Skip to content

Exempt the TensorGrid internal API from the global IP rate limit - #7

Merged
mrdjango merged 45 commits into
mainfrom
merge/upstream-20260908
Sep 9, 2026
Merged

mrdjango merged 45 commits into
mainfrom
merge/upstream-20260908

Conversation

@mrdjango

@mrdjango mrdjango commented Sep 9, 2026

Copy link
Copy Markdown
Owner

Why now

Production served ~48% 5xx for a stretch on 2026-09-09 because the gateway was returning 429 to every internal call from the TensorGrid backend. GlobalAPIRateLimit is installed on the whole /api group and keys on c.ClientIP(), so all server-to-server traffic — which arrives from a single container address — shared one 360-request / 180-second bucket. A dashboard load emptied it, and the 429s reached end users as 500/502/503.

The commit that fixes this is 5f0460402. It is currently only on this branch, so the :main image production runs does not contain it.

What changes

  • /api/internal/tensorgrid/v1 is exempt from the global limiter. Every request there is HMAC-signed against a shared secret with a bounded timestamp window (TensorGridServiceAuth), so it is a trusted first-party caller rather than anonymous traffic. The check is by path because the limiter runs before the route-level signature check that would otherwise mark the caller trusted.
  • Rate-limited responses now carry the standard {success, code, message} envelope. They were aborted with an empty body, so API clients could not tell backpressure apart from a malformed reply — the TensorGrid backend reported it as gateway_invalid_response and retried instead of backing off.

Two tests cover it: internal traffic passes well past the configured limit while the same client IP is still throttled on other /api routes and its budget is untouched, and a throttled response carries both the envelope and Retry-After.

Scope note

This branch is the merge/upstream-20260908 upstream sync, so merging it brings 45 commits, not just the rate-limit fix. If you would rather ship the fix on its own, say so and I will cut a branch off main with just 5f0460402.

Matching change

The TensorGrid side (which cuts the call volume that exhausted the bucket) is a separate PR in mrdjango/TensorGrid.

🤖 Generated with Claude Code

Calcium-Ion and others added 30 commits September 4, 2026 10:07
Model-name post-processing is rebuilt around an explicit trailing
@key:value modifier syntax (thinking/effort/temperature/topp) that
overrides request fields, survives model mapping, and records
conversion diagnostics on the consume log.

- Legacy naked aliases (-thinking, -nothinking, -thinking-<budget>,
  effort tails) now parse only for positively matched families
  (gpt-*/o-series, claude-*, gemini-*, incl. vendor/ namespaces);
  names like qwen-max stay opaque. EffortTailModelIDs remains the
  escape hatch for real in-family IDs such as gpt-5.1-codex-max.
- Billing identity resolves once in ModelPriceHelper via a ladder:
  configured request name first (legacy wildcard entries intact), then
  canonical billing names rebuilt from parsed intent
  (base@effort:E@thinking:S, then base@thinking:S; order, duplicates,
  and budget values are irrelevant; temperature/topp never priced),
  then base. Routing and token limits fall back through
  RoutingMatchModelName; pricing lookups stay wildcard-only.
- Pass-through stays byte-identical: modifiers and aliases are neither
  parsed nor validated there and forward verbatim for the upstream
  (or a chained gateway) to interpret.
- Unknown modifier keys and invalid known-key values are rejected with
  400; models whose real names contain @tag:value are exempted via the
  thinking-suffix blacklist, which now supports re:-prefixed Go regex
  entries.
- Claude reasoning render coerces unsupported combinations (disable,
  adaptive, budgets) with warning diagnostics instead of erroring;
  native-protocol requests without host syntax pass through untouched.

