feat: 新增 VolcAdapter (58) 渠道——火山方舟原生 API 兼容代理 - #4702
Closed
JAYotta wants to merge 35 commits into
Closed
Conversation
Add Volc task list route adaptation and explicit unsupported DELETE behavior, preserve Ark video payload fields, and update official seedream/seedance model IDs with focused test coverage. Made-with: Cursor
…dpoint types Task 1 – Move image route to relay-router.go: - Register /volc/api/v3 group in relay-router.go with only POST /images/generations (synchronous relay via controller.Relay). - Remove POST /images/generations from video-router.go volcV3Router (async task group). - Drop now-unused `types` import from video-router.go. Task 2 – Trim volcengine/constants.go: - Remove doubao-seedance-* and bare seedance-* video model entries from relay/channel/volcengine/constants.go (wrong channel — they belong on doubao-video). - Keep all doubao-seedream-* and bare seedream-* image entries plus LLM entries. - Move seedance-* bare aliases to relay/channel/task/doubao/constants.go and add all official date-suffixed IDs from the 2026-04-27 Ark API scan. Task 3 – Rewrite middleware/volc_adapter_test.go: - Extract runVolcMiddlewareCase helper to eliminate boilerplate. - Replace veo-1 with realistic doubao-seedance-2-0-260128 model. - Add TestVolcConvert_ImageGeneration_T2I, _I2I (doubao-seedream models, image array). - Add TestVolcConvert_VideoSubmit_T2V (action set), _I2V (action absent when image present). - Improve VideoFetchByID, VideoList, VideoDelete tests with full assertions. - Add TestVolcConvert_RequestKeyFallback: table-driven model/prompt fallback chain (model → model_name → req_key, prompt → content). - Add TestVolcConvert_InvalidBody: invalid JSON and empty body → 400. - Assert KeyRequestBody context key via deep-equal on all submit/image cases. - Assert metadata deep-equals the entire original request body (not spot-checked). Task 4 – Add volc endpoint types for pricing/marketplace UI: - Add EndpointTypeVolcImage = "volc-image" and EndpointTypeVolcVideo = "volc-video" to constant/endpoint_type.go. - Register default paths in common/endpoint_defaults.go: volc-image → POST /volc/api/v3/images/generations volc-video → POST /volc/api/v3/contents/generations/tasks - Add "seedream" to common/model.go ImageGenerationModels so IsImageGenerationModel matches doubao-seedream-* and seedream-* names. - Update GetEndpointTypesByChannelType in common/endpoint_type.go: ChannelTypeDoubaoVideo → [volc-video, openai-video] ChannelTypeVolcEngine + image model → [volc-image, image-generation, openai] ChannelTypeVolcEngine + LLM/other → fall through to default (openai) - Pricing API wiring verified: supportedEndpointMap is auto-populated via GetDefaultEndpointInfo; no controller changes needed. - i18n: PricingEndpointTypes.jsx uses raw endpointType string as label (getEndpointTypeLabel returns endpointType verbatim); no translation keys exist for any endpoint type in the JSON locale files, so "volc-image"/"volc-video" will display as raw strings — consistent with all other endpoint types. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Volc's official SDKs default to base URL https://ark.cn-beijing.volces.com and request /api/v3/...; mounting our compat gateway at the same path means users only need to change the domain. There's no collision with existing /api/* admin routes (those are all specific paths, none under /api/v3). Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Splits the volc-compat API endpoints onto a dedicated channel type so ChannelTypeVolcEngine (45) and ChannelTypeDoubaoVideo (54) revert to their pre-cleanup minimal scope (LLM/TTS and OpenAI-format video tasks respectively). The new VolcAdapter channel reuses existing volcengine and taskdoubao adaptors but ships its own default model list (seedream + seedance) and binds the volc-image/volc-video endpoint types. Display name chosen: "VolcAdapter" (matches the naming style of existing single-word entries like "VolcEngine", "DoubaoVideo", etc.) Channel type number: 58 — verified as next available after ChannelTypeCodex=57. ChannelTypeDummy becomes 58 (same value, per Go const repeat semantics). Changes by task: - Task 1.1: constant/channel.go — add ChannelTypeVolcAdapter=58, base URL, display name; common/api_type.go — map to APITypeVolcEngine - Task 1.2: relay/relay_adaptor.go — add VolcAdapter to taskdoubao case; relay/common/relay_info.go — add to streamSupportedChannels - Task 1.3: relay/channel/volcadapter/constants.go — new package with seedream + seedance model list; controller/model.go — register in openAIModels and override channelId2Models entry - Task 1.4: common/endpoint_type.go — remove VolcEngine/DoubaoVideo volc-* endpoint bindings; add ChannelTypeVolcAdapter case with per-model dispatch - Task 1.5: relay/channel/volcengine/constants.go — revert to pre-worktree state (LLM + seedream + seedance + seed-thinking); relay/channel/task/doubao/constants.go — revert (7 official + 2 legacy, no bare date aliases) - Task 1.6: relay/relay_task.go — change list filter to VolcAdapter platform; relay/relay_task_volc_list_test.go — update platform + add coexistence test - Task 1.7: middleware unchanged (no channel-type refs) - Task 1.8: web/src/constants/channel.constants.js — add type 58 to CHANNEL_OPTIONS; web/src/helpers/render.jsx — add case 58 icon (Doubao) Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Now that the volc-compat gateway lives on its own ChannelTypeVolcAdapter channel, the existing VolcEngine (45) and DoubaoVideo (54) channels no longer need to advertise the full set of seedream/seedance models. Match upstream/main's minimal model lists so this branch stays close to the upstream merge base — the new-and-shiny seedream/seedance entries are available on the dedicated VolcAdapter channel. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Adds RelayFormatVolc and routes the volc-compat image endpoint through a dedicated handler (mirroring the Gemini pattern) instead of cursor's body-rewriting middleware. Volc body fields (size 2K/4K, sequential_image_generation, optimize_prompt_options, watermark, etc.) flow through to upstream without translation. Changes: - types/relay_format.go: add RelayFormatVolc constant - dto/volc_image_request.go: new DTO that captures known Volc fields plus an Extra map for unknown fields (preserves byte-identity) - relay/channel/adapter.go: add ConvertVolcRequest to Adaptor interface - relay/channel/volcengine/adaptor.go: no-op pass-through implementation - relay/channel/*/adaptor.go (31 files): "unsupported" stub implementations - relay/helper/valid_request.go: add RelayFormatVolc case + validator - relay/volc_handler.go: new VolcImageHelper (mirrors GeminiHelper) - relay/common/relay_info.go: GenRelayInfoVolc + GenRelayInfo case - controller/relay.go: volcRelayHandler + RelayFormatVolc switch case - router/relay-router.go: image route uses RelayFormatVolc, removes VolcRequestConvert() middleware from the image route group The task path still uses the existing middleware in this commit; the next commit migrates it and deletes the middleware entirely. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Migrates the volc-compat task endpoints (submit, fetch, list, delete)
to native RelayFormatVolc pass-through, matching the image side from
the prior commit. Body bytes flow byte-identical to upstream.
