Relay:
- Each response holds 8,000 audio input and 4,096 audio output tokens of
the account's budget, at the session model's rates, from
response.created until its usage is recorded. A session opens only if
one hold fits.
- Session attempts are limited to 60 per minute per credential, checked
before authentication.
- A response the client requested that starts after the client leaves is
cancelled and billed.
- Input transcription is priced at the costlier of the transcription
models in force at commit and at completion.
- rate_limits.updated isn't forwarded, and provider organization IDs are
removed from error messages.
SDK (puter.ai.convo):
- microphone.stop() during the permission prompt rejects the pending
start() with session_closed.
- speaker.start() rejects with { code, message } for browser exceptions.
- connect() waits 20 seconds for the session, past the server's 15-second
setup.
- Convenience events fill IDs the provider leaves out; types cover
turn.done, error, and client events for send().
Relay:
- Upstream messages may be up to 4 MiB; client messages stay at 64 KiB.
An OpenAI event above 64 KiB (an echoed item near the client limit, a
long reply, a retrieved audio item) closed the upstream before in-flight
responses reported usage.
- An upstream that closes with a response still in flight raises a
warning alarm.
- A null usage counts as none instead of closing the session.
- Seats whose team requires 2FA need it for realtime too, at connect and
on the credential recheck.
SDK (puter.ai.convo):
- A reconnect reapplies every acknowledged setting, native session.update
events included, split to fit the message limit.
- connect() during a reconnect settles with it.
- Reconnect delays are jittered, and closing cancels a pending delay.
The credential-recheck and browser worklet tests now fail when the
behavior they cover breaks.
Relay:
- Cancel and bill a response still generating when the client leaves,
instead of cutting the upstream before response.done.
- Record realtime usage through the AI cost-factor hook.
- Drop the upstream byte and message caps; replies arrive faster than real
time and tripping them closed sessions and skipped billing.
- Price usage as it arrives, before any forwarding decision, so a closing
session never drops a usage event.
- Bill the model the session reports and refuse a session.update that
changes the model.
- Require a paid plan.
SDK (puter.ai.convo):
- Reconnect after a dropped connection with the same settings, keeping
listeners and media; emits reconnecting and ready { reconnected }.
- Errors reject only the operation they name; a failed send rejects only
its own request.
- configure() sends only the given settings and settles on its own ack;
commitAudio follows turn detection set natively.
- interrupt() right after sendText cancels the reply once it starts.
- Microphone: capture at the device rate, drop chunks instead of stopping
on stalls or a congested socket, no timeout on the permission prompt,
permission_denied on refusal.
- Speaker: 300 s playback bound without interrupting, played position
accounts for output latency and gaps between chunks.
- Audio control events keep headroom over queued audio.
- Upgrade prompt on subscription_required / insufficient_funds.
Tests that need the built SDK and Chromium move to the puter.js suite.
* feat(events): simplify onPersistent delivery
Replace delivery/targets with a boolean single. Without single, every
event runs the app's published handler in its events worker and is also
sent to every connected client; the server no longer routes between the
two. single keeps leasing each event to one consumer that acks.
delivery: 'single' | 'broadcast' is still accepted and targets is
ignored. Stored rows are read by their delivery column alone.
* docs(events): compare onLocal, onPersistent and single
* feat(events): rename single to queue
onPersistent({ queue: true }) is the work-queue mode: each event goes to
one consumer at a time and is offered again until acknowledged.
delivery: 'single' is still accepted as an alias, and subscription
objects report queue instead of single.
The docs explain the three modes (onLocal, default onPersistent, queue)
and each events recipe says which one it uses and why.
* docs(events): retries are the only repeat for a queued event
* feat(events): queue deliveries fall through to the worker on decline
A queued delivery gets one socket attempt under a 15 s lease instead of
two 60 s ones. A client with no handler for it, or whose handler fails,
sends events.nack and the worker takes it without waiting for the lease.
Declines are relayed to the region holding the lease like acks are, and
a decline for a lease that is no longer the client's is ignored.
* feat(events): drop thrown errors, retry only temporary ones
A queued handler that throws now drops the event and counts one failure
toward the pause. Network errors, timeouts, rate limits, 5xx responses
and errors thrown with retry: true are retried with the existing
backoff. One classifier is shared by puter.js and the events worker
runtime, which inlines it at build time. A dropped event is billed once,
like a completed one. ack() remains as an optional early release.
