So had a fun problem show up recently We have azure-openai and it only support the deployment flow url and it worked fine.. But newer openai models use the responses endpoint So was looking at the docs and it says to use openai-generic but is that /chat/completions or /responses compatible? https://docs.boundaryml.com/ref/llm-client-providers/microsoft-foundry So im guessing i need to setup the responses client in my registry https://docs.boundaryml.com/ref/llm-client-providers/open-ai-responses-api But docs are not super clear so just asking here if there is any guidance https://cdn.discordapp.com/attachments/1253172394345107466/1554244046367756391/image_4.png?ex=6abc2e11&is=6abadc91&hm=adc90648995f71f94d6618caa11ffbc4d25c10dff8a7b70511c669541bcd4945& https://cdn.discordapp.com/attachments/1253172394345107466/1554244047210807307/image_3.png?ex=6abc2e11&is=6abadc91&hm=1fb13d6dd5444f355f5dc0c550ffbfea67f31e05a2d7d9e75ea2d495ad548512&

I'm using a raw terminal with any CLI agent (codex, cursor cli, claude code cli)

yes. I am hosting my repos in a remote sandbox (self hosted) (I'm using coder, that's like github codespaces)

Have you guys been design a workflow where you exclusively supervise your ai agents on your phone?

The recent vid's are great. Are y'all actually building a software factory as well for public use, or just for internal use? I am wanting to formalize more my own software factory and thinking about doing it directly in BAML, but certainly starting with some point-solutions in BAML (such as a clean-room PLAN review workflow). Just curious if there are particular lanes that would be cleaner to stay in w/r/t BAML project roadmap...

I was watching a YT videos. it said you guys ran agents on MacBook. What are the specs of the MacBooks?

Is there any way to get the full prompt that is sent to the LLM? I want to log this to my database for now. But it looks like you can only get this from an error or bamls own internal logging.

Hey all, just wondering if in the baml toolchain there is a supported api or cli for static source analysis? I mean beyond compilation and symbol description? I am thinking something like tserver? or ts-morph? It's not a hard issue for me, but does shape a decision for me in the project I am using to explore baml

Does Jetbrains plugin have the same feature parity with vscode? I see that the versioning is the same, but checking the old documentation seems like jetbrains is behind.

Is the https://boundaryml.com/techdocs URL meant to be live? It's pointed to in several links at https://boundaryml.com/#describe

womp, looks like baml website is broken https://cdn.discordapp.com/attachments/1253172394345107466/1534644336694788179/image.png?ex=6a74e06a&is=6a738eea&hm=b7b92f42cc8bc6056e80d97f6ddbe6846219d155558e6b76b0155bb67179d0df&

Perhaps I got too used to the old UI, but I can't seem to find many things in the new Playground - Token usage is missing, test history cannot be accessed without rerunning the test, data is redacted etc. Am I missing something here? https://cdn.discordapp.com/attachments/1253172394345107466/1534286423119626310/image.png?ex=6a739314&is=6a724194&hm=2df9e2429c48255e13b332d4d9fa8f43e566ea448e26372eea6292397ec888e1&

Hey folks, is there anything special we have to do for the

allowed_role_metadata

flag for the gpt 5.6 models? From my testing, it seems those only work for claude & cache_control, but not prompt_cache_breakpoint

Hey! Quick question about the generated SDK folder. I've got a small Bun/TS service using BAML.

baml generate

outputs ~40k lines into

baml_ts/

, and I'm not sure whether that should be committed or gitignored — it's generated, so part of me says ignore it, but it's not a package manager artifact either (the CLI isn't a JS dep, and my Dockerfile just COPYs the folder instead of regenerating). What do you recommend? And if committing is the way, is there a recommended way to catch it going stale when someone edits a prompt and forgets to regenerate? Thanks!

Hey! Quick question about the VS Code extension. When I search "baml" in the marketplace I see two extensions, both published by Boundary: one called "BAML" (BAML language support — syntax...) and another called "Baml" (BAML is a DSL for AI applications, ~8K installs). I'm on Arch Linux with the toolchain installed via yay (

yay -S baml-bin

), and I installed the extension with

baml ide install --code

. That gave me the first one ("BAML language support"). What's the difference between the two? Is the one installed by

baml ide install --code

the correct/current one, and is the ~8K-install one the older marketplace version I should ignore? Just want to make sure I'm not running the wrong one. Thanks! https://cdn.discordapp.com/attachments/1253172394345107466/1532429005259800777/Captura_de_pantalla_20260730_184304.png?ex=6a6cd13a&is=6a6b7fba&hm=7625bbf68294f48df7f68e0a873466be8f176f37a75a4bd223200768b6570996&