Deletes middleware/volc_adapter.go and its rewritten test now that
both image and task paths bypass the middleware.
Key changes:
- controller/relay.go: add RelayTaskVolcSubmit (GenRelayInfo with
RelayFormatVolc), RelayTaskFetchVolc, RelayTaskVolcDelete (501)
- relay/common/relay_info.go: GenRelayInfo handles RelayFormatVolc
with nil request as task-style relay info (with TaskRelayInfo)
- relay/channel/task/doubao/adaptor.go:
- ValidateRequestAndSetAction: branches on RelayFormatVolc to parse
Volc-native body minimally (model + content[] action detection)
- BuildRequestBody: branches on RelayFormatVolc to forward raw bytes
byte-identical; model-mapping patches only the model field
- EstimateBilling: branches on RelayFormatVolc to read video_url
from raw body content[] instead of TaskSubmitReq.Metadata
- router/video-router.go: volcV3Router uses new controllers directly,
no VolcRequestConvert() middleware; relay_mode and task_id are set
inline via route-level closures
- middleware/volc_adapter.go: DELETED
- middleware/volc_adapter_test.go: DELETED
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Adds a mechanism for shipping default tiered-billing expressions in code. Default expressions activate only when: - Admin hasn't set BillingMode in DB - Admin hasn't set BillingExpr in DB - Admin hasn't customized ModelRatio in DB Architecture: B (lazy fallback). GetBillingMode/GetBillingExpr check defaultBillingExpr at lookup time with no startup mutation. Chosen over Architecture A because the DB load (loadOptionsFromDatabase) runs *after* InitRatioSettings, so at apply-time we cannot distinguish a DB-set ratio from a code-default ratio that the admin resaved; lazy lookup avoids the ambiguity entirely and keeps the hot path a simple two-map check. Added ratio_setting.IsModelRatioCustomized(model) to detect whether an admin explicitly changed ModelRatio for a model: if the model is absent from defaultModelRatio, any value in modelRatioMap came from DB; if it is present in defaultModelRatio, a differing value signals a DB override. The default map is empty in this commit; Step 4 populates it with seedance video expressions covering 1080p resolution and video-input discount tiers. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
This reverts commit e6d31ca8c. Defaults for Seedance billing will live in the UI's existing PRESET_GROUPS template list in TieredPricingEditor.jsx instead of a Go-side fallback mechanism. Existing deployments stay on legacy ratio mode unchanged; admins who want tiered billing for these models opt in via the UI preset. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Adds six tiered_expr presets to the Tiered Pricing Editor's "请求条件" group covering doubao-seedance models with their actual RMB/1M-output-token rates: - Seedance 2.0: 46/28 base, 51/31 at 1080p (text/with-video) - Seedance 2.0 Fast: 37/22 (text/with-video) - Seedance 1.5 Pro: 16/8 (with-audio/silent) - Seedance 1.0 Pro: 15/7.5 (online/flex) - Seedance 1.0 Pro Fast: 4.2/2.1 (online/flex) - Seedance 1.0 Lite: 10/5 (online/flex) Each expression coalesces native Volc body shape (top-level fields) and OpenAI-format-wrapped shape (metadata.*) so the same preset applies whether the request hits /api/v3/* or /v1/video/generations on a VolcAdapter channel. Token estimation formula per Volc docs: (input_video_duration + output_video_duration) × output_width × output_height × output_fps / 1024 Final billing uses upstream's usage.completion_tokens. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
The previous preset format `(cond) ? tier(...) : tier(...)` failed to
produce a valid cost estimate in the UI cost-estimator (evalExprLocally).
The local JS evaluator builds a minimal env {p, c, len, tier, math-helpers}
and calls `new Function(...)` on the raw expression string. When param()
appears before tier() — i.e. as the condition of the outermost ternary —
JavaScript throws `ReferenceError: param is not defined` before tier() is
ever reached, so the estimator shows an error and reports cost=0.
Investigation findings:
- BOTH old and new formats compile and run correctly in the Go billingexpr
engine (pkg/billingexpr); smokeTestExpr passes for both forms.
- The failure is UI-only: evalExprLocally in TieredPricingEditor.jsx did
not include param(), header(), has(), or nil in its JS eval environment.
- With the multiplicative form `tier("base", c * X) * (param(...) ? ...)`,
tier() executes first (JavaScript evaluates left-to-right) and returns a
valid float. param() is still evaluated as part of the multiplier, so the
error still occurs — meaning the fix also requires adding param/nil stubs
to evalExprLocally. Both changes are included in this commit.
- Root cause: evalExprLocally is a preview-only estimator; it intentionally
cannot inspect the real request body. param() stubs returning null allow
the expression to evaluate to the base-tier price (all multipliers resolve
to their else-branch with null inputs), giving a meaningful preview.
Changes:
1. Rewrites all 6 Seedance presets as `tier("base", c * <base>) * <mult>`
- doubao-seedance-2-0: 46 base × (res+video multipliers)
- doubao-seedance-2-0-fast: 37 base × (video multiplier)
- doubao-seedance-1-5-pro: 16 base × (silent 0.5 multiplier)
- doubao-seedance-1-0-pro: 15 base × (flex 0.5 multiplier)
- doubao-seedance-1-0-pro-fast: 4.2 base × (flex 0.5 multiplier)
- doubao-seedance-1-0-lite: 10 base × (flex 0.5 multiplier)
2. Adds param(), header(), has(), nil stubs to evalExprLocally so any
expression using request probes renders a valid base-price preview.
