Qodo 为 AI 生成代码新增策略驱动审批与会话审查
Qodo Adds Policy-Driven Approvals and Session Reviews for AI-Generated Code
Qodo 新增 Auto Approval Workflows 和 Agent Session Review,分别按团队策略自动处理符合条件的 PR,并集中评估编码智能体跨仓库产生的变更。前者于 11月18日开放 beta,可在工作区、Git 组织或单个仓库级配置规则与保护条件;后者以 research preview 推出,可关联实现决策、计划偏差、合并顺序及代码风险。Qodo 的审批仅计作一次批准,既有必需审查者和代码所有者要求仍然适用。
正文
Qodo is extending AI-assisted software development beyond code review with two capabilities aimed at deciding which changes can advance automatically and which still need an engineer’s attention. Announced on November 18, Auto Approval Workflows is available in beta, while Agent Session Review is being offered as a research preview.
The release addresses two different bottlenecks. Routine pull requests can wait for a human approval even after automated review finishes. Larger features created by coding agents can create the opposite problem: several apparently reasonable changes spread across repositories may be difficult to understand as one coherent implementation. Qodo’s response combines configurable approval policies with a consolidated assessment of an agent’s work.
A pull request, or PR, proposes a change to a codebase. Reviewing it identifies potential problems; approving it records a decision that it is acceptable to move forward under the team’s process. Automating the first step does not necessarily resolve the queue at the second.
According to Qodo’s announcement, Auto Approval Workflows lets engineering teams set rules for qualifying changes. When Qodo finishes reviewing a PR, a matching workflow can approve it or apply a “Qodo approved” label. Teams can start with templates for low-risk work and apply workflows at the workspace, Git organization or individual repository level.
The distinction between a label and an approval is important. A label provides a signal without itself authorizing a merge. In its product announcement, Qodo says its bot’s approval counts as one approval, while existing required reviewers and code-owner requirements still apply. The new feature should therefore not be confused with unrestricted automatic deployment.
This changes where the repeatable decision is made. Instead of requiring someone to rediscover the same acceptance criteria for every routine update, a team expresses those criteria in a workflow. The value depends on whether the policy captures the conditions that actually make a change acceptable, not simply on how many approvals the system can issue.
Qodo’s guards let teams identify repositories and authors that should never receive automatic approval. The company says each new commit triggers reassessment, with approval withdrawn if the updated pull request no longer qualifies. Its PR Triage view records approval status and the workflow or guard behind a decision.
That reassessment matters because approval is a judgment about a particular version of the code. A small fix can become a larger change when additional commits arrive. An approval that remains attached without considering those changes would say less about the final implementation than it appeared to.
These controls also make the boundary of automation visible. A team can delegate a narrow class of routine decisions while retaining direct review elsewhere. That is a more specific proposition than asking engineers to trust all AI-generated code, and it leaves ownership of the policy with the organization deploying the tool.
Reviewing an Agent’s Work as a Complete Feature
Agent Session Review tackles work that cannot be understood adequately from a single pull request. Running through Qodo’s Agentic Toolbox, it gathers PRs produced during a coding session, including work spanning multiple repositories, into one assessment. Qodo says the result connects implementation decisions, deviations from the original plan, merge order and risks tied to evidence in the codebase.
For an engineer, the useful unit of review may be the behavior of a complete feature rather than the boundaries of its repositories. A backend change and a client update can each appear reasonable in isolation while depending on different assumptions about how they will interact. A session-level assessment is intended to make those relationships easier to inspect.
The research-preview designation remains significant. Qodo is describing a new way to organize evidence for review, not demonstrating that every unintended behavior will be detected. Engineers still need to decide whether the resulting implementation matches the intended requirements.
The two additions point toward a practical division of labor: policies handle changes whose acceptance conditions can be stated clearly, while people examine the decisions and interactions that require broader judgment. The announcement does not establish independent evidence of faster delivery or fewer defects, and those outcomes should not be assumed from automated approval alone.
For development teams, the larger shift is from measuring how quickly an agent produces code to understanding how that code earns permission to proceed. Qodo’s beta workflows and session-review preview put that decision process closer to the center of the AI development toolchain.
AI 热点