Testing the bamlcode repo, and I keep running into this issue - Agent loop seems to give up with no traceback when it runs into lots of steps or large files. Is this normal? https://cdn.discordapp.com/attachments/1253172394345107466/1532086352458289293/image.png?ex=6a6b921b&is=6a6a409b&hm=5f5ec2b0c861b88910b3501fd5da4723b0546179c8c5d15ecaa7060e63e9bba4&

can i still get a shirt from filling out the form at https://discord.com/channels/1119368998161752075/1119375433666920530/1483184576141590612

When something is open on the right view it keeps the folder on the left where it comes from held open. However you can also click that folder to collapse it but it doesn't show any visual indication of anything changing. When the right view navigates somewhere else it'll close the folder nicely which is really nice but it should probably have some kind of visual indicator that the folder will be collapsed after navigating out if this is intended behaviour. Is this intended or not? If this isn't intended behaviour then imo it should be it's really nice to use hahaha https://cdn.discordapp.com/attachments/1253172394345107466/1526029023178785008/Code_mWO51mUHpp.png?ex=6a5588c7&is=6a543747&hm=18efba20c5019cf55dc7be62480bf2a33e0595f392fa1b1391ff68c2d1c95869&

Does anyone know of any attempts at describing the scaling of time/input token/output token complexities with respect to narrowly defined LLM prompts? It'd be nice to have a way to formally describe how semantically consistent LLM function primitives scale to have a useful way of building really complex but reliable things with them. Like a stochastic variant of Big O. Seems like something worth spending some time thinking about imo.

How does onboarding for sheep council work? I'd like to get access but I'm normally not awake around seattle business hours because I live in NZ. I will be awake around those times this week however because I ended up flipping my sleep schedule programming instead of sleeping after f1 was finished whoops.

how can i get early access of baml first-class standard library harness?

what will be in scope of 1.0 stdlib, i want to create a websocket lib, should i extent the baml runtime rust code, or write in pure baml? or maybe generate baml sdk, and use websocket in the host language

Hey, quick BAML namespace/client question. After moving modules into

ns_*

folders, I tried keeping shared clients in

baml_src/clients.baml

at root, e.g.

client<llm> FastLlm { ... }

. But inside

baml_src/ns_blog/...

,

client: FastLlm

does not resolve the root client, and

client: root.FastLlm

also seems unsupported in the

client:

field. If I duplicate

client<llm> FastLlm

inside multiple namespaces,

baml check

passes, but runtime fails with:

Client resolve function FastLlm$new matches multiple namespaced entries

What’s the intended way to share one LLM client across namespaces? Should root clients be referenceable from namespaced LLM functions, or do client names need to be globally unique per namespace?

Hi there! I have a question. I see the latest version of BAML is 0.223 right now. I saw you create a new version of language with new functionality. Does it mean old version of BAML will be deprecated and we will not have direct call BAML functions from typescript/python in the future like it was before in old version of BAML ? So it will be new language without support old approach ?

Where can we obtain the AI That Works notes? The recent github links are broken: https://github.com/ai-that-works/ai-that-works/tree/main/2026-06-23-software-factory-for-agent-tools

hey I saw in this video that there might be some demo code available on the discord?

> Is there a way to mock the LLM response and write test cases for unexpected outputs without actually calling the LLM? > > For example, I currently have an optional field defined as a

string

, which I later parse into JSON in my code. I'm planning to change it to

@@dynamic

, but before doing that I want to verify that my parser can handle hallucinated or malformed responses. > In practice, the LLM may return: > *

null

> * a valid JSON object (

{ ... }

) > * an empty array (

[]

) > * or other unexpected values. > Regardless of the input, I want my code to always produce a valid JSON object (

{}

). > I know I could modify the prompt during testing to intentionally make the LLM return invalid data, but I'm wondering if there's a better way to mock the LLM response or directly test the parser against these edge cases before switching to

@@dynamic

.

What are you guys finding the best way to track cost per call? I'm trying to get a sense of how much it costs us per page (we process pdf's). I've got something close to working but I seem to be consistently off what google is billing me for gemini. I'm not off by a fixed percent either.

is there a newsletter or email subscription list for upcoming 🦄 ai that works episodes/ to recieve the email.md summaries after the episode is released?

Hi Team! I'm struggling to get the Zed language server extension set up. I've tried uninstalling, restarting, reinstalling. Still get the same error:

Language server baml-language-server:

from extension "Baml" version 0.223.0: no asset found matching "baml-cli-baml-language-0.12.2-nightly.20260625.d-aarch64-apple-darwin.tar.gz"

Has someone had a similar issue? I think it's not downloading the language server when it should, but I don't know how to trigger it.