BREAKING(openrouter): drop the host-invented "-thinking" model-name
alias (added in 4f6d16e) that trimmed any *-thinking model on
OpenRouter channels and injected reasoning.enabled. It matched too
broadly and mangled real model IDs such as kimi-k2-thinking.
Migration: use some-model@thinking:on, or keep the old public name via
a channel model mapping {"some-model-thinking": "some-model@thinking:on"}.
Claude/Gemini family aliases (incl. anthropic/claude-*-thinking) keep
working via the family whitelist.
* perf(relay): bulk-copy Responses raw JSON fields

* perf(common): share RawMessage deep-copy optimization

---------

Co-authored-by: CaIon <i@caion.me>
…s#7211)

* fix(relay): treat gpt-5 and later generations alike for max_completion_tokens

IsOpenAIGPT5Model matched on the literal prefix "gpt-5", so gpt-6-astra
(and every generation after it) fell through the gpt-5 request rules:
max_tokens was forwarded as-is and the provider rejected it with
"Unsupported parameter: 'max_tokens' is not supported with this model.
Use 'max_completion_tokens' instead." The same gap left temperature,
top_p and logprobs untouched, each of which the provider also rejects,
and made the channel test button report a 400 for a healthy deployment.

Match on the major version instead (gpt-<n>... with n >= 5). Callers are
unchanged: ConvertOpenAIRequest, GetSystemRoleName, buildTestRequest and
the health check all go through this one helper. buildTestRequest now
sends max_completion_tokens for these models directly instead of relying
on the later conversion. gpt-4.1, gpt-4o, gpt-oss, gpt-image and
gpt-realtime names still do not match.

Verified against Azure OpenAI gpt-6-astra (2026-09-03): with the old
prefix max_tokens / temperature / top_p / logprobs each returned 400,
while gpt-5.6-luna with the same payload returned 200.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016bwz8o5UeoRtrtDusKaayp

* fix(relay): separate OpenAI chat model compatibility rules

---------

Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>
Co-authored-by: CaIon <i@caion.me>
Move account security settings into a dedicated page and add token status, rotation, revocation, and access history.

Store audit events with role snapshots and JSON metadata, add audit.read authorization and an independent audit page, and upgrade the ClickHouse driver to v2.46.0.
Upgrade the SQLite GORM driver to correctly distinguish unique indexes from constraints. Normalize equivalent MySQL decimal defaults and PostgreSQL CHAR metadata during migration comparison while preserving real schema changes.

Validation: 27 fresh-start and rc.26 upgrade scenarios using SQLite 3.50.4, MySQL 5.7.44, PostgreSQL 9.6.24, and ClickHouse 25.8.33.6; 15 upgraded databases passed uniqueness checks. Relational databases issue no DDL on unchanged restarts. Existing ClickHouse TTL synchronization remains unchanged.
Require single-use operation proofs for passkey enrollment, two-factor setup, and channel key access. Add password and OAuth verification flows, enforce session-bound enrollment, and redact OAuth callback secrets from logs.

Validation: affected Go packages pass; frontend typecheck, changed-file lint, and 111 tests pass. Security enrollment regressions pass on SQLite 3.50.4, MySQL 8.0.46, and PostgreSQL 16.15. Full frontend lint has pre-existing errors outside the changed files.
Use authorization code flow with PKCE and verified ID tokens for Telegram login, binding, and security verification. Preserve existing bindings and require administrator OAuth configuration.

Keep the restricted WeChat first-enrollment session proof, fix missing-target authentication errors, and preserve callback requests after OAuth popups close.
Require scoped, single-use verification for account bindings and password
operations. Bind OAuth authorization and email confirmations to the initiating
session; preserve the last usable login method and audit operation outcomes.

Apply Unicode-aware password length limits, Argon2id writes with bcrypt
compatibility, and long-password encryption.
Return has_password with the existing profile SELECT without extra queries.
Reuse the existing security dialogs and add all seven locale translations.

Validation:
- Go: go test ./common ./model ./service ./middleware ./controller ./router -count=1
- DB: SQLite 3.50.4, MySQL 8.4.11, PostgreSQL 16.15; separate main/log databases
- MySQL/PostgreSQL: TEST_SECURITY_DIALECT=<dialect> with TEST_<DIALECT>_DSN,
  go test ./controller -run '^(TestSecurityAccount|TestSecurityEnrollment|TestGenerateOAuthCode|TestOAuthBind|TestTelegramOAuth)' -count=1 -v
- Web: relevant Vitest suites, bun run typecheck, targeted oxlint/format,
  bun run i18n:sync, and bun run build

Roll out dual-format readers to every instance with
ACCOUNT_PASSWORD_HASH_ALGORITHM=bcrypt before enabling Argon2id writes
and the new UI. Rollbacks must retain Argon2id and v2 envelope readers.

