fix(cli): auto-select Codex model when provider resolves to openai-codex

[teyrebaz33]

Problem

Fixes #651

When a user sets up Hermes with OpenAI Codex, any non-Codex model in LLM_MODEL, OPENAI_MODEL, or the config default (e.g. anthropic/claude-opus-4.6) causes an immediate 400 BadRequestError on every request:

Error code: 400 - {'detail': "The 'anthropic/claude-opus-4.6' model is not supported when using Codex with a ChatGPT account."}

Root Cause

HermesCLI.__init__ sets self.model from LLM_MODEL / OPENAI_MODEL env-vars or the config default before the provider is resolved. When _ensure_runtime_credentials() later resolves the provider to openai-codex, the incompatible model was never corrected.

PR #661 addressed only the config-default path and used a hardcoded fallback model (gpt-5.3-codex). This PR fixes the root cause for all model sources including env-vars, which is the exact scenario reported in #651.

Fix

In _ensure_runtime_credentials(), after the provider resolves to openai-codex, check whether the active model is Codex-compatible. If not, replace it with the first model returned by get_codex_model_ids() and warn the user if they had explicitly set the model.

Changes

• cli.py: Add _model_is_default flag to track explicit vs default model selection
• cli.py: In _ensure_runtime_credentials(), normalise incompatible models when provider is openai-codex — covers CLI arg, env-var, and config default paths
• cli.py: Trigger agent rebuild when effective model changes
• tests/test_cli_provider_resolution.py: Add 3 tests covering config-default path, env-var path (exact [Bug]: hi borked it? #651 scenario), and explicit Codex model preservation

Testing

pytest tests/test_cli_provider_resolution.py tests/test_codex_models.py -q
9 passed in 2.92s

Full suite: 2063 passed, 0 failed.

[teknium1]

Thanks for the thorough fix @teyrebaz33! The _model_is_default flag and the comprehensive test coverage (4 scenarios) were both great ideas that we incorporated into the final implementation.

Unfortunately this PR couldn't be merged directly because the branch is 12 commits behind main and the diff includes unrelated regressions — it removes the /title command (session naming from PR #720), the /provider command, reverts image preprocessing back to raw base64 (undoing commit cf810c2), and removes session title display on resume.

We implemented the fix in commit 95b1130, combining the best of both this PR and #661:

• Clean _normalize_model_for_provider() method (from fix(cli): fall back to a Codex model for Codex runtime #661)
• _model_is_default flag for smarter warnings (from this PR)
• 4 test scenarios covering all paths

Credit noted in the commit message. Closing this PR — thanks for the contribution!