Repository navigation
0.8.0 - 2026-11-18
Immutable
release. Only release title and notes can be modified.
Release Notes
Added
ledger backfill,ledger anchor,ledger verify-anchorsandledger projectmaintain a ledger in process.backfillappends afindingrecord for each observation of a manifest pinned by its SHA-256, at the observation time, and a rerun appends nothing.anchorsigns the day's head and publishes it once asANCHORS/YYYY-MM-DD.head.jsonafterverifyproves it binds to the log and extends the latest anchor;verify-anchorsproves the chain.projectwritesindex.json,index.mdand per-package JSON and Markdown history from a log whose signed head covers every record. Each prints one JSON object and exits 2 when refused;backfillandanchorexit 75 when an append or a signature did not finish. Each has an MCP tool:eb_ledger_backfill,eb_ledger_anchor,eb_ledger_verify_anchorsandeb_ledger_project.
Fixed
- A
from-commitread uses a directory it created for that read. It no longer shareseb-stack-from-commit-<pid>with every other read in the process, and it does not remove a directory that was already there. - A failed verification command is a
sanityfinding, ortimeout, unless its output names a transport, executor, runtime, or resource cause, so an Lmoderror:line no longer makes itcompile. Finding summaries and campaign history name the class and exit status as the state file does,sanityandexit 1, notSanityandexit Some(1). - A generated stack policy is a new file for each solve. The MCP server no longer reuses
eb-stack-policy-<toolchain>-<pid>.json, and it removes the file when the solve returns. - A job
stack admitrefuses for want of inodes carries arefusal: the inodes it is short, then theavailable,reserved,footprintandmarginthat decided it. It wasnull, so only a failed quota or scheduler command said why. - A lock's timestamp is
SOURCE_DATE_EPOCHwhen that variable is a number of seconds. Two solves of the same inputs then write the same time. - A pin on a package the trees have only outside the solve's toolchain hierarchy, in a
stack solvepolicy or apackage plan/package bumpstack policy, now fails with the hierarchy searched and the toolchains the trees have the package on, instead of "unknown package" and "names a package with no candidate at all". A stack-policy exclusion gets the same message. - A step of the Cargo prelude that fails stops EasyBuild's install command again: the host triple is read inside the prelude's
&&chain, not after_triple="";, whose semicolon ended the chain. A missing or failingrustcstill leaves the triple empty, even underpipefail. - A verification command runs with the module tree of the attempt's install path,
<installpath>/modules/allby default, first onMODULEPATHand onEESSI_EXTRA_MODULEPATH, somodule load {module}finds the module the attempt built, also after the EESSI init replacesMODULEPATH. - Compatible-release, caret, and tilde requirements exclude a pre-release of the next series.
1.3.0rc1is outside~=1.2.0and~1.2.3, and2.0.0rc1is outside^1.2.0. - Date versions such as LAMMPS's
29Aug2024compare by year, month and day, and a following_updateNby its number, sostack gapnames a LAMMPS22Jul2025_update4recipe as the newest outside the generation instead of29Aug2024. Requirements such as>=29Aug2024compare the same way. - The CI smoke step on the eOn core fixture checks
eOn-2.17.10-foss-2026.1.eb, the recipe the fixture holds, and passes only whenrecipe checkfails on the stack dependencies the drafts leave out, with the drafts supplying quill, readcon-core and rgpot. It named the 2.17.2 recipe the fixture no longer has, so it passed on "No such file or directory" without checking anything.tests/eon_core_rgpot.rsruns the step as CI runs it. - The
