OpenAI-compatible gateway in front of free AI API providers. Invite-only: Luke and his friends get Aqua keys; provider keys never leave the server.
- PUBLIC to every user: real model names (
/v1/models, themodelfield in responses and stream chunks). Model maker/name is not a secret. - ADMIN-ONLY: which serving provider routes each model. Provider ids, base URLs, upstream model ids, provider headers, error bodies, and failover metadata never reach clients. Errors to clients are generic.
- The public model name never contains the provider id (enforced in config parsing and by test). Serving-provider content inside model prose itself is out of scope for redaction: the guarantee covers routing metadata, not what a model says.
Each provider contributes (public, upstream) model pairs via env:
ZRO_MODELS="deepseek-v4.1-flash=deepseek-v4.1-flash" # public=upstream, comma-separated
AMD_RADEON_MODELS="qwen-3.8-27b=Qwen3.8-27B,deepseek-v4.1-flash=DeepSeek-V4.1-Flash,glm-5.3-flash=GLM-5.3-Flash"
LOGFARE_MODELS="glm-5.3=glm-5.3,qwen-3.8-27b=qwen-3.8-27b,deepseek-v4.1-flash-1017=deepseek-v4.1-flash-1017"
# dated model revisions stay distinct; Flash is not the full model
NVIDIA_NIM_MODELS="deepseek-v4.1-flash=deepseek-ai/deepseek-v4.1-flash,glm-5.3-flash=z-ai/glm-5.3-flash"
Routing is model-first. A specific public model name gathers every eligible
provider carrying that exact model and fails over across those providers only -
a pinned request never drifts to a different model. Unknown names get a generic
400. The reserved auto alias walks AQUA_AUTO_SHORTLIST (ordered canonical
model names) model by model, provider by provider; with no shortlist configured,
auto is disabled rather than delegated to any provider-side router. Cross-model
fallback happens only inside auto.
Two providers carrying the same public name are bindings of one canonical model: the operator must verify the upstream ids really are the same model variant (same family is not the same model) before reusing a public name.
AQUA_MODEL_META attaches verified context windows: "coder-a=128000". The
only local size check is sound: a requested max_tokens larger than the window
gets a generic 400 without an upstream call or quota burn. Prompt size is never
estimated (no tokenizer); unknown context means no rejection.
Provider health is passive: real request outcomes cool an account briefly (15s after errors, 60s after a bare 429), and a provider-sent Retry-After (delta seconds or IMF-fixdate, bounded to 1h) marks the account hard-limited until it lapses: hard-limited accounts are never retried, even as a last resort. Cooldowns are account-scope: a 429 on one model keeps the gateway from hammering that account's other models. No probing or inference calls are made to test availability.
Public names are operator-configured from each provider's actual model list - verify against the provider before adding; do not invent ids.
- Upstream IO carries one aggregate deadline (AQUA_UPSTREAM_BUDGET_SECS, default 300s, clamped 1..3600) shared by every fallback attempt of a client request: a stall on each candidate cannot multiply the total upstream wait.
- Downstream writes carry the same value as a per-write-operation timeout, so a client that stops reading cannot pin a worker (and its in-flight quota slot) forever. Healthy long streams are never cut. There is no whole-request wall clock: these are IO bounds, not a total deadline.
- The SSE rewriter caps one pending (undelimited) event at 256 KiB, independent of how the network chunked it. An upstream that never terminates an event gets its stream cut: the client response headers are already committed at that point, so there is no failover mid-response: the stream ends honestly, without a fabricated [DONE].
- Responses are capped at 4 MiB total, streamed or not.
A "free" label is not proof a route costs nothing. Every provider's default
status is off; enable deliberately with AQUA_STATUS_<ID>=off|personal|friends
only after cost (no credits, no metered billing) and terms are verified. Status
friends means any valid gateway key can use it; personal is owner keys only.
Known concerns recorded for the review trail: openrouter free-tier key sharing for friends not cleared by ToS; zro free plan documented "for your own use"; gonka is a paid metered service.
