Skip to content

Releases

irit 0.6.1

Choose a tag to compare

@github-actions github-actions released this 15 Nov 14:08

SemiBin2 multi-sample binning, single-end reads for contig coverage, and cooccur tested on real
data (it did not find TM7's known host).

irit-mcp

  • SemiBin2 multi-sample binning: semibin2_concatenate (SemiBin2's concatenate_fasta, each sample
    named per file) and multi on semibin2 (multi_easy_bin); bins per sample and in one folder for
    CheckM2 and galah. On three 0.5 Gb Crystal Geyser time-series samples it fitted in 8 GB (4.3 GiB,
    48 min) and binned as well as the pretrained global model, which takes about 20 s per sample.
    The real-tools workflow runs it on the mock.
  • coverm_contig takes single-end reads (single), for runs ENA serves unpaired.

cooccur

  • Tested on real data (write-up): on 128 supragingival-plaque and
    136 tongue HMP MetaPhlAn2 profiles, Saccharibacteria (TM7) and their cultured host Actinomyces
    odontolyticus
    came out negatively associated (rho −0.25 in plaque), not positively; the
    shuffled-sample control was nominal (4.86–5.00 % of null pairs at p ≤ 0.05). Co-occurrence alone
    doesn't find this host.
  • A passes column in the edge list (the printed q is rounded), and a warning when the share of
    absent genomes differs a lot between samples, which pushes log-ratio associations around.

Download

Terminal (Linux or macOS), installs into ~/.local/bin without sudo:

curl -fsSL https://raw.githubusercontent.com/dzulfikar08/irit/main/scripts/install.sh | IRIT_VERSION=v0.6.1 sh
  • Linux command-line tools (static, any distribution): x86_64, aarch64
  • macOS command-line tools (Apple Silicon and Intel, Developer ID signed and notarised): irit-0.6.1-universal-macos.tar.gz
  • msascope for macOS (signed, notarised): msascope-0.6.1.dmg. Open it and drag msascope to Applications.
  • msascope for Linux (glibc 2.39 or newer: Ubuntu 24.04, Debian 13, Fedora 40 or later): x86_64, aarch64
  • Windows (sheetgen, leanqc, isotable, dblock): irit-0.6.1-x86_64-windows.zip. Not code-signed: SmartScreen may warn the first time you run them.

Check downloads with sha256sum -c SHA256SUMS --ignore-missing (macOS: shasum -a 256 -c SHA256SUMS --ignore-missing). More, including checking the macOS signatures: https://dzulfikar08.github.io/irit/install/

irit 0.6.0

Choose a tag to compare

@github-actions github-actions released this 15 Nov 05:10

New: the rest of a metagenome-to-MAG study on a laptop: SemiBin2 binning, galah dereplication, a
genome × sample abundance table, GTDB tree placement for genomes the ANI screen can't name
(gtdbprep place), and co-occurrence statistics (cooccur, a new tool). sheetgen verify checks
read files against their MD5s before analysis, and fieldsheet is an offline phone worksheet for
sample metadata. A replay now tolerates bins a binner names differently after a changed assembly.

irit-mcp

  • semibin2 (SemiBin 2.6.0, single sample): pretrained environment models from CoverM's depth
    table, or self-supervised training from BAMs. On the Crystal Geyser sample the pretrained models
    took about a minute in 0.9–1.5 GiB and gave 25–27 bins against MetaBAT2's 23, without better
    quality; self-supervised training was killed at 8 GB.
  • galah (0.6.3): dereplicate bins from several samples at 95 % ANI, choosing representatives
    by CheckM2 quality. 48 bins made 24 clusters in under a second.
  • coverm_abundance: several samples mapped to a set of genomes, one genome × sample table
    (relative abundance, mean depth, covered fraction, read count) for cooccur.
  • reads_verify, cooccur and gtdbprep_place: irit-mcp tools for sheetgen verify, cooccur
    and gtdbprep place, with a stdio test that runs the first two with the real binaries and replays.
  • coverm_contig gains bam_cache_dir, to keep each sample's sorted BAM for SemiBin2.

sheetgen

  • sheetgen verify checks read files before analysis: full gzip decompression (every member,
    CRC-32 and length) and FASTQ structure, with MD5 and SHA-256 in the same pass, compared with an
    md5sum file (--md5) or ENA's published MD5s and sizes (--ena, opt-in network). Each file is
    ok, mismatch, corrupt (with the byte offset, record and line) or not listed; exit 1 on any
    problem; --json for irit-mcp. Constant memory; 163–174 MB/s on a 568 MB file on a loaded laptop.

cooccur (new)

  • Which genomes rise and fall together across samples, from a coverm genome table: CLR
    proportionality rho (Lovell et al. 2015) or Spearman on CLR values, seeded sequential permutation
    p-values and Benjamini-Hochberg q-values, CPR / non-CPR labels from a GTDB-Tk summary. rho and
    Spearman match the R package propr 5.1.8 and R. The permutation null's calibration and the power
    were measured by simulation; 500 genomes × 60 samples run in about 6 s and 20 MiB.

gtdbprep

  • gtdbprep place: GTDB-Tk tree placement on the placement-only data dblock fetches
    (identify, align --skip_gtdb_refs, classify without the ANI screen), with gPlace as the placer
    under a memory cap. It checks that gPlace wrote every placement, skips genomes gtdbprep ani
    assigned (--skip-assigned), and writes place.tsv and a place.json run record. On the three
    Crystal Geyser bins the ANI screen couldn't assign it gave the same classifications as the earlier
    manual run, 5.6 GiB peak.

fieldsheet (new)

  • An offline web app, installable on a phone, for recording sample metadata in the field:
    projects, automatic sample IDs, range checks, GPS with a coarse or off option, photos kept on the
    device, CSV export with sample first for sheetgen and isotable, JSON backup and import, English
    and Bahasa Indonesia. At https://dzulfikar08.github.io/irit/fieldsheet/.

Download

Terminal (Linux or macOS), installs into ~/.local/bin without sudo:

curl -fsSL https://raw.githubusercontent.com/dzulfikar08/irit/main/scripts/install.sh | IRIT_VERSION=v0.6.0 sh
  • Linux command-line tools (static, any distribution): x86_64, aarch64
  • macOS command-line tools (Apple Silicon and Intel, Developer ID signed and notarised): irit-0.6.0-universal-macos.tar.gz
  • msascope for macOS (signed, notarised): msascope-0.6.0.dmg. Open it and drag msascope to Applications.
  • msascope for Linux (glibc 2.39 or newer: Ubuntu 24.04, Debian 13, Fedora 40 or later): x86_64, aarch64
  • Windows (sheetgen, leanqc, isotable, dblock): irit-0.6.0-x86_64-windows.zip. Not code-signed: SmartScreen may warn the first time you run them.

