Changelog

Merger Agent Stats gives engineering leaders a dashboard for the Merger agent's real-world track record. See evaluated PRs, the ready-to-merge rate, the share of PRs merged without human input, a breakdown of merge outcomes, and PR throughput over time, so trust in Merger is backed by numbers instead of anecdotes.

PR Inbox gives every developer a focused view of the work that needs them: a My Pull Requests tab with a Waiting for Review section for reviews requested of them, and a My Open PRs section for the Changes they own. Both sections share the same search, filters, pagination, and URL state as the main Changes list, so PR Inbox becomes a daily home for review work instead of another view to check separately.

Account settings now show current usage, let admins set spend limits, and surface Fixer costs alongside the rest of engineering work spend. Usage cards, limit forms, and status widgets make it clear how much has been spent, what is left, and what happens when a limit is reached, so usage-based pricing comes with the guardrails teams need to trust it.

Plans generated by Planner now go through a dedicated approval workflow before implementation starts. Reviewers can navigate plan versions, leave comments pinned to specific sections, view Mermaid diagrams inline, and approve or request changes on a plan the same way they would review code, giving teams a checkpoint before any code is generated.

Sessions now cover more of what Baz's agents do. Planner sessions show the full invocation, tool calls, final plan, and stats. Merger, Spec Reviewer, and full codebase security scans are now tracked as sessions too, so teams get a single, traceable view of what an agent did, which tools it used, and where the run ended, whether it ran against a PR or an entire repository.

Advanced Security can now scan an entire repository instead of only the files changed in a pull request. Start a full codebase scan from Advanced Security settings, track its progress through Sessions, and review persisted findings by severity once the scan completes. Full scans surface latent security risk that was already in the repository before the next PR ever touched it.

Admins can now scope Azure DevOps onboarding to the specific projects and repositories Baz should monitor, instead of connecting an entire ADO account at once. Connect Azure Repos, select projects and repos from a picker, and Baz keeps webhook coverage aligned with that selection as it changes. Existing Azure DevOps connections may need to reconnect to pick up project-scoped access.

Start a planning session directly from your coding agent with /baz:plan-with-baz to generate a repository-aware implementation plan before code is written. Plans capture the full implementation strategy, including affected repositories and files, change sequence, cross-repository coordination, open questions, and verification steps. Plans are versioned, support threaded discussions, review requests, approvals, and change requests, making them a collaborative artifact for engineering teams rather than a disposable prompt. The first release also introduces the Plans workspace, reviewer notifications, plan history, version-aware comments, markdown support, and a unified review experience that follows the plan from creation through implementation.

Teams can now see how long it takes for code to get picked up for review, and how long it takes to move from opened to merged. The graphs show median duration over time, per-period counts, weekly change, and Baz onboarding context. This makes review latency measurable. It gives teams a clean way to see where the review cycle slows down, and whether Baz is helping them move faster after rollout.

Merger Agent helps teams clear stale PR queues by analyzing open changes for merge readiness and returning a clear verdict: merge, reject, or escalate. It checks CI state, unresolved Baz findings, PR metadata, diff context, repo impact, git history, auto-approve rules, and org guidelines to identify changes that can safely move forward while escalating PRs with risk, failed or pending CI, open issues, or critical business impact for human review.

Added Session Logs to provide a structured execution record for Baz runs. Teams can now inspect how a run was triggered, where it executed, its current status, major completed stages, outcome, and cost. Session details are presented as a product-level timeline instead of raw logs, making reviewer runs, fixer runs, scheduled scans, benchmark runs, and future Codex or Claude Code workflows easier to understand and debug.

Baz can now update your AI coding guidelines, rule files, or skills when review feedback shows they are missing or out of date. For AI Coding Guidelines findings, Baz now gives developers two paths. They can ask Baz to fix the code with apply commit or apply pr, or they can ask Baz to update the underlying guidance by replying fix skills. The rule update is created as a separate PR to main, keeping guideline changes separate from the original code change.

Starting June 1, 2026, Baz is adding Engineering Work Credits on top of seat-based pricing. Seats continue to cover access and standard AI code review. Credits apply to high-model engineering work, including Fixer sessions, Advanced Security, Spec Reviewer, and AI SRE.

