- Drop selection_token (unsalted hash of a short grant ID, reversible) and
claudeGrantSelection from the read-only DTO; assert it never serializes.
- query() returns only the projected result; no raw upstream block escapes.
- Clear a cooldown_until that is already past, using the same `now` as the
redeemable check; the cell also shows the cooldown only when in the future.
- Invalidate in-flight queries when account id, platform or type changes.
- Sort the earliest-expiry fallback by parsed timestamp, invalid/missing last.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
- Show resets held (sum of resets_left) on the count button, matching the
Codex count; flag "not usable now" when none is redeemable yet, instead
of showing 0 while a grant exists.
- Pick the redeemable (next) grant for the expiry chip, falling back to
the earliest expiry, rather than assuming credits[0] is next.
- Render a single ClaudeResetCreditsCell across usage states so a query
started while usage loads is not lost on remount.
- Drop an unrelated google/subcommands go.sum entry.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The Claude reset credit query rendered as a bare "Claude 重置额度" link,
which reads like a redeem action, and dumped raw ISO timestamps and
window keys. Render it like OpenAIQuotaResetCell instead: a read-only
"次数 N" button sharing one row with the local query button via a
#pre-actions slot, plus an expiry chip whose tooltip carries the grant
label, cleared windows and limit requirement.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Codex expands a leading ~/ in config.toml paths on every platform but
does not expand %userprofile%. The Windows config generated by the Use
Key dialog wrote model_catalog_json = "%userprofile%\.codex\codex-models.json",
which Codex resolves relative to the config directory
(C:\Users\<user>\.codex\%userprofile%\.codex\codex-models.json), so it
fails to load the config with "The system cannot find the path
specified (os error 3)".
Always write ~/.codex/codex-models.json into config.toml. The download
hint still shows the Windows location to save the file.
MALFORMED_FUNCTION_CALL responses from upstream contain only
thoughtSignature without any text, thinking, or tool calls.
The previous fix still treated delta.Signature and block.Signature
as meaningful data, so the empty-stream failover was never triggered.
- Remove Signature from isMeaningfulAntigravityCompatEvent checks
- Let hasMeaningfulData() also consult processor.HasContent() as
a secondary guard for edge cases
RecordUsage treats a missing price as zero cost on the first pass, but
the Free Fast Standard re-evaluation returned the same pricing error.
Requests in a free_openai_fast group sent at priority/fast tier to a
model without pricing therefore lost their usage log entirely.
Ignore pricing-unavailable errors on the Standard pass and keep the
zero-cost entry, matching the first pass.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
UsageProgressBar returned 'now' whenever showNowWhenIdle was set and
utilization was 0%, ignoring resetsAt. OpenAI Codex 7d windows at 0%
usage still carry a future reset time, so the admin account list hid it.
Only fall back to 'now' when no reset time is known.
With the credit config now visible to the scheduler, every scheduling pass
notifies the auto-reset worker for paused credit-enabled accounts.
evaluateAccount forced an upstream query whenever the reset threshold was
reached, regardless of the last result. Accounts with no credits left and
7d >= 95% were therefore polled on /wham/usage every ~15-20s across
instances (previously once a minute by the scanner), and during upstream
outages the failed state caused the same storm.
- At the reset threshold, skip the forced query while a no_credit result is
still within the state TTL (10 min); re-query once it expires so newly
granted credits are still picked up.
- After a query-stage failure (RESET_CREDIT_QUERY_FAILED,
USAGE_SNAPSHOT_WRITE_FAILED, RESET_CREDIT_DETAILS_UNAVAILABLE), wait one
minute before querying again, including snapshot-staleness triggers.
- Redemption-stage failures are unaffected and still retry immediately with
the same idempotency key; accounts with available credits still query and
redeem immediately.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Chat Completions file parts already become Responses input_file, but the
Antigravity path dropped them when converting to Anthropic and then to
Gemini. Map data-URI input_file to a document block and emit it as
inlineData so Gemini receives the PDF.
Fixes#7583
The personal-access-token /v1/alpha/search fallback turned any upstream
Responses SSE body into a 200 search result and billed one web search call.
Error, response.failed and response.incomplete events, a response.completed
with a non-success status or without its response, a truncated stream and a
bare [DONE] were all accepted, returning partial text as a successful and
billable search.
Parse the stream strictly: those shapes now return an error, so nothing is
written to the client and no WebSearchCalls result is produced. A successful
response.completed keeps the existing output and citation handling. The
existing PAT metadata test now feeds a Responses SSE stream, which is what
that upstream path returns.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The API-key Responses capability probe treated every 404 as conclusive and
stored responses_supported=false. The probe model is the first mapped
upstream model in lexical order, so on third-party accounts it is often an
auxiliary entry such as codex-auto-review; a model_not_found answer then
permanently routed the account away from /v1/responses although the
endpoint works.
A 400/404 whose error code/type or message says the probe model does not
exist or is unavailable is now inconclusive, so the stored capability is not
overwritten. Plain 404/405 still mean the endpoint is missing. The probe
also prefers a general gpt-* text model (not an image model) among the
mapped models before falling back to the first one.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
When a channel pricing entry left image_output_price empty,
GetModelPricingWithChannel and the pricing resolver zeroed the image
output price and marked it explicit, so image output tokens were billed
at $0. image_input_price followed the same "unset means zero" rule and
fell back to the text input price instead of the catalog image input
price. A channel that lists an image model such as gpt-image-2 in token
mode without filling in prices therefore generated images without
charging for image output tokens, while the available-channels view
(#2475) displayed the catalog image price for the same entry.
#2930 wanted an explicit 0 on the channel to stop falling back to the
catalog; the change applied the same treatment to unset values as well.
Treat the two image fields like every other token price field:
- Set on the channel (including 0): override the catalog. An explicit
image_output_price keeps ImageOutputPriceExplicit, so 0 stays free.
- Left empty: keep the catalog price (LiteLLM or built-in fallback). If
the catalog has none, computeTokenBreakdown still falls back to the
text input/output price.
The admin UI saves an empty field as null and only an entered 0 as 0,
so the two cases stay distinguishable. The flat, interval and
GetModelPricingWithChannel paths now share applyChannelImagePriceOverrides.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The canonical Anthropic stream leaves tool_use input empty on
content_block_start and streams the arguments as input_json_delta.
Anthropic-compatible relays may instead put the complete arguments on
content_block_start and send no delta. The Anthropic-to-Responses stream
converter ignored that input, so function_call_arguments.done carried ""
and the closed item "{}", and clients executed the tool with no arguments.
Keep the inline input as a seed. A real input_json_delta supersedes it, so
canonical streams keep their exact event sequence and two JSON documents are
never concatenated. When no delta arrived by content_block_stop, emit the
seed as a single arguments delta before the done event, which still repeats
exactly what the deltas streamed. Empty, "{}" and null inputs keep the
existing "{}" fallback.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
BufferedResponseAccumulator.SupplementResponseOutput rebuilt the output only
when the terminal response had an empty output array. Upstreams can also end
with a non-empty output whose message carries no (or only blank) output_text.
The buffered Chat Completions and Anthropic Messages paths then returned an
empty reply although the text had streamed, while usage still billed the
terminal output_tokens.
When the terminal output has no usable message text, fill the first empty
output_text part (or append one, or append a message item) from the
accumulated deltas. Non-empty terminal text stays authoritative, and nothing
is synthesized when no text streamed.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
isReasoningModel matched the literal "gpt-5" prefix plus the GPT-6 Sol/Luna
spellings, so gpt-6 and gpt-6-astra (both already in the model catalog and
price list) fell into the sampling branch. Anthropic Messages and Chat
Completions requests converted to the Responses API then forwarded
temperature/top_p, which the Responses API rejects for these models with
"Unsupported parameter".
Key the check on the parsed GPT generation number instead, keeping the
existing Sol/Luna spelling match, so later families are covered without
another edit. Non-generation GPT families (gpt-image-1, gpt-audio) and
gpt-4.x are unaffected.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
OpenAI candidate filtering reads the Redis metadata projection, but
filterSchedulerExtra dropped auto_reset_credit_enabled, the 5h/7d
reset-credit thresholds and codex_auto_reset_credit_state. The scheduler
therefore always saw automatic reset credits as disabled, so accounts
were paused at the ops quota auto-pause threshold even when a fresh
state showed an available reset credit. The "keep scheduling until the
reset-credit threshold" branch never took effect, and accounts with
credits stayed idle until their window reset naturally.
- Keep the four keys in the scheduler projection; the test round-trips
the real JSON payload and resolves the credit config from it.
- Once visible, every scheduling pass notifies the auto-reset worker for
paused credit-enabled accounts, and each notification reloads the
account. Add a 30s per-account cooldown for the scheduler hot path
only; other event-driven notifications are unchanged and the
per-minute scan remains the fallback.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
#7395 made Codex (and Grok) imports store a base URL ending in /v1, but the
usage script still appends a hard-coded /v1/usage, so CC Switch requests
/v1/v1/usage, gets a 404 and shows "query failed" for every newly imported
Codex/Grok provider. Move the script into ccswitchImport.ts and strip an
existing trailing /v1 from {{baseUrl}} before appending /v1/usage, which
covers both the imported value and base URLs users edit either way later.
Close the existing catalog identity test before the new capability test and
assert that an explicit null capability remains present. Update the quota
snapshot fixture to the pointer parameter already used by the current base.
Verified six full backend unit packages, scoped lint and CGO-disabled build.
A single passthrough account made GetAvailableModels return nil for the whole
group, so every other account's model_mapping was dropped from /v1/models and
the Codex catalog. Treat passthrough accounts like unmapped ones: skip their
(possibly stale) mapping and let supplementUnmappedOpenAIModels add the
default set, while ordinary accounts' mappings still reach the list.
Refs #6647, #7517
Validate mapped targets for default image models as well as aliases. Preserve native image choices in passthrough and cover mapping changes over the shared catalog cache.
Image choices were resolved from DefaultModels IDs and from mapping keys, so an
alias whose target is a local image model (paint -> gpt-image-2.5-flare) never
reached the picker: IsModelSupported only matches keys, and a public name is not
an image model. Resolve the public name to its target before deciding, and only
synthesize an entry when the target is an image model.
Update the fixture of
TestFetchOpenAIAccountModelsOAuthLabelsLocalImageModelsLikeUpstream: it asserted
on an upstream slug the account did not configure, which conflicts with the
account-scoped projection. Coverage is preserved by configuring that slug and by
a new test asserting unconfigured text models stay out.
Vertex AI reports 429s with a standard google.rpc.RetryInfo detail
(retryDelay like "39s"), but ParseGeminiRateLimitResetTime only knew the
AI Studio quotaResetDelay metadata, so parsing always failed for Vertex
bodies. The fallback then lumped service_account accounts in with
API-key/AI-Studio ones and rate limited them until the next PST
midnight, so a transient per-minute throttle could take a pay-as-you-go
account (and with it the whole group) offline for up to a day.
- Parse google.rpc.RetryInfo retryDelay from 429 bodies, capped at 15
minutes to guard against absurd upstream values.
- When no reset time can be parsed, cool service_account (Vertex)
accounts down for 60s instead of until PST midnight: pay-as-you-go
quotas are short-window RPM/TPM with no daily reset.
Fallback behavior for Code Assist, Google One and API-key accounts is
unchanged, covered by new regression tests.
Adds a case with the real block layout (attribution block, identity sentence,
task instructions) and asserts that user content mentioning Claude Code and
claude-sonnet-4-6 is forwarded unchanged.
Stripping the x-anthropic-billing-header line is not enough. Claude Code's
Agent SDK prompts (session title generation, and other SDK-driven calls) open
a separate system block with
You are a Claude agent, built on Anthropic's Claude Agent SDK.
and the Google upstream still rejects those requests with 429 RESOURCE_EXHAUSTED.
Measured on Sub2API 0.2.7, which already contains the attribution fix, using an isolated single-account group, waiting for the per-model cooldown
to expire before every request so that a previous failure could not turn later
results into 503:
original request replayed 429 (8/8)
same request, generic system prompt 200 (8/8)
attribution also removed client-side 429 (3/3)
attribution removed + identity neutralized 200 (3/3)
only 'Claude' replaced in system 200 (2/2)
only 'Anthropic' removed from system 200 (2/2)
output_config / max_tokens / non-stream no effect
The rewrite is anchored at the start of a system block, so user instructions that
merely mention Claude or Anthropic, and model ids such as claude-sonnet-4-6, are
left untouched. Scoped to the Antigravity transformer, as with the existing
attribution handling.
Codex sends images as input_image.image_url. DeepSeek's live /responses
decoder requires a url field and returns 422 missing field url, including
on openai-platform accounts whose base_url host is api.deepseek.com.
Co-authored-by: Cursor <cursoragent@cursor.com>
Extend the local plugin system's host<->plugin contract with generic,
capability-neutral facilities, so transport plugins can be built on top
without embedding plugin-specific logic in the host.
Host services (plugin dials a broker-provided HostService gRPC):
- per-plugin namespaced KV store (key/value size and TTL bounded)
- account directory: ListAccounts + ResolveOutboundIdentity, restricted to
the accounts a plugin's declared capability already covers
- InitHostServices lifecycle hook; plugins that don't implement it run
unchanged and are simply offered no host services
Read-only runtime status channel:
- HealthResponse gains an opaque, plugin-defined status_json produced
passively (no config apply, no upstream call, no side effects)
- GET /admin/plugins/:id/status (ungated) and the plugin.status UI-bridge
verb map to Health, so a status UI can poll live state without a step-up
prompt or a "test" side effect
- config.test no longer force-shows a host success toast; the plugin UI owns
result presentation, so plugins that read status do not spam toasts
Document the UI-bridge verb taxonomy (read vs mutate/test gating) and the
status contract under pkg/pluginapi.
Kimi reports Coding Plan window exhaustion (5h/weekly) as a 403 with
error.type=access_terminated_error instead of 429. It fell into the
generic handleOpenAI403 escalation, so three 403s within 180 minutes
permanently disabled accounts that recover by themselves at the window
reset.
Detect the quota-exhausted 403 for CN coding-plan accounts and cool it
down exactly like the 429 path: SetRateLimited until the earliest future
window reset from the quota snapshot, falling back to a short
temp-unschedulable window when the snapshot is missing. Non-coding-plan
accounts keep the generic 403 escalation since the rolling-window
recovery mechanism does not apply to them.
Also extract cooldownCNProviderToQuotaSnapshotReset and
setCNProviderTempUnschedulable so the 429 (OpenCodeGo / Coding Plan)
and 403 branches stop carrying three copies of the same sequence.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Codex histories often replay encrypted-only reasoning items. After a
Responses→Chat Completions fallback, a cache miss left assistant
messages without reasoning_content and DeepSeek rejected the turn
with 400. Fill empty assistant fields with a single-space placeholder
for DeepSeek hosts, including openai-platform accounts whose base_url
points at api.deepseek.com. Existing plaintext and cache hits are
left unchanged.
