fix(vertex): prepend mandatory google/ publisher prefix to bare model names

[tangtaizong666]

What does this PR do?

Fixes the native Vertex AI provider (--provider vertex) failing every request with HTTP 400 INVALID_ARGUMENT ("Malformed publisher model") when the user supplies a standard bare model name such as gemini-3.1-flash-lite.

Vertex AI's OpenAI-compatible endpoint requires the model field in the mandatory <publisher>/<model> form (e.g. google/gemini-3.1-flash-lite). normalize_model_for_provider() in hermes_cli/model_normalize.py had no vertex branch, so bare names fell straight through the "pass through as-is" fallback and reached the API unprefixed. (The original Vertex feature PR #16010 — which contained normalization — was closed unmerged; the provider that landed as a plugin never picked up the prefix injection.)

The new vertex branch applies these rules:

• Bare name (gemini-3.1-flash-lite) → prepend the google/ publisher prefix.
• Explicit publisher prefix (google/…, anthropic/…, meta/…, mistralai/…) → pass through unchanged. This deliberately deviates from the snippet suggested in the issue, which rewrote any vendor/model to google/model — Vertex hosts non-Google publishers, so an explicit publisher is authoritative and must survive.
• Matching provider prefix (vertex/gemini-3.5-flash, google-vertex/… — the aggregator-style form users copy into config.yaml) → repaired to the google/ publisher form, consistent with how other direct providers repair provider/ prefixes.

Dots in model IDs are preserved (Vertex model IDs use them, e.g. gemini-2.5-flash), and provider aliases (google-vertex, vertex-ai, gcp-vertex, vertexai) resolve to the same behavior. google/ is never mistaken for a provider prefix because the google alias canonicalizes to the gemini provider, not vertex.

Relationship to #56805 and #56806

Both opened while this change was going through its test run — flagging the overlap up front and deferring to maintainers on which to take. The behavioral differences are material, not cosmetic:

Input (provider vertex)	#56805	#56806	this PR
gemini-3.1-flash-lite (the issue)	google/gemini-3.1-flash-lite ✓	google/gemini-3.1-flash-lite ✓	google/gemini-3.1-flash-lite ✓
anthropic/claude-sonnet-4.6	google/anthropic/claude-sonnet-4.6 — double prefix, malformed	unchanged ✓	unchanged ✓
meta/llama-3.3-70b-instruct-maas (Vertex MaaS)	google/meta/… — malformed	unchanged ✓	unchanged ✓
vertex/gemini-3.5-flash (copied into config)	google/gemini-3.5-flash ✓	unchanged — vertex isn't a Vertex publisher, still HTTP 400	google/gemini-3.5-flash ✓
bare name w/o detectable vendor	google/<name>	unchanged — bare names are always rejected by the endpoint	google/<name>

(On the last row: the endpoint's <publisher>/<model> form is mandatory, so passing a bare name through can never succeed; defaulting to google/ matches the provider's documented scope, "Gemini models via Google Cloud".)

Related Issue

Fixes #56778

Type of Change

• 🐛 Bug fix (non-breaking change that fixes an issue)
• ✨ New feature (non-breaking change that adds functionality)
• 🔒 Security fix
• 📝 Documentation update
• ✅ Tests (adding or improving test coverage)
• ♻️ Refactor (no behavior change)
• 🎯 New skill (bundled or hub)

Changes Made

• hermes_cli/model_normalize.py — new vertex branch in normalize_model_for_provider() (strip a matching vertex/ provider prefix, keep any remaining explicit publisher/, otherwise prepend google/); module docstring and function doctest examples updated.
• tests/hermes_cli/test_model_normalize.py — new TestIssue56778VertexPublisherPrefix regression class (12 cases: bare-name prefixing, google/ pass-through, non-Google publisher preservation, vertex/-prefix repair, alias handling, dot preservation, empty input).
• scripts/release.py — added my commit email to AUTHOR_MAP, as required by the contributor-attribution CI check for first-time commit emails.

