What's new in Cejel
Cejel keeps two version tracks. This is the CLI track —
the binary, npm package, GitHub Action, Docker image, and MCP server. Changes to the scoring
rubric are tracked separately, with a full before/after corpus delta, in the
rubric changelog.
Install or update any time with
npx @cejel/cejel@latest .
← Back to cejel.dev · Repository Evidence Review (Barg Labs)
v0.6.12
2026-11-19Removed, breaking: the @cejel/cejel/d-series subpath (#423). It was deprecated in 0.6.2 and has warned with CEJEL_DEP_D_SERIES on every load since. Importing it now fails with Node’s ERR_PACKAGE_PATH_NOT_EXPORTED, and the npm package no longer ships dist/packs/d-series/. D-series detection was retired on 2026-08-02: the D1–D6 rules are a taxonomy, not a production detector. Only code that imports @cejel/cejel/d-series is affected, and there is no replacement export. cejel scan, cejel-mcp, @cejel/cejel/decision-contracts, scores, bands and report fields are unchanged.
Added, under every rubric: B3 names a counted CI workflow in which a failing test command is ignored. Where a workflow the PR-gate CI workflow count includes runs its tests with continue-on-error: true on the test step, or ends the test command with a fallback that always exits 0 (|| true, || :, || exit 0, || echo …), one of B3’s disclosures names the workflow and the line, and says the count still includes it. Cejel does not check whether another step fails the run. It names only what the workflow shows: the command must be a test runner, not a line that mentions one, and every later line of the step must only print. No scoring change: the count, B3’s score and status, and every other score are unchanged, and a report with no such workflow is byte-identical.
Added: Cejel’s scores as a GitLab metrics report, so a merge request shows the scores that changed. cejel export gitlab-metrics <report.json> writes one line each for the overall score, process trust, code trust and every criterion. A value Cejel did not measure gets no line, never a 0. The GitLab template runs it after the Code Quality export and publishes it as artifacts:reports:metrics; GitLab shows metrics reports on Premium and Ultimate only. No scoring change: every report.json is unchanged. See the GitLab guide.
Changed: the GitHub Action’s job summary says what a result does not cover. Under the verdict, the summary now prints the sentence the certificate’s first page and the terminal already carry: Cejel read the repository’s files and history and did not run the code or its tests, so the result does not show that the code works as intended or that it has no vulnerabilities. A job summary is often read without the certificate. No scoring change, and the Action’s outputs and pass/fail result are unchanged.
Changed: every rubric’s behaviour fingerprint moved because the corpus it is measured on changed; the corpus change alone moved them, and it changed no rubric’s behaviour (#622). The corpus gains a depth-1 shallow clone and a publishing workflow that runs its tests before it publishes, two shapes it did not hold before. Scored without them, it reproduces all eleven previous fingerprints. A report.json differs from the previous build’s in rubricBehaviourFingerprint alone; no score, verdict or other field changes.
Fixed: a failed or timed-out git call no longer silently changes which evidence a scan reads (#574). A scan now asks git for the history depth once, and every reader uses that answer. Where git’s check fails, the depth comes from the repository’s shallow file, the file git itself reads, and each failed read is stated as a scan limitation. This changes v17 on the failure path only: where git does not answer the depth check, A2 and B2 read the depth the shallow file gives, where they read “not shallow” before, so a score can move there, and only toward the reading git’s own answer gives. Where git fails and the files cannot settle the depth either, A2 abstains under every rubric unless a confirmed critical secret finding stands, where the report already withholds its verdict. Where git answers, this change leaves every report byte-identical.
Fixed: scans that stopped without a report now write one. A scan with more than 16 scan limitations failed its own schema and wrote no report; the 16-entry cap now holds where the list is assembled, and the history-depth sentence is always kept (#719). A file path longer than the report’s 700-character limit also stopped the scan; such a path is now left out of the report with a stated count, and no path is shortened (#612). A file of many blank lines, or one long run of identifier characters, no longer stalls the scan; that fix is performance only, and every report is byte-identical (#550). Only scans that failed before change.
Changed, under the prospective v27 rubric only: B3 and A1 name more of what they cannot read. B3 counts a publishing workflow whose every publishing job declares a named environment, cites each workflow it does not count, and names CI configurations Cejel does not read, such as .travis.yml, where its publishing rule leaves no workflow that targets pull requests (#517). A1 names as not read the pytest runs an unexported Makefile assignment of PYTEST_ADDOPTS may reach, and its tox reader works within a stated budget (#674, #653). v17 and every other rubric are unchanged. v27 stays opt-in through --rubric-pin, the default stays v17, and it carries no precision, recall or calibration claim; no holdout or calibration ran, and no published score moved.
Inspect the v0.6.12 release → · Read the OCI and MCP release-chain proof →
v0.6.11
2026-11-09Changed: Cejel is a GitLab CI/CD Catalog component, and its GitLab template has run on a GitLab runner. The v0.6.10 remote include ran on a GitLab.com instance runner (pipeline 3048076526), with full history in the cejel job and the Code Quality report shown on a merge request. The same job is published in the GitLab CI/CD Catalog as gitlab.com/barglabs/cejel/cejel@0.6.10 (pipeline 3048107004). The GitLab guide and the README now give the component first and keep the remote include as the fallback. This supersedes the statement in the 0.6.4 entry that the template had not been run on a GitLab runner. No scoring change.
Changed: a score or count change at the same file and line keeps the GitLab Code Quality fingerprint (#700). The fingerprint covered the finding summary, which carries a metric-derived score and band, so on a demonstration merge request two improved scores read as “2 new findings and 2 fixed findings”. The fingerprint now covers the criterion, path, line and the finding’s identity, with no score or count in it, and the export carries the same number of entries as before. description, check_name, location and severity are unchanged, and the score stays in description. Every located finding gets a new fingerprint once, so the first merge request after upgrading shows each existing finding as resolved plus new, one time. Export only: every report.json is unchanged.
