buildUsage() only emits cache reads under prompt_tokens_details, so the
top-level-only read dropped the count for every Responses-format provider
(codex, grok-cli, ...), persisting cached_tokens: 0 and billing cache hits
at the full input rate. Mirror the cache_creation fallback already used
just above.
Registry/pricing/CLI catalog already had gemini-3.7-flash-{high,medium,low}
but MITM_TOOLS.antigravity.defaultModels was missing them, so the tiers
never showed up in the dashboard model-mapping table.
The 3000 ms timeout on /v1/compress was fixed, so busy or slow machines
timed out often and sent the LLM an inconsistently compressed body,
hurting prompt caching. Add a headroomTimeoutMs setting, thread it from
the chat handler down to compressWithHeadroom, expose it in the Token
Saver dashboard, and normalize invalid values back to the 3000 ms default.
Prefer the sourceFormat-matched runtime transport over a model's
declared targetFormat when both apply. MiniMax-M3 previously resolved
to a Claude-shaped body while being posted to the already-selected
OpenAI endpoint, silently dropping image_url blocks from OpenAI
clients. Fixes#3418.
Resolve body.size through sizeToAspectRatio and append the ratio as a
-WxH suffix so the executor's parseImageConfig picks it up. Also fall
back to gemini-3.1-flash-image when a non-image model reaches the
image handler.
Two problems on the credentialFallback path, where a search provider
borrows a chat provider's connection:
- the lock was attributed to the search provider id, but the connection
belongs to the chat provider, so markAccountUnavailable looked it up
under the wrong provider and read a stale backoffLevel
- with no model argument the lock key is `modelLock___all`, which
isModelLockActive treats as blocking every model — one failing search
would have taken the shared glm key offline for chat as well
Attribute the lock to the provider that owns the connection, and scope
it to `websearch:<provider>`, passed to getProviderCredentials too so
the lock is read back under the same key.
Add a Zed usage handler so connected Zed accounts appear on
/dashboard/quota. Reads GET /client/users/me for plan, edit
predictions, optional hosted model requests and billing-cycle reset.
Render unlimited rows as "N used · Unlimited" instead of 0 / ∞, and
surface overdue-invoice / token-billing messages.
CommandCode returns errors as a type:"error" event inside an HTTP 200
NDJSON stream instead of a non-200 status, so the existing combo/account
fallback logic (keyed off response.status) never triggered and the error
text was streamed to the client as if it were content.
Peek the first NDJSON events before committing to a stream; on a
type:"error" event, abort and return a proper 4xx/5xx Response instead.
Normal streams are replayed losslessly (buffered prefix + rest of the
stream) through the existing translator, so the happy path is unchanged.
Add CommandCodeExecutor.parseError() so parseUpstreamError() can extract
a clean message/status from the synthesized error body.
The separate zai-search entry showed "No connections" on the web search
page because credentials live on the `glm` connection, not on it. Every
other provider that does both chat and search (antigravity, kimi, xai,
gemini) declares webSearch on the provider itself, so do the same here.
- glm gains serviceKinds ["llm", "webSearch"] and the MCP searchConfig
- the request builder / normalizer move from "zai-search" to "glm"
- drop the zai-search registry entry and its svg logo, which also
removes the only need for svg logo support in getProviderIconSrc
ollama-search keeps its own entry and credentialFallback: its search
endpoint is unrelated to the ollama chat transport.
Register two web search providers that reuse an existing chat provider's
API key instead of requiring their own connection:
- ollama-search (POST ollama.com/api/web_search) reuses the `ollama` key
- zai-search (POST api.z.ai MCP web_search_prime) reuses the `glm` key
A new `credentialFallback` registry field drives this: when a search
provider has no connection of its own, the search handler falls back to
the linked chat provider's credentials.
Also teach getProviderIconSrc to serve .svg logos for providers that
ship vector art.
translateResponse() short-circuited untouched on claude->claude streaming,
so OAuth-cloaked tool names (CLAUDE_TOOL_SUFFIX) leaked to the client and
every tool call was rejected as unknown. Add decloakStreamChunk(), the
streaming counterpart of decloakToolNames(), and call it on the same-format
path using the already-plumbed state.toolNameMap.
GLM quota parsing only accepted TOKENS_LIMIT and wrote every limit to a
single "session" key, so credit-based plans showed nothing and later
intervals overwrote earlier ones. Accept CREDIT_LIMIT too and derive the
quota key from the limit unit (5h session, 7d weekly, tokens, custom).
