2026-08-01 — rabbit history

Source docs: none — direct chat request from jay (dictated, no task/design file).

docs: render all markdown docs to browsable HTML under docs/html/

jay wanted to read the docs index on his phone's browser, which can't render raw .md. Added scripts/generate-docs-html.mjs (uses the repo's existing marked dependency) to render every .md file in the repo (290 files, excluding node_modules/.venv/build dirs) to a self-contained HTML page under docs/html/, mirroring the source path 1:1 (docs/features/x.md → docs/html/docs/features/x.html) so relative links between converted docs resolve correctly by simple extension swap, and non-.md asset links (images) are re-pointed back at their original location instead of being copied. Source .md files are untouched. Added pnpm docs:html as a package.json script to re-run it later as new docs are added.

Repointed all 19 .md card links on docs/index.html to their generated .html counterparts, and added a new "Current Projects" section at the top linking to Rabbit's own README (docs/html/README.html) and Verex's README, generated the same way in the verex repo (verex/docs/html/README.html, linked cross-repo via a relative ../../verex/... path — only resolves when both repos sit side-by-side under ~/work on this machine).

Also committed a small pre-existing uncommitted diff on this branch separately first (the Jul 30 log-entry addition to the Logs section), so it didn't get tangled into this much larger change.

Scanned the newly-converted docs/zsub/*.md and docs/dev.md output for secrets before pushing; found nothing new — an Etherscan API key and some local-anvil trace hex blobs already existed in dev.md/dev2.md/error.md from an earlier commit (83dd7ff), just duplicated into the new HTML mirror, not newly introduced.

Per jay's explicit instruction, fast-forward merged claude/remove-nostra-core-section (5 commits, clean fast-forward, no divergence from origin/main) directly into main and pushed both rabbit and verex — bypassing the repo's own stated PR-review policy for this one-off docs-only change, at jay's repeated explicit request.

docs: scope correction — keep the Verex README copy inside rabbit, not in verex

jay clarified that all doc-conversion work should stay confined to the rabbit project only — pushing docs/html/README.html to the verex repo (in the change above) was out of scope. Reverted that commit on verex (git revert, pushed) and instead regenerated the same rendered Verex README into rabbit/docs/html/projects/verex/README.html, reading verex/README.md as a read-only source but writing the output only inside rabbit. Updated the "Current Projects" card in docs/index.html to link to the new in-repo path instead of the cross-repo one. verex's git history is now untouched by this work.

chore: pre-commit hook to auto-convert only changed .md files

jay asked to avoid re-converting all 290 files on every change — only the ones that actually changed. Extended scripts/generate-docs-html.mjs to accept explicit file paths as CLI args (converts just those; falls back to a full scan with no args). Added .githooks/pre-commit, enabled via git config core.hooksPath .githooks (chose this over Husky since it's a single-dev repo and needs no new dependency) — on each commit it regenerates docs/html/ for staged .md files only, removes the generated .html for deleted .md files, and stages the result automatically. Known limit: a brand-new .md with no card in index.html yet still needs a manual card added — the hook only keeps already-linked docs in sync.

docs(index): fix Current Projects — feature roadmap doc, not root README

jay clarified the original ask: the "Rabbit" card in Current Projects should link to docs/features/README.md (the feature roadmap/current-work index), not the generic project root README.md I'd used. Repointed the card to html/docs/features/README.html and retitled it "Rabbit — Feature Designs" to match the doc's own heading. No script changes needed — that file was already converted in the initial full pass.

docs(index): same fix for Verex

jay asked for the identical fix on the Verex card, and separately confirmed I hadn't touched the verex repo since the earlier revert (confirmed: git status clean there). Copied verex/docs/features/README.md — verex's own feature-roadmap doc, structurally identical to rabbit's — into docs/html/projects/verex/docs/features/README.html, read-only from verex, written only into rabbit. Repointed the Verex card and deleted the now-orphaned root-README copy (docs/html/projects/verex/README.html), since nothing linked to it anymore.

