns across plan and resolve views.
Batch Changes secrets pages now cross-link between per-user and global secrets. When Agentic Batch Changes is enabled, both pages explain how secret descriptions guide Agentic Batch Changes.
Site admins can now audit Agentic Batch Changes activity: tool approval decisions, agent deletion, and read token rotation are recorded in the audit log.
Batch change workspaces now display their last-known results (diff, steps, changeset specs, and timing) after re-preparing a spec but before re-executing, with a stale banner indicating the workspace needs to be re-run.
Improved the performance and accuracy of agentic batch changes when writing certain batch spec edge cases.
Added UI renderers for inner-loop coding agents to display execution progress and events in a human-readable format instead of raw JSON.
Logs are retained in the UI for cached batch change executions, improving visibility when the batch change agent automatically updates specs.
Improved step caching for coding agent steps to be independent of agent version and desugaring logic changes, as well as buildImage steps. Some cache entries will need to be recomputed when using mount on a step.
changesetTemplate template variables are now rendered in transformChanges group branch names.
Recommended a scoped CI-log token in the batch change agent prompt's CI-failure-hook guidance.
Agent guidance sources are now clickable links that open their configured repository paths.
Deep Search and Agentic Batch Changes now render timestamps in their transcripts.
Added a "Generating a ConnectRPC client" section to the API reference intro docs, showing how to download the proto schema (.binpb) — all services by default, or a comma-separated subset via ?service= — and generate typed stubs with buf. This renders as an always-visible section in the API reference page (built on the proto download added in #14309), so customers can self-serve ConnectRPC/protobuf client generation. The OpenAPI golden snapshot is regenerated to match.
Added proto schema download to the API reference page. The Download button now offers both OpenAPI schema (.yaml) and proto schema (.binpb) options, enabling ConnectRPC/protobuf consumers to generate typed clients directly from the external API definition (e.g. via buf generate sourcegraph-api.binpb).
Service account access token creation page now displays an informational alert recommending M2M OAuth credentials as an alternative that eliminates manual token rotation.
Improved Finder prompt to provide more complete answers with better file references.
Added a labeled "Viewed" chip for marking files as viewed in the compare/diff view, replacing the previous unlabeled circle control.
The new compare page is now generally available.
Improved Perforce compatibility for the compare page.
Added a commit/history panel to the compare page diff viewer, allowing users to view changes for individual commits within the selected range.
Added support for Claude Opus 5
MCP server now supports protocol version 2026-07-28 while maintaining backward compatibility with older clients.
Introduced a dedicated site config setting for blocking specific MCP tools. The setting is "mcp.tools.disabled": ["tool_name"] - you supply an array of tool names that should be disabled across all MCP endpoints.
Enabled Code Finder MCP tool by default for customers who have signed AI terms.
Added the mcp.deepSearchTool.enabled site configuration option for controlling the Deep Search tool on general-purpose MCP endpoints. The environment variable SRC_MCP_DISABLE_DEEPSEARCH_TOOL is deprecated and will be removed in a future release.
Microsoft Teams is now supported as a notification channel for observability alerts.
Trace spans for GitHub and GitLab requests now include the upstream request ID (github.request_id from X-GitHub-Request-Id for GitHub, gitlab.request_id from X-Request-Id for GitLab) to help correlate requests with code host support teams.
Audit logs now show whether the user is a service account or not.
Hide Ownership tabs for repositories without a CODEOWNERS file.
Data tables now show an edge shadow on the sticky row-number/checkbox column while horizontally scrolled.
Fixed spacing between Deep Search prompt editor and error/quota banners
Fixed layout shift in the search context picker
Fixed ambiguous file mentions by using ':' delimiter between repository name and file path instead of '/', preventing the model from misinterpreting multi-segment repository names as a single repository.
MCP search tools are now correctly attributed to MCP rather than Deep Search in backend telemetry.
Fixed filtering references by location in the code navigation Explore panel.
Fixed a panic in precise code navigation that could occur when a repository had multiple SCIP uploads indexing the same file path.
Fixed a deadlock between executor heartbeats and the unknown-commit janitor that could cause code intelligence indexing jobs to stall.
Fixed go-to-definition (cmd/ctrl+click) in the file view not scrolling to the definition when it was off-screen.
Fixed an issue where code navigation request cancellations would incorrectly be counted as errors in SLO statistics.
Fixed pagination of code intelligence data
Janitor sub-jobs for cleaning up deleted commits (e.g. on force-pushes) have been tuned to reduce unnecessary DB pressure.
Added per-janitor-job interval control knobs CODEINTEL_UPLOADS_CLEANUP_INTERVAL_<job>, which are useful when debugging Sourcegraph in production environments.
Optimized a precise codeintel commitgraph reconciliation query, reducing commitgraph lag up to ~80% on very large repositories.
