Which model, and what happens when it is unavailable
Anthropic’s Claude is the shipped provider, and it is used at two different strengths on purpose: a cheap fast model classifies what a merge is and how much it matters, and a stronger one writes the communication that people will actually read. Classification runs on every watched merge; writing runs only on the ones worth writing about.
Underneath sits a declared fallback ladder, and the first configured rung wins. If no Anthropic key is set, an OpenAI-compatible provider can be wired in its place for the analysis, and for the drafting too when that deployment configures a content model. If neither is present, the analysis is heuristic and the drafts are deterministic — plainer, and still produced.
You will not find a model version named anywhere on this site, and that is deliberate. The versions are launch choices overridable by an environment variable, not constants; printing one here would make this page a stale claim within a quarter. What you can hold us to is the behaviour: at the first model error the pipeline degrades to the deterministic path and says so, rather than inventing a result to fill the gap.
When a claim cannot be verified
A grounding validator reads what the model wrote against the evidence it was given. This is what happens when a sentence fails it.
- One retry, with the specific problem fed back
Not a blind second attempt: the structured issues go back into the request, so the model is told which sentence failed and why before it tries again.
- If it fails again, the draft is kept
Mergelay does not substitute a bland template and pretend nothing happened. It keeps the model’s draft, sanitises it, and marks the exact sentences that could not be supported.
- The marker follows the draft everywhere
It shows on the workspace draft, in Approvals, on the one-click review page and through the coding-agent tools. There is no surface where the claim appears without the flag beside it.
- A person may still approve it
The marker is disclosure, not a block. You can read the flagged sentence, decide it is right, and approve deliberately. Approving does not erase the marker from the record.
- What the marker does forbid is a machine deciding
Supervised autonomy, autopilot and scheduling all refuse to send a draft carrying unverified claims, whatever the workspace settings say.
- It retires one claim at a time
The marker survives cosmetic edits elsewhere in the text and clears only as each flagged sentence is actually rewritten, so a comma cannot dismiss it.
How this was produced
Two different things happen to every Update: it is classified, and then it is written. Mergelay states the provenance of each separately and never lets one sentence stand in for both — an Update and an approval each carry one line for the classification and one for the drafts.
That matters most on the one-click review link, where a reviewer decides from their inbox without opening the workspace. That page carries the drafts line, so the difference between a deterministic starter draft, model-written copy and a paragraph a colleague rewrote by hand is visible at the moment of the decision rather than afterwards.
The uncomfortable cases are stated too. When a model was called and did not complete, the page says exactly that, rather than claiming no model was used. An analysis stored before Mergelay recorded which classifier produced it says nothing at all rather than guessing from today’s configuration. And in Activity, a degradation that still produced a usable result reads Fallback, not Error — the result exists, and the technical reason stays in the entry’s details where an operator can find it.
The limits worth knowing
Four constraints that a demo would not show you.
- Model-driven regenerations are capped per Update — five by default. Manual editing is never counted against that cap and stays available after it is reached, so you are never locked out of your own wording.
- Repository content is treated as untrusted evidence, never as an instruction. A prompt written into a README, a commit message or a pull request description is data to be described, not a command to be followed.
- Mergelay holds the model key. There is no bring-your-own-key tier in the standard hosted path, which also means model cost is our problem rather than a surprise line on your bill.
- Every analysis and generation call is logged with its provider, role, prompt version, model, token counts and duration — and never the key. That log is an internal cost and abuse signal, not a customer-facing allowance.
Questions
Is my code sent to a model provider?
Bounded, sanitised parts of it. The title, the description, the labels and the commits go. Patches go too under the default mode, which keeps credential files, key files, infrastructure state, dumps, backups and fixtures to metadata alone and drops any patch where redaction spots a key, an email address or a high-entropy value. A deployment can be set to send no patch body at all. What never happens, under any setting, is a full clone leaving the platform.
Can it publish on its own?
Not without a person. The shipped default is automatic analysis and automatic drafting with human approval before publication, and that approval binds all the enabled drafts and destinations of one exact revision — edit the content and you get a new revision that needs a fresh decision. A draft carrying unverified claims cannot be published unattended at all, whatever else is switched on.
What if the model invents a feature that does not exist?
That is the failure the grounding validator is built to catch. It retries once with the specific issues fed back, and if the sentence still cannot be grounded it is marked wherever the draft appears. You see the invented claim and the flag together, before you approve, rather than reading about it from a customer.
Do I have to use an AI coding agent for this to be useful?
No. Agents raise how often the problem shows up — more merges, less context per merge — but nothing here assumes one. A team merging by hand hits the same gap between shipping and explaining.
Where to read next
- GitHub release notes generatorThe full pipeline, and the table of what leaves your repository.
- Customer-facing release notesWhat it takes to put a draft in a customer’s inbox.
- SecurityHow to report a vulnerability, and what we owe you back.
- PricingBoth plans, and what the plan actually changes.
Read a draft before you decide whether to trust one.
Your first Update is free. No card required.