Skip to content

Changelog

Every change to the reusable workflows, the theme package and the tools, newest first, from CHANGELOG.md.

[Unreleased]

[1.16.0] - 2026-11-11

Added

  • Docs theme: the chip at a glance in the README. docs/hardware.yml takes optional package data from the datasheet, so a driver's README can show the module view without opening the docs: package.outline (length_mm, width_mm, height_mm, pitch_mm, ref), package.variants (the other packages the chip is sold in: id, type, pins, size, ordering, note, ref) and supply (name, min_v, max_v, ref). kokeb-docs hardware --check refuses numbers that are not positive, a citation that does not resolve, a variant id listed twice, a supply range upside down and pins that do not fit their edge at the pitch.
  • Docs theme: the chip-outline figure. When package.outline is given, kokeb-docs figures draws the package body from above to scale, light and dark, with the pads (the pinout's colours, numbers and pin-1 dot), arrows for length, width and pitch, a side view with the height and a size label. Leads reach out of a leaded package and sit inside a leadless one.
  • Docs theme: the chip-card region. <!-- generated: chip-card --> in a README (written by kokeb-docs gen, checked by kokeb-docs check) holds a table of package, pin count, body size, pin pitch and supply range, and one of the other packages, each row with its datasheet citation. A chip without an outline or variants still builds: the card leaves those rows out. Existing files are unchanged: every new field is optional and the schema stays at 1.

[1.15.3] - 2026-11-11

Fixed

  • ru-mkdocs-publish: a pre-release as the first publish. When a vX.Y.Z-rc.N tag was the first publish of a site, no version held the latest alias, so mike set-default latest failed and nothing was published. A pre-release now claims latest when no other version holds it, as a default-branch publish already does; it never takes latest from a version that holds it, and a final release still moves it.

[1.15.2] - 2026-11-11

Fixed

  • Docs theme: sibling organisations. A repository owned by a sibling organisation named <org>-* (such as kokeb-labs-drivers) now gets the same defaults as one owned by the organisation: site_url on its owner's Pages host (https://<owner>.github.io/<repo>/latest/), the absolute "Go to the latest version" link of the older-version banner, and the organisation link. Before, these were left unset for such a repository, so the banner link was relative (../..), pointed outside the built site and failed the offline link check on every page, and the sitemap and canonical addresses were missing.

[1.15.1] - 2026-11-10

Fixed

  • Chip circuit figures: a right-side lane is as tall as its parts and their labels draw, so a load's label no longer lands on the capacitor in the lane above. The layout check now reports any text on a symbol's strokes (parts, supply and ground symbols, pin stubs, no-connect crosses), so this cannot pass kokeb-docs hardware --check.
  • Chip block diagrams: links into one face of a block spread over it, each toward its other end, and a link label is placed clear of every other link and pin wire; a line through another link's label is reported.
  • Pin type colours: analog, power-out and nc get their own figkit tones (magenta, brown, stone), so every pin type has a distinct colour in light and dark (before, analog and power-out were the same and matched open-drain in dark).
  • Package types with a vendor prefix or variant (PG-DSO-36, P-DSO-20, TSDSO, HVQFN, WQFN, TQFN) are recognised, so they no longer need an explicit layout.
  • figkit cards: a title word too wide for its card breaks after a / or -, then shrinks, before it is a problem.

[1.15.0] - 2026-11-10

Added

  • Chip hardware diagrams in the theme package (kokeb_mkdocs_theme/hardware/, guide: docs/chip-diagrams.md): a driver describes its chip once in docs/hardware.yml (package, every pin with its type, voltage domain, function and driver hook, the internal blocks, reference circuits, a datasheet section and page on every fact) and gets the figures chip-pinout, chip-blocks and chip-circuit-<id> (light and dark, drawn by a pure-Python orthogonal router with real schematic symbols, each MCU wire labelled with the port hook that drives it), the generated regions chip-pins and chip-parts, and kokeb-docs hardware: --check validates the schema, pin numbers, citations, hooks (against the port header when source and symbol are given) and every circuit endpoint, --kicad-netlist OUT --circuit ID exports a circuit as a deterministic KiCad netlist, --svg DIR writes the figures anywhere. kokeb-docs check runs the hardware check first, and kokeb-docs figures --check drift-checks the chip figures with the others. The template site shows the fictional XMPL100.
  • figkit: Fig.circle(), Fig.path() and Fig.text(rotate=-90).