Moves the parser into its own usage/glm.js, re-exported from misc.js.
Z.ai / GLM-5.2+ require a top-level reasoning_effort (low/high/max)
alongside thinking:{type:"enabled"} to control reasoning depth; the zai
branch previously only set thinking and dropped reasoning_effort, so every
GLM-5.x request ran at the model default (max). Gate the field behind
GLM-5.2+ (thinkingEffortSupported in capabilities.js) since older GLM
(4.x, 5.0, 5.1, 5-turbo, 5v-turbo) do not read it, and map client levels
to the exact low/high/max values z.ai accepts.
extractThinking now checks reasoning_effort/reasoning.effort before the
thinking object so a client-supplied effort is not overwritten by
thinking:{type:"enabled"} mapping to mode:auto.
Fixes#2721
Caveman/Ponytail injection now matches each target wire format instead of
assuming an OpenAI-shaped body:
- Chat arrays append a text block; Responses arrays append input_text and
create typed message items
- Claude inserts before the final cache-control block; Gemini preserves the
snake/camel systemInstruction wrapper
- Kiro updates systemPrompt and its mirrored first-user prefix atomically,
rolling back if the pair fails to converge
- Format label decides Claude/Gemini before the wire-shape sniff, since their
bodies also carry messages[]/contents[] and Anthropic rejects a "system"
role inside messages[]
- Delimiter-aware dedup makes injection exact-idempotent across retries, so
distinct prompts sharing a long prefix are no longer collapsed
- Every write is fail-open on frozen or proxied bodies
Saver order and X-9Router-Token-Saver: off behavior are unchanged.
Fixes#3202.
muse-spark-1.2-contributor-free returned HTTP 500 on /zen/v1/chat/completions.
The model is only served by /zen/v1/responses, so route it there via a per-model
targetFormat and normalize the Chat fields the Responses API rejects
(max_tokens -> max_output_tokens, reasoning_effort -> reasoning{effort,summary}),
clamping max/ultra down to the highest effort the model accepts (xhigh).
Routing stays per-model: the other free models (big-pickle, hy3-free, mimo,
nemotron, laguna) are not served by /responses and keep /chat/completions.
`document.fonts.ready` resolved before the 4MB Material Symbols woff2 even
started loading — it runs in <head>, ahead of any element that would trigger
the lazy fetch. The `fonts-loaded` class landed early, so icons rendered
blank until a second load served the font from disk cache.
Load the face explicitly and swap visibility for opacity, with a 3s fallback
so icons never stay hidden if the font fails.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Strict Anthropic-compatible gateways (e.g. MiniMax) reject Claude-format
requests with HTTP 400 when tools[].type is missing. Normalize each
missing/falsy tools[].type to "custom" before dispatch when the final
request format is Claude. Built-in tool types (computer_use, bash,
web_search_*) are passed through untouched.
createSSEStream splits on "\n" and keeps the remainder, which only flush()
parses. That call omitted targetFormat, so parseSSELine required a "data: "
prefix and dropped whatever an NDJSON provider left without a closing
newline. The !parsed.done guard compounded it: the SSE sentinel and an
Ollama final chunk both carry done:true, but the latter is the real last
chunk holding done_reason and the token counts.
Pass targetFormat and scope the sentinel check to formats that emit one, so
the tail reaches the translator. Accumulate its usage into state the same way
the transform loop does, so finalizeStream logs those tokens instead of null.
Route POST /v1/search with provider "antigravity" through Google Search
grounding on v1internal:generateContent, using the existing Antigravity
OAuth account pool. Grounding chunks become citations with the grounded
sentence as snippet and its surrounding answer text as content.
Upstream repeats a source across chunks, so citations are keyed by URL
and their snippets merged. A missing projectId is reported up front —
upstream answers a fabricated or absent project with a misleading
"no valid license" 403.
Based on the approach in #3437 by @Nautilaceae.
The runtime hook pinned better-sqlite3 12.6.2, whose prebuilds stop at
Node ABI 141 — on Node 26 the install fell back to a node-gyp source
build and failed on machines without build tools, silently degrading to
the sql.js fallback.
Node >= 22 now installs 13.0.3, which is N-API and ships per-platform
prebuilds inside the package. Two things were needed to make that
actually work:
- npm injects an implicit `node-gyp rebuild` for any package shipping a
binding.gyp, so the install still demanded build tools; `--ignore-scripts`
skips it and uses the bundled prebuild as-is.