Check downloads with sha256sum -c SHA256SUMS --ignore-missing (macOS: shasum -a 256 -c SHA256SUMS --ignore-missing). More, including checking the macOS signatures: https://dzulfikar08.github.io/irit/install/

irit 0.5.0

Choose a tag to compare

@github-actions github-actions released this 14 Nov 09:32

New in irit-mcp: assembly_stats, coverm_genome (depth by read mapping), kleborate and rasusa
(subsampling), and a warning for Flye contigs that hold a small plasmid several times. dblock
downloads in parallel (fetch --jobs), and msascope reads gzip faster. Fixes from a second AI-agent
paper redo (Amena et al. 2026, E. coli and Salmonella Kentucky, MRA e01429-25) and from three
more nanopore isolates.

irit-mcp

  • assembly_stats: QUAST-style assembly statistics (contigs, total length, largest contig, N50,
    L50, N90, L90, GC, Ns per 100 kbp, contigs >= 1 kb; contigs of 500 bp or more, QUAST's default),
    computed by isotable and recorded like any call.
  • coverm_genome (CoverM 0.8.0): an isolate's depth by read mapping, and the fraction of the
    assembly covered. Warns when Shovill's depth estimate differs by more than 1.5×.
  • kleborate (Kleborate 3.2.4): Klebsiella typing (ST, virulence and resistance scores,
    ybt/clb/iuc/iro/rmpADC, K/O loci with Kaptive's confidence). Its bundled databases are hashed into
    the record; warns about assemblies with no row or outside the KpSC.
  • rasusa (Rasusa 5.1.0): subsample reads to a depth, number of bases, reads or a fraction, seed
    1 by default (byte-identical output, so replayed and compared). Flye's description and guide now
    say that asm_coverage only limits its first stage: a 5.7 Mb K. pneumoniae at about 450x went
    over 8 GB in Flye's consensus stage, so deep long reads are subsampled first.
  • Outputs isotable won't read are allowed: mlst, AMRFinderPlus and CoverM genome outputs not
    named <sample>.<tool>.tsv now run with a warning instead of being refused, for side analyses
    that shouldn't join the table (found by the second paper redo).
  • Flye: plasmids closed as several copies are flagged. A circular contig up to 300 kb with more
    than half of its 21-mers repeated gets a warning (and the copy count when it is a clean
    multiple): AF38's 6,141 bp blaOXA-232 plasmid came out as three copies in an 18,377 bp circle,
    so AMRFinderPlus listed the gene three times. Flye's description mentions meta for small
    plasmids.
  • A tool whose bundled data (mlst's schemes, Kleborate's databases) can't be found now gets a
    warning that the record identifies it by version only.

isotable

  • isotable assembly-stats, and largest contig, L50, N90, Ns per 100 kbp, contigs >= 1 kb and
    all contigs as table columns (from assemblies and QUAST reports).
  • Reads CoverM genome and Kleborate 3 tables. Shows Shovill's depth estimate next to the mapped
    depth; --min-depth (default 20) grades the mapped depth (warn below it, fail below half), and a
    warning says when the two differ by more than 1.5×. New requirements depth and kleborate.

msascope

  • Faster gzip loading: gzip input is decompressed on its own thread while the parser encodes,
    BGZF (bgzip) files are decompressed 32 members at a time in parallel, and flate2 uses its
    zlib-rs backend. On a 20,000 × 30,000 alignment, measured on a busy machine: BGZF 5.4 → 4.4 s
    and one-member gzip 7.2 → 6.3 s (medians of 9), fastest runs 3.0 → 1.4 s and 3.1 → 2.6 s; peak
    memory unchanged. Tests cover plain, multi-member and BGZF gzip, progress and cancel.

  • Read QC named after the sheet's read files joins their sample: with --sheet, a leanqc or
    FastQC report of reads/SRR30733615_1.fastq.gz joins the sample the sheet lists that file for,
    instead of getting a row of its own (found by the second paper redo).

  • --min-contig (and irit-mcp's min_contig): the contig cut-off for assembly statistics,
    500 bp by default as QUAST; 200 bp compares with papers that report GenBank's assemblies.

dblock

  • Downloads side by side: fetch --jobs N (default 4) downloads the parts or files of an entry,
    and several databases, under one shared limit; --jobs 1 fetches one thing at a time as before.
    Measured: placement-lite (11 parts, 175 MB) 133 s → 68 s, AMRFinderPlus (38 files, 126 MB)
    414 s → 161 s. When a part fails, the parts still downloading stop and keep their partial files
    for the next run; the error names the part that failed. Progress stays on one live line.

Download

Terminal (Linux or macOS), installs into ~/.local/bin without sudo:

curl -fsSL https://raw.githubusercontent.com/dzulfikar08/irit/main/scripts/install.sh | IRIT_VERSION=v0.5.0 sh
  • Linux command-line tools (static, any distribution): x86_64, aarch64
  • macOS command-line tools (Apple Silicon and Intel, Developer ID signed and notarised): irit-0.5.0-universal-macos.tar.gz
  • msascope for macOS (signed, notarised): msascope-0.5.0.dmg. Open it and drag msascope to Applications.
  • msascope for Linux (glibc 2.39 or newer: Ubuntu 24.04, Debian 13, Fedora 40 or later): x86_64, aarch64
  • Windows (sheetgen, leanqc, isotable, dblock): irit-0.5.0-x86_64-windows.zip. Not code-signed: SmartScreen may warn the first time you run them.

