Repository navigation
AI: add InterlinedApiClient.Ai.cs service + artifact models #9
Description
Add the service layer for /api/ai/*. Everything after this is UI.
Endpoints
GET /api/ai/status→{subscriber, providers[], defaultModels{}, quota{usedToday,dailyLimit,remaining}}POST /api/ai/suggest→{ok, feature, artifact, usage{inputTokens,outputTokens,model}, quota{}}POST /api/ai/generate→{ok, feature, created{...}, quota{}}
Request shapes
/suggest: {feature, input, context?, model?, maxOutputTokens?}
/generate: {feature, artifact, provider?, model?, scheduleImmediately?, crossPost?}
Artifact kinds by feature
| feature | suggest artifact kind |
generate created |
|---|---|---|
powered_template |
list |
{listId} |
powered_document |
document |
{documentId} |
message_series |
message_series |
{listId} or {scheduledMessageIds,firstScheduledAt,lastScheduledAt} with scheduleImmediately:true |
article_series |
doc_series |
{folderId, documentIds[]} |
writing_assist |
message | thread | tags |
not persistable → 422 invalid_input |
Error codes to model as first-class states
401 unauthorized, 403 subscription_required, 409 no_provider_configured, 422 invalid_input, 422 invalid_ai_output, 429 quota_exceeded, 429 rate_limited (honor Retry-After), 500/502 provider_error.
InterlinedApiException should carry the machine-readable code so the UI can branch on it instead of string-matching messages.
Acceptance criteria
-
Services/InterlinedApiClient.Ai.cswithGetAiStatusAsync,AiSuggestAsync,AiGenerateAsync. -
Models/AiStatus.cs,Models/AiArtifact.cs(+ per-kind payloads),Models/AiQuota.cs,Models/AiUsage.cs. - The
codefield from error bodies is surfaced onInterlinedApiException. -
GET /api/ai/statusverified live against the test account.
Activity
Adron commentedon Sep 16, 2026
Implemented in #137 — work continues in the PR from here.
Three things differed from this issue's sketch, and they matter for #10–#15. First,
quotais not one shape:/statusreturns{usedToday, dailyLimit, remaining}but/suggestreturns only{usedToday, dailyLimit}— noremaining— soAiQuota.Remainingis nullable and UI should bindRemainingOrComputed. Second, AI is Anthropic-only app-wide with no per-user provider keys (removed server-side 2026-09-05): theproviderfield on/generateis audit-ledger-only and the openai/gemini entries indefaultModelsare unreachable, so there's no provider picker to build. Third, thecodefield is surfaced on a newAiApiExceptionrather than onInterlinedApiException— the latter issealedand is being rewritten in another open PR; folding the two together once that lands is noted as a follow-up in the code.Also worth knowing: input-validation 422s cost no quota (verified —
usedTodaystayed 0 across two), but any call that reaches the model does, including one that fails, so the client pre-flights word caps / the 10-word series gate / missing powered_document source refs / non-persistable artifacts locally and never auto-retries.POST /api/ai/generatewas deliberately never called (it writes to the shared test account), so its envelope is contract-transcribed and flagged unverified in the code.