NIM free inference was verified on 2026-11-16: rate-limited access, no credits consumed and no per-token billing. NVIDIA describes this access as development and prototyping. This does not establish permission to share an account or key. The router honors valid Retry-After headers on rate-limit responses.
Logfare free Standard inference was verified with qwen-3.8-27b on 2026-11-16.
It logs prompts and responses to evaluation datasets (18+, Discord required).
Keep AQUA_STATUS_LOGFARE=personal: no sharing the owner's key with friends.
Friends need their own provider connection before they can use this route.
The GLM full, Qwen and dated DeepSeek bindings above come from the verified catalog;
only Qwen inference is verified so far.
AMD Radeon Token Factory inference was verified on 2026-11-16. Its points
are usage units, not money or purchased credits. Keys expire after 30 days.
The adapter id is radeon, with AMD_RADEON_API_KEY, AMD_RADEON_MODELS
and AQUA_STATUS_RADEON. The retired AMD_TOKENLAB_* variables are not read.
Cloudflare Workers AI remains off until the account is verified to have no
paid fallback. Its free allocation is 10,000 neurons/day across models.
Only GLM 4.7 Flash is configured below, and no model marked
require_workers_paid=true may be enabled. The account-specific base URL
must be pinned by the deployment configuration; this registry entry does
not enforce account billing limits.
AQUA_OWNER_KEYS="luke:<key>" # name:key, comma-separated
AQUA_FRIEND_KEYS="sam:<key>"
AQUA_ADMIN_KEY=<key> # unlocks GET /v1/admin/providers
AQUA_DAILY_LIMIT=200 # per key per UTC day, atomic reservation
AQUA_ALLOWED_ORIGIN=https://aqua.goattech.us # single explicit CORS origin (the Pages console)
AQUA_BIND=127.0.0.1:8080 # listener; loopback by default, set 0.0.0.0:PORT explicitly at deploy
# console login (GitHub OAuth; optional - API keys work without it)
AQUA_GITHUB_CLIENT_ID=<oauth app client id>
AQUA_GITHUB_CLIENT_SECRET=<oauth app secret>
AQUA_PUBLIC_URL=https://aqua-api.goattech.us # this backend's own public URL (OAuth callback host)
AQUA_CONSOLE_URL=https://aqua.goattech.us # where the browser lands after login (Pages)
AQUA_GITHUB_OWNERS=goattech-42 # GitHub logins, lowercase, comma-sep
AQUA_GITHUB_FRIENDS=sam,alex
AQUA_AUTO_SHORTLIST=coder-a,coder-b # ordered models for the auto alias; empty disables auto
AQUA_MODEL_META=coder-a=128000 # verified context windows, tokens
AQUA_AUTH_FAILURE_LIMIT=5 # failed credential attempts per fixed window
AQUA_AUTH_WINDOW_SECS=60 # per-fingerprint window, see docs/AUTH.md
AQUA_AUTH_FAILURE_ENTRIES=10000 # bounded counters; full fresh tables preserve existing entries
AQUA_MAX_CONNECTIONS=64 # total active handlers, clamped 2..1024
AQUA_CONTROL_RESERVE=8 # slots held back from ordinary parsed requests
AQUA_UPSTREAM_BUDGET_SECS=300 # aggregate deadline shared by ALL fallback attempts of one request
OPENROUTER_API_KEY=... ZRO_API_KEY=... etc.
AQUA_STATUS_CLOUDFLARE=off # until no-paid-fallback proof
CLOUDFLARE_AI_BASE_URL=https://api.cloudflare.com/client/v4/accounts/<account-id>/ai/v1
CLOUDFLARE_API_TOKEN=...
CLOUDFLARE_MODELS="glm-4.7-flash=@cf/zai-org/glm-4.7-flash"
Empty or placeholder client keys are dropped at config build and never authenticate. Provider keys are only ever sent to their own provider.
A new auth-failure fingerprint at a full fresh table scans up to the configured entry limit under the shared mutex. Hostile token rotation can add CPU and lock pressure; this is not flood resistance.