package planresiduals for an inferred moduleclass and for a dropped build requirement, and the source finding for a tree that declares no version, read as one sentence instead of carrying a run of spaces where the source line was wrapped. - The cargo-on-EESSI prelude quotes
CCandLIBRARY_PATH.CC='gcc; echo pwned'is stored as the linker, and it is not run. - The import scan of a PyPI source tree skips standard-library modules, test suites, documentation, examples and maintainer scripts, and reads only
importstatements outside triple-quoted strings, so natsort 8.4.0 has 3undeclared-importresiduals instead of 29. An import whose name is part of one already reported, such asrequestsafterrequests_cache, is still reported. Cargo.lockpins rustls 0.23.46 and rustix 1.1.6, the patched releases for RUSTSEC-2026-0285, RUSTSEC-2026-0346, RUSTSEC-2026-0348, RUSTSEC-2026-0349, and RUSTSEC-2026-0486.component-reachresolves the closure over every recipe in the trees, so dependents ongfbfandgompitoolchains of the generation are in it. On easybuild-easyconfigs develop with the EESSI 2026.06 easystack, curl's closure gains 19 dependents ongfbf-2026.1andgompi-2026.1, such as R, netCDF, HDF5 and SciPy-bundle, and 12 recipes only they pull in, such as Tk through CMake, so the counts go fromreview=106 no-rebuild=16toreview=133 no-rebuild=20. Requirements no recipe resolves are listed inclosure_unresolvedinstead of being dropped.easyblock-reachheadsreach.tomlandreach.mdwith a line that says only thateb-stack easyblock-reachgenerated the file, instead of thepr-overlapwarning about a data branch that is never merged. The report is written to--out-dir, and no branch holds it.ledger showprints each record'sstageafteraction, so the started and finished records of one staged rebuild no longer print as two identical rows.package bumpandpackage plankeep a dependency's stated versionsuffix when the robot tree has a build with it that the solve can take:('HDF5', '1.14.2', '-mpi')and('HDF5', '1.14.2', '')no longer become the-serialbuild because that one ranks higher. The solve takes another suffix only when the target has no build with the stated one, as when-CUDA-12.9.1moves to-CUDA-13.3.0across a generation.scripts/eessi-extend-eb.shcounts a presentEESSI_*_INSTALLvariable as set when the value is empty, matching EESSI-extend. An emptyEESSI_SITE_INSTALLorEESSI_CVMFS_INSTALLstays in that mode and no longer also creates a user prefix.scripts/eessi-extend-eb.shdefaults to EESSI 2026.06, takes the repository fromEESSI_CVMFS_REPO, puts the compatibility layer'sldonPATHfromEESSI_EPREFIXwhether or notEESSI_VERSIONwas set, and creates a namedEESSI_USER_INSTALLas well as the default one. The Cargo prelude takes the compatibility layer fromEESSI_EPREFIXtoo.scripts/eessi-extend-eb.shstill reaches EasyBuild whenrustcis not onPATH, and the Cargo prelude takes a mirror's compatibility layer fromEESSI_CVMFS_REPOwhenEESSI_EPREFIXis unset. CI runstests/eessi_extend_eb_script.rs.stack gappicks the bump source of ano-reciperow among builds of the newest version by toolchain family and generation instead of file name: a build on a toolchain family of the generation comes first, then the newest toolchain generation, so QuantumESPRESSO 7.5 is bumped fromfoss-2025brather thanNVHPC-25.3-CUDA-12.8.0, and a recipe onGCCcore-14.3.0is preferred to one onGCCcore-11.3.0.stack shardcounts-f, EasyBuild's short form of--force, as forcing a rebuild, as it counts--forceand--rebuild. A line that rebuilt an installed module with-fwas planned as building nothing: the module had no weight and was in no closure, so a line that needs it could land in another shard and build whileeb -frebuilt it.target doctorpasses the EasyBuild probe of a target withEB_LOCAL_EESSI_EXTENDon only when--versionprintsThis is EasyBuild. eb-local 2.4.0 has no extend mode, answered--versionitself and passed; the failed check now carries areason.