[1.14.2] - 2026-11-10

Fixed

  • ru-mkdocs-publish: the Lighthouse step no longer runs to its 20-minute timeout. 1.14.1 started kokeb-docs serve in the background with its output attached to the step, and the runner waits for every process holding that output; the server's output now goes to /dev/null, as the previous http.server line did.
  • Every apt-get update and apt-get install in the reusable workflows is bounded and retried: an apt configuration sets Acquire::Retries 3 and 20-second HTTP and HTTPS timeouts, update runs under timeout 180 and install under timeout 600, and each install is tried three times. A stalled package mirror then costs minutes, not the job: a fuzz job sat in apt-get update for its whole hour.

[1.14.1] - 2026-11-10

Fixed

  • ru-mkdocs-publish: Lighthouse measures the site as GitHub Pages serves it. The budget step served the built site with python3 -m http.server, which sends no compression, so each page was scored on two to three times the bytes a reader downloads (a quickstart page: 449 KiB instead of 173 KiB) and the performance score fell below 0.8 on a slow runner (0.75 to 0.78 on the same page that scores 0.98 compressed). The new kokeb-docs serve SITE PORT sends HTML, CSS, JavaScript, JSON, SVG and XML gzip-compressed, as Pages does; the budgets are unchanged.

[1.14.0] - 2026-11-09

Added

  • ru-release-firmware: the firmware release of ru-release-with-artifacts, with the same inputs and behaviour, whose job asks for actions: read as well as contents: write. In a private repository the token without actions: read cannot list the build workflow's runs, so the release failed in the download step with Resource not accessible by integration (list workflow runs for a workflow). Callers grant actions: read and contents: write.

Deprecated

  • ru-release-with-artifacts: use ru-release-firmware; change the file name in uses: and grant actions: read next to contents: write. It keeps working for existing callers, in public repositories, until the next major. Its guide no longer claims that the token reads another run's artifacts without actions: read; that holds only for public repositories.

[1.13.0] - 2026-11-09

Added

  • ru-misra: MISRA C++:2023 analysis with the CodeQL coding-standards pack (2.62.0, CLI 2.21.4) and the Kokeb C++20 profile pack on traced host and Cortex-M builds, the handbook's deviation records applied through the pack's deviation mechanism (a marker must cite a record and deviate a finding of its guideline), clang-tidy's validated checks and cppcheck on the same compile databases, and a per-guideline report (step summary, SARIF, JSON keyed by guideline, CSV). Levels high and safety. Every scan analyses a canary and fails on incomplete extraction. The first job refuses private repositories without an Advanced Security licence (CodeQL terms). Needs contents: read.
  • Tool validation of the MISRA toolchain (docs/tool-validation/plan.md): requirements derived from the compliance matrix and the pack, a seeded-violation and compliant case per requirement (tests/misra-validation/), a traceability page, an anomaly log, and a validation report attached to each release. The self-CI fails when a tool-decided guideline has no passing test, or when the toolchain changes without re-validation.
  • codeql/kokeb-cpp20: the own C++20 profile query pack, with KB-CPP-046 (Result types are [[nodiscard]]).
  • ru-misra: open_source_prerelease lets a private repository whose code is under the OSI licence in its LICENSE file, and is to be published, run the analysis; the job prints the declaration as a warning.
  • Code scanning of ci-tools moves to an advanced setup (.github/workflows/codeql.yml, security-and-quality) that leaves out the test fixtures, which break rules on purpose and whose findings the self-CI asserts.
  • ru-misra passes cmake_args and paths to its scanner as --option=value, so a value that starts with -D is not read as an option; the fixture self-test now passes one.
  • ru-misra gives CodeQL an explicit heap (the runner's memory less 2 GiB); the analysis ran out of memory on the 2-core runners of private repositories.
  • A tool suppression must cite a deviation record only when it can silence a finding the matrix relies on (a cppcheck suppression, a bare NOLINT, or one naming a validated clang-tidy check); NOLINTEND never needs one.
  • A validated clang-tidy claim can carry message:, a regex the diagnostic must match to count: a check that reports more than its guideline (cppcoreguidelines-macro-usage and object-like macros) decides only that part.
  • ru-misra's extraction check skips a unit the build defines but never compiles (no object file: a target outside all, such as a compile-fail test, or one the Cortex-M target does not build) and lists it as not analysed.
  • ru-misra reads the tool validation's anomaly log: at the high level, findings of a rule with a recorded false positive are listed for review (with the anomaly ids) instead of failing the job; validated rules still fail.
  • Tool validation: every seeded violation is marked (// seeded) and every claimed tool must report on that line, so a bad file that seeds several forms proves each detection; an independent review of all cases against the guidelines corrected the cases that tested the wrong thing and added the missing forms. New cases record the misreadings of class-template code found by the first driver scans. The CodeQL databases are traced with -fno-exceptions -fno-rtti, as the scans build; clang-tidy claims accept known_misses.