Advanced Security is a dedicated security review flow that uses cyber-capable models to identify exploitable security issues in pull requests. Advanced Security goes beyond static pattern matching by analyzing the change with surrounding code, related repositories, package context, PR metadata, and other relevant signals. It uses Baz's Agent Harness and Context Broker to gather the context needed to validate whether a finding is real, reachable, and worth fixing. Use it to catch deeper risks across authentication, authorization, input handling, data exposure, dependency usage, unsafe defaults, insecure flows, and cross-service assumptions before they merge. Advanced Security is available for configured repositories and posts findings directly as Baz review comments on the PR.

Baz reviewers can now use relevant context from across your organization's repos when evaluating a PR. This helps Baz catch higher-signal issues that depend on shared APIs, security policies, infrastructure conventions, and downstream service behavior outside the current repository. Findings that use this context are marked with a cross repo tag so reviewers can see when Baz connected evidence across repo boundaries.

We redesigned the Configurations surface to make control over Baz more explicit, structured, and scalable. Settings are now clearly grouped into Auto Audit (how Baz reviews) and Auto Resolve & Approve (how Baz acts), making behavior easier to reason about. You can configure trigger modes to run Baz automatically or on-demand via comments (Baz review, Baz summarize), control summary generation independently, and set an agents recalculation cycle with visibility into last and next runs. Review visibility and behavior are now fully customizable, including GitHub Checks presence, PR graph generation, comment limits, draft Change handling, CI failure comments, branch ignore patterns, and fix suggestion options. Content filtering allows you to exclude specific entities from processing, while new rule-based systems for auto-resolve and auto-approve let you define conditions, scope them by repository and users, optionally reply on resolution, and automatically approve or label Changes.

Fixer now supports one-click Auto Fix and a Fix All flow. Auto Fix will apply a suggested fix for a single finding, and Fix All applies all safe, validated fixes across the current file or change. These features are available from the Fixer agent and were added as part of the Fixer rollout and sprint work.

Fixer now also discovers and auto configures your sandbox with proper settings that match your repo. Fixer will make multiple attempts to setup using the best configuration possible and eventually select the best setup to ensure fixes don't fail your CI.

Baz now fetches context from your engineering knowledge stored in Notion and Fibery and uses it to contextualize and rationalize the changes it observed in your code. If Spec Reviewer is enabled, the agent will build a temporary index with all the referenced documents to build a single checklist to validate if code was implemented according to spec.

Fixer is a sandboxed auto-fix agent that automatically applies and validates code fixes inside an ephemeral, isolated runtime. It builds a repo-compatible environment, installs dependencies, runs strict format, lint and build validations, and only commits changes when validation succeeds. Fixer currently supports JavaScript, TypeScript and Python projects, with more runtimes coming soon.

Baz now treats a skill directory as the unit of authoring, so AI reviewers ingest the full set of rules and examples that apply to a module. The result is fewer missed rules and false negatives, more accurate and consistent automated reviews across teams and monorepos, and less manual work and onboarding friction for reviewers and contributors, all while working with existing skill-style authoring, so teams do not need to rewrite their guidance.

Spec Reviewer Design Spotlight turns screenshot evidence into a single, review-ready artifact for each requirement. It automatically identifies the precise UI region, renders a color-coded, numbered bounding box and a short label, and makes the annotated image the canonical visual evidence. The feature reduces reviewer friction, speeds remediation by making issues traceable to code or tests, and supplies structured, machine-friendly artifacts for downstream automation. It fails safely, preserves originals when localization is uncertain, and renders annotations in a resolution-aware way so they remain legible across devices.

Reported bugs reflect the issues that escaped code review and surfaced after deployment. They are a lagging signal, but an important one. Reported bugs show where review coverage was insufficient, where risk was underestimated, or where certain change patterns consistently slip through. By tracking reported bugs over time, teams can see not just how many issues reach production, but how that trend evolves as review practices change. Engineering Impact treats reported bugs not as a failure metric, but as feedback for improving how and where review effort is applied.

Spec Reviewer now connects directly to Azure DevOps, Monday, Shortcut, and YouTrack so tickets from those systems become first-class inputs to your Changes: when a Change references a ticket, Spec Reviewer pulls title, description and acceptance criteria to seed or update the spec, links criteria to the Topics it extracts, and runs automated checks that compare code and Usage signals against those requirements, flagging mismatches as checks on the Change. Compact ticket cards surface title, status, assignee, labels and key acceptance items and link back to the original ticket, and project, component, sprint and custom fields can be mapped to Baz Topics and Change metadata so planning context appears in place. Teams may enable optional write scopes so Spec Reviewer can post review conclusions or update ticket fields when criteria are met.