3. Adds permanent regression test at pkg/billingexpr/seedance_presets_test.go
covering every tier × body-shape combination (31 cases total):
- 8 cases for sd2-0 (4 combos × 2 body shapes)
- 4 cases for sd2-0-fast (2 combos × 2 body shapes)
- 5 cases for sd1-5-pro (3 native + 2 wrapped, including default)
- 4 cases for sd1-0-pro, sd1-0-pro-fast, sd1-0-lite (2 × 2 each)
- 6 smoke-test cases (one per model via billing_setting.SmokeTestExpr)
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
The previous inline param() form failed in upstream new-api v0.13.1's
evalExprLocally because that version lacks param/header stubs. Switching
to structured `requestRules` (matching existing claude-opus-fast /
gpt-5.4-tiers pattern) keeps `expr` as just `tier("base", c * BASE)` so
preview evaluates cleanly while the backend still receives the full
conditional expression after combineBillingExpr.
Multiplicative composition is used instead of a 4-way matrix because
the rule system doesn't support negation. For Seedance 2.0, this
introduces ~0.13% drift on the 1080p+video tier (31.04 instead of
31.00); accepted as Volc may publish 31 as a rounded display price.
Each conditional dimension has 2 rules (top-level path + metadata.*
path) to handle both /api/v3 native and /v1/video/generations OpenAI-
format request shapes.
Also relocates the regression test from pkg/billingexpr (upstream
shared) to relay/channel/volcadapter (Graylight channel that hosts
seedance models) for cleaner upstream merge boundaries.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
1. Comment out EndpointTypeVolcImage/Video and the VolcAdapter case in
GetEndpointTypesByChannelType, matching upstream's pattern of leaving
task-style channels (Kling/Jimeng/Suno/Midjourney) without dedicated
endpoint types. Channel 58 falls through to the OpenAI default in
the marketplace UI, consistent with existing task channels.
2. Merge RelayTaskVolcSubmit / RelayTaskFetchVolc / RelayTaskVolcDelete
back into the existing RelayTask / RelayTaskFetch by reading
RelayFormat from a context key ("relay_format"). Removes ~90 lines of
duplicated retry-loop logic. Routes set the format inline before
dispatching; RelayTaskVolcDelete is replaced by a 5-line inline closure.
3. Translate VolcAdapter channel selector label to Chinese to match
the rest of channel.constants.js ('火山方舟兼容 (Seedream + Seedance)').
4. Remove the param/header/has/nil stubs added to evalExprLocally in
commit 81d37dced. They became dead code after switching to structured
requestRules — the expr field is now just `tier("base", c * X)` with
no param() calls.
5. Drop the redundant EndpointTypeImageGeneration entry from the
VolcAdapter image-model return list (now moot since the case is
commented out). Test cases updated to use string literals for
volc-image/volc-video absence assertions since the constants are
now commented out.
Plan Y (extending task billing with tiered_expr support) is still
being discussed; estimateBillingVolcNative, videoInputRatioMap, and
the seedance presets are intentionally untouched in this commit.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Plan B1 implementation: extends task billing flow to support BillingMode=tiered_expr, making the 6 seedance presets actually fire (previously dead config — see commit message for fcfbb6629 for the discovery). Investigation finding (Step 0): Option A — tiered_expr restricted to RelayFormatVolc requests. EstimatedBillingTokens==0 guard ensures OpenAI-format entry (/v1/video/generations) stays on ratio billing. metadata.* rules dropped from all 6 seedance presets accordingly. Pre-charge (relay/helper/price.go): - ModelPriceHelperPerCall branches on BillingMode==tiered_expr && EstimatedBillingTokens>0 → modelPriceHelperTieredForTask() - Runs billingexpr.RunExprWithRequest() with estimated tokens, freezes BillingSnapshot + BillingRequestInput on RelayInfo Token estimator (relay/channel/task/doubao/seedance_estimator.go): - EstimateSeedanceTokens(modelName, body []byte) int64 - Formula: (inputVideoDurSec + outputDurSec) × W × H × 24 / 1024 - Conservative over-estimate: input video capped at 15 s - Looks up model default resolution + max duration from tables - Returns 0 on unknown model → falls back to ratio billing Interface plumbing: - TaskAdaptor.EstimateBillingTokens(c, info) added to interface - BaseBilling provides zero default (no-op for non-seedance adaptors) - Doubao TaskAdaptor overrides; guards on RelayFormatVolc - relay_task.go sets info.EstimatedBillingTokens before price helper; wraps EstimateBilling OtherRatios in TieredBillingSnapshot==nil guard Settlement (service/task_billing.go): - RecalculateTaskQuotaByTokens branches on TieredSnapshot != nil - Replays RunExprByHashWithRequest with actual token count - Falls through to ratio settlement on expression error Persistence (model/task.go, controller/relay.go): - TaskBillingContext gains TieredSnapshot + TieredRequestBody fields - controller/relay.go copies snapshot from RelayInfo to task record Tests: - seedance_estimator_test.go: 15 subtests covering resolution parsing, duration/frames, video-input detection, unknown model fallback - seedance_presets_test.go: 22 tests updated for simplified expressions (metadata.* cases removed; Option A) - go build ./relay/... ./service/... ./model/... ./controller/... clean - go test ./relay/channel/task/doubao/... PASS (all 15) - go test ./relay/channel/volcadapter/... PASS (all 22) Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Plan B1 (commit 975494f6f) implemented task tiered_expr support but made three architectural mistakes: 1. Volc-specific code lived in taskdoubao.TaskAdaptor, polluting an adaptor shared by channels 45 (VolcEngine) / 54 (DoubaoVideo) / 58 (VolcAdapter). 2. Tiered_expr settlement was added to generic service/task_billing.go RecalculateTaskQuotaByTokens, when the proper extension point — adaptor.AdjustBillingOnComplete — already exists in task_polling.go:540-560 (called BEFORE the token-recalc fallback). 3. BillingContext stored ~1-2 KB of raw request body per task to feed billingexpr param() lookups at settlement, while task.Data already has the Volc fetch response with resolution/duration/service_tier. This commit: - Extracts relay/channel/task/volcadapter/ as the dedicated adaptor for ChannelTypeVolcAdapter, embedding taskdoubao.TaskAdaptor for shared task plumbing (ParseTaskResult, FetchTask, etc). - Cleans Volc-specific branches and helpers out of taskdoubao, bringing it close to its upstream/main state. - Implements tiered_expr settle in volcadapter.AdjustBillingOnComplete, reading expression hash + 3 flags from BillingContext and resolution/duration/service_tier from task.Data (Volc fetch result). - Reverts the tiered_expr branch in service/task_billing.go. - Replaces TieredRequestBody []byte with TieredVolcFlags struct (~30 bytes vs ~1-2 KB per task). Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…e merge 1. Drop the now-commented EndpointTypeVolcImage/Video references from common/endpoint_defaults.go, common/endpoint_type.go, and constant/endpoint_type.go. These three files now match upstream/main byte-for-byte, simplifying future merges. 2. Inline volcRelayHandler (a one-line wrapper calling relay.VolcImageHelper) into the switch case in controller/relay.go, matching the style of WssHelper / ClaudeHelper / etc. 3. Remove ConvertVolcRequest from the Adaptor interface. It only had one real implementation (volcengine.Adaptor's no-op pass-through); all other 30+ stub implementations existed solely to satisfy the interface. Replace with a local type-assertion in relay/volc_handler.go's VolcImageHelper. 4. Merge relay/channel/volcadapter/ into relay/channel/task/volcadapter/. The former held only the channel's ModelList/ChannelName constants; the latter held the task adaptor. Same package name, same domain — no reason to split. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