Check downloads with sha256sum -c SHA256SUMS --ignore-missing (macOS: shasum -a 256 -c SHA256SUMS --ignore-missing). More, including checking the macOS signatures: https://dzulfikar08.github.io/irit/install/

irit 0.4.0

Choose a tag to compare

@github-actions github-actions released this 13 Nov 18:06

New: seqplan (plan a sequencing run), long-read isolates (flye, medaka) and metagenomes to
MAGs (megahit, coverm_contig, metabat2) in irit-mcp, and msascope's menu bar and Mac App Store
preparation. Fixes from an AI agent reproducing Neumann et al. 2025 (K. pneumoniae from alfalfa
sprouts, MRA e01270-25) through irit-mcp alone (#14-#19), and from the depth experiment and the
nanopore run (#21-#25).

irit-mcp

  • A used run folder is refused (#14): irit-mcp won't start when <runs-dir>/<run-id>/ already
    holds a record. Reusing --run-id restarted call numbering and overwrote run.json, replay.sh,
    the RO-Crate and calls/NNN.*, losing the earlier session's record.
  • Databases fall back to dblock.lock (#16): with no path given and the variable unset,
    gtdbtk_classify_wf, gtdbprep_ani (data and index, now optional), checkm2, bakta and
    amrfinder use the one database of that kind locked in the working folder, and warnings names
    it; two candidates are refused.
  • Warnings from stderr (#18): CheckM2's scikit-learn InconsistentVersionWarning and
    AMRFinderPlus's WARNING: lines go into warnings. AMRFinderPlus from an unactivated conda
    environment gets CONDA_PREFIX set to it.
  • describe_run summarises by default (#19): tool, status, version, error, warnings and outputs
    per call; detail: true for the full records (170 kB for 26 calls), call_id for one.
  • amrfinder names Bakta's trimmed AMRFinderPlus copy instead of calling it an incomplete download;
    replay.sh says where a gtdbprep ani call's sketches come from on another machine (#19).

isotable

  • GTDB from the ANI screen (#15): classify/ani_screen/*.ani_summary.tsv (reference_taxonomy)
    counts as the GTDB classification, so a run where the screen assigned every genome no longer
    comes out incomplete under --require gtdb.
  • GC for Shovill assemblies (#15), over the same contigs of 500 bp or more as the contig count,
    length and N50 (QUAST's rule); the requirement help names the cutoff. On the paper's isolates:
    57.04 / 57.15 / 57.35 against the paper's 57.06 / 57.15 / 57.35.

dblock

  • No download without the prepare program (#17): fetch refuses a database whose prepare
    program (amrfinder_index) isn't on PATH before downloading it, instead of after; plan says
    it won't fetch, and notes databases whose tool isn't installed.

gtdbprep

  • The missing-sketch error says how much would be fetched and suggests --species-only (#19).

Fixes from the depth experiment and the nanopore run (#21-#25)

  • irit-mcp warns when Shovill's genome-size estimate collapses (#22). At 5-10× Shovill
    estimated 0.01-0.11 Mb for a 5.5 Mb genome, subsampled to it and assembled a fragment with
    exit 0. The server reads shovill.log and warns in three cases:

    • Shovill subsampled to a tiny size.
    • The assembly is over twice the estimate. The warning gives the real depth over the assembly,
      since Shovill's own figure is then too high (60× reported at 20×).
    • gsize was given and the assembly is under 70% of it.

    A separate warning comes when the depth over the assembly is under 15×.

  • gtdbprep ani screens the rest when skani can't sketch a genome (#23). Such a genome, e.g. a
    407 bp assembly, is listed as not screened (- markers; screened: false in irit-mcp) with a
    warning. Before, it failed the whole batch.

  • mlst: a warning when it chose the scheme itself and found no ST (#24). A fragmented
    K. pneumoniae got ecoli_achtman_4.

  • CheckM2: a warning above 20 genomes per call (#25). 48 peaked at 7.7 GiB on an 8 GB laptop.

  • Medaka's model file is recorded (#21). Medaka downloads a named model at run time, so the
    server hashes the file Medaka logged into the call's inputs, returns it as model_file, and
    replay.sh notes it.
    batch_size 50 peaked at 4.78 GiB and 25 at 3.66 GiB on the 6 Mb nanopore genome, against 7.36 GiB
    at Medaka's default 100. All three gave the same consensus and the smaller sizes were no slower,
    so the docs and the tool description recommend 50 or 25 on 8 GB machines.

  • isotable reads long-read QC (#21). For a nanopore-only isolate, leanqc's long-read numbers
    stand in for fastp:

    • long_reads and long_read_bases count as reads, and read_q_median counts as quality
      (warn below --min-read-q, default Q10). read_n50 is shown.
    • leanqc's short-read checks (per-position, per-tile, per-read quality and GC, read length) are
      not graded for long-read files, and a note lists them.
    • The nanopore run's ONT sample now passes instead of being incomplete.

seqplan (new)

  • Plan a sequencing run before sequencing: bases and read pairs per isolate at a target
    coverage, isolates per run, runs for N samples, and cost per isolate from a run price you give
    (no built-in prices). It pads each isolate by a barcode-imbalance margin (default 1.3) and counts
    only the usable part of a run (default 0.8 for Illumina). --barcodes caps isolates per run at
    the library kit's barcode count. Kits: MiSeq v2/v3/Micro/Nano, iSeq 100,
    NextSeq 500/550 mid/high, NextSeq 1000/2000 P1-P4, NovaSeq X 1.5B/10B/25B (Illumina
    specifications), plus MinION and Flongle R10.4.1 (reported typical yields, marked as such). There
    are 16 organism presets with NCBI's expected genome size. Every figure carries its source URL
    (seqplan/data/, --list-kits, --list-organisms). Example: K. pneumoniae 5.5 Mb at 50× fits
    29-33 isolates on a MiSeq v3 run and 67 on a NextSeq 2000 P1; at 25×, 59-67 and 134.
  • In the release archives (all platforms, including Windows) and the installer.

irit-mcp

  • seqplan tool: the same plan as JSON, so an agent answers from cited kit figures; it is told
    never to invent a run price.

irit-mcp: metagenomes to MAGs

  • megahit, coverm_contig and metabat2: MEGAHIT 1.2.9 assembly, CoverM 0.8.0 per-contig
    depth (methods: ["metabat"] writes MetaBAT2's depth table) and MetaBAT 2.18 binning, pinned and
    replayable like the isolate tools. Then checkm2 on the bins, gtdbprep_ani and isotable's MAG
    mode. Measured on the laptop in Metagenome to MAGs in 8 GB: a
    Crystal Geyser groundwater sample gave 23 bins (4 high*, 13 medium, 8 of them
    Patescibacteria), every step within an 8 GB cap.
  • MEGAHIT's memory defaults to 4 GB (MEGAHIT's own default, 90% of the machine, ignores memory
    caps); MetaBAT2's seed defaults to 1, so bins repeat. MetaBAT2's version comes from --help,
    in the record and in replay.sh.
  • A failed tool gets a hint when it ran out of memory: CheckM2's SIGKILLed DIAMOND step (use
    lowmem), or any tool killed by a signal.
  • The real-tools workflow runs the three on a two-genome mock metagenome, twice, and replays them.

msascope

  • Menu bar (macOS) with About, Acknowledgements, Hide, Quit, Open, Open Sample Alignment,
    Export, Close File, Close Window, Copy Selection and Select All. Quit, Open, Close Window, Export,
    Copy and Select All are key bindings (⌘ on macOS, Ctrl elsewhere), so the menus show them.
  • Closing the window: on macOS msascope keeps running and a click on its Dock icon opens a new
    window; elsewhere closing the last window quits. Before, it kept running without a window.
  • Finder opens files with msascope (Open With, a drop on the Dock icon): alignments and trees,
    registered as an alternative app, not the default one.
  • Acknowledgements: the licences of every dependency compiled into msascope, plus the Lucide
    icons and Catppuccin palettes, generated with cargo-about and shipped as Acknowledgements.html
    in msascope.app and the Linux archives. Neither shipped them before.
  • Mac App Store edition, prepared: it runs in the App Sandbox (security-scoped bookmarks for
    recent files, File ▸ Allow Access to Folder… for a tree next to the alignment, a clear message
    for a file it may not read); scripts/build-mas-pkg.sh and the msascope-mas.yml workflow build,
    validate and upload it. CI checks that msascope's dependencies stay permissively licensed
    (msascope/deny.toml) and that a sandboxed copy starts and can't read a file nobody chose.

irit-mcp

  • flye and medaka tools: nanopore (and PacBio) isolate assembly. Flye 2.9.6 (read_type
    nano-hq for R10.4.1 hac/sup reads, the default) and Medaka 2.2.2 (medaka_consensus), pinned
    and checked like Shovill: flags read from the release sources, inputs and outdir validated,
    link_as for CheckM2, GTDB-Tk and gtdbprep_ani, the assembly's contig count, length and N50
    recorded for replay. flye also returns each contig's depth and circularity. On one
    K. pneumoniae isolate, ONT reads alone gave the same ST, the same 49 AMRFinderPlus elements,
    the same CheckM2 completeness and the same GTDB species as Illumina, in an 8 GB cap: Flye 22 min
    (3.85 GiB), Medaka 9.4 min (7.36 GiB) (write-up).

isotable

  • Reads Flye's assembly.fasta (with flye.log beside it) and Medaka's consensus.fasta (with
    consensus_probs.hdf beside it) for contigs, length, N50 and GC; when a sample has both, the
    polished one counts and the draft is listed as skipped.

Download

Terminal (Linux or macOS), installs into ~/.local/bin without sudo:

curl -fsSL https://raw.githubusercontent.com/dzulfikar08/irit/main/scripts/install.sh | IRIT_VERSION=v0.4.0 sh
Read more

irit 0.3.5

Choose a tag to compare

@github-actions github-actions released this 12 Nov 17:04

irit-mcp

  • gtdbprep_ani tool (#4): gtdbprep ani through the server, recorded and replayable.
    source stays opt-in; the sketch cache, $GTDBTK_DATA_PATH and $GTDBPREP_ENV are resolved
    when the server starts and written into the command. It returns each genome's species,
    reference, ANI and AF, and <work>/gtdbtk_data, which gtdbtk_classify_wf takes as data_path
    (its skani-missing message now says so). On B. subtilis 168 with a warm cache: 15 s, then
    classify_wf 4.8 s, same species; the replay matched all 32 outputs. Since that folder is now an
    output of a recorded call, the run record shows where classify_wf's GTDBTK_DATA_PATH came from
    (#8).
  • GTDB-Tk's data is accounted for or flagged (#8): a skani/ database of up to 512 MB (the one
    gtdbprep_ani builds is a few MB) is hashed into gtdbtk_classify_wf's inputs under its real
    path, and the call warns when data_path is neither in a dblock store nor written by an earlier
    call of the run, since the record can't then say what the folder held.
  • Missing output folders are created (#2): trim/s1_R1.fastq.gz or an outdir of asm/s1
    no longer needs trim/ or asm/ to exist. Only folders of checked outputs are created (so inside
    the root), a file in the way is refused, the record lists them as dirs, and replay.sh runs
    mkdir -p on them, so a replay works from a clean folder.
  • Shovill link_as (#3): e.g. genomes/s1.fna becomes a relative link to
    <outdir>/contigs.fa, made before Shovill runs, recorded as the call's links and recreated by
    replay.sh (ln -sfn), so CheckM2, GTDB-Tk and gtdbprep_ani name the genome after the
    sample and replay works without hand-made copies.
  • Every Shovill output is volatile (#5): contigs.fa headers carry the date, and SKESA 2.5.1
    with 8 CPUs gave different assemblies from identical reads, as SPAdes did. replay.sh checks
    that they exist; the contig count, length and N50 are in the result and in isotable's row.
  • Replay explains outputs made from a volatile one (#5). The contig count, length and N50
    are now recorded with the Shovill call (assembly), and replay.sh prints the replayed
    assembly's beside them. Before each call it checks the inputs an earlier call wrote; when one
    differs (a new assembly), that call's different outputs (CheckM2's report, isotable's row) are
    reported as following it instead of failing the replay. Anything else that differs still fails.

isotable

  • Reads Shovill's contigs.fa (#7) when shovill.log is beside it, as the sample named after
    its folder: contigs, length and N50, counted as QUAST does (contigs of 500 bp or more), so a
    QUAST report on the same assembly agrees. The default isolate requirements (contigs, N50) can
    now be met without QUAST, also through irit-mcp.

leanqc

  • Reports are byte-identical between runs (#6): remaining_if_deduplicated_percent summed
    over duplication levels in a hash map's order, so its last digits could change from run to run
    and replay.sh's hash check of the JSON report failed. The levels are now summed in order.
    A test runs leanqc with 1 to 4 threads, eight times, on reads that push the duplication
    estimate past its tracking limit, and compares the JSON byte for byte (it fails on the old code).

gtdbprep

  • irit-gtdb-prep --version (and -V) works (#9); it said "unexpected argument". The release
    workflow now checks its version like the other binaries.

dblock

  • checkm2-db-v3 is 3.08 GB on disk (measured), not ~3.5 GB (#11).
  • A database folder moved or linked in from another store is reused with its .dblock.toml
    marker rewritten to its new place; the marker kept the old store's paths (#10). fetch now says
    what reuse checked (names and sizes) and that verify --deep re-hashes. Markers are written to
    a temporary file and renamed, so a marker hard-linked from the other store (a folder copied with
    cp -al, as on the laptop run) is replaced in this store only, not edited in both.

Download

Terminal (Linux or macOS), installs into ~/.local/bin without sudo:

curl -fsSL https://raw.githubusercontent.com/dzulfikar08/irit/main/scripts/install.sh | IRIT_VERSION=v0.3.5 sh
  • Linux command-line tools (static, any distribution): x86_64, aarch64
  • macOS command-line tools (Apple Silicon and Intel, Developer ID signed and notarised): irit-0.3.5-universal-macos.tar.gz
  • msascope for macOS (signed, notarised): msascope-0.3.5.dmg. Open it and drag msascope to Applications.
  • msascope for Linux (glibc 2.39 or newer: Ubuntu 24.04, Debian 13, Fedora 40 or later): x86_64, aarch64
  • Windows (sheetgen, leanqc, isotable, dblock): irit-0.3.5-x86_64-windows.zip. Not code-signed: SmartScreen may warn the first time you run them.

Check downloads with sha256sum -c SHA256SUMS --ignore-missing (macOS: shasum -a 256 -c SHA256SUMS --ignore-missing). More, including checking the macOS signatures: https://dzulfikar08.github.io/irit/install/

irit 0.3.4

Choose a tag to compare

@github-actions github-actions released this 12 Nov 15:11

msascope

  • The reference row is remembered per file: the reference and its options (pinned, ruler by
    reference, compare with it) are saved in the settings file for the last 50 files and restored
    when the same file (same size and modification time) is opened again and a row of that name is
    there. Stale or damaged entries are ignored.
  • Figures include the annotation tracks and the pinned reference row shown in the window, in
    PNG and SVG (SVG keeps track characters and labels as text). The figure's ruler counts the
    reference's residues when the view's does.

leanqc

  • Long-read views (ONT, PacBio), for files whose median read is over 1,000 bp or with the new
    --long: total bases, N50/N90, mean and median length, longest read, reads and bases per
    log-scale length bin; each read's quality from its mean error probability, its distribution,
    mean and median, reads (and bases) at or above Q10 and Q20, and mean read quality by length. In
    the HTML report, the JSON (long_reads, documented) and the terminal summary. The FastQC checks
    and the JSON of short-read files are unchanged. N50 and the other figures match an independent
    calculation on a public MinION run and a 1.5 GB synthetic file; peak memory stayed at 14–23 MB.

gtdbprep

  • ani fetches sketches in parallel: --fetch-jobs N range requests in flight (default 4,
    at most 16; 1 is the old one-at-a-time behaviour). Failed requests (connection errors, cut-off
    bodies, 408, 429, 5xx) are retried up to 5 times with growing, jittered waits and Retry-After;
    a request with no answer for 60 s, or a body slower than 32 KiB/s, is cut off and retried
    instead of waiting up to 10 minutes. Each sketch goes to the cache as soon as it is checked.
    From GTDB's server, 377 sketches (1.23 GB) took 474–562 s instead of 1,167–2,275 s; for a few
    dozen sketches the connection's variation was as large as the gain.
  • ani --species-only (opt-in) fetches only the references whose shared marker count shows
    they could reach ANI 95 (the smallest species radius) at AF 0.5, with a wide margin. GTDB-Tk's
    call is the same: identical on six test genomes, with 47 instead of 882 sketches fetched
    (147 MB instead of 2.8 GB; M. tuberculosis 1 instead of 377). The unassigned-hits file of an
    unassigned genome, and possibly the "other related references" column, list fewer references
    than GTDB-Tk's.
  • ani writes <work>/screen.tsv: every screened-in reference per genome with its shared marker
    count and whether it was used. With --species-only, the summary table gains a used column.

Download

Terminal (Linux or macOS), installs into ~/.local/bin without sudo:

curl -fsSL https://raw.githubusercontent.com/dzulfikar08/irit/main/scripts/install.sh | IRIT_VERSION=v0.3.4 sh
  • Linux command-line tools (static, any distribution): x86_64, aarch64
  • macOS command-line tools (Apple Silicon and Intel, Developer ID signed and notarised): irit-0.3.4-universal-macos.tar.gz
  • msascope for macOS (signed, notarised): msascope-0.3.4.dmg. Open it and drag msascope to Applications.
  • msascope for Linux (glibc 2.39 or newer: Ubuntu 24.04, Debian 13, Fedora 40 or later): x86_64, aarch64
  • Windows (sheetgen, leanqc, isotable, dblock): irit-0.3.4-x86_64-windows.zip. Not code-signed: SmartScreen may warn the first time you run them.

Check downloads with sha256sum -c SHA256SUMS --ignore-missing (macOS: shasum -a 256 -c SHA256SUMS --ignore-missing). More, including checking the macOS signatures: https://dzulfikar08.github.io/irit/install/

irit 0.3.3

Choose a tag to compare

@github-actions github-actions released this 12 Nov 12:51

irit-mcp

  • mlst's PubMLST data is in the run record. After locating mlst, the server asks it where its
    data is (mlst --help prints the resolved --datadir and --blastdb, so $MLST_DBDIR and
    wrapper scripts are followed; if it prints none, db/ next to the real program is used) and
    hashes VERSION.txt, the scheme map, the allele BLAST database and every scheme's profile table
    (177 inputs per call for Bioconda's mlst 2.35.0). $MLST_DBDIR is recorded and exported by
    replay.sh, and the mlst result gives database_version (the data's date, 2025-12-29). Two
    builds with different scheme snapshots no longer look the same in the record.
  • Shovill with SPAdes: contigs.fa is volatile. SPAdes 3.15.5 (Shovill's default assembler)
    gave different contigs.fa bytes in two identical runs (same length, same ST), so replay.sh
    only checks it exists; the hash is still recorded. SKESA's contigs stay hash-checked.
  • Checked against the real tools on a GitHub runner
    (run 43203640701): Shovill with
    SPAdes and its default read correction, stitching and Pilon (ST22 again), and
    dblock_verify with deep over the four databases, including the 162 files of the indexed
    AMRFinderPlus folder.