* docs(events): queued events reach every tab with a handler
* fix(kv): sign share handle ids and drop retired handles from every node's cache
New handle ids carry a short HMAC (kvh-<uuid>.<mac>); an id whose mac
does not verify is refused before any lookup. Unsigned ids keep working
but no longer fall back to the primary on a replica miss.
Retiring or changing a handle now broadcasts outer.pubsub.kv.share.forget
so every node drops its cached row. Event subscribes and delivery
rechecks refuse a handle whose owner is suspended, as the driver does.
* fix(kv): an app its user removed stops reaching shares
An app acting for a user exercises a share grant only while the user
still has the app authorized: at least one user-to-app permission row,
which every app token path leaves and revoke-user-app '*' removes. The
share itself stays; authorizing the app again restores access. The
answer rides the permission cache, keyed by the app-under-user
generation the revoke bumps, so live subscriptions go quiet too.
* feat(kv): list keys through a watch-level share handle
list with as: 'keys' through a handle needs only watch, which already
learns every key written in the region. Values and entries still need
read.
* feat(kv): filter share listings by app and by state
GET /kv/shares and GET /kv/shares/with-me take an optional appUid: an
app uid, or null for the app-less store. An account session may name
any app; an app is refused any namespace but its own. GET /kv/shares
also takes state=live|revoked|all, all by default.
* feat(puter.js): appUid and state options on the KV share listings
listShared() and listSharedByMe() pass appUid (null for the app-less
store), and listSharedByMe() passes state. Docs cover the new options
and that a watch share can list keys without their values.
* feat(kv): indexes and store counts for share-handle quotas and retention
Adds covering indexes for the owner's live count, the owner-holder pair,
app-filtered listings and retired rows, plus store methods to count a
pair's live handles and delete rows retired before a cutoff. The live
handle cache holds up to 50,000 entries; the miss cache stays at 10,000.
* feat(kv): higher share-handle caps, a per-holder cap, and sharing with several recipients
Live share handles per account go to 25,000 paid / 5,000 free, and per
app to 10,000 / 2,000. One owner may hold out 256 live handles to one
holder on every plan, which bounds the rows a revoke or unshare walks.
Re-sharing an existing region takes no slot and is never refused.
POST /kv/shares takes recipients (1 to 50) as an alternative to
recipient: one region, level and label for all. Caps and a new
per-recipient grant budget (300/min, 2,000/hour) are checked for the
whole batch up front; failures about one recipient are reported on its
item. The single-recipient request and response are unchanged and run
through the same code path.
* feat(kv): delete share handles 30 days after they stop
A daily job deletes retired share-handle rows older than 30 days, in
batches of 500. One node runs it per day; it stops on shutdown.
* feat(puter.js): share a KV prefix with several users in one call
puter.kv.share() takes an array of recipients, positionally or as
recipients in the options form, and resolves to one item per recipient:
the share, or an error with its message and code. The single-recipient
forms are unchanged.
* feat(kv): delete an app's data on uninstall, and close account-deletion gaps
Uninstall:
- POST /auth/revoke-user-app takes an optional `deleteData` boolean, only
with `permission: '*'` (400 otherwise). AuthService.uninstallApp revokes
every grant, deletes the app's worker sessions, then awaits
`app.uninstalled` { userId, userUuid, appUid, deleteData }.
- With deleteData, KvShareService retires the handles the user shares out
of that app's namespace (as the platform, in batches) and then empties
the namespace in the background, 5,000 items per step, stopping on
shutdown. Handles shared with the user are not touched.
- SystemKVStore.purgeNamespace(namespace, { maxItems }) deletes up to
maxItems items of one `v1:<uuid>:<appUid>` namespace, reading keys only,
with no mutation events or usage, and reports `done` once it is empty.
It refuses the system namespace and anything else. flush shares its
page-delete path.
- puter.perms.revokeApp(appUid, '*', { deleteData: true }); the desktop's
uninstall dialog has an unchecked "Also delete this app's data" option.
Account deletion:
- EventsService settles each of a deleted account's apps on `user.delete`
(they cascade without `app.changed`) and drops the owner's subscription
keys.
- The flat permission entries the account held are purged in pages.
Suspension:
- `user.suspended` / `user.unsuspended` events; the share-handle store
drops cached rows owned by that user here and on other nodes.
