When you use Markdown so much you start typing Markdown into Word documents automatically, without thinking! _(AUTOPILOT!)_

Strangler fig: retiring a legacy dispatch layer #patterns #practices #stranglerfig

Strangler fig: retiring a legacy dispatch layer | luke.geek.nz

Apply the strangler fig pattern to retire a legacy dispatch layer: domain-by-domain migration, route parity tests, ported tests, and safe cutovers.

luke.geek.nz

Building Execution Ledger - Durable Workflow Orchestration in Rust on Azure Container Apps Execution Ledger is an open-source durable workflow orchestrator I built in Rust on Azure Container Apps. #rust #aca #mvpbuzz #drasi #azure

Building Execution Ledger - Durable Workflow Orchestration in Rust on Azure Container Apps | luke.geek.nz

Learn how I built Execution Ledger, a durable workflow orchestrator in Rust running on Azure Container Apps with step-level replay, idempotency, compensation, KEDA autoscaling, and managed identity.

luke.geek.nz

Execution Ledger built on Rust with Drasi support. Durable job orchestration for Azure. Step-level idempotency, replay, skip, and compensate controls. Tenant-scoped API with Entra ID auth. Connector packs for REST, queue, blob, SFTP etc.

GitHub - lukemurraynz/Execution-Ledger: Execution Ledger

Execution Ledger. Contribute to lukemurraynz/Execution-Ledger development by creating an account on GitHub.

github.com

When Your AI Agent Lies: Silent LLM Fallbacks A system I was debugging had AI-generated content flowing through it for days. The responses came back fast. No errors in the logs. No alerts fired. Every single response was templated.

When Your AI Agent Lies: Silent LLM Fallbacks | luke.geek.nz

Detect silent LLM fallbacks: when AI agents return templated responses with 200 OK, monitor token usage and fallback rates to avoid hidden correctness failures.

luke.geek.nz

Treat prompt changes like code deploys Most teams running LLM-backed features would never ship a code change without a test suite and a CI gate. Then they tweak a prompt, eyeball one output, and send it straight to production.

Treat prompt changes like code deploys | luke.geek.nz

Prevent silent regressions by treating prompt edits like code deploys, with automated tests, CI gating, and continuous evaluation.

luke.geek.nz

Avoiding Reasoning Model Failures with Microsoft Foundry Swapping only the deployment name is not a safe change for reasoning-tier models. Even with the same endpoint and SDK, reasoning deployments can fail in at least four different ways.

Avoiding Reasoning Model Failures with Microsoft Foundry | luke.geek.nz

Learn four ways reasoning-tier models can break Azure OpenAI integrations and how to fix API versioning, params, token budgets, and reasoning_effort.

luke.geek.nz

🥇 AzureRM Provider v5.0.0 dropped! * RP registration defaults changed (none not legacy) * Enhanced validation moved into features block * Preflight validation at plan time (opt-in) * Removed resources, deprecated flags #terraform #azurerm

Releases · hashicorp/terraform-provider-azurerm

Terraform provider for Azure Resource Manager. Contribute to hashicorp/terraform-provider-azurerm development by creating an account on GitHub.

github.com

This is neat! Dragons! www.glasgowlive.co.uk/news/glasgow...

Game of Thrones dragon soars over Glasgow to celebrate Commonwealth Games opening

The eight metre flying model took to the skies again this morning.

glasgowlive.co.uk

My first Tauri app! Its not perfect but its bloody fast. v2.tauri.app

Saturday morning means brunch.

Microsoft Foundry hosted agents: lessons learned building a production multi-agent service Some of my thoughts and findings on my first deployment of Foundry Hosted Agents.

Microsoft Foundry hosted agents: lessons learned building a production multi-agent service | luke.geek.nz

Lessons and best practices for hosting production multi-agent services with Microsoft Foundry hosted agents—compute, protocols, sizing, and deployment tips.

luke.geek.nz

Nastiest bug of the month: AppConfigurationClient configured with trim_prefixes while every reader queried fully-namespaced keys. No errors - every single config read just silently returned the default. Config mismatches like this fail quiet. Log your resolved values.

I got a weird look when I told someone last week, I was refactoring something to RUST! Now I need to actually learn it!! Time to sleeves up, thanks @willvelida.com! rustup time!

Will Velida 🇺🇦@willvelida.com · 2w ago

I've created a free course for learning @rust-lang.org on my YouTube channel. If you want to learn how #Rust works, check it out! youtube.com/playlist?lis...

Reasoning-tier gpt-5 models (gpt-5, 5.1, 5.5, o1, o3-mini) reject temperature and most sampling params outright. gpt-5-chat still accepts it though. Make sampling params optional with drop-and-retry instead of hardcoding them.

Gotcha: Azure OpenAI's /openai/v1 Responses API doesn't take literal "v1" or "preview" as api-version — it pins to a moving preview date-string (now 2025-11-15-preview). Older date-style versions 400 immediately. Verify the current value before hardcoding it.

Assistants API hard-retires 2026-08-26 on OpenAI AND Azure OpenAI - same deadline. But migration paths differ: OpenAI devs move to the Responses API; Azure OpenAI devs move to Foundry Agent Service (built on Responses). Follow Microsoft's guide, not OpenAI's.

Azure SDK clients refuse bearer tokens over plain HTTP. If you're capture-replaying Azure SDK calls against a local mock server, your test doubles need AzureKeyCredential, not DefaultAzureCredential - otherwise it's not the mock failing, it's the transport.

Linking a Container App as a Static Web Apps backend caps you at a hard 45s timeout and HTTP-only - no WebSocket, no matter what your Container App supports. Long-running or persistent-connection agent endpoints? Go through Front Door or hit the Container App directly.

ACA gotcha before promising compliance a checkbox: there's no CMK setting for the managed environment itself. You can CMK dependent services (ACR images, mounted Files) but not the environment. Mandated CMK on the compute platform? That's an AKS conversation.

Building an MCP server on top of an internal agent framework? Don't hand-pick a subset of tools for MCP exposure - route the whole tool registry through the same audit/authorize decorators your other surfaces use, so MCP clients see the same capability set as your other clients.

TIL (the hard way): Container Apps ingress timeout defaults to 240s but isn't fixed — you can raise it up to 2400s (40 min) at the environment level via --ingress-timeout.

If prompts are part of your deployable surface, treat "promote to prod" like a merge to main: gate it on eval pass, not vibes. We added a hard PermissionError if the latest eval run for a prompt failed or never ran.

Debugging tip: if your agent pipeline has a deterministic fallback path, watch for it firing silently. Our tell: response time <5s + total_tokens_used: 0 = the LLM call never actually happened. Three stacked misconfigs caused this & every symptom looked like a working agent.

anyone done rooftop solar on a new roof? How does it affect warranty etc?

Awesome swag - just arrived GitKraken! I have become a #GitKraken Ambassador! GitKraken Desktop is making my git repo management easier across profiles for sure!!!

AzureFeeds - Newsletter for this week is out! 📅 Weekly Update: Saturday July 11 2026 to Saturday July 18 2026 #azurefeeds #newsletter azurefeeds.bmeurl.co...

Who needs a workout when your pushing a trolly at Costco

First GitHub Exam passed! GH-300 GitHub Copilot!

Azure Cloud Commanders - Azure Functions and network security.. Can it be done? #azure #mvpbuzz #azurecloudcommanders #azurefunctions #techcommunity Session with Rex is on now!