Fixed: the GitHub Action no longer reads the calling repository’s package.json to set up pnpm, and no longer changes the calling job’s pnpm (#685, #686). A caller pinning a pnpm other than 9.0.0 failed with “Multiple versions of pnpm specified”, and after a first fix every later pnpm step in the caller’s job ran the Action’s 9.0.0. The Action now fetches the exact pnpm its own package.json pins, checks the tarball against a pinned sha512, installs it into a fresh directory and runs it by path, adding no pnpm to the caller’s PATH. Everything from inputs: to the end of both Action manifests is now pinned byte for byte by the distribution validator (#689), and a pnpm version hardcoded in a workflow is refused in flow style too.
Fixed: cejel issue reproduces with the attested scan’s own date, so a certificate from an earlier year can be issued. v17’s B4 freshness check matches the scan’s calendar year, and the re-run used the issuer’s clock, so an issuance in a later year than the scan was always refused. The re-run now uses predicate.generatedAt from the attestation already accepted as bound to report.json, never from a flag, and refuses one that is missing, is not an ISO 8601 time, or is later than the issuer’s clock. issuance.json records it as reproduction.generatedAt, so the issuance format moves from 1.0 to 1.1; a 1.0 issuance still verifies. No score moves: a normal scan passes no date, and its report.json is byte-identical under every rubric.
Changed, under the prospective v27 rubric only: A1 and B3 read more of what CI runs, and name what they cannot read. A1 reads PYTEST_ADDOPTS in the environment of a CI command that runs make, since make passes it to the recipe, and names as not read a pytest run whose PYTEST_ADDOPTS is set by an expression or somewhere the reader does not follow, and a collection hook defined in a plugin class (#633, #634, #669, #673). B3 credits a test runner pointed at a path only where the path holds a test file Cejel counts (mocha index.js, and go test on a package with no test files, run nothing and exit 0), and a check that uv run launches only where uv does not build and install the project first (#659, #660). v17 and every other rubric are unchanged. v27 stays opt-in through --rubric-pin, the default stays v17, and it carries no precision, recall or calibration claim; no holdout or calibration ran, and no published score moved.
Fixed, in the tests only: the scaling check steps by 4 and refuses above 11 (#678, #680). At a step of 2 with a limit of 3, the same unchanged code measured 1.9 to 3.2 on CI, so the runner decided some runs. The check now compares L/2 with 2L, where linear is about 4 and the square 16, and refuses a ratio above 11 or below 2. The six readers above 4 were each timed alone: none is superlinear by algorithm, and their cost per character grows with the memory they touch, which a memory-bound stand-in in the self-test now shows beside them. No reader, rubric or report changes. The published leaderboard, scored from full-history clones, is not re-run for this release and declares a scorer version pin at 0.6.3.
Inspect the v0.6.11 release → · Read the OCI and MCP release-chain proof →
v0.6.10
2026-11-04Changed, under every rubric: a scan whose git history Cejel could not read gives no score. A shallow clone, or a checkout whose history depth Cejel cannot read, now abstains under v17, the default, as it has under v27 since 0.6.9: no score and no verdict, labelled “Unrated: shallow clone” or “Unrated: history depth unknown”, with the instruction to run git fetch --unshallow or check out with fetch-depth: 0. A directory with no .git, such as an uploaded snapshot or an archive export, is “Unrated: no git history”. Each criterion is still read and shown. Before, v17 scored a depth-1 clone, on which A2 cannot see a secret committed and later deleted, so a shallow clone could score higher than the full clone of the same commit and pass a --min-score gate the full clone fails. This is visible to every default user of the GitHub Action, because actions/checkout makes a shallow clone unless fetch-depth: 0 is set. BargLabs/cejel/action@v1 now declares the limited, limitation-count and limitations outputs the root action declares, and an Unrated run sets limited to true. This change leaves a complete clone’s score exactly as before, byte for byte, under v17 and v27.
Added: a plain-language first page. certificate.html now opens with a verdict line and four lists in plain words: what holds, what is wrong, what is not known, and what to check next, with a line saying what the result does not cover (Cejel does not run the code or its tests). Each item links to its source further down. An area that was not examined, or that the report says could not be fully read, is listed as not known, never as holding. The lists come from report.json alone; the terminal prints the four headings after the verdict, and summary.json gains one key, plainSummary. The certificate below the first page is unchanged apart from the anchor ids the links point to, and report.json, attestation.json, scores and verdicts are byte-identical for the same input, which tests pin against 644 renders captured before the change.
Changed, under the prospective v27 rubric only: A1 and B3 read more of what CI actually runs. A1 places every pytest collection hook that touches items in one of three outcomes (deselects everything, puts the items back, or not read and named), reads the pytest runs, options and configuration search pytest itself uses, through nox, tox, make, just and hatch, and names a run it cannot read instead of rating its files hollow. B3 credits a check before a publish only where it follows the check to a known test runner, or a lint form that fails on a finding, with flags and an environment known not to change what runs. Measured on 393 synthetic repositories: v17 to v26 byte-identical, and v27 never above v17. v27 stays opt-in through --rubric-pin, the default stays v17, and it carries no precision, recall or calibration claim; no holdout or calibration ran.
Fixed: the timing guard that stopped the 0.6.9 npm publish. A guard that times a v27 scan of very long CI lines against a 2,000 ms budget measured 2,063 to 2,165 ms on GitHub’s runners and failed the verify step of the 0.6.9 npm publish before the publish step ran. The shell-word scanner it exercised now compares character codes and copies the characters it keeps a run at a time; locally the guard’s scan fell from 650 ms to 520 ms, with the budget unchanged. A scaling check that could fail on a loaded host now times alternating pairs in CPU time (#642). No report, score, note or fingerprint changes under any rubric. The published leaderboard, scored from full-history clones, is not re-run for this release and declares a scorer version pin at 0.6.3.
Inspect the v0.6.10 release → · Read the OCI and MCP release-chain proof →
v0.6.9
2026-11-02Distribution: a GitHub Release only. 0.6.9 was published as a GitHub Release with its five binaries. Its npm publish stopped at the verify step on a timing guard, so it never reached npm, Docker / OCI or the Official MCP Registry. 0.6.10 carries every change below on every route, and its entry above describes the fix.