Documentation
- The EESSI target in
docs/orgmode/reference/targets.organdexamples/targets/eessi-extend.tomlneeds eb-local 2.5.0 or later, the first release withEB_LOCAL_EESSI_EXTEND. - The PyPI how-to and the
package-planskill plan the pure-Python beautifulsoup4 example onGCCcore/15.2.0, where EESSI 2026.06 installs Python,Python-bundle-PyPIandBeautifulSoup, with the newexamples/stacks/eessi-python-gcccore.tomlpolicy, and the fixture robot holds the GCCcore-15.2.0 builds of Python 3.14.2 andPython-bundle-PyPI2026.04. The numpy example stays onfoss/2026.1, becauseSciPy-bundleis never on GCCcore. - The README says when EasyBuild alone is enough, and which stackhands tool calls each
stackandledgerverb. - The
--helptext matches what the commands do: the top-level help names exit status 78 fromstack admit; the secondstack admitexample reads the quota and the scheduler inside--lockwith--quota-cmdand--live-jobs-cmd, instead of expandingsqueuebefore the lock;campaign --helpshowscampaign run;stack generation --helpmaps 2026.06 to 2026.1; andstack batch --installed-modulercsays it reads a module tree as well as a.modulerc.luafile. - The
stack shardandstack capacityhelp examples, the shard how-to, the capacity reference and the skills name a neutral example site, not one cluster: azen2.listorzen4.listbuild list, a commit cache under$SCRATCH, the node featurelocal-disk, and the site's build host. - The docs match the code again:
AGENTS.mdlistspackage latestandstack capacity; thecli.orgMCP list nameseb_stack_footprintsandeb_stack_admit;docs/campaign-boundary.mdcites the currentsrc/campaign.rsandsrc/foreign.rslines and therun_targetcalls; the shard how-to says to plan with EESSI-extend'sEASYBUILD_FILTER_DEPS, or pass that list as--filter-deps; the PyPI and CRAN how-tos say the robot tree is$EBROOTEASYBUILD/easybuild/easyconfigsfrom EESSI's EasyBuild module; the commented pins in the EESSI example policies name the EESSI 2026.06 bundles; and theprovidesandcampaignAPI docs have no broken link or stray///. - The shard how-to, tests and fixtures no longer carry one site's names: the
stack shardexamples buildsite-2026e.list, and the--include-systemsvalues in the tests aresite-zen2andsite. - The target reference and the eOn on EESSI how-to say that each campaign attempt installs into
<work_root>/install-attempts/<package>-<version>/attempt-<N>, which takes precedence over the prefix EESSI-extend derives fromEESSI_USER_INSTALL.
Removed
- The EESSI target runs
eb-localwithEB_LOCAL_EESSI_EXTEND=1instead ofscripts/eessi-extend-eb.sh, and the script is removed. eb-stack does not build eb-local; the container bind-mounts the binary the operator installed. A present emptyEESSI_*_INSTALLstill counts as that mode, whicheb-localenforces. With noEESSI_*_INSTALLset, installs move from the script's$HOME/EESSIto EESSI-extend's own$HOME/eessi; setEESSI_USER_INSTALLin[targets.easybuild.environment]to keep the old prefix.
Changed
- News fragments in
changelog.d/, assembled intoCHANGELOG.mdby towncrier at release, as in the other stackhands repositories. - The doc walkthrough tests cover
get-started.org: its bash blocks follow the same path rules as the other pages, and each block is run and must print the output the page quotes, withstack admitexiting 78. skills/golden-replay/run-on-builder.shhas no default builder: setEB_STACK_BUILDERto the builder's ssh destination (it wasrg.terra). WithoutEB_STACK_BINit runs the debug build of the builder's checkout at the same place under the home directory as the local one, instead of one fixed checkout path.
Install eb_stack 0.8.0
Install prebuilt binaries via shell script
curl --proto '=https' --tlsv1.2 -LsSf https://github.com/stackhands/eb-stack/releases/download/v0.8.0/eb_stack-installer.sh | shInstall prebuilt binaries via powershell script
powershell -ExecutionPolicy Bypass -c "irm https://github.com/stackhands/eb-stack/releases/download/v0.8.0/eb_stack-installer.ps1 | iex"Download eb_stack 0.8.0
| File | Platform | Checksum |
|---|---|---|
| eb_stack-aarch64-apple-darwin.tar.xz | Apple Silicon macOS | checksum |
| eb_stack-x86_64-apple-darwin.tar.xz | Intel macOS | checksum |
| eb_stack-x86_64-pc-windows-msvc.zip | x64 Windows | checksum |
| eb_stack-aarch64-unknown-linux-gnu.tar.xz | ARM64 Linux | checksum |
| eb_stack-x86_64-unknown-linux-gnu.tar.xz | x64 Linux | checksum |
| eb_stack-x86_64-unknown-linux-musl.tar.xz | x64 MUSL Linux | checksum |