Luke Duncalfe (0860cc86) at 18 Mar 07:17

Luke Duncalfe (698df651) at 18 Mar 07:15

Luke Duncalfe (bb5cb636) at 18 Mar 07:13

Luke Duncalfe (6f23d392) at 18 Mar 06:10

Luke Duncalfe (06ac2803) at 18 Mar 05:38

Luke Duncalfe (6127d696) at 18 Mar 05:22

Luke Duncalfe (6bf9a471) at 18 Mar 05:15

Since we are now in GA, I believe the 3rd point in the issue may no longer be needed, as we don't want to filter out projects based on Experiment/Beta settings anymore.

@Imjaydip Flows is still Beta, so a nice-to-have might be to have some support for this when the user is searching for projects to enable for Flows. But if it's difficult to do, we could create a follow-up issue with the idea.

Given these constraints, I don't think we can use this with_duo_eligible method as-is to filter projects for AI Catalog.

I wonder if we should change those existing constraints in the finder, as they may be incorrect now that DAP is becoming available to free tier for 18.10 https://release-18-10.about.gitlab-review.app/releases/2026/03/19/gitlab-18-10-released/#purchase-gitlab-credits-on-the-free-tier-on-gitlabcom @tgao3701908 @partiaga what are your thoughts on this as the author and reviewer of !209581 (merged)?

If the logic shouldn't be changed, I'd rather we change the name of the method #by_duo_eligible to describe more specifically what it is supposed to do, and add a new version of #by_duo_eligible which doesn't have the tier constraints.

After dogfooding the agent a couple of times, I can see some room for improvement (I'll 🎗️ see if I can make these improvements when I have time).

• It's currently instructed to be familiar with Flow Registry YAML, but Custom Flow YAML is a subset of that so the agent currently suggests properties that are invalid for Custom Flows, such as name and description. Recommendation We could perhaps instruct it to read the docs just added docs: Document Custom Flows schema differences ... (!227463 - merged) which describe how Custom Flow YAML differs from Flow Registry YAML.
• It seems to hallucinate around inputs - as a guess, it probably has no real insight into what data is coming through the context, for example, for any given trigger event. Recommendation We might need to document what context is available to Custom Flows from different trigger events, and update the agent to review those docs. We may need to instruct it not to guess at inputs.
• It would be great to continue to chat with it when a Custom Flow session fails to assist with debugging the problem and then iterating on the flow YAML to improve it, however, currently the agent does not have the tools to read pipelines jobs. Recommendation Give it that tool (this one's easy). Maybe need to add some prompting tips for this situation.
• Debugging failures is very hard because we currently don't expose enough information in the job log around what has gone wrong. Recommendation MVC Part 2: Session Debugging & Troubleshooting... (#592972) may help

Luke Duncalfe (780354f3) at 17 Mar 23:50

... and 1 more commit

Luke Duncalfe (1feab8e3) at 17 Mar 23:50

What does this MR do and why?

Custom flows are a subset of the Flow Registry v1 specification, with certain fields intentionally restricted. Add a new page documenting these restrictions and update custom.md to link to it. Without this, users may encounter validation errors and confuse when creating custom flows, as not all Flow Registry fields and features are supported for custom flows. For example : #587049 (comment 3118337142)

References

#587049 (comment 3118337142)

Screenshots or screen recordings

Before	After

How to set up and validate locally

MR acceptance checklist

Evaluate this MR against the MR acceptance checklist. It helps you analyze changes to reduce risks in quality, performance, reliability, security, and maintainability.

Related to #591995

@tbulva Sorry to bother you - this was a follow-up task created by incident.io around the MR gitlab-org/gitlab!225920 (merged), which we reverted after it landed in Canary. It's considered a "near-miss" in that it could have reached production.

Are you able to review whether we can improve our test coverage to prevent a similar problem from merging in future, and if so, add the missing test coverage? If this is something that requires E2E testing on a level that's not possible today, knowing this as a critical gap in our testing abilities would be very useful 🙇 Thank you!

@Imjaydip Thank you! Set to add to the merge train when checks pass 🚀

I'll resolve because this thread isn't blocking us merging these docs.

@Imjaydip I was imagining it might be used for stop keywords for agents to use as their final_answer for other agents, or something 🤔. @mikolaj_wawrzyniak Do you think AI Catalog Custom Flows (written by customers) would benefit from supporting stop?

@samdbeckham @amandarueda @justin_ho cc @knejad (as our resident database expert!) It might be worth us putting some initial backend effort to investigate whether advanced searching should all happen in PostgreSQL or whether ElasticSearch offers any advantage.

We've discussed this briefly before, see these comments with some small information from the search team:

• gitlab#555554 (comment 2634267909)
• gitlab#555554 (comment 2641097541)

Perhaps (if it hasn't been done already) signalling from a product sense what would bring the most value would be helpful too if we need to weigh efforts in the implementation for some over others.

🤖 Sorry for the noise - this is about 50% useful.