Fixed, under every rubric: A2 reads all of a large history, and says when it could not. A2’s whole-history reads that stopped at Cejel’s own 8 MiB output limit are now re-read in pages and batches (#567), each re-read within a 120-second time budget whose exhaustion is a scan limitation (#586). When A2 is not applicable after an incomplete read of the history (a shallow or partial clone, or a failed or capped read), the certificate says so under “What was not established” (#556). A failed git read is classified from git’s own messages, never from text that can include a file name, and a submodule at a credential-shaped path no longer counts as a failed read (#560). The print-only workflow reader claims only for GitHub-hosted Ubuntu and macOS runners, since it assumed bash (#564), and a path with control characters is called credential-shaped only when it is (#569). These can change a v17 report’s scan limitations, disclosures and badge gating; no v17 score moves.
Added: the GitHub Action takes an optional rubric-pin input. It takes the same value as the CLI’s --rubric-pin and is forwarded only when non-empty, so an empty or absent input leaves the invocation, the report and every output as they were. A v27 Action run on a shallow checkout no longer fails the step: the Action recognises the Unrated labels and leaves the score output empty, as for any abstention. When pinning v27, check out with fetch-depth: 0.
Changed, under the prospective v27 rubric only: shallow clones, publishing workflows and hollow test files. A shallow clone gets no composite score and no verdict (“Unrated: shallow clone”, or “Unrated: history depth unknown” where the depth cannot be read). B3 counts a publishing workflow with no pull-request trigger only when a failed check really stops every publish. A1 names every file it rates hollow and withholds a rule only on evidence, saying so each time. B2 and B3 stop crediting workflows that only check out, print or upload an already-built artifact (#527, #535). v17’s scores, statuses and fingerprint do not move, though every v17 certificate’s glossary entries for PR trace and the PR-gate CI count now describe v27’s rule. v27 stays opt-in and carries no precision or recall claim.
Measured: two preregistered public confirmations moved no public v17 score. On the 23 public corpus repositories, #567’s paged reads removed the history scan limitations of biomejs and react and gave react’s A2 one info finding, with no v17 score, status, placement or fingerprint moved; under v27, biomejs’s A2 no longer abstains (overall 2.5 to 3.0), and no score was higher under v27 than under v17. #586’s time budget left all 23 reports byte-identical under both rubrics. The published leaderboard is not re-run for this release and declares a scorer version pin at 0.6.3. Result (#567) · Result (#586).
v0.6.8
2026-10-22Security: some repository lines could stall a scan. Thirteen strip and match sites on repository and ingest text used regular expressions that retried from every position of a long run. An 80,000-character line took 3–4 s, and single 500 KB files scanned in 89–681 s. Each is now a linear scan with exactly the same result, which a test checks against the old expression. This was present in 0.6.7 and earlier; no score, status or report byte changes. Separately, A4 now quotes a CI command in Cejel’s own words only (uv sync … --locked), so a credential in the command cannot reach report.json or the certificate (#553). That defect was introduced while this release was in development and never shipped.
Fixed: certificates say what they set aside, did not check, or read less of. Under every rubric, including the default: A2 names credential-shaped values it set aside as documentation placeholders, by count and path, and no longer cites “.env files are gitignored” beside a committed env file; B2 labels each cited workflow by its own trigger; B3 names counted workflows whose only steps print text; Git LFS pointers are counted and labelled (#519, #520); and A4 no longer calls a lockfile “not required” when the repository’s own CI requires one. These sentences are a new criteria[].disclosures field (report format 1.8). The CI lockfile note is right in more cases, and the workflow reader is now linear in a workflow’s size and nesting. No v17 score, status or fingerprint moves.
Added: input binding for ingested evidence, and a prospective v27 rubric. When something is ingested, report.json gains inputBinding (report format 1.7): per document, the sha256 of its bytes, the producer and subject it declares, and whether that subject matches the scanned repository and commit. --strict-ingest refuses mismatched, stale and duplicate inputs before scoring, and cejel verify … --input fails on a swapped, missing or extra input. Default scoring is unchanged. witan-rubric-v27-prospective-2026-10-19 is v17 with five false assurances removed. It is opt-in through --rubric-pin, the public default stays v17, and it carries no precision, recall or calibration claim.
Measured: the preregistered confirmation moved no public v17 score. The 23 public corpus repositories were scored from full-history clones under v17 at v0.6.7 and at the release candidate, and under v27 at the candidate, which differs from this release only in records. Under v17, no score, status, metric value, finding count, placement or fingerprint moved; beyond the version, 17 reports differ only in B2’s evidence labels and A4’s evidence and finding text. Under v27, no composite or criterion score was higher than under v17, and one score moved, down. The published leaderboard is not re-run for this release and declares a scorer version pin at 0.6.3. Result.
Inspect the v0.6.8 release → · Read the OCI and MCP release-chain proof →
v0.6.7
2026-10-18Security: a scanned checkout could redirect where a scan writes its output (#508). The default output directory, .cejel, is created in the directory the scan runs from, usually the checkout being scanned. A checkout could commit .cejel, or a file in it such as .cejel/report.json, as a symbolic link, and a scan then wrote through the link, overwriting a file the user could write outside the checkout. scan, issue and export now refuse, before writing anything, when the output directory or an output file is a symbolic link, and exit non-zero. An explicit --out follows the same rule, so on macOS --out /tmp is refused, because /tmp is itself a link; name a directory inside it. This affects CLI scans of untrusted checkouts; the HTTP server is unchanged.
Added: v26 item 8 (#509). Under witan-rubric-v26-prospective-2026-10-14 alone, which stays explicit opt-in (--rubric-pin; the public default stays v17), A3’s “declares neither a build nor a typecheck script” finding fires only on the evidence that settles it. Where a place it must read was not read, the certificate says so instead. No metric changes, and no precision or recall claim.
Measured: the preregistered confirmation moved no public repository. The 23 public corpus repositories were scored under v17 and v26, at v0.6.6 and at the fix commit, which differs from this release only in version identity and records. All 23 reports were byte-identical under both rubrics. No public repository carries A3’s build finding at v0.6.6 under v26, so item 8 had nothing to narrow there; the run is not evidence about where item 8 stops the finding. The published leaderboard is not re-run for this release and declares a scorer version pin at 0.6.3. Result.
