Summary

Fixes #18961 — --oneshot CLI mode now respects the configured fallback_providers chain from config.yaml, bringing it to parity with gateway mode.

Problem

When the primary provider returns HTTP 429 (rate-limit) in --oneshot mode, the agent hard-fails instead of falling back to the next provider in the fallback_providers chain. Gateway mode already handles this correctly by passing fallback_model=self._fallback_model to the AIAgent() constructor.

Root Cause

hermes_cli/oneshot.py loads cfg = load_config() but never extracts fallback_providers / fallback_model from it, and the AIAgent(...) call at line 287 is missing the fallback_model= parameter entirely.

Fix

Two-line change:

1. Extract fallback config from the already-loaded cfg: fallback_model = cfg.get("fallback_providers") or cfg.get("fallback_model") or None
2. Pass it to AIAgent(...): fallback_model=fallback_model

AIAgent.__init__ already normalizes both the list (fallback_providers) and legacy dict (fallback_model) formats into a fallback chain, so this works with either config style.

Testing

• Manual: configure a fallback_providers chain in config.yaml, then run hermes --oneshot while the primary provider is rate-limited → previously hard-failed, now falls back.
• AIAgent.__init__ signature already accepts fallback_model parameter (line 954 in run_agent.py), so no downstream changes needed.

Files Changed

• hermes_cli/oneshot.py: Added fallback config extraction + parameter passthrough (5 lines)