By default at most 64 active connection handlers run. The 65th accepted socket is closed before a worker is started, without a response write in the accept loop. After the existing bounded request read (head and body), at most 56 ordinary requests run; excess ordinary requests receive 503. All requests, including console/control requests, remain subject to the total cap.
Eight slots are reserved from ordinary work for valid-session console pages,
/auth/me, logout, key management, settings and admin usage/caps/providers/resolve.
The existing route auth, owner checks and CSRF checks still apply. Inference,
model listing, image generation and unknown routes never use this reserve,
even with a valid session. Public landing, login and callback GETs can use it
to regain access. Cookie-free OPTIONS on those named console/control endpoints
can use the reserve for browser preflight; this grants no action permission.
AQUA_MAX_CONNECTIONS defaults to 64 and is clamped to 2..1024.
AQUA_CONTROL_RESERVE defaults to 8 and is clamped to 1..(total - 1), so
ordinary capacity is always at least one. Missing, invalid, negative or
overflowing integer strings use defaults. Programmatic configurations are
bounded at use too. Changes apply at process start; zero does not disable limits.
This is not an absolute no-lockout or denial-of-service guarantee. Unread or partially read sockets (including partial bodies) cannot be classified and can occupy all 64 handlers during the read phase for up to 30 seconds. Response writes and other work have separate timeouts, so this is not a promise that capacity returns at 30 seconds. Public login/preflight traffic can also fill the total cap. A reverse proxy with its own connection/rate limits is still needed for hostile traffic. No live deployment or environment change is implied by these defaults.
POST /v1/chat/completions- OpenAI shape,streamsupported. Whitelisted params forwarded: messages, temperature, top_p, max_tokens, stop, n, response_format, seed, presence_penalty, frequency_penalty, tools, tool_choice. Everything else (includinguser) is dropped.GET /v1/models- canonical public model names (deduplicated across providers, with verified context when known) visible to the caller's role, plusautowhen a shortlist is configured.GET /v1/admin/providers- admin only: provider ids, statuses, model mappings.
Streaming: SSE parsed per event (multi data: lines joined per spec), each chunk
rebuilt from a whitelist; upstream event:/id:/comment lines and non-SSE 200
bodies never reach the client. Hard 4 MB cap on bytes to the client.
Request framing: HTTP/1.1 with an explicit Content-Length only. Chunked bodies are rejected with 400 (never guessed); duplicate Content-Length is rejected; headers over 64 KB are rejected even when terminated. Quota is reserved only after a request parses and routes: malformed 400s and unroutable 503s never burn quota. Daily quota counters live in memory and reset on restart (beta). Failed auth is throttled per token (5/min, then 429) with a bounded failure map.
Logs: one line per request: timestamp, key name, provider id, stream flag. Prompts and responses are never logged.
Split hosting, per the owner's ask: static console on Cloudflare Pages at
https://aqua.goattech.us, this Rust backend on his own server at
https://aqua-api.goattech.us (Cloudflare-proxied record -> Cosmos -> container).
Same registrable domain, so the host-only SameSite=Lax session cookie flows on
credentialed cross-origin fetches; CORS answers exactly one origin with
access-control-allow-credentials: true, and cookie-authenticated mutations
reject Sec-Fetch-Site: cross-site. The backend still serves the
three pages itself for local dev and same-origin use; Pages is canonical in prod.
The pages pick their API base by hostname (aqua.goattech.us -> the api. host,
anything else -> same origin), with the localStorage override kept for testing.