The DraftTask struct and ContentItem.DraftTask field were added in cursor's feat/volc-adapter as part of the OpenAI-format draft_task content type passthrough for Seedance 1.5 pro. Upstream new-api does not have this feature. VolcAdapter (channel 58) already supports draft_task via raw body pass-through at /api/v3/*; users who want this feature should use the native Volc-compat path. The OpenAI-format /v1/video/generations entry on channels 45/54 reverts to upstream behavior (no DraftTask). Reduces fork divergence from upstream/main. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…tions route Path2RelayMode only handles /v1/... paths; the native Volc Ark image route at /api/v3/images/generations was leaving RelayMode=0 (Unknown), causing volcengine.Adaptor.GetRequestURL to return "unsupported relay mode: 0". Fix: explicitly set relay_mode context key to RelayModeImagesGenerations before calling controller.Relay, matching the pattern already used for task routes in video-router.go. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Two bugs found and fixed during Volc SDK e2e testing: 1. relay/relay_task.go: videoFetchByIDRespBodyBuilder now returns the Volc-native ContentGenerationTask JSON (patching id to public task ID) when relay_format is "volc". Without this the SDK's polling loop received an internal TaskDto wrapper with code/data fields instead of the flat id/status/usage/content structure the SDK expects. 2. middleware/distributor.go: GET /api/v3/contents/generations/tasks/:id no longer requires a model name in the request body. The new branch sets shouldSelectChannel=false for GET (task fetch) and reads model from body for POST (task submit). Without this every task status poll aborted with "Model name not specified". Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Adds unit tests for security boundaries identified during Tier 1 E2E review: - List filter user_id spoofing rejected (token-derived only) - Volc body validator handles malformed inputs cleanly - Task model not in any channel returns clean 4xx - Path traversal in task_id treated as literal lookup - Channel-Id header restricted to admin (verify upstream behavior preserved) Tier 1 tests live in scripts/volc-local-test/ as E2E integration. These are the unit-test counterparts. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…r channel
Reverses an earlier cleanup decision (commits f97a0c616 + 1f71ca830) that
commented out the volc-image and volc-video endpoint types to match
upstream's Kling/Jimeng/Suno/Midjourney pattern. UX feedback: with the
types removed, the model marketplace shows "openai" as the primary
endpoint for seedream/seedance models served by ChannelTypeVolcAdapter,
which is misleading — the gateway's native paths are /api/v3/*, not
/v1/*.
Restores:
- EndpointTypeVolcImage = "volc-image" → /api/v3/images/generations
- EndpointTypeVolcVideo = "volc-video" → /api/v3/contents/generations/tasks
- GetEndpointTypesByChannelType[VolcAdapter] dispatches per-model:
seedream → [volc-image, image-generation, openai]
seedance → [volc-video, openai-video]
- Tests guard that VolcEngine (45) and DoubaoVideo (54) still
fall through to defaults — only channel 58 surfaces volc-* types.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Replaces the 501 stub at DELETE /api/v3/contents/generations/tasks/:id. Verifies ownership, forwards DELETE to upstream Volc with channel API key, updates local task to cancelled status, and refunds pre-charge quota. Handler uses only common/constant/logger/model/service — no graylight deps. Ported from 5d9b05a2c (graylight/track-e-tos-signer), callback handler excluded. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Remove the 12 seedream-*/seedance-* alias entries (without the doubao- prefix) and their two comment headers. ModelList now contains only the 12 canonical doubao-prefixed IDs. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…S list
fix(billing): use gjson-standard double-quoted path in video_url check
The `param("content.#(type=='video_url')")` runtime check never fired because
gjson only accepts double-quoted string literals inside `#()` filters; the
single-quote form was treated as a literal apostrophe and never matched actual
values. This caused the 0.608696× video-input discount on
doubao-seedance-2-0-260128 and the 0.594595× discount on
doubao-seedance-2-0-fast-260128 to be silently skipped, overcharging admins
~64% for video-input requests.
Changes:
- TieredPricingEditor.jsx: update both seedance 2.0 PRESET_GROUPS entries to
use `content.#(type=="video_url")` (double-quoted gjson syntax, JS
single-quoted outer string).
- requestRuleExpr.js: replace all `[^"]+` path-capture regexes in
tryParseRequestCondition with `((?:[^"\\]|\\.)*)` plus unescapePath() helper,
so that round-tripping paths containing escaped double-quotes works correctly.
- requestRuleExpr.test.js: 7 vitest tests covering build→expr, expr→parse, full
round-trip, multi-group (seedance 2.0), and a regression test proving the old
single-quoted form was broken.
- web/package.json: add `test` script (`vitest run`).
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
ChannelTypeVolcAdapter (58) previously routed through APITypeVolcEngine to volcengine.Adaptor, which required adding a no-op ConvertVolcRequest to the volcengine package. This couples our new channel to the existing Volc channel and means any change to volcengine implicitly affects volcadapter. Add a dedicated APITypeVolcAdapter and channel.Adaptor implementation under relay/channel/volcadapter/. The new adaptor only supports the Volc-native /api/v3/images/generations endpoint; all other relay modes return a clear "not supported" error directing users to the volcengine channel for OpenAI-format requests. relay/channel/volcengine/ is now identical to upstream/main. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
VolcTaskDelete and its helpers (buildVolcDeleteResp, isTerminalStatus, volcDeleteMapStatus) previously lived in controller/graylight/, which the upstream PR cannot depend on. Move them to controller/ proper — the package was also empty after this move, so the directory is gone. router/video-router.go now imports only the single controller package instead of both controller and controller/graylight. No behavior change.
