tx-safety

Transaction safety guardrail

submit_transaction broadcasts a signed transaction to the real network. Nothing about the tool itself stops it from running against mainnet with the wrong amount, the wrong recipient, or while the user believed they were on a testnet. Once a transaction lands on-chain it cannot be undone. Treat every call to submit_transaction — and the build_* calls that precede it — as irreversible until proven otherwise.

Workflow

1. Before building any transaction, call get_network_info. Do not assume the network from context, prior conversation, or what the user "probably" meant.
2. State the active network out loud to the user before proceeding, plainly — not buried in a JSON blob. If is_mainnet is true, say so explicitly: this transaction will move real ADA/assets.
3. After building a transaction (build_payment_tx or build_smart_contract_tx), before calling submit_transaction, restate to the user in plain language: the network, the sender, the recipient, the exact amount (in ADA, not just lovelace), and any native assets. This restatement must happen even if the user already stated these values earlier in the conversation — confirm what the built transaction actually contains, not what was requested.
4. Get an explicit go-ahead from the user in the current turn before calling submit_transaction — a prior "yes, build it" is not the same as confirmation to submit. If the user has not responded to the restatement yet, stop and wait; do not call submit_transaction speculatively.
5. If get_network_info reports mainnet and the user has not clearly indicated they intend to use mainnet (e.g. they only ever mentioned "testnet" or "preprod"), flag the mismatch and ask before continuing rather than proceeding on the assumption they meant mainnet.

Why this exists

The eUTxO model has no reversible "undo" — a submitted transaction is final once confirmed, and a testnet/mainnet mix-up or a mistyped amount has real financial consequences. A model's confidence that it has the right sender, recipient, and network is not the same as verifying it — this skill forces that verification through get_network_info and an explicit human confirmation step, the same way aiken-verify forces compiler verification before showing Aiken code as correct.