How to Test

1. Automated: pytest tests/hermes_cli/test_model_normalize.py -q — the 12 new TestIssue56778VertexPublisherPrefix cases fail on main (bare names pass through unprefixed) and pass with this change; the pre-existing 86 cases stay green. scripts/run_tests.sh tests/hermes_cli/test_model_normalize.py tests/hermes_cli/test_vertex_provider.py tests/agent/test_vertex_adapter.py (CI-equivalent per-file isolation) — 125/125 pass. ruff check . clean; ty reports no new diagnostics.
2. Repro from the issue: hermes chat -q "echo" --provider vertex --model gemini-3.1-flash-lite (with GCP ADC configured) — previously died on the first turn with HTTP 400 "Malformed publisher model ('model': 'gemini-3.1-flash-lite')"; the request now goes out as google/gemini-3.1-flash-lite, the form the issue reporter verified works.
3. Workaround unchanged: --model google/gemini-3.1-flash-lite still passes through untouched.

Checklist

Code

• I've read the Contributing Guide
• My commit messages follow Conventional Commits (fix(scope):, feat(scope):, etc.)
• I searched for existing PRs to make sure this isn't a duplicate
• My PR contains only changes related to this fix/feature (no unrelated commits)
• I've run pytest tests/ -q and all tests pass
• I've added tests for my changes (required for bug fixes, strongly encouraged for features)
• I've tested on my platform: Ubuntu (WSL2, kernel 6.6.87.2-microsoft-standard)

Documentation & Housekeeping

• I've updated relevant documentation (README, docs/, docstrings) — or N/A
• I've updated cli-config.yaml.example if I added/changed config keys — or N/A (no config keys changed)
• I've updated CONTRIBUTING.md or AGENTS.md if I changed architecture or workflows — or N/A (no architecture change)
• I've considered cross-platform impact (Windows, macOS) per the compatibility guide — or N/A (pure string normalization, no I/O)
• I've updated tool descriptions/schemas if I changed tool behavior — or N/A (no tool behavior change)

Screenshots / Logs

Normalization through the real path (normalize_provider → normalize_model_for_provider), before vs. after:

# main (bare name reaches the API unprefixed → HTTP 400)
'gemini-3.1-flash-lite'        + provider 'vertex' -> 'gemini-3.1-flash-lite'

# this PR
'gemini-3.1-flash-lite'        + provider 'vertex' -> 'google/gemini-3.1-flash-lite'
'gemini-3.5-flash'             + provider 'vertex' -> 'google/gemini-3.5-flash'
'google/gemini-3.1-flash-lite' + provider 'vertex' -> 'google/gemini-3.1-flash-lite'
'anthropic/claude-sonnet-4.6'  + provider 'vertex' -> 'anthropic/claude-sonnet-4.6'
'vertex/gemini-3.5-flash'      + provider 'vertex' -> 'google/gemini-3.5-flash'

[alt-glitch]

This was generated by AI during triage.

Related Vertex normalization cluster: the former duplicate anchor #56806 is now closed unmerged, while current main still lacks a vertex normalization branch. This PR remains a valid focused repair for #56778: it prefixes bare Gemini/Gemma names, repairs matching Vertex aliases, and preserves explicit non-Google publisher prefixes. It is related to the broader open #56820 (which also covers native-route normalization), not a duplicate; maintainer review marked this PR keep_open with high salvageability. Related: #56778, #56805 (closed), #56806 (closed), #56820.

[tangtaizong666]

Acknowledged — overlap with #56806 is disclosed in the PR body (opened while my change was mid-test-run), including a behavior comparison table. The two deltas vs #56806: (1) bare names without a detectable Google vendor still get the mandatory google/ prefix there is no bare form the endpoint accepts, so passing them through can only reproduce the 400; (2) a matching vertex/-alias provider prefix copied into config is repaired to the google/ publisher form instead of reaching the API as a nonexistent publisher. Happy either way: take this PR, or fold those two cases + the non-Google-publisher pass-through tests into #56806 — deferring to maintainers.

[teknium1]

Thanks for the focused Vertex fix. Current main still lacks a vertex branch in normalize_model_for_provider() and falls through unchanged at hermes_cli/model_normalize.py:466-467; agent/agent_init.py:498-505 applies that normalizer to Vertex models before requests. The proposed branch directly corrects that path and is consistent with the documented Vertex requirement for a google/ model prefix at website/docs/guides/google-vertex.md:87-96.

The preservation of explicit publisher-qualified IDs and repair of aliases rely on the existing _strip_matching_provider_prefix() behavior (hermes_cli/model_normalize.py:226-244) and provider aliases (hermes_cli/models.py:1231-1237). The added tests cover bare names, explicit prefixes, aliases, and the empty-input guard. GitHub reports the PR cleanly mergeable against current main.

This is an automated hermes-sweeper review.