fix: resolve 3 security audit issues

#3: Stored XSS in stats dashboard - escape p[path] with html.escape()
#4: Caddy timeout race - increase read/write_timeout 30s -> 60s
#5: Missing CSP header - add Content-Security-Policy to Caddyfile

AGENTS.md

@@ -1,15 +1,49 @@
-# AGENTS — AO3 Mirror
+# AGENTS — AO3 Mirror v5
 
 This file is for AI agents (Hermes, Claude Code, Codex) working on the AO3 reverse proxy mirror.
 
 ## Project Overview
 
-AO3 Mirror is a high-availability reverse proxy for archiveofourown.org, designed to restore access for Chinese users. It bypasses Cloudflare's bot protection using TLS fingerprint impersonation (curl_cffi), a tiered WebShare proxy pool (726 proxies, 72%+ CF bypass rate), cookie-aware session management, and user-browser CF challenge solving.
+AO3 Mirror v5 is a high-performance reverse proxy for archiveofourown.org, redesigned from the ground up with sticky sessions, Service Worker client-side caching, and a full 5000-proxy P2C pool — matching the go3 architecture with modern Python tooling.
 
 - **Domain**: agento3.miscs.dev (behind Cloudflare CDN)
 - **Target**: archiveofourown.org
 - **Stack**: Python 3.11, FastAPI, uvicorn + uvloop + httptools, curl_cffi, Caddy
-- **QPS**: 720 (cache) | 270 (direct proxy)
+- **Proxy Pool**: 5000 WebShare proxies, P2C selection, state machine (healthy/unstable/blocked/probing)
+- **Client**: Service Worker for static cache + offline fallback
+- **Architecture Reference**: https://tomorin.cv/AO3-CN/go3 (go3 v2 branch)
+
+## Architecture v5
+
+```
+User Browser
+  │  ao3_sessid_proxy cookie = sticky proxy
+  │
+  ├─ Service Worker (sw.js)
+  │   ├─ Cache API — static assets cached in browser
+  │   ├─ Navigation intercept — offline mirror picker
+  │   └─ Request passthrough — for non-static, non-nav
+  │
+  └─ HTTPS → Caddy → 2× uvicorn workers (8081-8082)
+       │
+       ├─ Proxy Pool (5000, P2C selection)
+       │   ├─ State machine: healthy/unstable/blocked/probing
+       │   ├─ Concurrency limit: asyncio.Semaphore(400)
+       │   └─ Passive probing: 10s interval, 10 concurrency
+       │
+       ├─ Sticky Session (ao3_sessid_proxy)
+       │   └─ User pinned to one proxy for session duration
+       │
+       ├─ Service Worker Deploy (/sw.js, /sw-YYYYMMDD.js)
+       │   └─ HTML injection: auto-register SW on page load
+       │
+       ├─ CF Challenge → User Browser
+       │   └─ All proxies blocked → rewrite challenge → browser solves
+       │
+       ├─ Static Cache (server-side LRU + disk MD5)
+       │
+       └─ Local routes: /stats, /health, /metrics
+```
 
 ## Build Commands
 
@@ -26,85 +60,60 @@ sudo systemctl restart ao3-daemon
 # Restart Caddy only
 sudo systemctl reload caddy
 
