Module mooncake 

Mooncake JSONL primitives.

This module is producer- and consumer-agnostic: it defines the row schema, the block-hash-to-id mapping, the token-block hashing helper, and the JSONL writer. Workload-specific orchestration such as scheduling, tokenization, and parsing lives elsewhere.

The MooncakeRow schema deliberately matches the externally-authored Mooncake trace format: timestamp and delay are f64 milliseconds, and input_length/output_length/timestamp/delay accept the upstream aliases (input_tokens, output_tokens, created_time, delay_ms) on deserialization. Serialization emits the canonical names.

Structs

AgenticMooncakeRow

One row of an agentic Mooncake replay trace.

AgenticToolEvent

Harness tool span attributed to the LLM request that consumed it.

MooncakeJsonlWriter

JSONL writer for Mooncake rows plus an optional sidecar stream.

MooncakeRow

One row of a Mooncake replay trace.

RollingHashIdMapper

Maps sequence-aware block hashes to compact, stable u64 ids.

WriterStats

Counters for what a MooncakeJsonlWriter has emitted.

Functions

hash_token_blocks

Token-block hashing helper for the Mooncake replay schema.

ids_for_sequence_hashes

Map stable sequence hashes to compact Mooncake IDs with a shared mapper.

require_positive

Sentinel used by callers that want to bail when neither block_size nor worker count is allowed to be zero. Producers may also enforce this on their own configuration types.

sequence_hashes_for_tokens

Derive the sequence-aware block hashes recorded by Dynamo request traces.

try_hash_token_blocks

Fallible token-block hashing helper for callers that want to surface invalid block-size or request-shape errors.

write_empty_files

Create both files empty (touch-equivalent), preserving directory creation semantics for callers that want a “no rows produced” outcome to still emit well-formed (empty) JSONL files.