…ugh paths Both the image handler (relay/volc_handler.go) and the task adaptor (relay/channel/task/volcadapter/adaptor.go) were forwarding raw body bytes without applying model mapping or param override after reading from body storage. Image path: extract applyVolcImagePatches helper that operates on map[string]json.RawMessage so all unknown Volc-specific fields (sequential_image_generation, optimize_prompt_options, watermark, etc.) are preserved while patching "model" and injecting ParamOverride fields. Task path: add ApplyParamOverrideWithRelayInfo call after the existing patchVolcBodyModel / model-extraction block, before any future injectSafetyIdentifier / injectCallbackURL steps. Tests: 4 new unit tests for applyVolcImagePatches covering model mapping, no-op when not mapped, param override injection, and both combined; 1 new test for BuildRequestBody asserting ParamOverride is applied while preserving unknown fields. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Without an explicit p term in the expression, the frontend tier breakdown UI fails to render the prompt-cost line. Adding p*0 keeps the cost identical but lets the UI parse the expression. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
…billing.go The seed contained ¥-denominated rates which only make sense for CN deployments. Pre-seeding on every fresh install is wrong because USD-denominated deployments would silently get ~7× over-charges. Admin must configure billing per-deployment via the UI; the PRESET_GROUPS in TieredPricingEditor.jsx still expose curated templates as a starting point that admin can load and adjust before save. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Both edits crept in during the feat/volc-adapter merge but are pure
churn — the import reordering in video-router.go and the
make()→map[]{} swap + extra comment in tiered_billing.go don't
contribute to the seedance preset feature this branch carries.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
tryParseRequestCondition runs on every UI edit in the tiered-pricing editor. Compiling six RegExp objects per call is avoidable overhead; hoisting them to module-scope constants compiles them once at module load. Also factors out QUOTED_VALUE_RE for the two has()/== templates that share that fragment, and the numeric op→MATCH_* map. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
The quoted-value regex fragment accepts any \\. escape sequence, but
JSON.parse only handles JSON's escape grammar (\", \\, \/, \b,
\f, \n, \r, \t, \uXXXX). When admins type a half-finished
expression like has(header("x"), "\q") the regex matches but
JSON.parse throws, bubbling up and crashing the editor. Guard both
JSON.parse call sites with safeJsonParse → null on failure, mirroring
the unescapePath try/catch pattern, and add tests for both cases.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Squashed cherry of the PR QuantumNous#5 review feedback train (private fork). Generic-only files; graylight-specific callback wiring is left out. Key changes: - service/task_billing.go: settle now wraps task.Quota persistence, funding adjustment, used_quota / channel.used_quota updates and billing log in a single DB transaction. RowsAffected guard on the targeted UPDATE so a phantom task ID aborts the settle without applying side effects. Subscription updated_at now reuses the settle timestamp. Cache invalidation moved out-of-tx to keep the primary commit fast. - service/task_polling.go: SettleTaskBillingOnComplete is exported, accepts a nil adaptor (token fallback path), prefers TotalTokens but falls back to CompletionTokens when only the latter is reported. - model/task.go: TieredVolcFlags grows resolution / duration / service_tier fields so the tiered_expr settle path can evaluate param() lookups even on callback-only deployments where task.Data is just {"id":"..."}. - controller/relay.go: extractVolcFlags refactored to call the new parseVolcDuration helper. parseVolcDuration accepts float64, int, int64, json.Number and string forms; rejects negative, non-integer and out-of-int-range values. The decoder uses json.Decoder.UseNumber so numeric JSON values arrive as json.Number rather than float64. - relay/channel/task/volcadapter/adaptor.go: buildSynthesizedBody reads resolution / duration / service_tier from TieredVolcFlags first, falling back to task.Data only when flags are absent — fixes settle on callback-only deployments. AdjustBillingOnComplete comment updated to reflect exported SettleTaskBillingOnComplete name. Tests: - service/task_billing_test.go: regression tests for the targeted UPDATE persistence, RowsAffected==0 guard, funding rollback, CAS win/lose on both refund and settle, nil-adaptor token fallback, PerCallBilling skip with nil adaptor, effectiveTokenCount fallback. - controller/relay_volc_flags_test.go: parseVolcDuration string / json.Number / oversized / non-integer cases. - relay/channel/task/volcadapter/adaptor_test.go: two new cases for callback-path flags carrying resolution and flags taking priority over task.Data in buildSynthesizedBody. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Contributor
|
Caution Review failedThe pull request is closed. ℹ️ Recent review info⚙️ Run configurationConfiguration used: Organization UI Review profile: CHILL Plan: Pro Run ID: 📒 Files selected for processing (56)
WalkthroughAdds a new VolcAdapter channel (58) with Ark-native image and video task routes, adaptors, distributor/routing, token estimation, tiered billing, task delete/fetch/list, model catalogs, request validation, and extensive tests. UI gains channel option and Seedance tier presets. ChangesVolcAdapter Ark-native integration
Sequence Diagram(s)sequenceDiagram
participant Client
participant Gateway
participant Adaptor
participant VolcArk
participant Billing
Client->>Gateway: POST /api/v3/images/generations (Volc)
Gateway->>Adaptor: Convert/DoRequest (raw body)
Adaptor->>VolcArk: Forward request
VolcArk-->>Adaptor: Response
Adaptor-->>Gateway: OpenAI-like result
Gateway->>Billing: Post consume quota
Billing-->>Gateway: OK
Gateway-->>Client: 200 response
Estimated code review effort🎯 4 (Complex) | ⏱️ ~75 minutes Possibly related PRs
Suggested reviewers
Poem
✨ Finishing Touches🧪 Generate unit tests (beta)
|
There was a problem hiding this comment.
Pull request overview
This PR adds a new Volc Ark native pass-through channel (ChannelTypeVolcAdapter / API type 58) that forwards Volc Ark /api/v3/* requests byte-identically (preserving Volc-private fields), plus introduces tiered_expr billing integration and a transactional async-task settlement refactor. It also updates the Classic web UI to provide Seedance tiered pricing presets and improves request-rule expression parsing.
Changes:
- Add VolcAdapter (58) routing, adaptors, handlers, and Volc-native
/api/v3/images/generations+/api/v3/contents/generations/tasks*endpoints. - Integrate tiered_expr for Volc task pricing (token estimation + settlement snapshotting) and refactor async task settle/refund to commit accounting atomically.
- Enhance Classic UI tiered pricing editor presets and request-rule expression build/parse (with new tests).
Reviewed changes
Copilot reviewed 56 out of 56 changed files in this pull request and generated 10 comments.