Inspect the v0.6.7 release → · Read the OCI and MCP release-chain proof →
v0.6.6
2026-10-15Security: a repository could stall a scan with a regular expression or a workspace glob (#502). A pnpm run /…/ selector in a package script was compiled and run against the repository’s own script names, so a crafted pattern could backtrack for minutes. Cejel no longer evaluates it; the reference is treated as unresolved. Declared workspace globs, new in 0.6.5, are now matched without a regular expression, in time linear in the path and with a cap on their length; matching is otherwise unchanged.
The one scoring change. Because the selector is no longer evaluated, a repository whose only route to its coverage script is a pnpm run /…/ selector loses that A1 coverage credit, under every rubric. The “no coverage configuration” finding is withheld for it as well.
Security: an HTTP-server upload and the terminal certificate (#500). An uploaded snapshot containing a .git entry could make Git read another repository on the host, and the certificate could then carry that repository’s commit and a tracked file name. Such upload paths are now refused, and Git discovery cannot leave the snapshot directory. This affects HTTP-server deployments; CLI scans are unchanged. Separately, control characters from ingested findings reached the terminal; repository- and ingest-derived fields are now stripped of them before printing. Cejel’s own text and report.json are unchanged.
Measured: the preregistered confirmation moved no public repository. The 23 public corpus repositories were scored under v17 and v26, at v0.6.5 and at the fix commit, which differs from this release only in version identity and records. All 23 reports were byte-identical under both rubrics. This shows that neither fix alters these repositories’ reports; it does not test the fixes themselves. The published leaderboard is not re-run for this release and declares a scorer version pin at 0.6.3. Result.
Inspect the v0.6.6 release → · Read the OCI and MCP release-chain proof →
v0.6.5
2026-10-15Changed: four v17 recognition fixes (#479, #484, #487, #489). Under witan-rubric-v17-2026-07-24 and every selectable rubric, with no identifier bump, as in 0.6.3 and 0.6.4; weights and thresholds are unchanged. Each fix reads evidence Cejel did not read before. B3 reads the scripts of declared npm, yarn and pnpm workspace packages, not only the root package.json (#479). A1 recognises more documented coverage configuration before it says there is none (#484), and more frameworks’ assertion idioms (xUnit, NUnit, MSTest, XCTest, ScalaTest, Kotest, tape and tap) before it calls a test file hollow (#487). A2 flags a digest comparison when the same file computes that digest with a keyed HMAC, which narrows the cost 0.6.4 stated for #430 (#489). Recorded in the rubric changelog, with fingerprints and report goldens re-pinned.
Measured: the preregistered public before-and-after held every expected value. The 23 public corpus repositories were scored under v17 at v0.6.4 and at the #490 merge, which differs from this release only in version identity, records and experiments. Five moved at scoring level, all up or unchanged, none down: two in B3 through #479, both declaring workspaces, and three in A1. No other criterion moved, and the #489 rule changed no public repository. This measures movement, not whether any finding is correct. The published leaderboard is not re-run for this release and declares a scorer version pin at 0.6.3. Result.
Added: v26, an explicit opt-in prospective rubric (#478, #482). witan-rubric-v26-prospective-2026-10-14 is selected only with --rubric-pin; the public default stays v17. v26 carries no precision or recall claim. Under v26 alone, A5’s no-reconciliation finding and A2’s tenant-without-RLS finding fire only on the evidence that settles them, and otherwise abstain with one sentence in the certificate.
Also in this release. Presentation only (#490): the B3 citation now says how many delegated CI steps Cejel did not read, a reusable workflow in another repository or a local action. report.json and certificate bytes change for repositories whose workflows delegate that way, and for repositories that declare workspaces (#479).
Inspect the v0.6.5 release → · Read the OCI and MCP release-chain proof →
v0.6.4
2026-10-13Changed: three v17 scoring behaviour changes (#459, #427, #430). Under witan-rubric-v17-2026-07-24 and every selectable rubric, with no identifier bump and weights and thresholds unchanged. A3’s Express error-handler check recognises two ordinary spellings it missed: a lint-underscored parameter and a trailing comma after the fourth parameter (#459). AVA assertion calls in active inline callbacks now count toward A1’s non-hollow test share (#427). A2 no longer asserts a timing side-channel from a bare digest name, which supplies no evidence that the digest is secret (#430). Recorded in the rubric changelog, with fingerprints and report goldens re-pinned.
Measured: the preregistered public before-and-after moved no score or placement. All 23 public corpus rows completed in each of three arms: before #459, after #459, and with the corrected #427 and #430 fixes. No score or placement moved on any public row. This does not mean the changes cannot move a repository; the synthetic specimens written for them do move. The published leaderboard is not re-run for this release and declares a scorer version pin at 0.6.3. Result.
Changed: the certificate states what it read (#467). Presentation only: no score, weight, detector or file selection changes. When A3, B2 or B3 counted or read more workflow files than they cite, the citation now says so, and a withheld workflow is stated as not read. Where a metric reads package scripts from the root package.json only, it says how many other authored package.json files are present. Certificate and report.json bytes change for those repositories, so byte-reproducibility records made with earlier versions will differ for them.
Added: a warning when npx ran a different cejel than it resolved (#436, #462). Inside a checkout of the Cejel repository, npx can run whichever cejel is on PATH rather than the version requested. When the requested version is known and differs from the one running, Cejel prints one warning: line to stderr naming both. The run and its exit code are unchanged. This narrows #436; it does not fix it, and a cejel older than 0.6.4 on PATH prints nothing.
Also in this release. The GitLab CI template asks for full history on the cejel job, so A2 and B2 can read it, for an include pinned to v0.6.4 or later; it has still not been run on a GitLab runner (#456, #461). The certificate glossary and B6’s missing line now name the evidence the protected-path gate counts (#463, #465, #466). A new prospective rubric is available by explicit opt-in only; the public default stays v17.
Inspect the v0.6.4 release → · Read the OCI and MCP release-chain proof →
v0.6.3
2026-10-10Changed: Cejel no longer credits text that only describes a control (#433). A scoring behaviour change under witan-rubric-v17-2026-07-24, with no identifier bump and no version gate. B6’s fail-closed privilege check and kill switch no longer credit a test, fixture or calibration/ file, or a match only inside a comment. A3’s error-boundary filename rule is anchored to a path segment, and its Express error-handler check ignores comments. B6’s human-gate and required-review evidence no longer credits a changelog, an experiment or calibration record, a review note, a reports/ file, a test or fixture tree, or a committed Cejel certificate. CODEOWNERS counts only at the repository root, .github/ or docs/, the locations GitHub reads. Recorded in the rubric changelog, with the behaviour fingerprint re-pinned.
