mirai— agent skill

mirai is a minimalist R framework for async, parallel, and distributed evaluation, built on nanonext.

Core Principle: Explicit Dependency Passing

mirai evaluates expressions in a clean environment on a daemon process. Nothing from the calling environment is available unless passed explicitly — this is the #1 source of mistakes.

# WRONG: my_data and my_func are not available on the daemon
m <- mirai(my_func(my_data))

There are two ways to pass objects, and the names used must match the names referenced in the expression.

.args (recommended)

Objects in .args populate the expression's local evaluation environment — available directly by name inside the expression.

m <- mirai(my_func(my_data), .args = list(my_func = my_func, my_data = my_data))

... (dot-dot-dot)

Objects passed via ... are assigned to the daemon's global environment. Use this when objects need to be found by R's standard scoping rules (e.g., helper functions called by other functions).

m <- mirai(my_func(my_data), my_func = my_func, my_data = my_data)

Shortcut: pass the whole calling environment

# .args form — populates local eval env
process <- function(x, y) mirai(x + y, .args = environment())

# ... form — single unnamed environment, populates daemon global env
df_matrix <- function(x, y) mirai(as.matrix(rbind(x, y)), environment())

When to use which

Common Mistakes

Unqualified package functions

Daemons start with no user packages loaded. Same applies inside mirai_map() callbacks.

# WRONG: dplyr is not loaded on the daemon
m <- mirai(filter(df, x > 5), .args = list(df = my_df))

# CORRECT: namespace-qualify
m <- mirai(dplyr::filter(df, x > 5), .args = list(df = my_df))

# CORRECT: load inside the expression
m <- mirai({
  library(dplyr)
  filter(df, x > 5)
}, .args = list(df = my_df))

# CORRECT: pre-load on all daemons
everywhere(library(dplyr))
m <- mirai(filter(df, x > 5), .args = list(df = my_df))

Expecting results immediately

m$data accesses the value but may still be unresolved. Use m[] (or collect_mirai(m)) to block until done; use unresolved(m) for a non-blocking check.

m <- mirai(slow_computation())
result <- m[]                          # blocks until resolved
if (!unresolved(m)) result <- m$data   # non-blocking

Setting Up Daemons

No daemons required

mirai() works without calling daemons() first — it launches a transient background process per call. Setting up daemons is only needed for persistent pools of workers.

Local daemons

# Start 4 local daemon processes (with dispatcher, the default)
daemons(4)

# Direct connection (no dispatcher) — lower overhead, round-robin scheduling
daemons(4, dispatcher = FALSE)

# Concise programmatic statistics (vs. the richer status())
info()

# Reset (daemons otherwise persist for the session)
daemons(0)

Scoped daemons (auto-cleanup)

with(daemons(...), {...}) creates daemons and automatically cleans them up when the block exits.

with(daemons(4), {
  m <- mirai(expensive_task())
  m[]
})

Scoped compute profile switching

local_daemons() and with_daemons() switch the active compute profile to one that already exists — they do not create daemons.

daemons(4, .compute = "workers")

# Switch active profile for the duration of the calling function
my_func <- function() {
  local_daemons("workers")
  mirai(task())[]  # uses "workers" profile
}

# Switch active profile for a block
with_daemons("workers", {
  m <- mirai(task())
  m[]
})

Compute profiles (multiple independent pools)

daemons(4, .compute = "cpu")
daemons(2, .compute = "gpu")

m1 <- mirai(cpu_work(), .compute = "cpu")
m2 <- mirai(gpu_work(), .compute = "gpu")

Memory Backpressure (memory + try_mirai())

For high-throughput producers (Shiny, promises, ingest pipelines), use the memory argument to daemons() to cap the queued task payload at dispatcher (MB, metric). Pair it with try_mirai() so the host R thread never blocks on submission.

# 100 MB queue cap. mirai() blocks on submission once the queue is full.
daemons(4, memory = 100)

# try_mirai() returns NULL (invisibly) instead of blocking when the cap is hit.
m <- try_mirai(work(x), .args = list(x = x))
if (is.null(m)) {
  # backpressure: drop, retry later, or signal upstream
} else {
  # m is a regular mirai
}

# Inspect current and peak queue usage
status()$memory

memory requires dispatcher. Without dispatcher (or with memory = NULL), try_mirai() always returns a mirai.

mirai_map: Parallel Map

Requires daemons to be set. Maps .x element-wise over a function, distributing across daemons. Namespace-qualify any package functions used inside the callback (see Mistake 2).

daemons(4)

# Basic map — collect with []
results <- mirai_map(1:10, function(x) x^2)[]

# Constants via .args, helpers via ... (same passing rules as mirai())
results <- mirai_map(
  data_list,
  function(x, power) helper(x, power),
  .args = list(power = 3),
  helper = my_helper_func
)[]

# Flatten results to a vector
results <- mirai_map(1:10, sqrt)[.flat]

# Progress bar (requires cli package)
results <- mirai_map(1:100, slow_task)[.progress]

# Early stopping on error
results <- mirai_map(1:100, risky_task)[.stop]

# Combine options
results <- mirai_map(1:100, task)[.stop, .progress]

Mapping over multiple arguments (data frame rows)