Engineering teams struggle to get a clear, consolidated view of review impact across repositories. Conventions end up living in pull request threads and notifications, which makes it hard to see trends, measure whether suggestions are actually adopted, or prove that work is reducing risk. Without a single, reliable signal, teams cannot prioritize the highest-impact problems or show engineering managers the real effect of review work. The Engineering Impact dashboard fixes that by turning Baz findings into a straightforward, team-friendly view: a timeline of detected issues with a headline trend and a measure of remediation, plus a curated list of the most significant, addressed bugs with the human context needed to act. The result is faster prioritization, better visibility into remediation adoption, and a simple way for engineering leaders to track the concrete impact of Baz on code quality.

Baz can now run in Private Mode. When enabled, code is analyzed locally and not stored by Baz. Designed for teams with strict security and compliance requirements. Docs: docs.baz.co/account/private-mode

Teams author rules as agent files or Cursor rules but needed a safe, deterministic way to enforce them during code review. Baz can now convert agent rule files into auditable review guidelines and run an AI reviewer that comments per those guidelines. The reviewer appears as "AI Coding Guidelines".

Spec Reviewer now embeds unmet requirement recommendations directly inline as review suggestions, pinned to the exact code locations where the requirement is violated or incomplete. For behavior-based evaluations, review comments now include screenshots captured from the agent's sandbox session, showing the UI state or flow observed during validation. This makes it clear why a requirement failed and removes the need to infer behavior from text alone.

We're releasing the Baz CLI, introducing an interactive code review loop directly in the terminal. Instead of generating static review reports, the Baz CLI guides reviewers through a change as a structured flow. It supports walkthroughs, requirements-driven reviews, interactive chat, and clear comment resolution, helping teams reach confident merge decisions earlier in the development cycle.

We have expanded Spec Reviewer to support backend behavior validation. Spec Reviewer for Backend brings product requirements, system constraints, and runtime behavior into one review step. Traditional code review cannot detect backend regressions that emerge only when real data, queries, and configuration interact. These issues do not appear in a diff and often surface only after deployment. Spec Reviewer for Backend closes this gap by validating backend behavior directly in a running environment.

• Reads backend requirements from Jira, technical specifications, and system constraints
• Runs the real environment to observe runtime behavior, including queries, pagination, filtering, data windows, and response shapes
• Compares actual behavior to the specification and surfaces any deviations directly in the pull request with evidence

Automated spec validation for pull requests that checks a change against its ticket and design. The agent consumes the PR diff, ticket context and attachments, design artifacts, and a validated preview environment, and it can use stored preview credentials to perform authenticated browser-based UI checks and visual comparisons. Output is a single structured report and CI check that lists unmet requirements first, to reduce noise.

Our agentic reviewers now run on GPT-5-Codex with Sonnet 4.5 reflections, bringing better understanding of large, multi-file changes.

We've fine-tuned reflection confidence thresholds and introduced trace propagation across review sessions, so AI reviewers can now connect context between findings and learn from previous evaluations.

User roles now automatically include a default contributor role on creation, ensuring consistent permission levels across orgs.

Admins can view and update these roles in the refreshed Account Settings screen, which also includes UI refinements and improved sidebar logic.

This guarantees a smoother onboarding flow for new users and fewer manual permission fixes.

You can now connect Baz directly to your GitLab repositories. Baz mirrors your projects, syncs members, and brings AI-powered reviews into your GitLab workflow. What's included:

• Connect GitLab Projects - Link Baz to your GitLab org or repos using a Personal Access Token (PAT).
• Mirror Code + MRs - Baz syncs repositories, groups, and members, showing merge requests inside Baz.
• Seamless Reviews - Comments and merge actions in Baz sync back to GitLab, authored under the PAT's identity.
• Flexible Management - Add or remove projects and update credentials at any time.

GitLab support is now available to all Baz users.