docs(index): add Rabbit/Verex Tasks sections

jay asked for a new "Tasks" section per repo, latest-first, capped at 6 cards each. Ranked docs/tasks/*.md by last git commit date (not filename — the date embedded in filenames like jun-30-rabbit-design.md isn't reliably the actual edit date, and some files' names don't match their content date at all). Rabbit already had all 6 selected files converted from the initial pass; copied Verex's 5 task files into rabbit the same read-only-source way as the features README. Ties at the same commit are ordered arbitrarily (a few of Verex's task files were all touched in the same commit) — accepted as good enough, not worth over-engineering finer-grained ranking for.

docs: rename features README to current-design.md; merge index sections

jay wanted a stably-named "current design" doc per project instead of a generic README.md, and wanted the separate Tasks sections folded back into "Current Projects" (confirmed this was his original intention — the earlier separate-sections read was a misunderstanding on my part).

Renamed docs/features/README.md → docs/features/current-design.md via git mv (history follows the rename) and updated the live links pointing to it (README.md, docs/handbook/workflow.md, and the 3 design task files citing it as "IA:"). Deliberately left docs/history/*.md and one stale absolute-path mention in jun-26-rabbit.md unedited — those are dated records, not live navigation, and rewriting them to match a later rename would be revisionist.

For Verex: could not rename the actual source (would mean touching the verex repo again, ruled out earlier) — renamed only rabbit's read-only copy to match. Merged "Current Projects" / "Rabbit — Tasks" / "Verex — Tasks" into one "Current Projects" section: 1 current-design card + top-2-most-recent task cards per project = 6 total, matching jay's requested split. Verified every link resolves and confirmed verex's git status stayed clean throughout.

docs: fix — Current Design belongs in docs/tasks, not docs/features

jay corrected the above entry: the intended file was docs/tasks/jun-30-rabbit-design.md (the most recently-edited design doc — referenced by ~10 other files with section anchors like #s2/#s5b), not docs/features/README.md. Reverted the features rename and its link updates, then renamed docs/tasks/jun-30-rabbit-design.md → docs/tasks/current-design.md, fixing all live references across docs/rabbit-design.md, tasks/details/, tasks/jul-01-..., and 6 docs/features/*.md files. History entries left alone as before.

Applied the same correction to Verex's rabbit-side copy only: jul-28-verex-design.md (the verex equivalent by last-edit recency) is now current-design.html there. Verex's real repo stayed untouched throughout — confirmed via git status after each step.

docs: full correction — restore Feature Designs card, rename to current-plan.md, add View All Tasks

jay laid out the complete intent again after two rounds of partial misses: (1) keep the section merge (confirmed correct); (2) restore the docs/features/README.md "Feature Designs" card, which I'd wrongly dropped; (3) rename the task-design file to current-plan.md (changed his mind from "current-design"); (4) Current Projects = 6 cards total, 3 per project (Feature Designs + Current Plan + 1 latest task); (5) a "View All Tasks" link for files that don't make the cut, mirroring the Logs section's existing "View All Logs" pattern.

Given this was the third correction cycle on the same section, reflected the full understanding back in chat and got explicit confirmation ("Go!") before touching files again, rather than risk a fourth rework. Built docs/tasks/summary.md (real file, flows through the normal pipeline) and a hand-built Verex-side equivalent (docs/html/projects/verex/docs/tasks/summary.html, since verex's repo isn't touched) listing every task newest-first.

Noted and worked around a side effect: my own cross-reference edits earlier today had bumped some files' git-commit dates, which would have skewed a fresh "latest" recompute — used the ranking already established earlier in the session instead of re-deriving from the now-polluted dates. Verified all links resolve, exactly 6 cards in the section, and verex's git status stayed clean throughout.

feat: Jay Chat — public About-Jay persona chat (branch: claude/jay-chat-public)

jay wanted the "AI Chat" nav item restored and renamed to "Jay Chat" — turns out the real gap (already flagged in docs/features/ai-chat.md) was that /chat has always required login, so logged-out visitors (employers/clients — the actual intended audience) never had access. Talked through the design first: public keyless persona-only endpoint, GitHub + LinkedIn as additional corpus sources (LinkedIn manually, never scraped — against its ToS), production model (OpenAI) over local LLM for now, RAG technique — all per jay's explicit calls.

