2026-02-27 Session Log

Topics Worked On Today

• wrap-up
• blog-logs-activity
• lingua-rag
• competitive-agents

---

wrap-up

Implemented wrap-to-blog skill + expanded blog ActivityCalendar to all collections

Done

• feat: implemented wrap-to-blog skill (plugins/wrap-up/skills/wrap-to-blog/SKILL.md) — automatically saves session records to blog logs collection after wrap-up completes
• feat: added blog_log section to config.yaml (enabled flag, blog_dir, collection)
• feat: added Step 6 to SKILL.md — prompts whether to generate blog log after checking blog_log.enabled
• feat: added blog logs collection schema (src/content/config.ts)
• feat: added logs list/detail pages to blog (src/pages/logs/index.astro, [...slug].astro)
• feat: implemented ActivityCalendar component — GitHub grass style, 52 weeks×7 days, pure CSS Grid
• feat: added Activity section to blog home + added Logs nav item to Header
• feat: expanded ActivityCalendar to support digests/learnings/logs combined
• feat: added ActivityEntry type + getAllByDateMap() to posts.ts — unified date map across 3 collections
• feat: rendered each post title as a clickable link in tooltip, tooltip stays on mouse move with 150ms hover delay
• chore: registered wrap-to-blog symlink (~/.claude/skills/wrap-to-blog)
• feat: added Session Management rules to CLAUDE.md — suggest wrap-up on topic switch

Key Decisions

• wrap-to-blog as a separate skill: can be auto-invoked from Step 6 or run independently as /wrap-to-blog
• ActivityCalendar unified dateMap: expanded from logs-only → unified across 3 collections, abstracted as ActivityEntry { title, url } type
• Tooltip hover delay approach: implemented with pointer-events: auto + mouseenter/mouseleave + 150ms timeout to allow clicking links inside tooltip
• config.yaml reused: was marked as “unnecessary file” in a previous session, but now has a purpose with the blog_log section — kept

Next

• Commit project changes (wrap-to-blog skill + blog changes)
• Test actual wrap-up → wrap-to-blog end-to-end flow
• Verify ActivityCalendar behavior after blog GitHub Pages deployment
• Update plugin.json version

---

blog-logs-activity

Added Logs collection + GitHub-style ActivityCalendar to AI Learning Blog

Done

• feat(blog/content): added logs collection — defined logSchema (title, date, description, tags), registered in config.ts
• feat(blog/pages): logs/index.astro — log list page in reverse chronological order
• feat(blog/pages): logs/[...slug].astro — individual log detail page
• feat(blog/components): ActivityCalendar.astro — GitHub-style annual contribution heatmap (month labels, day labels, 4 intensity levels, tooltip, year selector, date click drill-down, Show more pagination)
• feat(blog/utils): added getAllLogs(), getLogsByDateMap(), getAllByDateMap() to posts.ts
• feat(blog/pages): integrated Activity section into index.astro home
• feat(blog/components): added “Logs” navigation item to Header.astro
• chore(blog/content): added 4 sample logs (2026-02-23 ~ 2026-02-27)

Key Decisions

• getAllByDateMap() aggregates logs + digests + learnings by date into a single calendar
• Calendar rendering uses client-side JS (data-cal JSON payload) — compatible with Astro SSG
• Uses UTC date arithmetic (T12:00:00Z) — prevents date shift from KST timezone offset (+9)
• Future dates rendered transparently (not clickable), full year rendered through December 31st
• Year selector supports multi-year switching; defaults to current year

Next

• Dark mode support — current cell colors are fixed GitHub green (need to replace with CSS variables)
• Add tag filtering to Logs list page
• Consider search functionality (e.g. Pagefind)
• Keep adding log entries to fill the calendar
• Verify behavior on GitHub Pages after deploying changes

---

lingua-rag

History UI, Supabase DB migration, and Google OAuth auth implementation across 3 sessions

