AWS Flywheel

Autonomous multi-branch release engine — 3 bug + 1 UX fix per branch, every hour, self-learning, self-healing.

21 branches 24h cooldown 1 release/hour 253+ releases

Full source: github.com/unclehowell/datro/tree/cnei/flywheel

Architecture

A cron job (0 * * * *) on an AWS EC2 instance selects one branch via round-robin rotation, runs 4 automated fix passes against its working tree, commits and tags, and publishes a GitHub release with structured notes. An executive-coach meta-review runs separately at 00:00 UTC daily on a command machine, reviewing AWS output and self-improving the flywheel code.

Components

File	Role
multi-branch-release.sh	Bash orchestrator — cron entrypoint, branch selection, pass dispatch, git/release
intelligence.py	Python intelligence pipeline — builds LLM prompts, applies fixes, manages profiles, self-learns
gather_review_data.py	Log parser — extracts fix source ratios (AI vs Pool vs Fallback) from AWS logs for meta-review
meta-review.sh	Executive coach — fetches AWS state, runs intelligence.py --type meta, syncs profile learnings
release-state.json	State file — rotation index, cooldowns, release counters, fix/ux rotation pointers
profiles.json	Per-branch profiles — LLM rotation lists, learned patterns, skill libraries, branch knowledge
agent/branches/*.md	Branch context files — soul/manifest/memory per branch, category, cornerstone mission

21 Branches

Each branch deploys independently via Cloudflare Pages at {branch}.datro.xyz:

Branch	Category	Domain	Releases
financecheque	E-commerce	financecheque.uk	150
ccan	Community	ccan.datro.xyz	16
althea	Community	althea.datro.xyz	7
bpvsbuckler	E-commerce	bpvsbuckler.datro.xyz	5
carfinancecheque	E-commerce	carfinancecheque.datro.xyz	5
ceo	Platform	ceo.datro.xyz	7
dash	Platform	dash.datro.xyz	5
datro	Meta	datro.xyz	6
dcc	Community	dcc.datro.xyz	6
gui	Platform	gui.datro.xyz	6
hbnb	E-commerce	hbnb.datro.xyz	6
library	Documentation	library.datro.xyz	6
llmwiki	Documentation	llmwiki.datro.xyz	5
subrepos	Platform	subrepos.datro.xyz	4
ui	Platform	ui.datro.xyz	5
wave	Platform	wave.datro.xyz	5
wayback	Archive	wayback.datro.xyz	4
whitepaper	Documentation	whitepaper.datro.xyz	3
archives	Archive	archives.datro.xyz	5
cnei	Meta	cnei.datro.xyz	—
greathousefarm	Advocacy	greathousefarm.datro.xyz	1

All content lives under static/{branchname}/ in the repo. Cloudflare Pages root_dir is set per project to that prefix.

Pass Strategy — Fix Diversity Enforcement

The original loop (for pass in 1 2 3) always let the AI win, resulting in 105/105 console.log removals. The strategy was redesigned into individually-controlled passes:

Pass	Source	Purpose
1 (Bug)	AI → Rotating Pool → Fallback	AI finds professional improvements (SEO, structured data, accessibility). If AI fails, pool rotates through 32 fix functions. Last resort: guaranteed fallback.
2 (Bug)	Rotating Pool only	Guarantees pool diversity — skips AI entirely so the rotating pool always gets a turn.
3 (Bug)	Guaranteed Fallback	Always runs — catches triple-blank-lines via rg -U multiline matching.
4 (UX)	AI → Rotating Pool → Fallback	User-experience improvements (viewport, tap targets, hover styles, skip links, print styles, etc.).