Connect Baz directly to your GitLab repositories. With this new integration, Baz mirrors your projects, syncs members, and brings AI-powered reviews into your GitLab workflow. What's included:

• Connect GitLab Projects - Link Baz to your GitLab org or repos using a Personal Access Token (PAT).
• Mirror Code + MRs - Baz syncs repositories, groups, and members, showing merge requests inside Baz.
• Seamless Reviews - Comments and merge actions in Baz sync back to GitLab, authored under the PAT's identity.
• Flexible Management - Add or remove projects and update credentials at any time.

GitLab support is currently in Beta. To request access, head to your Integrations settings in Baz.

We've completely redesigned Baz Evaluations to move from simple visibility into actionable insights that help you improve reviewer impact. Highlights:

• Impact Dashboard: A central hub connecting adoption rates, category-level performance, and reviewer comparisons with clear recommendations.
• Smarter Outcomes: Comments are now classified as Accepted, Rejected, or Unaddressed, giving you a true view of adoption.
• Sentiment Analysis: Capture developer reactions (including emoji responses), weighted by seniority.
• Reviewer Memory + Prompt Versioning: Reviewers now adapt to your standards automatically and retain full version history for safe iteration.

This upgrade makes Evaluations not just a reporting tool, but a feedback loop that continuously improves AI reviewers across your org.

Evaluations is a new insights feature now live under the Baz Agents tab. Evaluations give teams clear visibility into reviewer performance with detailed usage metrics and comment tracking. Designed for engineering managers and teams who want better visibility into reviewer interactions and the effectiveness of their AI-assisted code review processes.

• Get Reviewer Performance Reports: View a breakdown of all comments per reviewer agent and track which ones were addressed vs. unaddressed.
• Analyze by Reviewer Type: Filter performance for different reviewers in real time, making it easy to see where improvements or prompt tuning are needed.
• Optimize Agent Prompts: Gain data-driven insights to refine reviewer strategies and improve code review quality.

Baz now includes a Reviewer Memory system that learns from every interaction you have with a Baz Reviewer: each time you provide similar feedback more than once, a new memory is stored in that reviewer's dedicated memory bank, the reviewer's system prompt is automatically updated, and the collected memories surface on the Reviewer Card and in the drawer. All prompt changes driven by these memories are versioned, so you can always roll back to a previous prompt if you disagree with the community-contributed feedback.

Automatically spot flawed conditionals, infinite loops, precision loss, and other logic errors. Catch bugs before they ship.

Customers can now connect their Linear workspace to Baz. Ticket metadata and context from Linear are surfaced in every review, ensuring each change includes the original intent.

Baz's MCP server enables developers to receive structured, context-aware code reviews directly in their IDE by exposing Baz's review tools over the Model Context Protocol. Once connected, Baz analyzes file usage, change history, and test coverage in real time, returning high-signal insights inline, with no configuration, labels, or manual setup needed.

Scale your team's best review practices with AI agents trained on your actual PR history through Custom Reviewers with prompt optimization: auto-generated from past comments and suggestions in your codebase to reflect the real conventions your team already enforces, each reviewer including example comments and links to the PRs that inspired it, with editable prompts to fit your team's tone and style and consistent scaling across multiple repositories.

In this release, you get automatically generated reviewers from past PR discussions, linked examples and context for transparency, and editable prompts, plus the Prompt Playground, a review agent test lab that lets you edit Custom Reviewer prompts directly (or describe changes in plain language), test reviewer behavior on real change requests before rollout, and tune the tone, scope, and strictness of any review agent without pushing untested AI into production.

We've just launched the new Notifications Center, your streamlined hub for staying on top of code reviews. Whether it's a review invite, a new comment, or a resolved thread, you'll get notifications that are clear, relevant, and easy to act on. Each update includes helpful context like avatars, timestamps, comment previews, and direct links to the CR. You can filter unread items and mark things as read in one click, with no clutter, just what matters.

Baz now integrates Jira ticket data directly into review workflows, enhancing the context and quality of code reviews.

• Baz incorporates Jira fields like title, description, type, and comments into all change request context windows. This improves the model's understanding of the change's purpose and scope.
• Reviewers now see prompts shaped by Jira context, enabling evaluations that consider ticket goals, user stories, or bug descriptions.
• Better Summaries & Topics: The diff summary workflow merges Jira context with PR metadata and flow context to generate more relevant, concise summaries and topic breakdowns.

