Skip to content
Trazum

2.7.11

Released 2026-11-17

Added#

  • An issue comes in with a form, labels and a triage state. Three new forms cover what had none:

    • a wrong price or a missing model, which asks for the provider's own page, the quote and the day it was read;
    • a feature request, which starts with the problem;
    • a documentation fault.

    Every form, including the two existing ones, starts its issue as needs-triage. The labels and what each means now live in .github/labels.json, and a Labels workflow makes GitHub's list match it using only the gh CLI. Until now the only list was GitHub's settings page, and no issue carried a label. CONTRIBUTING.md gains an Issues section: which form to use, the four triage states, and that an issue is never closed for going quiet. SUPPORT.md links every form. issue-intake.test.js holds it together:

    • every label a form or dependabot.yml applies exists;
    • every label is explained in CONTRIBUTING.md;
    • every form is linked from SUPPORT.md;
    • every page a form links to exists;
    • the workflow never deletes a label.

    Eight faults planted one at a time were each caught. The bug form's version hint no longer says there is no npm package. (#665)

  • A pull request is checked for its issue and for where its changelog line landed. The new PR readiness check, run by scripts/pr-readiness.mjs on every pull request, refuses three things:

    • a description without a closing keyword (Fixes #<n>);
    • a line added inside a released version's section of CHANGELOG.md;
    • a pull request that adds no line under Unreleased.

    The second is what happened on 2026-11-17: a branch merged main after the 2.7.9 and 2.7.10 releases, git placed its entry inside ## 2.7.9 without a conflict, and changelog-coverage.test.js, which can only ask whether Unreleased is empty, had nothing to say.

    Exemptions:

    • Dependabot;
    • the release pull request, which adds a version heading;
    • two labels only a maintainer applies, no-issue and changelog-history.

    Rebuilt from that merge, the check refuses it. It accepts all 40 of the latest commits on main and the 2.7.10 release itself. The description reaches the script through the environment, never inlined into the shell. CONTRIBUTING.md now says how a pull request lands: issue first, squash only, current with main, and a look at the changelog line after a release. The PR template opens with Fixes #. .github/rulesets/main-branch.json now allows only squash, requires PR readiness, and carries the code_scanning and code_quality rules that the ruleset GitHub enforces already had. An admin re-imports it. (#666)

Fixed#

  • Google's prices were re-read, and gemini-2.5-pro bills at Gemini 3.8 Flash's rates from 2026-12-01. Google's Gemini API docs moved to aistudio.google.com (the old ai.google.dev pricing address now redirects there), the November review could not reach them, and Google's figures stayed dated 2026-10-10, so the CLI would have warned about every one from 2026-11-25. The pricing, deprecations and caching pages and each row's model page were read on 2026-11-16: every price, cache and batch price, storage price, context window and cache minimum matched, and Google's review date moves. What is new is that "effective December 1, 2026" requests to the retired gemini-2.5-pro route to Gemini 3.8 Flash and are billed at its rates, $0.75/$3.75 through 2026-12-31 and $1.50/$7.50 after, against $1.25/$10 today. The id now carries both periods, as grok-4 does, and a log spanning the day is billed both ways. gemini-2.5-flash routes to Gemini 3.5 Flash-Lite that day, whose text rates are the ones it already carries. trazum models and the MCP list_models tool print a row's current figures, so until 2026-12-01 they show 2.5 Pro at 3.8 Flash's price; a call in a log is priced on its own date. Not changed: the Batch tabs of 3.1 Pro Preview and both 2.5 models charge the standard cache price, and a batch call's cache reads are still billed here at half of it.

  • A release never asked npm about the credential that uploads it. The release job checked the trusted publisher, which refused this workflow at 1.86.0 and refuses it still, and said nothing about the granular token the uploads go through on. That token has an expiry. 2.7.10's release pull request merged on 2026-11-17, the job ran five minutes of checks, and the first upload answered E404 Not Found - PUT, under a failure note that said of a job holding the token that "this failure is something else". Nothing was half-published, and main named a version npm did not have. Three changes, none of which alters anything installable. The job now asks npm whoami with the stored token before verify, on a manual dry run too, so a dry run from a release branch answers before the merge; a refusal is an error on the run page and in the job summary with the steps that fix it, and the failure note asks again and ends on the answer, as does a dry run, in one word. Measured against the live registry: a token npm does not know is answered 401, and so was the stored token itself, on a dry run of this change 42 minutes after the failed upload; the token that replaced it was answered 200 that afternoon, and thirteen minutes later was refused the upload with EOTP, because it had not been made to publish without a one-time code. So the step reports and does not fail, an accepted token is worded as less than a promise, and the failure note names EOTP beside E404. The token's expiry date is now recorded in .github/release-credentials.json (null until the owner writes it down), and scripts/freshness.mjs calls it overdue from fourteen days before, which turns the weekly scheduled run red on main; a date nobody recorded is printed as a note that says so. And scripts/prepare-release.mjs refuses to prepare a release once the recorded date has come. The token reaches two more steps of release.yml than it did, both running one script of this repository that imports nothing outside Node and sends the token to npm's registry or to nowhere, and security.test.js holds the list of steps to exactly those. Sixty-one faults planted in the check, the workflow, the date arithmetic and the two scripts that read it were each caught. docs/releasing.md gains "When the token expires", and loses the paragraph that said there was no token to rotate.