Co-authored-by: Cursor <cursoragent@cursor.com>
Gemini/Antigravity delivers tool arguments as one complete object, so this
converter synthesizes the Responses tool-call stream itself and therefore owns
its self-consistency. The done event was built without Arguments, leaving it ""
while the preceding delta carried the entire JSON. Clients that reconcile the
done payload against the accumulated deltas (OpenAI Responses SDK); Qoder's
reconcileToolCall) reject that mismatch as inconsistent_tool_call and abort the
turn non-retryably, so every tool call through the gateway failed.
Emit state.CurrentArgs, which closeCurrentResponsesItem already uses for the
item, restoring sum(deltas) == done.arguments == item.arguments.
Regression test asserts the delta sum equals the done payload and fails closed
when Arguments is dropped.
DeepSeek native Responses rejects Codex view_image results when function_call_output.output contains input_image parts. Move media into a following user message and keep the tool output as a string so call_id pairing remains valid.
Paused (schedulable=false) active accounts were excluded from the
background token refresh candidates, so their stored access_token
silently expired. The admin usage-window quota probe uses the stored
token verbatim (no token provider / no refresh), so the probe then gets
HTTP 401 UNAUTHENTICATED and the dashboard reports a false "needs
re-auth" for a perfectly healthy account, while Test Connection and live
traffic (which refresh on demand via the token provider) keep working.
Pause is not permanent rejection: error accounts are already excluded by
the status = 'active' filter, and accounts whose refresh genuinely fails
are rate-limited by the retry-cooldown exclusion (10-minute retry-
exhausted cooldown). Refresh success only persists credentials and never
touches schedulable, so a paused account cannot be silently un-paused.
Drop the schedulable filter and update the SQL-filter test, the
integration pagination test, and the service-level candidate fake.
- Replace unchecked single-value type assertions with comma-ok form
guarded by require.True (errcheck check-type-assertions)
- Remove trailing blank line flagged by gofmt
Co-Authored-By: Claude Code <noreply@anthropic.com>
The Responses to Chat Completions bridge mapped both the instructions field and role developer items to system messages. Codex always sends both, and also injects developer notices mid-conversation when switching models, so the forwarded request contained consecutive leading system messages and mid-conversation system messages. Strict Qwen-family upstreams reject this with HTTP 400 "System message must be at the beginning."
Leading system/developer messages are now merged into a single leading system message, and later system/developer messages are downgraded to user so the content is preserved without violating upstream role ordering.
Real Ollama Cloud 429s are driven by ollama.com usage windows, not by
response headers that OpenAI/Anthropic/CN rate-limit parsing understands.
Handle them separately: apply a never-shortening temporary cooldown first,
then schedule an async probe that learns the true usage-window reset point
with a controlled, coalesced per-api_key refresh.
- ratelimit_service: route real Ollama Cloud usage accounts to a dedicated
429 path before the CN-provider branch.
- ollama_cloud_usage: add a background probe coordinator (started in Start,
waited by Stop) that fans out model-429 probes to one controlled refresh
per api_key group, reusing the recent group snapshot for coalesced members.
- repository/account_repo: add SetRateLimitedIfUnchanged, a write-back CAS
counterpart to ClearRateLimitIfObserved, so an async reset never overwrites
a newer 429, an admin clear, a re-armed generation, or a key/state change
observed by another writer between read and write.
- wire: inject OllamaCloudUsageService into RateLimitService.
Tests: unit tests for service and repository pass; targeted race tests pass;
build and vet clean; git diff --check clean. The CAS integration test compiles
but is not run (no Docker available).
- Fix test assertion target location for whereItems prefixItems
- Support const keyword normalization to enum and infer scalar type
- Infer scalar type for enum-only schema to prevent object reason injection
- Recursively clean properties and items after anyOf/oneOf merge
- Ensure empty prefixItems is safely deleted
- Handle JSON Schema Draft 2020-12 prefixItems (tuple arrays) by converting to a best-match items schema
- Ensure array schema always defines a valid items schema to prevent Gemini 400 'missing field: items.items' error
- Fix recursive traversal when a schema defines both properties and items
- Add comprehensive unit test cases for tuple and array schemas
Fixes#7061
monitorProviders and probeCapableProviders were missing
MonitorProviderMiniMax even though the constant, adapter, ent schema and
error message all already included minimax, so provider=minimax was
rejected by validateProvider.
Return the updated account alongside partial refresh warnings so the admin UI can keep its state current and tell operators when project ID retrieval must retry.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Use shared model routing for image forwarding and account tests, retain Luna orchestration for legacy models, and add Image 2.5 pricing and usage accounting.
Skip OpenAI messages-dispatch gpt-5.x defaults for OpenCode groups and
composite OpenCode targets. Filter unpriced claude-* billing candidates,
estimate count_tokens locally, sanitize body session IDs before setting
X-OpenCode-Session, and treat Zen 402 as recoverable payg cooldown.
Stop rewriting the inbound body before protocol routing so account
model_mapping applies once and usage_logs.model stays the client name.
Apply account model_mapping to the request body before OpenCode
protocol routing so /v1/responses, /v1/chat/completions, and
/v1/messages all send the mapped ID as the real upstream model.
PR #6499 merged without CI ever running on its head commit (fork run stayed
action_required), landing three bare `defer db.Close()` calls in
internal/setup/setup_test.go. The repo's errcheck config sets
disable-default-exclusions: true, so those are reported and golangci-lint has
been red on main ever since, failing every subsequent PR run.
Switch them to the codebase's existing convention
`defer func() { _ = db.Close() }()`.
Keep the internal platform ID opencode_go. Product surface is OpenCode
with account_mode zen|go, model-aware routing across Chat Completions,
Responses, and Anthropic Messages, and X-OpenCode-Session derived from
prompt_cache_key / metadata.user_id so prompt cache can hit.
Unmatched models fall back to /v1/chat/completions. Related to #6742.
prolite -> Pro 5x (violet, same as Pro), pro/chatgptpro -> Pro 20x, self_serve_business_prolite -> Business Premium (indigo, same as Team), team -> Business Standard.
The tier naming stays OpenAI-only so Antigravity Pro and Grok pro keep their own labels; planTypeDisplayLabel now shares openAIPlanTypeLabel with PlatformTypeBadge instead of duplicating the mapping, and the manual plan_type dropdown gains the Pro 5x / Business Premium presets.
The streaming Responses-to-Anthropic converter emitted assistant text only
from output_text.delta, so a stream carrying its answer on
response.output_text.done or in the terminal response.output produced a
protocol-valid message with no text.
Record emitted text per output_text part in a ledger that survives content
block closure, and recover only the suffix a finished payload adds, as #6629
does for function-call arguments. Where the part a payload finishes cannot be
established, decline to recover rather than repeat delivered text: terminal
recovery runs only when nothing was emitted, and a done payload indexed
differently from every delta is ignored.
Give output_text.done its own handler so the shared block-closing helper and
its tool-path caller stay unchanged, and so a payload arriving after
message_stop cannot open a content block past the end of the message.
The claude_code_only probe exemption only fired for models containing
"haiku". Claude Code also sends max_tokens=1 requests without a system
prompt to the model currently in use, e.g. when switching models or
refreshing the context usage, so on sonnet/opus groups those requests failed
the system-prompt check and were rejected as non-Claude-Code clients.
Recognise the probe from the request body itself (max_tokens == 1) once the
User-Agent check has passed, regardless of model, and carry max_tokens into
the ParsedRequest projection so the hot path sees it too. The haiku context
flag and the local haiku interception are unchanged; non-haiku probes reach
the pool as ordinary one-token requests. One output token offers nothing to
a client that already forges the CLI User-Agent, so this widens no attack
surface.
Fixes#6591
Keep always-on toolConfig from main (reasoning models without tools) while
retaining the PR behavior that drops mixed built-in tools when client
function tools are present, so includeServerSideToolInvocations stays unused.
Co-authored-by: Cursor <cursoragent@cursor.com>
Update the leftover buildTools unit test that still expected
googleSearch to coexist with client function declarations.
Co-authored-by: Cursor <cursoragent@cursor.com>
Resolve three conflicts from main's #6680 (supplementUnmappedOpenAIModels):
1. gateway_handler.go — keep PR's allowlist refactoring (ModelAllowlistEnabled,
writeAllowlistedModelsList, modelListingSource); main's OpenAI branch in
writeModelsList was auto-merged.
2. group_models_list.go (modify/delete) — delete file per PR intent; relocate
supplementUnmappedOpenAIModels to group_model_allowlist.go where gateway_service
and openai_codex_models_service can still call it.
3. openai_codex_models_service.go — combine main's supplementUnmappedOpenAIModels
wrapping with PR's ModelAllowlistEnabled() guard.
Also fix two auto-merged test files that still referenced the old
GroupModelsListConfig / ModelsListConfig types.
- TestCodexAliasFailoverMappingHonorsModelRouting/no_routing_rule: expect
["text","image"] instead of ["text"] — image input is resolved per-account
via isOpenAICodexImageInputModel, independent of routing rules
- TestCodexAliasFailoverMappingHonorsModelRouting/routing_rule: expect
effort levels to include "ultra" (added by #6718)
#6669 renamed the constant from redeemMaxErrorsPerHour to
redeemMaxFailedAttempts (fixed 10-minute window), but #6670's tests
still referenced the old name, breaking compilation on main.
- Resolve conflict in openai_codex_models_service_test.go: keep both
GPT-6 Astra instructions template assertion (from this PR) and ultra
reasoning level + multi_agent fields (from #6718 on main)
- Fix compilation error on main: add missing `group *Group` parameter
to buildCodexModelsManifestForAccounts calls in
openai_codex_model_metadata_test.go (lines 24, 32) — caused by #6688
changing the function signature while #6718 added new callers with
the old signature
Give operators a V2 setting (channel_monitor_hide_user_ranking) to hide
the user ranking tab and the /users payload from non-admin channel
monitor v2 viewers. Default false (ranking stays visible). Admin
endpoints and views always keep the ranking.
Also stop painting first-token '-' as a red/empty failure when traffic
is sync-only and has no first-token samples: a missing TTFT is
"not applicable" (health-unknown), not a breached latency budget.
Changes:
- new setting key + fail-open parse (only literal "true" hides), wired
through public/admin settings DTOs, handlers, and the V2 runtime
- ChannelMonitorV2Service.Users returns an empty list to non-admin
viewers when the toggle is on
- frontend: feature flag + SettingsView toggle + user V2 view hides the
ranking tab and skips the /users call
- monitorFormat: isTtftUnavailable/ttftDisplayState keep missing TTFT
neutral
Tests:
- setting parse defaults to visible and only literal "true" hides
- V2 Users honors the toggle for non-admin viewers
- frontend MetricCell/monitorFormat keep missing TTFT neutral
Use encoding/json instead of Go string quoting when promoting message content to cached text blocks. Add regression coverage for control characters, Unicode, literal backslashes, text preservation, and idempotence.
An OpenAI group that keeps a failover account for a shared public alias
produced a Codex model descriptor with no reasoning levels and text-only
input, which leaves the client's reasoning-effort slider with nothing to
select.
Two accounts explicitly mapping one alias to different upstream models
makes codexExplicitModelTargetsConflictForPlatform true. With no
upstream_model_metadata snapshot to intersect, groupCodexModelMetadata
short-circuits to {reasoningConflict, inputModalitiesConflict}, so
DefaultReasoningLevel is nil (the key vanishes under omitempty) and
SupportedReasoningLevels is an empty array. mergeMissingCodexModelFields
fills only missing-or-null fields, so an empty array is never repaired.
Failing closed is right when the alias target is genuinely ambiguous, but
a group that declares model_routing for the alias has already resolved
that ambiguity: the operator stated which accounts serve it and which is
primary. Treating that as a conflict rejects a deliberate failover setup.
Manifest side: when a routing rule names the alias, restrict capability
declaration to the routed accounts and skip the conflict override.
Capabilities then fall back to the same by-name derivation used for a
single account without a metadata snapshot. Groups with no rule, a
disabled rule, or a rule for another alias keep failing closed.
Scheduling side: model_routing was stored and read platform-agnostically
(admin_group.go, Group.GetRoutingAccountIDs) but four gates in
gateway_scheduling.go only consulted it for Anthropic, so an OpenAI
group's rules were accepted and displayed yet never honored. Without
this, the manifest would honor a rule that scheduling ignores. Both
gates now share one predicate, with the allowed platform set defined in
a single place.
Selection inside a routing list is unchanged: the list is a membership
set, and the winner is still decided by Priority, then load, then
LastUsedAt.
Tests: alias-failover manifest recovery plus the three negative cases,
the platform predicate table, and an OpenAI routed selection that must
beat a higher-priority unrouted account.
Upstream returns 400 INVALID_ARGUMENT when toolConfig is absent. The guard
`if !isReasoning || len(tools) > 0` skipped it for reasoning models called
without tools, contradicting the intent stated in the comment above it.
Set toolConfig unconditionally and cover the three cases with a unit test.
Payment and admin fulfillment shared the public Redeem entry point, so
fulfillment retries were blocked by, and contributed to, the user's
public redeem failure counter, and lookup errors were silently treated as
"code does not exist" on a partial re-run.
- resolveRedeemAction now distinguishes a genuine not-found lookup from
other lookup errors (which now fail the fulfillment instead of being
swallowed and re-created).
- Add validatePaymentRedeemCode: before reusing an existing code during
fulfillment, verify it belongs to the order (code id, balance type,
amount, user, status), so a stale or foreign code can never be
credited against the wrong order/user.
- Split Redeem into a policy-based core: public redemption enforces the
failure counter; payment fulfillment (redeemForPaymentFulfillment) and
admin fulfillment (RedeemForAdminFulfillment) bypass it, while all code
validation and transactional updates remain identical. Admin
fulfillment keeps the normal redeem-level affiliate rebate.
- The admin CreateAndRedeem endpoint now uses RedeemForAdminFulfillment.
Tests: updated resolveRedeemAction cases, new TestValidatePaymentRedeemCode,
code-used-by-another-user rejection, payment/admin rate-limit bypass
coverage, and public-path regression coverage.
The redeem failure counter used INCR followed by Expire(24h) on every
attempt, so each failed attempt refreshed the key TTL and the window
never actually closed for a user who kept retrying within 24h.
- Count attempts in a fixed 10-minute window via a small Lua script
(INCR, then PEXPIRE only when the key is created or has no TTL), which
is atomic and also repairs keys that lost their TTL.
- Move the counter to the redeem:ratelimit:v2: key space so pre-existing
24h counters do not seed the new window.
- Rename the service threshold to redeemMaxFailedAttempts (30) to reflect
that it bounds attempts per window, not per hour.