- the binary check only looked at build/Release, which 13.x no longer
creates, so every start re-ran npm install; it now also accepts
prebuilds/<platform>-<arch>.node.
Node < 22 stays on 12.6.2 (13.x requires Node >= 22), and an existing
working install is left untouched either way.
Each card kept its own copy of the localStorage preset logic inside
BaseUrlSelect, so an endpoint saved on one card was invisible to the
others until a reload, and a URL typed into the custom field was
forgotten the moment the card collapsed.
Move the store into cliEndpointPresets.js and publish a change event
so open cards resync live. Applying settings now remembers the
endpoint unless it matches a built-in option, and each card passes
its configured URL as currentUrl so BaseUrlSelect can preselect the
matching preset instead of always falling back to 127.0.0.1.
Deleting a preset falls back to the first real option rather than
clearing the field.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Responses-API clients (codex, droid) close the socket on
response.completed because the protocol has no [DONE] sentinel, so
every successful request printed "⚡ DISCONNECT: ResponseAborted"
after its own "📊 done" line. Keep the dbg("CTRL", …) trace and drop
the console line; ABORTED and ERROR still print.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The Responses API has no [DONE] sentinel, so codex closes the socket
as soon as response.completed arrives. That cancels the reader before
flush() runs — and flush() held every usage side effect, so a fully
successful request logged nothing: no 📊 done line, no token stats,
no request detail.
Extract that tail into a once-guarded finalizeStream() and also call
it right after the terminal event is forwarded, in both passthrough
and translate mode. flush() still calls it; the guard makes the
second call a no-op. Streams that end normally are unaffected, and a
terminal event carrying no usage falls through to the existing
estimate/null path rather than blocking.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Codex only authenticates a custom model provider from env_key,
http_headers, env_http_headers or a token command — auth.json is
read solely by the built-in openai provider. Writing OPENAI_API_KEY
there left every request unauthenticated (401 Missing API key) while
clobbering an existing ChatGPT login.
Put the key in [model_providers.9router.http_headers] instead, and
drop the auth.json write. Also move the subagent model to the
agents.default_subagent_model scalar: agents.<role> now declares a
custom role and requires a description, so the old [agents.subagent]
table was discarded with a startup warning. DELETE still clears
auth.json to repair machines configured by the previous version.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
collectEntries() computed each model's "current" capabilities with the
previous catalog still installed, so every delta was measured against the
last one. An upstream value that still agreed with what we had written
looked like no change and was dropped: the second run cut `providers`
from 20 entries to 5, taking glm-5.3's 1M context correction with it.
The baseline has to be the hand-written tables alone, so the reader is
detached for the snapshot and restored in a finally — a mid-sync failure
must not leave capabilities.js without it.
Two smaller corrections:
- One vote per provider in the modality tally. Ids that normalize to the
same model (claude-opus-4-thinking:1024, :8192, :32768 …) were each
counted, giving nano-gpt five votes where other gateways had one. No
model's result actually flipped — the variants agree with each other —
but the majority rule only means something if the denominator does.
- Restore the etag from the file on startup. It lived only in module
state, so every restart re-downloaded 4.3MB to be told nothing changed.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The worker resolved its own path through import.meta.url, which the
bundler rewrites — so the running server looked for the file at a path
that does not exist there:
[modelCatalog] sync failed: Cannot find module
'/Users/Working/router4/9router/src/lib/modelCatalog/worker.js'
It was guarding against a 23ms JSON.parse that runs once a day, 60s after
boot. Inlining it into sync.js costs that 23ms on an otherwise idle tick
and removes both the failure mode and a whole file.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Extract Spark rate limit windows from the Codex usage response and expose
them as spark_session/spark_weekly quotas, reusing the existing prefix
mechanism. Map codex quota types to readable dashboard labels.
Fixes#3431
Capability tables are hand-maintained, so a model gains vision or a wider
context only when someone notices and edits the file. This adds a daily
sync that fills the gap for models already in the registry.
How it decides:
- Modalities (vision/pdf/audio/video) belong to the MODEL — every gateway
serving glm-5.3-flash serves the same weights — so they are keyed by
model id and shared. A majority of sources must declare one, which keeps
out lone mis-declarations: minimax-m2.5 (1 of 45), glm-4.7 (1 of 44) and
gpt-oss-120b (2 of 76) are text-only despite a reseller claiming vision.
- Context/output limits belong to the GATEWAY — each truncates differently
(glm-5 ships as 202752/16384 on one host and 204800/131072 on another) —
so they are keyed by provider + model and only the matching provider's
own numbers are trusted.
