Notes from the plumbing.
Every post leads with runnable code and a real run — measured token counts, cent-level costs, and the governance that actually fired. If a number appears here, the recipe that produced it is public.
More posts
Your OpenAI bill is a black box. Here is the seam that opens it
A step-by-step journey from an unwrapped OpenAI client to a governed one: exact tokens, a Decimal cost, a pre-flight budget that refuses the call that would cross your cap, and a redaction that fires before the request leaves your process. Every step has a run command, its real output, and a screenshot. Python and TypeScript, no API key needed.
Governing Microsoft Foundry: four ways to call it, and only three can be governed
On Foundry, how much governance you get is decided by one thing — who holds the model client. The Foundry SDK hands you a plain OpenAI client, so all seven Cendor libraries work with no Foundry-specific code. The Azure AI Inference SDK hands you a different shape, and instrument() returns it untouched: zero events, budgets that never bind, an empty audit chain, and an app that looks fine. Plus the deployment name that makes a USD cap silently do nothing. Python and TypeScript, runs offline with no Azure account.
Your Teams agent's policy refusal is showing up as "the agent hit an error"
A Microsoft 365 Agents SDK custom engine agent holds its own model client, so it is ordinary library work — except that the guardrail call RAISES on a block rather than returning one, three separate exception types reach a governed handler, and on the TypeScript port the session cap silently never accumulates. A step-by-step journey with the real Agents Playground on screen, a mid-stream break, a verifiable audit chain, and the whole agent replayed for $0. No tenant, no tunnel, no key. Plus the governance Adaptive Card that makes all of it visible to the person in the chat, because `channelData` is not.
Introducing Cendor: production plumbing for LLM applications
Seven composable libraries and a governed agent SDK for the layer beneath your agent framework — context, cost, guardrails, testing, and audit, in Python and TypeScript. Local-first, Apache-2.0.