Show a summary per file
| File | Description |
|---|---|
| web/classic/src/pages/Setting/Ratio/components/TieredPricingEditor.jsx | Adds Doubao Seedance tiered pricing presets. |
| web/classic/src/pages/Setting/Ratio/components/requestRuleExpr.js | Improves request-rule expression parsing/perf and escaping behavior. |
| web/classic/src/pages/Setting/Ratio/components/requestRuleExpr.test.js | Adds vitest coverage for expression round-trips/escaping. |
| web/classic/src/helpers/render.jsx | Adds icon mapping for channel type 58. |
| web/classic/src/constants/channel.constants.js | Adds UI option for VolcAdapter channel type 58. |
| web/classic/package.json | Adds vitest run test script. |
| types/relay_format.go | Introduces RelayFormatVolc constant. |
| setting/billing_setting/tiered_billing.go | Minor import formatting change. |
| service/task_polling.go | Exports settle helper, adds nil-adaptor + token fallback logic. |
| service/task_billing.go | Adds transactional settle for async task quota/funding/stat updates. |
| service/task_billing_test.go | Adds extensive tests for new transactional settle logic and edge cases. |
| router/video-router.go | Adds Volc-native task routes under /api/v3/.../tasks*. |
| router/relay-router.go | Adds Volc-native image route /api/v3/images/generations. |
| relay/volc_task_test.go | Adds tests around Volc task routing/body passthrough (but contains stale DELETE stub expectation). |
| relay/volc_handler.go | Implements Volc image handler with byte-level patching. |
| relay/volc_handler_test.go | Adds tests for Volc image handler and patching behavior. |
| relay/relay_task.go | Adds Volc-native task fetch/list response builders and tiered token estimation hook. |
| relay/relay_task_volc_list_test.go | Adds tests for Volc task list builder filtering/security invariants. |
| relay/relay_adaptor.go | Registers APITypeVolcAdapter adaptor + task adaptor routing for channel 58. |
| relay/relay_adaptor_test.go | Adds tests validating adaptor routing for VolcAdapter. |
| relay/helper/valid_request.go | Adds Volc image request validation dispatcher. |
| relay/helper/valid_request_volc_test.go | Adds tests for Volc image validation parsing/robustness. |
| relay/helper/price.go | Adds tiered_expr pricing branch for async tasks using estimated tokens. |
| relay/constant/relay_mode.go | Adds RelayModeVideoFetchList constant. |
| relay/common/relay_info.go | Adds RelayFormatVolc relay-info creation and EstimatedBillingTokens field. |
| relay/channel/volcadapter/adaptor.go | New image adaptor for VolcAdapter channel type 58. |
| relay/channel/task/volcadapter/volc_helpers.go | Volc-native task request parsing + body patch helpers. |
| relay/channel/task/volcadapter/seedance_presets_test.go | Adds regression tests for Seedance tiered_expr presets. |
| relay/channel/task/volcadapter/seedance_estimator.go | Implements Seedance token estimation formula. |
| relay/channel/task/volcadapter/estimator_test.go | Adds tests for Seedance estimator corner cases. |
| relay/channel/task/volcadapter/constants.go | Adds VolcAdapter model list constants (Seedream/Seedance). |
| relay/channel/task/volcadapter/constants_test.go | Adds tests for model list sanity (duplicates/type classification). |
| relay/channel/task/volcadapter/adaptor.go | New VolcAdapter task adaptor: byte passthrough + tiered_expr settlement. |
| relay/channel/task/volcadapter/adaptor_test.go | Adds unit tests for task adaptor settlement/body behavior. |
| relay/channel/task/taskcommon/helpers.go | Extends base billing helpers with EstimateBillingTokens default. |
| relay/channel/task/doubao/adaptor_test.go | Adds regression guards for legacy doubao task adaptor paths. |
| relay/channel/adapter.go | Extends TaskAdaptor interface with EstimateBillingTokens. |
| model/task.go | Adds TieredSnapshot + TieredVolcFlags to task billing context. |
| middleware/distributor.go | Adds Volc-native task path handling in distributor model extraction. |
| middleware/distributor_test.go | Adds tests for distributor security boundaries and Volc path behavior. |
| dto/volc_image_request.go | Adds Volc image request DTO with Extra-field capture. |
| controller/volc_task_delete.go | Adds DELETE task cancellation handler with refund + upstream forwarding. |
| controller/volc_task_delete_test.go | Adds unit tests for delete response/status mapping helpers. |
| controller/relay.go | Wires RelayFormatVolc image helper + snapshots Volc flags for tiered_expr. |
| controller/relay_volc_flags_test.go | Adds tests for Volc flags extraction and duration parsing. |
| controller/model.go | Publishes VolcAdapter model list for model listing endpoints. |
| controller/channel-test.go | Updates channel test routing for seedream models under VolcAdapter/VolcEngine. |
| constant/endpoint_type.go | Adds volc-image and volc-video endpoint types. |
| constant/channel.go | Adds ChannelTypeVolcAdapter (58) and base URL mapping. |
| constant/channel_test.go | Adds tests for VolcAdapter registration and base URL. |
| constant/api_type.go | Adds APITypeVolcAdapter constant. |
| common/model.go | Marks “seedream” as an image-generation model prefix. |
| common/endpoint_type.go | Adds endpoint type resolution for VolcAdapter (image vs video). |
| common/endpoint_type_test.go | Adds tests for VolcAdapter endpoint type mapping. |
| common/endpoint_defaults.go | Adds default endpoint infos for volc-image/volc-video. |
| common/api_type.go | Adds ChannelTypeVolcAdapter → APITypeVolcAdapter mapping. |