# Each row becomes arguments to the function
params <- data.frame(mean = 1:5, sd = c(0.1, 0.5, 1, 2, 5))
results <- mirai_map(params, function(mean, sd) rnorm(100, mean, sd))[]

Process as completed (race_mirai)

race_mirai() returns the integer index of the first resolved mirai in a list (or 0L if empty). Useful when you want to handle results in completion order rather than submission order.

remaining <- mirai_map(jobs, run)
while (length(remaining) > 0) {
  idx <- race_mirai(remaining)
  process(remaining[[idx]]$data)
  remaining <- remaining[-idx]
}

everywhere: Pre-load State on All Daemons

daemons(4)

# Load packages on all daemons
everywhere(library(DBI))

# Set up persistent connections
everywhere(con <<- dbConnect(RSQLite::SQLite(), db_path), db_path = tempfile())

# Export objects to daemon global environment via ...
# The empty {} expression is intentional — the point is to export objects via ...
everywhere({}, api_key = my_key, config = my_config)

# .min = N forces a synchronization point: the call must complete on at least
# N daemons before subsequent mirai evaluations proceed. Useful when launching
# remote daemons that connect over time.
everywhere(library(arrow), .min = 4)

Error Handling

m <- mirai(stop("something went wrong"))
m[]

is_mirai_error(m$data)       # TRUE for execution errors
is_mirai_interrupt(m$data)   # TRUE for cancelled tasks
is_error_value(m$data)       # TRUE for any error/interrupt/timeout

m$data$message               # Error message
m$data$stack.trace           # Full stack trace
m$data$condition.class       # Original error classes

# Timeouts (requires dispatcher)
m <- mirai(Sys.sleep(60), .timeout = 5000)  # 5-second timeout

# Cancellation (requires dispatcher)
m <- mirai(long_running_task())
stop_mirai(m)

Shiny / Promises Integration

ExtendedTask pattern

library(shiny)
library(bslib)
library(mirai)

daemons(4)
onStop(function() daemons(0))

ui <- page_fluid(
  input_task_button("run", "Run Analysis"),
  plotOutput("result")
)

server <- function(input, output, session) {
  task <- ExtendedTask$new(
    function(n) mirai(rnorm(n), .args = list(n = n))
  ) |> bind_task_button("run")

  observeEvent(input$run, task$invoke(input$n))
  output$result <- renderPlot(hist(task$result()))
}

For high-traffic apps, set daemons(4, memory = ...) and submit with try_mirai() to apply backpressure without stalling the Shiny event loop.

Promise piping

library(promises)
mirai({Sys.sleep(1); "done"}) %...>% cat()

Remote / Distributed Computing

SSH (direct connection)

daemons(
  url = host_url(tls = TRUE),
  remote = ssh_config(c("ssh://user@node1", "ssh://user@node2"))
)

SSH (tunnelled, for firewalled environments)

daemons(
  n = 4,
  url = local_url(tcp = TRUE),
  remote = ssh_config("ssh://user@node1", tunnel = TRUE)
)

HPC cluster (Slurm/SGE/PBS/LSF)

daemons(
  n = 1,
  url = host_url(),
  remote = cluster_config(
    command = "sbatch",
    options = "#SBATCH --job-name=mirai\n#SBATCH --mem=8G\n#SBATCH --array=1-50",
    rscript = file.path(R.home("bin"), "Rscript")
  )
)

HTTP launcher (e.g., Posit Workbench)

daemons(n = 2, url = host_url(), remote = http_config())

Converting from future

The key conversion step: identify all objects the expression uses from the calling environment and pass them explicitly via .args or ....

Converting from parallel

Drop-in replacement via make_cluster

For code that already uses the parallel package extensively, make_cluster() provides a drop-in backend:

cl <- mirai::make_cluster(4)
parallel::parLapply(cl, 1:100, my_func)
mirai::stop_cluster(cl)

# R >= 4.5: native integration
cl <- parallel::makeCluster(4, type = "MIRAI")

Random Number Generation

# Default: L'Ecuyer-CMRG stream per daemon (statistically safe, non-reproducible)
daemons(4)

# Reproducible: L'Ecuyer-CMRG stream per mirai call.
# Results are the same regardless of daemon count or scheduling.
daemons(4, seed = 42)

Debugging

# Synchronous mode — runs in the host process, supports browser()
daemons(sync = TRUE)
m <- mirai({
  browser()
  result <- tricky_function(x)
  result
}, .args = list(tricky_function = tricky_function, x = my_x))
daemons(0)

# Capture daemon stdout/stderr
daemons(4, output = TRUE)

Advanced Pattern: Nested Parallelism

Inside daemon callbacks (e.g., mirai_map), use local_url() + launch_local() instead of daemons(n) to avoid conflicting with the outer daemon pool.

mirai_map(1:10, function(x) {
  daemons(url = local_url())
  launch_local(2)
  result <- mirai_map(1:5, function(y, x) x * y, .args = list(x = x))[]
  daemons(0)
  result
})[]