Public redeem behavior is unchanged apart from the window semantics:
the check fails open when the counter read fails, and a failed
increment never masks the underlying domain error.
The Antigravity native Gemini route (generateContent) expects each
contents entry to carry an explicit "user" role, matching the request
shape used by the established Antigravity account-test flow. Without it
the probe body is rejected with 400 INVALID_ARGUMENT, so Gemini channel
monitors report false failures.
Add the explicit role to the Gemini provider adapter body and cover it
with a regression test that parses the generated request body.
The Antigravity service fingerprints clients by the version embedded in the User-Agent header. The built-in default (1.23.2) had drifted far behind current real client builds, so requests using the default UA started failing OAuth and probe flows.
Bump DefaultUserAgentVersion from 1.23.2 to 2.9.1 (current client build) and update the constant test accordingly. The default remains overridable via ANTIGRAVITY_USER_AGENT_VERSION and the admin setting.
The Grok cross-client model mapping is enabled by default: InitializeDefaultSettings writes "true", and the parser treats a missing/empty value as enabled ("Default true (missing/empty -> enabled)"). However the comment above the defaults and the admin settings UI hint (EN and ZH) both claimed "disabled by default", contradicting the actual behavior.
Align the comment and both UI hints with the real default. No behavior change.
Claim scheduler-acquired slots before media eligibility checks and task-owner
validation. Release abandoned attempts immediately and retain once-only
cleanup on cancellation and forwarding completion.
Select only the resolved video owner, reuse bounded account waiting without
sticky escape, and preserve the task binding and TTL. Keep generation
failover and owner/billing isolation intact.
Regression tests reproduce unreleased account slots on upstream ab99d56e9
and pass with this change. Validation: go test ./..., unit and integration
suites, focused race tests, backend build, and golangci-lint (0 issues).
Cyber-policy detection only inspected response.failed, and the ingress check ran after error-specific early returns. As a result, error-shaped failures and the connection-pool, HTTP-bridge, and WSv2 paths could miss cyber-policy hits.
Centralize event marking and invoke it before retry and failover returns across the WebSocket implementations. Preserve the recorded guard during account failover and defer connection blocking until the logical failover chain completes, avoiding duplicate records while ensuring subsequent turns are blocked.
Official OpenAI hosts and models.dev entries without an api field left
Astra falling back to reasoning effort none and empty Fast tiers. Persist
complete per-model metadata, advertise Astra reasoning/Fast locally, and
surface partial sync warnings when image models remain incomplete.
Co-authored-by: Cursor <cursoragent@cursor.com>
Anthropic gates newer models on a minimum Claude Code client version. With
CLICurrentVersion pinned at 2.1.220, claude-fable-5-1 fails every request with
"Claude Code 2.1.220 does not support this model; version 2.1.251 or newer is
required", and the only fix is a release that bumps one constant.
Add SUB2API_CLAUDE_CLI_VERSION, mirroring what internal/pkg/xai already does
with XAI_GROK_CLI_VERSION. CLICurrentVersion remains the built-in pin; the new
CLIVersion() is what every consumer reads. With no env var set the resolved
value equals the pin, so existing deployments are unaffected.
The value is resolved once at package init rather than per call: the User-Agent
header and the request body's cc_version are written by different code paths for
the same request, and reading different values would make that request
contradict itself.
Validation is strict (three numeric segments, not below the pin) because
identity_service.go documents what a malformed version does once it reaches an
account's persistent fingerprint: sustained 429s without reset headers, 5s
fallback cooldown, pool collapse into a 503 storm, and no reset entry point.
Invalid values warn and fall back rather than being silently dropped.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0168uDMAj9anaJrnH8CeFze6
Google released Gemini 3.7 Flash on 2026-08-13. Register it alongside 3.8
using the same shape as the existing 3.6 entries: bare id plus the
Antigravity -high/-low/-medium/-tiered thinking tiers, the model catalog,
the custom-mapping passthroughs, fallback pricing and the tier alias
normalizer.
Pricing matches 3.8 at $0.75 input / $3.75 output / $0.075 cached input
per MTok. Both are promotional through 2026-12-31 and double on
2027-01-01; the constants note this.
Also corrects the 3.8 catalog timestamps to the documented GA dates.
Google released Gemini 3.8 Flash on 2026-09-02. Register the model and its
Antigravity thinking tiers the same way Gemini 3.6 Flash is registered:
- DefaultAntigravityModelMapping: bare id plus -high/-low/-medium/-tiered
- geminiModels catalog exposed through /antigravity/v1/models
- ensureAntigravityDefaultPassthroughs so accounts with a custom
model_mapping still resolve the new ids
- fallback pricing and the thinking-tier alias normalizer, so the tiers
share the base model price card instead of failing pricing lookup
Fallback rates are Google's published prices: $0.75 input / $3.75 output /
$0.075 cached input per MTok. These are promotional through 2026-12-31 and
double on 2027-01-01; the constant carries a note to that effect.
Refs #6523
A collapsible sidebar group was force-expanded whenever the active route
was one of its children, so clicking the chevron did nothing: the toggle
only updated the internal set while isGroupActive() kept reporting the
group as expanded. Any click that produces no visible change reads as a
bug to users.
Track per-group expand/collapse overrides instead. A group without an
entry keeps the automatic behavior (expanded while active, so context is
not lost on load or navigation), but a chevron click records the user's
choice and that choice wins, letting an active group be collapsed and
re-expanded at will.
Accept service_tier=ultrafast on the OpenAI/Codex gateway, advertise it
on GPT-5.6 Sol in the Codex model catalog, and include it in ChatGPT
OAuth routing hints so Codex clients can select Ultrafast.
Follow-up to the event-time proxy snapshot: every upstream attempt now
carries attribution, the direct/unknown contract is enforced everywhere,
and the queued upstream_errors array has hard bounds again.
Attribution coverage
- compact fallback retry helper and the streaming compact failover event
now stamp proxy_id/proxy_name; the non-streaming SSE compact retry
previously dropped its first attempt on `continue` and now records it
- raw chat-completions stream truncation events (new on main) attributed
- Anthropic/Bedrock transport errors get attribution inside the shared
handleUpstreamTransportError helper, so all five callers inherit it
- OpenAI first-output timeout takes attribution from the caller: HTTP
callers pass the HTTP rule, the WS passthrough relay passes the WS rule
(no managed proxy => unknown, never direct, because http.DefaultClient
may honor environment proxies)
Contract hardening
- single opsUpstreamProxyAttribution returns (id, name) from one decision;
a hydrated Proxy without a durable ID is unknown, not id=null+name
- both normalizers (struct at append time, JSON on detail reads) enforce
proxy_id==null <=> sentinel name; blank managed name => "proxy"
- detail-read normalization edits only the attribution keys via sjson, so
keys from older struct versions and stored key order are preserved
Queue bounds
- keep at most 256 attempts and ~512 KiB serialized per entry, newest
first; the oldest retained event records dropped_earlier_attempts
- attempts outside the 16-event body window trim url/message to 512 bytes
- restore the empty-event drop rule, judged before the body window clears
detail so older real attempts survive
Tests cover each site above (managed proxy, direct, WS unknown), the
normalizer invariants, key/order preservation, and the count/byte bounds.
Long-running LLM inferences (such as high-thinking Gemini models) may experience
silent TCP disconnections across intermediate proxies, NAT gateways, or firewalls
when no data frames are emitted for tens of seconds to minutes.
Go's default http2.Transport sets ReadIdleTimeout=0 (disabling active PING checks).
This commit introduces HTTPUpstreamProfileLongStream with:
- ReadIdleTimeout = 10s
- PingTimeout = 5s
When a pooled H2 connection stays idle for 10s during long streams, active PING frames
are dispatched. Dead connections failing to respond within 5s are proactively evicted,
preventing client generation requests from hanging on dead sockets or failing with EOF/502.
Signed-off-by: WUJI-Labs <zoujincool@gmail.com>
The streaming forward loops treated upstream EOF as the only normal exit
condition; response.completed/[DONE] only set a flag. With keep-alive /
HTTP/2 connection reuse to the upstream, the upstream may hold the
stream open for 8-46s after the terminal event, during which the
gateway could only emit downstream keepalives, inflating tail latency.
Finish the stream as soon as the terminal event frame has been fully
written downstream (usage is already parsed from the terminal event),
and close the upstream body to unblock the reader goroutine. Applied
to the guarded/async, sync, and passthrough streaming paths. The WS
HTTP bridge already behaved this way.
Adds a regression test with an upstream that hangs after
response.completed.
Moonshot pay-as-you-go and Kimi For Coding both expose native
/v1/responses. Stop converting Kimi Responses traffic to Chat
Completions, and treat Kimi like DeepSeek for protocol selection while
keeping /v1/responses URL construction.
classifySelectionFailureError upgrades a selection failure to 429
rate_limit_error whenever the scheduler's compact reason string matches
model_rate_limited=(\d+) with a positive count. It does so unconditionally,
including when the fallback it was handed is the authoritative 404
model_not_found classification.
Selection failure reasons are additive, so both can appear at once, e.g.
"pool=9, filtered: model_not_supported=8 model_rate_limited=1": eight accounts
cannot serve the model at all and the ninth happens to sit in a per-model
cooldown. classifyNoAccountError only produces ModelNotFound through
DiagnoseModelAvailabilityForPlatform, a dedicated database query over
persistent eligibility (active + schedulable + model_mapping), so it has
already established that no account in the group can serve the model. One
transient cooldown does not make "All available accounts are currently
rate-limited" true.
The mislabel is not cosmetic:
- Retrying can never succeed, yet 429 explicitly invites a retry.
- Clients that treat 429 as a rate limit retry hard and swallow the body;
Codex surfaces only "exceeded retry limit, last status: 429", discarding the
one message that names the real problem. Operators then look for capacity
and rate limits instead of a model whitelist.
- Ops attribution becomes self-contradictory: classifyNoAccountErrorFromGin
has already marked the request business-limited for local model
configuration, while call sites gate markOpsRoutingCapacityLimitedIfNoAvailable
on ModelNotFound, which the override clears, so the same request is also
counted as routing capacity limited.
Return the fallback untouched when it is ModelNotFound. The 429 upgrade is
preserved for every other fallback, which is the case it was written for:
accounts that can serve the model exist but are all cooling down, where the
fallback is 503 and retrying is meaningful.
Claude-Session: https://claude.ai/code/session_01LwvyJVV2sdqiu41Z9JDNUq
filterSchedulerExtra is an allowlist and never carried openai_passthrough /
openai_oauth_passthrough, while credentials.model_mapping is preserved in full.
Candidate filtering (ListSchedulableAccounts) reads that trimmed projection
(sched:meta:<id>), so there a passthrough account looks exactly like a
whitelist-mode one: Account.IsModelSupported falls back to the model_mapping
whitelist and the account is dropped as model_not_supported. Forwarding, which
hydrates the full account from sched:acc:<id>, still honors passthrough - hence
"testing the account directly succeeds, but the gateway reports no available
accounts supporting model X".
The practical impact is that a stale leftover whitelist silently shrinks the
candidate pool of accounts that were explicitly configured to let the upstream
decide. When no account in the group lists the requested model, the group is
fully dark for it even though every account can serve it.
This is not a regression of #4936: that fix corrected the predicate
(Account.IsModelSupported short-circuits passthrough before the mapping check),
this one restores the input the predicate is fed. The projection allowlist
predates #4936 by three and a half months and has never carried the flag, so
the fix simply never took effect on the snapshot path.
Add both keys to the allowlist, plus a regression test that round-trips the
projection through JSON (the real sched:meta write/read path) and asserts the
model gate still lets a non-whitelisted model through for a passthrough account.
Claude-Session: https://claude.ai/code/session_01LwvyJVV2sdqiu41Z9JDNUq
Clients now send body.fallbacks; upstream rejects it without
anthropic-beta: server-side-fallback-2026-07-01. Mirror the
context_management sanitize so OAuth mimic / default API-key
beta paths no longer 400 with "fallbacks: Extra inputs are not permitted".
OpenAI/Composite groups can now reject requests that exceed the max
reasoning effort instead of always rewriting them down to the ceiling.
Existing groups keep the previous auto-downgrade behavior.
Allow OpenAI/Composite group mappings to match by exact model name,
prefix, or suffix. Empty type and model still apply to every model.
The admin form groups multiple request-to-forward pairs under one
type+model scope.
CN provider accounts (Kimi/Zhipu/DeepSeek) and the monitor quota view each
hand-rolled their own usage bar, so the same account column rendered three
different label styles, bar widths and threshold palettes. Both now reuse
UsageProgressBar, the component the other five platforms already share.
Also move the default utilization thresholds from >=100 red / >=80 amber to
>=90 / >=75. Turning red only once the quota is exhausted is an after-the-fact
notice rather than a warning. The remainingCapacity mode keeps its 20/50 pair,
which is already a warning semantic.
- UsageProgressBar: add labelWidth 'fixed' | 'auto' so the monitor view can
render composite labels ("Pro/7d") without breaking column alignment; the
fixed branch is byte-identical, so the existing platforms are untouched
- drop the superseded utilizationColor/utilizationTextColor/formatReset copies
- shorten CN window labels to 5h / 7d and remove the now-dead resetSoon keys
The cache token breakdown tooltip in `UsageStatsCards` was hidden with
`opacity-0`. An element with zero opacity is invisible but still takes part
in layout, so its fixed `w-56` (224px) width kept contributing to the
document's scrollable overflow.
The tooltip is anchored with `left-1/2 -translate-x-1/2`, and on mobile the
stats grid falls back to `grid-cols-2`, which puts the trigger icon close to
the right edge of the viewport. Centering a 224px box on that point pushes it
roughly 26-49px past the right edge, so `/usage` and `/admin/usage` gained a
phantom horizontal scroll with nothing visible to explain it.
Switching the hidden state to `hidden` (`display: none`) takes the tooltip out
of the flow entirely, while `group-hover:block` / `group-focus:block` keep the
existing hover and keyboard-focus behaviour. The 150ms opacity transition is
dropped, since `display` cannot be transitioned.
Verified over Chrome DevTools Protocol at 360/390/414px: `documentElement.scrollWidth`
goes from 424 back to 390 (equal to `clientWidth`) with no other overflowing
elements remaining, and the tooltip still renders at 224x86 on focus. A sweep of
all 39 application routes confirmed these were the only two affected pages.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
grok-build CLI recap-style side-calls (turn-summary / title-refresh) send
a fresh per-call X-Grok-Conv-Id label but set body prompt_cache_key to
the parent session id, so a direct xAI connection shares the main turn's
server-side prompt cache prefix.
sub2api's explicitGrokCacheSeed preferred X-Grok-Conv-Id over the body
key, which fragmented the cache identity on every side-call and forced a
full-price replay of the whole conversation (~300K+ tokens each, ~2.3%
cache hit rate observed over 44 side-calls / 4.9M input tokens in 7
days — about a third of all grok-4.6 input).