Relevant controls: ASVS 5.0.0 6.2.1-6.2.3, 6.2.5-6.2.9, 6.3.7, 7.4.3, 7.5.1;
this change does not assert application-wide ASVS certification.
Record successful and failed API token operations with safe target metadata. Capture quota adjustments in a transaction, synchronize committed cache differences, and correlate audit and top-up records.

Show operation targets, changes, quota balances, and failure details consistently across audit and usage logs, with translations for all seven locales.

Validated controller, middleware, and model tests; 78 frontend tests; typecheck and lint; real SQLite 3.50.4, MySQL 8.4.11, and PostgreSQL 16.15 with shared and separate log databases.
Treat TOTP and Passkey as alternative enrolled factors across login and
sensitive account operations. Gate every primary login transport before
issuing a session, require WebAuthn user verification, and consume login
challenges atomically with session creation.

Reuse the shared verification UI for login, 2FA management, and account
deletion. Require scoped, single-use deletion proof; recheck the session
inside the deletion transaction and revoke all sessions afterward.

Validation: controller/service/model/middleware tests; real SQLite 3.50.4,
MySQL 8.4.11, and PostgreSQL 16.15 security regressions; frontend tests,
TypeScript, targeted lint, formatting, and production build.

Deploy the frontend and all backend nodes together. No schema changes.
Introduce a unified model management experience: catalog metadata
validation, vendor management, batch delete with channel/pricing
cleanup, model pricing snapshot editing with optimistic concurrency,
and an upstream ratio-sync flow with price cells. Move configuration
into dedicated pricing config/metadata-sync/vendor-management backend
services and add audit records for model/vendor/pricing mutations.

Rework the models page around vendors and model connections, add
model-pricing and vendor-management dialogs, and replace the shared
Select usages with the Combobox component across subscriptions,
plugins, OAuth presets, audit filters, and settings. Add the model
pricing panel and verify behavior with focused tests.
* feat(ali): support wan3.0 all-in-one video models

Extend the Alibaba task plugin for wan3.0-video / wan3.0-video-prime
using DashScope media, resolution/ratio, duration rules, and billing ratios.

* fix(ali): make wan3.0 smart duration billable and alias-safe

Review fixes on top of QuantumNous#7240 (qiuliw):

- duration -1 never reached the plugin: the host rejects negative
  canonical duration/seconds facts before any hook runs, on every
  entry point. Decoders now normalize -1 into an auto_duration marker;
  convert emits -1 upstream and bills 30s up front; non-wan3.0 models
  reject the marker instead of silently defaulting to 5s.
- extractUsageOnComplete read output.duration/output.resolution, but
  wan3.0 reports usage.output_video_duration and numeric usage.SR, so
  smart-duration and resolution settlement never reconciled. Read the
  usage block first, keep the legacy output fields as fallback.
- convert keyed default resolution on the client model name, so a
  channel-mapped alias fell to 720P while the direct request got 1080P.
  Every model-shaped decision in convert now uses ctx.upstreamModel.
- Unknown wan3.0 size values were silently coerced to 1080P; reject them.
- Image-only openai_responses input was accepted for every model,
  regressing t2v models into pre-consume then upstream rejection.
  Restore the guard and allow image-only for i2v and wan3.0 only.
- Native passthrough dropped wan3.0 parameters (ratio, audio) unless
  media was present; forward them via metadata for wan3.0.
- Hoist the triplicated size-to-resolution map; add wan3.0 contract tests.

---------

Co-authored-by: qiuliw <a1807191473@qgmail.com>
Kimi K3 injects tools mid-conversation via a system message that carries
a `tools` array. `dto.Message` had no such field, so the tools were
silently dropped during the parse/re-marshal round trip and the upstream
rejected the request with `'tool_choice'='required' requires a 'tools'
field`.

- add `Message.Tools` (json.RawMessage passthrough)
- omit the `content` key only for tool-loading messages with nil content,
  as Kimi rejects `tools` next to `content`; all other messages keep
  emitting `"content": null`
- count message-level tools in token estimation
- skip tool-loading messages in channel system prompt injection and make
  the compatible handler reuse applySystemPromptIfNeeded
- add kimi-k3 to the moonshot model list

Fixes QuantumNous#7235
`common/json.go` and `relaykit/relayconvert/kitutil/json.go` were two
hard-wired copies of the same encoding/json wrapper, so swapping the JSON
engine required editing both modules.