cargo test # 46 tests incl. hostile-upstream redaction probes
cargo run # serve
This is usage accounting, not billing. Aqua never buys provider credits. Friends start at $2/day and $15/month in virtual USD; the owner is tracked but uncapped. Daily and monthly windows reset at UTC midnight. Prices are list-price benchmarks, not a promise of what a provider charges. Default input/output USD per million tokens, checked 2026-11-16:
| Model | Input | Output | Source |
|---|---|---|---|
| glm-5.3 | 1.40 | 4.40 | https://docs.z.ai/guides/overview/pricing |
| glm-5.3-flash | 0.15 | 0.50 | https://docs.z.ai/guides/overview/pricing |
| glm-4.7-flash | 0 | 0 | https://docs.z.ai/guides/overview/pricing |
| deepseek-v4.1-flash-1017 | 0.30 | 1.20 | https://api-docs.deepseek.com/quick_start/pricing/ |
| qwen-3.8-27b | 0.50 | 3.00 | http://docs.modelstudio.console.alibabacloud.com/en/model-studio/qwen3-8-27b |
DeepSeek uses peak, uncached list prices for that dated identity only. Qwen uses the maker's international price. Unknown models use a conservative 10/30 estimate and are owner-only by default. Output prices of $10/million or more default to owner-only. These settings do not claim that hypothetical future models exist.
Configuration:
AQUA_SPEND_STATE_PATH: persistent local JSON ledger, defaultaqua-spend.json. Requires Rust 1.89+ for its exclusive OS file lock. One server process per ledger. Back up the ledger; deleting it resets usage. A missing parent directory, corrupt ledger or competing process prevents startup. Never put it on ephemeral storage.AQUA_SPEND_DEFAULT_CAP=2/15: daily/monthly virtual USD for friends.AQUA_SPEND_CAPS=sam=2/15,alex=1/10: initial per-account caps.AQUA_MODEL_PRICES=custom-model=0.15/0.50: input/output USD per million.AQUA_MODEL_ACCESS=custom-model=admin: restrict explicitly.allexplicitly permits friends, still subject to provider eligibility and caps.
The admin API has GET /v1/admin/usage and POST /v1/admin/caps with JSON
{"name":"sam","daily":"2","monthly":"15"}. Decimal amounts are strings.
Use the admin bearer key or an owner browser session. Cookie mutations require an
explicit approved Origin and reject cross-site fetches. Owner caps cannot be set.
Names match configured account names, not API tokens. Cap edits persist in the
ledger and take precedence over the initial environment settings.
Before every friend dispatch, Aqua durably reserves a verified context window of
input plus a full output window per completion (n=1..8). Missing verified context
metadata fails closed. Large windows can require a cap larger than one request's
actual cost: no token-estimation shortcut weakens the cap. Successful JSON usage
or a completed SSE usage report replaces the reservation with reported cost.
Missing usage, truncated streams and ambiguous network failures keep the full
hold. Explicit request/auth/model/rate rejections (400, 401, 403, 404, 422, 429)
release it; ambiguous server failures do not. Concurrent requests share the same
account budget. A report above the reserved bound is recorded and blocks further
friend dispatches for that account until the ledger/bound is repaired offline.
Persistence failure also fails closed for friends. Owner requests stay uncapped;
check accounting_available and unknown_requests before treating totals as
complete. unknown_owner_requests_since_startup counts owner admissions skipped
because storage failed, and resets on restart. Audit logs are needed across such
outages. Owner usage cannot be reconstructed when an upstream omits usage.
These adapters add no credentials, enablement or deployment. New variable names
are source mappings, not permission to widen a loader's allowlist. Keep every
new AQUA_STATUS_* unset/OFF until its account, model mapping, user scope and
reviewed deployment config are bound together. Shared friend access is not
implied by an owner's successful test; existing per-account restrictions remain.
| Adapter | API base | Key / models environment variables | Exact upstream allowlist |
|---|---|---|---|
| OrcaRouter | https://api.orcarouter.ai/v1 | ORCAROUTER_API_KEY / ORCAROUTER_MODELS |
deepseek/deepseek-v4.1-flash-free only |
| Z.AI standard API | https://api.z.ai/api/paas/v4 | ZAI_API_KEY / ZAI_MODELS |
glm-4.7-flash, glm-4.5-flash only |
| Inferbase | https://api.inferbase.ai/v1 | INFERBASE_API_KEY / INFERBASE_MODELS |
requested deepseek-v4-flash only |
Status variables are AQUA_STATUS_ORCAROUTER, AQUA_STATUS_ZAI and
AQUA_STATUS_INFERBASE. Unknown or paid upstream IDs in these three adapters are
dropped during model parsing, rather than treated as fallback candidates.