Measured: the preregistered public-corpus before-and-after. All five preregistered expected values held. Two public rows moved: cejel’s B6 became not applicable with its overall unchanged at 2.8, and sinatra’s B6 became not applicable with its overall moving from 2.4 to 2.2. No ranked placement changed. Preregistration · result.
Not in this release: the published leaderboard is not regenerated. Board rows still show the 0.4.9-era scoring until the full-history regeneration (#431). The before-and-after above measures the change; it does not update the board. Update, 2026-10-11: the board is now regenerated on 0.6.3 with full-history clones and shared ranks; see the leaderboard and the rubric changelog’s 2026-10-11 entry.
Plugin: the #433 known issue is retired. The Claude Code plugin README now says the three false citations it named on Cejel’s own certificate are fixed in 0.6.3. The note on #436 stays; #436 is not fixed.
Inspect the v0.6.3 release → · Read the OCI and MCP release-chain proof →
v0.6.2
2026-10-08Added: a Claude Code plugin (#429). /plugin marketplace add BargLabs/cejel, then /plugin install cejel@cejel, gives Claude Code the scan tool and a cejel-scan skill on when to scan and how to read a certificate without overstating it. The tool is started through npx, pinned to this release, and carries a title and read-only, non-destructive, idempotent, closed-world annotations. Scoring is unchanged and stays offline.
Added: cejel mcp serves the MCP server over stdio (#432). It is the same server as the cejel-mcp bin, which keeps working, and it is how the Claude Code plugin launches the server. cejel mcp takes no arguments; a repository directory named mcp is scanned with cejel scan mcp or cejel ./mcp.
Added: a shallow clone is disclosed on the certificate (#428). When the scanned repository is a shallow clone, report.json carries historyLimitations naming the two criteria that read commit history, A2 and B2, and the terminal, HTML and Markdown certificates show it. It is display-only: no score, status, abstention, badge or --min-score outcome changes. Report format 1.5 (additive). Who is affected: GitHub Action users on the default actions/checkout depth of 1, whose certificates now show the limitation; check out with fetch-depth: 0 to give A2 and B2 the full history.
Changed: the Docker image builds dist/ once, natively (#421). Both platforms now build on the builder’s own architecture and run nothing under emulation, which removes the arm64 build hangs and means the SBOM and both images describe the same dist/ bytes. Who is affected: nobody using the image; its contents and behaviour are meant to be unchanged.
Deprecated: the @cejel/cejel/d-series subpath, removed in 0.7.0 (#419). Loading it emits one Node DeprecationWarning (CEJEL_DEP_D_SERIES). Every export still works until 0.7.0. cejel scan, cejel-mcp, scores, bands and report fields are unchanged.
Inspect the v0.6.2 release → · Read the OCI and MCP release-chain proof →
v0.6.1
2026-10-06Changed: release SBOMs list what each binary and the image bundle, and are attested (#404). Each standalone binary’s SPDX SBOM is now generated from the build: it lists the packages bundled into the binary with their exact versions, the Node.js runtime it embeds, and the binary’s SHA-256. The 0.6.0 SBOMs listed only the scanned directory and the binary. Each SBOM is attested to its binary and can be checked with gh attestation verify <binary> -R BargLabs/cejel --predicate-type https://spdx.dev/Document/v2.3. The Docker image keeps its SBOM of base-image packages and gains a second SPDX attestation listing the packages Cejel bundles. A release whose SBOM is empty, or is missing a required package, now fails before anything is published or attested. The npm package does not carry an SBOM. No score, band or report field changes.
Changed: the certificate reads correctly the first time (#415). Five changes to wording and layout on the terminal, HTML and Markdown certificates, from a first read of the 0.6.0 certificate for expressjs/express. A static coverage percentage of 0 is no longer shown as 0/100 percent, which read as “0% coverage”; the certificate now says that Cejel does not run tests and how the rubric scores the absence. The summary box and “What was established” count dimensions the same way and say how many do not apply. “What to do next” no longer tells you to make files readable when Cejel skipped them only for their file type. A Conditional, At risk or Unverified verdict with no critical or warning finding now says it comes from the overall score’s band. The note about the product name is in plain English. No score, band, verdict or abstention moves. report.json, summary.json, badge.json and badge.svg are byte-identical, so attestation.json binds the same report digest; reportFormatVersion stays 1.4.
Inspect the v0.6.1 release → · Read the OCI and MCP release-chain proof →
v0.6.0
2026-10-03Added: cejel export gitlab-codequality <report.json> [-o gl-code-quality-report.json] (#401). Writes a Cejel report’s located findings as a GitLab Code Quality report, offline. Only findings with both a file and a measured line are exported. Findings scoped to a whole file (line: null) and ingested third-party findings are not exported; a footer on stderr counts them, so nothing is silently dropped. Severity maps critical to critical, warning to major and info to info; minor and blocker are never emitted. No score, band or report field changes.
Added: a GitLab CI template, ci/gitlab/cejel.gitlab-ci.yml (#401), documented in docs/gitlab-ci.md. It is an includable job that runs the scan, publishes the Code Quality report as artifacts:reports:codequality, keeps .cejel/ as an artifact, and fails the job below an optional CEJEL_MIN_SCORE. Include it from this release’s tag. The template is covered by structural tests in this repository; it has not yet been run on a GitLab instance by us.
Changed: the GitHub Action’s description is shorter (#402), to fit GitHub Marketplace’s limit of 125 characters. Inputs, outputs and behaviour are unchanged.
Inspect the v0.6.0 release → · Read the OCI and MCP release-chain proof →
v0.5.0
2026-10-02--min-score now exits non-zero when fewer than half of the applicable dimensions were measured, even when the score clears the minimum. Before this, a score resting on a handful of measured dimensions passed the gate as readily as a fully measured one. Now the gate refuses if the measured share of applicable (not not_applicable) dimensions is below one half, overall or in any category. Who is affected: CI jobs that run --min-score on low-coverage repositories, such as a small or newly created codebase where most dimensions return insufficient_data. Those jobs used to pass and will now fail. How to see why: stderr reads Cejel: cannot evaluate the required minimum N/4.0 because fewer than half of the applicable dimensions were measured in at least one score (…), with the measured/applicable counts in the parentheses. The scores in the report do not change; only the exit code does.
