Summary

_load_config() in tools/delegate_tool.py reads cli.CLI_CONFIG before falling back to hermes_cli.config.load_config(). Because CLI_CONFIG is populated exactly once at cli.py module import and never refreshed, long-running gateway processes (Discord / Telegram / Slack) can't see edits made to ~/.hermes/config.yaml's delegation.model / delegation.provider after startup. The result is a silent fall-back: users edit the config, restart nothing, and subagent dispatches silently inherit the parent session's model instead of the configured override.

This flips the fallback order so load_config() (which reads the file on every call) is tried first, and CLI_CONFIG is only the backup.

Observed Symptom (production Discord gateway, 2026-04-25)

• Gateway running as anthropic/claude-opus-4.7
• ~/.hermes/config.yaml: delegation.model: anthropic/claude-sonnet-4.6, delegation.provider: openrouter
• delegate_task dispatches returned envelopes with "model": "anthropic/claude-opus-4.7" despite the pin
• SQLite sessions table confirmed: all child sessions recorded the parent's model, not the configured one

Verified root cause:

1. _load_config() returned CLI_CONFIG["delegation"], a dict frozen at April 22 import (when delegation.model / delegation.provider were still empty strings).
2. _resolve_delegation_credentials() read configured_model = None, configured_provider = None.
3. Hit the "No provider override — child inherits everything from parent" branch (line ~2307).
4. creds["model"] = None → effective_model = model or parent_agent.model (line ~321 in _build_child_agent) → child ran Opus.

Cost Impact

In the incident that triggered this investigation, ~$700 of unintended Opus burn on Sonnet-suitable mechanical coding work (4 PRs: CSV streaming importer, data migration command, Filament Infolist grouping refactor, doc updates) before the routing miss was diagnosed. A user who's pinned a cheaper model is probably doing so because the task doesn't need Opus; they should not have to kill their gateway just to pick up a config edit.

Fix

hermes_cli.config.load_config() reads ~/.hermes/config.yaml on every call. Try that first. Only if the disk read fails / returns empty do we fall back to the frozen CLI_CONFIG.

Cost of the extra read: a ~10 KB YAML file parsed once per delegate_task call (not per API hit, not per token). ~5 ms on a modern machine, on the delegation cold path. Rounding error relative to any LLM API call.

Diff Semantics

CLI users: No change. load_config() reads the same file as load_cli_config, so whatever CLI_CONFIG had, load_config() returns an equivalent delegation block.

Gateway users: Config edits take effect on the next delegate_task call instead of requiring a gateway restart.

Testing

Reproduced in isolation before the fix:

# With stale CLI_CONFIG simulating a long-running gateway:
creds = _resolve_delegation_credentials(_load_config(), parent)
# creds["model"] == None   <- bug

After the fix:

# Same setup, but _load_config now reads the live file:
creds = _resolve_delegation_credentials(_load_config(), parent)
# creds["model"] == "anthropic/claude-sonnet-4.6"  <- fixed

Full delegate_task flow exercised in /tmp/trace_delegation_full.py (not included in commit) confirms the child AIAgent's self.model stays sonnet-4.6 through construction and past the first API call.

Docstring update

Added a section explaining the fallback order and the historical failure mode so future readers see why disk-first matters.