Priority is now: CC session > generic session headers > body
prompt_cache_key > X-Grok-Conv-Id > explicitKey > previous_response_id.
IDE header-stability tests updated; new regression test locks in the
side-call/parent key-sharing contract.
Team GLM Coding Plan accounts fail the quota probe with
"当前用户不存在coding plan" because the personal-plan query path
does not carry team context.
- quota probe: when credentials contain zhipu_organization, switch the
Zhipu quota endpoint to the team variant (?type=2 + bigmodel-organization
/ bigmodel-project headers); personal plans keep the existing path
- account create/edit modals: optional organization/project ID inputs for
zhipu coding-plan accounts, persisted in credentials; edit modal
backfills them and clears back to personal plan when emptied
- add a click-to-open tooltip tutorial (zh/en) explaining how to grab the
IDs from the bigmodel.cn web console via the /api/biz/v1/organization
network request, with a sample URL mapping org-/proj_ segments to fields
- Treat TempUnschedulableUntil/RateLimitResetAt/OverloadUntil as source of truth during account selection
- Drop stale local account-level runtime blocks with generation+deadline CAS
- Leave model-scoped transient and Grok quota blocks unchanged
The passthrough streaming loop buffers leading events in `pendingLines` until
the first visible output, so a client receives **zero bytes** — not even the
HTTP response headers, since gin's ResponseWriter only commits them on the
first write — for as long as the upstream takes to produce its first visible
event.
The Forward path already recognises this exact hazard and solves it with a
downstream keepalive:
// Track downstream writes separately from upstream reads: pre-output
// failover can buffer response.created / response.in_progress, so
// keepalive must be based on downstream idle time.
The passthrough path has the same buffering but never starts a keepalive.
Reasoning models routinely think for several minutes before emitting a visible
event, so intermediate proxies close the connection on idle timeout.
Production evidence from one deployment (12h window, gpt-5.6 family):
* 44 `/v1/responses` requests produced their first visible output only after
600–900 s, each carrying 30k–50k output tokens — the upstream had finished
the work.
* Every one of them was killed by an intermediate nginx `proxy_read_timeout`
(600 s) and returned 504. The user received nothing.
* Cutting early does not even save resources: the code deliberately keeps
draining the upstream for usage after the client disconnects.
Fix: start the existing keepalive helper for the passthrough stream as well.
* `StartOpenAICompactSSEKeepalive` keeps its compact-marker gate; the gate-free
body is extracted as `startOpenAISSEKeepalive` for callers that already know
they are in an SSE context.
* The passthrough loop starts it when `gateway.stream_keepalive_interval > 0`
and stops it immediately before its first real write, so the ResponseWriter
is never shared.
Pre-output failover is unaffected: keepalive bytes are already excluded by
`OpenAICompactKeepaliveAdjustedWrittenSize`, which `openAIStreamClientOutputStarted`
on this very path already consults (#3887). No pendingLines and no
account-specific headers are ever emitted by a keepalive.
Behaviour is unchanged when `gateway.stream_keepalive_interval` is 0.
Public groups have always been bindable by every user: CanBindGroup returned
true for any non-exclusive group, and user_allowed_groups only ever carried the
exclusive groups an admin had granted. Admins had no way to hand a single user
a subset of the public groups short of converting a group to exclusive, which
changes it for everyone already using it.
A user now carries restrict_public_groups. While it is false, which is the
default and what every existing row migrates to, nothing changes: every public
group stays bindable. Once an admin turns it on for a user, that user's public
groups are narrowed to the ones listed in user_allowed_groups, the same table
that already gates exclusive groups.
The flag is an administrative control, so it rides on the admin user DTO only
and leaves the shape of the end-user endpoints alone.
The model plaza filter honours the flag as well, so a restricted user is not
shown groups they would be refused when binding a key. Anonymous visitors have
no user record and keep the previous view.
Enforcement rides on the existing CanBindGroup choke point, so it covers key
creation, key updates, and per-request authorization together. An API key bound
to a group that is later withdrawn stops working at request time rather than
lingering as a key that can be listed but not used.
The admin dialog gains a toggle over the public group list. Turning it off
re-checks every public group, so an admin cannot save a list that reads as
restrictive while the restriction itself is disabled.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
When running a custom build or pre-release version (e.g. v0.1.183-custom),
parseVersion failed to parse the patch segment because of the non-numeric suffix,
causing compareVersions to incorrectly report an available update.
This commit strips any hyphenated suffix before splitting SemVer components.
Signed-off-by: WUJI-Labs <zoujincool@gmail.com>
Merge origin/main into contrib/routed-codex-model-catalog.
Single conflict in openai_gateway_forward.go: main added
ClearActualOpenAIUpstreamEndpoint + SetActualOpenAIUpstreamEndpoint
(commit 4795650d2) at the same insertion point where the PR added
filterOpenAIResponsesNoneReasoningEffortForAccount. Both changes
are independent — endpoint observability first, then body filtering.
Copy cached API-key manifests before group-specific mutation, use DeepSeek
model IDs for Codex fallbacks, omit unsupported config.toml effort, and
drop wildcard mapping keys from generated catalogs.
Co-authored-by: Cursor <cursoragent@cursor.com>
Keep catalog membership on schedulable accounts and intersect advertised
capabilities across all active group members that map an alias. Preserve
upstream plugin, service-tier, and models-list body-limit changes.
Co-authored-by: Cursor <cursoragent@cursor.com>
Intersect advertised alias capabilities across all active group members
that map the model, not only currently schedulable accounts, so rate-limits
cannot widen image input or context windows.
Co-authored-by: Cursor <cursoragent@cursor.com>
Users only see the reasoning effort they requested. Admins still see
the requested value plus the forwarded mapped value, matching the
model column's requested vs upstream split.
The raw Chat Completions passthrough returned a nil error whenever the upstream
body ended, so an upstream that cut the SSE stream mid-generation surfaced as
HTTP 200 with usage 0/0: the client kept a half-finished answer, the request was
not counted as an SLA failure, and ops saw nothing. Observed against
opencode.ai/zen/go (deepseek-v4-pro): 11 of 1346 streams in 24h recorded a first
token but no final usage, alongside a concurrent `connection reset by peer`
against the upstream edge.
Track whether the upstream ever emitted a terminal signal — [DONE], a usage
chunk, or a finish_reason — and treat a stream that ends without one as an
upstream failure. Matching only [DONE] would misfire on compatible upstreams
that EOF right after their last chunk and would drop billing for tokens that
were actually delivered, so all three count.
- nothing written to the client yet: UpstreamFailoverError, retried on another
account with no client-visible half stream
- bytes already on the wire: typed openAIUpstreamStreamReadError, so the
handler appends an SSE error frame and marks the request an SLA failure
Client cancel/disconnect keeps the previous semantics — there the upstream read
error is indistinguishable from a truncation, so partial usage is still billed
rather than blamed on the account.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
normalizeOpenAIParallelToolCallsWithoutTools only looked at the top-level
"tools" array, but normalizeOpenAIResponsesLiteTools moves namespace tools
into an input item of type "additional_tools" and deletes the top-level key.
A Responses Lite request that carries tools therefore looks like it has none,
and the parallel_tool_calls:false that ensureOpenAIResponsesLiteParallelToolCalls
had just pinned gets deleted on the way out.
OpenAI then applies its default of true and rejects the request:
400 unsupported_value: "X-OpenAI-Internal-Codex-Responses-Lite requires
`parallel_tool_calls` to be false."
Note the field cannot simply be pinned to false unconditionally: without tools
OpenAI rejects it with "'parallel_tool_calls' is only allowed when 'tools' are
specified", so the two constraints have to be honoured together.
Reuse the same tool-detection standard the Lite path already uses by adding
openAIRequestBodyHasTools, the []byte counterpart of openAIResponsesLiteHasTools,
so both sides of the repo agree on what "has tools" means.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The upstream names one offending index per response, but a replayed
conversation routinely carries dozens of items of the same type, each with a
status the upstream schema does not accept. Clearing a single index per round
trip needs one retry per item, so a conversation with more than
maxOpenAIResponsesRejectedFieldRetries such items exhausts the bounded budget
and the 400 reaches the client. Reported against tool_search_output items,
where the rejection surfaced as "Unknown parameter: 'input[60].status'".
Clear the status of every input item sharing the rejected item's type in the
same pass. Items of other types keep theirs: the rejection only proves that
the rejected item's type has no status field. When the rejected item carries
no type to match on, fall back to clearing the named index alone.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Add gateway.models_list_read_max_bytes with the existing 8 MiB behavior as its default, and apply it consistently to generic, Codex, and Antigravity model-list reads.
Read one sentinel byte for Codex manifests so oversized responses return an explicit bounded upstream error instead of malformed JSON.
- cc_pipeline: observe Chat Completions chunks/bodies as untyped payloads
(empty event type) so the upstream-echoed service_tier is trusted, matching
the upstream constraint that only terminal events and untyped bodies report
the actual processing tier.
- tests: add model field to terminal SSE frames (observation only triggers on
model-bearing frames) and assert response.created tier echo is ignored.
- openai_ws_forwarder_ingress: resolve billing tier from the local
upstreamResponseModelObserver (upstream echo first) instead of the raw
request payload, matching the HTTP->WS bridge and WS v2 forwarder.
- openai_ws_http_bridge_test: add fast-alias + upstream default case
proving the local observer's echoed tier wins.
- handler tests: keep only invalid service_tier -> 400 (short-circuits at
validation); valid/omitted semantics covered by the pure service-level
validation tests, avoiding real account selection in tests.
- Accept fast|priority (canonical priority), flex|auto|default|scale on
/v1/responses and /v1/chat/completions; reject unknown/empty/non-string
with HTTP 400; omitted and null stay compatible.
- Propagate service_tier through JSON/SSE, Responses<->Chat conversions,
fallback paths and HTTP->upstream WebSocket bridge.
- Billing prefers the upstream terminal tier; the outbound (policy-
transformed) tier is used only when upstream omits the field.
Explicit upstream default bills Standard even when Fast was requested.
- Pricing: Fast premium 2x Standard for gpt-5.6-sol/terra/luna and
gpt-5.4; 2.5x for gpt-5.5; channel FastMultiplier stays authoritative.
- Live verification (official Codex 0.149.0 + gateway, HTTP & WS):
upstream ChatGPT backend may return terminal default even when the
account catalog advertises priority; billing follows the actual tier.
DashScope/DeepSeek later tool_call deltas send empty id and
function.name. Clients that merge with !== undefined overwrite
the first delta's identity and dispatch unknown tool "". Drop
those empty fields on the raw Chat Completions SSE path.
DOMPurify <=3.3.1 (and the mermaid-transitive 3.3.3) carry ~18 disclosed
sanitizer-bypass/XSS advisories, including GHSA-cj63-jhhr-wcxv
(CVE-2026-65913): with USE_PROFILES enabled, ALLOWED_ATTR is rebuilt as a
plain array and looked up via ALLOWED_ATTR[lcName], so a polluted
Array.prototype property (e.g. onclick) is treated as an allow-listed
attribute and survives sanitization -- this app calls
DOMPurify.sanitize(svg, { USE_PROFILES: { svg: true, svgFilters: true } })
in src/utils/sanitize.ts, whose output is rendered via v-html in
ImageUpload.vue's SVG upload preview.
Bumped to 3.4.14 (latest, OSV-clean) and pinned via pnpm.overrides so the
mermaid-transitive copy dedupes to the same patched version instead of
staying pinned at 3.3.3. Lockfile-only regen via pnpm 9, no other package
changes.
A terminal event that arrives with an empty output was rebuilt from delta
accumulation. BufferedResponseAccumulator models only one reasoning item, one
message, and N function calls, and records no item id, status, or phase, so a
turn carrying several items collapsed into a single fabricated message: the
reasoning item disappeared, the real message id was replaced, and phase was
lost.
reconstructResponseOutputFromSSE already prefers the raw output_item.done
items over accumulation for buffered responses. The streaming path had no
equivalent because it never sees the whole body at once. Collect the raw item
of each output_item.done keyed by output_index and rebuild from those, falling
back to accumulation only when the stream reported no done item at all.
Items are stored as raw JSON, so vendor extensions and item types this gateway
does not model survive the rebuild verbatim.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Composite groups are a routing layer. The usage-side aggregation already
maps their traffic to the concrete account platform via
usageLogEffectivePlatformExpr, but the error aggregation read
ops_error_logs.platform verbatim. Pre-routing failures on composite groups
fall back to the group platform 'composite', which is never an enabled
monitor v2 config platform, so those error facts were filtered out of
every channel monitor v2 query and composite groups looked permanently
healthy.
Join groups/accounts in the error dedup CTE and resolve the effective
platform the same way as the usage side: account platform first, then a
real recorded platform, else 'unknown'. The window delete+rewrite
recompute rewrites historical facts with the corrected key.
Recognize GoUsageLimitError responses and parse both single and composite reset durations from their human-readable messages. Keep API-key accounts rate-limited until the parsed reset instead of falling back to the short default cooldown.
In Docker + cgroup v2 with no memory limit set, /sys/fs/cgroup/memory.current
returns a small container number while /sys/fs/cgroup/memory.max is "max".
readCgroupMemoryBytes then returned (used=<container>, total=0, ok=true).
collectSystemStats used that container "used" but, being unable to derive a
cgroup total, filled the total from the host via gopsutil. The dashboard then
computed container_used / host_total, e.g. ~60MB / 23GB ≈ 0.3% — wildly
understating real usage.
Fix: introduce resolveMemoryStats, which picks a single self-consistent
(used, total, percent) trio from ONE source. cgroup metrics are used only when
the cgroup exposes both a current usage AND a concrete limit (memory.max != max,
so total > 0); otherwise used/total/percent all fall back to the host reading.
The two sources are never mixed.
- memory.current valid + memory.max = "max" -> all host metrics
- memory.current = 512MiB + memory.max = 2GiB -> ~25% from cgroup
- no cgroup (bare metal) -> all host metrics
CPU metric behavior is unchanged (cgroup attempt then host fallback).
Adds ops_metrics_collector_memory_test.go covering all branches.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The Responses client-tool restore raised an upstream function_call back to
custom_tool_call and tool_search_call without touching its item ID, so the
client received a custom_tool_call carrying the upstream's fc_ ID. That
combination is invalid: once the item lands in the client's history, any
later replay to an upstream that validates item IDs fails the whole request
with "Invalid 'input[N].id': 'fc_...'. Expected an ID that begins with
'ctc'", and every following message re-submits the same bad item.
Re-prefix the ID to match the type it is raised to, keeping the suffix so
the ID stays stable and unique per upstream item. IDs without a known
tool-call prefix are left alone rather than guessed at.
