Protocol Deployment

Prepare and execute evidence-backed deployments of Sablier Airdrops, Flow, and Lockup to a new EVM chain.

Arguments

• Resolve the first argument as the target chain name or numeric chain ID.
• Default the package selection to all; otherwise limit work to the named packages and their prerequisites.

Contract

• Treat deployment preparation, simulation, and reporting as read-only or local work.
• Treat --broadcast, contract verification submissions, Safe creation, role changes, and other on-chain writes as external writes. Require explicit user authorization immediately before the first such action.
• Never read, print, log, or echo an existing .env, mnemonic, private key, keystore password, or API key. Ask the user to populate secrets; inspect only variable names, example files, permissions, and signer addresses.
• Preserve deterministic deployment inputs: repository revision, package version, optimizer profile, constructor arguments, salt logic, CREATE2 deployer, and admin selection.
• Unless the user specifies another address, propose 0xcB88fBf459000853F22a7296b23d163901BB385E as the initial admin for new deployments. Do not assume the broadcaster is also the admin. Confirm the proposed admin before broadcast. If a legacy bootstrap requires a transient upgrader, disclose it and transfer control to the proposed admin before deploying dependent protocols.
• For an existing protocol release, resolve the canonical deployment commit, package version, compiler settings, creation bytecode, runtime bytecode, and salt version from ~/sablier/sdk/deployments when that checkout is available. Treat checked-in artifacts, broadcasts, and on-chain code as authoritative when their README prose conflicts.
• A newer branch is acceptable only after its creation and runtime bytecode match the canonical SDK artifacts exactly. Matching bytecode does not permit a different package version to silently change BaseScript's CREATE2 salt; pin the documented release version separately and record how it was pinned.
• Stop before broadcast when the chain ID, administrator, broadcaster, CREATE2 factory, Comptroller address, package selection, or funding is unresolved.
• Finish only after recording transaction hashes, deployed addresses, receipt status, runtime code, constructor relationships, and any incomplete verification or repository bookkeeping.

Workflow

1. Inspect the Deployment Surface

Read the root and applicable package AGENTS.md files. Do not read TODO.md. Inspect rather than assume:

• Root and package justfile deployment recipes and their package ordering.
• .env.example, root .gitignore, foundry.base.toml, and package foundry.toml files.
• utils/src/tests/BaseScript.sol and utils/src/tests/ChainId.sol.
• The selected packages' deterministic and non-deterministic deployment scripts.
• Flow and Lockup NFT descriptor address maps.
• Current package versions, salts, constructor arguments, and broadcast artifact paths.
• The matching SDK deployment README, broadcast commit, artifacts, and transaction inputs for an already-deployed release. Inspect the deployment commit's pinned Foundry version and use it when current tooling does not reproduce the canonical artifacts.

Snapshot git status --short before generators, builds, simulations, or broadcasts. Ignore unrelated changes from other agents and do not modify or report them.

Completion criterion: identify the exact scripts, prerequisite graph, signer-selection path, admin-selection path, canonical release commit and salt version, and expected deployment artifacts for every selected package. Either use the canonical deployment checkout or prove byte-for-byte creation and runtime parity for every deployed contract.

2. Resolve and Probe the Chain

Use $evm-atlas to resolve the target mainnet, native token, authoritative chain ID, public RPC, RouteMesh support, provider route, and explorer. If Atlas does not support the chain, surface that limitation and request authoritative chain data rather than silently substituting another chain or RPC.

Prefer the Atlas public RPC when routeMesh is false. Before any live or API command, run harmless prerequisites and identify local files it may create.

Run the bundled read-only helper from the repository root:

bash '.agents/skills/protocol-deployment/scripts/preflight.sh' '<RPC_URL>' '<CHAIN_ID>'

It verifies eth_chainId and reports whether Foundry's canonical CREATE2 deployer 0x4e59b44847b379578588920cA78FbF26c0B4956C and the expected Sablier Comptroller proxy 0x0000008ABbFf7a84a2fE09f9A9b74D3BC2072399 already have code. Pass a third argument only when repository evidence selects a chain-specific Comptroller address.