Illustrative mappings, not an active environment file:
ORCAROUTER_MODELS="deepseek-v4.1-flash=deepseek/deepseek-v4.1-flash-free"
ZAI_MODELS="glm-4.7-flash=glm-4.7-flash"
INFERBASE_MODELS="deepseek-v4-flash=deepseek-v4-flash"
Inferbase returned deepseek-v4-flash:fp8 in its qualified test. The public name
is deepseek-v4-flash at Luke's request, but the fp8 build remains this binding's
internal identity. Do not pool it with V4.1 or assume another V4 build is identical. An Orca free suffix is a route requirement, not permission to
remove a model's version. Verify identity before joining any canonical pool.
Z.AI has a fixed local concurrency limit of one active inference across all its models, not one request per minute. A nonblocking account permit is held through the complete upstream JSON/SSE response and downstream writes; all exits release it. Local saturation skips that account and may use another binding of the exact requested model. If no binding succeeds, the existing generic 503/429 rules apply. Local saturation alone returns 503, creates no provider cooldown, and invents no Retry-After. The permit covers this Aqua server instance only; other processes or external clients using the same account are outside it.
Qualification evidence received from the server operator on November 16, 2026:
one real inference each succeeded for Zro deepseek-v4.1-flash, OrcaRouter's exact
allowlisted route, Z.AI glm-4.7-flash, and Inferbase's requested model above.
These are individual route smoke tests, not coding benchmarks or statistical
reliability claims. glm-4.5-flash has pricing/catalog evidence only, not a smoke
pass, and is intentionally absent from the example mapping. No new inference
was made while implementing this registry.
- Orca's explicit free alias must stay intact: no paid alias or wallet fallback. Standard free limit recorded as 10 RPM / 50 daily requests. Sources: https://docs.orcarouter.ai/routing/free-models and https://www.orcarouter.ai/api/free-package/public .
- Z.AI's exact two Standard API models had free input/output/cache pricing; FlashX, other Flash generations, Coding Plan, tools, paid fallback and top-ups are not included. Recorded account balances were zero with auto-recharge off and no payment method. Current pricing is not a forever-price guarantee. Sources: https://docs.z.ai/guides/overview/pricing and https://docs.z.ai/legal-agreement/terms-of-use .
- Inferbase qualification was for the verified existing Google account, not an earlier separate account's Playground. Recorded Free plan, zero credit/spend, 100 requests/day, 5,000/month, 10 RPM, 8,192 input/1,024 output tokens, $0.50/day provider-cost allowance. Allowance metering is not paid cost. Its saved policy allows Inferbase Managed only, known output price at most $2/million, one attempt and no cross-model/direct/BYOK fallback; request-content logging OFF. This price-eligibility policy is NOT a permanent no-spend cap after a top-up or upgrade. No saved-card list was available to prove historical card absence. Keep zero credit and no top-up, upgrade or paid fallback. Sources: https://inferbase.ai/docs/api and https://inferbase.ai/models/deepseek-v4-flash .
- NextRouter remains parked OFF after Cloudflare 1010 on both tested paths. No qualified model, no further workarounds, and no inference that its key is invalid. Groq remains held. Neither gets a new active registry route here.
- Existing AMD/NIM/Logfare/OpenRouter and the held Cloudflare registry retain their previous gates. Cloudflare remains OFF pending no-paid-fallback proof.
Account balances, policy and free pricing are time-bound evidence. No loader, secret transport, upgrade, top-up, provider enablement or service start is part of this source change. Future friend routing still needs verified context metadata for spend reservations; missing metadata continues to fail closed.
OpenRouter uses only these exact upstream IDs: cohere/north-mini-code:free,
nvidia/nemotron-3-ultra-550b-a55b:free, nvidia/nemotron-3.5-super-vl:free. The google/gemma-4-31b-it:free
candidate is held out of the gate pending the owner's A9a decision. Requests
force zero prompt/completion max price,
no provider fallback, and required parameter support. User plugins, tools,
model-fallback arrays, and search options are not forwarded. Initial input is
text-only, at most 64 messages and 8192 serialized bytes. This is a transport
limit, not a claim of exact token counting. Output defaults to 512 and requests
above 1024, zero, noninteger values, or n != 1 are refused for this route.