[1.12.1] - 2026-11-08

Fixed

  • ru-build: the native build no longer downloads ccache on every run. The pinned, hash-checked binary is kept in the Actions cache by version; a download retries on every error with a delay, and one that still fails falls back to the runner's ccache package with a warning. GitHub release downloads returning HTTP 504 had failed whole ESP32 jobs.
  • Self-CI caches its pinned shellcheck and actionlint by version and hash and retries their download on any error; ru-cross-build and ru-docs-publish retry their release downloads on any error (HTTP 5xx included).
  • ru-docs-linkcheck and ru-mkdocs-publish run lychee 0.24.2 installed by the job (pinned, SHA-256-checked, kept in the Actions cache) instead of lycheeverse/lychee-action, which downloaded it from GitHub releases on every run. Same arguments, same failure rule; the link report goes to the step summary.
  • ru-docs-publish (legacy): Doxygen 1.10.0 is kept in the Actions cache by version (falling back to the runner's package with a warning when GitHub releases fail), and its link check runs the same cached lychee.

[1.12.0] - 2026-11-08

Added

  • ru-build: input idf_install (default native). The job installs ESP-IDF for the matrix target only (a shallow clone and install.sh <target>, about 1.1 GB) and keeps it in the Actions cache keyed by the commit the IDF version names, with a pinned, hash-checked ccache. A warm restore takes about 20 s, against 70 to 160 s to pull the 9.4 GB espressif/idf image; a Debug and Release build of one app dropped from 175 s to 93 s. idf_install: docker keeps the image and espressif/esp-idf-ci-action.

[1.11.0] - 2026-11-08

Added

  • ru-repo-checks: DCO, the organisation standard, Markdown lint, the handbook rules, template drift, cppcheck and YAML lint as steps of one job, with the same scripts, pinned actions, failure rules and summaries as dco, org-standard-check, ru-markdown-lint, ru-handbook-check, ru-template-drift, ru-cpp-analysis and ru-yaml-lint. DCO runs first, before any checkout, and passes with nothing to check outside a pull request, so the gate needs no allow_skipped for it. Needs contents: read and pull-requests: read. The self-CI (check_repo_checks.py) fails when a copied step differs from its single workflow. The single workflows stay.
  • ru-cross-build: input per_core_jobs (default false) keeps one job per core.
  • ru-host-quality: input separate_jobs (default false) keeps one job per check.

  • Trigger policy (docs/triggers.md), enforced by the self-CI (check_triggers.py): workflows start only from push, pull_request, merge_group and workflow_call, with release.yml's workflow_run as the one exception.

Changed

  • The self-test callers (ci-*.yml) and the mkdocs-site template no longer have a workflow_dispatch trigger: a manual run is a re-run. A dispatched run executes in the default branch's context and can write the cache that runs on main restore, which made every checkout of a pinned ci-tools or driver-template commit a CodeQL actions/cache-poisoning/poisonable-step finding. The examples in the docs drop it too.
  • ru-cross-build: every core is built in one job, one after the other, each in its own build directory and log group; a core that fails does not stop the others. The toolchain is downloaded once, and a private repository is billed one rounded-up job instead of four (the plan job and one per core).
  • ru-host-quality: complexity, coverage and sanitizers run as steps of one job, each still running when another failed, with the same limits, summaries, artifacts and outputs; the limits job is gone (each job resolves the limits itself). Fuzzing stays a job of its own and no longer waits for the limits job. A private repository is billed one rounded-up job instead of four.

[1.10.0] - 2026-11-08

Added

  • ru-build: input group_build_types (default true) builds every build type of an app and IDF version in one job, so the runner is prepared and the ESP-IDF image pulled once per app instead of once per build type. The job name and artifact name list the build types together (Debug Release).