Completion criterion: the RPC returns the expected chain ID and the CREATE2 deployer and current Comptroller states are known. An absent CREATE2 deployer makes vanity deployment unavailable but does not block the bespoke CREATE fallback.

3. Prepare New-Chain Support

If ChainId.isSupported rejects the chain, make the minimum conventional change in utils/src/tests/ChainId.sol:

• Increment the relevant count.
• Add the chain constant.
• Insert it into the corresponding chain array without leaving uninitialized entries.
• Add its canonical lowercase name to getName.

Add an RPC alias to foundry.base.toml only when it improves the workflow; a full --rpc-url is sufficient. Do not add a RouteMesh endpoint when Atlas reports routeMesh: false.

Do not pre-populate Flow or Lockup NFT descriptor maps with predicted addresses. Their scripts use a zero mapping entry to decide that the descriptor must be deployed. Add the actual addresses after successful deployment.

Validate the narrow change with the Utils tests/checks and optimized builds for only the selected packages. Inspect the relevant recipes first and prefer just when they provide the needed signal.

Completion criterion: getComptroller() accepts the target chain and all selected optimized deployment scripts build without changing unrelated files.

4. Separate Broadcaster from Administrator

Establish both roles explicitly:

• Broadcaster: pays gas and signs deployment transactions.
• Comptroller admin: controls protocol administration after deployment and is selected by getAdmin(); it is not automatically the broadcaster.

Inspect getAdmin(), including default EOA and shared multisig branches. Check whether a selected Safe address has runtime code and is controllable on the target chain. Do not select an undeployed or uncontrolled Safe merely to match another chain.

Use the repository-root .env, which just setup symlinks into packages. For mnemonic mode, instruct the user to set only:

MNEMONIC="..."

Remove ETH_FROM; it takes precedence. BaseScript derives mnemonic index 0 and remembers that key for Forge.

For a raw private key, instruct the user to set:

ETH_FROM="0xBROADCASTER"
PRIVATE_KEY="0xPRIVATE_KEY"

Export the file in the user's shell and pass --private-key "$PRIVATE_KEY" to every Forge/just deployment command. The Solidity scripts read ETH_FROM, not PRIVATE_KEY; the Forge CLI needs the latter to sign. Prefer a Foundry keystore or interactive signer when the user wants to avoid a raw key in process arguments.

Ensure .env is ignored and mode 600. Derive or obtain the public broadcaster address without displaying secret material, then verify sufficient native-token funding.

Completion criterion: the user confirms the intended Comptroller admin, and the selected broadcaster is available to Forge and funded.

5. Deploy the Comptroller Prerequisite

Every selected protocol constructor obtains and validates a Comptroller. Read getComptroller() and probe its runtime code before explicitly choosing one of three workflows:

• Vanity deployment: freshly deploy the exact legacy vanity proxy at 0x0000008ABbFf7a84a2fE09f9A9b74D3BC2072399. Run utils::deploy-vanity, which targets DeployVanityComptrollerProxy.s.sol. It must reject a chain whose configured Comptroller differs from the vanity address or whose vanity address already has code. The script bootstraps the canonical v1.0 implementation and proxy, initializes the proxy with a transient atomic coordinator, deploys and upgrades to the current implementation, transfers administration to getAdmin(), and verifies state in one top-level broadcaster transaction.
• Bespoke deployment: freshly deploy the current implementation and a configured chain-specific proxy. Run utils::deploy-bespoke, which targets DeployBespokeComptrollerProxy.s.sol. It must reject a chain configured for the vanity address, reject an already-deployed configured proxy, and require its predicted proxy to equal getComptroller(). It uses the canonical CREATE2 factory when available and falls back to the atomic CREATE coordinator when the factory is unavailable; either path produces one top-level coordinator CREATE transaction. The transaction type alone therefore does not identify the selected branch: inspect its internal deployments. The proxy constructor initialization must bind the final admin and configuration.
• Use existing: validate and reuse the configured already-deployed proxy, whether vanity or bespoke. Run utils::use-existing, which targets ValidateExistingComptroller.s.sol. It must validate the implementation, admin, oracle, and minimal interface without broadcasting a Comptroller transaction.