Security discussion before building: agreed on 5 required (not optional) safeguards — dedicated API key, OpenAI account-level hard spending cap, hourly token budget, per-IP burst guard, scoped system prompt. Then refined the budget scope with jay: hourly token total (not per-minute, not per-message-count) since a real employer conversation shouldn't hit an artificial wall, and —jay's call, given traffic is currently rare—global pool rather than per-visitor, accepting that tradeoff as fine for now and cheaply upgradable later if traffic picks up.

Built as an isolated new feature (/jay-chat, /api/jay-chat, lib/jay-chat.ts) rather than retrofitting the existing private /chat, so there's zero risk to it. Reused the already-working RAG core (lib/about-me.ts) unchanged. Added content/profile/github-summary.md from GitHub's real public API (linked0) — picked up automatically by the existing corpus loader, no code change needed. Also fixed a real deploy gap found along the way: content/ wasn't being copied into the Cloud Run image, so the markdown corpus depth silently only worked in local dev — updated the Dockerfile and corrected the now-stale note about it in the feature doc.

Verified end-to-end locally (temporary dev server on a spare port — port 3100 was already in use by what looked like jay's own running server, left untouched): real grounded answers including a GitHub-specific question pulling the new corpus file, oversized-message rejection, and the burst guard tripping at exactly the configured threshold.

Committed on claude/jay-chat-public, not pushed — this is real app code (not the docs-only work from earlier today), so held off pending review. Also not yet deployable: needs a real dedicated JAY_CHAT_OPENAI_API_KEY from jay (not his existing AI_API_KEY) with its own spending cap set in the OpenAI dashboard.

deploy: Jay Chat live on production

jay provided the dedicated key. Verified it directly against OpenAI before wiring it in, then verified the actual /api/jay-chat route end-to-end locally with the real key (not the borrowed AI_API_KEY used for the earlier smoke test) — asked "How do I contact him?", got back the correct grounded answer with jay's real email.

Extended scripts/deploy.sh with the same upsert_secret pattern already used for AI_API_KEY etc. — new rabbit-jay-chat-key Secret Manager secret, wired in only when JAY_CHAT_OPENAI_API_KEY is set in .env.local. jay confirmed he didn't want to wait for the OpenAI-side spending cap before deploying, since the app's own safeguards (hourly token budget, burst guard, request caps — all already tested) provide real protection on their own; the account-level cap is a defense-in-depth backstop, not a blocker.

Merged claude/jay-chat-public into main and pushed first (keeping "what's live" in sync with "what's on main," same as today's earlier pattern), then ran the actual deploy — rabbit-jay-chat-key secret created, new revision (rabbit-00014-v7w) serving 100% of traffic.

Verified live on the real production domain, not just Cloud Run's *.run.app URL: https://www.jaylabs.xyz/jay-chat loads (200), /api/jay-chat returns a real grounded answer, and the nav genuinely shows "제이 챗" (Jay Chat) publicly — confirmed ALLOW_CHAT=true was already set locally so the menu item didn't silently stay hidden behind its visibility flag.

feat: Telegram notifications for site visits + Jay Chat starts, deployed

jay asked for the same Telegram-notification pattern built for verex earlier today to also cover rabbit: notify on home-page/other-page visits and on chat starts. Added lib/visitor-notify.ts (notifyPageView, notifyChatStart), fire-and-forget, 5-minute per-(page, visitor) debounce so refreshes don't spam duplicates. Wired into the home page (app/home/page.tsx, served at / via the existing rewrite), the Jay Chat page, and the first message of a new Jay Chat conversation. Scope assumption: these two entry points, not every page site-wide — flagged for jay to confirm or expand later.

Local testing hit a real but environment-local snag: Node's fetch to api.telegram.org timed out from this sandbox specifically (both IPv6 and IPv4 connection attempts failed at the socket level, confirmed with --dns-result-order=ipv4first), while curl to the same host succeeded instantly. Diagnosed as a sandbox-local networking quirk, not a code bug — didn't chase it further since the identical fetch pattern already proved working end-to-end against verex's real Cloud Run deployment earlier today (jay confirmed receiving that message). Deployed anyway on that basis, and confirmed live afterward (home + jay-chat both 200).

scripts/deploy.sh: TELEGRAM_BOT_TOKEN through Secret Manager (rabbit-telegram-bot-token, same upsert_secret pattern as everything else), TELEGRAM_CHAT_ID as a plain env var.

fix(nav): EN/KO menu labels had drifted apart

jay noticed the top-menu labels didn't match between languages. Found two real mismatches: /market said "마켓" (KO) vs "Hyperliquid Trading" (EN), and /xyz said "XYZ 데모" (KO) vs "PBS" (EN) — both pre-existing, not introduced today. Used each page's own <h1> (which already agrees between languages) as the source of truth rather than guessing: /market's h1 is just "Market" in both languages, /xyz's is "XYZ Demo — PBS consumer track" in both — so aligned the nav's EN labels to match KO (and the pages' own titles) rather than the other way around. Deployed and confirmed live.

Correction: jay wanted the opposite direction — keep the English labels and align Korean to them. Redid it: /market → "하이퍼리퀴드 트레이딩" / "Hyperliquid Trading", /xyz → "PBS" in both (kept as the acronym, matching how "AP2" stays untranslated elsewhere in the nav).

content: enrich the profile tagline

jay wanted the tagline to cover more than the generic specialization line. Drafted an added sentence for review first (he'd asked to see it before committing), iterated once on his feedback (drop "설계" — keep just "구현"), then committed the approved version: L1 blockchain engine built from scratch + hands-on end-to-end experience across DAO governance, NFT marketplace, and prediction market platforms. Feeds both the homepage headline and Jay Chat's RAG corpus.

feat: merge Featured+Projects on home, inline Jay Chat, add /projects

jay's homepage restructure. Reflected the full design back in chat before touching code (it's a live public portfolio, and the request had several interacting parts), and confirmed two open points with him: URL = /projects, and /jay-chat removed entirely rather than kept unlinked.

Home now has one "수행 프로젝트" section replacing the separate Featured + Projects sections, split half/half: left = Verex featured card with the rotating 3D sphere, right = compact text-only project list + a "전체 보기 →" link. Jay Chat is inline below it. The full Featured card and image-card grid moved to the new /projects page.

Notable details: /api/jay-chat is intentionally kept (the inline chat still calls it) — only the page route is gone. Created a new ALLOW_PROJECTS flag rather than reusing ALLOW_CHAT for a differently-named menu item, since silently repurposing the old code would have been confusing later.

Checked both pages visually in a browser at 1280px before deploying, not just via build output — which caught a layout flaw the build couldn't: the Verex card only filled the top of its column, leaving a large empty gap next to the taller project list. Fixed with equal-height columns. Deployed and confirmed live (home 200, /projects 200, old /jay-chat 302, nav shows "수행 프로젝트", inline chat returns a real grounded answer).

style(home): Projects section cut to ~1/3 height

jay's screenshot feedback: the section was too tall, and the per-project description line under each title was what made the right column so high. Measured before/after rather than eyeballing (his target was 30–40% of original): 537px → 179px = 33.3%. Three changes got it there — dropped the description lines (titles only), showed 3 projects instead of 7, and moved "전체 보기 →" from below the list up beside the section heading (~30px saved, and it reads better). Checked at 1280px and 390px.

feat: résumé + LinkedIn corpus, one-click prompts, curated home projects

jay reported Jay Chat refusing "어떤 학교를 나왔어요?" and supplied his résumé (PDF ×3) and a LinkedIn MHTML export. No PDF tooling was installed (no pdftotext/poppler, and pip is externally-managed) — used a throwaway venv with pypdf, and parsed the MHTML via Python's email module.

Built content/profile/resume-career.md from them. Fixes the reported failure, and closed a gap nobody had noticed: the corpus had no mention of his current job — it ended at ZeroOne (2023). Now includes Blockchain Lead at Sapiens AI (2025.07~), the full career back to 1997, certifications and skills.