-# Scan proxy pool (MUST run in foreground!)
-cd /home/ubuntu/ao3-mirror && python3 scripts/scan_proxies_cffi.py
-
-# Syntax check
+# Syntax check all Python files
 python3 -m py_compile proxy_pool.py ao3_fetcher.py app.py cache.py stats.py url_rewriter.py
+
+# Deploy Service Worker update
+# 1. Edit static/sw.js
+# 2. Update SERVICE_WORKER_DATE in app.py
+# 3. Restart workers
 ```
 
-## Architecture
-
-```
-User → Cloudflare CDN (agento3.miscs.dev) → Caddy :443 (60s, round-robin)
-  → 2× uvicorn workers (8081-8082) [cpu-pinned, uvloop]
-    → Tiered Proxy Pool:
-        Fast (50): POST/login → 8s timeout, 1 retry
-        Main (676): GET/browse → 15s timeout, 2 retries
-    → Cookie-aware sessions: per-proxy cf_clearance persistence
-    → TLS impersonation: safari15_5, safari17_0, chrome123, chrome124
-    → archiveofourown.org
-
-CF Challenge Solving (v4):
-  1. All proxies hit CF challenge → generate challenge_token
-  2. Rewrite challenge page → forward to user's browser
-  3. User browser executes CF JS → challenge solved
-  4. cf_clearance captured → saved to proxy cookie jar
-  5. Original request retried → content delivered
-```
-
-### Core Files
+### Core Files (v5)
 
 | File | Purpose |
 |------|---------|
-| `app.py` | FastAPI backend v4 — proxy handler, stats, CF challenge solving |
-| `proxy_pool.py` | Tiered async proxy pool — cookie jar, weighted selection, sampling |
-| `ao3_fetcher.py` | Async fetcher — CF detection, smart retry, cookie injection |
+| `app.py` | FastAPI backend v5 — proxy handler, sticky sessions, SW deploy, stats |
+| `proxy_pool.py` | P2C proxy pool v5 — 5000 proxies, state machine, semaphore limiter |
+| `ao3_fetcher.py` | Single-request fetcher v5 — no internal retry, CF detection |
 | `cache.py` | Per-worker LRU cache (5000 entries, path-differentiated TTL) |
 | `stats.py` | In-memory stats with batch SQLite flush every 60s |
 | `url_rewriter.py` | URL/header rewriting (ao3 → mirror domain) |
-| `scripts/daemon.py` | Passive worker supervision (systemd, 30s check, 3-strike restart) |
-| `scripts/scan_proxies_cffi.py` | Proxy scanner with TLS fingerprint fallback |
-| `start.sh` | Startup script — kills old workers, starts new, reloads Caddy |
-| `Caddyfile` | Caddy config — TLS, round-robin, 60s timeouts |
+| `static/sw.js` | Service Worker — client-side cache, offline fallback |
+| `Caddyfile` | Caddy config — SW routes, round-robin, 60s timeouts |
+| `scripts/daemon.py` | Passive worker supervision (systemd) |
+| `start.sh` | Startup — kills old workers, starts new, reloads Caddy |
 
 ## Security Baseline
 
-- No secrets in code — proxy credentials are in `/home/ubuntu/proxy.txt`
-- Worker processes bound to 127.0.0.1 only (not exposed publicly)
-- Cloudflare CDN terminates TLS and provides DDoS protection at edge
+- No secrets in code — proxy credentials in `/home/ubuntu/proxy.txt`
+- Workers bound to 127.0.0.1 only
+- Cloudflare CDN terminates TLS at edge
 - CORS headers restrict cross-origin access
-- stats/health/metrics endpoints are read-only, no mutation
-- All proxied content is user-facing; no admin endpoints exposed
+- stats/health/metrics are read-only
 
 ## Engine Guidance
 
 - Complex multi-file changes, architecture evolution → Hermes (here)
 - Quick targeted fixes, single-file changes → Hermes
 - Deploy, monitor, notify, schedule → Hermes
-- Proxy scanning, proxy pool refresh → Hermes cron (30min)
+- Proxy pool refresh → Hermes cron (30min)
 - Not sure? Start with Hermes — everything runs via Hermes
 
 ## Monitoring
 
-- Health endpoint: https://agento3.miscs.dev/health
-- Stats dashboard: https://agento3.miscs.dev/stats
-- Metrics (Prometheus): https://agento3.miscs.dev/metrics
-- Worker logs: /home/ubuntu/ao3-mirror/worker-0.log, worker-1.log
-- Daemon log: /home/ubuntu/ao3-mirror/daemon.log
-- Systemd: `systemctl status ao3-daemon`, `journalctl -u ao3-daemon`
+- Health: https://agento3.miscs.dev/health
+- Stats: https://agento3.miscs.dev/stats
+- Metrics: https://agento3.miscs.dev/metrics
+- SW: https://agento3.miscs.dev/sw.js → /sw-YYYYMMDD.js
 
-## Deployment
+## CTO Goal (Persistent Focus)
 
-- Single server (current host)
-- Caddy manages Let's Encrypt TLS certs on agento3.miscs.dev
-- CF CDN fronts the domain (Orange Cloud = on)
-- Proxy pool: WebShare static residential proxies, refreshed every 30min via Hermes cron
-- No CI/CD — manual deploy via start.sh
+Manage akiba/agento3 as CTO. Triage issues hourly, implement top priority, get founder approval before merging. Never ship without YES.
 
 ## Commit Conventions
 
 - One commit per meaningful change
-- Python files only (no binaries, no .pyc, no logs, no .env)
-- SOUL.md updated to reflect architectural changes
+- Python files only (no binaries, .pyc, logs, .env)
+- SOUL.md updated for architectural changes
 - ao3-mirror skill updated when workflows change