[Bug]: Profile model configuration is ignored when using -p flag (timing bug)

[eloklam]

Issue: Profile model configuration is ignored when using -p flag

Bug Description

When running Hermes with a specific profile (hermes -p <profile> chat), the model configured in that profile's config.yaml is not being used. Instead, Hermes falls back to the hardcoded default model (anthropic/claude-opus-4.6).

Environment

• Hermes version: 0.6.0 (2026.3.30)
• OS: Linux
• Python: 3.12.3

Reproduction Steps

1. Create a profile with a specific model:

hermes profile create orchestrator

2. Configure ~/.hermes/profiles/orchestrator/config.yaml:

model:
  provider: fireworks
  model: fireworks-ai/accounts/fireworks/routers/kimi-k2p5-turbo
  api_key: <your-key>

3. Verify profile shows correct model:

$ hermes profile show orchestrator
Profile: orchestrator
Model:   fireworks-ai/accounts/fireworks/routers/kimi-k2p5-turbo ✅

4. Run with profile:

$ hermes -p orchestrator chat
⚕ claude-opus-4.6 · 0:05          ❌ Wrong model!

Expected Behavior

When running hermes -p orchestrator chat, the CLI should use the model from ~/.hermes/profiles/orchestrator/config.yaml.

Actual Behavior

The CLI uses the hardcoded default model claude-opus-4.6, ignoring the profile config.

Root Cause Analysis

The issue is a timing problem in the config loading sequence:

1. In cli.py line 465:

   CLI_CONFIG = load_cli_config()

   This is evaluated at module import time.

2. In hermes_cli/main.py:

   _apply_profile_override()  # Sets HERMES_HOME based on -p flag
   # ...later...
   from cli import HermesCLI  # cli.py is imported AFTER profile override

   BUT load_cli_config() at line 135 uses:

   user_config_path = _hermes_home / "config.yaml"

   Where _hermes_home is evaluated at module load time (line 76):

   _hermes_home = get_hermes_home()  # Cached before profile override

3. Result: CLI_CONFIG loads from the default ~/.hermes/config.yaml, not the profile's config.

Affected Code Locations

cli.py line 135:

user_config_path = _hermes_home / 'config.yaml'  # Uses cached value

cli.py line 76:

_hermes_home = get_hermes_home()  # Evaluated before profile override

cli.py line 1097-1099 (HermesCLI.__init__):

_model_config = CLI_CONFIG.get("model", {})
_config_model = (_model_config.get("default") or _model_config.get("model") or "")
self.model = model or _config_model or _DEFAULT_CONFIG_MODEL  # Falls back to opus

Workaround

Until this is fixed, users must:

1. Set the model in the root ~/.hermes/config.yaml instead of profile configs
2. Or add default: <model> to profile configs (in addition to model:)

Suggested Fix

Change cli.py line 135 to use get_hermes_home() directly instead of the cached _hermes_home:

-    user_config_path = _hermes_home / 'config.yaml'
+    user_config_path = get_hermes_home() / 'config.yaml'

This ensures the current HERMES_HOME env var is read each time load_cli_config() is called.

Additional Context

This bug defeats the purpose of profile-based configurations. Users expect profiles to be self-contained, but currently the root config's defaults interfere with profile configs due to the merge behavior + hardcoded fallback.

Labels

• bug
• config
• profiles
• high-priority