Module token_batch 

Batch (255): several token instructions in one CPI.

The p-token build of SPL Token accepts one instruction whose data is [255] followed by, per inner instruction, a two-byte header ([account count][data length]) and that instruction’s data; the accounts of every inner instruction follow each other in the account list. One CPI then pays one invocation overhead instead of one per instruction.

TokenBatch is a TokenSink: builders in crate::token or crate::token_2022_ix are appended with TokenBatch::push through the same emit that its invoke() uses, so the batched bytes and the single-CPI bytes are identical by construction. The buffers are const generic and live on the stack; nothing is allocated. Each inner instruction must fit the wire format’s 255-byte data and 255-account limits as well as the batch’s capacities. A larger payload must be sent separately.

One account may appear in several inner instructions (a transfer there and back); the batch is sent through crate::cpi::invoke_signed_batch_with_bounds, which keeps every per-meta check of the default tier and only waives the refusal of one account behind two writable metas, since for a batch that repeat is the contract rather than the footgun. Within an inner instruction, repeated writable accounts are still refused with AccountBorrowFailed (including self-transfers).

Both programs accepted a batch of two TransferCheckeds on devnet on 2026-09-28 (SPL Token at 2,472 CU for the whole instruction, Token-2022 at 5,575); the Token-2022 build bundled with Mollusk 0.15 refuses discriminator 255 with InvalidInstruction, so a local test cannot stand in for the cluster on that point.

Structs

TokenBatch

A stack-resident batch of token instructions.

Constants

BATCH_DISCRIMINATOR

The Batch discriminator.

BATCH_INSTRUCTION_HEADER_LEN

Bytes of header in front of each inner instruction: its account count and its data length, one byte each.