* fix(permission): await a peer's flat-permission deletes so emitAndWait sees them applied
* fix(kv): strongly consistent purge pages, and an uninstall purge stops once the app is authorized again
A stale empty page could end a purge as done while items remained. The
background purge after an uninstall now checks before each step that the
user hasn't authorized the app again, so data written after that is kept.
* fix(kv): flush keeps eventually consistent page reads; only a purge reads strongly consistent
* fix(user): emit user.suspended and user.unsuspended when the user store changes an account's suspension
* feat(kv): share key-value prefixes with other users, and bound what apps can be granted
- KV share handles move out of Events into KvShareService, served at /kv/shares
(the /events/kv-handles routes stay as an alias).
- Shares have watch, read and write levels, encoded in the permission root
(kv-share, kv-share-read, kv-share-write) with write implying read and watch.
- Every puter.kv method takes optConfig.handle; keys are relative to the shared
prefix, cursors are sealed, and private entries stay hidden.
- New puter.kv.share, unshare, listShared and listSharedByMe; optional label;
kv.share.received notice in the app's mailbox only; recipient blocks apply.
- puter.kv.set gains onlyIfAbsent.
- fs and kv events gain viaWorker, and kv events report self from the writer.
- /auth/grant-user-app only accepts canonical permissions from a fixed set of
families; the consent dialog, puter.perms and docs follow.
- Migrations: kv_share_handles gains revoked_by, label, and grantee and app
indexes (mysql_mig_49, postgres_mig_38, sqlite 0095).
* fix(kv): never name a private key in a cross-app or handle cursor, and bound flat-permission scans
- Listings of another namespace (cross-app or through a handle) resume after
the last entry they returned, and their cursors are sealed, bound to the
namespace or handle, and padded to one length. Old plain cross-app cursors
are still read when they name the target namespace; other namespaces get 400.
- PermissionStore scans flat permissions a few roots at a time and skips
roots nested under another root in the same call.
* docs: easy events recipe (PUT-1961)
* docs: catch-up-on-notifications events recipe (PUT-1961)
* docs: events recipe corrections from a run against a live server (PUT-1961)
* docs: medium events recipes (PUT-1971)
* docs: simpler wording in medium events recipes (PUT-1971)
* refine
---------
Co-authored-by: Daniel Salazar <daniel.salazar@puter.com>
Cache writes now meter under `cache_write_tokens` at the model's write
rate, split out of `prompt_tokens` by the shared usage splitter, so K3 gets
the same breakdown. The credit hold reserves the prompt and any replayed
reasoning at the input rate, and a stream that ends without a usage chunk
is charged the estimate like every other provider. The per-TTL response
headers, the 502 accounting error and the one-hour write prices are gone:
the cache is scoped to the account and nothing here requests a one-hour
TTL, so those writes never occur and Moonshot publishes no rate for them.
The catalog entries carry the release date so responses normalize by
default, the provider suite no longer imports the SDK, and the docs cover
the reasoning efforts, replay and cache usage keys in their existing
sections.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
* feat(ai): MiniMax video generation (H3 and Hailuo)
Adds a minimax-video-generation provider to the video driver. H3 models go
through MiniMax's v2 task API and are billed per second of output; Hailuo
models go through the v1 API, exchange a file id for a download URL, and
are billed per clip. The shared `minimax` chat key is reused, with the
chat block's `/v1` base URL reduced to its host.
WIDTHxHEIGHT sizes now map onto tiers for catalogs that include a 2k or
4k tier.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
* fix(ai): validate MiniMax video inputs and resolution tiers
---------
Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>
* docs: easy events recipe (PUT-1961)
* docs: catch-up-on-notifications events recipe (PUT-1961)
* docs: notification examples use notif:app-user, which apps can read
* docs: events recipe corrections from a run against a live server (PUT-1961)
* refine and update agents md
---------
Co-authored-by: Reynaldi Chernando <reynaldichernando@gmail.com>
- Providers send upstream calls with a driver-owned `upstreamSignal`
instead of the caller's `abortSignal`. A disconnect detaches the chat
stream: output stops, the provider keeps reading until the upstream
reports usage, and that usage is metered.
- A detached stream gets SDK_TIMEOUT_MS to finish. After that the driver
aborts the upstream, and the stream is metered like any other stream
that ends without usage.
- The disconnect check also runs right before dispatch.
- The OpenAI stream handlers report usage that arrived before an abort.
