[docs, ci] feat: add bump-dependency skill and fix L2 trigger docs - #3686
Conversation
Signed-off-by: oliver könig <okoenig@nvidia.com>
Light Code ReviewDocs-only PR adding skills/bump-dependency/SKILL.md. Cross-references (@CONTRIBUTING.md, @skills/build-and-dependency/SKILL.md, @skills/cicd/SKILL.md, @skills/testing/SKILL.md) all resolve. The active/flaky directory layout, override-dependencies pattern, and needs-more-tests label semantics match the repo. Minor issue — Nested code fences in Step 4 (lines 100-130): The PR body template contains triple-backtick code fences (lines 109-111, the Lockfile delta block) inside an outer bash fence. When rendered by GitHub markdown parser, the inner backticks close the outer fence prematurely, causing the rest of the block (the gh pr create command) to render as regular text. Fix: use four backticks for the outer fence, or indent the heredoc body so the inner backticks are not at the same nesting level. Everything else looks good - the watchdog script logic is correct, the pitfalls table is accurate, and the anti-patterns section is well-motivated. Suggested test cases: No perf tests impacted. |
Signed-off-by: oliver könig <okoenig@nvidia.com>
Signed-off-by: oliver könig <okoenig@nvidia.com>
Signed-off-by: oliver könig <okoenig@nvidia.com>
|
|
||
| ### Watchdog script | ||
|
|
||
| Save to `/tmp/watchdog-<PR>.sh` and chmod +x: |
There was a problem hiding this comment.
may want to prefix this with repo as well to avoid potential clash?
…3686) Signed-off-by: oliver könig <okoenig@nvidia.com>
…VIDIA-NeMo#3686) Signed-off-by: oliver könig <okoenig@nvidia.com> Signed-off-by: Vasudevan Rengasamy <vrengasamy@nvidia.com>
Claude summary
What
skills/bump-dependency/SKILL.md— end-to-end workflow for shipping a dependency-bump PR (TE / MCore / git-pinned override) and driving it to green.full-test-suitelabel, not only on schedule /workflow_dispatch.@<path>references. The skill now keeps only the bump-specific narrative — orchestration arc, watchdog script, iteration loop — and trims sections like worktree creation,uv lockmechanics, sign-off boilerplate,/ok to testsemantics, and thegit mvquarantine recipe to one-line pointers (-97 lines, +61 lines = net -36 lines).Why
Distilled from the recent TE
release_v2.15bump (#3672). Two things were notable enough to bake into the skill:Monitortool) — never a cronjob or scheduled wakeup. Wakeups poll on a clock; a watchdog reacts on every job edge in theCICD NeMoworkflow and tears down cleanly viaTaskStoponce the run is green.CICD NeMountil green — quarantine real flakes bygit mv-ing launch scripts into the matchingflaky/dir, append them to the PR body's "Quarantined tests" section, repeat.While writing the skill I confidently propagated a wrong claim — that L2 is schedule-only. Reading
.github/workflows/cicd-main.ymlshowsEXPECT_L2=trueis also set whenFULL_TEST_SUITE=true, i.e. when the PR carries thefull-test-suitelabel (line 152). Fixed everywhere it was wrong.Files touched
skills/bump-dependency/SKILL.md— new skill,full-test-suiterecommended for high-blast-radius bumps; trimmed against cross-referenced skills.skills/testing/SKILL.md— Tier Semantics table updated.skills/cicd/SKILL.md— pipeline tree updated.CONTRIBUTING.md— Tier table, Risk Labels table, and PR labeling checklist all gainedfull-test-suite..github/workflows/cicd-main.yml— comment abovecicd-functional-tests-l2corrected (no behavioural change).Test plan
@CONTRIBUTING.md,@skills/build-and-dependency/SKILL.md,@skills/cicd/SKILL.md,@skills/testing/SKILL.md).pyfiles touched → no functional CI required (docs-onlylabel)