The lowering direction has to reverse this, because the WebSocket HTTP
bridge replays restored items back to the upstream and would otherwise lose
the upstream's own ID. normalizeLoweredFunctionItemID now maps a ctc_/tsc_
call ID back to fc_ instead of dropping it, which makes the round trip
lossless; output item IDs (ctco_/tso_) still have no function-protocol
counterpart and are still dropped. The helper was renamed from
dropInvalidLoweredFunctionItemID to match what it now does.
The stream restorer keeps the upstream ID for its own bookkeeping so later
upstream events still match the call, and emits the client-facing ID on
output_item.added/done and the custom_tool_call_input events.
Closes#6071
Account model mappings decide availability, but generated Codex catalogs previously inferred capabilities from local model-name tables. Compatible upstreams can expose unknown models such as OpenCode x-preview-f-free, so reasoning levels, image input, and context limits were missing or wrong.
Sync capability metadata from the account model endpoint, enrich incomplete lists from the Models.dev provider matching the account base URL, and persist only complete snapshots. Mixed groups consume the safe intersection across schedulable accounts. When IDs sync without complete metadata, return an explicit warning and preserve the previous snapshot.
Third-party Responses providers often omit /models. If that endpoint returns 404 or 405 and the account already has concrete model mappings, use those configured upstream IDs for capability enrichment. Authentication, rate-limit, server, and network failures remain hard errors, and metadata is never guessed across providers by model name alone.
Advertise a single none choice for non-reasoning models so current Codex can select them, then omit that catalog placeholder when forwarding to compatible Responses upstreams while preserving official OpenAI request semantics.
The admin user edit modal rejected concurrency < 1, so a user whose
concurrency is already 0 could not be saved at all — the guard runs
before the request, blocking notes, password, role and RPM edits on that
user too.
Everywhere else already treats 0 as unlimited: the gateway skips slot
limiting when maxConcurrency <= 0 (ConcurrencyService.AcquireUserSlot),
the batch limits endpoint binds concurrency with min=0, and the bulk edit
modal only rejects negative values.
Reject negative and non-integer values instead, mirror the RPM field with
min/step and a "0 = unlimited" placeholder and hint, and rename the error
key to match its new meaning. Account concurrency is unchanged.
The quick-add parser rejected every IPv6 proxy: the host group [^:]+
cannot match IPv6 literals (colons) and the pattern had no bracketed
form, so lines like socks5://[2001:db8::1]:1080 were reported invalid.
Add a bracketed-IPv6 host alternative and strip the brackets before
storing; the backend re-brackets via net.JoinHostPort when building the
proxy URL. Bare (unbracketed) IPv6 stays rejected because it is
ambiguous with host:port. Also add a regression test.
OpenAI groups can mix OAuth accounts with Responses-compatible API key accounts whose upstreams do not implement /models. The manifest handler previously selected one account and fetched its catalog before merging model_mapping aliases, so a permanent 404 returned 502 and hid models the administrator had already configured.\n\nTreat explicit model mappings from schedulable accounts as the authoritative Codex catalog. Build that catalog locally before account selection, preserve group filtering and final-body ETags, and retain the existing upstream discovery and failover path only when no mappings exist. Composite groups keep their existing account-first catalog behavior.
Anthropic fallbacks previously reused Antigravity defaults, leaking Gemini models into Claude-only groups. Composite aliases were also emitted with generic metadata and could be scheduled to accounts that did not own the exact mapping, while generated configs could default to a model absent from the downloaded catalog.
Keep provider defaults separated, derive alias capabilities from unique upstream targets with conservative fallback, require exact mapping ownership during scheduling, filter media targets, and choose generated config defaults from the fetched catalog.
Composite model catalogs expose exact account model_mapping aliases, but request routing only considered explicit routes or model-name detection. Unprefixed aliases were selectable but unroutable.
Resolve unique exact account ownership before the detector, keep explicit routes authoritative, and fail closed when an alias is claimed across platforms.
Codex treats the downloaded model catalog as a capability contract. Slug-only entries hide supported reasoning levels, can advertise invalid image inputs, and expose automatic or media-only models in the model picker. Generate complete route-aware metadata, preserve official manifests, filter non-agent models, and compute validators from the final catalog.
Codex model discovery expects a top-level models manifest, but Composite and other non-OpenAI groups either reached the OpenAI live-manifest handler or had no Codex-specific response at all. API key users also had no supported way to fetch that manifest and reference it from config.toml.
Generate minimal manifests from each group's effective model list while preserving the official OpenAI live path and the ordinary /models response. Add the Use Key flow for authenticated catalog download and model_catalog_json configuration without writing the API key into the downloaded file.
The model plaza route already supports public access, but both built-in /home headers omit its entry. Add the link to compact and default headers while keeping the existing feature and authentication settings authoritative, then cover the visibility matrix with focused component tests.
Constraint: Keep the change frontend-only and preserve router-owned access control
Rejected: Add the link to AppHeader only | /home renders its own headers and never mounts AppHeader
Confidence: high
Scope-risk: narrow
Reversibility: clean
Directive: Keep the model plaza entry gated by the existing opt-in flag and require-auth setting
Tested: HomeView focused Vitest, full frontend Vitest (223 files / 1554 tests), ESLint, vue-tsc, production build
Not-tested: Manual browser click-through against a running backend
Related: #5524
ConfigManager.refreshLoop reloads the Prompt Guard config every 5s and
Reload logged config_loaded on every successful load, so an unchanged
config produced up to ~17k identical lines per instance per day and
buried real configuration changes.
Log the event only when the reload carries news: the first snapshot, a
new config version (every admin save bumps it under the advisory lock in
UpdateConfig), a flip of the global risk control gate (a separate setting
that leaves the version untouched), or a recovery from a failed reload so
the degraded-to-healthy transition stays visible.
This mirrors logInvalidTokenEndpoints, which already warns once per
change rather than on every refresh.
CN-provider accounts explicitly configured with api_protocol=anthropic
(the frontend offers per-provider presets such as GLM Anthropic) fell
through the account-test router to the generic Claude tester, which:
- appended ?beta=true, which the native passthrough deliberately avoids
for third-party Anthropic endpoints, and
- defaulted a missing base_url to https://api.anthropic.com — sending
the provider's API key to Anthropic instead of the provider's own
Anthropic-compatible endpoint (guaranteed 401 plus cross-provider
credential exposure).
Route them to a dedicated probe that mirrors the adaptive Anthropic
check: resolve the endpoint via GetAnthropicProtocolBaseURL (same
resolution as real /v1/messages forwarding, including per-platform
defaults), use the shared API-key auth header, and mark 401/403 as the
account error state.
Also fail fast with an actionable hint when the anthropic-protocol
base_url still points at an OpenAI-compatible endpoint (paas path,
version segment, or chat/completions suffix): the naive {base}/v1/messages
join would 404 (e.g. /api/paas/v4/v1/messages) without any pointer to
the actual misconfiguration.
The quota cell used the noun label "5-hour window/weekly window" and the
balance cell the balance value itself as the click target for a manual
refresh. Both read as passive captions, so users could not discover the
manual probe and reported the feature as missing ("only OpenAI has a
refresh button").
- Split data display from the refresh action: bars/balance render as
static rows (snapshot already paints on mount), with a dedicated
action button below, aligned with the OpenAI "Query" and Grok
"Probe" buttons (same icon, blue action styling).
- Label the control with a verb (cnProviders.probe: 查询 / Query) and
give the balance cell a tooltip (it previously had none).
- Keep the tier/label layout classes and data-test hooks intact.
- Extend the quota cell spec and add a balance cell spec covering
snapshot rendering, explicit action labeling and failure passthrough.
sanitizeGroupMessagesDispatchFields forces AllowMessagesDispatch=false for
every non-openai platform including composite, and the gate checked only the
group platform — so composite requests resolved to grok/CN targets were
rejected 403 on /v1/messages and /v1/messages/count_tokens before reaching
the composite target whitelist, leaving the CN rollout unusable for
Claude-protocol clients.
- allowOpenAICompatibleMessagesDispatch: composite groups resolved to a
grok/CN target get the same exemption as the standalone platforms;
openai-resolved and unresolved targets keep requiring the switch
- resolveOpenAIMessagesDispatchMappedModel: skip the group-level dispatch
model mapping (openai-specific gpt defaults) for composite grok/CN
targets — model rewriting stays with account-level model_mapping
- Responses WebSocket composite whitelist stays openai+grok: CN accounts
cannot pass the WSv2 ingress transport filter and the WS HTTP bridge
has no Responses conversion for them, so admitting CN targets only
turns a clear policy rejection into a misleading no-available-account
- widen the responses/input_tokens and messages/count_tokens composite
gates to the shared openai-compatible text whitelist so CN targets get
the same local-estimate token counting as generation
- restore the Claude default-model fallback for standalone CN groups
(admin candidates and custom models list) while keeping composite
listings limited to CN account mapping keys
- refresh the scheduler bulk-rebuild comment and bucket capacity hints
for the 8-platform canonical set
The star history chart was broken due to GitHub stargazer API restrictions. Update the chart URLs in README.md, README_CN.md and README_JA.md so the chart renders again.
P2-2/P2-4/P2-5 from review:
- validateLinkedAccount/revalidateLinkedAccount now check
monitorAccountQuotaCapability after the platform match, blocking combos
that would permanently error at runtime: deepseek coding and
custom-domain kimi coding (no quota endpoint), zhipu payg (no balance
endpoint), anthropic/openai API-Key accounts (usage query requires
oauth; anthropic setup-token still allowed via local estimation).
gemini/grok/antigravity stay permissive. Quota mode surfaces
CHANNEL_MONITOR_ACCOUNT_NOT_SUPPORTABLE on edit; probe mode silently
unbinds (same as platform mismatch).
- applyMonitorUpdate re-runs validateCheckMode on the effective
provider+check_mode whenever either field is patched, so a
provider-only update can no longer persist antigravity+probe (the
check is conditioned on provider/check_mode presence so legacy illegal
rows can still be renamed/disabled).
- normalizeMonitorPrimaryModel only substitutes the "quota" placeholder
for pure quota mode; quota_probe with an empty model now fails with
MissingPrimaryModel instead of probing model="quota". The quota branch
also moves ahead of the grok default, so grok+quota now gets "quota"
rather than "grok-4.5" (review L-item).
Tests: capability matrix (15 cases incl. anti-over-blocking rows),
linked-account wiring, revalidate quota/probe split, provider-only
update bypass + legacy-row rename, quota_probe missing-model create,
normalize matrix.
P2-1/P2-3 from review:
- fetchCNQuota: credential-invalid now judged by StatusCode 401/403
(aligned with fetchCNBalance) instead of `!Success && !CredentialValid` —
CN quota service only sets CredentialValid=true on the success path, so
500/429/zhipu business errors were all misclassified as failed instead
of error.
- fetchCNBalance: snapshot carries new BalanceLow flag computed with the
scheduler's exact criterion (`!Available || allCNBalancesBelowThreshold`)
against Gateway.CNProviders.BalanceThreshold (ctor now takes cfg; wire
regenerated). quotaDegradedHint reports "balance low" instead of the old
`<=0` check, so an account already paused by the scheduler (balance 5 /
threshold 10) no longer shows green in the monitor.
- threshold helper falls back to viper default 0.5 for nil/<=0 config to
avoid a zero-threshold regression where balance=0 stops alerting.
Tests: CN quota status-code matrix (rewrites the test that cemented the
old behavior), balance-low matrix (below-threshold / unavailable /
multi-currency healthy), threshold-from-config; PayG stubs now set
Available explicitly (zero-value trap).
- ChannelsView platformOrder now includes the three CN provider
platforms so channel pricing can be configured for them; composite
group expansion/attachment stays limited to the five main platforms
(matches backend isConcreteRequestPlatform and composite-routes
target_platform validation)
- SyncPricingModels maps kimi->moonshot, zhipu->zhipu,
deepseek->deepseek; also fixes gemini mapping to "google" which
matched zero catalog entries (provider key is "gemini")
- Add CN platform colors to channel pricing model tag classes
- Replace ApplyCodexCanonicalIdentity with CodexCanonicalAuthIdentity /
ApplyCodexCanonicalAuthIdentity: the credential face (auth.openai.com
token exchange / refresh / PAT whoami) now sends the originator +
canonical User-Agent pair and no version header, matching codex-rs
default_headers(); the version gate (#3901) only exists on the
/backend-api/codex inference face. whoami keeps its original header
shape (originator + UA) with the canonical UA source.
- Token exchange and refresh send the full pair instead of a bare UA,
eliminating the half-identity (UA without originator) combination no
real client ever emits.
- Codex models manifest: the Version header now follows the client's
own client_version when it is valid and >= the upstream floor (same
source as the query param, restoring the pre-refactor consistency),
falling back to the canonical version otherwise; the query param
keeps its verbatim passthrough contract.
- Drop the now-unreferenced openAICodexProbeVersion constant and its
vacuous consistency assertions; probes resolve their version through
resolveCodexOutboundIdentity at runtime.
The struct field alone never reached the wire: the raw passthrough
pipeline is covered by enableMixedGeminiToolInvocations (#5711), but
TransformClaudeToGeminiWithOptions builds GeminiToolConfig from scratch
and never set the flag, so gemini-* models entering through the Claude
format gateway could still hit the upstream 400 from issue #5709.
- Set IncludeServerSideToolInvocations=true when the built tool
declarations mix functionDeclarations with googleSearch, matching the
raw-path injection semantics.
- Replace the marshal-roundtrip-only test with behavior tests that
drive TransformClaudeToGeminiWithOptions: mixed tools set the flag,
function-only and web-search-only requests leave it unset.
- user_repo.create(): keep the TxFromContext fast path, but restore
tolerance for dbent.ErrTxStarted in the self-owned-transaction branch.
ent's Client.Tx only inspects the driver type, so a repository built
from a tx-bound client (client-injected transactions, e.g. the
integration fixture testEntTx + tx.Client()) hits ErrTxStarted; reuse
that client instead of failing. Fixes the two red integration tests in
allowed_groups_contract_integration_test.go.
- createUserAndClaimInvitation: roll back via defer (matching the OAuth
registration precedent) so a panic inside the transaction cannot leak
the connection.
- settingRepoStub: guard call counters and state with a mutex; the new
concurrency regression test exercises it from multiple goroutines and
the unsynchronized counters were flagged by -race.
Reduce filtered admin usage statistics from four scans to one GROUPING SETS query so every breakdown shares the exact same filters. Add concurrent expression indexes for requested and upstream model filters on large usage_logs tables.
Token exchange, PAT whoami, models, probes, and pre-writes now follow
the same UA/version chain as Codex inference instead of hardcoded
codex-cli/0.91.0 or compile-time constants.
PR #1463 removed the Sora platform, but some references survived:
- README/README_CN/README_JA kept the 'Sora status (temporarily
unavailable)' sections and gateway.sora_* docs that the removal PR
never touched.