- Docs: stopping a stream stops the output, and a request that reached
the model is billed in full. Detached streams get 10 more minutes.
* refactor: let core run without a teams registration
Teams becomes optional: `services.team` and `stores.team` may be absent, and
core stops assuming either is there.
Seventeen call sites across share, signup, account deletion and the homepage
now degrade instead of throwing: a team recipient that cannot be resolved is
"no such recipient" rather than a fault, a group share folds in nothing, and
account deletion stops asking what a seat was charged for. This generalises
the pattern `gates.ts` and `SocketService` already used.
Teams' own files take a private handle to their store and service, so the
feature's code stays free of optionality that only core needs.
No behaviour changes where Teams is registered.
* refactor: put a contract between core and teams
`core/teams.ts` holds the shapes core reads and the calls it makes, and the
registries declare the pair as provided rather than built in. The service,
store and controller are registered by whatever runs the feature.
Typing them properly turned up eight AuthController call sites the previous
commit could not see: the controller reads its stores from the built-in
registry, so an absent `team` fell to an index signature and typed as
`unknown` instead of erroring. They now go through the same optional path as
the rest.
The feature's own tests move with it. The cases in core suites that build a
team to exercise a core rule still need relocating, so those suites are red
here.
* test: stub core's seat rules, rehome the rest with the feature
A core suite can no longer build a team, so the rules that refuse a seat get
`testFixtures/seatStub.ts` instead: it answers the contract, which is all
those rules read, and puts the registry back when the test ends.
Seventeen cases move that way. Four could not: group shares and a seat's 2FA
enrolment over HTTP need the real feature and the `/teams` route, so they go
with it, as does the whole of `GroupGrants.test.ts`.
Two whoami cases asserted a seat's team through `extension.import`, which
resolves the registry rather than the stores object. They now assert the
absent case, and the behaviour they covered is asserted against
`buildUserDetails`, which takes its stores as an argument.
Suite: 12252 passed, 38 skipped.
* refactor: take the teams console out of the built-in tabs
`TabTeams`, `teamsConsole` and `teamActionIcons` are no longer built in. The
dashboard already invites a tab through its `dashboard-will-open` event, so
whatever provides Teams contributes the console the same way.
Its 113 strings go with it and register into the shipped English dictionary
at load, so `i18n('teams_*')` resolves as it did. Only `en.js` ever carried
them, so no translation is lost.
Three modules stay, because things other than the console use them:
`teamBadge` (the dashboard shell), `orgSeat` (Home, Account, Security) and
`shareTeams` (the share picker). They already render nothing when the
account has no team. `teams_account_of` stays with `teamBadge` for the same
reason, and the `share_*` and `plan_team_*` strings stay with their callers.
The GUI builds clean with no console at all, and builds it in when the tab
is contributed.
* docs: seats past the owner's current limit get no allowance
* fix(fs): support offset and byte_count in read operations
Allow partial content reads via offset and byte_count parameters in puter.fs.read and backend FS controllers.
- In puter.js read operation, check options.offset and options.byte_count against null/undefined so offset: 0 is preserved in query params.
- In handle_resp, accept any 2xx status code so 206 Partial Content responses resolve successfully.
- In backend FS controllers (LegacyFSController and FSController), derive the HTTP byte range from query offset/byte_count parameters when an explicit Range header is not present.
- Add unit and integration tests across controllers and SDK test suites.
* fix(fs): strict range params, 416 past EOF, drop format churn
- resolveReadRange accepts only decimal digit strings and safe integers
- readContent maps S3 InvalidRange to 416 instead of a 500
- revert unrelated union-type reformatting, restore truncated test comment
---------
Co-authored-by: Daniel Salazar <daniel.salazar@puter.com>
* feat(ai): add Kling for video, image and speech generation
One `kling` API key now registers a provider on three drivers:
- txt2vid: Kling 3.0 Turbo, 3.0, 3.0 Omni, O1, 2.6 and 2.5 Turbo,
with first/last frames, reference images on the omni models, native
audio, and editing a source video on 3.0 Omni and O1. Billed per
second at the resolution, audio and video-input rate; an edit is held
at the longest source and billed for the clip Kling returns.
- txt2img: Kling Image 3.0, 3.0 Omni, O1 and 2.1, billed per image.
- txt2speech: Kling TTS with its English and Chinese voices, billed per
request.
