Config-driven override for the Gemma picker-visibility filter

[rrschott-ai]

Summary

• _GOOGLE_HIDDEN_MODELS (agent/models_dev.py) hides several Gemma IDs from the interactive model/fallback picker by default — their low TPM quotas trip under agent-style traffic for most users, which is a reasonable default.
• The only way to see a hidden model in the picker was editing the frozenset directly in source. A hermes update resets that file to upstream, silently wiping the edit — a user who deliberately wants one of these models has to reapply a source patch after every release, with no warning it happened.
• Adds models.show_hidden_google_models to config.yaml (default [], no behavior change unless set): a list of model IDs to force-show despite the built-in filter. Config lives outside the package, so it survives updates.
• Raw config entries (fallback_providers, a manually-typed model id) already bypassed this filter entirely and worked fine — this only changes what the picker UI shows, matching the intent of the original filter (a default for typical users, not a hard block).

Test plan

• Both touched files parse clean (ast.parse)
• Verified directly: with no override, a hidden Gemma id stays hidden; with models.show_hidden_google_models containing it, it's shown; an unrelated hidden id (not in the override list) stays hidden; a non-Google provider is unaffected; a config-load failure fails safe (stays hidden, no exception)
• Confirmed end-to-end against a real local config.yaml with the override set — picker filter correctly un-hides only the configured ids
• Reviewer: confirm models.show_hidden_google_models is the right home/name for this vs. e.g. folding it into the existing providers config dict

[teknium1]

Thanks for making the existing picker filter configurable; the underlying limitation is present on current main (agent/models_dev.py:567-597).

Problems

• The new override un-hides every _GOOGLE_HIDDEN_MODELS member. That set also contains stale/retired Gemini slugs that are explicitly documented as 404ing at agent/models_dev.py:557-563, not only low-TPM Gemma IDs. Preserve the stale-model filter while allowing the intended Gemma subset.
• Commit d5b3da2 also adds agent.skill_auto_patch, which is outside the PR summary and has no implementation consumer on current main (rg -n 'skill_auto_patch' . returned no matches). Please remove this unrelated config addition from the change.
• No tests are included. Extend tests/hermes_cli/test_gemini_provider.py:309-356 to cover the configured Gemma override on both catalog paths and verify stale Gemini IDs remain hidden.

Suggested changes

• Separate the low-TPM Gemma IDs from stale Google IDs, and consult the opt-in list only for the former.
• Keep the change scoped to the picker override and add targeted regression coverage.

Automated hermes-sweeper review.

[teknium1]

_GOOGLE_HIDDEN_MODELS also contains stale/retired Gemini IDs that main documents as 404ing (agent/models_dev.py:557-563). This permits users to re-add known-invalid models; restrict this override to the low-TPM Gemma subset and keep stale IDs unconditionally hidden.

[rrschott-ai]

Addressed all three points, pushed as a new commit (b9ef402):

1. Scoped the override correctly. Split _GOOGLE_HIDDEN_MODELS into _GOOGLE_LOW_TPM_MODELS (what the config override is meant for) and _GOOGLE_STALE_MODELS (retired slugs that 404 — a fact, not a posture). _should_hide_from_provider_catalog now only consults show_hidden_google_models for the low-TPM set; a stale slug in that config list is silently ignored rather than un-hiding a model that doesn't work. Verified directly: a stale id stays hidden even when explicitly listed in the override, while a low-TPM Gemma id in the same override list is correctly un-hidden.

2. Removed the stray agent.skill_auto_patch hunk. You're right that it's unrelated — it leaked in from a diff-extraction mistake while splitting several concurrent local changes into separate PRs (that config key belongs to Add agent.skill_auto_patch config gate for skill_manage patches #58353, not this one). Confirmed zero occurrences of skill_auto_patch in the current diff.

3. Added test coverage to tests/hermes_cli/test_gemini_provider.py (TestGeminiModelsDev): the override un-hiding a configured Gemma id while other Gemma ids stay hidden, a stale slug in the override list staying hidden regardless, and default (no config) behavior unchanged. All pass, along with the pre-existing suite in that file (4 unrelated pre-existing failures in TestGeminiAgentInit are a missing concurrent_log_handler dependency in my local environment, not touched by this change).

Thanks for catching the cross-contamination — verified with git diff that this PR is now scoped to just the picker-override change.