How this site is built.

Tech Stack

Technology stack.

• Astro 6 Web framework (zero JS, static generation)
• Cloudflare Workers Hosting (Static Assets + SSR)
• Tailwind CSS 4 Styling (Vite plugin)
• Workers AI Llama 3.3 70B wiki chat (free tier)
• TypeScript 6 Type system (astro check)

Cloudflare

System configuration on Cloudflare.

The site is a single Cloudflare Worker serving Astro build artifacts. Here's what we use, how, and why.

1. Workers (Static Assets)

   How: Astro build generates static assets, served as a single Worker via the @astrojs/cloudflare adapter.

   Why: Static hosting alone has alternatives, but Workers lets us add AI chat (/api/chat) and bindings (KV, Vectorize) within the same runtime — so we chose Workers from the start.

2. Custom Domain (apex)

   How: The apex domain jamjet-iqbal.com is assigned directly to the Worker via wrangler.jsonc routes. Cloudflare auto-creates DNS records and TLS certificates.

   Why: Maintain apex-level publishing without manual DNS or certificate management.

3. wrangler (config as code)

   How: name, compatibility_date, nodejs_compat, and routes are consolidated in wrangler.jsonc and merged with adapter-generated deploy config for wrangler deploy.

   Why: Keep deploy configuration in git-managed code, enabling identical procedures from local and CI.

4. GitHub Actions auto-deploy

   How: Weekly/monthly observation workflows commit collected notes, push, then run wrangler deploy. API tokens are scoped to Workers edit + target zone only.

   Why: Keep wiki updates and production publishing running regardless of PC state. Manual changes still deploy via npm run deploy.

5. Free tier first + binding reservations

   How: Default path stays within Cloudflare free tier (100K requests/day, etc.). AI / KV / R2 / D1 / Vectorize are unimplemented, designed as plug-in points in wrangler.jsonc for roadmap phases P3–P7.

   Why: Low cost is the top priority. Paid/metered features are "slot ready, default off" for staged introduction.

AI Chat

Wiki chat (/chat).

A chat feature where Workers AI answers questions using published note excerpts as evidence. Responds via SSE streaming.

Model

@cf/meta/llama-3.3-70b-instruct-fp8-fast

Search

Bigram match scoring (zero-dependency Japanese support)

Context

Top 4 note excerpts injected into system prompt

Cost

Operated within Workers AI free tier (10,000 Neurons/day)

Model Tiers

Three-tier model hierarchy.

Models are assigned by role nature, balancing reasoning depth with token cost.

1. claude-opus-4-6 [1m]

   Core of debate. Handles lead argumentation, mutual counterarguments, auditing, attacking shared premises, and managing academic citations — roles requiring deep reasoning and 1M-context long-document comprehension.

   4 agents

2. claude-sonnet-5

   Collection, review, and utility. Read-only structured extraction and clearly scoped inspections achieve sufficient quality with the latest Sonnet, optimizing cost.

   5 agents

3. claude-fable-5

   Too token-intensive for standing assignments. Reserved for especially deep theoretical rounds (e.g., SO theory arbitration), specified via launch-time override on a case-by-case basis.

   0 agents (no standing assignment)

Agents

Nine subagents.

Organized into four groups spanning collection through critique. Each agent switches roles at launch via dispatch headers (profile / persona / lens / axis / canon / mode), covering the perspectives of the former 27 agents with just 9. Badges indicate each agent's model.

Pipeline

Four stages from collection to publication.

1. 01 — Sources

   Primary source layer (read-only)

   Collection corpora and citation ledgers in source/review/ serve as the primary source layer. raw/ is a reserved immutable area (write-blocked by hook).

2. 02 — Collect

   Collection & corpus building

   Collection agents follow a protocol prioritizing primary sources, tracking citations, and filtering position-talk to produce citation-ledgered corpora.

3. 03 — Debate

   Debate & critique (R1–R5)

   Debate leads, auditors, provocateurs, and academic critics exchange arguments and counterarguments across multiple rounds. The deliverable is a cited opposition table, not a smoothed consensus.

4. 04 — Publish

   Published as wiki/note

   Integrated results are written back as notes with full references, updating the index and log for publication on this site.

Guardrails

Safety measures.

• raw/ immutability (PreToolUse hook blocks writes)
• Unpublished research control (pages without draft / so_theory flags are not published)
• README and asset list sync (PostToolUse hook prompts updates)
• Debate reproducibility (debate leads stay off the web, arguing only from pre-fixed evidence)
• Debate cost logging (debates run one round, recording per-agent consumption in the debate record)
• CI gate (wiki health check, type check, and build run on every push / PR)
• Compact recovery (decision structures saved to state file before compaction; PreToolUse gate forces recovery after)
• Early context alerts (PostToolUse / UserPromptSubmit suggest /compact-prep above 60% usage)