Done

History UI + Cold Start Handling

• feat(backend): added message_count LEFT JOIN to list_by_session query
• feat(frontend): added /api/conversations, /api/conversations/[id]/messages, /api/health proxy routes
• feat(frontend): useChat — auto-loads DB history on unit switch
• feat(frontend): useBackendHealth hook — detects backend warm-up (3s polling, max 20 attempts)
• feat(frontend): cold start banner — “Server is starting” amber banner at top

Supabase DB Migration (Render PostgreSQL → Supabase)

• chore(db): migrated Render PostgreSQL → Supabase (PostgreSQL 17.6, Singapore)
• feat(db): schema.sql — added CREATE EXTENSION IF NOT EXISTS vector (v0.2 RAG preparation)
• fix(backend): connection.py — sslmode=require DSN parsing + automatic asyncpg ssl='require' configuration
• fix(deploy): resolved Render deployment failure (OSError: Network is unreachable) → switched to Session Mode Pooler URL

Google OAuth + JWT ES256

• feat(auth): implemented Google OAuth — Supabase Auth + @supabase/ssr, middleware route protection
• feat(auth): deps/auth.py — ES256 token verification via PyJWT JWKS client (cryptography + certifi)
• feat(db): dropped sessions table, replaced with conversations.user_id (auth.users.id)
• feat(frontend): added lib/supabase/{client,server}.ts, middleware.ts, app/login/page.tsx, app/auth/callback/route.ts
• feat(frontend): 3 API proxies — switched from lingua_session cookie to Authorization: Bearer header
• feat(frontend): sidebar bottom user card — initial avatar + name display, click shows popover (email + logout)
• fix(auth): Supabase new project JWT ES256 — switched to JWKS endpoint + PyJWKClient
• fix(auth): macOS Python 3.13 SSL — explicitly injected certifi CA bundle into JWKS client
• fix(deploy): middleware.ts Next.js 15 type error + missing setup/page.tsx commit → resolved 404

Key Decisions

• Chose Supabase Auth: integrates DB + Auth on the same platform — no additional service needed
• JWKS/ES256: new Supabase projects use ES256 instead of HS256. Removed SUPABASE_JWT_SECRET, configured JWKS via SUPABASE_URL
• Explicit certifi injection: macOS Python.org install doesn’t include system CA → use certifi.where() explicitly. Same behavior on Render (Linux)
• Session Mode Pooler: Supabase direct connection is IPv6-only → use Session Pooler (5432) compatible with both Render and macOS
• Removed session cookie entirely: lingua_session httponly cookie → JWT Bearer. Simultaneously resolved OQ-1 (expiry) and OQ-6 (multi-device)

UX Stabilization + Stream Persistence

• fix(frontend): unified session cookie name session_id → lingua_session across 3 proxy locations
• feat(frontend): restructured onboarding routing (/ level select → /setup → /chat?unit=&level=)
• feat(frontend): useChat generationRef — orphan stream handling on unit switch. Drains the stream without aborting so the backend can persist the assistant message to DB
• feat(frontend): persistent ChatPanel via display:none — preserves stream state and React state on unit switch, resumes real-time streaming when user returns
• feat(frontend): sidebar drag-to-resize (160px–480px range, mousedown/move/up events, user-select: none during drag)
• fix(backend): MAX_TOKENS 1024 → 2048 to prevent truncation of long grammar explanations
• feat(frontend): answer copy button — copy icon on hover with 1.5s “Copied” feedback

v0.2 RAG: PDF Indexing + Vector Search

• feat(rag): scripts/index_pdf.py — PDF extraction via pdftotext (subprocess), 204 pages → 250 chunks
• fix(rag): fixed 2 infinite loop bugs in chunk_text() that caused 50GB memory usage / exit 137
• feat(rag): generated embeddings with OpenAI text-embedding-3-small, inserted into Supabase document_chunks
• fix(rag): max_distance 0.5 → 0.7 (chunks contaminated with copyright text pushed distances to 0.61–0.63)
• chore(rag): skipped ivfflat index — sequential scan is faster than index for 250 rows

