You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
layer.generate({ tools }) sends every tool you declare, write and
destructive tools included, as the framework adapters and untyped catalogues
already did. Until now a catalogue that declared risk lost its write tools on
any request that did not contain one of 11 verbs ("Refund order 4411" was sent
without refund_order). The gate is now opt-in: new ToolSelector({ withholdWriteTools: true }). A write or destructive tool left
out by the selector is never credited as a saving: it does not count in tokens.eliminatedInput or in cost.*, and the tool-selector trace lists it
under withheldWritesNotCredited. If you set maxTools, it keeps pruning read
tools exactly as before, and outside an action request the write tools are sent
on top of them, so a request can carry more than maxTools tools.
The exact cache starts cold. Its key generation moves from n2 to n3 (see
the key fix below), so no entry written by an earlier version is read again:
expect fewer exact hits for the first days. Old entries in Redis, KV or
PostgreSQL stay until their TTL runs out.
The exact cache keeps capital letters. The exact key and in-flight coalescing
no longer merge questions that differ only in a capital that changes the meaning: curl -O and curl -o, git branch -D and -d, 5 MW and 5 mW, 3 Mb and 3 MB, [A-Z]+ and [a-z]+, Name and name. Only the capital that opens a
turn is still folded, along with accents, sentence punctuation and whitespace.
Near-duplicates that differ in other capitals (ALL CAPS, a lower-cased proper
noun) are now separate entries: expect fewer exact hits on those. An exact repeat
always hits. The semantic cache (opt-in) is unchanged.
Requests that transform their own text are keyed verbatim. When any turn,
the system prompt included, asks to translate, rewrite, summarize, fix, convert,
format or correct text, each turn enters the exact key as written, trimmed only at
the ends: "Translate to French: I love Polish food" and "…polish food", or "Can you
help me?" and "Can you help me!", never share a cached answer. The word list is
wide on purpose; with one of those words in your system prompt, only exact
repeats hit.
Empty answers and calls without a full usage counter are not cached. They are
served as before, but not written to the exact or semantic cache, so a repeat goes
back to the provider. An endpoint that never reports usage therefore stops using
the cache.
An HTTP 200 without an answer is an error. A 2xx body that carries an error
({"error": …}, Anthropic's {"type": "error"}) or no completion (no choices[0].message, no content, no finishReason) makes generate() throw
instead of returning empty text. Batch results are unchanged.
Models without a list price are estimated. A model with no catalog row and
none in costModels or defaultCostModel is still priced at the illustrative
rate, but cost.basis is "estimated", not "measured", and a price trace
carries priceSource: "illustrative". Ids with : (gpt-4o-mini:free) and
relatives whose suffix is not an alias of the same model (gpt-4o-mini-tts
no longer takes gpt-4o-mini's row) have no list price; dated, -latest, -preview and -exp ids keep their model's.
Random output and the current time are not served from cache. A request
whose user turns ask for random or freshly generated output (a random pick, a new
UUID or password, a dice roll, a coin toss) or for the current time or date skips
the exact and semantic cache. With seed, a random request is cacheable again; a
time question never is. The phrases are in English and Spanish; a false positive
costs one cache hit.
The arithmetic resolver answers only a bare question: one user message, with
no system prompt, no earlier turns, no tools and no structured data, phrased as a
question with an operator. It declines dates, phone numbers, percentages, 2.500,
leading zeros, numbers a double cannot hold exactly, and results it cannot write
exactly in nine decimals (10 / 3). A Mastra agent with instructions sends the
question to the model. It also declines superscripts, subscripts, circled numbers
and fraction characters: "What is 3² + 4²?" was answered locally as 74 and now goes
to the model. The quickstart (What is 2 + 2 * 3? → "8") is unchanged.
Fixed
Exact-cache key. Normalization no longer merges different code questions: text
in backticks or code blocks is kept as written, a sign attached to what follows is
an operator (if (ok) and if (!ok) are two questions), and quoted text keeps its
case. It runs in linear time on adversarial input, and the tool selector's
relevance folds backticks and case again.
npm install with current peers. Installing next to @anthropic-ai/sdk 1.x
or gpt-tokenizer 4 no longer fails with ERESOLVE: the optional peer ranges are
now >=0.103.0 <2 and ^3 || ^4, and bvsala init picks the Anthropic SDK
adapter for SDK 1.x. Projects on AI SDK 5, LangChain 0.3 or Mastra below 1.68
still need --legacy-peer-deps; the README lists each adapter's minimum version.
Counting tokens with gpt-tokenizer is linear. A long run of one kind of
character (400 KB of spaces took 70 s) no longer blocks the process. Counts for
text without such runs are unchanged; inside a run longer than 512 characters the
count can be at most 2 tokens higher per 512 characters, never lower.
content as a list of parts. OpenAI-compatible endpoints that return content as a list of parts (for example Mistral reasoning models), also when
streaming, no longer fail with TypeError: me.trim is not a function: the answer
is the text of the text parts.
Gateway and bvsala CLI
Without a key the gateway meters locally. In community mode every request is
measured on this machine with the provider's usage: log line, exit summary and the
local /_bvsala report routes. Only sending it to BiVelio needs BIVELIO_SERVICE_KEY. The exit summary counts the inference requests that went
through, and metering in /_bvsala/v1/status is now always true. With BVSALA_GATEWAY_USAGE_INJECT=1, include_usage is injected without a key too.