The HTTP MCP server is stateless: certificates and badges come back with the scan that produced them (#369, merged as #388). The cejel://last-scan/certificate.html and cejel://last-scan/badge.svg resources are removed. They never worked over HTTP: each request builds a fresh server, so a resource read could not see the scan from an earlier request and always answered "No scan has run yet". The scan tool now takes an optional artifacts list (certificate, badge) and returns the requested HTML certificate and SVG badge in the same response. Who is affected: HTTP MCP clients that read those two resources after calling scan; pass artifacts to the scan call instead. The stdio MCP server and the CLI are unchanged.
Report format 1.4: each metric records the weight share it actually contributed (#370, merged as #389). Metrics gain appliedWeightShare, and the HTML certificate reads it from the report instead of recomputing it at render time, so the weights a certificate shows are the ones the signed report carries. reportFormatVersion moves from 1.3 to 1.4. No score, band or behaviour fingerprint changes; a test pins that the field moves neither. Who is affected: consumers that validate reports against a closed 1.3 schema.
The HTTP MCP summary response no longer rejects valid scan results (#369, merged as #388). Its strict schema did not allow the per-finding dimensionBand and displaySummary fields or the content-read summary that scan results carry, so some valid scans failed schema validation on the way out.
Inspect the v0.5.0 release → · Read the OCI and MCP release-chain proof →
v0.4.11
2026-09-25A2 under the default rubric missed a populated PostgreSQL password in a committed .env file's DATABASE_URL (#336). A DATABASE_URL=postgres(ql)://user:password@… in a non-template .env/.env.* file is now a committed-secret finding, in both the current-tree and the history scan, and an unchanged value is not reported twice. A placeholder-shaped password is not flagged, and .env.example-style templates are not read. This is a change to A2 under the calibrated default. It moved no row of the published corpus, which contains no non-template .env file. See the 0.4.11 entry in leaderboard/RUBRIC_CHANGELOG.md.
A3 prod_readiness_primitives under the default rubric missed a real, registered Express global error handler. An independent evaluator ran 0.4.10 against five pinned revisions of his production system; every revision reported no error boundary despite a real handler being present and wired in. We do not have his code, so this widens on a fixture catalogue of public Express idioms instead of his specific file (src/witan/__tests__/a3-error-boundary-idioms.test.ts has the full table). Now credited: a handler with its first parameter renamed _err or its fourth renamed _next; a class method registered via .bind(this) from an export class (TypeScript); and a handler file under a conventional server/ root using a .mjs/.cjs extension. Also fixed a related false positive found while building the catalogue: a commented-out // app.use(errorHandler); line was wrongly making an otherwise-dead, never-registered stub count as reachable, because the reachability check read raw file text without stripping comments. Stated limits, not silently widened: a three-parameter handler with next omitted entirely stays uncredited (Express recognizes error middleware by arity, and matching on names alone there would be the arity-only false positive this detector's shape pattern exists to avoid). This is a recall change on prod_readiness_primitives under the default rubric; it carries a paired before/after delta record at unchanged corpus commits in docs/experiments/v17-behaviour-delta-post-0.4.10-2026-09-23/. No fixture in the existing v17/v22 byte-stability or A3 suites changed output — this widening only reaches repositories that hit the newly-credited idioms.
The (prospective, witan-rubric-v23) health_readiness_route info-severity finding missed a real Express health route under several ordinary, public-documentation idioms , cataloged and regression-guarded in src/witan/__tests__/a3-health-route-idioms.test.ts. This finding is info-severity, not a scored metric — no A3 score moves; a certificate's "what to do next" does. Now also credits:
- a segment-internal
_/-prefix directly after the route's leading slash (/_health); - the Rails convention
/up, slash-anchored only — never as a bare decorator string, which would also match unrelated direction/toggle literals ('up' | 'down'); - the NestJS decorator idiom with no leading slash (
@Get('health')), requiring@Get(...)context for barehealth; context-free bare keywords are limited tohealthz,readiness, andliveness, so ordinary'ready','live', and'health'strings do not match; - a
functions/serverless-functions directory (Firebase/Netlify convention), added to this one signal's file-selection predicate rather than to the shared implementation-file allowlist other A3 signals also use.
Deliberately left as stated limits, not silently dropped: /ping, /status, and /alive (too generic, or not a named convention the way /up and /healthz are); a trailing query string (/health?probe=1 — crediting it would also credit an outbound probe/test URL, and a false assertion is worse than a miss here); and a health-check library imported with no literal path anywhere in the repo (express-healthcheck, @godaddy/terminus, lightship) — whether the import itself should count as evidence is an open decision, not made in this change.
Two of the three file-selection examples this gap was originally reported against (src/server/health.controller.ts, app/api/health/route.ts) turned out to already be admitted by the existing predicate once checked against the running detector; only the functions/ case was a genuine gap.
A certificate that withheld content could print no disclosure line at all, and a reader could not tell a correct no-intersection ("no signal would have read it") from a disclosure that simply failed to print. On the 2026-09-21 reruns, a revision with two oversized withheld files produced identical certificates under the default rubric and the prospective witan-rubric-v23 — no withheld-path line under either — leaving open whether v23's withheld-path abstention correctly found no intersection or should have fired and silently did not. The certificate's "What was not established" section now carries exactly one withheld-path sentence on every certificate, never conditionally: "No content was withheld from any signal" when nothing was withheld; a "no signal that would have read them selected them" sentence when something was withheld but earns no signal's own file-selection test; the existing per-signal abstention disclosure, plus the count, when the intersection is earned and witan-rubric-v23's mechanism acted on it; and, new for every other rubric (including the calibrated public default), a sentence naming which signal would have read the withheld content and stating plainly that this rubric reports on what it read while the prospective v23 rubric abstains instead — the exact sentence the 2026-09-21 default-rubric certificate was missing if the second reading is the true one. report.json (report format 1.3) gains an additive withheldPaths array — one entry per withheld path, always present (empty when nothing was withheld), each carrying its skip reason, every signal whose own file-selection test would have read it, and whether the running rubric's mechanism acted on that intersection — computed from the same predicate calls the abstention itself consults, so the disclosure and the abstention can never drift apart. No scored metric moved on any fixture under any rubric.
