Summary

The hardcoded _PROVIDER_MODELS["xai"] list in hermes_cli/models.py is stale, and the pattern of hardcoded per-provider model lists is structurally fragile across the file.

Concretely:

# hermes_cli/models.py (current main: 4a9ac5c3)
_PROVIDER_MODELS: dict[str, list[str]] = {
    ...
    "xai": [
        "grok-4.20-reasoning",        # ← xAI no longer accepts this name
        "grok-4-1-fast-reasoning",    # ← never existed in xAI's catalog
    ],
    ...
}

xAI's actual current model IDs (per models.dev cache, fetched today) use date-suffixed names like grok-4.20-0309-reasoning, grok-4.20-0309-non-reasoning, grok-4.20-multi-agent-0309. None of those match the hardcoded list.

Reproduction

1. Configure xAI provider via env: XAI_API_KEY=…
2. Run hermes and select xAI from /model picker
3. Picker shows only the 2 stale entries; selecting either yields:

HTTP 400: Unknown Model, please check the model code.

Same shape as the open Z.AI bug #7922.

Root cause

_PROVIDER_MODELS hardcodes per-provider lists that don't track upstream catalogs. Of the ~15 providers in this dict, only openai-codex is currently sourced dynamically (_codex_curated_models(), added in PR #7844 to fix the analogous bug #6595). Every other provider — xai, nous, openai, copilot, gemini, zai, nvidia, kimi-coding, stepfun, moonshot, minimax, minimax-cn, etc. — keeps drifting silently.

The drift surfaces in two places:

1. The /model picker (this issue, also Bug: Duplicate provider entries in /model picker #14057, Missing GPT-5.5 model support across provider catalog, Codex image-gen default, tests, and docs #16161)
2. detect_provider_for_model() validation, which uses _PROVIDER_MODELS membership and emits "Could not reach … to validate" warnings for any model not in the curated list (same false-warning bug as [Bug] Telegram /model picker missing gpt-5.4 for openai-codex provider (stale _PROVIDER_MODELS hardcoded list) #6595).

Suggested fix

Short-term: apply the _codex_curated_models() pattern to xai (and the other most-active providers). For xai specifically, derive from the existing models_dev_cache.json that hermes already maintains:

```python
def _xai_curated_models() -> list[str]:
import json, os
cache = os.path.expanduser("~/.hermes/models_dev_cache.json")
try:
if os.path.isfile(cache):
with open(cache) as fh:
data = json.load(fh)
ids = list(data.get("xai", {}).get("models", {}).keys())
if ids:
return sorted(ids)
except Exception:
pass
return [...static fallback...]

_PROVIDER_MODELS["xai"] = _xai_curated_models()
```

Longer-term: the pattern of per-provider helpers will keep growing. Worth considering a single _provider_models(provider: str) resolver that:

• Reads models_dev_cache.json for any provider available there (covers most),
• Falls back to provider-specific helpers (e.g. _codex_curated_models) where the cache isn't authoritative,
• Falls back to a small per-provider static list only as a last resort.

Related: the model-allowlist feature request #16608 would let users opt into a known-good subset rather than relying on the curation embedded in source.

Workaround

Patch locally by replacing the xai literal with a _xai_curated_models() helper modeled on _codex_curated_models(). We're maintaining a small Ansible playbook that re-applies this on each managed-host install; happy to share if helpful.

Environment

• Hermes Agent main @ 4a9ac5c3 (2026-04-27)
• Python 3.12, Ubuntu 24.04 LTS
• Provider: xAI direct (XAI_API_KEY, https://api.x.ai/v1)

References

• [Bug] Telegram /model picker missing gpt-5.4 for openai-codex provider (stale _PROVIDER_MODELS hardcoded list) #6595 (closed → PR fix: add gpt-5.3-codex-spark to openai-codex curated model list #7844, the openai-codex precedent)
• Gateway /model command sends provider-prefixed slug as raw model name for non-curated models (HTTP 400 Unknown Model) #7922 (open, same shape for Z.AI: stored slug → HTTP 400 Unknown Model)
• Bug: Duplicate provider entries in /model picker #14057 (open, duplicate provider entries in picker)
• Missing GPT-5.5 model support across provider catalog, Codex image-gen default, tests, and docs #16161 (open, missing GPT-5.5 across catalog)
• [Feature]: Model allowlist — let users control which providers appear in /model list #16608 (open, model allowlist feature request)