- kitutil defines a `Codec` interface with a standard-library default and
  a `SetCodec` hook, mirroring the existing SetLogging host hook; every
  kitutil JSON helper and relaykit DTO (un)marshal method goes through it
- `common/json.go` forwards to kitutil and injects `hostJSONCodec` from
  init() so tests run on the same engine as production; swapping the
  engine now touches only this type in the root module
- route the remaining direct encoding/json calls inside relaykit
  (dto/values.go, responses stream validation) through kitutil
- add a codec routing test and a host codec conformance test locking the
  encoding semantics the DTOs depend on

Direct encoding/json call sites in the root module are left for a
separate cleanup.
- Portal Combobox and Select popups into the vaul DrawerContent via a
  portal-container context so they stay inside the Radix modal layer
  instead of inheriting body pointer-events: none
- Provide an in-memory localStorage/sessionStorage in test-setup when the
  Node 25+ global accessor resolves to undefined and shadows jsdom
- Add regression tests for popups rendered inside and outside the drawer
The perf summary API only returned the last three non-empty buckets as
bare success rates, so the 24-slot status strip on model cards never lit
more than three bars and could not show hours without traffic.

Replace recent_success_rates with recent_success_series: one timestamped
point per hour that had requests, aligned to hour start regardless of
the configured bucket size. The badge now anchors its 24 slots to the
client's current hour and places each point by timestamp, leaving hours
without data gray.
… plugin

Disabling a task plugin that has both a factory built-in and an override
row only flipped the override flag, so the built-in kept routing the same
models and same-name uploads (e.g. minimax-h3 vs MiniMax-H3) still hit a
routing conflict. Now the key also enters the disabled-factory set and the
list reports "disabled" instead of "disabled_fallback" when nothing serves.
…t-ins

Decode on ctx.upstreamModel || ctx.model and echo ctx.model; fix the
lyrics/music render branch that read a nonexistent ctx.requestBody.model;
stop sending empty Accept/Content-Type. Bump sunoapi to 1.0.2. Add an
alias-echo table test covering every built-in.
TaskPluginOverrideEnabled had no UI since the master switch landed, yet
when left off it marked every third-party plugin "disabled; platform
unavailable" regardless of its own toggle. Drop the option, env var,
registry flag, and the dead branch in ListTaskPlugins; the master switch
and per-plugin toggles are the only two levels now.
The model card grid and its loading skeleton only reached three columns at
2xl, leaving a two-column layout on common desktop widths. Use the xl
breakpoint for both and lock it with a test.
Calcium-Ion and others added 15 commits September 8, 2026 15:15
Include configured channel models without creating metadata, derive square
visibility from live routes and metadata policy, and filter before pagination.

Share pricing display with the model square, show expression tiers and task
unit prices, preserve zero rates, and expose full pricing and visibility
reasons from compact responsive rows. Complete all seven locale translations.

Validated frontend tests, typecheck, lint and production build, plus the
model database matrix on SQLite 3.50.4, MySQL 5.7.44 and PostgreSQL 9.6.24.
Show remaining and used API key quota with a progress bar, and use consistent mobile cards, group multiplier badges, and activity timestamps. Present available user balance with used quota underneath and translate the new labels.

Resolve full API keys only for explicit copy or chat actions. Reviewed OWASP Authentication and Session Management guidance and ASVS 5.0.0 V14.2.6 and V8.3.1; backend authorization is unchanged, and regression tests cover refused and denied key resolution. This frontend change does not assert application-wide ASVS compliance.

Validation: 55 related component tests passed; the latest mobile group and quota changes passed 35 focused tests. TypeScript, scoped lint, formatting, production build, and git diff checks passed. Responsive previews verified narrow screens and finite, unlimited, exhausted, and inactive quota states.
Reuse the shared Combobox to suggest groups for the active log view while allowing historical group names to be entered manually. Exclude the automatic routing pseudo-group from suggestions and preserve masking, reset, URL navigation, and mobile drawer behavior.

