Skip to contents

Changelog

Source: NEWS.md

fmisc 0.3.0

Breaking changes

  • get_flir_rules() is removed, so calling it errors and listing fmisc in the from-package field of flir/config.yml no longer loads any rules. The bundled rules themselves, deprecated-sample-n and deprecated-sample-frac (which rewrote dplyr’s superseded sample_n() and sample_frac()), are gone with no replacement: write the rules into your own project instead, or drop the from-package entry. flir is no longer in Suggests, and the using-fmisc vignette that documented the integration is removed, so vignette("using-fmisc") errors (#62).

New features

  • agent_run() runs the Claude Code CLI (claude -p) once, read-only, in a directory and returns its answer, which must match a JSON Schema given as a string, as an fmisc_agent_result. The agent gets only the Read, Glob and Grep tools, no repo or user customisations, and the prompt goes on stdin. The model, turn and budget caps are fixed, and timeout bounds the whole run and kills the process tree. It uses the CLI’s existing login, intended to be a Claude subscription, and removes API-key, gateway and cloud-provider variables (ANTHROPIC_API_KEY, ANTHROPIC_AUTH_TOKEN, ANTHROPIC_BASE_URL, CLAUDE_CODE_USE_*, …) from the child environment so a run never bills elsewhere by accident. Fixed Read deny rules keep common secret files (.env*, .Renviron*, *.pem, *.key, SSH keys, .netrc, .git-credentials) away from the agent; that is defense in depth, not a sandbox. Errors are classed (fmisc_agent_error and subclasses). Needs processx and jsonlite (Suggests) and Claude Code CLI >= 2.1.259; not supported on Windows yet (#75).

  • ci_failures() fetches the latest failed GitHub Actions run of the current branch and returns one row per finding (job, os, step, kind, file, line, message, sha, url). It recognises R CMD check ERRORs, WARNINGs and NOTEs, testthat failures and errors, lintr annotations and air format-check failures; a failed job it cannot explain is kept as kind "unparsed" with the raw log of the failing step, and a timed-out or cancelled job as kind "timed_out" or "cancelled". gh and gert join Imports; both were already installed as dependencies of usethis (#76).

fmisc 0.2.0

Refactoring

  • with_cache() is now a thin facade over memoise::memoise() backed by cachem::cache_mem(). .max_size maps to max_n (LRU eviction) and .ttl to max_age; the hand-rolled environment backend and the deparse()-based key scheme are gone (memoise hashes internally). The .key argument is removed: custom cache keys are not supported by memoise, so there is no drop-in replacement. cache_info() now reports backend = "cachem" with size, max_n, and max_age (the old "env"/"memoise" backends and keys element are gone). memoise and cachem moved from Suggests to Imports.

New features

  • with_retry() gains two conveniences: .on_error now also accepts a character vector of condition classes to retry on (e.g. c("http_error_503")), mirroring httr2’s is_transient idiom; and a .message = TRUE toggle emits a message() before each retry.
  • undecorate() gains a depth argument (default 1); depth = Inf unwraps the full decorator chain. New vignette vignette("decorators") compares the decorators against purrr::insistently(), memoise + cachem, and ratelimitr, and documents the deliberate skip of burst-tolerant rate limiting (use httr2::req_throttle() for HTTP burst needs).
  • Decorated functions now record each applied decorator (name + key parameters) in an "fmisc_stack" attribute, accumulated as wrappers compose via decorate() or the native pipe. print() on an "fmisc_decorated" object shows the wrapper stack outermost-first.

Function Decorators (Function-Operator Toolkit)

  • New composable “function operators” for cross-cutting concerns. Each with_*() takes a function as its first argument and returns a wrapped function with class "fmisc_decorated", making them curry- and pipe-friendly:
    • with_retry() — retry with exponential backoff, jitter, .on_error predicate, .on_retry callback, and .max_delay cap. Raises "fmisc_retry_exhausted" on final failure while preserving the original condition’s classes.
    • with_timing() — wall-clock timing via proc.time(), reported as cli message, "elapsed" attribute, or callback; reports on both success and failure.
    • with_logging() — call/success/error events via a default cli logger or a custom .logger(event, data).
    • with_cache() — memoisation facade over memoise::memoise() + cachem::cache_mem() with LRU eviction (.max_size) and time-to-live (.ttl). Companions: cache_clear(), cache_info().
    • with_rate_limit() — throttle to n calls per rolling period seconds; either wait or raise "fmisc_rate_limit_exceeded".
  • Composition helpers: decorate() composes unary decorators left-to-right; is_decorated() and undecorate() inspect the wrapper chain.

Chunking and RAM Management Framework

  • New chunking and RAM management framework for processing large datasets that exceed available memory:
  • Features include:
    • Automatic chunk size calculation based on data size and RAM constraints
    • Cross-platform RAM monitoring (Windows, Linux, macOS)
    • Disk spillover with temporary file management
    • Flexible result combination via custom combine_fn parameter
    • Progress monitoring with optional verbose output

Smart Parallel Computing Framework

  • Added comprehensive parallel computing framework that automatically selects the best backend based on OS and available packages
  • New functions:
  • Features:
    • OS-aware backend selection (fork-based for Unix, socket-based for Windows)
    • Support for multiple backends: mclapply, parLapply, doParallel, doMC, furrr, future
    • Automatic fallback to sequential processing
    • Resource cleanup with on.exit() guarantee
    • Windows support with automatic clusterExport()
    • Input validation and robust error handling

Bug fixes

  • chunk_processor()$get_results() now correctly combines chunks in chronological order, combining disk-saved chunks before in-memory chunks.

  • create_chunk_iterator() now properly excludes list objects from vector validation, preventing incorrect handling of list inputs.