Skip to content

add google/vertex-model-retirements - #238

Open
adssib wants to merge 1 commit into
pl5-geminifrom
pl5-vertex
Open

adssib wants to merge 1 commit into
pl5-geminifrom
pl5-vertex

Conversation

@adssib

@adssib adssib commented Nov 21, 2026 •

Copy link
Copy Markdown
Contributor

Stacked on #237

Part of #37 (the Google provider; audit item google/vertex-model-retirements)

What this changes

New source google/vertex-model-retirements (kind notices, layer ai-models): Google Cloud's model versions page, https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/model-versions (the old cloud.google.com/vertex-ai/…/model-versions redirects there; re-checked 2026-11-21, 200). It has HTML tables | Model ID | Release date | Retirement date | Replacement model ID | (or "Recommended upgrade"), and gives one model event per model with a plain retirement date (39 today).

  • The retire reader reads HTML tables (<code> becomes a code span, headings kept).
  • A GA model's "October 13, 2027 or later" / "No sooner than May 20, 2028" is a minimum lifetime, not a notice: it doesn't parse as a date, so no event.
  • A model id with a footnote mark (textembedding-gecko@003 *) isn't read.
  • The existing google/vertex-ai-models (the release-notes feed) stays. It announces; this page is where the dates are kept. The same model can be an event in both, under different source ids.

Checklist

  • Metadata complete in the registration (packages pypi:google-genai, pypi:google-cloud-aiplatform, npm:@google/genai)
  • Golden over an excerpt of the real page: every heading and table line, unchanged. It gives the same 39 retirements as the full page.
  • testdata/README.md row
  • Pinned drop test: the Gemini image models' table, four "… or later" rows and one dated retirement, gives one event. The same table with plain dates would give five.
  • providers/google-vertex-model-retirements.md
  • make check passes

Related PRs

🤖 Generated with Claude Code

@claude

claude Bot commented Nov 21, 2026 •

Copy link
Copy Markdown

Claude finished @adssib's task in 1m 23s —— View job


✅ Ready to merge: the source is registered with complete metadata, and it has a golden test over a real excerpt, a meaningful pinned drop test and a provider page. Every check passes locally.

Rules

Rule Status Evidence
Metadata in the registration (Rule 1) ✅ plugins/notices/notices.go:85-90: id, layer ai-models, Watches, URL, packages (pypi:google-genai, pypi:google-cloud-aiplatform, npm:@google/genai), page, Wayback archiver. No hand-kept list was added. I read the words and they match what the code does.
Provider page ✅ providers/google-vertex-model-retirements.md has every section of _template.md. Why this source has the link, the date checked (2026-11-21), the reason, three alternatives and a quote. I found the quote (veo-3.0-generate-001 … June 30, 2026 … veo-3.1-generate-001) in the saved HTML at lines 803-806. The cross-reference in google-gemini-models.md is updated too.
Golden test + dated quote ✅ testdata/golden/google-vertex-model-retirements.json is a new file, not an edit. vertex-model-versions.html is an excerpt of the page, and testdata/README.md has its row: URL, date, license (CC BY 4.0) and how it was trimmed. Every event has a Wayback source_url and a source_excerpt that includes the heading, the header row and the model's row.
Pinned drop test ✅ TestVertexDropsAMinimumLifetime runs on the Gemini image models' table and expects exactly 1 event. It then strips " or later" and checks that the same table gives ≥2. That second check proves the drop really comes from the "or later" wording.
Fetch only via env ✅ No net/http, no filesystem access and no database in the new code. The new reader only transforms bytes (htmlPipes), and fetching still goes through the existing Plugin.
Nothing private (Rule 2) ✅ The only URLs are public docs.cloud.google.com and Wayback links. No names, tokens or internal hosts. leak_guard.py --all is clean.
Contract fixtures ✅ Not touched.

Bugs

None found. What I checked about precision:

  • dateHeader matches only ^(shutdown|retirement) date$, so the "Release date" column can never be read as the retirement date.
  • "… or later", "No sooner than …" and "No retirement date announced" all fail parseDate, so they give no event. These rows also mark the model as seen, so a different row for the same model can't put a date back.
  • In textembedding-gecko@003<sup>*</sup>, the tags are replaced by spaces, so the cell becomes `textembedding-gecko@003 *`. That fails googleModel, so the model is dropped, as the PR says.
  • If a table uses rowspan/colspan, its rows come out shorter than the header and Parse returns an error. The source then fails loudly instead of reading the wrong column. The current page uses neither.

Verified

  • go build ./...: ok
  • go vet ./...: ok
  • gofmt -l .: printed nothing
  • go test -race ./...: every package ok, including plugins/notices and plugins/notices/retire
  • python3 scripts/leak_guard.py --all: "no banned terms (8 hashed terms checked)"

Non-blocking (optional)

  • Quote in testdata/README.md: "<td>June 30, 2026</td>" appears in several rows, so it doesn't point to one place. A full row, like the veo-3.0-generate-001 row you quote on the provider page, would show more clearly what the plugin reads.
  • rowspan/colspan: htmlPipes ignores both. That's fine today. A one-line comment saying they aren't supported, and that Parse would then fail loudly, would help the next person who reads an HTML page with this reader.
  • PR body: "Related PRs: Combined: plugins#…" is still a placeholder. Fill it in or remove it.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant