Nhật Báo Tiên Tri

Key improvements for visual impact at production panel sizes (500+px): - Door: real SVG stroke-dashoffset with dynamic viewBox sizing - Pensieve: 6 ripples (was 3), 700px expansion (was 350), + vortex - Floo: 30 flames + 20 embers (was 14 flames), brighter glow - Patronus: 16 rays + central flash core, rays extend 400px (was 200) The D1 mirror pattern: This site follows a consistent architecture — external systems (Todoist, GitHub Issues, Linear, Notion) are the systems of record, but D1 mirrors everything for display. The RoR page never calls external APIs at request time. This keeps edge latency fast and removes external dependencies from the critical render path. | CLAUDE.md as a symlink to AGENTS.md: This repo uses a single source of truth for agent instructions. AGENTS.md is the real file, CLAUDE.md symlinks to it. Edits to either land in the same place. Supports both Claude Code and other agents. | Scaling UI from cards to lists: Card layouts look great with 2-6 items but break down at 20+. The key is progressive disclosure — one dense row per item by default, native HTML details/summary for drill-down. Avoids JavaScript while keeping the page scannable. Stats bar at top gives orientation. | CSS custom properties for per-item theming: Using style={--tech-color} on each element lets you drive all color variations from a single CSS rule. color-mix(in srgb, var(--tech-color) 6%, transparent) creates tinted backgrounds. Scales to any number of items without N separate CSS classes. Files: .superpowers/brainstorm/22808-1776186108/content/ror-production-v2-beefed.html,seed/seed.json,src/pages/room-of-requirement/index.astro

Added 0.97 scale dip at 4% of pensieveDive before the spring to 1.08. Gives the panel that "pressed in, springs out" physical feel that makes clicks feel connected to the animation instead of floating. The D1 mirror pattern: This site follows a consistent architecture — external systems (Todoist, GitHub Issues, Linear, Notion) are the systems of record, but D1 mirrors everything for display. The RoR page never calls external APIs at request time. This keeps edge latency fast and removes external dependencies from the critical render path. | CLAUDE.md as a symlink to AGENTS.md: This repo uses a single source of truth for agent instructions. AGENTS.md is the real file, CLAUDE.md symlinks to it. Edits to either land in the same place. Supports both Claude Code and other agents. | Scaling UI from cards to lists: Card layouts look great with 2-6 items but break down at 20+. The key is progressive disclosure — one dense row per item by default, native HTML details/summary for drill-down. Avoids JavaScript while keeping the page scannable. Stats bar at top gives orientation. | CSS custom properties for per-item theming: Using style={--tech-color} on each element lets you drive all color variations from a single CSS rule. color-mix(in srgb, var(--tech-color) 6%, transparent) creates tinted backgrounds. Scales to any number of items without N separate CSS classes. | Compositing multiple visual layers for depth: smoke z-index 1, vortex z-index 2, primary particles z-index 3, flashes z-index 4. mix-blend-mode: screen on static noise brightens content beneath instead of overlaying. | Astro scoped CSS silently breaks dynamically-created particles. Styles like .pensieve-ripple compile to .pensieve-ripple[data-astro-cid-xxx] and JS-created elements don't have that attribute. Use <style is:global> for any class targeted via document.createElement + classList.add. | Click-time feedback bright flash at t=0 is crucial — humans expect <200ms visual response. A pre-animation radial flash gives immediate feedback while main effect ramps up. | Completion moments add narrative weight: preview animations felt impressive partly because destination label fades in after panel dissolves, givi Files: src/pages/room-of-requirement/index.astro

- Switch all log scripts from `echo | jq -Rs` to `jq -n --arg` to eliminate trailing newlines in JSON string values - Add 3-layer guard to post-commit-diary.sh: command filter, attribution check, and dedup via last-diary-commit file - Remove non-standard `if` and `async` fields from settings.json hook config Auto-diary hook architecture: PostToolUse on Bash with if filter for git commit. Runs async so commits are not blocked. Reads prompts from .session/prompts.jsonl, insights from .session/insights.jsonl, calculates tokens from session JSONL. Only clears prompts per commit, insights persist across commits within a session. Files: .claude/hooks/log-insight.sh,.claude/hooks/log-plan.sh,.claude/hooks/log-prompt.sh,.claude/hooks/log-task.sh,.claude/hooks/post-commit-diary.sh,.claude/settings.json,session-log.md

PostToolUse hook fires after `gh pr create`, reads .session/ JSONL buffers (prompts, insights, plans), and appends them as a "Session Logs" section to the PR body. Idempotent — skips if logs already present. Auto-diary hook architecture: PostToolUse on Bash with if filter for git commit. Runs async so commits are not blocked. Reads prompts from .session/prompts.jsonl, insights from .session/insights.jsonl, calculates tokens from session JSONL. Only clears prompts per commit, insights persist across commits within a session. Files: .claude/hooks/post-pr-logs.sh,.claude/settings.json

With 268 posts, EmDash's SEO + byline hydration at limit=100 exceeds Cloudflare Workers CPU time. Cap all content list requests above 50 at the Worker level before they reach the Astro/EmDash handler. Also removes the unused Astro middleware from earlier attempts. Auto-diary hook architecture: PostToolUse on Bash with if filter for git commit. Runs async so commits are not blocked. Reads prompts from .session/prompts.jsonl, insights from .session/insights.jsonl, calculates tokens from session JSONL. Only clears prompts per commit, insights persist across commits within a session. Files: src/worker.ts

Auto-diary hook architecture: PostToolUse on Bash with if filter for git commit. Runs async so commits are not blocked. Reads prompts from .session/prompts.jsonl, insights from .session/insights.jsonl, calculates tokens from session JSONL. Only clears prompts per commit, insights persist across commits within a session. Files: src/worker.ts

Auto-diary hook architecture: PostToolUse on Bash with if filter for git commit. Runs async so commits are not blocked. Reads prompts from .session/prompts.jsonl, insights from .session/insights.jsonl, calculates tokens from session JSONL. Only clears prompts per commit, insights persist across commits within a session. Files: src/worker.ts

Auto-diary hook architecture: PostToolUse on Bash with if filter for git commit. Runs async so commits are not blocked. Reads prompts from .session/prompts.jsonl, insights from .session/insights.jsonl, calculates tokens from session JSONL. Only clears prompts per commit, insights persist across commits within a session. Files: src/worker.ts

With 268 posts, EmDash's SEO + byline hydration at limit=100 exceeds Cloudflare Workers CPU time. The Worker now caps content list requests to limit=50 before passing to the Astro/EmDash handler. Root cause of earlier failed deploys: wrangler deploy was shipping a Auto-diary hook architecture: PostToolUse on Bash with if filter for git commit. Runs async so commits are not blocked. Reads prompts from .session/prompts.jsonl, insights from .session/insights.jsonl, calculates tokens from session JSONL. Only clears prompts per commit, insights persist across commits within a session. Files: src/middleware.ts,src/worker.ts

Worktree isolation: each worktree gets its own .session/ directory, session data is isolated per branch | Test insight: session tracker validation run | Token auto-calc in pre-push reads ALL session JSONLs for today from shared project dir, not per-worktree. Numbers represent cumulative daily cost. | Diary pipeline evolved from agent-dispatch-per-commit to direct SQL write in shell. priori-incantatem agent now only used as fallback via pre-push gate. | Clean insight test — no trailing newline | Verify: clean JSON without trailing newline | jq -Rs vs jq -n --arg: Rs reads stdin where echo adds \n; --arg takes shell arg directly with no trailing newline, also safer with special chars | post-commit-diary now has 3-layer guard ordered by cost: (1) command guard via stdin JSON, (2) attribution grep, (3) dedup via last-diary-commit file. Old version ran git log on every Bash call. Files: src/data/site-routes.json

Wrangler login, created D1 database (pensieve-db), KV namespace (SESSION), R2 bucket (pensieve-media). Wired bindings into wrangler.jsonc.

Connected Chrome browser via MCP. Downloaded Messenger conversation photos. Built a plant gallery with dark theme, image grid, lightbox, and cart panel.

Migrated plant gallery to Pensieve. Added quantity controls to cart. Deployed to huuloc.com/plant-gallery/. 68 files uploaded.

Created the repository and initial project structure for the Pensieve blog.

Scaffolded the Pensieve blog with per-category themes and deployed to huuloc.com/pensieve. Built a Facebook narrative extractor and imported 14 real posts into the seed file and remote D1.

Redesigned the Pensieve home page with a themed hero section, category pills for navigation, and grouped content sections.

Created session-log.md and session-logger subagent. Set up UserPromptSubmit hook for automatic prompt capture. Dual system: hook for capture, agent for summaries.

Created plugins/plugin-resend/ — provides email delivery via Resend API and sends notifications on new comments. Fixed capability requirements and Block Kit admin settings page.

Replaced VI/EN links with a dropdown menu matching Categories style. Language stored via cookie. Home link changed to root domain.

Configured EmDash GitHub OAuth. User created a GitHub OAuth App, set secrets via wrangler. GitHub login now available at the admin UI.

Created .github/workflows/deploy.yml with Node 22, bun, and wrangler-action. Automated deployment on push to main.

Created moderate-comments subagent at .claude/agents/moderate-comments.md. Verified EmDash comments work out of the box. Added moderation notice below comment form.

Enabled Email Routing for huuloc.com. Created me@huuloc.com forwarding to loctruongh@gmail.com via wrangler CLI.

Expanded the corpus to 155 posts with Vietnamese and English translations. Established the Linear-inspired visual system and built content pipeline v2 with date handling, deduplication, R2 media uploads, and incremental sync.

Added an Elsewhere section with social links on the root landing page and a moderation notice below the comment form.

Added Vietnamese support to Priori Incantatem (dates, type labels: Phép Thuật, Lời Tiên Tri, Reparo, Hiện Hình). Created room-of-requirement subagent for updating tech/plugins/agents on the build page.

Built about page at /Trương with origin story bio, profile links (GitHub, LinkedIn, Facebook, email), and live GitHub infographic fetched from GitHub API at request time — stats, activity chart, recent repos grid, and activity feed timeline.

Planned plugin-pensieve-engage: email subscriptions with welcome emails, newsletter dispatch on publish with per-post opt-out, click/open tracking in emails, and reading engagement analytics (scroll depth, reading time, section focus via IntersectionObserver). 5-phase implementation. Single trusted plugin using page:fragments for tracker injection.

Created site-routes.json manifest as single source of truth for SEO. Sitemap reads static pages from manifest + queries EmDash for dynamic content. Added llms.txt and .well-known/ai-plugin.json for AI agent discovery. Updated CLAUDE.md and agents to enforce manifest updates.

Added Room of Requirement and Trương links to the huuloc.com static landing page. Fixed nav label from 'Build' to 'Room of Requirement'. Updated seed menu to point About to /Trương.

Built the build section (Room of Requirement) with architecture overview, tech stack, and subagents. Created Priori Incantatem — a timeline diary recording every Claude Code session with model versions.

Removed presumptive bipolar disorder mention from the about page. That's the author's story to tell.

Added blocking rule to CLAUDE.md: priori-incantatem agent must run before every git push. No more forgotten diary entries.

Added prominent authorship banners on every post. 'Penned by Quill' (green) for original human-written content. 'Conjured Translation' (accent) for Claude-translated posts. Bilingual EN/VI. Each banner explicitly states whether content is human or AI-translated.

Added GITHUB_TOKEN support to Trương page for higher GitHub API rate limits (60 → 5,000 req/hr). Token set via wrangler secret.

Created Geminio (Doubling Charm) subagent — i18n police that ensures all user-facing text has both English and Vietnamese versions. Ran full audit: found ~100 hardcoded strings across 12 files.

Fixed data-switch-lang JS handler that reloaded current page instead of navigating to sibling URL. Renamed toggle: 'Read the Original Manuscript' / 'Đọc Bản Gốc' for originals, 'Read the Conjured Translation' / 'Đọc Bản Dịch Phép Thuật' for translations.

Corrected coding eras from GitHub commit history: Solo 2006-Jun 2022 (1,755 commits), GitHub Copilot Jun 2022-Jan 2026 (8,106 commits, public preview), Claude Code Jan 29 2026-now (232 commits). First Claude Code commit: Jan 29, 2026.

Removed build:relocate-assets script that moved /_astro/* to /pensieve/_astro/. With base: '/', assets must stay at root or CSS/JS links break.

Replaced per-visitor GitHub API calls with Cloudflare edge caching (1hr TTL via cf fetch options). Renamed session-logger agent to priori-incantatem.

Added content signals meta tags (contentsignals.org). Added 'What Claude Code Does' section to Room of Requirement. Added coding eras timeline to Trương page: Solo (2006-2021), GitHub Copilot (2021-2026), Claude Code (Apr 2026-now, zero human code).

Merged the separate authorship banner and translation toggle into a single unified block.

Refactored Astro base from /pensieve to /. Moved blog pages to src/pages/pensieve/ to preserve URLs. Converted static landing page to Astro page. Simplified worker — removed URL rewriting hack. Root-level pages (Room of Requirement, Trương) now native Astro routes.

Migrated build diary from JSON file to EmDash diary collection with token/cost tracking fields (input_tokens, output_tokens, cache_read, cost). 28 entries seeded to production D1. Priori Incantatem page now queries EmDash.

Full bilingual treatment across all 12 page files. Base layout, Room of Requirement, Truong, blog home, memories, search, 404, categories, category, tag pages. ~100 hardcoded strings translated to Vietnamese.

Built subscription Phase 5 admin dashboard (sends page, analytics page, subscriber widget). Created OG image generator (/og/[slug].svg). Created Floo Network cross-posting agent. Fixed production DB (menu label, notify_subscribers column).

Added Owl Post (HP-themed newsletter) subscribe forms to post pages and blog footer. Bilingual EN/VI. Success message: Owl dispatched!

Created Nixie — a floating wand icon in the bottom-right corner of every page. Hover reveals Conjured by CC with accent glow. Links to Room of Requirement. Bilingual.

Built Lumos — HP-themed guest like button with golden glow animation. Spam prevention via cookie + IP hash combo. Plugin routes: lumos/cast and lumos/count.

Added Total Lumos stat and per-post Lumos column to the analytics dashboard alongside pageviews, scroll depth, and reading time.

Fixed 5 hardcoded English strings in client-side JS for Owl Post subscribe forms (error/success messages). Both post page and footer handlers now bilingual.

Added SVG icons throughout the site: Pensieve bowl in breadcrumbs, sprout icon as the new site-level favicon and home eyebrow, and distinctive icons for each Space card. Scoped the bowl to /pensieve/ and the sprout to site-wide identity.

Added authenticated GitHub API support to the About page so private repo stats are included in counts. Redacted private repo names in the activity feed. Added an arched door SVG icon to the Room of Requirement breadcrumb and refactored the breadcrumb icon system to be generic.

Planned cross-posting feature: EmDash plugin with content:afterPublish hook calling Facebook Graph API, LinkedIn Share API, Threads API. Requires OAuth tokens per platform — manual setup needed.

Renamed post sidebar labels: Author to Penned By, Reading time to Spell Duration, Tags to Enchantments. Bilingual EN/VI.

Subscription Phase 3 (click/open tracking with redirect URLs and tracking pixel) and Phase 4 (reading engagement analytics with page:fragments tracker injection, scroll depth, heartbeats, beacon storage). Updated Room of Requirement with new agents and plugins.

Added reading progress bar on posts (3px accent bar tracking article scroll). Newsletter dispatch on publish (Phase 2). HP search placeholder (Accio memories). Added HP origin mention to Trương page.

Added breadcrumbs to all pages. HP-themed nav (Sorting Hat, Enchantments, Memories, Accio search). Built subscription plugin Phase 1 with subscribe/unsubscribe/welcome email/admin.

Added collapsible day groups to the Priori Incantatem build diary page using native HTML details/summary elements. Latest day stays expanded, older days collapsed by default. Includes entry count badges and chevron indicators, pure HTML/CSS with no JavaScript.

Rewrote micro-copy across 11 site surfaces with Harry Potter theming: 404 page, search, theme toggle (Lumos/Nox), empty states, footer headings (Marauder's Map / Owl Directory), home sections (The Grounds / Beyond the Gates), post page, comments, RSS feeds, and Room of Requirement subagents.

Renamed /posts to /memories and /truong to /trương with proper capitalization. Built a custom sitemap with correct URLs and committed Claude Code memories.

Fixed three issues on the Priori Incantatem diary page: added date display to individual entry cards, changed sort order from alphabetical title to published_at timestamp for proper chronological ordering, and fixed formatModel to preserve dots in version numbers (displaying "Opus 4.6" instead of "Opus 4 6"). Also backfilled 6 missing diary entries from git history.

Built a Patronus creature system with 6 HP-themed silhouettes that roam the site as silvery ghosts with layered flip/bob/gallop animations and ghost trails. Added an Expecto Patronum easter egg on the footer name, fixed Lumos sparkle positioning, and rewrote all 78 git commits for proper Claude Code attribution.

Ultraplan session to address the problem of Claude Code forgetting to resolve pending tasks. Deep research into options: Linear, GitHub Issues, Todoist, plain-file approaches. Decision pending — plan produced at https://claude.ai/code/session_01959w2yF9vBNnezfhbHocs7 for review. Core problem: tasks created during sessions get abandoned when context shifts. Need a system that persists across sessions and integrates with the Claude Code workflow.

Built automated diary system with three hooks: UserPromptSubmit logs prompts to .session/prompts.jsonl, log-insight.sh captures insights, PreToolUse on git push blocks until diary is written. Also: collapsible diary days with details/summary, date+time on entries with Vietnam timezone, fixed Opus 4.6 version display, fixed Lumos like button (unwrap EmDash data envelope), added web app manifest.json, and plan capture as first-class diary entries.

11-task plan to implement all 6 task tracking systems with HP theming. Systems: Remembrall (Todoist), Marauder's Map (GitHub Issues), O.W.L.s (Linear), Room of Requirement (Notion), Pensieve Strands (session-tracker), Marauder's Parchment (TASKS.md). Architecture: MCP servers for external writes, D1 tasks table for unified reads, log-task.sh hook for mirroring. 3 showcase surfaces: RoR dashboard cards, Task Observatory dedicated page, diary integration via pre-push hook. Implementation order: D1 schema -> log-task.sh -> MCP config -> TASKS.md -> CLAUDE.md routing rules -> Task Observatory page -> RoR dashboard section -> site routes -> diary hook extension -> Geminio pass -> E2E verification. Key decisions: raw D1 table (not EmDash collection) because tasks have CRUD lifecycle; MCP for Todoist/Linear/Notion, native gh CLI for GitHub Issues; session-tracker runs automatically on SessionEnd; routing is CLAUDE.md rules (not automated hook) to match existing log-insight/log-plan pattern.

Research plan from ultraplan session. SOUL.md is OpenClaw's foundational concept — a plain markdown file defining who an AI agent is. The celestial body metaphor: it's the gravitational center around which all behavior orbits. Key findings: (1) Philosophy not config — captures values, opinions, tone, boundaries, not procedures. (2) Four core sections: Core Truths, Boundaries, Vibe, Continuity. (3) Seven-file architecture: SOUL.md (foundation) → IDENTITY.md → AGENTS.md → USER.md → TOOLS.md → HEARTBEAT.md → MEMORY.md. (4) Bootstrap problem solved via three-phase persisted interview. (5) Design principles: files not databases, philosophy not configuration, specificity over safety, living document, portable, tripartite separation.

Built /room-of-requirement/task-observatory — full page querying D1 tasks table directly. Shows 6 HP system cards (Remembrall, Marauder's Map, O.W.L.s, Room of Requirement, Pensieve Strands, Marauder's Parchment) with per-system counts, status filtering, and task list with color-coded system badges. Fixed Astro v6 breaking change: replaced Astro.locals.runtime.env with import { env } from cloudflare:workers.

Created TASKS.md as the in-repo task checklist with HP theming ("I solemnly swear that I am up to no good"). Added Task Routing section to CLAUDE.md and AGENTS.md mapping 6 task types to 6 HP systems with commands for each.

Created D1 tasks table with indexes (remote + local) for mirroring task data from 6 external tracking systems. Built log-task.sh hook that inserts mirror rows into D1 and appends to .session/tasks.jsonl for diary integration. Each HP system slug maps automatically to its external tool name.

Fetched ultraplan research on OpenClaw's SOUL.md concept via Chrome browser. Created SOUL.md for the pensieve agent with four sections: Core Truths (have opinions, be resourceful, earn trust, move fast), Boundaries (writing is sacred, bilingual first-class, Claude builds / Loc architects), Vibe (concise, HP theming is joyful not precious), Continuity (every session leaves traces, you have a Pensieve). Completes the tripartite separation: SOUL.md = who, CLAUDE.md = what, memory = context.

Configured Todoist, Linear, and Notion MCP servers. Used Chrome browser automation to create Linear workspace (Pensieve) and O.W.L.s project — fought icon pickers and discard dialogs, solved with JavaScript DOM click. Created Remembrall project in Todoist via MCP API. Created Room of Requirement database in Notion via MCP with Type/Status/Tags columns. Installed claude-session-tracker for Pensieve Strands.

Added Task Observatory section to Room of Requirement index page with 6 HP system cards showing live task counts from D1. Each card displays system name, external tool, routing description, and open/done counts. Reuses existing tech-grid CSS with colored left borders.

Extended pre-push-diary.sh to read .session/tasks.jsonl buffer. Modified log-task.sh to also append to session buffer. Updated cleanup instructions to include task log. Diary entries now capture which HP systems were used during the session.

