You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
ci: a release-check workflow, run on demand on the release PR: update path, configs, audit log, mixed versions, stress, changelog, timing report #199
Before a release, the release PR is checked for compatibility with the last releases. Today an agent does this by hand from written steps, on another machine. It should be a workflow on GitHub runners, run only when the owner decides to release, not on each update of the release PR.
What to build
.github/workflows/release-check.yml, started by workflow_dispatch on the release-please branch. It records the commit it checks (the candidate) and writes one result file (JSON artifact + job summary) with each check's result and evidence.
One build script for release.yml and this workflow (scripts/build-release.sh or a reusable workflow): the same Go version, CGO_ENABLED=0, architectures, -trimpath, ldflags, files, tar and gzip settings, and SHA256SUMS. release.yml uses it too; signing and publishing stay in release.yml.
N-1 and N-2: the two highest published, non-draft, non-prerelease vX.Y.Z releases below the candidate (not the nearest tag: a failed release can leave a tag or a draft).
Update path: the real N-1 and N-2 binaries run nodkey update (and their own install.sh) against a local HTTPS fixture that serves the candidate's files. Also: a missing or wrong checksum, and a missing file, are refused and keep the old binary. The production signer cannot sign a candidate, so the signature step is simulated here; ci: check the draft release's files before publishing (checksums, attestation, version, clean install) #200 checks the real one.
Configs: a corpus of configs that N-1 accepts (test/compat/configs/, from the docs of each release, plus refused cases) loads with the candidate the same way.
Audit log and state: N-1 writes them, stops; the candidate goes on with no tamper alarm; audit verify passes with both; a changed or cut copy is refused.
Mixed versions: the end-to-end test takes NODKEY_E2E_BROKER_BIN and NODKEY_E2E_CLIENT_BIN (default: build from source, as today), and runs candidate/candidate, N-1 broker/candidate client, candidate broker/N-1 client. Release builds send version notices: the tests expect them.
Stress: the full suite with -race -count=5 -cpu 1,2,4.
Timing report (never fails the workflow): an agent-browser login (plugin → broker → a fake bw that sleeps 6 s per call), cold and after a pre-approval: the time against PluginResolveTime (12 s), and the number of bw calls. A reader judges it; timing on shared runners is not stable enough to fail on.
The workflow never needs a secret beyond the default token, and never publishes.
Done when
release-check.yml runs checks 1-9 by workflow_dispatch and writes the result file.
release.yml builds with the shared script (the files of a build are the same as before: checked byte for byte on one version).
The end-to-end test takes the two binary variables.
docs/releasing.md says how to run it and how to read the result.
gh workflow run release-check.yml --repo code6x/nodkey --ref release-please--branches--main
release.yml builds with the shared script, and the files are the same. Yes. The old inline build of release.yml and scripts/build-release.sh, both for version 0.8.0 with Go 1.27.2, gave the same four tarballs and SHA256SUMS, byte for byte (diff -r empty; the hashes are in ci: a release-check workflow, run on demand on the release PR #215).
The end-to-end test takes the two binary variables. Yes: NODKEY_E2E_BROKER_BIN and NODKEY_E2E_CLIENT_BIN, plus NODKEY_E2E_FIRST_BROKER_BIN for check 5. Without them it builds from the source, as before.
docs/releasing.md says how to run it and how to read the result. Yes: "Check a release candidate".
The timing report never fails the workflow. In the first run, the login after a pre-approval took a few milliseconds, with 0 runs of bw. The cold pre-approval took about 18 s, with 3 runs of bw at 6 s each.
Activity
groupsky commentedon Nov 18, 2026
hq-jobs
nodkey-release-ci-2(runr-20261118-210019-7f96)Done in #215 (merged). What each "Done when" box shows:
workflow_dispatchand writes the result file. Yes. A first run, onmain: https://github.com/code6x/nodkey/actions/runs/44382300500 (artifactrelease-check-result, and the job summary).mainthe manifest says 0.7.0, so the candidate was "0.7.0" with N-1 0.6.0 and N-2 0.5.0. Checks 1-6 and 9 passed.internal/brokerTestLoginRememberedTimer. It also failed CI onmain. Filed as test: fix three flaky tests that the release check and CI found #217, with two other flaky tests.workflow_dispatchuses the workflow file of the chosen branch.release-please--branches--maindoes not haverelease-check.ymlyet: release-please changes that branch only when the release notes change, and ci: a release-check workflow, run on demand on the release PR #215 and ci: verify the draft release's files before publishing #216 areci:(hidden). After the nextfeat:,fix:ordocs:merge, run:release.ymlandscripts/build-release.sh, both for version 0.8.0 with Go 1.27.2, gave the same four tarballs andSHA256SUMS, byte for byte (diff -rempty; the hashes are in ci: a release-check workflow, run on demand on the release PR #215).NODKEY_E2E_BROKER_BINandNODKEY_E2E_CLIENT_BIN, plusNODKEY_E2E_FIRST_BROKER_BINfor check 5. Without them it builds from the source, as before.docs/releasing.mdsays how to run it and how to read the result. Yes: "Check a release candidate".The timing report never fails the workflow. In the first run, the login after a pre-approval took a few milliseconds, with 0 runs of
bw. The cold pre-approval took about 18 s, with 3 runs ofbwat 6 s each.