- deploy/config.example.yaml still documented ~130 lines of sora_*
gateway keys, the top-level sora: direct-client/storage block, and
token_refresh.sync_linked_sora_accounts - none of which map to any
field in the config structs anymore.
- The OIDC login PR (02a66a01c, branched off pre-removal main and
merged 4 days after #1463) re-added the dead
PublicSettings.SoraClientEnabled field, which no code ever sets.
- A release sync (748a84d87) re-introduced sora i18n keys that the
later i18n split (d9e514f98) faithfully carried into
locales/{zh,en}/admin/{overview,settings}.ts. No component references
any of these keys.
This drops all of the above. Pure deletions, no behavior change.
- Add IncludeServerSideToolInvocations field to GeminiToolConfig to prevent dropping client tool settings.
- Fix HTTP 400 error when mixing built-in tools (e.g. Google Search) with function calling on Gemini 3.6/3.7 models.
- Add serialization/deserialization unit test TestGeminiToolConfig_IncludeServerSideToolInvocations.
Fixes#5709
RegisterWithVerification checked CanUse() and then marked the code used in
two separate, non-transactional steps; the second step's failure was
swallowed ("invitation code mark failure does not affect registration").
Concurrent registrations with the same invitation code could all pass the
check and each create an account, turning a one-time invitation code into
an unlimited account factory (TOCTOU race).
Fix:
- AuthService: create user and claim the invitation code inside one DB
transaction (createUserAndClaimInvitation). The claim reuses
redeemRepo.Use's conditional UPDATE (WHERE status='unused'); losers are
rejected with INVITATION_CODE_INVALID and their transaction (including
the user insert) is rolled back. No-code registration path unchanged.
- userRepository.create: explicitly join an outer ent transaction via
TxFromContext instead of relying on Client.Tx returning ErrTxStarted
(ent's Tx never inspects the context, so the old reuse branch was dead
code and user inserts always committed in their own transaction,
leaving orphan users behind when the outer transaction rolled back).
Regression tests:
- unit: concurrent register with one invitation code must succeed exactly
once (8 goroutines -> 1 success, 7 x INVITATION_CODE_INVALID)
- integration: outer-tx rollback removes user and releases the claim;
commit persists both atomically
89d826be2 raised backend/go.mod to `go 1.26.6` and updated the three CI
workflows' version assertions, but left the Go builder image in all three
Dockerfiles pinned at 1.26.5. Since the official golang images set
GOTOOLCHAIN=local, the toolchain is not auto-downloaded and any image build
fails hard at `go mod download`.
CI does not catch this: the workflows build with actions/setup-go, not with
these Dockerfiles.
Also extend the Go-upgrade checklist in DEV_GUIDE.md, which listed only the
CI files -- that omission is why the Dockerfiles were missed.
Default codex_fingerprint_mode to off. v0.1.175 treated a missing key as
"session", so upgrading silently rewrote installation/session/thread/turn/
window identifiers for every existing OAuth account that had never configured
this field. The quota regressions in #5555, #5556 and #5582 line up with that
version boundary, with A/B reports that rolling back to v0.1.173 restores
quota. Convergence is now explicit opt-in (#5610).
Only accounts that never set the field change behaviour; explicit off /
device / session / full keep working exactly as configured. That required
flipping the persistence condition in all three account modals from
"!== 'session'" to "!== 'off'": the old rule deleted the key when it equalled
the default, which after the flip would have silently discarded an
administrator's explicit opt-in to session.
Also extend convergence to the passthrough path, which previously left client
identifiers untouched:
- resolve the ids once in forwardOpenAIPassthrough and rewrite
client_metadata on the raw bytes (gjson extract + sjson splice) because
passthrough is a hot path that must not fully unmarshal multi-MB bodies;
a shared core keeps the raw and map variants from drifting
- both request builders apply the staged ids at the same relative position
(after session isolation, before identity enforcement) so headers and body
share one id set and turn_id stays consistent
- stage the ids unconditionally, including nil: a failover from a converged
account to an off account must not leave the previous account's ids behind
OpenAI sunset the legacy unary /responses/compact endpoint (404, #5598,
#5624), so the account "compact probe" in the admin UI kept failing even for
healthy accounts, and the beta-feature negotiation header was only attached
to compaction turns.
Beta features (codex-rs session/mod.rs build_model_client_beta_features_header
+ client.rs build_responses_headers): the header is a session-level constant
attached to every /responses request, the WS handshake and /responses/compact.
Enumerating FEATURES shows no Experimental feature is enabled by default, so a
default install sends exactly "remote_compaction_v2". Mirror that:
- OAuth requests without a client-declared header get the default shape, so we
no longer produce a "header only on compaction turns" pattern real Codex
never emits (#5586 chains that strip the header)
- a client-declared header is preserved as-is: non-empty without v2 means the
user disabled the feature and the gateway must not rewrite that
- native v2 turns (compaction_trigger in body) always ensure v2 is present
- non-OAuth upstreams keep the compaction-turn-only behaviour
- the WS injection sits outside the client-header copy block so prewarm and
turn handshakes cannot land in different pool compatibility buckets
Compact probe now exercises native v2 (streaming /responses +
compaction_trigger) instead of the dead endpoint. Success requires an actual
compaction output item — scanning output_item.done/added, the terminal
response.output[] and the whole-JSON fallback — so a 2xx that silently drops
the trigger is reported as unsupported (the "got 0 items" class, #5478,
#5648). Probe identity is now UUID-shaped and applies the account's
convergence, matching real traffic on the same endpoint.
Codex captures x-codex-turn-state from /responses SSE, /responses/compact
JSON and the WS handshake (codex-api sse/responses.rs, endpoint/compact.rs),
then echoes it back on later requests of the same turn. The HTTP path dropped
it because the header is not in the generic response allowlist, while the WS
path already relayed it — an inconsistency that broke the protocol chain.
Relay it explicitly at every commit point instead of widening the global
allowlist (which would leak it into Anthropic/Gemini responses):
- streaming, non-streaming and SSE-to-JSON handlers relay it, clearing any
value left over by a previous failover attempt when upstream sends none
- under the first-output guard the header is only staged; provenance is
recorded when applyAttemptResponseHeaders actually writes it, because a
first-output timeout discards the staged headers and the client never
receives that blob
- record (api key + client session) -> minting account, TTL-bounded with an
opportunistic sweep, and strip echoes known to come from another account
before they go upstream. Stripping only: injection is the Claude bridge's
job. Same-account or unknown provenance passes through unchanged.
The responses->chat streaming converter re-emits the tool name as an empty
string on every function_call_arguments.delta because ChatFunctionCall.Name
has no omitempty. OpenAI-compatible clients accumulate the name from the
first tool-call delta, so the trailing "name":"" overwrites it and the
call fails dispatch (unknown tool ""). Empty names also break the next
round-trip upstream (missing field 'name').
Fixes#5631
The channel cache holds a groupID -> platform map with a 10 minute TTL, and
only channel Create/Update/Delete call invalidateCache(). Changing a group's
platform through the admin API therefore leaves the cache pointing at the old
platform for up to 10 minutes.
Channel pricing, model mapping and the model whitelist are all matched per
platform, so during that window the lookups silently miss: pricing falls back
to the global LiteLLM price list, renames stop applying and the whitelist
stops restricting. Nothing is logged.
Inject a narrow ChannelCacheInvalidator into the admin service (same shape as
the existing APIKeyAuthCacheInvalidator) and call it from UpdateGroup only when
the platform actually changed. The dependency is optional -- when it is nil the
cache simply rebuilds on TTL expiry, as before.
`validateNoConflictingModels` used `toModelEntry`, which only lowercases,
while `expandPricingToCache` keys the cache with
`normalizeChannelPricingModelName` (lowercase + TrimSpace + `.` -> `-`
for `claude-*`). Two pricing entries that the validator considers distinct
therefore collapse onto the same cache key, and the one written later
silently overwrites the other.
Add `toPricingModelEntry`, which reuses `normalizeChannelPricingModelName`,
and use it for pricing conflict detection. `toModelEntry` is left untouched
for model *mapping* validation, because `expandMappingToCache` keys the
mapping cache with plain `strings.ToLower` -- normalizing mappings the same
way would reject configurations that the cache keeps separate.
This completes the fix for #4754, which added the normalization to the
lookup side only.
Replace if-then-t.Fatal nil checks with testify require.NotNil so
staticcheck can prove the pointer is non-nil before field access.
Fixes SA5011 in payment_handler_test, parse_test, and
openai_images_incomplete_test.
Resolve conflict in backend/internal/handler/openai_gateway_handler_test.go.
main and this branch each appended a passthrough upstream stub plus a test at
the same two insertion points:
main openAIHTTPPassthroughSSERateLimitUpstream
TestOpenAIResponses_APIKeyPassthroughSSERateLimitUsesConfiguredPoolRetry
branch openAIHTTPPassthroughAuthFailoverUpstream
TestOpenAIResponses_APIKeyPassthroughPoolAuthFailureRetriesThenSwitchesToHealthyAccount
Both sides are kept verbatim; the only edit is giving each stub its own
calls() body instead of sharing the trailing one. No assertion was changed.
openai_gateway_passthrough.go and openai_oauth_passthrough_test.go merged
automatically.
The security audit group (Risk Control + Prompt Audit) was hidden from
the sidebar via `hideInSimpleMode`, but nothing else in the stack
restricts it in simple mode:
- `router/index.ts` simple-mode `restrictedPaths` does not cover
`/admin/risk-control` or `/admin/prompt-audit`
- the `/risk-control` and `/prompt-audit` admin route groups have no
`RunMode` gate, and neither does `internal/securityaudit/` nor
`service/content_moderation.go`
- `SettingsView.vue` renders the risk control toggle together with a
`<router-link to="/admin/risk-control">` shortcut, and the settings
page stays visible in simple mode
The feature is therefore fully usable in simple mode and already
reachable through the settings page — only the sidebar entry was
missing. Drop the flag so the menu matches actual behaviour.
The group remains gated by `risk_control_enabled` (opt-in, default
false) through `featureFlag: flagRiskControl`.
Three guards on the response_model billing basis, all scoped to the opt-in
channel mode so existing channels are unaffected.
1. Per-unit billing gate was stale. Audio (AudioUsage) and the search
surcharge (SearchCount) reached the billing paths after this branch was
cut; both are priced per unit rather than per token, so they must be
excluded like image/video/web-search already are. Audio pricing ignores
the model entirely, so the previous code "adopted" a basis switch that
changed nothing and emitted a misleading audit log for it.
2. Never zero out a billable request. A catalog entry whose token prices are
explicitly 0 still passes the identified-pricing gate (TokenPricingAbsent
only means both prices are missing), so an upstream could declare a free
model name and drop the bill to zero. Reject a zero (or negative)
recomputation whenever the baseline was billable; an already-zero baseline
is unaffected.
3. Never cross from channel pricing to the global table. Channel pricing
matches exact keys and prefix wildcards and does not strip date suffixes,
while the global table's identified lookup does. Upstreams routinely
declare dated model IDs (claude-opus-4-5-20251101), so allowing a
cross-source comparison would silently bypass an administrator's channel
markup on essentially every request. Admins who want a downgrade target
discounted can price it explicitly on the channel.
Also skip the recomputation entirely when the declared model equals the
baseline: it is provably the same cost and only burned a pricing resolve.
The identified-pricing helpers now return whether the model resolved to
channel pricing so the third guard costs no extra resolve.
DEV_GUIDE.md and the three READMEs still advertise Go 1.25.7 and
golangci-lint v2.7, but CI has since moved on:
- backend/go.mod declares go 1.26.5, and backend-ci.yml / release.yml /
security-scan.yml all resolve the toolchain via
`go-version-file: backend/go.mod` and then hard-assert
`go version | grep -q 'go1.26.5'`.
- backend-ci.yml pins golangci-lint to v2.9.
So a contributor following DEV_GUIDE.md installs a linter two minor
versions behind CI (different findings locally vs. in CI) and expects a
Go version that the workflow's own assertion step rejects.
Update all ten stale references, and note in the CI section which files
the Go version assertion lives in so future bumps don't miss one.
Docs only, no code or workflow changes.
A Responses stream that ends with response.completed but carries no output,
no usage and no error (only response.created + response.completed observed)
is a silent upstream refusal. It was previously treated as a successful
turn, recording 0 input / 0 output tokens without triggering failover, so
clients saw an empty reply and usage was never billed.
Detect the empty terminal event in both the standard and passthrough
streaming paths while nothing has been written to the client, and return an
UpstreamFailoverError so the next account is tried; when failover is
exhausted the client gets an explicit Responses-format error instead of an
empty success. Streams with any semantic output, usage or error are
untouched. Fixes#5009.
OpenAI upstream requires reasoning input item ids to begin with 'rs' and
rejects replayed item_* ids with 400 ("Expected an ID that begins with
'rs'."). shouldStripOpenAIResponsesInputItemID handled message (msg) and
tool call (fc) ids but not reasoning items, so long Codex sessions restored
through the OpenAI API key passthrough path failed with a converted 502.
Strip non-rs reasoning ids while preserving valid rs_* ids, mirroring the
existing message/tool-call handling. Fixes#5410.
- Change loadRuntimeSnapshot failure from fail-open to fail-closed
- Return block decision with 500 status when config cannot be loaded
- Add regression test for snapshot failure scenario
- Fixes#5388
- Add 'composite' to imagePricingPlatforms set in groupsImagePricing.ts
- Allow Composite groups to toggle allow_image_generation in admin UI
- Backend already supports allow_image_generation field for all platforms
- Add test case to verify Composite platform is included
Fixes#5369
When OpenAI Responses streaming has compact keepalive commit 200 headers
but no semantic SSE event has been written, upstream 4xx errors were
silently swallowed, causing clients to hang until timeout.
The fix distinguishes between 'headers committed' and 'meaningful SSE
output written'. Compact keepalive may commit headers for heartbeat
purposes without writing actual SSE events. In that case, the Responses
stream still needs its protocol-correct terminal response.failed event.
Preserves existing behavior when a complete upstream response or
meaningful SSE payload was already communicated, and preserves image
keepalive semantics.
Fixes#5289
UpdateS3Config only encrypted the SecretAccessKey on the branch where the
caller supplied a new one. The other branch inherits the stored secret
through loadS3Config, which *decrypts* it — so that path wrote the
plaintext straight back to the settings row and overwrote the ciphertext.
It is the normal path, not an edge case. UpdateS3Config clears the secret
in its own return value, so the admin form's field is empty afterwards.
Any later save — changing the endpoint, saving the schedule, pressing save
twice — sends an empty secret and takes the inheriting branch. One extra
save is enough to leave the secret in plaintext at rest.
