Skip to content
Trazum

1.80.0

The doors somebody else walks through

Added#

  • DCO sign-off, enforced in CI. .github/workflows/dco.yml walks base..head on every pull request and fails on any commit without a Signed-off-by: line, naming the SHA and its subject. The failure prints both remedies verbatim: git commit -s for work not yet committed, git rebase --signoff <base> for commits that already exist. A guard that fails without a remedy is one people learn to route around.

    Read-only permissions, the plain pull_request event, and fetch-depth: 0, because a shallow checkout would walk an empty range and report clean while examining nothing. Merge commits are skipped: GitHub writes those itself when a branch is updated from the web.

    Proved by planting the violation, in a scratch repository for the script and in the workflow file for the guard: an unsigned commit fails and names itself, a sign-off with no address still fails, a merge commit does not fire, an empty range passes rather than erroring, and each of the seven pinned facts fails when removed. Two of those plants found the guard reading the wrong thing rather than the workflow being wrong: \s+ matched the job-level contents: read when the top-level one was flipped to write, and the sign-off assertion was satisfied by a header comment rather than by the grep. Both are tightened.

  • A sign-off section in CONTRIBUTING.md, stating plainly that the DCO is a certification of origin and not a copyright assignment, linking the text being certified, and giving the reason a single-maintainer project bothers: a repository that cannot account for the provenance of its own history cannot answer a corporate legal review and cannot change its own licence. Guarded, so the document and the workflow cannot drift apart.

    CodeQL flagged that last guard twice, and was right twice. It began as /developercertificate\.org/, which also matches https://notdevelopercertificate.org/; anchoring it to the full origin fixed that and left the real complaint standing, since an unanchored URL-shaped pattern still matches inside a longer URL. The subject was never a URL: it is a markdown document, and the question is whether it links the text. So the assertion is a containment check on the exact link target now, with no regex at all. It rejects a bare mention in prose, a subdomain impostor, and the domain embedded in somebody else's query string, none of which the first version rejected.

  • docs/licensing.md, for somebody deciding whether to depend on this. It answers the three questions LICENSE leaves open: what the licence covers today, what will never move out of it, and what it never covered at all.

    The promise it makes is that no analysis this repository can perform today moves out of the open set, and the argument is not generosity: the case for trusting a figure here has always been that the code producing it can be read, so a withheld capability would break the product's thesis before it broke its licence. What is not open is anything needing infrastructure somebody else operates. None of it exists, and if it is ever built it ships proprietary from its first commit rather than being moved out of the open set. What is reserved is the name, which the licence has never covered.

    The list of what is open is derived, not typed. licensing.test.js reads every non-private package manifest plus each shipping surface directory and fails when one is missing from the page, checks that every published package really declares MIT, pins the three promises by the shape of their claim so the prose can be rewritten and the promise cannot vanish with it, and asserts the page is linked from the section a deciding reader actually reads. This repository has four recorded cases of a hand-typed fact going stale (1.51.2, 1.53.2, 1.53.3, 1.60.3), which is why nothing here is typed.

    Proved by planting eight violations: a published package dropped from the page, a shipping surface dropped, a brand new package nobody named, a package declaring GPL-3.0, each of the three promises reworded away, and the page unlinked from the index section. All eight fail, and each names what went missing.

    One of those six assertions bounded a section of docs/README.md by naming the heading that follows it, which is the failure publish.test.js has ratcheted against nine times before. It caught this one on the commit that tracked the file, not before: that guard walks git ls-files, so an untracked test is invisible to it and the rule bites the moment the file is staged. Rewritten to use sectionOf, and checked by inserting a section between the two headings, which the old form would have swallowed.

    Two open items are named on the page rather than implied: Apache-2.0 is under consideration and undecided, and the licences of vendored and development dependencies have not been audited. Runtime dependencies outside this repository are zero by design and a guard already holds that.

Nothing installable changes: no package code, no command, no flag.

  • packages/mcp/server.json and an mcpName on the MCP package, the two files the official MCP registry reads. The registry hosts metadata rather than artefacts: it records that a name maps to a package on npm and verifies the claim by reading mcpName back out of that package, so three facts in two hand-edited files have to agree.

    publish.test.js holds them in the same lockstep as the six version manifests, and docs/releasing.md names the file in the release checklist. The namespace prefix is pinned as well: GitHub authentication grants only io.github.<user>/*, and a name off that prefix is refused at publish time by a registry the maintainer must already be logged into to find out. Cheaper to fail in CI.

    Proved by planting six violations: a version left behind in either place, mcpName removed, the two names drifting apart, both names moved off the granted namespace consistently, and a pointer at a package this repository does not publish.

    Publishing itself is not done here. It needs mcp-publisher login github, which is the maintainer's account. This is the repository half.

Nothing installable changes: no package code, no command, no flag.

Changed#

  • The Action's one line names both of its gates. action.yml builds two argument vectors, check against a token budget and profile against a spend budget, and seven of its eighteen inputs exist only for the second. Its description said "Token budget for prompts" and stopped, so the single sentence the Marketplace shows described a version that stopped being the whole story several releases ago. Every other assertion about this file is about its inputs and its script; nothing looked at the sentence on top.

    security.test.js now reads the modes out of the script, requires the description to name what each one gates, and checks the 125-character limit the Marketplace enforces. Proved by planting four violations: the old description, the spend gate dropped from the sentence, a description over the limit, and a third mode with no phrase pinned for it, which fails naming the mode rather than passing quietly.