Both layers are strictly additive and sit BELOW the hand-written tables,
which short-circuit first. A capability already true stays true.
Mechanics: worker thread (the 4MB parse would block the loop ~20ms),
ETag so an unchanged catalog costs one empty request, 60s startup delay,
30min backoff on failure, MODEL_CATALOG_SYNC=off to disable. Only the
~57KB delta is kept; lookups cost ~0.1us via an mtime-guarded cache.
capabilities.js is bundled into the browser through useModelCaps, so it
cannot import node:fs — the server injects the reader via
setCatalogSource() from instrumentation.
visionPatterns.js is the last resort: a model nobody has catalogued yet
still accepts images when its id says so (qwen3-vl-plus, glm-4.6v, llava),
with image-generation and embedding ids excluded.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Vendors shipped four multimodal models the registry did not carry:
- glm-5.3-flash — z.ai's first natively multimodal GLM-5, 1M context,
image + video + pdf input (glm, glm-cn, opencode-go)
- deepseek-v4-flash-vision-exp — image input at V4-Flash text parity,
1M context / 384k output (deepseek, opencode-go)
- grok-4.6, grok-4.5 — 500k context; 4.6 has no text output limit (xai)
Capabilities needed hand entries because the existing globs mis-matched:
*glm-5* and *deepseek-v4* carry no vision, and *grok-4* would have capped
grok-4.6 at 256k instead of 500k. The grok-4.6 pattern sits above the
generic *grok-4* so it wins the first-match lookup.
Also corrects glm-4.6v / glm-4.5v, which were missing video input and
declared no maxOutput, and backfills glm-4.6v on glm-cn — zhipuai serves
it and the sibling provider already listed it.
tests/unit/opencode-go-models.test.js pins the opencode-go model list, so
its expected array moves with the registry.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
On a 409/429 from Antigravity, fetch live quota to learn the exact
per-model resetAt instead of guessing a backoff, then skip only the
exhausted account/model pair until that time.
- antigravityQuota.js: in-memory quota cache, coalesced concurrent
refreshes, 30s throttle per connection (applied to failures too),
keeps known cache when upstream returns 401/403 error payloads
- auth.js: pre-filter exhausted account/model pairs; report the
earliest quota reset when every account is blocked; skip the
30-minute cooldown cap so the upstream resetAt is not truncated
- chat.js: antigravity 409/429 falls back on the RAM cache only, no
persistent modelLock_* for this path
- Logs identify accounts by id prefix, never email or name
Closes#3561
Xquik needs a GET request with x-api-key auth and a tweets envelope
normalizer, neither of which the generic search fallback provides. Adds a
dedicated request builder and normalizer, cursor pagination passthrough,
result-based credit usage reporting, and a validateUrl probe so key
validation hits the no-charge credits endpoint.
The openai-responses branch of compressWithHeadroom returned null without
recording a reason, leaving the diagnostics panel blank and making Codex
translation failures indistinguishable from a successful compression.
Add a "Bulk Add" flow for the grok-cli provider, mirroring the existing
codex one: paste a JSON array/object or drag-drop multiple .json files,
then create all OAuth connections in a single call.