Download

Terminal (Linux or macOS), installs into ~/.local/bin without sudo:

curl -fsSL https://raw.githubusercontent.com/dzulfikar08/irit/main/scripts/install.sh | IRIT_VERSION=v0.3.3 sh
  • Linux command-line tools (static, any distribution): x86_64, aarch64
  • macOS command-line tools (Apple Silicon and Intel, Developer ID signed and notarised): irit-0.3.3-universal-macos.tar.gz
  • msascope for macOS (signed, notarised): msascope-0.3.3.dmg. Open it and drag msascope to Applications.
  • msascope for Linux (glibc 2.39 or newer: Ubuntu 24.04, Debian 13, Fedora 40 or later): x86_64, aarch64
  • Windows (sheetgen, leanqc, isotable, dblock): irit-0.3.3-x86_64-windows.zip. Not code-signed: SmartScreen may warn the first time you run them.

Check downloads with sha256sum -c SHA256SUMS --ignore-missing (macOS: shasum -a 256 -c SHA256SUMS --ignore-missing). More, including checking the macOS signatures: https://dzulfikar08.github.io/irit/install/

irit 0.3.2

Choose a tag to compare

@github-actions github-actions released this 11 Nov 17:48

msascope

  • Annotation tracks: Stockholm #=GC lines of every tag are shown under the track header,
    aligned to the columns and following hidden columns: SS_cons as helix and strand bands, RNA
    base pairs and coil, PP_cons as bars, other tags (seq_cons, RF, …) as text. The Tracks
    menu switches each one; the hover card over a track gives its character in words. #=GR
    lines are read too: the hover card over a cell lists the row's values, and the reference row's
    lines can be shown as tracks. Lines of the wrong length (or for a sequence that isn't in the
    file) are left out with a warning in the status bar and the terminal; at most 64 MiB of #=GR
    lines are kept.
  • Reference row: set from the Tracks or Sort menu, by right-clicking a name, or with m (the
    cursor's row); marked "ref" in the name panel. It can be pinned under the tracks, the ruler can
    count its residues (from its /start-end), and Differences, the identity track, the density
    view, the overview and figures compare with it instead of the consensus. It stays with its
    sequence through sorting, filtering and tree order.
  • Cheaper track header: the consensus, identity and gaps tracks (and the new ones) are drawn
    as cached image tiles instead of quads every frame; on a 20,000 × 10,000 alignment the paint
    median at the fit zoom went from 2.7–4.6 ms to 0.2 ms, at 1 px per cell from 3.9 ms to
    0.3–0.4 ms.
  • --stats prints the annotation read and any lines left out.
  • The --bench peak RSS difference reported for 0.3.1 was measured again: 0.3.0 and 0.3.1 peak
    within 2 MiB of each other, at the end of loading.

dblock

  • gtdbtk-r232-mockup is fetched again: GTDB corrected the stale MD5 in its MD5SUM.txt
    (GTDBTk#726), and the entry now uses it.
    gtdbtk-r232-client keeps the same file and now cites the published MD5.
  • amrfinderplus-db-4.2-2026-10-29.1: AMRFinderPlus's database, release 2026-10-29.1 (format
    4.2), from NCBI's FTP: the 38 files amrfinder_update 4.2.7 downloads, with MD5 and SHA-256 we
    computed (NCBI publishes none). 126 MB download, 256 MB once indexed.
  • Multi-part entries can be plain files ("format": "file"): each part is kept under its name.
  • prepare: a fixed step run on the downloaded files before the folder is sealed. The only one is
    amrfinder_index (needs AMRFinderPlus on PATH); the lock records the program and its version
    (prepared).
  • The Bakta entries' notes now say how to give Bakta this database as amrfinderplus-db/latest
    through a folder of links, instead of running amrfinder_update into the dblock folder.

gtdbprep

  • gtdbprep ani no longer reads GTDB's server unless asked. --source has no default:
    pass an unpacked sketches.db (path or URL), or --source gtdb to range-read GTDB's archive,
    as GTDB's maintainers asked (GTDBTk#727).
    A run whose sketches are all cached needs no --source.
  • Fetched sketches go to one cache per user ($XDG_CACHE_HOME/irit/gtdbprep-ani, or
    ~/.cache/irit/gtdbprep-ani) instead of <work>/cache, so a sketch is fetched once per machine,
    not once per work folder. --cache still overrides it. Runs sharing a cache no longer write
    through the same temporary file.

irit-mcp

  • shovill (Shovill 1.4.2, the current Bioconda release): paired-end Illumina assembly with
    SPAdes, SKESA, MEGAHIT or Velvet. Only Shovill 1.4.2's flags (--opts, which goes into a shell
    command, and --tmpdir are left out); reads checked as FASTQ, ram at least 8 GB as Shovill
    requires, gsize, namefmt and kmers checked the way Shovill parses them, an outdir whose
    path a shell would split refused. Returns contig count, total length, N50, longest contig and GC.
  • mlst (mlst 2.35.0): sequence type from the PubMLST schemes bundled with mlst. outfile must
    be named so isotable finds it (<sample>.mlst.tsv), and with one assembly its sample part becomes
    mlst's --label, so Shovill's contigs.fa joins to the right sample.
  • amrfinder (AMRFinderPlus 4.2.7): resistance genes and point mutations from an assembly,
    proteins, or both with a GFF (Bakta's with annotation_format: bakta). The database is checked
    like amrfinder does (files, indexes, format not newer than 4.2.7, data at least 2025-09-22.2,
    organism one the database knows) and its version.txt is hashed into the record; --update is
    not offered. output must be named <sample>.amrfinder.tsv (or contain amrfinder) for isotable.
  • AMRFINDER_DB is the default database, as for amrfinder itself; --tool-path takes the three
    new names.
  • The real-tools workflow also runs fastp → Shovill → mlst → AMRFinderPlus on real
    Staphylococcus aureus reads (ERR044595, nf-core/bacass's test data) and AMRFinderPlus on
    Bakta's proteins, then isotable over all of it, twice, then replay.sh.

isotable

  • parse::mlst and parse::amrfinder are public, for irit-mcp's summaries. No change in what
    isotable reads.