Kling's business error codes map onto the driver error contract, so an
exhausted resource pack reports as credit exhaustion and a submit-time
content block as `moderation_flagged`. "Risk control" failures now
count as content-filter refusals.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(ai): treat Kling's unknown-model codes as a Puter-side failure
Kling answers 1202/1203 when the method or model isn't one it serves.
Puter picks both, so these meant catalog drift but surfaced as a 400 that
blamed the caller and raised no alarm. They now return 502 upstream_failed
with a warning alarm.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(ai): tighten Kling input checks and error mapping
- 3.0 Turbo takes at most 2,500 prompt characters with a first frame,
the image-to-video endpoint's limit.
- O1 refuses reference images together with a first and last frame.
- 3.0 Omni accepts a source of up to 15.5 seconds, and an edit is held
at 16 seconds to cover it.
- A source video Kling can't decode fails as 400 upstream_bad_request
instead of a Puter-side 502.
- A 1201 about the model counts as catalog drift, like 1202/1203.
- Auth failures no longer pass Kling's console wording to the caller.
Adds integration tests for the image, video and TTS providers, gated
on PUTER_TEST_AI_KLING_API_KEY.
* feat(ai): wait up to 14 minutes for a video job
Video jobs, Kling edits in particular, can take over 10 minutes. The
poll window and the Gemini Omni deadline move to 14 minutes, under the
SDK's 15-minute idle timeout, with holds and signed source links
lengthened to match. The keep-alive default moves to 920 seconds so it
stays above a 15-minute proxy idle timeout.
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Daniel Salazar <daniel.salazar@puter.com>
* feat(ai): edit a source video with txt2vid on Gemini Omni
txt2vid takes an `inputVideo` option (sent as `input_video`): a URL, data
URI or Puter file path to an MP4, QuickTime or WebM video of up to 32 MiB.
Gemini Omni receives it inline next to the prompt, which describes the
edit or extension. Google enforces the 10-second limit.
Models declare `videoInput` in the catalog, and the driver rejects a
source video for any other model instead of letting the provider drop it.
The credit hold adds 16,000 input tokens for a source video, since
countTokens reports far fewer tokens than the reply bills.
loadFileInput takes an optional `timeoutMs` for web URLs, and cancels
the body of a failed response.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* feat(ai): edit a source video with FLUX 3 on Together AI
FLUX 3 takes `input_video` as Together's `media.source_video`. A web URL
goes through for Together to fetch; a data URI or Puter file is read with
the caller's access, capped at 32 MiB, sniffed as a video and sent inline.
Editing is estimated and billed at the 1080p rate of $0.54/s, since FLUX 3
jobs report no cost. Image inputs can't be combined with a source video.
Seedance 2.0 ignores `source_video`, treats `reference_videos` as a loose
reference rather than an edit, and Seedance 2.5 fails on reference videos,
so neither declares `videoInput`.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* docs(ai): name FLUX 3 as a video editor in the AI overview
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(ai): match Hailuo 02 to Together's price and document estimate billing
Together quotes $0.56 for a Hailuo 02 clip; the catalog had $0.49.
Together's video jobs no longer report what they cost, so every Together
clip is billed at the catalog's highest published rate. The docs said the
charge was the amount Together reports; they now describe the estimate.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix: enforce video edit durations and propagate source fetch cancellation
* fix(ai): pass media inputs as Puter files and stream source videos
Source videos were read whole into the API process, base64-encoded and
sent twice to Gemini (countTokens and the interaction), which held about
128 MiB per in-flight edit for up to ten minutes.
- Video inputs are a URL, a Puter file path or a data URI of at most
1 MiB. Together gets a Puter file as a 15-minute presigned S3 URL and
fetches it itself; Gemini streams it into the request body with an
exact Content-Length, so it's never held as one string.
- Gemini counts only the prompt and image for a source video, and both
providers read the source only after the credit hold.
- txt2img and txt2vid image inputs accept Puter file paths, read within
the caller's access right before the provider call. Existing URL,
data URI and base64 inputs are unchanged.
- puter.js writes a Blob, File or data URI over 1 MiB to a temporary
file under .tmp, sends its path, and deletes it once the call settles.
- FLUX 3 edits bill at the 720p editing rate, matching their 1280x704
output.
* fix(ai): stage inputs after website sign-in, keep images working without a write
- puter.js chose the temporary file path before signing a website visitor
in, so it fell outside the app's data directory and the upload was
refused. It now signs in first, then picks the path.
