Two-phase scheme instead of the sequential gate: the build job runs in
parallel with test/e2e jobs and only warms the buildx GHA cache
(push:false, cache-to mode=max); a new publish job (needs: test,
e2e-server, e2e-mcp, build) rebuilds from the warm cache (near-instant
on hit, full rebuild on eviction — same as the old sequential timing)
and pushes :develop. GHCR login moved to publish; build-args blocks are
kept textually identical between the two jobs so the cache hits.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Reverse the previous policy where e2e jobs only turned the run red
without blocking the image publish: build.needs now lists test,
e2e-server and e2e-mcp, so a failing test of any kind stops the
:develop image from being built and pushed. Stale policy comments
updated accordingly.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
PR #333 deliberately changed the canonical markdown export of callout
nodes to the Obsidian-native format ('> [!type]' + blockquote body,
pinned by packages/prosemirror-markdown unit tests); the importer still
parses both ':::type' fences and '> [!type]'. The get_page e2e assertion
was missed in that switch and still expected ':::info', failing the
e2e-mcp job on develop since 4369bbc5.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
agent_vscode