workflow_dispatch with version/push/latest inputs: push defaults to
false, so the runner, socket and registry login can be verified without
creating a tag or a package version. latest is never moved unless asked.
The release-please job never had a checkout because the action clones into its own temp directory. The new reconcile step runs a file from the workspace, so the first post-merge run failed with MODULE_NOT_FOUND for scripts/reconcile-release-please.mjs. Add actions/checkout@v4 before the label/reconcile steps; the action still uses its own temp clone for the release branch.
Both PR-Agent findings were confirmed. A tag at the merge commit now counts only when it is exactly v<manifest version>; foreign CI or annotation tags are logged and ignored. Pending release PRs are fetched through the autorelease: pending label filter instead of paging every closed PR, so a concurrent close cannot shift a page and hide one.
Also fail loudly on a merged PR without merge_commit_sha or a manifest without a version, instead of producing vundefined.
Rebased on main after the separate audit fix (e0c69c1); the dependency overrides and lockfile changes from the earlier local commit were dropped. Verified on the rebased tree: lint, typecheck, build, the exact audit command, actionlint and a live dry run.
The registry endpoint is the Gitea hostname itself, not a separate
registry.noelsabia.dev host. Point IMAGE, docker login and the v2
manifests check at git.noelsabia.dev and record the correction.
Renovate fetches changelogs and release notes from github.com. Anonymous
requests are rate-limited, so the first run skipped them and printed the
"no github.com credentials" warning in the PR body.
RENOVATE_GITHUB_COM_TOKEN now comes from a repository secret. It is
independent of RENOVATE_TOKEN: any github.com account with read-only
access is enough, and no scopes are needed for public repositories.
Run PR-Agent as a pinned webhook server on the Gitea Docker network and review on PR open plus slash commands. The per-PR opencode installer and verdict-posting step are removed.
The server-side .env stays outside the repository; deploy/pr-agent/ holds the compose definition and operator notes.
Gitea has no hosted Renovate, so the repository runs the official image
from a scheduled Gitea Actions workflow. The platform settings stay in the
workflow environment because Renovate accepts them only in the self-hosted
config, not in renovate.json.
renovate.json keeps the repository config minimal: schema plus
automerge=true, which also covers major updates. The bot account and the
RENOVATE_TOKEN secret are the remaining manual steps.
release-please finds merged release PRs by the 'autorelease: pending'
label. Gitea resolves label names with GetLabelIDsInRepoByNames, a SQL IN
filter that silently drops unknown names, so on a repository without the
labels the PR is created unlabeled and every run finds zero merged release
PRs: no tag, no release, and the next release PR instead - with a green
run each time. That is what happened to #20 and #21.
The workflow now ensures both labels exist (idempotent: list, create the
missing ones) and gained workflow_dispatch so a missed release can be
replayed.
The review's blocker was real: runs-on: ubuntu-latest is the label
pr-checker.yml uses, and the job needs a runner-wide Docker socket, so the
only working configuration violated the trust boundary in the decision
doc. The job now runs on a dedicated 'publish' label that only a second
act_runner registration (docker_host: automount) may carry.
Version tags are now immutable: before building, the registry v2 API is
asked for manifests/<version> and 200 aborts the run. docker manifest
inspect cannot do this job - the pinned Debian 20.10 client fails on OCI
indexes for existing and missing tags alike (measured).
Also: the smoke test requires a rendered body (10 kB floor) on top of the
2xx, .github/actionlint.yaml declares the custom label, and the actionlint
hook passes that config explicitly because this repo keeps workflows in
.gitea/workflows, where actionlint does not auto-discover it.