Subscription traffic keeps its TTL. The cache-write TTL lever (1h→5m) no
longer touches Claude Max/Pro traffic: there is no write premium to save. Traffic
billed per token is unchanged.
bvsala claude (and qwen, goose, aider) keeps the base URL you already
had. A pre-set ANTHROPIC_BASE_URL (ANTHROPIC_HOST for Goose, OPENAI_BASE_URL or ANTHROPIC_API_BASE for Aider) becomes the gateway's
upstream, with a line that says so. One with a path prefix, or one that disagrees
with BVSALA_GATEWAY_UPSTREAM, stops the wrapper with exit 1 and says why.
A provider cut mid-response reaches the client as a cut, not as a complete
body, and the request is counted as not metered with its reason.
The request path never changes the upstream host. Repeated leading slashes
are collapsed (a base URL ending in / works), an absolute-form target keeps only
its path, and a target that would resolve to another origin, or is not a path
(OPTIONS *), gets a 400 and is not forwarded.
--port=N works like --port N; any other --port… form is an error.
bvsala gateway and bvsala doctor reject what they do not know. An unknown
or misspelled flag or a stray argument (--upsteam, -p, --Port) exits with 1,
names the flag and suggests the closest one, instead of starting with defaults.
bvsala doctor and the README no longer tie the levers to a PRO license: they run
free, with no key. The same now holds for the gateway's startup lines, bvsala --help and the terminal guides: in community mode the levers are on and
free, and the startup line names the full switch, BVSALA_GATEWAY_ELISION=0. Only
with an organization license does that license's feature list apply.
Wrappers keep stdout clean.bvsala <terminal> prints its banner, the
terminal note and the per-request log on stderr, so bvsala claude -p … --output-format json | jq gets clean JSON. bvsala gateway
on its own is unchanged.
Ctrl-C belongs to the agent. Inside bvsala aider, bvsala goose or any
wrapped agent that keeps running after Ctrl-C, the gateway stays up until the
agent exits, then exits with the agent's code. SIGTERM is forwarded to the agent.
bvsala gateway always stops. Open connections are cut 5 seconds after
Ctrl-C, a stuck streaming response included, and a second Ctrl-C exits at once
(code 130; 143 for SIGTERM).
bvsala gemini, codex, opencode and kilo keep the endpoint you already
had (GOOGLE_GEMINI_BASE_URL, OPENAI_BASE_URL, or a baseURL in the inline
opencode/Kilo config), as bvsala claude already did: the gateway forwards
there, or the wrapper stops with exit 1 and says why. For codex, opencode and
Kilo the base URL must end in /v1.
Claude Code through --upstream deepseek is metered and counted in the exit
summary; before, every request was forwarded but reported as not metered.
bvsala init wires the cloud from the environment. The code it prints or
writes (and examples/stacks) has licenseServerUrl and serviceAccountKey
enabled: with BIVELIO_SERVICE_KEY set, usage reaches the Savings Cloud dashboard
and the license renews itself; with no keys, nothing is sent to BiVelio. The
README's adapter snippets carry licenseServerUrl too.
Python client
pyproject.toml links the public repository, and the README's LICENSE link is an
absolute GitHub URL (it was a 404 on pypi.org). From the extracted sdist, the
parity tests are skipped with their reason instead of failing.
Catalog
List-price rows for claude-fable-5-2 and gemini-2.5-flash-lite (list prices
checked on 2026-11-19), so both are measured instead of estimated.
List-price rows for gpt-5.1, gpt-6.1-sol, gpt-6.1-terra, gpt-6.2-sol, gpt-6.2-astra, gemini-4-carbon, gemini-4-argon, gemini-4-neon and gemini-4-hydrogen, with their Flex and Batch rates and the 1 December 2026 price
step of Gemini 4 Carbon and Argon. Their dollar figures no longer fall back to the
illustrative price.
Claude Sonnet 5.5 cache reads are priced at $0.10 per million tokens ($0.05 in
batches), following Anthropic's 2026-10-07 price cut: measured spend on Sonnet 5.5
requests with cache hits no longer overstates the bill.
GPT-5.6 Sol Flex and Batch savings are credited through 30 November 2026, the end
date OpenAI publishes for its promotional price. They used to stop on 21 November.
Docs
README and website: "install it and it saves" refers to the SDK. With a coding
agent, the local gateway meters your spend on your machine, and its cache-write
TTL lever saves only when the client pays per token and asks for the one-hour
cache. Claude Code on an API key asks for five minutes unless promptCacheTtl is
set to 1h.
The npm README fits under the registry's 64 KiB limit, so the package page no
longer ends mid-sentence: the 14 optional sections moved word for word to docs/guias/ and the README links each one; "Licensing & privacy" now follows the
prices. The first screen says what the package is, how to install it, what is free
and what the paid cloud adds.
docs/terminales (in the tarball): the "not metered" reasons, the TTL lever's
scope and what community mode meters are up to date.
SECURITY.md and docs/PUBLISHING.md say what is true today: the repository is
public, every version since 1.0.0 carries npm provenance, the SBOM is attached to
every GitHub Release, and the package is open core.
Website (savings.bivelio.com)
/pricing and /privacy are shared as "BiVelio Savings Layer" in the page title and
the social card, in all 12 languages.
A card payment whose Stripe events arrive out of order within the same second no
longer leaves the customer without an active plan: the webhook reads the
subscription fresh from Stripe before saving it.
A seat change on a yearly plan keeps the license expiry at the end of the paid
period, also when Stripe sends the billing period in the format used before API
version 2025-03-31.
On the Team plan, the license and the dashboard count at least the seats included
in the plan, even if the quantity is lowered in the billing portal.
A sign-up confirmation link that was already used (often by an email security
scanner) or has expired lands on the sign-in page in your language and says that
a confirmed account can sign in with email and password.
The dashboard's install snippet reads the keys from the environment, and its
closing note no longer says that the layer refuses to run without a license: it
refuses only a key that is set but invalid.