Skip to content

MASTERPLAN.md

Latest commit

 

History

History
93 lines (78 loc) · 3.79 KB

MASTERPLAN.md

File metadata and controls

93 lines (78 loc) · 3.79 KB
Raw

DATRO Flywheel — Master Plan

Architecture Overview

The flywheel is an autonomous agentic system running on Cloudflare Workers that continuously improves all branches of the DATRO monorepo.

Key Components

1. Agentic Loop (ReAct Pattern)

Replaces the old single-prompt AI flow. Each release cycle runs a chain-of-thought agent:

  • Think: Agent recieves full context (HTML, wing files, brain memory, Honcho, research)
  • Tool Call: Agent picks a tool to investigate or modify the codebase
  • Observe: Tool results are fed back to the agent
  • Repeat: Up to 6 iterations per cycle
  • Done: Agent outputs SUMMARY and commits are tagged/released

2. MCP Tool System

The agent has these tools at its disposal:

Tool Description
read_file Read any file from any branch
write_file Write + commit changes to any branch
list_branches List all branches and their latest commit
search_code Search codebase for patterns
run_scan Run EAA WCAG / AccessScore scans
honcho_memory Read/write Honcho memory per branch
wallet_info Get crypto wallet + compute budget
brainstorm Ask LLM for creative thinking

3. Deterministic Crypto Wallets (per Branch)

Each branch gets a unique deterministic wallet derived via HKDF-SHA256:

MASTER_SEED + HKDF(salt="agentic-flywheel-v2", info="branch:<name>") → 256 bits
  ├── First 32 bytes → Wallet private key
  └── Last 4 bytes → Compute budget (0-9999 units)

Wallets are cached in KV for 24h. Used for:

  • Compute emulation (higher budget → faster gear)
  • Branch identity verification
  • Future: signing releases, proof-of-work

4. Compressed Release Cycle

  • Cron: */10 * * * * (every 10 min, down from 30 min)
  • Gear 1 cooldown: 600s (10 min, down from 3600s)
  • Gear 10 cooldown: 35s
  • Burst Mode: Progressive bias (1-2) → gear +2; High risk (4-5) → gear +1; High compute → gear +1
  • cnei: Gets half-gear cooldown for faster self-improvement
  • No-branch fallback: Forces cnei when all regular branches are on cooldown

5. Honcho Integration

  • Read per-branch memory for durable facts
  • Write observations after each release cycle
  • Used in system prompt for agent context

6. Brain / Memory System

  • MEMORY.md per branch: Records each cycle's changes
  • BRAIN_CONSOLIDATED.md on brain branch: Aggregated memory from all branches
  • Honcho: Durable facts that persist across deploys
  • KV: State (rotation, bias, config, wallets, MCP cache)

7. Release Pipeline

Cron trigger (10 min)
  → Acquire lock (KV)
  → aggregateMemory to brain branch
  → Check bias/risk/wallet for burst mode
  → Find available branch (off cooldown)
  → processBranch():
    1. Apply self-improvements (cnei only)
    2. Run MCP scans (EAA, AccessScore)
    3. Inject master record visuals from wing files
    4. Run agentic loop (ReAct, up to 6 iterations)
    5. OR fallback to best-practice engine
    6. OR audit-only release
    7. Create git tag + GitHub release
  → Save state + release lock

Wallet System (Compute Emulation)

Each branch wallet's computeBudget (0-9999) is derived from the deterministic seed. When computeBudget > 5000, the gear advances by 1, giving faster releases to branches with "more compute."

Wallet addresses are hex strings derived from SHA-256 of the per-branch key (first 20 bytes, lowercased).

To inspect: GET /__wallet?branch=<name> or GET /__wallets

Future Roadmap

  • MCP server endpoint for external tool access
  • Agent-to-agent communication between branches
  • Merge request generation via cnei
  • Distributed compute across branches (Kruskal's / Merkle)
  • Honcho background reasoning for cross-branch insights
  • Auto-deploy cnei improvements to production flywheel