Nothing surfaced because loadS3Config falls back to the raw value when
decryption fails, for compatibility with pre-encryption rows. Backups kept
working and the only trace was a line per read:
[Backup] S3 SecretAccessKey 解密失败(可能是旧的未加密数据):
decrypt: cipher: message authentication failed
Seen on a live install: the stored value was 64 characters, exactly the
plaintext length of an R2 secret, where AES-GCM + base64 would have been
124. The compatibility fallback turns a one-time write bug into a
permanent plaintext secret.
It also feeds itself: pg_dump includes the settings table, so the backup
uploaded to the bucket carries the credentials for that same bucket.
Fix: hoist the encryption out of the if/else so it runs for a non-empty
secret whichever way it arrived. The ephemeral-key guard (#4524) moves
with it, and the no-secret-at-all case still skips encryption entirely, so
an update with no stored secret keeps working under an auto-generated key.
The existing KeepExistingSecret test passed either way, because it only
asserted loadS3Config returns the plaintext — true with the fallback too.
The new test asserts on the row itself: still ciphertext after a second
save, not double-wrapped after a third.
Checked the new test fails without the fix and passes with it.
The release workflow builds the frontend once, then runs each configured Go target on its own Linux runner. `CGO_ENABLED=0` permits cross-compilation. The target matrix is read from `.goreleaser.yaml`, including its exclusions; simple releases select Linux amd64 only.
Each build uses GoReleaser OSS in snapshot mode with the selected release version and one target. Archive naming, bundled files, Go flags and release templates remain in the existing GoReleaser configurations. Every archive is accompanied by its source commit, target, version and SHA256. The publishing job verifies the complete matrix before building images or publishing. It uses GoReleaser's `extra_files` support to publish existing archives and checksums, with builds disabled. No Pro license is needed.
Go caches are isolated by target and refreshed on each source commit, with fallback to the preceding target cache. Save uses the original restore key, even if a build hook changes `go.sum`. Matrix jobs upload uniquely named artifacts. The publishing job extracts only the regular Linux binary from each verified archive and restores its executable permission before constructing Docker contexts. QEMU remains limited to runtime-image instructions. DockerHub images are omitted when its credentials are absent; GHCR is always retained. Simple mode still publishes only the amd64 GHCR image and the simple release description.
All build jobs use the commit resolved by `prepare`, including a manual release's selected tag. Helper scripts come from the workflow revision and are passed as a run-local artifact, so older application tags do not need to contain the new scripts. The workflow serializes release runs to prevent simultaneous updates to moving image tags.
A dry run builds all selected archives and both runtime images, verifies artifact provenance and produces the final checksum file. It exports images locally as OCI archives instead of pushing them. It skips registry logins, GitHub Release publication, DockerHub description updates, Telegram notifications and VERSION synchronization. Test the simple path separately with `simple_release=true`.
Dry-run artifacts are available in the Actions run, including `release-dry-run-report`. Compare job start/end times, GoReleaser's build duration and cache restore results. Do not present an initial cold-cache run as a warmed-cache benchmark; publishing network time is not measured by dry runs.
<td>Thanks to APIKEY.FUN for sponsoring this project! <a href="https://apikey.fun/register?aff=SUB2API">APIKEY.FUN</a> is one of the core contributors to the sub2api open-source project, dedicated to providing open, stable, and cost-effective AI API access. The platform supports API relay services for Claude, OpenAI, Gemini, and other popular models, with pricing starting from as low as 7% of the original rate. Register via the exclusive link: <a href="https://apikey.fun/register?aff=SUB2API">APIKEY</a> to enjoy up to 5% off on all recharges.</td>
<td>Thanks to APIKEY.FUN for sponsoring this project! <a href="https://apikey.fan/register?aff=SUB2API">APIKEY.FUN</a> is one of the core contributors to the sub2api open-source project, dedicated to providing open, stable, and cost-effective AI API access. The platform supports API relay services for Claude, OpenAI, Gemini, and other popular models, with pricing starting from as low as 7% of the original rate. Register via the exclusive link: <a href="https://apikey.fan/register?aff=SUB2API">APIKEY</a> to enjoy up to 5% off on all recharges.</td>
</tr>
<tr>
@@ -58,6 +58,11 @@ Please read the following carefully before using this project:
<td>Thanks to AIGoCode for sponsoring this project! AIGoCode is an all-in-one platform that integrates Claude Code, Codex, and the latest Gemini models, providing you with stable, efficient, and highly cost-effective AI coding services. The platform offers flexible subscription plans, zero risk of account suspension, direct access with no VPN required, and lightning-fast responses. AIGoCode has prepared a special benefit for sub2api users: if you register via <a href="https://aigocode.com/invite/SUB2API">this link</a>, you'll receive an extra 10% bonus credit on your first top-up!</td>
<td>Real GPT-5.6 series at 3% of OpenAI pricing — <a href="https://codex-everywhere.com">CodexEverywhere</a> is democratizing access to frontier models for developers worldwide. We believe in transparency and honesty, with model quality verified by active community oversight for months. USD and crypto friendly. Start with a free $20 trial at <a href="https://codex-everywhere.com">codex-everywhere.com</a>.</td>
<td>Huge thanks to BmoPlus for sponsoring this project! BmoPlus is a highly reliable AI account provider built strictly for heavy AI users and developers. They offer rock-solid, ready-to-use accounts and official top-up services for ChatGPT Plus / ChatGPT Pro (Full Warranty) / Claude Pro / Super Grok / Gemini Pro. By registering and ordering through <a href="https://shop.bmoplus.com/?utm_source=github">BmoPlus - Premium AI Accounts & Top-ups</a>, users can unlock the mind-blowing rate of 10% of the official GPT subscription price (90% OFF)</td>
@@ -91,54 +96,23 @@ Please read the following carefully before using this project:
<td>Thanks to Suixiang AI Gateway for sponsoring this project! <a href="https://sui-xiang.com/">Suixiang AI Gateway</a> is a reliable and efficient API relay service provider offering relay services for Claude, Codex, Gemini, and more. A privacy-focused relay — no data reselling, no model dilution; privacy, transparency, and lightning-fast after-sales support. New accounts get ¥0.5 in trial credit daily by signing in; top-ups are 1:1, no subscription required, pay-as-you-go. Multi-line redundancy, cross-region disaster recovery, automatic failover, and uninterrupted long-link SSE. 99.9% availability — critical calls never fall behind.
<td>Thanks to Proxy4Free for sponsoring this project! Proxy4Free is a data proxy service provider for developers and AI applications, offering residential proxies, static residential proxies, ISP proxies, and datacenter proxies for scenarios such as Web Scraping, Browser Automation, and AI Agents. With global IP resources, stable connections, and flexible switching, it helps developers improve data collection success rates and reduce the risk of IP bans. Register via <a href="https://www.proxy4free.com/?keyword=4yjqecpc">this link</a> to get started and easily build more stable and efficient automation workflows.
<td>🎉 Thanks to FastAIToken for sponsoring this project! <a href="http://www.fastaitoken.com/register">FastAIToken</a> is an AI API aggregation platform for developers, supporting mainstream large models such as OpenAI, Claude, and Gemini. Top-up at 1:1 — 1 CNY = 1 USD of API credit — letting developers use the world's leading large model services at lower cost and with greater convenience.<br>
🚀 The platform offers a variety of channels to choose from: an ultra-low-price 0.02x OpenAI promotional group (limited time), groups as low as 0.25x OpenAI, 0.7x Claude with 95% fixed cache, and a 1.2x Claude Max channel. It also provides a public status page showing real-time availability, latency, and operating status of each group for transparent and reliable service, plus 7×24 human technical support (not bots) with fast responses to developer needs.
<td>Thanks to Aimzoon for sponsoring this project! <a href="http://aimzoon.com">Aimzoon</a> provides stable, cost-effective AI API access services, enabling developers to quickly connect popular AI services to coding tools such as Codex, Claude Code, and Gemini CLI. No complex configuration — faster onboarding, more stable calls, and lower costs. Ongoing promotions including discounted Codex rates and special pricing, with free trial credits upon registration, bringing AI coding into your daily workflow. <a href="http://aimzoon.com">Click here</a> to register and try it out!
<td>Thanks to Claude API for sponsoring this project! <a href="https://console.claudeapi.com/agent/register/drTKjyn6wGLK061Z">Claude API</a> is an official-channel API provider focused on Claude models. Built on official Anthropic keys and the official AWS Bedrock channel, it delivers a stable integration experience for Claude Code and Agent applications, supports the full Claude model lineup, and retains official capabilities such as Tool Use and long context. The service involves no reverse engineering and no model degradation, making it a great fit for heavy Claude Code users, Agent engineers, and enterprise engineering teams. Register via the <a href="https://console.claudeapi.com/agent/register/drTKjyn6wGLK061Z">exclusive link</a> and contact customer support to receive free trial credits; invoicing and team onboarding are also supported.
<td>Thanks to code0.ai for sponsoring this project! <a href="https://code0.ai/agent/register/LgpIgl9JHtVG53V1?utm_source=zcf&utm_medium=partner&utm_campaign=zcf_2026&utm_content=default">code0.ai</a> is an AI coding workbench for developers and engineering teams, aggregating mainstream agent coding capabilities such as Claude Code and Codex, and covering common development scenarios including code generation, project understanding, debugging and fixing, code review, and documentation generation. It suits independent developers, Agent engineers, open-source maintainers, and enterprise R&D teams, with invoicing and team onboarding supported. Register via the <a href="https://code0.ai/agent/register/LgpIgl9JHtVG53V1?utm_source=zcf&utm_medium=partner&utm_campaign=zcf_2026&utm_content=default">exclusive link</a> and contact customer support to receive free trial credits and experience a more efficient AI coding workflow.
<td><a href="https://nagora.ai/">Nagora</a> is a multi-model AI API gateway built for developers and teams. With a single account and API key, you can access more than 26 leading text and image models through one unified interface. It is compatible with OpenAI, Anthropic, and Gemini protocols and integrates seamlessly with development tools such as Claude Code, Codex, and Gemini CLI. The platform provides intelligent routing, automatic failover, transparent pricing, and consolidated billing, along with budget management, rate limiting, and concurrency controls. This makes AI usage more reliable and manageable across individual development, team collaboration, and production environments. No changes to your existing application are required. Simply replace the Base URL and API key to complete the integration in as little as one minute.</td>
<td>Thanks to <a href="https://www.novada.com/?sub2api/">Novada</a> for sponsoring this project! Novada provides residential, ISP, datacenter, and mobile proxies, along with Web Unlocker and Scraper APIs for developers building AI applications and automation workflows. With global IP coverage, flexible rotating and sticky sessions, and precise geo-targeting, Novada helps teams access web data reliably for AI agent workflows, cross-region testing, web research, and browser automation. Explore Novada to build more stable and scalable AI workflows.</td>
<td>Thanks to Qiniu AI for sponsoring this project! Qiniu AI is the enterprise-grade large-model MaaS platform under Qiniu Cloud (02567.HK), offering one-stop access to 150+ mainstream models worldwide, compatible with the protocols of major global model providers, and covering full-modality capabilities including text, image, audio, video, and file processing, serving over 1.69 million enterprises and developers. Qiniu AI offers an exclusive benefit for Sub2API users: register via <a href="https://s.qiniu.com/u6rQrq">this link</a> — enterprise users get 12 million tokens free, and developers get 3 million tokens free.</td>
@@ -164,6 +138,36 @@ Please read the following carefully before using this project:
<td><a href="https://hao.ai">hao.ai</a> is a high-speed, stable unified large-model API gateway for developers and teams. With a single API Key and a unified interface, you can access mainstream models such as GPT, Claude, and xAI Grok, with compatibility for common protocols and SDKs including OpenAI and Anthropic. The platform provides model routing, failover, team management, and complete request logs, with model prices as low as 15% of official reference pricing, helping users build AI applications more simply, more reliably, and at lower cost.</td>
<td>Swiftproxy is a high-performance proxy solution built for developers, providing stable and reliable residential and static residential proxy services. With 90M+ clean residential IPs, global coverage, flexible rotation, and precise geo-targeting, it helps projects such as web scraping, AI automation, browser automation, SEO monitoring, and multi-account management overcome access restrictions and improve workflow efficiency. It supports HTTP(S) and SOCKS5 protocols, integrates with popular automation tools like Playwright, Selenium, and Puppeteer, with dynamic proxy traffic that never expires until used and free testing available — <a href="https://www.swiftproxy.net/?ref=sub2api">start your free test now</a>!</td>
<td><a href="https://www.duckip.cn/?keyword=cu7oog6y">DuckIP</a> - 90M+ global residential network resources across 195+ countries and regions, with rotation and sticky sessions for public data collection, RAG updates, model evaluation, and multi-region data workloads. 🟢Residential Proxy - 20% Off; 🟢Static Residential Proxy - Starting at ¥50.00/IP; 🟢Unlimited Residential Proxy - Starting at ¥19.8/Hour. ✅Get 500M Free Trial.</td>
<td>Thanks to APIMart for sponsoring this project! <a href="https://go.apimart.ai/gh-sub2api">APIMart</a> is a low-cost API platform for AI image and video generation — GPT-Image-2 from $0.006 per image, with 160+ images per dollar. One async API covers both image and video: submit a task, get an ID, and retrieve results via polling or callback. Batch tens of thousands of images without timeouts, and switch models without changing code. Pay as you go with no monthly fee — <a href="https://go.apimart.ai/gh-sub2api">sign up here</a> to get started.</td>
<td>Thanks to AxisNow for sponsoring this project! <a href="https://www.axisnow.io/">AxisNow</a> protects and accelerates websites and APIs, delivering an optimal access experience across mainland China and globally, while extending acceleration and security capabilities to native/mobile apps through client SDKs — <strong>self-hosted private-deployment CDN</strong> | <strong>subscription-based DDoS-protected CDN</strong> | <strong>autonomous, flexibly composable CDN network</strong>.</td>
<td><a href="https://pp.dog/register?aff=SUB2API">PP.dog</a> is a source-level API gateway with its own massive account pool, purpose-built as an upstream relay for downstream transit stations and high-frequency developers — saving you the hassle of building and maintaining your own account pool. ✅ Direct from source: self-operated account pool with no middleman markup; 🧧 Cost crusher: combined rate multiplier as low as 0.03x, just 0.35% of official pricing; 🚀 Blazing fast: first-token latency < 1s, on par with the official native API. <a href="https://www.pp.dog/register?aff=SUB2API">Get started with PP.dog now</a></td>
<td>ColaProxy provides high-quality residential proxies built for web scraping, automation, and multi-account management. Get a free trial with non-expiring traffic, pricing starting as low as $0.3/GB, unlimited concurrent connections, and intelligent IP rotation for a smoother, more stable proxy experience. Use discount code COLA10 to get 10% off and start scaling your projects with reliable residential proxies today. <a href="https://colaproxy.com/?utm_source=sub2api&utm_medium=sub2api&ref=sub2api">Get started with ColaProxy</a></td>
</tr>
</table>
## Overview
@@ -196,7 +200,7 @@ Community projects that extend or integrate with Sub2API:
| Component | Technology |
|-----------|------------|
| Backend | Go 1.25.7, Gin, Ent |
| Backend | Go 1.27.0, Gin, Ent |
| Frontend | Vue 3.4+, Vite 5+, TailwindCSS |
| Database | PostgreSQL 15+ |
| Cache/Queue | Redis 7+ |
@@ -540,12 +544,6 @@ default:
rate_multiplier:1.0
```
### Sora Status (Temporarily Unavailable)
> ⚠️ Sora-related features are temporarily unavailable due to technical issues in upstream integration and media delivery.