Download

Terminal (Linux or macOS), installs into ~/.local/bin without sudo:

curl -fsSL https://raw.githubusercontent.com/dzulfikar08/irit/main/scripts/install.sh | IRIT_VERSION=v0.3.2 sh
  • Linux command-line tools (static, any distribution): x86_64, aarch64
  • macOS command-line tools (Apple Silicon and Intel, Developer ID signed and notarised): irit-0.3.2-universal-macos.tar.gz
  • msascope for macOS (signed, notarised): msascope-0.3.2.dmg. Open it and drag msascope to Applications.
  • msascope for Linux (glibc 2.39 or newer: Ubuntu 24.04, Debian 13, Fedora 40 or later): x86_64, aarch64
  • Windows (sheetgen, leanqc, isotable, dblock): irit-0.3.2-x86_64-windows.zip. Not code-signed: SmartScreen may warn the first time you run them.

Check downloads with sha256sum -c SHA256SUMS --ignore-missing (macOS: shasum -a 256 -c SHA256SUMS --ignore-missing). More, including checking the macOS signatures: https://dzulfikar08.github.io/irit/install/

irit 0.3.1

Choose a tag to compare

@github-actions github-actions released this 11 Nov 12:06

msascope

  • Loading progress and Cancel: the loading window shows a progress bar of the bytes read
    (compressed bytes for gzip), the sequences read so far, both passes of Clustal and Stockholm,
    and the steps after reading; the status bar repeats it. Cancel or Escape stops the load and
    frees what it read; the file that was open stays open. Opening another file cancels a load.
    Clustal and Stockholm files are now opened twice instead of three times.
  • PNG and SVG export (Export menu): the view, or the selection (rows × shown columns), with
    names and ruler optional. SVG has the residue letters, names and ruler as text and is limited
    to 200,000 cells (larger regions get a message to export a PNG); PNG at 1×, 2× or 4×, cells
    drawn like the screen's tiles and text with resvg, at most 16,384 px a side and 100 megapixels.
  • Sorting and filtering rows: Sort menu by name, identity to the consensus or to a reference
    row (set from the cursor), gap share or length, reversible; filter bar (Ctrl+L) by name or
    accession (substring or /regex/), minimum identity and maximum gap share, and "Show only this
    clade" from the tree; "N of M rows shown" and Clear. Filters keep tree order; a sort turns the
    tree panel off until tree order is chosen again. Copy and export take the rows shown.
  • Go to name:position (Ctrl+G): a residue in the sequence's own numbering (from its
    /start-end), by name, accession or unique prefix, with messages for ambiguous or missing
    names and positions outside the sequence; a name alone selects its row.
  • Motif search (Find box in Residues mode, Ctrl+Shift+F): plain letters (IUPAC codes for
    nucleotides), PROSITE patterns (C-x(2,4)-C) or /regex/, over the ungapped sequences so
    matches span gaps; run in the background, matches highlighted over the shown columns, Enter
    or F3 for the next one, a count in the status bar.
  • Tree menu (right-click a node): collapse or expand, collapse to a depth, expand all, select
    the clade's rows, show only the clade, ladderize, reroot on the branch above the node.
  • Recent files: the last 10 in the settings file; the empty window lists five.
  • --stats --motif PATTERN and --stats --sort KEY time a motif search and a sort and filter.
  • Ctrl+G, Ctrl+F and Ctrl+L also work while a text box has the focus.

Download

Terminal (Linux or macOS), installs into ~/.local/bin without sudo:

curl -fsSL https://raw.githubusercontent.com/dzulfikar08/irit/main/scripts/install.sh | IRIT_VERSION=v0.3.1 sh
  • Linux command-line tools (static, any distribution): x86_64, aarch64
  • macOS command-line tools (Apple Silicon and Intel, Developer ID signed and notarised): irit-0.3.1-universal-macos.tar.gz
  • msascope for macOS (signed, notarised): msascope-0.3.1.dmg. Open it and drag msascope to Applications.
  • msascope for Linux (glibc 2.39 or newer: Ubuntu 24.04, Debian 13, Fedora 40 or later): x86_64, aarch64
  • Windows (sheetgen, leanqc, isotable, dblock): irit-0.3.1-x86_64-windows.zip. Not code-signed: SmartScreen may warn the first time you run them.

Check downloads with sha256sum -c SHA256SUMS --ignore-missing (macOS: shasum -a 256 -c SHA256SUMS --ignore-missing). More, including checking the macOS signatures: https://dzulfikar08.github.io/irit/install/

irit 0.3.0

Choose a tag to compare

@github-actions github-actions released this 11 Nov 08:08

msascope

  • msaview is now msascope; the name MSAView belongs to an existing viewer, GMOD's. The binary,
    msascope.app (bundle id com.dzulfikar.irit.msascope), msascope.dmg and the
    msascope-<arch>-linux.tar.gz archives take the new name. The installer replaces an msaview
    it installed earlier with msascope, and still installs msaview from releases up to 0.2.1.
  • Colour schemes: Clustal X with its per-column rules (Jalview's table), percent identity,
    hydrophobicity (Kyte–Doolittle), nucleotide and differences from the consensus, each with a
    "by conservation" fade, in light and dark. Letters on every fill keep at least 4.5:1 contrast
    (a test). Chosen from the toolbar's colour menu; s cycles, d differences, c conservation.
    Residues a scheme leaves unfilled are drawn as muted letters on the background, and gaps are
    the background with a faint dot, also in the dark theme.
  • Low zoom shows density, not sampled cells: below one pixel per cell each pixel takes the
    majority residue and the share of differing cells of the cells it covers (from the pyramid or
    the cells, up to 32 samples). "Fit all" on Pfam PF00042 full (32,112 × 698) now shows conserved
    columns as stripes instead of noise.
  • Overview panel on the right: the whole alignment, the visible area as a box (click or drag
    to pan), the visible range and the scheme's legend. Toggle in the toolbar.
  • Hover card next to the pointer: sequence name, accession (#=GS AC) and range, residue
    with its position in the ungapped sequence (from the name's /start-end), column in the
    alignment's numbering and among the shown columns, consensus, identity and gaps. It replaces
    the hover text in the status bar.
  • Insert and gappy columns can be hidden (toolbar, i and g). Insert columns are now kept
    at load: the columns where every row has lower case or ., or #=GC RF when the file has it.
    Residues are still stored as upper case, at 1 bit per column extra. Gappy means at least 25, 50
    (default), 75, 90 or 100% gaps. The ruler keeps the original column numbers and marks hidden
    runs; go-to takes original numbers. Inserts are hidden by default when a file has them.
  • Selection: click names for rows (Ctrl adds or removes, Shift extends, drag extends), click
    or drag on the track header for columns, Shift+click or Shift+drag in the alignment for a
    block; plain drag still pans. Ctrl+A selects all, Escape clears. Selected rows show in the name
    panel, columns in the ruler, blocks with an accent wash and outline; the status bar gives the
    size.
  • Copy and FASTA export: Ctrl+C copies the selection as FASTA, Ctrl+Shift+C as residues only;
    the Export menu writes the selection or all rows to a FASTA file. Hidden columns are left out
    unless "Include hidden columns" is on; "Leave out gaps" writes ungapped sequences. Rows come in
    the order shown. Above 64 MiB the copy goes to a file.
  • Cell cursor: arrows move it (Shift extends a block, Ctrl jumps over runs of residues),
    Page Up/Down and Home/End move it by a page or to the row's ends; Alt+arrows pan as the arrows
    did before. The cursor shows in the ruler and the name panel, and the hover card follows it.
  • Drag-and-drop: drop an alignment, a tree, or both on the window. Open… also takes trees.
  • Newick tree panel: from --tree, Open or a drop, Stockholm #=GF NH, or a sibling
    .nwk/.newick/.tree/.treefile with the alignment's stem. Rows follow the tree (Sort:
    tree / file, t); clicking a node collapses its clade into one row showing the clade
    consensus; the status bar counts collapsed clades and reports unmatched leaves and sequences.
    The parser handles quoted labels, branch lengths, comments and NHX, and internal labels, without
    recursion. A synthetic 230,000-leaf tree reads and parses in 0.1–0.5 s and matches in
    0.3–0.8 s on the test laptop, in the background.
  • The ruler's marks for hidden columns are now thin ticks, merged when closer than 6 px; the hover
    card over the ruler lists the hidden columns next to the column.
  • msagen --tree TREE.nwk writes a tree matching the generated alignment.
  • Toolbar icons that were not in the bundled icon set (the columns button) now use bundled ones,
    so they no longer show blank.

