Prompt library
Endpoint now accepts referenceAssetKind as empty string or null. Admin form still sends empty string.
Contract drift: schema expects image, video, audio, null, or empty compatibility. Caller impact: admin form depends on empty string handling. Validation gap: test null and empty string separately. Rollout note: normalize to null at read boundary. Risk: strict parser may reject existing draft records.
Plan removes live database writes after moving prompt content into reviewed release files. Old rows remain in production.
Data loss risk: low if reads no longer depend on the old table. Backfill order: confirm release-file coverage before disabling writes. Rollback limit: re-enabling database writes may not recreate missed edits. Verification query: count published file templates against old rows. Decision: keep old rows read-only until one release passes.
Request: clean asset ownership, retire an old compatibility path, update docs, and improve audit scripts.
Ticket 1: audit current asset paths and temporary link references. Ticket 2: replace final URLs and verify public pages. Ticket 3: remove the compatibility path only after example coverage holds. Ticket 4: update governance docs and release checklist. Verification: prompt checks, media audit, typecheck, and build.
Prompt rail Use button updates URL query, but textarea keeps the previous prompt after client navigation.
Symptom: URL changes but textarea state does not refresh. Likely stale state: query hydration runs only on first mount. Reproduction: click two prompt cards on the same model page. Minimal fix: observe search params and sync only when relevant values change. Test: direct load and same-page navigation both refill textarea.
Unauthenticated users visiting /zh/studio/image should land on sign-in and return to the localized studio path.
Redirect loop risk: sign-in should not redirect to itself. Locale handling: preserve zh in return path. Protected route leak: studio content remains noindex and gated. Test case: unauthenticated localized studio request. Regression check: default locale and zh should behave consistently.
Need to backfill result_primary_url from result_urls_json for old AI tasks without changing new task writes.
Source of truth: result_urls_json first item for old completed tasks. Dry run: count missing primary URL by status. Write order: old completed tasks only, batch by ID. Verification: compare counts before and after. Rollback limit: primary URL can be cleared only if original JSON remains intact.
Video tasks failed for one provider, but logs only show generic upstream error and support could not see provider code.
Missing log: provider error code and request ID. Missing metric: failure rate by provider and model. Missing trace: upload to generation handoff. Alert gap: no provider-specific spike alert. Next step: persist normalized upstream error source and code for support views.
Playwright version changed and screenshots fail because the matching Chromium revision is not installed.
API change: none confirmed yet. Generated files: no app files should change from browser install. Browser requirement: install matching Chromium revision. Fallback plan: use existing cached revision only if version matches. Verification: run the screenshot command after install and record revision.
Release changes prompt static paths and adds 58 chat templates. No schema change. Build must include new pages.
Switch point: before deploy, rollback is git revert; after deploy, redeploy previous build. Data risk: none from schema, but sitemap count changes. Owner: release engineer for deploy; content owner for template validation. Verification: prompts:check, i18n checks, typecheck, build. Abort condition: missing locale file or prompt static route failure.
Prompt list page feels slower after adding many templates. Server render is static, but client filtering has more items.
Likely cause: client filtering and card rendering scale with item count. Measurement plan: compare hydration time and filter input latency before and after. Safe experiment: memoize search values or virtualize only if needed. Rollback trigger: interaction latency exceeds target on mid-range mobile. Do not change: SEO static generation without evidence of server bottleneck.
Prompt cards now have compact chat previews and action buttons below a clipped conversation block.
Focus order: card link should not trap action buttons. Target size: copy and run buttons need at least 24px target or spacing. Reduced motion: hover sweep should be decorative only. Label check: buttons need visible or accessible action names. Mobile risk: conversation bubble text must not overlap actions.
User opens a model page, clicks a related chat prompt, and the run panel should prefill with that prompt.
User path: model detail to related prompt to run panel. Data boundary: prompt text travels through client navigation. Failure mode: textarea keeps stale prompt. Test case: click two different prompt cards and assert latest value. Verification target: URL and textarea stay in sync.
Checkout completed events add credits. Retry events can arrive twice. Wallet page reads credit ledger.
Idempotency: event ID must be unique before credit write. Replay safety: verify signature and timestamp tolerance. Credit write: ledger entry should reference checkout session. Customer-visible failure: show pending review if payment succeeded but credit write failed. Test gap: duplicate event and out-of-order event cases.
Need to align model config between Rivya and adjacent seed scripts without changing runtime behavior first.
Order: audit current config, compare generated facts, then update seed script. Contract: model slug, category, and provider ID must stay stable. Verification: parity check before runtime change. Rollback boundary: config generation can revert independently from UI content. Risk: changing display fields may affect SEO pages.
Changed prompt template content only. Existing docs had uncommitted edits. Need next owner to review SEO wording.
Touched files: prompt template source and locale files. Invariants: no code path or route behavior changed. Known risk: new pages increase static prompt count. Verification: prompts:check and SEO title audit. Next owner decision: whether to run full build before merge.
Feature lets users upload reference images, keep history, and reuse prompts across studio sessions.
Auth question: who can access reused prompts and uploaded references. Storage question: where reference assets live and when they expire. User data question: whether prompts can contain private customer data. Abuse path: public sharing could expose private media. Review owner: security and product need retention rules before launch.
prompts:check passes, but i18n:check fails after generated message files changed in the working tree.
Changed-file failure: inspect locale JSON shape first. Environment failure: not likely if prompts:check passed. Flaky test: unlikely for deterministic i18n:check. Next command: run i18n:generate, then i18n:check again. Do not do: revert generated files without understanding the source mismatch.
Decision: keep reusable prompt templates reviewed before release instead of editing them directly in a live admin screen.
Context: public prompt pages need static, reviewable content. Options: database CMS, file source, or hybrid writeback. Decision: file source with admin diagnostics only. Consequences: deploy needed for edits, but SEO and review stay stable. Revisit trigger: operations need safe non-developer write workflow.
