Tags

testing-rc-2026-08-26.48

testing-rc-2026-08-26.47

testing-rc-2026-08-26.46

testing-rc-2026-08-24.45

RC 45 — shared capacity, end to end

Twenty-one commits since RC 44. The arc is Operations' generalized
shared-capacity contract, from the identifiers underneath it to a
prompt that holds itself while a pool starts.

CAPACITY
- Four readiness states, `standby` among them, resolved offering-exact:
  the approved catalog record, its pool, that pool's observed state.
  Warming never implies model readiness, and standby is never called
  warming before Operations accepts a warm request.
- Warm and standby intent for one approved offering. The lease, the
  identity and the request id are the server's; standby carries no
  lease, no allowance check and no launch gate, because an expired
  entitlement must never prevent stopping spend.
- Recurring schedules expand into bounded UTC windows — wall-clock
  hours, midnight crossings, both DST directions, overlapping rules
  merged.
- A prompt aimed at sleeping capacity is held, starts it once, and is
  released exactly once when the exact offering is ready — on every
  surface, not only chat.
- LOGOS_SHARED_COMPUTE_LAUNCH_ENABLED is honoured, defaulting to off.
  It was read by nothing before this.

MONEY AND ITS EVIDENCE
- The Compute panel says what compute costs, keeping the monthly
  allowance and a sponsored grant apart.
- Warm pools are capacity among the workers, with what they cost while
  warm and the schedule that stops it.
- Several dedicated workers at once, and a list that survives a reload.
- The usage section in Billing renders at all, for the first time.
- Payments in Settings: saved cards, the billing configuration, and no
  field anywhere that takes a secret.

BOUNDARIES HELD
- Six Operations identifiers preserved end to end; the stored offer is
  a record and never the authority for provisioning, authorization,
  settlement or metering.
- Integer micro-USD throughout, parsed rather than grepped.
- The operator-only DeepSeek endpoint is fenced off by test.
- Capabilities a deployment switched off are no longer offered.

Media stays unpublished.

testing-rc-2026-08-24.44

RC 44: the sponsored path, ready to prove itself against live Operations.

Sponsored beta entitlements are fetched by ORGANIZATION over the
product identity, the returned billing account and snapshot digest are
persisted on the warm request, and the grant is spent from its own
window ledger rather than a monthly counter. None of it needs Stripe.

Cut so the service-to-service identity and network path can finally be
exercised: the contract is verified against Operations' published shape
and their confirmed record, but nothing has yet minted a token as
logos-backend@ and called them. That needs a deployed backend carrying
this code.

Also: platform workers delete in two stages like personal ones, the
enrollment token has no browser proxy left to reach through, and no
service-specific Anthropic fallback flag exists anywhere.

MIGRATION-BEARING: 66, additive.
MEDIA NOT PUBLISHED — PUBLISH_MEDIA_RELEASE=false.
GATES UNCHANGED: shared-compute launch, reconciler and Vast volume
delete all stay false.

testing-rc-2026-08-24.43

RC 43: shared Muse capacity, and money that knows what kind it is.

The product asks one reviewed pool to warm, mints a platform enrollment
under the same request id, follows Operations' nine states, and tells
them when its OWN answerer says the model would take a prompt — only
then are the linked requests ready. A prompt for that model is HELD
server-side until the warm request is ready, the model resolves ready,
and the pinned offering still matches.

Two accounting models kept apart: a subscription is a calendar-month
allowance; a sponsored beta grant is a finite pot credited to a ledger
keyed by its snapshot digest, so a month boundary cannot refill it and a
renewal cannot inherit it. Shared inference is metered because the
EXECUTION ROUTE says shared_compute, never because the transport is
local_worker.

No service-specific automatic Anthropic fallback remains anywhere in
code or configuration.

MIGRATION-BEARING: 62-65, all additive.
MEDIA NOT PUBLISHED — PUBLISH_MEDIA_RELEASE=false.

testing-rc-2026-08-23.42

RC 42: the dedicated purchase, end to end.

Offer discovery against Operations' contract, a quote a person confirms,
saved cards entered on Stripe's own page, an ownership check before
anything immutable exists, the Operations billing account resolved by
UUID, and a manual-capture hold that only a verified webhook may
authorize. The four Anthropic fallback flags are gone: a provider that
fails now raises instead of quietly spending elsewhere.

MIGRATION-BEARING: 60-61, all additive.
MEDIA NOT PUBLISHED — PUBLISH_MEDIA_RELEASE=false.

testing-rc-2026-08-22.41

RC 41: compute resolution rebuilt — one resolver, the person's choice outranking mode, per-user scoping, canonical provider fields, and image work routed to media rather than the prompt. MIGRATION-BEARING: 57-59, all additive.

testing-rc-2026-08-21.40

RC 40: dedicated compute bound to what authorized it — offering catalog, reservation binding and lifecycle, model-exact routing with a readiness preflight, shared compute visible; durable project storage steps 1-6. MIGRATION-BEARING: 50-56, all additive.

testing-rc-2026-08-19.39

RC 39: durable project storage — a catalog that holds room before bytes, imports that become objects with opaque IDs, byte-time delivered exactly once; an invited member can create a project

testing-rc-2026-08-18.38

RC 38: dedicated-worker enrollment automated (token never reaches a browser), cloud imports visible, every Python image on one Bookworm base

testing-rc-2026-08-18.37

RC 37: the deployment merge plus the base-image CVE sweep — every published Python image on one Bookworm base

testing-rc-2026-08-17.36

RC 36: the runtime stops narrating itself — conversational answers, channels on the project, a sleeping engine reported as state

testing-rc-2026-08-16.35

RC 35: the smoke asks the right python — RC 34 re-cut with the venv-interpreter fix

testing-rc-2026-08-16.34

RC 34: the pipeline fix lands — GPU-free ComfyUI runtime smoke green

testing-rc-2026-08-16.33

RC 33: the missed merge rides along — ComfyUI outbound worker runtime, plus the selector-may-decline and generate-from-image routing fixes

testing-rc-2026-08-15.32

RC 32: the one-surface release — five stops everywhere, gear in the brand bar, New at the prompt, one converged log

testing-rc-2026-08-15.31

RC 31: the frame holds — carousel axis fix, hydration fixes, dedicated worker card, Operations contract

testing-rc-2026-08-15.30

RC 30: mobile-first chat, attachments, vision routing, decline fallback, ComfyUI device channel

testing-rc-2026-08-14.29

RC 29: the local connector, done right

Three commits on top of RC 28. The local default model is discovered
from the running Ollama (env override → live discovery → last-resort
constant) so a newly pulled model moves the whole local family without
a code change. The outbound worker's model derives from the enrolled
device's heartbeat capabilities — never an Ollama probe a hosted
backend cannot reach — enqueueing exactly what the device advertised,
with explicit LOCAL_WORKER_MODEL as the cloud-testing pin. Worker
failures reach the UI as structured JSON with specific codes (offline,
model mismatch, timeout, inference error, nothing-advertised).

Three serving routes stay independent: vllm = hosted DeepSeek
(unchanged), local_worker = the user's enrolled Mac/Gemma,
platform/Vast worker = Muse Glimmer. No paid fallback fires when
LOCAL_WORKER_FALLBACK_PROVIDER is empty. Gates unchanged; media jobs
manual; worker artifact unchanged in intent since RC 23.