HarvestBot

An agent that harvests tax losses on Robinhood tokenized equities — bounded by an onchain mandate it cannot exceed, and slashed when it lies.

---

✅ Verification

Every row is on chain 46630 (Robinhood Chain testnet) and every hash is a live explorer link. The agent key and the owner key are different keys. Check it live, no wallet: site · /verify (your browser reads these receipts from chain) · /judge (the 30-second path) · deck.

Tests	91 Foundry (forge test) — 100 % line · 100 % function · 98.3 % statement · 89.2 % branch coverage on src/ product contracts (make coverage; the bench-only memory-copy ledger is 100 % line-covered on its own), 5 fuzz properties, regression tests named for the defect they pin · 6 Stylus native (cargo test)
Static analysis	Slither 0 high · 0 medium — every remaining finding triaged in docs/SLITHER.md · forge fmt · cargo clippy -D warnings
Killer number	64-lot HIFO harvest: Stylus 1.69× less L2 compute than a memory-optimized Solidity ledger — the fair engine-to-engine figure (857,559 → 507,979 gas; 1.79× at 128 lots, 0.86× at 8 where uncached Stylus loses). Against the straightforward Solidity twin we shipped: 1,456,905 → 507,993, 2.87× (3.1× at 128) — the number we first published, ~41 % of it our baseline re-reading storage. Same chain, same inputs, identical output across all three (bench)
Beat 1 · HARVEST	0x689c4d34…6123 — agent-signed, router-gated, HIFO recomputed on the Stylus ledger, −$24.531251 realized on 9 lots, real AMZN → real NFLX
Beat 2 · BLOCKED REBUY	0xd486468e…cf10 — reverted on chain, WashSaleViolation
Beat 3 · SLASHED	attempt 0xd8b6e577…94c8 reverted OffMapSubstitute → challenge 0x6b896b6e…7dcb — bond 1,000 → 800 mUSDC, challenged by a third-party key (0x0Fec…28e2, neither owner nor agent). The hardened bond also refuses stale proof: the expired beat-2 envelope (0x673f8459…ed51) and the already-executed beat-1 envelope (0xbbf04002…c049) both revert EnvelopeNotLive

Contract	Address	Kind
TaxLotLedger	0xEff7…5a21	Stylus (Rust → WASM) — deploy
HarvestMandate	0x15FD…a6E2	Solidity
AgentBond	0xBA92…e273	Solidity — hardened build, redeployed 2026-09-27 (details); retired pre-fix V1 0x822f…0736 kept for history
ExecutionRouter	0x2a12…19aE	Solidity
WashSaleGuard	0x1A98…A29C	Solidity
SubstituteMap	0x1e8D…7CfE	Solidity
MockOracle · MockSwap · MockUSDC	0xBaDD…2598 · 0xFd7f…9E6a · 0x0E8A…3f89	MOCK, labeled
Assets	AMZN 0x5884…9E02 · NFLX 0x3b82…8C93 · PLTR 0x1FBE…98d0	real Robinhood testnet Stock Tokens (faucet)

Deployer/owner 0x72cd…557b · agent 0xF783…457E · all addresses in deployments/46630.json, all receipts in receipts/46630.json.

---

💡 The problem

In December, Maya sees her tokenized-equity position down $3,140 on paper and does nothing — because banking that loss properly means picking which of 64 tax lots to sell, then policing a 30-day calendar so she doesn't trip the wash-sale rule and void the deduction. Wealthfront and Betterment sell exactly this service ("direct indexing") for 0.25 %/yr. On chain, where Robinhood now issues tokenized stocks, none of the plumbing exists: no cost-basis ledger, no specific-lot identification, no wash-sale enforcement — and no safe way to hand the job to an agent.

