fix: skip first-segment slash strip for URI-scheme model IDs (#3429)

[b3nw]

fix: getModelLabel breaks URI-scheme model IDs (#3429)

Closes #3429

Thinking Path

• Hermes WebUI supports custom providers that expose models via URI-shaped identifiers (e.g. Yandex GPT uses gpt://${FOLDER_ID}/model/version)
• PR fix: model picker snaps to wrong model with multi-slash IDs (#3360) #3366 (shipped in v0.51.210) fixed multi-slash model ID collisions by replacing split('/').pop() with first-segment-only stripping in getModelLabel, _normalizeConfiguredModelKey, and _norm_model_id
• That fix assumes the first slash-segment is always a provider prefix — but URI-scheme IDs like gpt://folder/model/latest have :// in the scheme, and the first / after gpt: is part of the URI authority, not a provider delimiter
• The result is that getModelLabel("gpt://b1g12345/deepseek-v4-flash/latest") strips gpt: and returns //b1g12345/deepseek-v4-flash/latest — a broken label shown in the composer chip
• This PR guards all four slash-stripping locations with a URI-scheme detection check (:// presence), preserving both the Bug: Model picker snaps to wrong model when multiple multi-slash IDs share the same base name #3360 multi-slash fix and correct handling of URI-shaped model IDs
• The benefit is that users with Yandex GPT, custom HTTP endpoint proxies, or any URI-scheme-based model IDs see correct labels again

What Changed

Fix 1 — getModelLabel (static/ui.js)

Added a regex check (/^[a-z][a-z0-9+.-]*:\/\//i) before the first-segment slash strip. When the model ID has a URI scheme, the entire ID is passed through without stripping.

Fix 2 — _normalizeConfiguredModelKey (static/ui.js)

Wrapped the existing colon-prefix strip and slash-strip logic in a URI-scheme guard. URI-scheme IDs normalize with their full structure preserved (hyphens still replaced with dots for consistency).

Fix 3 — _norm_model_id (api/config.py)

Added a "://" not in s check before the first-segment slash strip. Simple, dependency-free — no regex import needed in the nested function scope.

Fix 4 — _get_label_for_model (api/config.py)

Added a _has_scheme lambda ("://" in s) that guards both the _norm helper used for group label lookup and the bare fallback used for title-casing.

Files changed

File	Lines	Description
static/ui.js	+40/−16	URI-scheme guards in getModelLabel and _normalizeConfiguredModelKey
api/config.py	+29/−11	URI-scheme guards in _norm_model_id and _get_label_for_model
tests/test_issue3429_uri_scheme_model_ids.py	+258 (new)	7 regression tests across 4 test classes

Why It Matters

Any user running Yandex GPT (gpt:// scheme), a custom HTTP-endpoint proxy (https:// scheme), or any provider that uses URI-formatted model identifiers will see broken labels in the composer chip and potentially mismatched normalization in badge assignment and configured-entry dedup. The regression was introduced in v0.51.210 and affects all users with URI-scheme model IDs.

Verification

Automated tests

pytest tests/test_issue3429_uri_scheme_model_ids.py \
       tests/test_issue3360_multi_slash_model_collision.py \
       tests/test_norm_model_id_trailing_empty_guard.py \
       tests/test_issue1188_fuzzy_match.py \
       tests/test_issue1228_model_picker_duplicate_ids.py \
       -v
# 49 passed

New tests (test_issue3429):

• TestGetModelLabelUriScheme — 3 tests: Yandex gpt:// preserved, multiple URI schemes preserved, non-URI slash IDs still stripped
• TestNormalizeConfiguredModelKeyUriScheme — 2 tests: URI scheme preserved in normalization, regular slash IDs still normalize
• TestBackendFrontendUriSchemeParity — 1 test: Python _norm_model_id produces identical output to JS _normalizeConfiguredModelKey for URI and non-URI inputs
• TestBackendGetLabelUriScheme — 1 test: Python _get_label_for_model does not strip URI scheme

Existing test suites — all pass unmodified:

• test_issue3360_multi_slash_model_collision (9 tests) — multi-slash fix preserved
• test_norm_model_id_trailing_empty_guard (5 tests) — trailing-empty guard preserved
• test_issue1188_fuzzy_match (11 tests) — fuzzy match unaffected
• test_issue1228_model_picker_duplicate_ids (17 tests) — provider-aware preference unaffected

Risks / Follow-ups

• Low risk: The URI detection uses simple :// substring check (backend) and regex (frontend). This correctly distinguishes gpt://folder/model from openai/gpt-5.5 — no known false positives since normal provider/model IDs never contain ://.
• Edge case: A hypothetical model ID like custom:// (URI scheme with no path) would skip stripping and normalize to itself — strictly better than the broken empty-string behavior.
• Not addressed: The _findModelInDropdown function doesn't need a URI guard because it operates on option values via exact match or normalized comparison — the normalization is already guarded.

AI Usage Disclosure

• Provider: Anthropic
• Model: Claude Opus 4.6
• Tool: Cursor IDE (agentic pair-programming)
• Usage: Root cause analysis of the Regression (#3366): getModelLabel breaks gpt:// model IDs (shows /${YANDEX_FOLDER_ID}/...) #3429 regression, implementation of URI-scheme guards across all four affected functions, test authoring, and verification were done collaboratively. All changes reviewed and tested by a human before commit.

[nesquena-hermes]

Thanks @b3nw — shipped (the backend half) in v0.51.219 (Release GM).

You spotted that the #3429 URI-scheme bug also lived in the model-IDENTITY paths, not just the visible chip. The front-end label (getModelLabel) was already fixed in v0.51.218 (#3429), so I took your complementary backend + normalization halves: api/config.py _norm_model_id / _get_label_for_model and static/ui.js _normalizeConfiguredModelKey now skip the first-segment slash strip for scheme:// ids — preventing the #3360-class identity collision/mislabel for URI-shaped model IDs in dropdown matching, dedup, and badge assignment. Your backend↔front-end parity tests are included.

The one part I didn't take was your getModelLabel() rewrite: v0.51.218's version extracts the actual model name from the URI path (e.g. gpt://${FOLDER}/deepseek-v4-flash/latest → deepseek-v4-flash) rather than returning the whole id, which is friendlier in the chip. Everything else is yours — Co-authored-by attribution preserved in the commit. Verified: full suite 7306 passed / 0 failed, Codex SAFE, Opus APPROVE.