Privacy call — the important part. Those documents contained a phone number, home address, birth year/age/gender, GPA, a certificate ID, and a personal gmail. This corpus feeds a public, unauthenticated endpoint, so all of it was deliberately excluded (kept only the already-public linked0@me.com). Verified two ways: grepped the corpus, and asked the live production bot for each — it refuses. This matches the privacy line jay drew earlier (relationships/feelings/private life off-limits); he was told explicitly so he can ask for any of it back if he actually wants it.

Also: example prompt buttons now send on click instead of just filling the input — refactored send() to take the text as an argument so the form and buttons share one path. And the home project list is now a fixed set of the three jay called most important (NFT Marketplace, DAO Governance, EVM-based Bosagora Mainnet) rather than the 3 most recent by date.

Verified in a real browser (a scoped locator was needed — the first button.ghost on the page is the EN language toggle, not a chat prompt) and again on production after deploy.

content: Verex as the current prediction market

Asked "예측 시장 프로젝트가 뭐죠?", Jay Chat only described Nostra — the older project — since that was the sole prediction-market entry in the corpus. Added content/profile/prediction-market-verex.md covering both: Nostra as the first take, Verex as the current improved successor (live at verex.jaylabs.xyz, built at Sapiens AI).

Facts pulled from the verex repo rather than written from memory, and deliberately split into "built and running" (CTF markets, CLOB+AMM, multi-outcome groups, async settlement, MM agent, Sepolia on Cloud Run/Cloud SQL) vs "planned/exploratory" (EIP-7702 AA, negative-risk MM, Chainlink CCIP, markets-as-tokens, MCP), following the repo's own S1✅/S2-current/S3–S10-planned status. This bot talks to potential employers — claiming unshipped features as done would be the worst failure mode. Verified: asking whether AA/CCIP already work returns "don't have that detail", not a false yes.

fix(rag): corpus docs were being silently truncated

Found while testing a new career-summary instruction that the bot kept ignoring. Root cause: markdownChunks() produced ONE chunk per file capped at 1200 chars. resume-career.md is ~8.6k, so everything past the first 1200 chars — the entire guidance section, and most of the career history — was unreachable by retrieval. The model wasn't disobeying; it could not see the text.

Fixed by splitting each doc on its ##/### headings so every section (and each employer entry) is separately retrievable, with the cap now per-chunk. Verified every section of resume-career.md now fits under the cap. This silently improves answers across the whole corpus — the older project write-ups had been cut off the same way.

feat: career-summary prompt + honest model attribution

First example prompt is now "주요 경력을 요약해주세요" / "Summarize his career", with corpus guidance for the answer's shape (per jay, refined once): 1997–2013 compressed to one sentence — Myriad folded in at his request, so the detailed part now starts at People & Technology (2013) exactly where his blockchain work begins — then one line per role through Sapiens AI, closing with project highlights, and never omitting the current role.

Chat subtitle now names the model actually in use: "OpenAI ChatGPT(gpt-4o-mini) ... (RAG)". jay asked whether to also write "local LLM"; deliberately did NOT, since Jay Chat only calls OpenAI today (lib/jay-chat.ts has no local path) and advertising an unwired backend on a public page would be false. Worth adding once it actually exists.

note: cost exposure review (Telegram vs OpenAI)

jay asked whether frequent Telegram notifications could get expensive. Telegram's Bot API is free (only rate limits, no billing), and the extra Cloud Run egress is negligible — the real cost driver is OpenAI. While checking, found and corrected an earlier overstatement of mine: the hourly token budget in lib/jay-chat.ts is in-memory and therefore per instance, and both services run maxScale=20, so the effective ceiling is up to 20×30k tokens/hour, not 30k. jay has since set an $80/month hard cap on the OpenAI side, which bounds the worst case (~9 days of a saturated attack to reach it), so the offered Postgres-backed global counter was deliberately NOT built — it would add DB complexity to defend an already-bounded risk. Revisit if Jay Chat ever gets real traffic.