Applied the findings that held up against the code:
- strict vMAJOR.MINOR.PATCH guard; v[0-9]* accepted v1, v1-foo and
prereleases, all valid registry tags
- latest now follows the newest release, so replaying an older tag cannot
drag it backwards; skipped with a warning if the tag list is unavailable
- smoke test starts the built image and requires 2xx on / before pushing
- timeout-minutes: 20 on the job, so a hung build cannot hold the
capacity-1 runner indefinitely
- the decision doc records the socket's blast radius: root-equivalent,
runner-wide, only on a label PR jobs cannot use
Rejected with evidence: the .dockerignore leak (.env*, *.pem, .git are
excluded) and the Docker client/daemon API mismatch (20.10.24 client
against a 29.9.2 daemon builds over the socket).
Build Dockerfile on every v* tag push (created by release-please when the
release PR merges) and push registry.noelsabia.dev/cv_website/nuxt as the
version, sha-<12> and latest. Main pushes do not build: the registry holds
releases, not commits.
The job runs in node:22-bookworm and needs only the Docker client, so the
act_runner must expose its host socket (container.docker_host: automount);
a preflight step fails with that cause when it does not.
The diff filter drops pnpm-lock.yaml to keep the patch small; the model read
that as an uncommitted lockfile and raised a BLOCKER on PR #16. Say so in the
prompt instead of widening the diff.
- review model: opencode-go/claude-haiku-5-5#medium instead of
deepseek-v4.1-flash#max; flash has no medium variant
- merge lint and typecheck into one job with a single pnpm install
- cache the pnpm store and opencode binary via actions/cache@v4.1.0,
keyed on a sha256 of the lockfile (hashFiles is empty on this runner)
- update the decision records and log
The 401 in run 89 was the final isomorphic-git push inside
code-suggester's createPullRequest, not an API call: tree, commit and
local ref update all ran through the token-authenticated octokit API,
then 'git push' used the git-username/git-password inputs - unset in
our workflow, so the push was anonymous. The remote branch
release-please--branches--main still sat at an old commit, proving the
push never landed. Pass the account name and the same token.
The dedicated RELEASE_PLEASE_TOKEN_CV_SITE_NUXT secret authenticated for
reads but threw when creating the release PR (exit 1, releases_created
false, no PR created): it lacks write:repository / write:issue. Use the
proven PR_GITEA_TOKEN instead, per the one-well-scoped-token policy in
OKF_Nuxt/decisions/snyk-paused-and-secrets-policy.md.
Findings validated from review comments 71 and 75:
- [BLOCKER] the no-verdict branch exited \${EXIT_CODE:-1} which is 0
when opencode succeeded without a verdict - job went green and the
post step published non-review output. The else branch now exits 1
unconditionally, the verdict is anchored to the last five lines (the
prompt itself contains VERDICT lines), and the post step re-verifies
the contract before posting.
- [BLOCKER] the v1 'permission' object is ignored by the opencode v2
CLI - shell commands executed in both review runs despite deny. The
config now uses the v2 'permissions' array with 'shell' as the
action name, denying edit/shell/webfetch/websearch/external_directory.
- hygiene: droast linting prunes node_modules, audit comment names the
exact key to re-add, shell blocks verified with sh -n.
- release-please-config.json: node release type, root package,
plain v-tags, default versioning
- .release-please-manifest.json: track current version (1.0.0)
- package.json: add version field required by the node release type
- release-please.yml: use the config/manifest files
- droast: chmod +x was lost in the checksum rewrite (would have
failed with exit 126); checksums now pinned in the workflow
instead of trust-on-first-use
- permission config migrated to the OpenCode v2 schema (permissions
array, bash renamed shell, websearch and external_directory denied)
- posting gate is now the output contract (verdict line present)
instead of the exit code: opencode can exit 1 on benign tool
denials after a completed review
- find -print -quit instead of piping into head (SIGPIPE, exit 141)
- audit comment now says to re-add continue-on-error
- [BLOCKER] remove opencode.json from the checkout: it held the API
key in plaintext, readable by the review agent, and a malicious PR
diff could prompt-inject the agent into quoting it in the publicly
visible review comment. Env-only auth (verified headless). The
permission config moves to $HOME/.config/opencode - outside the
checkout and secret-free.
