feat(phase6b): LLM analysis layer — DeepSeek + insights + cost ledger + budget guard
Schema:
- insights (fingerprint dedup, occurrence_count, related_session_ids[], priority_score)
- cost_ledger (per-call, cache_hit/miss/output token split + USD)
- prompt_templates (versioned, per-tag, with JSON schema + tier + temperature)
- budget_settings (global key/value: monthly_hard_cap_usd, daily_soft/hard, per_call_max, analysis_paused)
Worker:
- lib/deepseek.ts: Anthropic-compat endpoint client (deepseek-v4-flash | deepseek-v4-pro),
usage→USD with V4 promo pricing, cache_read_input_tokens awareness, extractJson() helper
- lib/budget.ts: checkBudget() returns active|soft_throttled|hard_paused|monthly_paused,
forces flash tier on soft cap, halts on hard/monthly cap or analysis_paused
- lib/prompts.ts: 5 seed templates (bug_triage P, ux_friction F, payment_issue P,
onboarding_stuck F, provider_quality F) with embedded JSON schemas + Sase.tr context;
pickPromptTag() maps session tags → prompt
- lib/json-validate.ts: lightweight schema validator (no ajv dep)
- lib/seed-runtime.ts: idempotent upsert of prompts + default budget settings on worker boot
- lib/minio.ts: +getText() for compressed timeline fetch
- jobs/analyze.ts: budget guard → 6h fingerprint cache (attach session to existing insight) →
template lookup → severity-based tier override → DeepSeek call → JSON parse + validate →
insert insight (or aggregate occurrence) → write cost_ledger
- scheduler: analyze@*/4min on insight-pipeline queue
UI:
- /insights (was pipeline view) → now Insight Inbox: priority-sorted list w/ KPI strip
(new/in_backlog/shipped/today $/month $), severity badges, link to detail
- /insights/i/[id]: insight detail with structured body render, related sessions,
per-session cost breakdown, raw JSON collapsible
- /insights/costs: KPI cards (today, month, avg, cache hit), daily 30d bar table,
by-model + by-prompt breakdowns, top 10 expensive, recent errors
- /insights/pipeline: moved old session-pipeline view here
- /insights/sessions/[id]: unchanged session timeline viewer
Defaults:
- monthly cap $30, daily soft $1.50 / hard $3, per-call $0.20, analysis_paused=false
- Severity→Tier: P0/P1=pro, P2/P3/INFO=flash; budget soft-cap forces flash
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
95c7e8e321