_Issue created by GitLab Duo, rewritten on Sep 8, 2026 using :robot: to reflect what was actually found._ ## Problem to solve A customer on GitLab Self-Managed 18.11 had Code Review Flow failing on large merge requests, with the AI Gateway returning `prompt is too long: 204579 tokens > 200000 maximum`. Their reviews were running on Claude Sonnet 4.5, which they had never selected — it was the default. Claude Sonnet 5, which the docs listed as the default for Code Review Flow, was not available to them, and nothing explained why. The fix, switching to Claude Sonnet 4.6, was available in their dropdown, but nothing in the docs pointed to it. Diagnosing this took roughly a week across several GitLab team members. See the threads below for extra details. ## Findings Three elements to the issue here: 1. What setting to use to select a model for Code Review Flow - The model setting for Code Review Flow (`review_merge_request_dap`, listed as **Agentic Code Review** in the UI) was added in 19.1 ([MR 236876](https://gitlab.com/gitlab-org/gitlab/-/merge_requests/236876)). - On 19.0 and earlier, the classic non-agentic GitLab Duo Code Review and Code Review Flow share one setting, listed as **Code Review** under **GitLab Duo for merge requests** (`review_merge_request`). - Code Review Flow's default model is locked to the non-agentic code review's default, currently Claude Sonnet 4.5 Vertex. - Its selectable models are the non-agentic list, so Claude Sonnet 5, GPT-5.2, and GPT-5.3 Codex cannot be reached. - Neither limitation is clearly documented. - Only GitLab Self-Managed is affected. GitLab.com is always on the latest version. Affected readers use the instance model selection page. We don't currently point users to that page from the model selection or Code Review Flow page, so it's easy to assume those pages contain the latest guidance for SM. 1. What models are available for use - The history items in the docs list milestones for when models are added or removed. But models and defaults come from the AI Gateway and reach every instance immediately. The GitLab version you are on does not affect availability. "Added in GitLab 19.3" reads as an upgrade requirement when the change reached every instance the day it shipped. - The settings in the UI **are** gated by version, however (see the Code Review Flow example in item 1). So we have to be careful about this specific case. 1. Which model selector is for classic GitLab Duo Code Review - The selector is labelled **Code Review**, which is easy to mistake for Code Review Flow. ## Proposal 1. **Clarify the behaviour.** Document that Code Review Flow uses the classic setting on 19.0 and earlier, on the pages self-managed administrators use, and state that model changes come from the AI Gateway regardless of GitLab version. Avoid a version matrix because it implies models are version-gated. 1. **Rename the UI setting.** Change **Code Review** to **Non-Agentic Code Review** so it pairs with **Agentic Code Review** and the two dropdowns disambiguate themselves. Docs describing 19.0 and earlier keep **Code Review**. 1. **Use dates for model changes.** Switch history entries for gateway-delivered model changes from milestones to dates. Entries about settings keep their milestones, because a GitLab version genuinely gates those. ## Out of scope Backporting. No single target reaches the affected population: the affected range is roughly 18.7 to 19.0, which sits outside the patch window, and archived and instance-bundled model tables are frozen while the AI Gateway keeps adding models, so a correct note would sit beside a stale list. _Issue created by GitLab Duo, rewritten on Sep 8, 2026 using :robot: to reflect what was actually found._ ## Problem to solve A customer on GitLab Self-Managed 18.11 had Code Review Flow failing on large merge requests, with the AI Gateway returning `prompt is too long: 204579 tokens > 200000 maximum`. Their reviews were running on Claude Sonnet 4.5, which they had never selected — it was the default. Claude Sonnet 5, which the docs listed as the default for Code Review Flow, was not available to them, and nothing explained why. The fix, switching to Claude Sonnet 4.6, was available in their dropdown, but nothing in the docs pointed to it. Diagnosing this took roughly a week across several GitLab team members. See the threads below for extra details. ## Findings Three elements to the issue here: 1. What setting to use to select a model for Code Review Flow - The model setting for Code Review Flow (`review_merge_request_dap`, listed as **Agentic Code Review** in the UI) was added in 19.1 ([MR 236876](https://gitlab.com/gitlab-org/gitlab/-/merge_requests/236876)). - On 19.0 and earlier, the classic non-agentic GitLab Duo Code Review and Code Review Flow share one setting, listed as **Code Review** under **GitLab Duo for merge requests** (`review_merge_request`). - Code Review Flow's default model is locked to the non-agentic code review's default, currently Claude Sonnet 4.5 Vertex. - Its selectable models are the non-agentic list, so Claude Sonnet 5, GPT-5.2, and GPT-5.3 Codex cannot be reached. - Neither limitation is clearly documented. - Only GitLab Self-Managed is affected. GitLab.com is always on the latest version. Affected readers use the instance model selection page. We don't currently point users to that page from the model selection or Code Review Flow page, so it's easy to assume those pages contain the latest guidance for SM. 1. What models are available for use - The history items in the docs list milestones for when models are added or removed. But models and defaults come from the AI Gateway and reach every instance immediately. The GitLab version you are on does not affect availability. "Added in GitLab 19.3" reads as an upgrade requirement when the change reached every instance the day it shipped. - The settings in the UI **are** gated by version, however (see the Code Review Flow example in item 1). So we have to be careful about this specific case. 1. Which model selector is for classic GitLab Duo Code Review - The selector is labelled **Code Review**, which is easy to mistake for Code Review Flow. ## Proposal 1. **Clarify the behaviour.** Document that Code Review Flow uses the classic setting on 19.0 and earlier, on the pages self-managed administrators use, and state that model changes come from the AI Gateway regardless of GitLab version. Avoid a version matrix because it implies models are version-gated. 1. **Rename the UI setting.** Change **Code Review** to **Non-Agentic Code Review** so it pairs with **Agentic Code Review** and the two dropdowns disambiguate themselves. Docs describing 19.0 and earlier keep **Code Review**. 1. **Use dates for model changes.** Switch history entries for gateway-delivered model changes from milestones to dates. Entries about settings keep their milestones, because a GitLab version genuinely gates those. ## Out of scope Backporting. No single target reaches the affected population: the affected range is roughly 18.7 to 19.0, which sits outside the patch window, and archived and instance-bundled model tables are frozen while the AI Gateway keeps adding models, so a correct note would sit beside a stale list. Support engineers and solutions architects diagnosing older instances can still benefit from updates to the current documentation. ## References - Source discussion: https://gitlab.com/gitlab-com/account-management/western-north-america/cisco-group/meraki/meraki-sm-dap-pov/-/work_items/48 - Separate DAP model selection introduced in 19.1: https://gitlab.com/gitlab-org/gitlab/-/merge_requests/236876 - Claude Sonnet 4.5 deprecated for the Code Review flow: https://gitlab.com/gitlab-org/modelops/applied-ml/code-suggestions/ai-assist/-/merge_requests/6456 ## References - Source discussion: https://gitlab.com/gitlab-com/account-management/western-north-america/cisco-group/meraki/meraki-sm-dap-pov/-/work_items/48 - Separate DAP model selection introduced in 19.1: https://gitlab.com/gitlab-org/gitlab/-/merge_requests/236876 - Claude Sonnet 4.5 deprecated for the Code Review flow: https://gitlab.com/gitlab-org/modelops/applied-ml/code-suggestions/ai-assist/-/merge_requests/6456