We've significantly upgraded Baz's support for C# projects:

• Baz now deterministically identifies C# elements like classes, functions, enums, and more, regardless of file name.
• Description graphs, summaries, and topics are now more accurate as they incorporate C# structure and context across files.
• Review comments are enriched with the surrounding encapsulating elements (e.g. class or method) to clarify each change's placement in your codebase.
• Support added for identifying C# packages (modules) and NuGet icons.

We've added two specialized reviewers to help teams ship safer and more consistent backend code.

• 🛡️ Basic Security Patterns: Automatically flags hardcoded secrets, SQL/command injection risks, and identifiable PII exposure in backend code. Designed to avoid noise by ignoring placeholders, test data, and non-sensitive logs, this reviewer focuses only on precise, actionable issues.
• 🌐 REST API Best Practices: Enforces clear and maintainable API design by checking for consistent route naming, correct HTTP method usage, logical parameterization, and versioning structure. Applies exclusively to backend API implementations to ensure standards without interfering with client-side code.

These reviewers are now available in your workflow configuration to improve code quality and reduce manual review effort.

We've added a new set of filters to help you stay focused during code review. Great for cutting down notification noise and zeroing in on what still needs your input.

• Open on me - See threads where you're mentioned or participated, but haven't had the last word.
• Show open - Filters out outdated and resolved threads.
• Show all - View all comments, as before.

We've added a new reviewer workflow metric to track the acceptance rate of suggestions per workflow type. This metric appears in the Preferences panel and aggregates the number of accepted suggestions and total suggestions generated by Baz Reviewer across all repositories connected to your workspace.

Module Memory is a new foundational upgrade to Baz's review infrastructure that enhances code change contextualization. This persistent memory layer captures the nuanced implementation details used by various modules in your codebase. Module Memory bridges the gap of developers relying on mental models of architecture, features, and implementation. Module Memory outputs can now be found in descriptions, reviews and the newly added Change Request Graph.

Developers can now pinpoint exact locations in the codebase, whether it's a tricky function or a nuanced review note, ensuring teammates land exactly where intended without sifting through files or scrolling endlessly. These links are dynamically generated and fully integrated into the interface, supporting smoother collaboration by bridging context gaps across distributed teams. Whether you're referencing a bug, suggesting an optimization, or highlighting a discussion point, this feature saves time and boosts clarity in your workflow.

AI coding tools have a tendency to repeat themselves, especially when they use your codebase for reference. As a result, you may start to see similar implementations of the same business logic across multiple areas of your codebase. This is where our new Reviewer agentic workflow kicks in. Reviewer continuously analyzes the diffs in your PRs and flags possible instances of duplicated implementations. This workflow is powered by a state-of-the-art RAG data pipeline that leverages Voyage AI's code embeddings models.

Baz Reviewer now offers an interactive layer of functionality: you can directly reply to its automated comments within GitHub pull requests, transforming one-way feedback into a dynamic conversation. Whether you're seeking clarification on a suggestion, requesting a concrete code example to better understand its reasoning, or challenging a recommendation that doesn't fit your use case, this feature empowers you to engage with the AI in real time. Responses are generated with context from the original comment and your codebase, ensuring relevance and precision. This enhancement fosters a more collaborative and iterative review process, reducing friction and accelerating resolution of complex feedback loops directly in your GitHub environment.

We're introducing configurable Reviewer workflows in Baz. Organization admins can now enable or disable specific workflows to better tailor Baz to their team's needs. Simply navigate to Baz Settings, then Preferences, to adjust which workflows are active. This gives teams more control over their review process, reducing noise and focusing on the insights that matter most. Looking ahead, we're working on repository-specific path exclusions, allowing teams to exclude certain files or directories from analysis to further refine results. Note that Breaking Changes and CI Errors workflows remain always enabled, ensuring critical issues are never missed.

The new Baz Reviewer introduces an AI-powered system for automated code review, focusing on detecting critical issues like type inconsistencies, syntax errors, and logical changes across functions and classes. It ensures updates are consistently propagated, reducing the risk of integration failures while improving overall reliability.

Beyond error detection, Reviewer provides actionable optimization suggestions, such as refining code structure and adopting modern language idioms. It enforces best practices, ensuring code is efficient, maintainable, and aligned with industry standards.

Reviewer also enhances testing by identifying gaps in coverage and ensuring functional changes are properly validated. With iterative and context-aware reviews, Reviewer streamlines development, minimizes technical debt, and accelerates high-quality code delivery.