DeepSeek v4-pro reasoning model rejects tool_choice parameter.
All 4 endpoints now use system prompts with JSON schema
instructions and parse content as JSON instead of tool_calls.
Co-Authored-By: Claude <noreply@anthropic.com>
DeepSeek API deprecated deepseek-chat model. Returns error:
'the supported API model names are deepseek-v4-pro or deepseek-v4-flash'
Affected all 4 AI endpoints:
- risk-flags (90-day dashboard)
- triage-suggest
- rca-draft
- quality-check
Co-Authored-By: Claude <noreply@anthropic.com>
Convert all 11 incident API routes from Supabase PostgREST to Drizzle
ORM with withUser/asAdmin/writeAuditLog patterns and RLS enforcement.
Only uploadEvidenceFile retains supabase client (Phase 5 storage work).
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Convert lib/settings.ts, lib/notifications/{in-app,email,capa-escalation,
effectiveness-recheck}.ts from Supabase PostgREST to Drizzle asAdmin queries.
Drop supabase arg from all call sites in app/api/ and cron routes. Rewrite
notification unit tests to mock @/lib/db/with-user instead of SupabaseClient.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Custom auth stack: bcryptjs password hashing (cost 10, GoTrue-compatible),
jose JWT session cookies (edge-safe, 8hr TTL), new API routes for
login/logout/reset/change-password, middleware rewritten to JWT-only
verification with no DB access. All 38 protected pages and API routes
migrated from supabase.auth.getUser() to getSession(). Supabase .from()
queries retained for Phase 4. lib/db/index.ts refactored to lazy Proxy
singleton to avoid module-level throw during Next.js build.
tsc: clean, build: clean, tests: 4/4 passed
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- lib/claude/client.ts: replace Anthropic SDK with openai package pointed at DeepSeek baseURL
- 4 AI routes: port tool definitions, tool_choice, and output parsing to OpenAI function-calling format
- Drop thinking:{type:'adaptive'} (no DeepSeek equivalent); model string → deepseek-chat
- settings/route.ts: add DEEPSEEK_API_KEY to ALLOWED_KEYS
- migration: seed DEEPSEEK_API_KEY placeholder row in app_settings
- tests: update 3 AI route tests to mock createDeepSeekClient + OpenAI response shape
Voyage AI embedding path untouched (DeepSeek has no embeddings endpoint).
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CPf5Rc8QPx6V8KLEEgfKEQ
- Add new migration 20260711000013_match_incidents_auth_guard.sql that
replaces match_incidents with an inline auth guard: callers without
hse/admin role receive PGRST301 Forbidden, closing the SECURITY
DEFINER RLS bypass.
- Wrap anthropic.messages.create() in try/catch returning 503 in all
four AI routes: quality-check, triage-suggest, rca-draft, similar.
- Wrap JSON.parse(inc.embedding) and embedText() in similar/route.ts
in a shared try/catch returning 503 Embedding service unavailable.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FFDuBhMKvoWjrWT3ZnGmmr
Adds POST /api/incidents/ai/quality-check using Claude tool_use to score incident descriptions 1-10. Report form blocks submission when score < 6, shows amber feedback with suggestions, and offers a 'Submit anyway' override.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FFDuBhMKvoWjrWT3ZnGmmr
adminandClaude