- BulkImportGrokCliModal: flexible JSON parsing (array, single object,
{accounts:[...]}, concatenated objects) + multi-file upload
- POST /api/oauth/grok-cli/bulk-import: serial createProviderConnection,
snake_case/camelCase token fields, email backfilled from id_token or
access_token, authMethod "device_code" to match the login flow
## Features
- **Auth**: native SAML 2.0 SSO alongside OIDC — AuthnRequest generation, ACS
assertion handling, SP metadata export, admin config test, replay-protected
via a `saml_state` cookie matched against `InResponseTo`
- **Providers**: add Alibaba Token Plan (`token-plan.ap-southeast-1`) — the
fourth Alibaba key type, Singapore-only and OpenAI-compatible transport only
- **Providers**: add `glm-5.3` to GLM Coding and GLM (China)
- **Providers**: Kimchi accepts API keys as well as OAuth (dual auth), with a
working Test Connection for both modes
- **Antigravity**: add Gemini 3.7 Flash and its tiered high/medium/low variants
(also in the Gemini registry) with pricing and quota tracking
- **TTS**: add Fish Audio — model id travels in an HTTP `model` header, voice
is a `reference_id` (preset or cloned voice model)
- **OpenCode-Go**: route by request format via declared transports instead of
forcing every client into `/messages` — Codex/OpenAI clients no longer pay a
lossy Responses→OpenAI→Claude double translation. Per-model `supportedFormats`
guard; the bespoke executor is gone (its shared `_lastModel` cache could cross
auth headers between concurrent requests)
- **Usage**: dedup + cache Claude quota calls (120s TTL keyed by access token,
in-flight promise dedup, last-good read on soft failure) to stop multiple
tabs tripping 429; manual refresh (↻) sends `force=1` to bypass the cache
## Fixes
- **Docker**: ship `sql.js` in the image so the pure-JS DB fallback can start —
file tracing carried the package's JS without `dist/sql-wasm.wasm`, so a
container with no native driver aborted with ENOENT and never got a database
(#3248)
- **Usage**: read Gemini `usageMetadata` out of the antigravity `{ response }`
envelope — every non-streaming antigravity request logged `IN 0 | OUT 0`
(#3260)
- **Claude**: re-anchor passthrough cache breakpoints — the client's own
`cache_control` markers point at pre-normalization offsets, so the tail was
re-cached every request. Last system block and last tool pinned at 1h TTL,
last assistant turn at 5m, mid-conversation system messages folded into the
neighbouring user turn instead of hoisted into `body.system`
- **Combos**: detect images from Hermes and attachment payloads (`images[]`,
`experimental_attachments`, message-level `image_url`/`audio_url`, inline
`data:` URIs) so the Vision Adapter auto-switch fires for Hermes/Ollama/
Vercel AI SDK shapes
- **Kiro**: intercept chat via `x-amz-target` — Kiro IDE 1.0.228+ moved
`GenerateAssistantResponse` to `POST /` + header, bypassing MITM. Also emit
the now-mandatory initial-response frame and map the `auto` model slot
- **Kiro**: report real output tokens and stop discarding usable turns
- **Qoder**: detect billing blocks at stream start and return a synthetic 403
so combo/account fallback triggers instead of leaking the error into chat
- **Antigravity**: strip competitive system prompts (Zed IDE's Claude-agent
prompt) that Antigravity flags with a 429 Quota Exhausted
- **OpenCode**: send the official client fingerprint on free-tier requests so
the Console stops classifying traffic as unidentified and rate-limiting it;
session id resolves conversation-stable to preserve prompt caching
- **Responses**: don't close the message on an empty `tool_calls` array — some
providers attach one to every chunk, and the truthy check ended the message
on the first content token (#3234)
- **Translator**: preserve `prompt_cache_key` when converting chat to responses
- **Models**: expose snake_case token limits on `/v1/models`
- **Combos**: strip `stream_options` from the Fusion panel fan-out to avoid a
DeepSeek 400 (#3024); raise the dashboard model-test probe budget to 1024 and
soft-pass reasoning-only responses (#3010)
- **Headroom**: the toggle reflects the `headroomEnabled` setting even when the
proxy is down — it previously showed OFF while the engine kept calling
`/v1/compress`; proxy status stays visible via the status chip
- **Hermes**: add the `api_key` parameter to the model block in YAML config
- **Providers**: add llm7 to provider test support
## Docs
- **i18n**: add Spanish, French, and Brazilian Portuguese README translations
## Security
- **Real IP**: `x-9r-real-ip` and the Host fallback were trusted from
client-controlled headers whenever `custom-server.js` was not in the request
path (`npm run start`, `start:bun`), letting a remote caller pose as local to
skip API key auth and reach `LOCAL_ONLY_PATHS` (`/api/mcp/*`,
`/api/tunnel/enable`, `/api/auth/reset-password`). The server now stamps a
per-process `x-9r-peer-token` on every request it sanitizes and only trusts
`x-9r-real-ip` behind it — falling back to Host in development and failing
closed in production (GHSA-pjm4-8fpg-f9p6). Also fixes IPv6 loopback
detection (`::1`, `::ffff:127.0.0.1`) and routes `npm run start` /
`start:bun` through `custom-server.js`
- **Search**: `resolveBaseUrl()` rejects client-supplied non-public baseUrls
(SSRF guard on `/v1/search`)
- **Login**: fresh-install remote login with the default password returns 403
without issuing a JWT
- **Usage**: `/api/usage/request-details` redacts request/response payloads
opencode-go hard-coded targetFormat: claude per model, so every client
format was force-routed to /messages (Codex/OpenAI clients paid a lossy
Responses->OpenAI->Claude double translation). Declare the existing
upstream multi-endpoint transports [openai, claude, openai-responses]
and guard per model via registry supportedFormats: kimi/glm/mimo only
support /chat/completions, minimax/qwen add /messages, deepseek adds
/responses. Undeclared models keep the upstream default.