Admission requires the existing durable AQUA_SPEND_STATE_PATH store plus
AQUA_ACCOUNT_VERIFIED_UTC_DAY=YYYY-MM-DD for the current UTC day and
AQUA_OPENROUTER_VERIFIED_REMAINING from an operator's current account check.
The latter must be an integer 0..50. Local allowance is remaining minus 10,
clamped to at most 40. Missing/stale verification leaves this route closed and
unlisted. There is no automatic daily renewal or remote-check bypass.
These are proposed loader inputs, not a change to a deployed loader.
All OpenRouter models, owners and friends share one persistent daily counter. The account has one in-flight attempt across JSON and SSE. Every admitted attempt consumes one call permanently, even on error, timeout, missing usage, client disconnect or restart. This deliberately underuses uncertain allowance. A friend has at most 3 calls per UTC day across all models. Each exact upstream model has at most 10 calls per day including owner calls. A local preflight failure after admission can also consume a call. There are no refunds or key rotations to regain allowance. Outside traffic is not observed by this ledger; use a dedicated account/key scope or keep remote checks current. The snapshot is not a live claim of remaining provider capacity.
The durable account resource ledger also supports integer micro-neuron reservations, with unit tests for a 10,000-neuron shared account cap and an 8,000-neuron aggregate friend subcap. Cloudflare dispatch and listing remain closed in source: GLM-4.7-Flash's full-input token bound and forced total-output cap, including reasoning, are not yet verified. No byte-to-token guess is used for neuron admission. GPT-OSS and Qwen are not wave 2 routes. A bounded smoke alone does not prove worst-case cap enforcement. Activating Cloudflare needs a separate reviewed adapter change after that evidence, Workers Free active, current account usage and no-paid-fallback checks. No paid plan upgrade.
For owner-development-only AMD Radeon, NVIDIA NIM and Logfare, retain
AQUA_STATUS_RADEON=personal, AQUA_STATUS_NIM=personal, and
AQUA_STATUS_LOGFARE=personal. The existing status filter excludes their
provider bindings from friend catalog, direct routing and auto candidates.
This does not grant production use or override provider terms. Gonka stays off.
GET /v1/models includes display-only maker_pricing metadata from a reviewed
registry, with source URL, checked date and estimate basis. These are the model
maker's official API rates, not serving-provider rates, Aqua charges or savings.
Uncached standard online list rates are used; peak rates are used where time
bands exist. Free maker API tiers are shown as zero only where documented.
No published paid token tariff or unresolved build identity stays unavailable.
The browser calculates a virtual USD estimate from confirmed prompt/completion
token counts. Reasoning already included in completion tokens is not added again.
The registry never changes virtual accounting, friend access, caps, provider
allowances or ledger data. Historical usage is not repriced. Prices are a dated
snapshot, not a claim that a maker will keep its rate or offer this route forever.
Runtime mappings for this release use one deepseek-v4.1-flash public binding
for Zro and OrcaRouter. Luke accepted advertised same base weights for pooling;
exact serving build/quantization equality was not independently proved. Existing
provider eligibility and same-model failover rules still apply. Both routes stay
personal/owner-only under the current runtime packet. The hidden legacy request
name deepseek-v4.1-flash-route-b resolves to that canonical pool. The old
deepseek-v4-flash:fp8 request name resolves only to deepseek-v4-flash.
Update the runtime bindings with this source release; no counter or ledger reset.
Keep maker-price metadata unavailable for generic DeepSeek aliases until the
exact official API snapshot is grounded. A self-identification answer is not
model provenance.
NIM, Radeon and Logfare are owner-only even if their runtime status is mistakenly
set to friends. Only the exact upstream IDs in trial_model are eligible. NIM
DeepSeek is excluded (410 on the current key). Runtime still ships off by default;
a source allowlist is not a deployment or account-entitlement claim.