RAG Indexing Quality Improvement

• fix(rag): rewrote Korean unit header regex patterns for 100% detection accuracy
• fix(rag): LESSON_START_PAGE=10 (skip TOC), LESSON_END_PAGE=178 (exclude appendix)
• fix(rag): MAX_UNIT_STEP=5 monotonic guard — prevents page numbers from being falsely detected as unit numbers
• fix(rag): copyright watermark line filter (“License Number”, “Zusammen A1”, “독독독 독일어”)
• feat(data): full replacement of all 56 unit titles to match actual textbook content
• chore(rag): re-indexed 244 → 186 chunks (removed 58 appendix noise chunks)

Key Decisions

• Chose Supabase Auth: integrates DB + Auth on the same platform — no additional service needed
• JWKS/ES256: new Supabase projects use ES256 instead of HS256. Removed SUPABASE_JWT_SECRET, configured JWKS via SUPABASE_URL
• Session Mode Pooler: Supabase direct connection is IPv6-only → use Session Pooler (5432)
• Persistent Panel via display:none: simpler and more effective than a stream registry. Preserves both React state and in-flight fetch
• pdftotext (CLI) for PDF extraction: pdfplumber/pypdf both consumed 50GB memory due to the chunk_text bug. pdftotext uses per-page subprocess calls to minimize Python heap
• ivfflat index unnecessary: pgvector sequential scan is faster for 250 rows. Revisit at 10K+ rows

Next

• Production E2E verification — confirm Render redeployment, Google login → chat → multi-device history sharing
• Start v0.2 RAG — select embedding model (ADR-003), PDF parsing, pgvector vector search

---

competitive-agents

Ran full launch-kit plugin competitive pipeline (2 rounds) — used planning-interview outputs as mission context, selected Fuse A+B, deployed v1.1.0

Done

• feat(competitive-agents): completed full launch-kit plugin pipeline (2 rounds × 2 agents)
  • Phase 1: Alpha(Pragmatist) + Beta(Architect) v1 generated in parallel
  • Phase 2: Cross-review Round 1 — Alpha 81/100, Beta 88/100
  • Phase 3: Improve Round 1 — v1→v2 (Alpha 89/100, Beta 97/100)
  • Phase 4: Cross-review Round 2 — v2 evaluation complete
  • Phase 5: Improve Round 2 — v2→v3
  • Phase 6: Judge evaluation (Opus) — Alpha wins 88.5/100 vs Beta 82.0/100
• feat(launch-kit): selected Fuse A+B → generated fused v1.1.0
  • Alpha structure (correct file format, symlink-aware template paths) + Beta features integrated
  • Integrated Beta features: notes_partial 3-tier routing, Korean slug handling, email preview text rules, CLAUDE.md developer guide
• chore(launch-kit): copied fused v1.1.0 → plugins/launch-kit/ + npm run link complete
• chore(competitive-agents): handled large temp output files (51–87KB) via Python parsing to extract v2/v3 files

Key Decisions

• Why Beta v3 was overtaken: Beta (97/100 at v2) wrapped all files in ```markdown``` code fences in the v3 improver → YAML frontmatter unparseable → Convention Compliance dropped 15→7.5 → Alpha took the win
• Chose Fuse approach: combined Alpha’s correct file format + Beta’s functional superiority. Margin was small (6.5 points) and Beta features represent genuine UX improvements, making Fuse superior to Use Winner
• notes_partial 3-tier routing: fully extracted (skip) / partially extracted (targeted follow-up only) / not extracted (full question)
• Korean product name slugs: contains_non_ascii() detection → explicitly requests Latin slug, includes empty slug fallback

Next

• Real-world test of launch-kit v1.1.0 — notes_partial 3-tier routing, Korean slugs, email preview text
• competitive-agents: add “do not wrap entire file content in code fences” warning to improver prompt
• Verify competitive-agents Step 11.5 behavior (select generate_docs = Yes) — not tested this session (selected No)