fix(acp): preserve explicit model providers

[khushal-jagota]

Summary

• preserve every recognized explicit provider:model choice in Hermes ACP
• keep provider auto-detection for unqualified model text and supported slash, custom, aggregator, and variant forms
• add deterministic resolver-collision coverage and a no-turn session/set_model persistence regression

Root cause

ACP parsed an explicit provider correctly, then ran bare-model detection when the parsed provider matched the current provider. A catalog collision could therefore replace the provider that the client selected.

This builds on #59092 and preserves @BROCCOLO1D's original commit authorship.

Impact

ACP clients can select a provider-qualified model without Hermes silently routing the model through another provider. Unqualified selections retain automatic provider detection.

Validation

• scripts/run_tests.sh tests/acp_adapter/ tests/hermes_cli/test_model_validation.py tests/hermes_cli/test_models.py — 84 passed
• .venv/bin/ruff check . — passed
• git diff --check — passed

[khushal-jagota]

Ready for maintainer review. This salvages #59092 with BROCCOLO1D's original commit and adds the real resolver-collision plus no-turn session/set_model regressions.

The fork workflows await maintainer approval:

• CI: https://github.com/NousResearch/hermes-agent/actions/runs/31568823161
• Docker: https://github.com/NousResearch/hermes-agent/actions/runs/31568822829

Local verification passed 84 relevant tests, repository-wide Ruff checks, and git diff --check.

[alt-glitch]

This was generated by AI during triage.

Duplicate of #59092: both PRs prevent ACP session/set_model from re-detecting a recognized explicit provider:model selection and replacing it after parsing. #59092 is the earlier open canonical PR.

[Enough1122]

AI code review — automated review for reference, author can ignore or act on any point.

PR: fix(acp): preserve explicit model providers

1. Slash vs colon asymmetry (acp_adapter/server.py, _resolve_model_selection): parse_model_input_details treats any recognized provider:model as explicit regardless of the current provider, but the new slash branch only treats provider/model as explicit when the provider equals the current one. google:model while on anthropic switches provider; google/model while on anthropic falls through to detect_provider_for_model. Same user intent, different outcome. Either apply the colon semantics to the slash form too, or document why the slash form is current-provider-only.

2. Slash support lands only in ACP, not the CLI (hermes_cli/models.py, parse_model_input): parse_model_input_details still only parses colons, so /model anthropic/model in the CLI won't get the slash handling the ACP resolver now has — two entry points, two syntaxes. If slash syntax is intended to be general, normalize it inside parse_model_input_details (a shared helper) so both callers get it and the duplicated slash logic in the ACP resolver disappears.

3. is_known_provider_name depends on _KNOWN_PROVIDER_NAMES staying lowercase (models.py): the set is built from _PROVIDER_LABELS.keys() + _PROVIDER_ALIASES.keys() + {"openrouter", "custom"}; the new helper lowercases the input before membership, so any future label/alias key added with uppercase silently breaks lookups. A frozenset(name.lower() for ...) at construction would make the invariant structural. (The old inline check had the same property, so this is hardening, not a regression.)

4. Good coverage: the collision fixture plus parametrized cases and the no-turn persistence test are exactly the regressions this needed. The openrouter_collision monkeypatch of a module-level function is clean.