These three adapters have local concurrency one, held through the full JSON/SSE response, and a shared rolling two-admissions-per-minute limit per provider. Local rate state is bounded and in memory; restart clears it. It is not a spend, token or daily account quota and does not promise that an upstream will accept a call. Owner asked to rely on provider-side free quota exhaustion rather than add spend limits. This does not alter existing Aqua virtual USD tracking, friend caps, ledger history, or Gonka/Cloudflare eligibility.
Trial adapters reject non-text messages, tools, multiple choices and requests over 8192 serialized UTF-8 bytes. A byte limit is transport hygiene, not a token bound. Redirects are not followed. NIM upstream I/O is limited to120seconds and never extends the remaining aggregate request budget. Downstream writes remain bounded per operation, not by a promised whole-request wall clock. Output limits are whatever the upstream actually enforces; no hard token guarantee is claimed.
Healthy same-model bindings prefer observed successful response-header latency: EWMA with25%latest sample, per(provider,canonical model). A bounded512entry cache lives until restart. Unmeasured bindings retain config order after measured ones; there are no active probes or synthetic performance claims. Passive cooldowns still win; pinned requests never switch model and auto keeps its model shortlist order. This measures headers, not first generated token, throughput or answer quality. It cannot rank an untried provider or guarantee the next call's latency.
GET/POST /v1/settings use the canonical authenticated account. Writes save
atomically in the existing ledger, require a durable store, reject unknown fields
and cannot select another user. Cookie writes require the existing safe-origin
checks. The settings page offers server-enforced ZDR-only routing, a default
playground model and a playground maker-estimate visibility option. Strict ZDR
filters both catalog and pin/auto dispatch and rejects UNKNOWN as well as NON_ZDR.
No current route is VERIFIED_ROUTE, so strict mode currently leaves no usable
models. Changing a default never changes entitlement. Hiding cost never changes
accounting. Existing ledgers default to ZDR-off, auto, estimates-visible.
Retention source URLs and checked date are exposed only in owner admin provider metadata. Broker-only promises do not certify upstream retention. Logfare, Radeon and NIM trial are NON_ZDR; remaining routes are UNKNOWN until full-route terms and live configuration are verified. No training or consent setting is changed by this source.
Thread creation uses fallible OS spawn results: bind returns an error if the accept thread cannot start, and a handler spawn failure closes that connection. Existing admission/drain tests pass, but the OS thread-creation failure branch has not been exercised by fault injection.
Disabled by default. Source registration does not enable this route. After independent QA and deployment approval, opt in explicitly with:
AQUA_STATUS_POLLINATIONS=personal
POLLINATIONS_MODELS="gpt-oss-20b=openai-fast"
No API key is used. The destination is fixed to
https://text.pollinations.ai/openai, with no endpoint override, bearer token,
cookie, forwarded client headers, redirect following, or gen.pollinations.ai
continuation. Only the exact mapping above is accepted. The route stays
owner-only even if its status is accidentally set to friends. Within the
canonical gpt-oss-20b model it is in the lowest provider-priority group. It
cannot make another canonical model silently resolve to GPT-OSS 20B.
This first version is JSON-only. Streaming requests skip this route, and JSON
responses must report model: gpt-oss-20b before their contents reach clients.
Missing or changed response identity fails closed rather than claiming the
pinned model was served. The anonymous alias can change upstream; this guard
checks the returned label, not the underlying model weights.
One process-wide request may be in flight, held until the response body finishes or the attempt fails. Dispatch starts are at least 60 seconds apart, including failed attempts. This is a conservative local limit, not a published quota or a remaining-balance claim. Multiple Aqua processes do not share this local gate; keep only one enabled instance per egress while evaluating the anonymous tier. Upstream limits may apply to the shared IP. No paid fallback is configured for this adapter. Retention is unknown, so strict privacy mode excludes it.
Source checks: https://text.pollinations.ai/models, https://pollinations.ai/privacy, https://pollinations.ai/terms.