Inspect the v0.4.11 release → · Read the OCI and MCP release-chain proof →
v0.4.10
2026-09-17The npm provenance gap is repaired prospectively. The published 0.4.10 package carries npm attestations from GitHub Actions, and its registry publisher is GitHub Actions rather than the package account. The immutable 0.4.9 package remains disclosed as published without npm provenance; a new release establishes the corrected artifact without rewriting that historical record. 0.4.9 was published by hand, by the operator, from a machine holding the package token and outside the workflow that generates provenance; the cause, the token inventory and the publishing-access change that followed are stated in the release notes, and manual publication with two-factor authentication remains possible, so workflow publication is the stated policy and not a registry-enforced guarantee.
Release identity is checked across each published route. The signed tag, public GitHub Release binaries, OCI image, Official MCP Registry record, Homebrew formula, and consumer-facing Action all resolve to 0.4.10. The published Windows, Linux aarch64, and npm package runs produced byte-identical reports for the same fixture. The floating Action v1 tag now resolves to this immutable release commit.
Inspect the v0.4.10 release → · Read the OCI and MCP release-chain proof → · Read the release notes →
v0.4.9
2026-09-16Every certificate states what it can see. Certificates produced by 0.4.9 and later describe the repository tree at its pinned revision. Evidence outside that tree, including a separate test repository, is neither seen nor claimed to be absent. Earlier certificates do not gain this disclosure retrospectively; their existing attestations remain valid.
Scoring changes under the same calibrated rubric are disclosed. The public paired 0.4.8 → 0.4.9 record at unchanged corpus commits found metric changes on four of 24 rows and one headline change: django, 3.2 → 3.1. React’s observability count moved 68 → 108 and alfred’s 64 → 73, with unchanged scores in both cases. These are wider pattern matches, not evidence that either repository became more observable. The test-script check also misses delegating test commands; that limitation remains disclosed in the rubric record. Default scans remain on v17, including its unresolved withheld-file scoring limitation; the newer abstention treatment is prospective.
Report format and scoring are separate comparisons. All 24 reports differ at byte level because report format 1.2 adds the optional rubricBehaviourFingerprint. This format change does not mean all 24 scores changed. A synthetic-corpus guard now detects measured scoring changes under an unchanged rubric identifier; matching fingerprints establish agreement on that corpus, not detection coverage.
The npm provenance gap remains visible. The published npm 0.4.9 package executes as 0.4.9, but its metadata omits gitHead and dist.attestations; the npm attestation endpoint returned HTTP 404 on 16 September. The OCI image has separately verified source provenance. The immutable npm version and release tag will not be replaced to repair this omission; the next release must restore npm provenance.
Inspect the v0.4.9 release → · Read the full public rubric delta and limits →
v0.4.8
2026-09-08A file Cejel declined to read under its own size limit was scored like one it couldn’t read at all. Measured effect: a design partner’s same-commit comparison at rubric v17 found 3.1/4.0 measuring 9 of 9 criteria under 0.4.5, 1.5/4.0 measuring 5 of 10 under 0.4.7 — same repository, same rubric. Two compounding defects, both present since 0.4.6. First, a file over the 512,000-byte content limit was mapped to affected rubric criteria by path shape (test-shaped, doc-shaped, and so on) before that criterion’s own collector ran on the content that was readable, wiping criteria even when other readable evidence fully answered them; a criterion now abstains only when its own collector genuinely has no readable evidence.
Declining to read a file is a disclosed limit, not evidence loss — and this release stops scoring it like evidence loss. Second, that abstention carried the same insufficientData flag as a genuine read failure, so the scorer kept it in the composite denominator at a punitive 0 instead of excluding it like ordinary insufficient_data. report.json’s signals now carry the skip reason: unreadable/denied_path stay a read failure and remain scored; too_large, excluded_by_extension, and non_regular_file are Cejel’s own disclosed coverage limits, excluded from the composite and never scored. The certificate’s disclosure text no longer calls a declined-to-read file “could not be read.” ADR-0001: coverage is disclosed, never discounts a score.
v0.4.7
2026-09-07The certificate now leads with what is wrong, shows the weight it actually applied, and folds away what a reader does not need first. Critical and warning findings across all criteria open certificate.html, critical first, before the scope prose, and the section says plainly when there are none (#268). Every rendered metric shows the weight applied after renormalisation across surviving metrics, identically on HTML, Markdown and terminal, instead of its nominal declared weight (#271). The glossary and the not-applicable group collapse by default behind native <details>, zero JavaScript, so the certificate stays one offline file and a machine reader still sees everything (#285). The “what to do next” field is derived from the same gap logic as “what was not established”, ranked so an unmeasured criterion that could change the verdict comes first, and every sentence names an absence rather than promising a score (#283). Presentation only: report.json, attestation.json, summary.json and badges are byte-unchanged.
A certificate can now say which CI attempt produced it. --run-attempt <n> records the run attempt; the GitHub Action forwards GITHUB_RUN_ATTEMPT automatically and nothing else ever fabricates one. It appears on the HTML and Markdown certificates and as an additive-optional predicate.githubRunAttempt on attestation.json, never inside report.json, whose byte-reproducibility for a pinned revision excludes per-invocation values (#282). --rubric-pin also accepts the prospective witan-rubric-v23-prospective-2026-09-06, explicit opt-in only; the public default stays calibrated v17 (#276, #277, #278).
The release itself was fixed. The MCP Registry publish validator compared server.json against the registry’s last observed state rather than the release being cut, which is how 0.4.6 published as 0.4.5 twice. It now compares against package.json, and release-identity metadata is bumped in the same commit that gets tagged; 0.4.7 is the first release for which that is true (#273). The release-currency check of this site was split so the homepage asserts every pinned @cejel/cejel@<version> string names the release, instead of grepping for a marker that had moved (#257). A defect-class census maps the shipped rule inventory against the 2024 CWE Top 25, the 2021 OWASP Top 10 and Cejel’s own D1–D8 taxonomy, with a CI check against drift; breadth only, no recall or precision claim (#274, #275).
v0.4.6
2026-09-02A versioned contract for bringing in your own signals. The CLI adds a generic --ingest contract v1 (version: "1.0") with a published JSON Schema and explicit additive-minor / breaking-major compatibility rules — a malformed or unversioned ingest document is now rejected loudly instead of guessed at. New attestations carry predicate.reportFormatVersion: "1.0".