Changed

  • ru-build: ccache's directory is now kept between runs (actions/cache, skipped with clean_build). Before, IDF_CCACHE_ENABLE left it inside the container and every build started cold.
  • ru-build: the runner's unused SDKs are removed in parallel, and only when less than 30 GB is free.
  • ru-mkdocs-publish: the Lighthouse and screenshots job caches pip's wheels, the @lhci/cli download and Playwright's Chromium.

[1.9.1] - 2026-11-08

Fixed

  • ru-template-drift: with enforce: false the job failed instead of warning when the scaffold printed only to stderr (an empty report made the summary's grep fail). The scaffold's errors now reach the log and the summary.

[1.9.0] - 2026-11-08

Added

  • ru-template-drift: compares the files driver-template manages (workflows, _config/, skills, CMake helpers) with what the template makes today from the caller's .kokeb/template.yml, using the template's own scaffold.py --check-drift. It warns by default; enforce: true fails on drift. The signed-release job had already drifted: two drivers had it and the template did not.

[1.8.1] - 2026-11-07

Fixed

  • ru-mkdocs-publish: a pre-release tag (vX.Y.Z-rc.N) no longer takes the latest alias, and with release_version: minor no longer overwrites the docs of its minor; it publishes under its own tag. The first release candidate of driver-pcal9555 moved latest off dev, against the release standard.

[1.8.0] - 2026-11-07

Added

  • ru-attest-release: signed build provenance (GitHub artifact attestation, SLSA, Sigstore) for every file of a published release, after the files are checked against its SHA256SUMS. Callers add it as a job after ru-release or ru-release-with-artifacts and grant id-token: write and attestations: write. Public repositories, or GitHub Enterprise Cloud.

[1.7.0] - 2026-11-07

Added

  • ru-release and ru-release-with-artifacts attach supply-chain files to every release: a source archive with a stable checksum (<name>-<version>.tar.gz), an SPDX 2.3 SBOM (<name>-<version>.spdx.json: the repository, its licence and tagged commit, each git submodule at its pinned commit, each firmware zip) and SHA256SUMS. Built by the new tools/release_assets/release_assets.py (standard library only, unit-tested in the self-CI); the checksums also go to the run summary. Signed provenance waits for public repositories (docs/ru-release-with-artifacts.md#provenance).

Changed

  • ru-release-with-artifacts attaches one zip per firmware artifact, holding what flashes and debugs each ESP-IDF build (flasher_args.json, flash_args, the files they flash, the app .elf and .map), instead of every file of the build folders. The old upload sent about 2,900 files for a driver with a Debug and a Release build, past GitHub's limit of 1,000 per release, and files of the same name in Debug and Release overwrote each other.

[1.6.0] - 2026-11-07

Added

  • ru-host-quality: a fuzz job. With fuzz_cmake_args set (for example -DMYCHIP_BUILD_FUZZERS=ON), it builds with clang 18 and libFuzzer, ASan and UBSan, and fuzzes every fuzz_<name> executable for fuzz_seconds (by level: 60, 120, 600 s). Each target starts from its committed seeds in tests/fuzz/corpus/fuzz_<name>/ plus a corpus cached between runs. The step summary has runs, coverage edges and corpus size per target, and a failing input is uploaded as <artifact_name>-fuzz-crashes. Without fuzz_cmake_args nothing changes.

[1.5.3] - 2026-11-07

Changed

  • ru-handbook-check: the rules snapshot follows the handbook's ADR-0021 (EnsureInitialized() is the one start-up entry point: KB-CPP-043, KB-CPP-058, KB-CPP-059) and the new datasheet-trail rule KB-DOC-065.

[1.5.2] - 2026-11-07

Fixed

  • ru-host-quality: coverage merges the instances of a template (gcovr --merge-lines), so a line, branch or condition counts as covered when any instance ran it. A header-only driver is compiled once per port type, and an example's instance that runs only the happy path counted every failure path again, as uncovered: the AS5047U driver measured 73.8 % of branches instead of 94.5 %.