dblock

  • Identical files inside a database are hard-linked after unpacking, so they take disk space
    once. Files are matched by size, permissions and the SHA-256 dblock already records, so nothing
    is read twice. gtdbtk-r232-placement takes 2.7 GB on disk instead of 7.4 GB, and
    gtdbtk-r232-placement-lite 1.7 GB instead of 5.9 GB (measured with du on Linux).
    verify, verify --deep and GTDB-Tk folder hashes give the same results. The marker and lock
    entry record linked_files and linked_bytes. If the file system refuses hard links the copies
    are kept, with a note. fetch --no-dedupe keeps copies.
  • gtdbtk-r232-ani-screen (2.1 GB download, 2.3 GB unpacked): the screen index for gtdbprep ani,
    from a new unofficial irit-data release derived from GTDB R232 (CC BY-SA 4.0).

irit-mcp

  • Third-party tools for the microbial isolate path: fastp, bakta, checkm2 and
    gtdbtk_classify_wf. Each one offers exactly the flags of a pinned upstream version (fastp 1.4.0,
    Bakta 1.12.2, CheckM2 1.1.0, GTDB-Tk 2.8.0), read from that release's source. The calls go into
    the same run record and replay.sh as the irit tools, so a whole run from reads to taxonomy can
    be replayed without the AI.
    • The program is found on PATH or through the new --tool-path NAME=PATH. If it's missing,
      the call fails with a Bioconda install hint.
    • Inputs are checked before running (FASTQ or FASTA, plain or gzip, not empty). So are the
      databases: Bakta's version.json and files, CheckM2's .dmnd, and GTDB-Tk's data folders,
      including skani/, which classify_wf always needs, and the split-tree files unless
      full_tree is set.
    • The program's --version is recorded. A version other than the pinned one adds a warning to
      the result; replay.sh warns on a mismatch.
    • Exit 0 without the files the program always writes is an error (missing_outputs), recorded
      as a failed call.
    • Results carry small summaries: fastp totals, Bakta feature counts, CheckM2 completeness and
      contamination, and GTDB-Tk classifications. fastp, CheckM2 and GTDB-Tk files are read with
      isotable's parsers.
  • version in run records is now the first line of --version, as the docs said.

isotable

  • The output parsers are now also a library (isotable::parse), so irit-mcp reads tool outputs the
    same way isotable grades them. The command-line tool is unchanged.

gtdbprep

  • gtdbprep ani: GTDB-Tk's ANI species screen without the 85 GB skani database. The genomes
    are screened locally against a compact index of every R232 reference's marker k-mers. Only the
    references skani would compare in full are fetched, by HTTP range requests, and GTDB-Tk's own
    ANIScreener then runs on just those. Sketches can come from GTDB's own .tar.gz (through a
    gzip seek index, the default), an unpacked sketches.db (local or hosted), or one split into
    parts. On test genomes the output was byte-identical to GTDB-Tk against the full database. See
    the research page.
  • gtdbprep ani-index build | gz-scan | gz-attach make the screen index and the seek index from
    a full skani database and GTDB's archive (maintainers).
  • Reading sketches from GTDB's archive is a stopgap (about 5.7 times the sketch bytes, and load
    on GTDB's server): keep a --cache. Peak memory 1.9 GB for 9 genomes; GTDB-Tk on the full
    database takes 14.7 GB.

Download

Terminal (Linux or macOS), installs into ~/.local/bin without sudo:

curl -fsSL https://raw.githubusercontent.com/dzulfikar08/irit/main/scripts/install.sh | IRIT_VERSION=v0.3.0 sh
  • Linux command-line tools (static, any distribution): x86_64, aarch64
  • macOS command-line tools (Apple Silicon and Intel, Developer ID signed and notarised): irit-0.3.0-universal-macos.tar.gz
  • msascope for macOS (signed, notarised): msascope-0.3.0.dmg. Open it and drag msascope to Applications.
  • msascope for Linux (glibc 2.39 or newer: Ubuntu 24.04, Debian 13, Fedora 40 or later): x86_64, aarch64
  • Windows (sheetgen, leanqc, isotable, dblock): irit-0.3.0-x86_64-windows.zip. Not code-signed: SmartScreen may warn the first time you run them.

Check downloads with sha256sum -c SHA256SUMS --ignore-missing (macOS: shasum -a 256 -c SHA256SUMS --ignore-missing). More, including checking the macOS signatures: https://dzulfikar08.github.io/irit/install/