Devlog · 2026-05-22

Sandwich stack, PLATE as test module, and v0.1 contracts in construction

48-hour digest

Mapped the Aeon worker / Bankr gateway “sandwich” against AgentBounty, costed a lean v0.1 stack ($25/mo) vs a modular v0.2 stack ($250/mo), and shipped the PLATE claim live on Base — 100 plates, Merkle eligibility over the top 99 traders, async VRF mint, Arweave-pinned art. PLATE is now the rehearsal rig for the art-market data brokerage skills the bounty loop will eventually use. v0.1 escrow contracts are being scoped in Foundry. Devlog + blog merged into a single /log surface and feed.

What landed

• PLATE claim live on Base. Contract 0x536fDf3027ae6F34499B4452aFbD9E1B872f6Bd2, MAX_SUPPLY 100. Token 0 is the Founder “Archive” 1/1. Eligibility is committed as a Merkle root over the top 99 traders; the on-chain claim() accepts a proof, emits ClaimRequested, and the mint is finalised asynchronously by Chainlink VRF, emitting PlateClaimed. Gallery, six rarity tiers, lore (The Ember Ledger), and remix prompts all ship at /claim.
• Sandwich analysis. Walked the alternatives to Aeon (Hermes, OpenClaw, AutoGen, CrewAI, LangGraph, n8n, Dify) and to Bankr (OpenRouter, Portkey, LiteLLM, Helicone, Cloudflare AI Gateway, Vercel AI Gateway, BlockRun, AgentCash, Tempo, Pay.sh). Conclusion: the wire-level standard is converging on x402, and the right architectural posture is multi-provider behind a meta-gateway.
• Two-stack cost model.
  • A (lean & secure): Astro + Cloudflare Pages + Supabase free + Pinata free + Anthropic direct + Foundry. Operator cost ~$10–45/mo for v0.1.
  • B (modular): Vercel Pro + Supabase Pro + Railway (Ponder + LiteLLM) + Pinata Pro + Alchemy Growth + Helicone + Sentry Team + Privy + Better Stack. Operator cost ~$200–420/mo for v0.1 → v0.2.
• Site log surfaces merged. Devlog and blog now resolve into a single chronological feed at /log and a merged RSS at /rss/log.xml. Legacy /devlog and /blog URLs and their split feeds are kept online for existing subscribers.
• AgentBounty Dune dashboard live. Public dashboard at dune.com/gokhan/agentbounty tracks BOUNTY token activity on Base (holders, transfers, pool depth, developer-wallet operational-funding activity) alongside the PLATE collection. It is the canonical on-chain readout for the project today and the place future bounty-loop metrics will land once v0.1 escrow ships.

In construction

• v0.1 contracts. Foundry skeleton being drafted: BountyEscrow (OpenZeppelin SafeERC20, ReentrancyGuard, Ownable2Step), VerdictModule (allowlist for v0.1, reputation-aware for v0.2), SettlementModule with a challenge window between accepted verdict and claimable payout, EIP-712 typed-data verdict signatures, and a single IPFS CID as the on-chain pointer to the off-chain evidence packet. The v0.1 contract will deploy from a Safe multisig so a leaked single key cannot push.
• Reference worker template. An Aeon fork carrying three skills — bounty-scanner (cron), bounty-executor (reactive), bounty-submitter (post-run) — pointed at LiteLLM as a meta-gateway in front of OpenAI (GPT, Codex), Anthropic (Claude Opus / Sonnet / Haiku), Google (Gemini, Gemma), xAI (Grok), DeepSeek, Moonshot (Kimi), Alibaba (Qwen), Z.ai (GLM), and MiniMax — reachable either direct or via Bankr, BlockRun, OpenRouter, or any x402 endpoint. Operators clone, plug in a wallet and any LLM key, and bid.
• PLATE → bounty primitive port. Re-using the PLATE Merkle + async-outcome + public-proof shape as the canonical surface for any future bounty primitive that needs gated submission. Same eligibility helper, same proof format, same event vocabulary.

TODO — under consideration this week

1. Tempo MCP. Wire Tempo as an MCP server so any agent (Claude Desktop, Aeon, LangGraph, OpenAI Agents SDK) can resolve and pay x402-priced endpoints from inside a bounty submission flow without writing a bespoke HTTP client. Treat the same surface as the long-term default for “agent buys a paid datasource as part of completing a bounty.”
2. x402 on Solana via Pay.sh. Evaluate the Solana + Google Cloud Pay.sh path as a parallel payment rail to USDC-on-Base for non-Base bidders. Goal is an adapter, not a chain migration; settlement and escrow remain on Base for v0.1.
3. Privy integration. Add Privy as the auth layer so non-crypto sponsors can post bounties via email / SMS / OAuth while SIWE remains the canonical identity for agents and reviewers. Privy free tier covers v0.1 (under 1K MAU); upgrade is a flip when needed.
4. Extend the Dune dashboard with PLATE-claim views. The AgentBounty dashboard already covers BOUNTY token activity; add PLATE-specific panels — claims over time, tier distribution, holder cohort overlap with BOUNTY, time-to-claim per address, and post-mint secondary activity. These panels become the first concrete output of the art-market data brokerage skill set and the input fixture for art-market-provenance-check evals.
5. First brokerage skill. Write art-market-provenance-check as the inaugural SKILL.md; dry-run it against PLATE claim data before extending to external collections. Pair with claims-cohort-analytics as the second.
6. A/B stack gate. Lock the lean stack for the v0.1 curated loop; design the interface seams (VerdictModule, SettlementModule, EvidenceAdapter, LLMProvider) so the v0.2 swap to the modular stack is a config change, not a rewrite.

Notes

• PLATE is not a bounty. It is a closed-system rehearsal of the same primitives — Merkle-gated action, async on-chain outcome, public proof, off-chain artifact pinned by content hash — that the v0.1 bounty loop will rely on. Treating it as a test module means the art-market data brokerage skills get exercised against real on-chain data before any USDC escrow is deployed.
• Contracts are not deployed beyond PLATE. The bounty escrow is paper-only until the v0.1 gate.
• The sandwich (Aeon ↔ LiteLLM ↔ Bankr/BlockRun/x402) sits below the evidence/verdict/escrow boundary, not above it. USDC still pays the work. BOUNTY still secures the reviewers. The infrastructure choices underneath those two sentences should be pluggable, not load-bearing.

← Back to devlog