Fixed phantom next page appearing in batch spec library pagination when the number of records is an exact multiple of the page size.
Fixed a swallowed marshalling error in batch changes that could lead to downstream inconsistencies
Fixed an issue where the Batch Changes reconciler would silently mark a changeset as undrafted even when the operation failed to reach the code host due to remote repository resolution errors. The reconciler now properly surfaces these errors and schedules retries, preventing changesets from being incorrectly left in draft state.
Clarified when SSH push key needs to be configured for Batch Changes
Surface OAuth errors when connecting code hosts for Batch Changes
Batch Changes now correctly shows UNKNOWN CI status instead of transient pending/passed states when required check classification is supported but the initial sync has not yet completed.
Fixed review state for merged and closed changesets that could incorrectly show as pending approval
Redacted secret values in executor logs are now annotated as {{ REDACTED secrets.NAME }} instead of ${{ secrets.NAME }} to make clear the value was removed.
Fail validation with a clear error when a hook step's files target path contains invalid characters.
Ignores obsolete cancelled GitHub Actions check suites when a newer run of the same workflow succeeds.
Restored schema-aware suggestions in the batch spec editor.
Batch spec step files now require absolute paths and cannot be added to /work directory
Fixed scheduled changesets becoming permanently stuck after a batch change or its namespace was deleted.
Changesets whose reconciler hit a transient error (e.g. a code-host timeout) are no longer reported to the agent as a terminal failure while they are still being retried automatically.
Changesets created by Agentic Batch Changes no longer include forced line breaks in their descriptions.
Fixed an issue where workspaces would disappear from the UI when switching between batch specs until the page was refreshed.
Fixed thread panel on the /new page to always start expanded instead of using the previous session's collapsed state.
Batch Change Agent messages now stop after 24 hours if they do not finish, surfacing a recoverable failure instead of remaining active indefinitely.
Agentic Batch Changes can now present multiple-choice questions, each allowing free-form input.
The publish_workspaces tool now distinguishes between workspaces that completed with no changes and workspaces that are not ready for publishing. Workspaces that finished execution (COMPLETED/SKIPPED) but produced an empty diff are reported separately as workspace_ids_no_outcome. The tool now errors immediately when a selected workspace has not successfully finished (still queued/processing, or failed/canceled), preventing silent failures and allowing the agent to wait or retry appropriately.
Avoided redundant SSE streams while BCA agents run, reducing browser connection usage
Fixed blank multi-file diffs in workspace diff view
Transcript step timestamps now adapt to available space: wider layouts show timestamps beside timeline dots on hover, while narrower layouts display them in tooltips.
Fixed an issue where paused agent streams could exhaust browser connection limits and stall unrelated requests.
Validate buildImage.baseImage against container registry allow/denylist to prevent unauthorized base images in Agentic Batch Changes.
Fixed approval action buttons overflowing in narrow layouts.
Thread rename UI now uses standard UI components and improved navigation icon
Message stream retries now use exponential backoff (2/4/8/16/32 seconds) instead of retrying every 2ms, reducing unnecessary reconnect bursts.
Fixed Code Insights charts failing to load when data points render at nearly identical positions, preserving access to insight edit and delete controls.
Fixed Code Insights dashboards failing to load when an errored insight contains chart data that cannot be rendered, allowing the affected insight to be edited or deleted.
Fixed Cody Web chat failures when browsers advertise zstd compression support.
The markChangedFilesAsViewed mutation now rejects requests that contain files with both srcOID and dstOID set to nil.
batch-exec helper binary from published artifacts, ensuring version compatibilitybatch-exec binary and provides installation instructionsExecutor jobs, including Kubernetes-backed jobs, now have a default maximum runtime of 60 minutes.
batch-exec helper binary from published artifacts, ensuring version compatibility.batch-exec binary and provides installation instructions.GitLab API error messages are now logged during permissions sync to aid troubleshooting.
Fixed an issue where having a large number of code host connections could cause other pages in Sourcegraph to load extremely slowly.
Clarified that changing the gitMaxConcurrentClones site configuration setting requires restarting the worker service or pod to take effect.
Added upper limit to number of log events that can be sent per batch
Fixed smart hover summary bugs that could result in additional codeintel resource consumption.
Fixed page layout title alignment issue
Browsers no longer auto-translate the Sourcegraph web app, which could corrupt code and interface text.
Fixed an edge case where entitlement release could fail if the context was cancelled during the operation.
Fixed notification badge spacing for double-digit counts in the sidebar.
Hybrid search now has a separately configurable changed-path length limit via SEARCHER_HYBRID_MAX_TOTAL_PATHS_LENGTH, with a default of 300 KB.
Fixed an issue where users who signed in through an external provider could not create a password.