GPT-6 Astra now orchestrates Mergestorm Tempest PR reviews

GPT-6 Astra is live as an orchestrator on Mergestorm Tempest. You can select Astra in Agents, point Tempest at a pull request, and run an on-demand deep review for 30 credits at launch.

Tempest is where we send the questions that need more than another fast pass over a diff. It looks across files for contracts that no longer line up, configuration that drifted from the code it controls, and architecture changes whose failure only becomes visible when the pieces are read together. Astra now sits at the center of that pass.

Tempest goes deep on demand

Mergestorm has three review loops with different jobs.

Vortex reviews every push. Its fast fleet gives changed code continuous coverage, with specialist lanes hired where the diff needs them. Cyclone can patch clear findings. Tempest is the deliberate deep pass you start when a pull request deserves a wider read.

That separation matters. We did not turn every review into a large-model run, and we did not replace the fast fleet. Vortex V2 remains the sibling system built for frequent, convergent review. Tempest is for the point where an operator wants to step back and ask whether the pull request works as a whole.

In Agents, choose the orchestrator you want and run Tempest on the PR. The orchestrator directs the review across the relevant context and brings the result together. The GitHub comment names the orchestrator that actually ran, so the record on the pull request matches the choice you made.

What Astra is looking for

A deep review earns its cost when the important defect is not isolated to one changed line. A handler may accept a new field while a serializer silently drops it. An environment variable may be renamed in application code but not in deployment configuration. A shared type may still compile while two consumers have started assigning it different meanings.

Those are cross-file problems. They require following contracts and assumptions through the repository, not just scoring individual hunks.

Astra gives Tempest another orchestrator for that work. It can coordinate a broad read of the pull request, inspect the relationships that carry risk, and return findings grounded in the code. The useful question is not whether one model can produce a longer review. It is whether the review finds the boundary where the implementation and the system disagree.

There is an important detail behind the word "orchestrator." Selecting Astra does not mean every stage or specialist lane runs on Astra. Scope-specific work can still run on a cheaper model such as DeepSeek when that is the right tool. Astra coordinates the deep review. Mergestorm can still spend smaller models on narrower jobs instead of applying the most expensive model everywhere.

Launch price: 30 credits

An Astra Tempest review costs 30 credits at launch. Its later list price is 50 credits, in the same class as Fable, but that increase is not live today.

Free accounts can use Tempest. The free pool is 30 credits per month, so one Astra review fits in a free month's allocation. That is a review budget, not an unlimited plan. Paid accounts draw from the monthly credit pile associated with their plan. We explain the current piles and why they changed in our review credits post.

Tempest also keeps a range of orchestrators for different budgets and jobs:

The point of that catalog is choice. A routine deep pass may call for DeepSeek. Another PR may justify Astra, Sol, or Fable. Cost is visible before the run, and the resulting GitHub comment records which orchestrator performed it.

When to pick Astra

Use Astra when the uncertainty lives between files.

That can be a large refactor with old and new paths running side by side, an infrastructure change coupled to application configuration, or a feature whose public API, persistence layer, and UI must agree. It can also be a smaller pull request with an unusually expensive failure mode, where one deliberate deep pass is cheaper than discovering the broken contract after merge.

Tempest does not need to replace your normal review rhythm. Let Vortex cover each push and converge with the pull request. Use Cyclone when a grounded finding has a clear patch. Then call Tempest when you need the repository-level question answered before the change moves forward.

Pick Astra. Go deeper.

Open Agents, select GPT-6 Astra, and run Tempest on the pull request that needs the wider read.