> Please do not rely on Sora in production at this time.
> Existing `gateway.sora_*` configuration keys are reserved and may not take effect until these issues are resolved.
Additional security-related options are available in `config.yaml`:
-`cors.allowed_origins` for CORS allowlist
@@ -648,6 +646,32 @@ Or set `GATEWAY_OPENAI_WS_MODE_ROUTER_V2_ENABLED=true` in the environment.
Use `http_bridge` for client-WebSocket/upstream-HTTP operation when rolling out
or mitigating upstream WebSocket issues.
#### Force OpenAI upstream HTTP/SSE
When an egress proxy or network repeatedly reconnects OpenAI Responses
WebSockets, set the global fallback in the persisted deployment configuration:
```yaml
gateway:
openai_ws:
force_http:true
```
For Compose and Apple container deployments, the equivalent `.env` setting is:
```bash
GATEWAY_OPENAI_WS_FORCE_HTTP=true
```
This selects HTTP/SSE for OpenAI upstream Responses traffic that would
otherwise use WebSocket. It does not change the client-facing protocol or force
`GATEWAY_OPENAI_HTTP2_ENABLED=false`) separately when a proxy is incompatible
with HTTP/2. Unlike the account-level `http_bridge` mode, this global fallback
takes effect without enabling `mode_router_v2_enabled`. Keep the setting in the
deployment's persisted `.env` or `config.yaml`, rather than inside a running
container, so it is read again after an image update or container recreation.
#### ⚠️ Important: Creating the Admin Account
The initial admin account is **only created via the setup wizard** (served at `http://<host>:8080` on first run). The `default.admin_email` / `default.admin_password` fields in `config.yaml` are **not used** to create it — they exist in the template for historical reasons.
@@ -701,7 +725,10 @@ go generate ./cmd/server
Simple Mode is designed for individual developers or internal teams who want quick access without full SaaS features.
- Enable: Set environment variable `RUN_MODE=simple`
- Default groups are seeded on each startup. Set `SIMPLE_MODE_AUTO_CREATE_DEFAULT_GROUPS=false` (or YAML `simple_mode.auto_create_default_groups: false`) to manage groups yourself. The default is `true`; disabling it does not delete existing groups or change runtime auto-binding or admin concurrency setup.
- Difference: Hides SaaS-related features and skips billing process
- Optional key windows: Set `SIMPLE_MODE_KEY_RATE_LIMIT_ENABLED=true` to enforce each API key's configured 5-hour, daily, and 7-day spending windows. The default is `false`; balance and subscription debit remain bypassed when enabled.
- Window enforcement uses the database as its source of truth and records only API-key window usage. It is a post-request soft cap, so concurrent in-flight requests can overshoot by their final costs. Historical simple-mode usage is not backfilled.
- Security note: In production, you must also set `SIMPLE_MODE_CONFIRM=true` to allow startup
---
@@ -726,7 +753,7 @@ Sub2API supports both Grok subscription accounts through xAI OAuth and standard
- Codex CLI style Responses WebSocket ingress is accepted on the Responses targets and bridged to xAI HTTP/SSE Responses upstream
- Text models: `grok-4.5`, `grok-4.3`, `grok-build-0.1`, `grok-composer-2.5-fast`, `grok-4.20-0309-reasoning`, `grok-4.20-0309-non-reasoning`, and `grok-4.20-multi-agent-0309`
- Media targets for Grok groups: `/v1/images/generations`, `/images/generations`, `/v1/images/edits`, `/images/edits`, `/v1/videos/generations`, `/videos/generations`, `/v1/videos/edits`, `/videos/edits`, `/v1/videos/extensions`, `/videos/extensions`, `/v1/videos/{request_id}`, and `/videos/{request_id}`. Generation, editing, and extension requests require the group image-generation permission.
- Media models: `grok-imagine`, `grok-imagine-image-quality`, `grok-imagine-image`, `grok-imagine-edit`, `grok-imagine-video`, and `grok-imagine-video-1.5`
- Media models: `grok-imagine`, `grok-imagine-image-quality`, `grok-imagine-image`, `grok-imagine-image-2.0`, `grok-imagine-edit`, `grok-imagine-video`, and `grok-imagine-video-1.5`
- JSON image-edit and video-generation requests accept image references in `image`, `images`, `reference_images`, and `mask` objects. Use `url` for xAI-compatible payloads; the legacy `image_url` field remains accepted and is normalized to `url` before forwarding.
- Out of scope for this provider: TTS, transcription, browser automation, cookies, and Grok web scraping
@@ -794,7 +821,7 @@ xAI quota is passive. Sub2API does not invent subscription quota values; it reco
`401` responses temporarily remove accounts with invalid credentials from scheduling. `403` responses are treated as access or entitlement failures instead of token-refresh loops. `429` responses use `Retry-After` or a short cooldown to temporarily remove the account from scheduling.
New Grok image and video generation requests use a media-specific eligibility check. API-key accounts remain eligible. OAuth accounts require positive paid-entitlement evidence from the xAI billing probe; Free, forbidden, missing, malformed, and inconclusive billing observations are excluded from new media generation. Unobserved OAuth accounts are probed before the first media request is forwarded, and imports run the billing-first quota probe proactively. Chat requests and video status lookups are not affected by this media-only quarantine. If no eligible account remains, the media endpoint returns HTTP `503` with error type `grok_media_no_eligible_account`.
New Grok image and video generation requests use a media-specific eligibility check. API-key accounts remain eligible. OAuth accounts with explicit Free or forbidden billing evidence are excluded from new media generation. Missing or malformed observations are probed before dispatch; a successful but incomplete billing response is treated as `billing_inconclusive` and remains eligible for backwards compatibility, because an unknown billing schema is not proof that the account lacks media entitlement. Operators can quarantine a known-bad account with `extra.grok_media_eligible=false` or force-enable a verified account with `true`. Imports run the billing-first quota probe proactively. Chat requests and video status lookups are not affected by this media-only quarantine. If no eligible account remains, the media endpoint returns HTTP `503` with error type `grok_media_no_eligible_account`.
Administrators can override automatic media eligibility through the account create/update API by setting `extra.grok_media_eligible` to `false` (exclude) or `true` (force eligible). On update, set it to `null` to remove the override and return to automatic probe-based behavior; omitting the field preserves the current override. A weekly allowance period alone is not treated as a paid tier signal. Successful image responses must contain at least one actual image output; empty HTTP `200` responses trigger account failover instead of being counted and returned as successful generations.
<td>Real GPT-5.6 series at 3% of OpenAI pricing — <a href="https://codex-everywhere.com">CodexEverywhere</a> is democratizing access to frontier models for developers worldwide. We believe in transparency and honesty, with model quality verified by active community oversight for months. USD and crypto friendly. Start with a free $20 trial at <a href="https://codex-everywhere.com">codex-everywhere.com</a>.</td>
<td>感谢 BmoPlus 赞助了本项目!BmoPlus 是一家专为AI订阅重度用户打造的可靠 AI 账号代充服务商,提供稳定的 ChatGPT Plus / ChatGPT Pro(全程质保) / Claude Pro / Super Grok / Gemini Pro 的官方代充&成品账号。 通过<a href="https://shop.bmoplus.com/?utm_source=github">BmoPlus AI成品号专卖/代充</a>注册下单的用户,可享GPT 官网订阅一折 的震撼价格!</td>
<td>感谢 Proxy4Free 赞助本项目!Proxy4Free 是面向开发者和 AI 应用的数据代理服务商,提供住宅代理、静态住宅代理、ISP 代理及数据中心代理等多种代理解决方案,适用于 Web Scraping、Browser Automation、AI Agent 等场景。支持全球 IP 资源、稳定连接与灵活切换,帮助开发者提升数据采集成功率,降低 IP 封禁风险。通过<a href="https://www.proxy4free.com/?keyword=4yjqecpc">此链接注册</a>即可开始体验,轻松构建更稳定、高效的自动化工作流。
<td>🎉 感谢 FastAIToken 对本项目的赞助! <a href="http://www.fastaitoken.com/register">FastAIToken</a> 是面向开发者的 AI API 聚合平台,支持 OpenAI、Claude、Gemini 等主流大模型,充值 1:1,1 元 = 1 美元 API 额度,让开发者以更低成本、更便捷地使用全球领先的大模型服务。<br>
🚀 平台提供多种渠道自由选择:超级低价的0.02x OpenAI 福利分组(限时)、低至 0.25x OpenAI 分组、0.7x Claude 95%固定缓存、1.2x Claude Max 渠道;同时提供公开状态页,实时展示各分组的可用率、延迟及运行状态,服务透明可靠,并提供 7×24 小时真人技术支持(非机器人),快速响应开发者需求。
<td>感谢 Aimzoon 对本项目的赞助! <a href="http://aimzoon.com">Aimzoon</a> 提供稳定、高性价比的 AI API 接入服务,支持开发者将常用 AI 服务快速接入 Codex、Claude Code、Gemini CLI 等编程工具。无需复杂配置,更快接入,更稳调用,更省成本。codex倍率优惠,特价倍率等促销不断,注册即送免费体验额度,让 AI 编程真正进入日常工作流。<a href="http://aimzoon.com">点击这里</a>注册体验!
<td>感谢 Claude API 对本项目的赞助! <a href="https://console.claudeapi.com/agent/register/drTKjyn6wGLK061Z">Claude API</a> 是专注 Claude 模型的官方渠道 API 服务商,基于 Anthropic 官方 Key 与 AWS Bedrock 官方渠道,提供稳定的 Claude Code 与 Agent 应用接入体验,支持 Claude 全系列模型,保留 Tool Use、长上下文等官方能力。服务非逆向、非降智,适合 Claude Code 深度用户、Agent 工程师与企业技术团队使用。通过<a href="https://console.claudeapi.com/agent/register/drTKjyn6wGLK061Z">[专属链接]</a>注册后联系客服,可领取免费测试额度,并支持开票和团队对接。
<td><a href="https://nagora.ai/">Nagora</a> 是专为开发者和团队打造的多模型 AI API 网关。通过一个账户和一枚 API Key,即可统一调用 26+ 款主流文本与图像模型,兼容 OpenAI、Anthropic 与 Gemini 协议,并可无缝接入 Claude Code、Codex、Gemini CLI 等开发工具。平台提供智能路由、自动故障转移、透明计费与统一账单,同时支持预算、限速、并发控制,让个人开发、团队协作和生产环境中的 AI 调用更稳定、更可控。无需改造现有应用,只需替换 Base URL 与 API Key,最快 1 分钟即可完成接入。</td>
<td>感谢 <a href="https://www.novada.com/?sub2api/">Novada</a> 赞助本项目!Novada 为构建 AI 应用与自动化工作流的开发者提供住宅代理、ISP 代理、数据中心代理与移动代理,以及 Web Unlocker 和 Scraper API。凭借全球 IP 覆盖、灵活的轮换与粘性会话以及精准的地理定位,Novada 帮助团队在 AI Agent 工作流、跨区域测试、网络调研与浏览器自动化等场景中稳定获取网络数据。立即体验 Novada,构建更稳定、更可扩展的 AI 工作流。</td>
<td>本プロジェクトにご支援いただいた BmoPlus に感謝いたします!BmoPlusは、AIサブスクリプションのヘビーユーザー向けに特化した信頼性の高いAIアカウントサービスプロバイダーであり、安定した ChatGPT Plus / ChatGPT Pro (完全保証) / Claude Pro / Super Grok / Gemini Pro の公式代行チャージおよび即納アカウントを提供しています。こちらの<a href="https://shop.bmoplus.com/?utm_source=github">BmoPlus AIアカウント専門店/代行チャージ</a>経由でご登録・ご注文いただいたユーザー様は、GPTを 公式サイト価格の約1割(90% OFF) という驚異的な価格でご利用いただけます!</td>
<td>Proxy4Free のご支援に感謝します!Proxy4Free は開発者と AI アプリケーション向けのデータプロキシサービスプロバイダーで、住宅プロキシ、静的住宅プロキシ、ISP プロキシ、データセンタープロキシなど多様なプロキシソリューションを提供しており、Web Scraping、Browser Automation、AI Agent などのシナリオに適しています。グローバル IP リソース、安定した接続、柔軟な切り替えをサポートし、開発者のデータ収集成功率の向上と IP ブロックリスクの低減を支援します。<a href="https://www.proxy4free.com/?keyword=4yjqecpc">こちらのリンクから登録</a>して、より安定した効率的な自動化ワークフローを簡単に構築しましょう。
<td>Novada のご支援に感謝します!<a href="https://www.novada.com/?sub2api/">Novada</a> は、AI アプリケーションや自動化ワークフローを構築する開発者向けに、レジデンシャル、ISP、データセンター、モバイルプロキシに加え、Web Unlocker と Scraper API を提供しています。グローバルな IP カバレッジ、柔軟なローテーション/スティッキーセッション、精密なジオターゲティングにより、AI エージェントワークフロー、クロスリージョンテスト、Web リサーチ、ブラウザ自動化などのシーンで、チームが Web データへ確実にアクセスできるよう支援します。Novada で、より安定しスケーラブルな AI ワークフローを構築しましょう。</td>
// LongContextPricingEnabled applies equality check predicate on the "long_context_pricing_enabled" field. It's identical to LongContextPricingEnabledEQ.
// MaxReasoningEffortOverLimit applies equality check predicate on the "max_reasoning_effort_over_limit" field. It's identical to MaxReasoningEffortOverLimitEQ.
return&ValidationError{Name:"max_reasoning_effort_over_limit",err:fmt.Errorf(`ent: validator failed for field "Group.max_reasoning_effort_over_limit": %w`,err)}
return&ValidationError{Name:"max_reasoning_effort_over_limit",err:fmt.Errorf(`ent: validator failed for field "Group.max_reasoning_effort_over_limit": %w`,err)}
return&ValidationError{Name:"max_reasoning_effort_over_limit",err:fmt.Errorf(`ent: validator failed for field "Group.max_reasoning_effort_over_limit": %w`,err)}
// group.MaxReasoningEffortOverLimitValidator is a validator for the "max_reasoning_effort_over_limit" field. It is called by the builders before save.
Some files were not shown because too many files have changed in this diff
Show More
Reference in New Issue
Block a user
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.