- [WARN] verify the droast binary against the publisher's
sha256sums.txt of the same pinned release
- [WARN] pin the opencode v2 installer to 2.0.32
Findings validated with the validate-pr-review skill; findings 4/5
are moot with the blocker fix.
- deny edit/bash/webfetch in the generated opencode.json: headless
permission prompts are the most likely cause of the review agent
working silently for minutes and then dying; denying write tools
also makes the reviewer provably read-only
- capture the exit code without triggering 'set -e' ('|| EXIT_CODE=$?')
so the captured output is always printed before failing - the
previous version aborted before 'cat' and produced silent failures
The classic installer (opencode.ai/install) ships only OpenCode 1.x
- v1.19.3 on arm64 - which cannot resolve model variants and failed
with ProviderModelNotFoundError for deepseek-v4-pro#max. Use the v2
installer (currently v2.0.32, same version as local dev); verified
working in a fresh arm64 container with key-only auth.
The dockerfile-lint job skipped on every PR despite Dockerfiles
existing: the act runner's hashFiles() evaluation returned empty.
Replace it with an explicit file-discovery step. Also run the droast
binary in node:22-bookworm (arch-aware install) instead of the
vendor image, which ships no node and would fail actions/checkout
with exit 127 - same failure mode as the gitleaks image.
The headless CI container has no stored credentials, so auth depends
entirely on the injected key. Generate an opencode.json at runtime
that binds OPENCODE_API_KEY to the opencode-go provider explicitly,
removing any dependence on env-var-to-provider mapping behavior.
- audit no longer runs with continue-on-error: a failed audit must
turn the pipeline red, not pass silently
- opencode review runs with --print-logs so auth and provider errors
(e.g. invalid API key vs missing Go subscription) appear in the job
log instead of an opaque exit 1
main changed the review model to the default thinking budget; keep
opencode-go/deepseek-v4-pro#max (plan-included, max effort) which
supersedes both earlier variants.
The 'opencode' provider prefix points at OpenCode Zen (pay-per-call,
no funds), which failed every review run with 'Insufficient account
funds'. The opencode-go $10 plan covers deepseek-v4-pro with high
and max variants - use opencode-go/deepseek-v4-pro#max.
#max exceeded the available opencode account funds; the default
thinking budget keeps the flagship model at lower cost. #max remains
available by re-adding the variant.
Passing the PR diff as a command-line argument exceeded
MAX_ARG_STRLEN (128KiB per single argument; the diff was 190KB,
mostly pnpm-lock.yaml). The shell failed with E2BIG ('Argument list
too long') and the tee pipeline masked the failure, so the raw error
log got posted as a 'review' comment.
- exclude pnpm-lock.yaml from the review diff (pnpm audit covers deps)
- write the diff to .pr-review-diff.patch in the workspace; the agent
reads it with file tools
- redirect output instead of piping through tee and fail the job when
opencode fails, so errors are never posted as review comments
- model: opencode/deepseek-v4-pro#max (requires funds on the opencode
account; flash fallback documented in a comment)
- review prompt rewritten: SSR/hydration/reactivity pitfalls, Nuxt
server-route security, runtime config leaks, Go error handling,
concurrency and resource hygiene, frontend/backend contract checks
- strict output format with [BLOCKER]/[WARN]/[NIT] severities mapped
to the approve/request-changes verdict
- Snyk job disabled via 'if: false': free plan only has ~100 test
credits/month and the job runs 2 tests per PR push
- PR comment step no longer fails silently: curl prints the HTTP
status and response body, making token/URL errors visible
Dockerfile linting is already covered by the droast job; the second
Snyk step now scans our own source for code vulnerabilities (XSS,
injection, unsafe patterns) instead of the base image.
The pnpm-workspace.yaml makes Snyk treat the project as a package
workspace; without --all-projects it aborts with SNYK-CLI-0000
(422 Unprocessable Entity).
The snyk/snyk:node image is amd64-only and cannot start on an arm64
runner. Install the Snyk CLI through npm (supports linux/arm64) in
the multi-arch node:22-bookworm image instead.