At the root, use the matching explicit stack recipe: deploy-protocol-stack-vanity, deploy-protocol-stack-bespoke, or deploy-protocol-stack-existing. Do not restore an ambiguous recipe or infer "existing" merely from whether the configured address is vanity.

The exact legacy vanity address commits to proxy creation code with empty initialization calldata. Combining deployment, initialization, upgrade, admin transfer, and verification in one transaction removes the initialization gap inside that transaction, but does not make the public payload safe from predeployment: an observer can deploy the same empty proxy first and initialize it with different authority. Before a vanity broadcast, confirm immediately that the address has no code, disclose this residual risk, and obtain explicit authorization for the vanity workflow. If the vanity proxy has already been deployed incorrectly, use a constructor-initialized bespoke deployment instead. Do not treat private submission or --slow as eliminating the risk.

For a bespoke deployment, require the proxy constructor data to call initialize with the final intended admin and configuration. The implementation deployment, proxy deployment, constructor delegate call, and state checks must execute in one top-level broadcaster transaction. A copied deployment is harmless because changing the admin changes the initcode and address. For CREATE2, pin and record a release- and chain-bound proxy salt. For the CREATE fallback, record the broadcaster nonce and resulting coordinator, implementation, and proxy addresses. Add the resulting proxy address to getComptroller(). The implementation may retain its canonical release salt when its bytecode and constructor arguments match. On the CREATE2 branch, the coordinator address may vary with the broadcaster nonce, but the implementation and proxy addresses derive from the canonical factory, salts, and initcode and are not nonce-dependent.

Simulate the selected explicit workflow with the optimized profile and exact arguments. For either fresh deployment, omit only --broadcast and verification flags; inspect the trace, predicted addresses, exactly one transaction, admin, initializer values, and gas. For use-existing, confirm that validation produces no transaction. After explicit broadcast authorization, rerun the selected root stack recipe.

For either fresh deployment, confirm the single successful Comptroller receipt and runtime code at the expected proxy before continuing. If the transaction reverts, all state changes revert and it can be investigated before retrying. For use-existing, record the validated proxy and implementation but no Comptroller transaction hash. Use --resume only for an applicable multi-transaction upgrade script whose earlier transactions succeeded.

Completion criterion: the expected Comptroller proxy has code, its implementation and initialized admin are correct, and the deployment transaction hashes are recorded.

6. Deploy the Selected Protocols

Default to this human-readable order after the Comptroller:

1. Lockup
2. Flow
3. Airdrops

There is no deployment-time dependency among these three in the current scripts; Airdrops factories do not require the fresh Lockup address. Reconfirm this from constructors whenever the scripts change.

Use each package's optimized deterministic recipe. Simulate each exact command first, then request broadcast authority once the simulations and total funding estimate are available. Broadcast sequentially with --slow. Include --private-key "$PRIVATE_KEY" only for raw-key mode.

Do not use root deploy-all without inspecting its current package set and order. It may include Bob and may run Utils after dependent protocols, which expands scope and violates the prerequisite order.

On a partial run, use that script's broadcast artifact with --resume. A descriptor may already exist even while its chain map still returns zero, so rerunning from the beginning can revert at CREATE2.

Completion criterion: every selected script has only successful receipts and every reported contract address has runtime code on the target chain.

7. Verify and Reconcile

Extract contract names, addresses, and transaction hashes from each package's broadcast/<Script>.s.sol/<CHAIN_ID>/run-latest.json. Verify on-chain:

• Comptroller proxy implementation, admin, oracle, and initialized fee settings.
• Each protocol or factory's Comptroller address.
• Flow and Lockup NFT descriptor relationships.
• Runtime code at every deployment address.

Add the deployed descriptor addresses to the Flow and Lockup address maps after the first successful deployment, then run narrow formatting/build validation. Keep these bookkeeping edits distinct from already-broadcast bytecode inputs.

When contract verification is requested, use $explorer-contract-verification with the Atlas-resolved explorer. Treat verification as a separate external write and report partial verifier failures without implying deployment failure.

Report the resolved chain/provider route, repository revision, broadcaster address, admin address, deployed contract table, transaction hashes, explorer links, validations, verification status, and any incomplete history or manual follow-up. Never include secret values.