Extend the existing Combobox with keyboard event forwarding and accessible labels, preserve the selected custom value on focus, and close suggestions on blur. Enter confirms a selection before submitting the filter.

Validation: 25 tests passed across the shared Combobox and log group, type, and mobile filter suites. TypeScript, scoped lint, formatting, and git diff checks passed.
Place remaining quota on the left and used quota on the right above the progress bar, without visible labels in desktop rows. Preserve accessible descriptions and the existing mobile labels and stacked layout.

Validation: updated regressions failed before the change and all 26 API key listing tests passed afterward. TypeScript, scoped lint, formatting, production build, and git diff checks passed.
Increase the quota column width and minimum width from 220px to 260px to give the side-by-side amounts more room.

Validation: all 26 API key listing tests, TypeScript, scoped lint, formatting, and git diff checks passed.
Bound desktop quota content to 180px so column width provides whitespace before the group column. Remove the quota-only padding override to use the same cell padding as other columns. Preserve the mobile quota layout.

Validation: layout regressions failed before the fix and all 26 API key listing tests passed afterward. TypeScript, scoped lint, formatting, production build, and git diff checks passed.
Add confirmed multi-select deletion through one batch API request and one
soft-delete statement. Record the affected count and requested IDs in a
separate batch audit event, and identify legacy events with missing counts.

After creation, offer an unchecked Save as a file option with TXT/Markdown
formats and optional name/quota fields. Keep Done as the default completion
action. Include translations for all seven frontend locales.

Validation:
- Redemption and audit frontend regression tests, typecheck, and scoped lint.
- go build ./...
- go test ./controller -run '^TestDeleteRedemptionBatch$' -count=1 -v
  with TEST_MYSQL_DSN, TEST_MYSQL_LOG_DSN, TEST_POSTGRES_DSN,
  and TEST_POSTGRES_LOG_DSN set to isolated primary and log databases.
- Real SQLite 3.50.4, MySQL 8.4.11, and PostgreSQL 16.15 passed, including
  deletion of 15 records, duplicate/missing IDs, zero-row retries, audit
  deduplication, invalid input, and exclusion of credentials from logs.
Brings in 42 upstream commits (32c2619..ea7cb0b) and reconciles them
with the TensorGrid fork logic.

Conflict resolved:

- model/user.go: took upstream's new User fields (HasPassword,
  AccessTokenCreatedAt, password max=128) while keeping our bigint
  column types on Quota/UsedQuota/AffQuota/AffHistoryQuota. Upstream's
  new migration dialector only normalizes decimal defaults (MySQL) and
  bpchar/char (PostgreSQL), so it does not fight the bigint declaration.

TensorGrid integration verified intact after the merge:

- Credit-event enqueue in RecordConsumeLog still runs regardless of
  LogConsumeEnabled, with the durable request id allocated first.
- consumeParams threading survives on every SettleBilling /
  PostConsumeQuota / postConsumeQuotaWithResult call site.
- Wallet reserve/refund/adjust hooks, IsTensorGridUser gating, catalog
  and outbox migrations (including the identifier-width prologue in both
  migrateDB and migrateDBFast) are unchanged.
- Audio-transcription endpoint type and its frontend pricing/model
  constants survived upstream's model and pricing rework.

Behavioral note: upstream's canonical billing identity
(GetBillingModelName) now feeds the consume log, so TensorGrid usage
events record the canonical pricing name rather than the raw request
name. This is metadata only; it is not used as a lookup key.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
…ettles

"does not show a completed setup entry when the key lookup fails" failed
on both CI runs of this PR, and intermittently in local full-suite runs,
while passing whenever the file ran alone.

The assertion held the node returned by findByRole across an async
boundary. The guide subtree is remounted while the remaining overview
queries settle, which detaches the captured element, so toBeVisible saw
a detached node — the reported element had aria-expanded="true" but no
children, which is what an unmounted subtree leaves behind.

Re-query inside waitFor so a transient remount cannot fail the
assertion. A guide that never expands still fails, so the test keeps its
meaning, and this follows web/AGENTS.md: async tests wait for an
explicit UI state rather than a captured node.

Verified: the file passes alone and the full suite passes 995/995.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
@mrdjango
mrdjango merged commit 0e4cbab into main Sep 9, 2026
2 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

4 participants