The GitHub Action publishes only what it verified. Every invocation now runs in a fresh directory outside the scanned repository, so an early failure can no longer fall through to a committed workspace file. Public ingest across every entry point now shares one file validator with symlink and containment checks, and a scheduled full-tree mode extends the existing disclosure-boundary guard.
v0.4.5
2026-08-24Stable caller-supplied product identity and explicit rubric selection. The CLI adds --product-name so the same pinned repository emits the same identity fields across differently named checkouts, and --rubric-pin so a caller can explicitly select a supported rubric. Default scans remain on the calibrated v17 rubric.
Every certificate measurement now explains itself. Human-readable certificates state when score labels agree, explain why comparison does not apply, rename the bounded file-count ratio so it cannot be mistaken for prose verification, and give enumerable, binary, and conditional metrics exact reader-facing explanations. HTML metric values also stay inside their criterion cards.
v0.4.4
2026-08-18Certificate text stays inside its tooltip. Long unbroken metric descriptions now wrap inside the HTML certificate tooltip instead of overflowing its fixed-width box.
A named rubric must actually run. The sealed scoring path now rejects an explicitly supplied rubric version that is not wired into public dispatch, rather than issuing a certificate that names a rubric which did not execute. The release also adds clean-runner checks for the Smithery and OpenClaw MCP routes.
v0.4.3
2026-08-17Closes the stale-version trap. Every documented CLI invocation across the README, leaderboard site copy, and the calibration issue template now pins @latest or an explicit version, so npx @cejel/cejel can no longer silently resolve a stale cached package. A guard test fails the build if an unversioned invocation is reintroduced.
Readability fixes. Certificate tooltips now sit clearly above the page instead of blending into the content behind them. The Markdown certificate now carries the CLI version, matching the HTML and JSON certificates. Several metric labels and finding lists were corrected for consistency and accuracy across all three certificate formats. The release also states the recognized-CI boundary: which systems Cejel treats as real CI signal, and how that set can only change through a version-gated rubric change.
v0.4.2
2026-08-13A read-only currency verifier independently checks every supported release surface — npm, OCI, the MCP Registry, and GitHub — and fails closed on stale, unreachable, or inconsistent state, rather than trusting any one surface’s self-report.
Certificate clarity. Human-readable terminal, HTML, and Markdown certificates now include relying-party summaries, consistent measurements, and a plain-English glossary informed by external reviewer feedback. OCI distribution readback now authenticates before verifying the published image’s attestation and tagged source, and public-distribution validation rejects location fields and private-path-shaped text in public transparency artifacts.
v0.4.1
2026-08-12Byte-identical reports regardless of checkout location. report.json no longer embeds the absolute checkout path, so identical repository content and revision scanned in different directories produce byte-identical report artifacts. Reports from v0.4.0 and earlier keep their recorded path and remain verifiable.
Nothing else changed. This is a single-fix patch release: no scoring, rubric, certificate, or CLI behavior change. Prospective rubrics and in-development detector proposals remain unreleased.
v0.4.0
2026-08-09Stronger publication verification and two explicitly bounded prospective tools. The distribution workflow now reads the published MCP Registry record back, requires its immutable OCI digest, and verifies that digest’s signed provenance against the exact release tag and source commit. The release also carries prospective rubric v19 and a resource-bounded v2 discovery collector for explicit evaluation harnesses.
No silent scoring promotion. The v19 paired rescore completed all 24 rows with no score, status, coverage, placement, or non-B4 changes. It remains prospective; the public CLI default is still the separately holdout-calibrated v17 rubric. The v2 collector does not change the SHA-pinned v1.9 contract or ship the reserved Free LLM command family.
Inspect the v0.4.0 release → · Read the release evidence boundary →
v0.3.2
2026-08-07The published npm, Docker / OCI, standalone-binary, and MCP Registry release. Its published GitHub Release includes five native binaries, SPDX SBOMs, SHA256SUMS, and signed provenance. The source record names the v0.3.2 commit.
v0.3.1
2026-08-07The v0.3.1 npm and Docker / OCI release. Its source record names the v0.3.1 commit, so a reader can inspect the tagged source rather than rely on a release claim.
v0.2.2
2026-07-29Windows binaries — Cejel now ships a standalone single-file executable on all five platforms: macOS (Apple silicon + Intel), Linux (arm64 + x64), and Windows (x64). No Node, no npm, nothing installed. Every binary has an SPDX SBOM and SHA-256 checksum; the release also includes GitHub-signed build provenance, while own-platform verification receipts remain attached to the guarded build — because a trust tool you can’t verify is just another black box.
It runs inside your AI coding agent. Cejel is now documented as an MCP server you can add to OpenClaw (or any MCP client) through the shipped cejel-mcp bin, so an agent can check a repository’s evidence before you rely on it. It scans code you point it at — it does not watch or govern the agent’s actions.
Sharper certificates. Certificates now show both the producing CLI version and the exact rubric version, so reports from different installs explain their scoring identity. A dimension band that differs from the weighted score now carries an inline reconciliation instead of a bare number next to “verified.” A scan of a source tarball now warns when Git history was unavailable and a recent-PR metric may undercount. And the install docs force @latest to sidestep a stale npx cache. Thanks to the early testers who reported these.
Also folded in from main: hardened Git transport at the clone/checkout sinks, the free-core v50 multiple-comparisons disclosure, and a cross-repo preflight gate. Full commit-linked detail: CHANGELOG.md.
Earlier releases
0.2.1 and before0.2.1 was an npm-only certificate-presentation patch. 0.2.0 established the scoped public distribution: standalone binaries for macOS and Linux, the OCI image, and the Official MCP Registry listing. The complete history lives in the repository changelog.
Run the latest
Free, offline, deterministic. Nothing installed, nothing leaves your machine.