jackyu66gitandClaude Opus 4.6
cef29d40bd
Full-stack inner development operating system built on Next.js 16 + Prisma 7 + PostgreSQL + OpenAI + Clerk. Generates personalized daily mental training tasks, collects reflection, updates user state, and provides internal reward feedback. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
44 lines
2.9 KiB
Markdown
44 lines
2.9 KiB
Markdown
# Inner OS — Deferred Work
|
|
|
|
## From /plan-eng-review (2026-05-20)
|
|
|
|
### T1: Define state capture mechanism
|
|
**What:** Specify the UX and formulas for capturing attention_level, emotion_stability, and stress_level.
|
|
**Why:** These are inputs to task generation, but no data entry path exists. Session reflection has focus_score and emotion_score (1-5 user self-report), but no mapping to the state model.
|
|
**Context:** The session reflection form needs clear self-report questions ("How focused were you? 1-5", "How emotionally balanced? 1-5"). Map formulas: attention_delta = (focus_score - 3) * 5, emotion_delta = (emotion_score - 3) * 5. For the empty-reflection case (user writes nothing), default deltas to 0 (no change).
|
|
**Effort:** S (human: ~3h / CC: ~15min)
|
|
**Priority:** P1
|
|
**Depends on:** Session reflection form design
|
|
|
|
### T2: Add LLM output content safety
|
|
**What:** Add a lightweight safety layer on LLM-generated tasks before showing them to users.
|
|
**Why:** GPT-4o-mini could generate harmful suggestions (extreme practices, dangerous advice). A single bad output is a reputational event.
|
|
**Context:** Implement a keyword blocklist (self-harm, extreme behaviors, medical advice) and an optional LLM self-review prompt ("Does this task contain harmful or dangerous suggestions? Reply yes/no."). If flagged, fall back to a template task.
|
|
**Effort:** S (human: ~2h / CC: ~10min)
|
|
**Priority:** P1
|
|
**Depends on:** Task generation pipeline
|
|
|
|
### T3: Define stage advancement rules per level
|
|
**What:** Define concrete thresholds for each stage transition (0→1, 1→2, 2→3, 3→4, 4→5).
|
|
**Why:** The LLM prompt needs current stage to set task difficulty. Without per-level rules, advancement is arbitrary.
|
|
**Context:** Proposed rules:
|
|
- Stage 0→1: 7 completed sessions (any consistency)
|
|
- Stage 1→2: 14 completed sessions + consistency >= 60
|
|
- Stage 2→3: 30 completed sessions + consistency >= 80 + stability >= 60
|
|
- Stage 3→4: 60 completed sessions + consistency >= 90 + stability >= 70
|
|
- Stage 4→5: 120 completed sessions + consistency >= 95 + stability >= 80
|
|
**Effort:** S (human: ~1h / CC: ~10min)
|
|
**Priority:** P2
|
|
**Depends on:** —
|
|
|
|
### T4: Add Clerk user deletion webhook handling
|
|
**What:** Handle Clerk user.deleted webhook to cascade-delete user data, plus a periodic cleanup job for orphaned rows.
|
|
**Why:** Clerk webhooks are at-least-once (not exactly-once). If delivery fails, UserState/Sessions/Tasks/RewardEvents become orphaned with no owner.
|
|
**Context:** Add `onDelete: Cascade` on User→UserState Prisma relation. Implement Clerk webhook handler at `/api/webhooks/clerk`. Add a weekly cleanup job that deletes rows where user_id doesn't exist in Clerk (query Clerk API).
|
|
**Effort:** M (human: ~4h / CC: ~15min)
|
|
**Priority:** P1
|
|
**Depends on:** Clerk integration
|
|
|
|
## From /plan-ceo-review (2026-05-20)
|
|
- (all scope items accepted as "Build now" — nothing deferred from CEO review)
|