Drop the bespoke OpenCodeGoExecutor (its shared _lastModel cache could
cross auth headers between concurrent requests); DefaultExecutor already
consumes runtimeTransport and injects reasoning content.
Next file tracing follows JS imports, and sql.js loads dist/sql-wasm.wasm by
path at runtime, so the standalone output carries the package's JS without its
wasm binary. When both native drivers fail the last-resort adapter then aborts
with ENOENT on the missing binary and the container never gets a database.
The CLI bundle already guards this explicitly (build-cli.js step 3b,
ensureModuleInBundle("sql.js")); the image just never got the same treatment.
Copy the package the same way node-forge and next already are.
Fixes#3248
Antigravity and gemini-cli wrap their payload in { response: {...} }.
extractUsageFromResponse only tested top-level usageMetadata, so every
non-streaming antigravity request logged zero usage (IN 0 | OUT 0) and
zeroed rows in the usage dashboard. Read the envelope the same way
usageTracking.js and nonStreamingHandler.js already do; top-level
metadata keeps priority and the OpenAI/Claude branches are untouched.
Fixes#3260
x-9r-real-ip and the Host fallback were trusted from client-controlled
headers whenever custom-server.js was not in the request path (npm run
start, start:bun), letting a remote caller pose as local to skip API key
auth and reach LOCAL_ONLY_PATHS (/api/mcp/*, /api/tunnel/enable,
/api/auth/reset-password).
custom-server.js now generates a per-process secret at boot and stamps it
as x-9r-peer-token on every request it sanitizes. hasTrustedPeerHeaders()
(src/lib/auth/trustedPeer.js) gates trust in x-9r-real-ip on that secret;
otherwise the guard falls back to Host only in development, and fails
closed in production. Same gate on loginLimiter.getClientIp() so a spoofed
header cannot rotate the login lockout bucket.
Also: fix isLoopbackHostname for IPv6 (::1, ::ffff:127.0.0.1) which the
old split(":")[0] reduced to empty string; route npm run start /
start:bun through custom-server.js (postbuild copies it into
.next/standalone, build-cli.js fails without it) so documented deployments
keep passwordless local access.
Fourth Alibaba key type — Coding Plan (alicode/alicode-intl) and Model Studio
(alims-intl) both reject Token Plan keys. Registry entry only; PROVIDER_MODELS
builds from providers/registry so no executor or translator work is needed.
Singapore-only (eu-central-1 answers IllegalEndpoint) and OpenAI-compatible
transport only (the Anthropic surface is not authorized for this plan).
Closes#2754Closes#2806
Add gemini-3.7-flash and its tiered high/medium/low variants to the
Antigravity and Gemini registries, with matching capabilities, pricing
and Antigravity quota tracking.
extractModel now recognises gemini-3.7-flash-tiered alongside 3.6 and
derives the version from the request, so thinkingLevel still maps to the
right tiered alias.
Closes#3286Closes#3281
Registry entry plus one config-driven FORMAT_HANDLERS handler. The model id
travels in an HTTP `model` header rather than the JSON body, and the voice is
a reference_id (preset or cloned voice model).
Closes#2411
Zhipu released GLM-5.3 on both api.z.ai and open.bigmodel.cn coding
endpoints. Verified live against both, returning model:"glm-5.3" with
native reasoning_content.
No other changes needed: the '*glm-5*' family pattern in capabilities.js
and 'glm-5*' in pricing.js already cover it.
Passthrough kept the client's own cache_control markers, which point at
pre-normalization offsets. Once normalize/dedupe reshaped system and tools,
the breakpoints landed mid-array and the tail was re-cached every request.
- Pin the last system block and last tool at ttl 1h (was the client's 5m)
- Anchor the last assistant turn at 5m, falling back to the final message
so a first turn still gets a breakpoint
- Fold mid-conversation system messages into the neighbouring user turn
instead of hoisting them into body.system, where the volatile token
counters invalidated the prefix on every request
- Run the anchoring after every token saver, at the final body
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Inspect images[], experimental_attachments/attachments, message-level
image/image_url/audio_url, and inline data:image|audio|pdf URIs on trailing
user turns so Vision Adapter auto-switch fires for Hermes/Ollama/Vercel AI
SDK shapes. stripOpenAI now also drops msg.images and image attachments when
the active model lacks vision support.