Discovered Claude Code stores per-message token usage in session JSONL files at ~/.claude/projects/. Rewrote pre-push-diary.sh to auto-scan today's session JSONLs, sum input_tokens/output_tokens/cache_read_input_tokens via python3, and calculate cost in cents. Eliminates manual /stats checking — the single biggest gap in diary data quality (53/55 entries had zero tokens before this fix).

Split single combined token count into input/output/cache breakdown on Daily Prophet page. Fixed double dollar sign in cost display — was using template literal inside JSX expression causing $$ prefix. Added cacheRead to DiaryEntry interface and Release aggregation.

Built /room-of-requirement/daily-prophet — auto-generates release notes from diary entries. Each date = one chapter. Chapter names come from the day's most significant diary entry title (infinite by nature). Versioned by Hogwarts academic year (Sept 1 reset). Rejected fixed HP chapter lists in favor of work-derived naming. Shows entry types, token breakdown (input/output/cache), and cost per release.

Git commit stdout doesn't include the message body/trailers. Read from git log instead. Files: .claude/hooks/post-commit-diary.sh

Insights should accumulate and appear in every diary entry. Only prompts get consumed per commit since they're sequential. Files: .claude/hooks/post-commit-diary.sh

Auto-diary hook architecture: PostToolUse on Bash with if filter for git commit. Runs async so commits are not blocked. Reads prompts from .session/prompts.jsonl, insights from .session/insights.jsonl, calculates tokens from session JSONL. Only clears prompts per commit, insights persist across commits within a session. Files: .claude/hooks/post-commit-diary.sh

Auto-diary hook architecture: PostToolUse on Bash with if filter for git commit. Runs async so commits are not blocked. Reads prompts from .session/prompts.jsonl, insights from .session/insights.jsonl, calculates tokens from session JSONL. Only clears prompts per commit, insights persist across commits within a session. Files: .claude/hooks/post-commit-diary.sh

The EmDash v0.1.0 admin UI requests posts with limit=100, which triggers a CONTENT_LIST_ERROR (500) on the server. All other limit values work fine. This middleware rewrites limit=100 to limit=99 as a workaround until EmDash is upgraded to v0.3.0. Auto-diary hook architecture: PostToolUse on Bash with if filter for git commit. Runs async so commits are not blocked. Reads prompts from .session/prompts.jsonl, insights from .session/insights.jsonl, calculates tokens from session JSONL. Only clears prompts per commit, insights persist across commits within a session. Files: src/middleware.ts

The Astro middleware doesn't run for /_emdash/api/ routes since EmDash handles them internally. Moving the limit=100→99 rewrite to the Worker level where it intercepts all requests before Astro. Auto-diary hook architecture: PostToolUse on Bash with if filter for git commit. Runs async so commits are not blocked. Reads prompts from .session/prompts.jsonl, insights from .session/insights.jsonl, calculates tokens from session JSONL. Only clears prompts per commit, insights persist across commits within a session. Files: src/worker.ts

Auto-diary hook architecture: PostToolUse on Bash with if filter for git commit. Runs async so commits are not blocked. Reads prompts from .session/prompts.jsonl, insights from .session/insights.jsonl, calculates tokens from session JSONL. Only clears prompts per commit, insights persist across commits within a session. Files: src/worker.ts

The pre-rewrite approach didn't work because EmDash may read the limit from internal routing state rather than the rewritten URL. This approach lets the original request fail, then retries with limit=99 which avoids the bug. Auto-diary hook architecture: PostToolUse on Bash with if filter for git commit. Runs async so commits are not blocked. Reads prompts from .session/prompts.jsonl, insights from .session/insights.jsonl, calculates tokens from session JSONL. Only clears prompts per commit, insights persist across commits within a session. Files: src/worker.ts

Auto-diary hook architecture: PostToolUse on Bash with if filter for git commit. Runs async so commits are not blocked. Reads prompts from .session/prompts.jsonl, insights from .session/insights.jsonl, calculates tokens from session JSONL. Only clears prompts per commit, insights persist across commits within a session. Files: src/worker.ts

Server-side URL rewrite doesn't propagate auth cookies correctly. Instead, inject a <script> into the admin HTML that patches window.fetch to rewrite limit=100 → limit=99 client-side. Auto-diary hook architecture: PostToolUse on Bash with if filter for git commit. Runs async so commits are not blocked. Reads prompts from .session/prompts.jsonl, insights from .session/insights.jsonl, calculates tokens from session JSONL. Only clears prompts per commit, insights persist across commits within a session. Files: src/worker.ts