- Add PlatformOpencode constant (domain + service), account accessors
  (IsOpencode/GetOpencodeBaseURL/GetOpencodeAPIKey), extend
  IsOpenAICompatible and normalizeOpenAICompatiblePlatform
- Route opencode groups to the OpenAI-compatible Chat Completions
  gateway; guard messages/responses with a clear 404
- Resolve upstream URL to opencode.ai/zen/go/v1 (credentials base_url
  override supported), strip opencode-go/ model prefix
- Add opencode liveness probe (1-token deepseek-v4-flash chat
  completion) reusing the chat-completions test path
- Add config-gated non-streaming response cache for OpenAI-compatible
  chat completions (X-Cache: HIT on repeat requests, ~0 first token)

- GroupsView: opencode option in platform select + filter
- CreateAccountModal/EditAccountModal: opencode platform button,
  base_url/api_key placeholders (opencode.ai/zen/go/v1), credentials
  builder fallback
- PlatformIcon: opencode terminal icon
- useModelWhitelist: opencode preset with 11 chat-completions models
  (grok-4.5, glm-5.1/5.2, kimi-k3/k2.7-code/k2.6, deepseek-v4-pro/flash,
  mimo-v2.5/pro, hy3)
- i18n en/zh: admin.groups.platforms.opencode
- types: opencode added to GroupPlatform/AccountPlatform

embed_on.go carries //go:build embed; without the tag the binary has no
frontend hosting and / returns gin's default 404 (page not found).

- group_handler: allow opencode in Create/UpdateGroupRequest platform
  oneof (admin UI can create OpenCode Go groups via API)
- opencode_quota.go: fetch OpenCode Go dashboard usage windows
  (5H/weekly/monthly percent + reset) via auth_cookie, ported from
  account-pool's RSC parsing; tolerant of SolidJS unquoted keys
- account_test_service: opencode probe now refreshes quota windows into
  account.extra (opencode_quota_* fields) when workspace_id +
  auth_cookie are present; quota fetch failure never blocks liveness
- frontend: OpencodeQuotaBadge shows 5H/W/M utilization in the
  accounts list for opencode accounts

- platformColors: opencode added to Platform type + all 13 style records
  (teal accent) + isPlatform()
- i18n: opencode labels in admin.accounts.platforms (en/zh) and
  landing providers (en/zh)
- channel-monitor MonitorSettingsPanel, user PlatformUsageBreakdown and
  UserDashboardStats platform label maps
- ErrorPassthroughRulesModal: opencode platform option (backend
  error_passthrough_rule.AllPlatforms now includes opencode)

PlatformTypeBadge defaulted unknown platforms to 'Gemini', so opencode
accounts rendered as Gemini in the admin account list. Add explicit
opencode label (OpenCode Go) + teal badge class; also add opencode to
platformColors.platformLabel and ProxiesView probe-target label.

SyncPricingModels now aggregates LiteLLM catalog models across the
providers behind OpenCode Go (xai/zhipu/moonshot/deepseek/mimo/hunyuan)
instead of failing with unsupported platform.

OpenCode Go accounts with empty extra resolve ResponsesSupport=Unknown,
so ForwardAsChatCompletions skipped the raw path and forwarded to the
default api.openai.com URL, producing 401 (Incorrect API key). opencode
accounts now always take forwardAsRawChatCompletions, which resolves the
upstream to opencode.ai/zen/go/v1/chat/completions.

…ateway

OpenCode Go accounts now serve Anthropic Messages and OpenAI Responses
requests via the raw chat-completions conversion path (opencode.ai only
exposes /v1/chat/completions), so Claude Code can use opencode models
(e.g. deepseek-v4-flash, 1M context). Remove the opencode 404 guards and
force the raw conversion branch for opencode APIKey accounts in both
ForwardAsAnthropic and the responses forwarder.

resolveCCFallbackTarget used GetOpenAIApiKey which returns empty for
opencode accounts, breaking the messages/responses raw chat conversion
path (account missing api_key).

Brings Grok 4.6 support, JWT subscription tier detection, per-group
per-model pricing, the x_search endpoint and ~10 upstream bug fixes on
top of the OpenCode Go platform patches.

Only conflict was group_handler.go, where upstream added
LongContextPricingEnabled/ModelPricing to the create/update group
requests while the fork had added "opencode" to the platform oneof
validation. Kept both.

Verified: backend builds clean, handler/admin/dto/quotaview tests pass.
The one failing service test (ContentModerationRuntimeSnapshotRefresh
FailureKeepsStaleConfig) fails identically on pre-merge fork main, so it
is pre-existing and not a merge regression.

Co-authored-by: Cursor <[email protected]>

The OpenCode Go platform was wired into groups and accounts but never into
the subsystems that keep their own platform list, so several features were
silently unavailable for it.

Per-user platform quotas (was unusable end to end):
- service.AllowedQuotaPlatforms, the ent build-time validator and the
  user_platform_quotas CHECK all omitted opencode, so a DeepSeek limit was
  rejected at the DB even if the UI had offered it
- the frontend duplicated the platform list in three places; replace them
  with one exported PLATFORM_QUOTA_PLATFORMS constant that the type, both
  components and the spec derive from, so adding a platform is one edit

Composite routes:
- the handler binding and CHECK constraint rejected opencode while the
  frontend typed target_platform as Exclude<GroupPlatform,'composite'>,
  which includes it — picking OpenCode returned a 400

Badges:
- GroupBadge had no opencode branch and GroupOptionItem's default branch
  painted opencode with antigravity's violet; both now use the teal already
  assigned to opencode in platformColors.ts

Tests that hardcoded "5 platforms" now derive the count from the constant.

Co-authored-by: Cursor <[email protected]>

Dashboard HTML scrape stays stale after weekly 401/429, so admin accounts could not refresh 5H/week/month windows. Query GET /zen/go/v1/usage and persist the snapshot.

Co-authored-by: Cursor <[email protected]>