fix(models): validate OpenRouter presets structurally on custom endpoints (#97907)

[Tranquil-Flow]

What

OpenRouter dashboard presets (@preset/<slug>) are account-scoped server-side objects — they never appear in a /models listing. On OpenRouter-compatible custom providers, validate_requested_model()'s custom-endpoint branch probed the catalog, always missed, and soft-accepted with a false "not found in this custom endpoint's model listing" warning embedding the raw catalog URL. Telegram's link preview fetched that URL and rendered the large JSON response as a models.json document on every preset switch.

Note the premise: validate_requested_model() reclassifies openrouter + non-openrouter.ai base_url to custom (models.py:6508-6509), so the custom branch is exactly the path this issue reports.

Fix

Handle preset references structurally in the custom branch instead of catalog-probing them:

• Bare @preset/<slug> — accepted without probing (OpenRouter validates the slug at request time), with a URL-free message. Nothing for a link preview to fetch.
• <model>@preset/<slug> — the BASE model is validated against the live listing; the preset suffix is preserved verbatim on the persisted/corrected id (same rule as the existing :nitro routing-variant handling, which the fuzzy corrector would otherwise strip).
• Malformed slug (empty / invalid charset) — rejected with the slug contract ([A-Za-z0-9._~-]+), not a misleading catalog warning.
• Multiple @preset/ markers — not a defined form; keeps existing behavior.
• Non-preset identifiers keep the existing custom-branch behavior untouched.

No gateway/Telegram adapter change: the document trigger is removed at the source (no URL in the message), and normal link previews remain enabled per the issue's Expected section.

Verification

• tests/hermes_cli/test_openrouter_preset_custom_validation.py — 13 new tests covering all four layers (bare preset, combined form, malformed slug, non-preset unchanged).
• Fail-without-fix proven: on clean upstream/main the suite runs 8 failed / 5 passed (the 5 passing pin unchanged non-preset behavior); with the fix 13/13 pass. Re-verified independently by the reviewer on upstream 0d02f0d1dc.
• Nearby suites (test_model_validation.py, test_model_switch_custom_providers.py, test_custom_provider_model_switch.py, test_user_providers_model_switch.py, test_openai_codex_model_validation_fallback.py): 140 passed, 2 failed — both failures reproduce identically on clean upstream/main (pre-existing, unrelated).
• Branch is exactly one commit ahead of upstream/main (git rev-list --left-right --count → 0 1).

Competitor analysis

Open PR #89129 "fix(models): support OpenRouter preset references" handles @preset/ only in the normalized == "openrouter" branch — correct for native OpenRouter, but the custom-provider scenario of this issue never satisfies that gate (it is reclassified to custom upstream of the branch). Scopes are complementary with no gated-code-path overlap; the slug-charset rule mirrors #89129's for cross-branch consistency.

Fixes #97907.

Auto-published by Moonsong via Path B automated pipeline.

[Enough1122]

AI code review — automated review for reference; please use your judgment.

PR #97951 — fix(models): validate OpenRouter presets structurally on custom endpoints

Summary: OpenRouter @preset/<slug> identifiers are server-side objects never present in /models. Adds _split_openrouter_preset and _PRESET_SLUG_RE (letters/digits/._~-) and structural handling in validate_requested_model for custom providers: slug charset is validated, bare @preset/... is accepted without a probe (no URL in message to trigger Telegram link previews), combined <model>@preset/... validates the base against the listing and preserves suffix through auto-correction (mirrors :nitro handling). Multiple markers (a@preset/x@preset/y) fall back to ordinary catalog validation.

Non-blocking observations:

• _split_openrouter_preset (hermes_cli/models.py:35) counts exactly one @preset/ marker; empty slug @preset/ still returns a suffix and is then rejected by the charset check — correct per contract.
• Suggestion generation now uses _lookup (base) not original id, so close-match hints are for the base — appropriate.
• Leading/trailing whitespace in requested is not stripped before marker count; upstream validate_requested_model trims elsewhere, but if a raw " @preset/foo" slipped through it would probe incorrectly. Low risk.

Verdict: Well-tested fix for #97907 (four layers: bare, combined, malformed, non-preset unchanged). No blocking issues.