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
Comment on lines
+265
to
+279
| } else if strings.HasPrefix(c.Request.URL.Path, "/api/v3/contents/generations/tasks") { | ||
| // Volc-native task routes (/api/v3/contents/generations/tasks and .../tasks/:id). | ||
| // GET requests (task fetch and list) do not need channel selection. | ||
| // POST requests (task submit) extract model from the request body. | ||
| if c.Request.Method == http.MethodGet { | ||
| shouldSelectChannel = false | ||
| } else if c.Request.Method == http.MethodPost { | ||
| req, err := getModelFromRequest(c) | ||
| if err != nil { | ||
| return nil, false, err | ||
| } | ||
| if req != nil { | ||
| modelRequest.Model = req.Model | ||
| } | ||
| } |
Comment on lines
+13
to
+28
| // TestVolcTaskDelete_Returns501 verifies that the DELETE route for volc tasks | ||
| // returns 501 Not Implemented with a recognizable message. | ||
| // The route uses an inline handler (no dedicated controller function). | ||
| func TestVolcTaskDelete_Returns501(t *testing.T) { | ||
| gin.SetMode(gin.TestMode) | ||
| w := httptest.NewRecorder() | ||
| router := gin.New() | ||
|
|
||
| // Mirror the inline handler registered in router/video-router.go. | ||
| router.DELETE("/api/v3/contents/generations/tasks/:id", func(c *gin.Context) { | ||
| c.JSON(http.StatusNotImplemented, gin.H{ | ||
| "code": "not_implemented", | ||
| "message": "DELETE /api/v3/contents/generations/tasks/:id is not supported yet", | ||
| "status_code": http.StatusNotImplemented, | ||
| }) | ||
| }) |
Comment on lines
+317
to
+321
| // Unescape a JSON-string-escaped path (e.g. content.#(type==\"video_url\") → content.#(type=="video_url")). | ||
| // Only handles the two escape sequences that JSON.stringify emits for path strings: \" and \\. | ||
| function unescapePath(raw) { | ||
| return raw.replace(/\\"/g, '"').replace(/\\\\/g, '\\'); | ||
| } |
Comment on lines
+344
to
+356
| // GetAndValidateVolcImageRequest parses the native Volc Ark image request body. | ||
| // Only model presence is strictly validated; prompt/image validation is relaxed | ||
| // because Volc supports both t2i (prompt required) and i2i (image required) in | ||
| // the same endpoint and we want to pass all other fields through byte-identical. | ||
| func GetAndValidateVolcImageRequest(c *gin.Context) (*dto.VolcImageRequest, error) { | ||
| req := &dto.VolcImageRequest{} | ||
| if err := common.UnmarshalBodyReusable(c, req); err != nil { | ||
| return nil, err | ||
| } | ||
| // Accept model, model_name, or req_key as model identifier (Volc uses all three) | ||
| if req.Model == "" { | ||
| return nil, errors.New("model is required") | ||
| } |
Comment on lines
+448
to
+454
| if json.Unmarshal(t.Data, &probe) == nil { | ||
| if _, hasStatus := probe["status"]; hasStatus { | ||
| // task.Data is an upstream Volc response — return it with the | ||
| // public task ID so the SDK's polling loop can match responses. | ||
| probe["id"] = json.RawMessage(`"` + t.TaskID + `"`) | ||
| if patched, err := json.Marshal(probe); err == nil { | ||
| return patched |
Comment on lines
+488
to
+491
| b, err := common.Marshal(synth) | ||
| if err != nil { | ||
| return []byte(`{"id":"` + t.TaskID + `","status":"` + arkStatus + `"}`) | ||
| } |
Comment on lines
+576
to
+583
| func countFilteredVolcVideoTasks(userID int, queryParams model.SyncTaskQueryParams, modelFilter string, taskIDsFilter map[string]bool) int64 { | ||
| totalBase := model.TaskCountAllUserTask(userID, queryParams) | ||
| if totalBase <= 0 { | ||
| return 0 | ||
| } | ||
| // DB layer has no model/task_ids columns for direct filtering; fetch then filter in-memory. | ||
| allTasks := model.TaskGetAllUserTask(userID, 0, int(totalBase), queryParams) | ||
| var filteredCount int64 |
Comment on lines
+84
to
+88
| req, reqErr := http.NewRequestWithContext(context.Background(), http.MethodDelete, deleteURL, nil) | ||
| if reqErr != nil { | ||
| logger.LogError(c, "VolcTaskDelete: build DELETE request failed: "+reqErr.Error()) | ||
| c.JSON(http.StatusInternalServerError, gin.H{"error": "internal error"}) | ||
| return |
| } | ||
| b, err := common.Marshal(synth) | ||
| if err != nil { | ||
| return []byte(`{"id":"` + t.TaskID + `","status":"` + arkStatus + `"}`) |
Comment on lines
+148
to
+166
| // On model-patch failure the function logs and continues with the un-patched body | ||
| // (conservative: avoids introducing a new error path for a non-critical patch). | ||
| // On param-override failure the function returns an error. | ||
| func applyVolcImagePatches(rawBytes []byte, info *relaycommon.RelayInfo) ([]byte, *types.NewAPIError) { | ||
| // 1. Model mapping patch — byte-level, preserves all unknown fields. | ||
| if info.IsModelMapped && info.UpstreamModelName != "" { | ||
| var bodyMap map[string]json.RawMessage | ||
| if err := common.Unmarshal(rawBytes, &bodyMap); err == nil { | ||
| if newModel, err := common.Marshal(info.UpstreamModelName); err == nil { | ||
| bodyMap["model"] = newModel | ||
| if patched, err := common.Marshal(bodyMap); err == nil { | ||
| rawBytes = patched | ||
| } | ||
| // Marshal of bodyMap failed: log and continue with un-patched body. | ||
| } | ||
| // Marshal of model string failed: log and continue. | ||
| } | ||
| // Unmarshal failed: log and continue with un-patched body. | ||
| } |
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
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.
