修复 OpenAI TLS 指纹模板 User-Agent 回退#29
Open
lyen1688 wants to merge 742 commits into
Open
Conversation
…l, ban exemptions - Attention threshold: a non-hit highest-score >= the configured percent is stored as a non-hit summary only (no hit/ban/block side effects). Configurable via record_attention_inputs and attention_threshold. - API key accounts: keys can be paired with an account_email so the admin UI surfaces which account each key belongs to. - Auto-ban exemptions: auto_ban_exempt_user_ids and auto_ban_exempt_user_emails skip the disable step but still block and email on hit. - Frontend wires the new fields through the riskControl API, RiskControlView UI, and matching spec coverage.
vue-i18n parses a bare "@" as the start of a linked-message reference
("@:foo" / "@.lower:foo"), so example email addresses like
"email@example.com" embedded in placeholder strings raised
"Invalid linked format" and prevented the locale from initializing.
Wrap every bare "@" with {'@'} (matching the existing escape style used
elsewhere in the same files) so the literal stays in the string.
…revious_response_id and orphan tool outputs - Stop retrying response.failed events whose error indicates context window overflow; let the failed frame pass through so clients see the real cause. - Align HTTP /v1/responses invalid_encrypted_content recovery with the WS path: also drop previous_response_id when there is no function_call_output in flight. - Add a one-shot retry that drops previous_response_id when upstream returns "Unsupported parameter: previous_response_id". - In codex compat fallback for "no tool call found ..." errors, prune orphan function_call_output items whose call_id has no matching function_call / item_reference, and drop the now-stale previous_response_id. - Append a synthetic data: [DONE] terminator after streaming errors on the chat-completions path so OpenAI SDKs close the stream cleanly instead of waiting for read timeout. /v1/responses keeps response.failed.
…election failure messages - Add account_response_rewrite_rule service with credentials-aware persistence and a matcher keyed on raw upstream status code + body keywords. - Wire the matcher into OpenAI failover-exhausted handling so per-account rules override the default error mapping; passthrough rules still apply afterwards. - Replace the generic "Service temporarily unavailable" with a context-aware buildOpenAISelectionFailureMessage that surfaces "no account supports model X" versus other selection failures across /v1/responses, /v1/chat/completions, /v1/embeddings and /v1/images. - Plumb the active account through handleFailoverExhausted so the rewrite matcher can consult the account configuration. - Add response rewrite rules UI to Create/Edit Account modals with i18n.
- Add account_scheduling_threshold_eval/_reason services that pause an account when its native usage window crosses an admin-configured percentage; reuse the temp_unschedulable_until column to keep history with structured reason payloads. - Plumb the threshold filter through GatewayService and OpenAIGatewayService selection paths (snapshot list, recheck, getSchedulableAccount) so paused accounts are excluded uniformly across schedulers. - Persist account_scheduling_thresholds as a system setting (per platform, 100 = disabled), expose admin endpoints, and add a Settings UI section with per-platform inputs and i18n strings (EN/ZH). - Store a structured fallback reason payload when SetTempUnschedulable receives an empty reason so downstream UIs always have something parseable.
…uler snapshot
- After fetching Kiro quota for an account, write {kiro_sched_utilization,
kiro_sched_usage_updated_at, kiro_sched_reset_at} into account.extra so the
per-platform scheduling threshold filter has a fresh signal to act on.
- After fetching Antigravity model quotas, choose the worst (highest
utilization, then latest reset) scope and persist
{antigravity_sched_utilization, antigravity_sched_usage_updated_at,
antigravity_sched_reset_at} the same way.
- Skip persistence when the fetch errored, was forbidden or needs reauth so
stale data does not get promoted into the snapshot.
…s scheduling When an OpenAI account hits a 429 with "Rate limit reached for gpt-image-* ... on input-images per ..." the limit is image-route specific (codex + web2api share the per-account input-image quota). Previously we still set RateLimited / ran the OAuth runtime block on the whole account, so the same account could not serve chat/responses traffic until the cooldown elapsed. Now the image-rate-limit message is detected and only marks both image route prefixes' reset_at; the global rate-limited timestamp and runtime block path are skipped so text routes keep flowing.
- MediaObjectStore: add PresignGetObject so private downloads are served via short-lived signed URLs instead of streaming through the gateway. - MediaRepository: add GetByObjectKey for resolving an asset from its bucket+key, used to materialize public URLs on read paths. - MediaService.Create*DownloadURL now passes the asset to the URL builder so it can presign per-asset; PublicURL falls back to GetByID lookup before composing the public base URL. - Update mocks/stubs across handler and service tests so the new repo and store methods compile.
…r inputs - Insert the existing OpenAI overload special branch into the error-handling layering order so it is preserved alongside the new account-level rewrites. - Carve out the explicit local selection-failure cases (compact_not_supported, No available compatible accounts, etc.) from the generic 'no available accounts' message rewrite so already-precise messages are not regressed. - Clarify that account-level rewrite rules match against the raw upstream status code and body, not the gateway-mapped client-facing status.