[1.5.1] - 2026-11-07

Fixed

  • ru-host-quality: the complexity job now measures header-only implementation files (.ipp, .tpp, .inl). lizard picks the files of a directory by extension and skipped them, so a driver whose code lives in src/<chip>.ipp had its implementation unmeasured. The job lists every C and C++ file under library_paths itself; the self-test fixture has a too-complex function in an .ipp file as well and must report both.
  • ru-cpp-lint: extensions defaults to c,cpp,cc,cxx,h,hpp,ipp,tpp,inl, so clang-format checks header-only implementation files. clang-tidy skips those files as translation units and reports on them through the sources that include them, as before. Callers that set extensions add ipp themselves.

[1.5.0] - 2026-11-06

Added

  • ru-handbook-check: checks a repository against the Kokeb Labs handbook rules a script can decide. It reads the .kokeb/handbook.yml declaration (handbook version, kind, assurance level, waivers) and checks the root layout, driver.yml against the build settings, C++20 and no file(GLOB) in CMake, no heap, exceptions, RTTI or virtual functions in driver code, the target library subset of headers, and no committed sdkconfig or datasheet. Each finding names its rule, quotes its title and links to it; a rule still in its warning period warns, a waived rule is a notice, and an expired waiver fails. The checker (tools/handbook_check/) and its snapshot of the rules are fetched at this workflow's own commit; 15 unit tests and good and bad fixture repositories run under the CI gate.

Fixed

  • tools/docs-migrate.py: a character range listed twice in the emoji pattern (CodeQL py/overly-large-range).

[1.4.0] - 2026-11-06

Added

  • ru-cross-build: builds one CMake target for Cortex-M0+, M3, M4F, M7F and M33F cores with Arm's GNU toolchain 14.3.rel1 (pinned, SHA-256 checked; an Ubuntu mirror hung one install for 20 minutes), one job per core, with the caller's own warning and language flags, and reports the objects' code and data size. A driver proves its code compiles for a microcontroller, not only on the host.

[1.3.0] - 2026-11-06

Added

  • ru-release-with-artifacts takes a commit input: the commit whose CI artifacts are released, for callers that release from a finished CI run (workflow_run), where github.sha is the default branch's head. It also checks that the release tag points to that commit before releasing. The guide shows a caller that releases only after CI passed on the tag.

Fixed

  • ru-build no longer fails pull requests from forks: their token cannot comment, so the size comment is skipped for them (and never fails the job), and the size table is always written to the run summary. On same-repository pull requests the comment is updated on each push instead of added again.

Security

  • ru-yaml-lint installs yamllint 1.38.0 pinned with SHA-256 hashes on ubuntu-24.04, instead of the latest yamllint from an unpinned pip install.
  • ru-host-ci passes artifact_name through the environment instead of expanding it inside a shell command.
  • ru-markdown-lint with use_docker: true pulled markdownlint/markdownlint-cli2:latest, an image that does not exist; it now uses the author's image davidanson/markdownlint-cli2, pinned to v0.23.3 by digest.

[1.2.1] - 2026-11-06

Changed

  • ru-cpp-lint runs clang-format 20.1.8 and clang-tidy 20.1.0 itself, from LLVM's builds on PyPI pinned with their SHA-256, instead of cpp-linter-action. That action installed clang from LLVM's apt server, fetching and running its install script as root without a pinned version or checksum, and the install hung for 20 minutes several times a day. The job now needs only contents: read; thread_comments and lines_changed_only are accepted but ignored (whole files are checked). Findings are annotations and a run-summary table, as before.

Security

  • The cpp-linter/cpp-linter-action and hustcer/setup-nu actions are no longer used, so they can leave the organisation's allow-list.

[1.2.0] - 2026-11-06

Added

  • ru-host-quality: complexity (lizard, modified cyclomatic complexity, length, parameters), coverage (GCC 14 and gcovr: lines, branches and MC/DC from -fcondition-coverage) and AddressSanitizer/UndefinedBehaviorSanitizer jobs, with the limits the Kokeb Labs handbook sets per assurance level (level: standard | high | safety). Tools are pinned and hash-checked. Its self-test runs under the CI gate, including a fixture over every complexity limit that must be reported.