概述
新增
ChannelTypeVolcAdapter(channel type 58) 渠道，作为火山方舟（Volc Ark）原生 API 的兼容代理。该渠道与现有ChannelTypeVolcEngine(45) 共存，不修改 volcengine 渠道的任何现有行为：/v1/chat/completions等）。/api/v3/contents/generations/tasks*（视频/图片/3D 任务）和/api/v3/images/generations（图片生成）。请求体字节级转发，保留所有火山私有字段（sequential_image_generation、optimize_prompt_options、watermark、safety_identifier等）不被解析丢弃。适用场景：用户使用官方 Volc SDK 或 cURL 直接调用火山方舟 API 时，可以把 baseURL 指到 NewAPI 实例，请求体一字不差地转发到上游。
主要新增功能
1. 新渠道类型 ChannelTypeVolcAdapter (58) + APITypeVolcAdapter
constant/channel.go: 新增渠道类型 58VolcAdapter，base URL 默认指向https://ark.cn-beijing.volces.com。constant/api_type.go+common/api_type.go: 新增APITypeVolcAdapter，渠道 58 路由到独立的 image adaptor，不复用 volcengine.Adaptor。relay/relay_adaptor.go: 注册APITypeVolcAdapter → volcadapter.Adaptor。2. 独立 image channel.Adaptor
relay/channel/volcadapter/adaptor.go(新建，~100 行)：实现完整channel.Adaptorinterface。/api/v3/images/generations，其他 relay mode 返回明确错误"volcadapter does not support relay mode N; for OpenAI-format requests use the volcengine channel"。ConvertVolcRequest作为 no-op pass-through（请求体已经是 Volc 原生格式）。Convert*Request方法返回"not supported"引导用户走 volcengine 渠道。DoResponse委托给openai.Adaptor.DoResponse（火山图片响应与 OpenAI 兼容）。GetModelList复用 task adaptor 的模型列表（同一个上游服务）。3. Task adaptor
relay/channel/task/volcadapter/(新建)：实现TaskAdaptorinterface 处理异步视频/图片/3D 任务。adaptor.go: 字节级请求体透传（保留全部 Volc 字段），仅在info.IsModelMapped时通过patchVolcBodyModel改写model字段。seedance_estimator.go: 7 个 Seedance 模型的 token 估算（Volc 公式：(input_video_dur + output_dur) × W × H × FPS / 1024），用于 tiered_expr 计费的预扣 token 上限。volc_helpers.go: 字节级 JSON patch 工具（保留未知字段）。constants.go: 模型列表（Doubao Seedance 1.0/1.5/2.0 系列 + lite-i2v/lite-t2v/pro/pro-fast 等 7 个完整 ID）。4. Native pass-through handlers
Image:
relay/volc_handler.go(新建)VolcImageHelper处理/api/v3/images/generations，模式参考GeminiHelper：dto.VolcImageRequest（仅捕获 size、prompt 等已知字段，body 仍按原样转发）。volcImageConverterinterface 检查渠道支持，不支持返回 400。applyVolcImagePatcheshelper 在转发前应用 model mapping 和 param override（byte-level JSON patch，不破坏未知字段）。Task:
router/video-router.go+relay/relay_task.go新增
/api/v3/contents/generations/tasksPOST/GET/DELETE 路由：relay_format=volc，走RelayFormatVolc分支）。tasks/:id）和列表查询（tasks?...），返回 Volc 原生ContentGenerationTaskshape。controller/volc_task_delete.go处理任务取消（owner 校验 + 上游 DELETE 转发 + 本地状态更新 + 退款）。5. Tiered_expr 计费在 volcadapter 上的接入
pkg/billingexpr/（已在上游存在）的 tiered_expr 引擎首次接入 task channel：relay/channel/task/volcadapter/adaptor.go: 实现EstimateBilling、EstimateBillingTokens、AdjustBillingOnComplete，使任务在提交时按param("resolution") == "1080p" ? ... : ...等表达式计算预扣 quota，完成时按实际 completion_tokens 重算。model/task.go: 新增TieredVolcFlags字段（generate_audio、draft、has_video_input、resolution、duration、service_tier），在提交时从请求体抓取，settle 时通过param()暴露给表达式。这在「callback-only 部署」（task.Data仅含{"id":"..."}的场景）下保证表达式能拿到正确参数。controller/relay.go: 新增extractVolcFlags+parseVolcDuration，支持float64/int/int64/json.Number/string多种 duration 形态，拒绝非整数和负数（避免15.9被静默截断成15）。6. 异步任务结算（settle）链路重构
service/task_billing.go+service/task_polling.go重构异步任务的差额结算：tasks.quotaUPDATE + 钱包/订阅扣款 +users.used_quota/channels.used_quota更新 + billing log 一次性提交，避免「task UPDATE 成功但钱包写入失败」的不一致。RowsAffected守卫：targeted UPDATE 匹配 0 行（task ID 错误或行被删）时立刻 abort，不应用任何副作用。SettleTaskBillingOnComplete导出，接受niladaptor（callback handler 在 init 顺序未跑完时仍可结算，走 token fallback 路径）。effectiveTokenCounthelper：优先TotalTokens，回退CompletionTokens（部分火山响应只填 completion 一侧）。task.UpdateWithStatus(snap.Status)返回won标志，polling 和 callback 两条路径都检查，防止双重结算。model.InvalidateUserCache(task.UserId)确保 Redis 用户余额不滞后。7. 计费表达式编辑器 UI
web/classic/src/pages/Setting/Ratio/components/（已有的分档计费编辑器）增强：TieredPricingEditor.jsx: 加 7 个 Doubao Seedance 模型预设（PRESET_GROUPS list）。requestRuleExpr.js: hoist regex 到 module 作用域（性能）；has()表达式安全 parse 带引号值（避免输入\q之类崩溃）；用双引号代替单引号修复 gjson 路径匹配（Volc 视频字段如content.#(type=="video_url")）。requestRuleExpr.test.js: 新增针对 has() 安全 parse 的测试。改动文件清单
总计 56 文件，+5685/-88 行（其中测试约 3000 行）。
测试覆盖
relay/channel/task/volcadapter/: 16 个测试（adaptor、estimator、Seedance presets、constants）relay/volc_handler_test.go: 8 个测试（body pass-through、ConvertVolcRequest 支持检测、model mapping、param override）controller/relay_volc_flags_test.go: 3 个测试（duration 多种形态解析、非法输入拒绝）controller/volc_task_delete_test.go: 4 个测试service/task_billing_test.go: 19 个测试（事务结算、CAS、token fallback、PerCallBilling 跳过、各种回滚场景）relay/relay_task_volc_list_test.go: list 端点构造测试relay/volc_task_test.go: 完整 RelayTask 流程common/endpoint_type_test.go、constant/channel_test.go、relay/relay_adaptor_test.go: 注册一致性测试middleware/distributor_test.go: routing 测试web/classic/src/pages/Setting/Ratio/components/requestRuleExpr.test.js: 编辑器表达式工具测试兼容性
relay/channel/volcengine/*：现有 volcengine 渠道（45）行为完全不变。APITypeVolcAdapter不影响其他 API type 的路由。BillingMode = tiered_expr时仍走原来的 ratio 计费。TieredVolcFlags）作为task.PrivateData.BillingContext的子结构以 JSON 存储，不影响 schema migration。配置说明
要使用 volcadapter 渠道：
VolcAdapter(58)。https://ark.cn-beijing.volces.com，可改。doubao-seedance-2-0-260128等 ID（参考relay/channel/task/volcadapter/constants.go）。请求示例（视频生成）：
Summary by CodeRabbit
Release Notes
New Features
POST /api/v3/images/generations) and video task management (POST/GET/DELETE /api/v3/contents/generations/tasks,GET /api/v3/contents/generations/tasks/:id).Improvements