…rors Kiro upstream transport failures (DoWithTLS returning err — typically AWS Q RST manifesting as `context canceled`) used to short-circuit the failover loop: handler returned 502 immediately without switching accounts and the account was never marked. Subsequent requests kept landing on the same dead upstream. Now the transport-err branch: - Distinguishes real client disconnect (gin request context already canceled) from upstream-induced cancellation; the former still passes through silently. - Marks the offending account temp-unschedulable for 2 min (with cache sync + scheduling-blocker notification) so the scheduler skips it during recovery. - Returns *UpstreamFailoverError so the gateway main loop fails over to the next compatible account. Removes the in-service c.JSON(502) — the handler main loop owns the final response so failover-exhausted paths can format it consistently. Covers both the streaming entry and the free-thinking double-request path.
…happened Chat-completions / embeddings / images handlers used to fall back to a generic 502 "Upstream request failed" when the failover loop exited without ever talking to upstream (e.g. all accounts filtered out before the first dispatch). 502 misleads clients into thinking upstream actually failed; the real cause is local routing capacity. Switch the no-failover branch to 503 + a routing-capacity-limited ops marker so dashboards and clients can tell the two situations apart. The post-failover branch keeps the upstream-derived status via handleFailoverExhausted.
…essage guard - dropOrphanFunctionCallOutputs only filtered "function_call_output" items, but Codex tool-call outputs come in several variants (tracked by isCodexToolCallOutputItemType). Use the helper so all of them participate in the orphan-pairing pass instead of only the legacy type slipping through. - isOpenAIImageGenerationRateLimitMessage matched bare "gpt-image" in any message, which over-triggers when "gpt-image*" appears as a model name in non-image account-level 429s. Require an accompanying quota keyword (limit / quota / per ) so only true image-quota messages activate the image-route bypass.
Add archive file support to the account import dialog, allowing users to upload .zip/.cpa files directly. The frontend auto-detects file type and routes to the appropriate backend endpoint. Also fix codex session dedup logic to use email+account_id composite key, preventing false duplicates for team accounts sharing the same account_id.
…fallback name OpenAI org selection now checks the access_token `poid` claim first to pick the correct team workspace, falling back to default-org and then role/title heuristics. Kiro OAuth account naming extracts the profile ID from user_id when name/email are absent.
… selection exhausted message Add shouldSuppressForwardErrorResponse to skip writing fallback error bodies when the client has already disconnected (context.Canceled). Split buildOpenAISelectionExhaustedMessage from the failure variant so that local-exclusion retries return the generic fallback instead of leaking model-specific details.
Strengthen idempotency key generation for Airwallex, tighten Stripe refund error classification, and add wxpay duplicate-refund guard.
Add upstream source metadata to ops error logger, extend stream error event classification, and surface forward-error detail in diagnostics.
Broaden content moderation exemption logic and add tests covering boundary conditions for group-based bypass paths.
Refine account selection logic, improve channel availability checks, and add comprehensive scheduler rotation tests.
Simplify request conversion, fix response field mapping for tool-use content blocks, and add cc_to_anthropic_response tests.
Introduce messages-to-chat-completions fallback path, extend endpoint routing logic, and add raw/stream test coverage for edge cases.
Extend delta-shadow continuation logic with strict ordering, add forwarder success-path tests, and fix pool/state-store edge cases.
Adjust config defaults, add http_upstream dial tests, harden deploy script restart logic, and extend account modals with new form fields.
…idge - Map web_fetch_* tool types in anthropic-to-responses conversion - Convert unsupported Responses tool types to function fallback instead of error - Deduplicate native server tools (web_search, web_fetch) in CC bridge - Add default parameters/descriptions for known server tools
…covery for function_call_output - Raise default session_idle_ttl_seconds from 30 to 600 for better delta-chain hit rate - Unify neutral idle TTL with session idle TTL (remove hardcoded 60s constant) - Add configurable outbound_payload_http_fallback_threshold_bytes (0 inherits read limit) - Enhance previous_response_not_found recovery to replay full payload with function_call_output context - Make shouldPreflightFallbackOpenAIWSPayloadToHTTP a method on service for config access
…am accounts - Broaden TextEndpointAutoRouteEnabled to support both OpenAI and Anthropic platforms - Add raw CC/Responses forwarder for openai_compat_cc_upstream accounts - Allow CC-upstream accounts with auto-route in OpenAI compat ingress filter - Guard ShouldForwardOpenAITextMessagesViaChatCompletions to OpenAI-only path - Generalize frontend toggle visibility via supportsTextEndpointAutoRoute helper
- Implement GeminiMessagesCompatService.ForwardAsResponses for Responses-shaped clients - Surface cache_read_input_tokens and cache_creation_input_tokens in Claude-format response
…test Fixes test setup to supply schedulable account via repo mock, matching the runtime path that now checks account-level model mappings.
|
Important Review skippedAuto reviews are disabled on base/target branches other than the default branch. Please check the settings in the CodeRabbit UI or the ⚙️ Run configurationConfiguration used: defaults Review profile: CHILL Plan: Pro Run ID: You can disable this status message by setting the Use the checkbox below for a quick retry:
✨ Finishing Touches🧪 Generate unit tests (beta)
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
|
All contributors have signed the CLA. ✅ |
Author
|
I have read the CLA Document and I hereby sign the CLA |
Author
|
recheck |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
变更内容
user_agent字段，支持创建、更新、导入采集样本时保存对应 User-Agent。验证
go test ./...pnpm vitest run src/utils/__tests__/tlsFingerprintYaml.spec.tspnpm run buildgit diff --check