Rotating Fix Pool — 32 Bug Functions

Category	Fix Functions
SEO	fix_meta_description, fix_canonical_url, fix_og_tags, fix_twitter_card, seo_structured_data, seo_keywords_meta
Accessibility	fix_lang_attribute, fix_aria_label, fix_label_for, fix_button_type, fix_heading_hierarchy
HTML Standards	fix_doctype, fix_charset_meta, fix_viewport_meta, fix_self_closing, fix_br_syntax, fix_http_equiv, fix_form_charset, fix_bom
Security/Performance	fix_link_noopener, fix_script_defer, fix_img_dimensions, fix_inline_handlers
Code Quality	fix_console_log, fix_commented_code, fix_trailing_whitespace, fix_blank_lines, fix_tabs_vs_spaces, fix_duplicate_ids, fix_404_title

Rotating UX Pool — 17 UX Functions

Category	Fix Functions
Mobile	ux_viewport, ux_mobile_tap, ux_touch_action, ux_table_responsive
Interaction	ux_hover_styles, ux_button_states, ux_smooth_scroll, ux_focus_visible, ux_keyboard_nav
Layout	ux_css_order, ux_z_index, ux_spacing, ux_type_scale, ux_cls_fix
Accessibility/UX	ux_skip_link, ux_breadcrumb, ux_print_styles, ux_loading_indicator

Each pool function:

• Uses find or rg to locate matching files in the branch checkout
• Runs an inline Python script to detect and apply a specific improvement
• Sets POOL_DESC and POOL_FILE globals on success
• Advances a persistent rotation index (fix_rotation / ux_rotation) so the next run picks up where the last left off

Cornerstone Missions

Instead of telling the AI to "find a bug", the system prompts it to "Benchmark this site against professional standards for its category." Each branch has a category (ecommerce, community, platform, documentation, advocacy, archive, meta) with a custom checklist:

Category	Focus Areas
E-commerce	JSON-LD Product/Offer, OG tags, cookie consent, payment security, trust signals, GDPR privacy policy, returns policy, sitemap.xml, robots.txt, checkout UX, search filter
Community	JSON-LD DiscussionForum, signup/login flow, moderation tools, privacy policy, member profiles, notifications, search, pagination, onboarding
Platform	JSON-LD WebApplication, API docs, status page, rate limiting, auth flows, dashboard UX, error pages, changelog
Documentation	JSON-LD TechArticle, table of contents, search, breadcrumbs, code copy buttons, dark mode, version selector, last-updated timestamps

The --constraint CLI parameter injects blacklist rules: the AI is told what NOT to do (no console.log removal, no whitespace cleanup, no comment deletion — those belong to the pool).

Self-Learning Profiles

Every branch has a profile in profiles.json with:

• LLM Rotation: Ordered list of API endpoints with failover — financecheque.uk proxy → localhost → child proxy → NVIDIA → OpenRouter → Gemini → DeepSeek
• Rotation Index: Advances per attempt to distribute load across providers
• Successful/Failed Fixes: Immutable log of every fix (timestamp, source, tool, file, description, commit message)
• Skill Library: Deduplicated skills learned from successful fixes (name, description, tool, file type, use count)
• Learned Patterns: Patterns derived from fix file extensions and tools
• Reflections: Outcome log for every fix attempt (success/failure, tool, source)
• Branch Knowledge: Concatenated commit messages forming an evolving knowledge base of what the branch has been through
• Total Learning Cycles: Monotonically increasing counter of profile updates

After every fix, learn_from_fix() updates the profile atomically. The AI prompt dynamically injects recent reflections, successful patterns, branch knowledge, and the skill library so each run builds on prior context.

AI Fix Pipeline

1. intelligence.py loads the branch context (branch.md + category checklist), profile (history + skills + reflections), and builds a prompt with the Cornerstone Mission
2. The prompt is sent to the LLM rotation (first available API endpoint)
3. The LLM returns a JSON fix specification (file_path, bug_description, tool, new_str or sed old_str)
4. intelligence.py --apply applies the fix to the working tree
5. If package.json exists, npx eslint runs on the changed file for build validation
6. If lint fails, the fix is reverted (git checkout -- file) and the error is fed back for up to 3 self-correction attempts
7. On success, the branch memory, global memory, and profile are updated

Meta-Review (Executive Coach)

Runs daily at 00:00 UTC on a separate command machine via meta-review.sh:

1. SSH into the AWS server and fetch the last 200 log lines and release-state.json
2. Run gather_review_data.py to parse log for fix source classification (AI vs POOL vs FALLBACK, unique fix types per session)
3. Run intelligence.py --type meta --branch aws to analyze AWS performance and suggest improvements
4. Apply suggested changes to the flywheel scripts
5. Sync back: download AWS profiles, merge successful_fixes/skill_library/reflections/learned_patterns/branch_knowledge
6. Push updated code to GitHub

Versioning & Releases

Tags follow {branch}-v0.0.{patch}.{build:02d}:

release_number = total_releases[branch] + 1
patch  = release_number / 100          # increments every 100 releases
build  = release_number % 100          # 0–99, zero-padded
version = 0.0.{patch}.{build:02d}

Each GitHub release contains:

## [{tag}] - YYYY-MM-DD

### Fixed
- fix(branch): Add JSON-LD structured data for SEO
- fix(branch): Add meta description tag for SEO
- fix(branch): Remove duplicate blank lines (guaranteed fallback)

### Changed
- ux(branch): Add skip-to-content link for keyboard accessibility

After release, the system:

• Verifies the release appears on GitHub (up to 5 retries, 3s apart)
• Verifies Cloudflare deployment responds at {branch}.datro.pages.dev (up to 5 retries, 10s apart)
• Runs runtime quality checks (console errors, viewport issues via curl)
• Prunes old releases (keeps last 3 per branch)
• Advances global rotation_index for branch selection

Cooldown & Branch Selection

24-hour cooldown per branch, enforced by comparing last GitHub release timestamp against current time. Selection is round-robin via rotation_index in state file. If all branches are on cooldown, the script finds the branch closest to expiry and sleeps up to 1 hour. A FORCE_BRANCH env var bypasses rotation for manual targeting.

Cloudflare Pages Integration

Every branch has a Cloudflare Pages project with:

• root_dir set to static/{branchname}/
• Custom domain {branch}.datro.xyz (CNAME to {branch}.datro.pages.dev)
• Auto-deploys on git push
• World site (datro.xyz) has no root_dir constraint (serves from repo root)

Deploy Your Own

Server Setup

# Ubuntu 24.04 EC2 (t3.small+)
sudo apt update && sudo apt install -y git python3 ripgrep nodejs npm
npm install -g wrangler
sudo apt install -y gh && gh auth login

Clone & Configure

git clone https://github.com/YOUR_USER/YOUR_REPO.git ~/repo
mkdir -p ~/.fcukproxy
cp multi-branch-release.sh intelligence.py profiles.json ~/.fcukproxy/
# Edit BRANCHES array and GITHUB_REPO in multi-branch-release.sh
# Add 24h cron:
0 * * * * /home/ubuntu/.fcukproxy/multi-branch-release.sh >> ~/logs/multi-branch-release.log 2>&1

Required Files

• ~/.fcukproxy/multi-branch-release.sh — main orchestrator
• ~/.fcukproxy/intelligence.py — AI pipeline + profile management
• ~/.fcukproxy/agent/profiles.json — per-branch profiles with LLM rotation lists
• ~/.fcukproxy/agent/branches/{branch}.md — per-branch context files (soul, manifest, memory, cornerstone mission)
• ~/.fcukproxy/agent/aws-supervisor.md — executive coach context for meta-review

Security Model

• No secrets in scripts — API keys loaded from environment variables or external config
• All LLM API calls use timeout guards (60s AI, 30s apply, 10s learn)
• Lock file (/tmp/multi-branch-release.lock) prevents concurrent cron runs
• Stderr from tools captured to log only
• Atomic file writes via tempfile+rename for state and profiles (prevents corruption)
• IAM permissions: EC2 read/reboot only (for recovery, not exposed scripts)
• GitHub tokens scoped to repo via gh CLI OAuth

State File Reference

{
  "rotation_index": 1,
  "total_releases": { "financecheque": 150, "archives": 5, ... },
  "last_release":   { "financecheque": 1779506355, "archives": 1779500752, ... },
  "fix_rotation": 23,
  "ux_rotation": 0
}

Licence

MIT — use freely, adapt for your own repos. Source at github.com/unclehowell/datro/tree/cnei/flywheel.

---

Generated by the cnei branch — source · Last update: May 2026