BLUECROWINTERACTIVE

Build Log

Development

Development updates for BlueCrow Studio, posted as they happen — including the ones that aren’t especially glamorous.

BUILD STATUS

In Development

CURRENT FOCUS

BlueCrow Forge — the engineering agent that works inside a repository. Operator-controlled Read Only / Edit / Commit modes, answers that have to be backed by files it actually opened, and a restricted Python capability for the work a language model shouldn't do in its head.

LAST UPDATED

September 2, 2026

NOTE

BlueCrow runs daily on internal development machines during active development. It is not yet packaged or distributed to outside testers.

Recent updates

Forge learns to prove its work

An operator-controlled Read Only / Edit / Commit mode system, answers that have to be backed by files Forge actually opened, and a training programme where every failure is graded against evidence rather than against how confident the answer sounded.

forgeverificationtrainingmilestone

Forge, and the model as a swappable part

BlueCrow Forge is the engineering agent that works inside a repository. It now runs on a local model with no account at all, or on a hosted one when a harder task justifies it — with the agent, the tools and the safety rules identical either way.

forgearchitectureprovidersmilestone

A voice that cannot spend money by accident

INGRID's everyday speech is now local and free. Anything that could bill has to be authorised first, through a single policy that every speech path answers to.

ingridvoicepolicyengineering

INGRID acts, and proves it worked

Ask for an application by name and INGRID opens it, finding what is genuinely installed rather than reading from a hand-maintained list. Underneath, a control loop that observes, decides, acts, and then checks whether what you asked for actually happened.

ingridautonomylaunchermilestone

The unglamorous week: shutdown, threading, and things that only fail sometimes

A week spent on defects that produce no screenshot: a shutdown that occasionally ignored the first close, background work nobody owned, and an imaging library being called from several threads at once.

reliabilityengineeringquality

INGRID Link: remote video that survives a real connection

A working demo and a dependable feature are different things. INGRID Link's remote video now adapts to the connection it actually has, recovers from a stalled camera on its own, and has been run overnight to prove it.

ingrid-linkremotereliabilitymilestone

INGRID Call: live-verified on a real phone

A real two-way voice call with INGRID over a private, encrypted network link, with her live camera feed staying up throughout, has now been verified end to end on a real phone — including automatic recovery after the phone locks or switches apps.

ingridcallvisionmilestone

INGRID starts to see: a camera pipeline and Vision panel

A provider-neutral camera capture pipeline, a new INGRID Vision panel with a live preview and honest status, and a matching Observing avatar state — built and unit-tested ahead of a physical camera arriving.

ingridvisioncameramilestone

INGRID's first slice of conversational fluidity

A new semantic follow-up recovery layer, built on top of stabilized conversational memory, search context, and transient-failure isolation — the first tested slice of INGRID understanding rephrased follow-ups naturally, not just her original exact phrases.

ingridconversationmilestone

INGRID can hear you now

Live, end-to-end voice conversation with INGRID: real microphone input, OpenAI speech recognition, and a natural reply in Marin, an OpenAI voice — verified against real hardware, not mocked calls.

ingridvoicemilestone

Teaching INGRID to reason about heat

INGRID can now tell you whether your workload is CPU-bound or GPU-bound, and whether your cooling looks like it’s keeping up.

ingridreasoningmilestone

One health score, computed once

The dashboard, the health score, and INGRID used to compute their own numbers independently. Now there’s exactly one.

architecturehealth-scorestability