HarvestBot is that plumbing, with the agent bounded by contracts instead of by trust:

• 📒 Onchain HIFO tax-lot ledger — a Stylus (WASM) contract records every lot, selects the highest-basis lots for a sale, prorates the last one, and computes the realized loss deterministically. The compute-heavy part lives in Rust because that is where WASM beats EVM opcodes — 1.69× less L2 gas at 64 lots than a memory-optimized Solidity ledger (2.87× vs our straightforward Solidity twin), measured, not asserted (bench).
• 🧱 Contract-enforced wash-sale rule — after a harvest, any rebuy of the same or a substantially-identical asset reverts for 30 days. The agent cannot route around it; the chain says no.
• 🔐 A mandate the agent cannot exceed — the agent's key can call exactly one function through a router; the mandate re-verifies signature, nonce, deadline, substitute map, wash-sale window, HIFO selection and realized loss at the oracle mark before anything moves. No path exists from the agent key to a withdrawal.
• ⚖️ A slashable bond — anyone holding an envelope the agent signed that the mandate must reject can slash it: 20 % of the bond, 10 % of that to the challenger, the rest to the owner. Being wrong costs the agent more than it can ever gain.

🎬 The demo, in three beats

	Beat	What happens on chain
1	HARVEST	Agent signs the HIFO picks → router → mandate recomputes them on the Stylus ledger → real AMZN rotates into real NFLX, loss realized, window opened
2	BLOCKED REBUY	Same agent tries to rotate back into AMZN → WashSaleViolation, transaction reverted by the contract
3	SLASHED	Agent signs an off-map rotation into PLTR → reverted → anyone submits the signed envelope to challenge() → bond slashed, bounty paid

Numbers, hashes, and the reproduce commands: DEMO.md. The 30-second version: JUDGE.md.

🏗️ Architecture

flowchart LR
    A[agent key] -->|"sign EIP-712 decision"| R[ExecutionRouter<br/>1 target · 1 selector · value 0]
    R --> M[HarvestMandate<br/>owner custody]
    M --> S[SubstituteMap] & G[WashSaleGuard<br/>30 days] & O[MockOracle · MOCK]
    M -->|"computeHarvest → realize"| L[TaxLotLedger<br/><b>Stylus / WASM</b>]
    M --> X[MockSwap · MOCK]
    C[anyone] -->|"challenge(envelope)"| B[AgentBond]
    W[owner key] -->|"deposit · withdraw"| M

Loading

Layer	Choice	Why
Ledger	Rust → WASM on Arbitrum Stylus (stylus/ledger)	HIFO over 64 lots is a compute loop; WASM is cheaper per instruction than EVM — measured, not asserted
Ledger twin	Solidity (contracts/src/TaxLotLedger.sol)	Same ABI, same integer semantics; the benchmark control and the fallback for non-Stylus chains
Mandate · router · guard · map · bond	Solidity 0.8.28 + OpenZeppelin 5.4	Ownable, SafeERC20, ReentrancyGuard, EIP712, ECDSA
Assets	Robinhood testnet Stock Tokens	Real faucet tokens (AMZN/NFLX/PLTR), 18-dp ERC-20s
Oracle · swap · bond token	MOCK contracts, labeled	No Stock-Token AMM or USDC exists on the testnet

Seven invariants (INV-1 custody … INV-7 determinism) each have a named test — docs/ARCHITECTURE.md.

🏆 Why only Arbitrum + Robinhood Chain

• cargo stylus deploy → WASM contract on Robinhood Chain. The ledger is not Solidity. ArbWasm.stylusVersion() = 3 on chain 46630; the deploy tx above is the single cargo stylus deploy transaction. Remove Stylus and the HIFO scan costs 1.69× more L2 gas at 64 lots (1.79× at 128) on the best Solidity we wrote for it — a memory-copy ledger with the Rust structure — and 2.87× (3.1×) on the straightforward twin we first benchmarked. Measured on chain 46630 with the program uncached (the testnet has no ArbOS cache manager yet), so these are worst-case Stylus numbers.
• Robinhood Stock Tokens are the asset. The mandate holds 0x5884…9E02 (AMZN) and 0x3b82…8C93 (NFLX) from the official faucet — a tokenized-equity tax ledger only makes sense on the chain that issues tokenized equities.
• Arbitrum precompiles as a capability probe. cast call 0x…71 'stylusVersion()' answered the "is WASM live on this Orbit chain?" question in one RPC call, before any code was written (DX report).
• Honest limitation: there is no Stock-Token AMM on the testnet, so the rotation executes through a labeled MOCK fixed-rate venue at the oracle mark. The tokens are real; the venue is not.
• Scope of the wash-sale rule: the contract walls re-buys for the 30 days after a harvest; the IRS look-back (a purchase in the 30 days before the loss sale) is not yet enforced. This and the other known limits are listed in docs/ARCHITECTURE.md.

