Changes in version 0.1.1.9000 - mizu_as_arrow() now also accepts an atomic vector, exported as a single Arrow array; a matrix exports flat, in column-major order. - mizu_as_arrow() exports the output of a mizu_map() collected with .collect = "view" without copying, and likewise received raw vectors. - mizu_as_arrow() now gives the same Arrow data whether or not a column is still backed by shared memory: NaN in a copied double column stays a value (it became a null), and a copied character column is a large_string (it was a string). - Faster re-sending of Arrow data obtained from mizu_as_arrow(): columns left unchanged are no longer copied. - mizu_submit_batch() failing mid-batch now attaches the handles of the tasks already submitted to the raised condition as accepted; they run and stay collectible. - Faster mizu_map() collection and lower shared-memory use for a (non-prepared) map collected by value. - Faster mizu_map() on small and medium inputs. - Idle workers now release the shared memory of finished maps promptly. - Streaming mizu_map() with a mizu_call() spec no longer retains completed chunk regions on R workers. - Native mizu_map() character-template results are validated and flattened per worker batch; a type or length mismatch now fails fast with its element index. Spec maps retain submitter-side validation. - mizu_cancel() on a collected map handle no longer disrupts a later run of the same prepared map. - New mizu_recv_batch_async(): the promise counterpart of mizu_recv_batch() — one promise resolves with a list of up to n messages, the high-throughput async receive. - Fixed pool shutdown often waiting an extra 50 ms after its workers had exited cleanly. - Fixes mizu_send() of an Arrow table failing with "attempt to apply non-function" once a mizu_map() result had been collected in the session (#19). - Faster message delivery for the async channel verbs (mizu_recv_async(), mizu_recv_batch_async(), mizu_recv_handler()). - Faster sends of large payloads and of Arrow tables with missing values or factor columns, and faster first reads of large zero-copy views on Linux. - Faster by-value sends of Arrow tables, including nested tables and sends that cannot use a shared-memory frame region. - Faster resolution of pool task promises with mizu_collect_async() and as.promise(), particularly when many tasks are awaiting collection. Changes in version 0.1.1 (2026-11-08) - Faster sends of ALTREP atomic vectors to a Python peer. Changes in version 0.1.0 (2026-11-04) Initial release. - mizu_channel(): lock-free SPSC channels between R processes — or between R and a Python peer — over POSIX shared memory or Win32 file mappings, with sentinel outcomes for a full ring, timeout, close, or peer death. - mizu_pool(): work-stealing task pools over spawned workers — mizu_submit() / mizu_collect(), batch submission, mizu_collect_any() / mizu_collect_all(), worker retirement and respawn, nested submission, and classed remote errors. - mizu_map(): a parallel map — streaming and prepared maps, .template output areas with a zero-copy .collect = "view", and reproducible per-element random streams (.seed). - Zero-copy vectors and data frames — received as ALTREP views, and a view sent on again crosses by reference. - Zero-copy Arrow interchange — 'arrow' tables and record batches travel without copying, converting back with mizu_as_arrow(). - Python interop — channels and pools mix with Python processes running pymizu: mizu_py_launcher() spawns peers and workers, mizu_call() task specs drive them, and data frames cross with full type fidelity. - Event-driven async — mizu_recv_async(), mizu_recv_handler(), mizu_collect_async(), and as.promise() methods for channels, tasks, and maps over later and promises, ready for Shiny. - Agent skill — an agent skill for AI coding assistants ships in the installed package at skills/r-mizu, picked up automatically by btw and similar tools. - Housekeeping — mizu_prune() reclaims shared memory orphaned by dead processes; mizu_pool_dump() introspects a running or hung pool.