[1.1.0] - 2026-11-06

Changed

  • A release waits for the whole self-CI on the merged commit (release.yml runs on workflow_run of "🧪 Workflow Self-CI"). The self-tests of the reusable workflows (ci-*.yml) are now called from ci-self.yml, so the required "CI gate" covers them: a red reusable workflow can neither merge nor reach @v1. The Markdown lint self-test reports issues instead of fixing them in the runner.
  • ru-cpp-lint fails the job when clang-format or clang-tidy reports a finding (fail_on_findings, default true); before, findings were only annotated and the job always passed. A new cmake_args input configures CMake with compile_commands.json, so clang-tidy sees the project's include paths; without it clang-tidy could not find the project's own headers. A findings output gives the count.
  • ru-cpp-analysis runs cppcheck from Ubuntu's archive instead of the unpinned ghcr.io/facthunder/cppcheck:latest image, no longer ignores cppcheck's own exit status, counts findings from the XML report (strict used to fail on every run, because the report always has a header), defaults std to c++20, takes an artifact_name, and has a findings output.

  • Releases of this repository are cut from CHANGELOG.md: a pull request that should be released moves the Unreleased entries to a new ## [X.Y.Z] - YYYY-MM-DD section (python .github/scripts/changelog.py bump patch), and merging it tags that version, publishes the section as the release notes and moves v1. Before, every merge was tagged and the changelog was left behind. The self-CI checks that a new section is the next version, is dated, and is a major when it has ### Breaking.

[1.0.5] - 2026-11-06

Fixed

  • The bundled fonts' @font-face rules moved from kokeb.css to assets/kokeb/fonts.css, which loads without blocking the first paint (as Google Fonts did before). Referenced from the render-blocking stylesheet, the font files sat on Lighthouse's critical path and the portal's performance score fell from about 0.89 to 0.78.

[1.0.4] - 2026-11-06

Changed

  • Bundled fonts used font-display: optional without a preload. This did not restore the Lighthouse score and was replaced in 1.0.5.

[1.0.3] - 2026-11-06

Changed

  • The kokeb docs theme ships Inter and JetBrains Mono itself (variable woff2, SIL Open Font License) instead of loading them from Google Fonts, so no Kokeb page sends a reader's address to Google. ru-mkdocs-publish copies the fonts beside the Doxygen stylesheet for the API pages; the root 404 page uses system fonts.

Security

  • Mermaid, still loaded on demand from unpkg, is pinned to 11.17.2 and checked with Subresource Integrity, so a changed file on the CDN is refused instead of run.

[1.0.2] - 2026-11-05

Security

  • ru-docs-publish installs Doxygen from its own release, checked against a pinned SHA-256, and Graphviz with apt, instead of two small third-party actions (ssciwr/doxygen-install, tlylt/install-graphviz).

[1.0.1] - 2026-11-06

Changed

  • figkit draws in the Kokeb palette (copper on light, star gold on dark), and its legend can explain solid and dashed arrows. A region can put its caption on a second line.
  • The workflow figure shows the push lane and the tag lane apart, with labelled arrows.
  • The style guide says how a figure reads: one direction, a verb on every arrow, colour for ownership.

[1.0.0] - 2026-11-05

First release of kokeb-labs/ci-tools.

Added

  • Reusable workflows for C/C++ (ru-host-ci, ru-cpp-lint, ru-cpp-analysis), documentation (ru-mkdocs-publish, ru-docs-publish, ru-docs-linkcheck), Markdown and YAML lint (ru-markdown-lint, ru-yaml-lint), releases (ru-release) and the single required check (ru-ci-gate).
  • ESP-IDF firmware workflows in the same repository: ru-build (matrix build from app_config.yml, size report and budget) and ru-release-with-artifacts (release with the tested firmware), plus the ensure-tools-directory composite action. The build finds or clones espidf-tools.
  • dco.yml: fails a pull request when a commit lacks a Signed-off-by: line matching its author.
  • org-standard-check.yml: fails when README.md, LICENSE, CHANGELOG.md or .github/CODEOWNERS is missing, or when LICENSE is not the MIT licence.
  • packages/kokeb-mkdocs-theme: the documentation theme (theme: kokeb), plugin (plugins: [kokeb]), figkit and the kokeb-docs command, with the pinned MkDocs toolchain. The colours sit in one palette block (--kb-palette-* in assets/kokeb/kokeb.css).
  • Self-CI: actionlint and shellcheck (pinned, SHA-256 verified), shellcheck of the composite action, the permissions contract, the README inputs tables, breaking-change detection, unit tests of the checks (the CI gate, the DCO script and the organisation-standard script included), the theme package's tests and the template site.

Security

  • Every reusable workflow sets permissions: {} at the top and the least each job needs; the self-CI enforces both. Every checkout sets persist-credentials: false except the two jobs that push (the release tag and the documentation branch).