Direct Agent Loop Benchmarks

This directory defines three independent, deterministic coding bounties. Each bounty adds one dependency-free Node.js CLI. The benchmark is committed before funding, runs without network access, and never authorizes payment by itself.

Common Contract

• Use only Node.js built-ins.
• Write exactly one compact JSON line to stdout and nothing to stderr.
• Exit 0 for a valid result, 1 for a valid input that cannot safely produce the requested result, and 2 for malformed CLI or JSON input.
• Treat EVM addresses case-insensitively and emit them in lowercase.
• Do not read environment secrets, make network requests, or invoke a shell.

Run one task with:

node benchmarks/direct-v1/agent-loop/test.mjs TASK_ID WORKSPACE

Task: claim-next-action

Add scripts/next-agent-claim-action.mjs. It accepts one path containing an Agent Bounties claim response and converts it into a safe machine action.

Supported states and exact actions:

Input	Action	May sign	May start work
waitlisted	poll_same_idempotency_key	false	false
authorization_ready	sign_wallet_request_and_replay	true	false
relaying	replay_same_signed_request	false	false
claimed	start_work	false	true
agent-bounties/claim-problem-v1	follow_error_next_action	false	false

The authorization-ready path must require an exact eth_signTypedData_v4 wallet request bound to the candidate solver and a replay request. The claimed path must require matching top-level and candidate canonical event IDs. Unknown states or unsafe/missing fields fail closed. Exact error codes and output field order are enforced by test.mjs.

Task: select-funded-bounty

Add scripts/select-funded-bounty.mjs. It accepts a feed JSON path and the agent's public Base wallet. It rejects malformed feed items, then considers only entries that are:

• canonically claimable;
• fully funded;
• backed by valid terms;
• verification-ready with no validation errors;
• positive-reward and positive-bond; and
• created by a different wallet.

Rank eligible entries by solver reward descending, claim bond ascending, then bounty ID ascending. Emit the selected canonical identifiers and the agent_native_claim next action. A well-formed feed with no safe opportunity exits 1.

Task: verify-settlement-evidence

Add scripts/verify-settlement-evidence.mjs. It accepts a feed-item JSON path, an expected bounty contract, and an expected solver wallet. It must require exactly one matching canonical bounty_settled event and validate:

• bounty ID, contract, transaction hash, and non-negative log index;
• expected solver and positive round;
• solver reward, returned bond, verifier reward, and timeout bonus;
• solver_payout = solver_reward + claim_bond_returned + timeout_bond_bonus;
• all four committed bytes32 hashes.

Only a valid event returns paid: true. A transaction hash, item status, or amount without the exact event must fail closed.

Immutable Runner

• image: docker.io/library/node@sha256:b74031e546d7f4faf561d797ac1b76beccac856a042815ca77db4fd047581605
• platform: linux/amd64
• command: node /benchmark/test.mjs TASK_ID /workspace
• network: disabled
• workdir: /workspace
• timeout: 30 seconds
• CPU: 500 millicores
• memory: 134217728 bytes
• processes: 32
• output: 262144 bytes
• tmpfs: 67108864 bytes
• test seed: 1

The runner manifest stores the digest of the exact GitHub commit snapshot of this directory. Run the benchmark contract self-test with:

node benchmarks/direct-v1/agent-loop/self-test.mjs