- An image that can't be written (no access, storage full) is sent inline
as before instead of failing the call. Videos still fail, since a
large inline video isn't accepted.
- Telling raw JPEG base64 from a path decodes only the first 24
characters instead of the whole input.
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: 404oops <oops404@util.404oops.com>
Co-authored-by: Daniel Salazar <daniel.salazar@puter.com>
Neither direction applied backpressure: a client that stopped reading
grew a queue up to 64 MiB, and a client streaming 16 MiB frames was read
and forwarded as fast as it could send. A few such sockets pinned the
event loop and pushed RSS past 1 GB.
Each side now stops reading the other once 1 MiB is queued toward it,
and a client that keeps the upstream paused for 30 s is closed. Client
uploads are budgeted per session (1 MiB/s sustained, 16 MiB burst,
realtime.maxClientBytesPerSecond), pre-open buffering pauses reads at
1 MiB instead of rejecting at 32 MiB, and clients get a 30 s ping so
half-open sockets release their upstream and slot. Upstream frames use
a zero mask and skip UTF-8 validation, which the relay never needed.
The stats line gains elu. tools/realtime-stress gains abuse.mjs and a
runbook phase for event-loop abuse.
* Add Gemini Omni video generation with token metering
* Keep explicit Gemini video model selection within provider
* Settle reported Gemini video usage before validating output
---------
Co-authored-by: 404oops <oops404@util.404oops.com>
* feat(ai): add Gemini unary transcription
* Clarify transcription language results and omit pricing
* style(ai): keep Gemini transcription sources prettier-clean
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(ai-speech2txt): charge Gemini transcriptions that fail after Google answers
Google bills an interaction it processed whether or not it completed, so an
incomplete, failed or malformed reply was free to the caller and paid by
Puter; audio whose transcript overran the output cap could be resubmitted at
no cost. Usage is now recorded before the reply is checked, falling back to
the counted audio tokens when the reported usage can't be read.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* Preserve transcript text with optional malformed annotations
* Keep speech provider spacing consistent
---------
Co-authored-by: 404oops <oops404@util.404oops.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* fix(ai): keep Gemini speech instructions separate from transcript
* Clarify Gemini speech instructions
* style(ai-tts): keep Gemini cost comment prettier-clean
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* Validate speech requests and settle reported Gemini usage
* Clarify speech settlement for unreadable responses
* Reserve the published Gemini speech token budget
* Clarify speech accounting for bounded responses
* Limit only Gemini TTS text to 3000 characters; charge overrun replies
The 3000-character text cap moves from the TTS driver into the Gemini
provider, so other providers keep their own limits.
The Gemini reply cap rises to 64 MiB, which fits the full 16,384-token
audio budget. A reply that still overruns it or times out reports no
usage, so it is charged the held budget and the error tells the caller
to split the text.
---------
Co-authored-by: 404oops <oops404@util.404oops.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Daniel Salazar <daniel.salazar@puter.com>
* feat(ai): restore available Gemini models and align pricing
* Exclude deprecated Flash Image additions and share scheduled pricing
* Bill Gemini 2.5 Pro at its current list rates
Cache reads are $0.125/M, and prompts over 200k tokens bill every rate at
the long-context tier ($2.50 in, $15 out, $0.25 cached).
---------
Co-authored-by: 404oops <oops404@util.404oops.com>
Co-authored-by: Daniel Salazar <daniel.salazar@puter.com>
* feat(ai): add Gemini 4 Argon with native usage metering
* test(ai): isolate Argon HTTP smoke test from optional extensions
* fix(ai): preserve Gemini native completion and refusal reasons
* Refer model specifications and pricing to the directory
* Bound unary Argon requests and reject unsupported content
* Keep Gemini tool metadata out of Claude requests
* Describe provider-specific chat timeout boundaries
* Keep Gemini extra_content off other providers; report Gemini timeouts as 504
Replayed tool calls only carry extra_content for providers that opt in
(Gemini), so OpenAI-compatible and Responses vendors don't see an unknown
field. The native Gemini client now owns its request deadline: the SDK's
own timeout aborted without a reason and read as an internal error. A
deadline timeout returns 504 upstream_timeout and leaves the route in
rotation.
---------
Co-authored-by: 404oops <oops404@util.404oops.com>
Co-authored-by: Daniel Salazar <daniel.salazar@puter.com>
Daniel Salazar