🚀 Run it

git clone --recurse-submodules https://github.com/edycutjong/harvestbot && cd harvestbot
cd contracts && forge test                     # 91 tests, incl. exactly −$3,140.000000 on the seed scenario
cd ../stylus/ledger && cargo test              # 6 native tests on the WASM engine

# live, read-only, no wallet — ask the deployed Stylus ledger for Maya's HIFO picks
cast call 0xEff7B46049fC677F58264e0ebb19dF1a39195a21 \
  'computeHarvest(address,address,uint256,uint256)(uint64[],int256)' \
  0x72cd3cB98A5d9B830b386EeBA7B2340132Ba557b 0x5884aD2f920c162CFBbACc88C9C51AA75eC09E02 638501157098660213 412300000 \
  --rpc-url https://rpc.testnet.chain.robinhood.com
# → the NEXT nine HIFO lots, [55, 54, …, 47]: beat 1 already sold lots 63–56 and part of 55 —
#   its own picks [63, …, 55] and −24531251 are in that tx's HarvestReport event

Your own deploy (testnet key + faucet ETH, see .env.example): forge script script/Deploy.s.sol → cargo stylus deploy → scripts/seed.sh → scripts/beats.sh → python3 scripts/bench.py.

🧪 Testing & CI

make ci = forge fmt --check · forge test · forge coverage · gas-snapshot gate (make snapshot, fails on a >2 % regression) · cargo fmt/clippy -D warnings/test on both crates · ruff + shellcheck on scripts/ · gitleaks over full history + Slither (make security-scan, fails on medium+) · readiness check. GitHub Actions run the same gates in parallel, plus Slither SARIF, TruffleHog, CodeQL (Rust · Python · Actions), Lighthouse CI on the live site after each deploy (lighthouserc.json), Dependabot (Cargo · submodules · Actions), and semver releases cut from conventional commits only after CI is green on that commit (grouped notes, live links and the deployment/receipt/bench JSON attached).

Category	Where	What it proves
Three beats end-to-end	Mandate.t.sol, Bond.t.sol	the demo, deterministically
Invariants INV-1…INV-6	Mandate.t.sol, Router.t.sol, WashSale.t.sol	custody, loss integrity, wash-sale monotonicity, on-map, authenticity, bond dominance
Fuzz properties	testFuzz_* (5)	HIFO never worse than lowest-basis-first · window is exactly 30 days for any timestamps · breach is negative-EV for any bond · memory-copy and shipped Solidity ledgers pick identical lots and losses on random books, before and after realize
Regressions named for the defect	e.g. test_reenabling_a_disabled_pair_does_not_duplicate_it_in_substitutesOf	found and fixed during the build
Engine equivalence	bench/results.json (lotIds + realizedLoss identical across all three engines at every size) · MemCopy.t.sol (differential, 8–128 lots + fuzz)	INV-7

📁 Layout

contracts/      Foundry — src/ (6 contracts + 4 labeled mocks + bench/ memory-copy ledger), test/ (91), script/ (Deploy, Seed, Envelope)
stylus/ledger/  Rust — the Stylus TaxLotLedger (+ stylus/spike, the day-0 activation receipt)
scripts/        seed.sh · beats.sh · bench.py · check_submission_readiness.py   (cast-driven: forge cannot simulate WASM)
deployments/    46630.json (the system) · 46630-bench.json (3 bench ledgers)
receipts/       46630.json (the three beats) · envelopes/*.hex (the signed decisions, incl. the rogue one)
bench/          results.json · RESULTS.md
docs/           ARCHITECTURE.md · DX-REPORT.md · SLITHER.md · assets/
DEMO.md · JUDGE.md

📄 License

MIT © 2026 Edy Cu

🙏 Acknowledgments

Built solo for the Arbitrum Open House Singapore Online Buildathon (HackQuest, Sept 2026). Thank you to Offchain Labs for Stylus and cargo-stylus, to Robinhood for a testnet faucet that hands out real Stock Tokens, and to the judges for their time reviewing this project.