Bug Description

hermes config set delegation.model <X> writes config.yaml on disk and reports success, but the change does not take effect for delegate_task in any already-running process (CLI/gateway/cron). The runtime keeps using the previous delegation.model until the process is restarted.

This is silent — there's no warning that a restart is required, and the next delegate_task call happily runs on the old model with no indication anything is stale.

Steps to Reproduce

1. Start a CLI/gateway session with delegation.model: google/gemini-2.5-flash in config.yaml.
2. From a separate shell (or mid-session), run hermes config set delegation.model google/gemini-3-flash-preview. Confirm config.yaml on disk now shows google/gemini-3-flash-preview.
3. From the original session, call delegate_task(...). The subagent runs on google/gemini-2.5-flash, not the new model.
4. Restart the process. Now delegate_task correctly uses google/gemini-3-flash-preview.

Minimal reproducer

This 30-line script reproduces the bug deterministically without spinning up a real subagent:

import os, sys, yaml, tempfile, pathlib

tmp = pathlib.Path(tempfile.mkdtemp(prefix="hermes-repro-"))
cfg_path = tmp / "config.yaml"
cfg_path.write_text(yaml.safe_dump({
    "model": {"default": "claude-opus-4-7", "provider": "anthropic"},
    "delegation": {"model": "google/gemini-2.5-flash", "provider": "nous"},
}))
os.environ["HERMES_HOME"] = str(tmp)
sys.path.insert(0, "/path/to/hermes-agent")

# Simulate gateway startup: CLI_CONFIG loaded once at import time
from hermes_cli.config import load_config
CLI_CONFIG_v1 = load_config()

# Simulate `hermes config set delegation.model google/gemini-3-flash-preview`
data = yaml.safe_load(cfg_path.read_text())
data["delegation"]["model"] = "google/gemini-3-flash-preview"
cfg_path.write_text(yaml.safe_dump(data))

# Inject the cached CLI_CONFIG to simulate running gateway state
import cli as cli_mod
cli_mod.CLI_CONFIG = CLI_CONFIG_v1

# Call the delegate_task code path
from tools.delegate_tool import _load_config
result = _load_config()
print(f"Returns:  {result['model']}")
print(f"On disk:  google/gemini-3-flash-preview")
assert result["model"] == "google/gemini-3-flash-preview", "stale cache wins"

Output:

Returns:  google/gemini-2.5-flash
On disk:  google/gemini-3-flash-preview
AssertionError: stale cache wins

Root cause

tools/delegate_tool.py line 2322 — _load_config() checks the in-memory cli.CLI_CONFIG first and only falls through to a fresh disk read when CLI_CONFIG is empty:

def _load_config() -> dict:
    """Load delegation config from CLI_CONFIG or persistent config."""
    try:
        from cli import CLI_CONFIG
        cfg = CLI_CONFIG.get("delegation", {})
        if cfg:
            return cfg          # <-- runtime cache wins; disk ignored
    except Exception:
        pass
    try:
        from hermes_cli.config import load_config
        full = load_config()
        return full.get("delegation", {})
    except Exception:
        return {}

CLI_CONFIG is initialized once at cli.py:600 (CLI_CONFIG = load_cli_config()) and never refreshed. hermes config set mutates the file on disk but has no IPC channel to notify the running process. Result: stale cache silently wins for the lifetime of the process.

Why this is more than cosmetic

This makes A/B testing delegation models on a live session impossible without a full gateway restart, and the failure is silent — hermes config set reports success, the file on disk is correct, but the running agent reads stale values. Users will wrongly attribute behavior changes (or non-changes) to the wrong model.

It also affects any other delegation.* knob (provider, max_iterations, max_concurrent_children, reasoning_effort, etc.) since they all flow through the same _load_config() path.

Proposed fix

Smallest change with no protocol/IPC work: have _load_config() always reload from disk (the disk read is cheap — config.yaml is small and reads dozens of times per session is fine). The CLI_CONFIG path was a micro-optimization; the cache-staleness footgun is a worse tradeoff.

Alternative: stat config.yaml's mtime in _load_config() and only reload when it has advanced since the last read. Slightly more code, fewer disk reads.

Happy to submit a PR with the simpler "always reload" version plus a regression test in tests/tools/test_delegate_tool.py.

Environment

• Hermes Agent v0.12.0 (2026.4.30)
• Reproduces on current main (HEAD as of issue filing)
• OS: Linux

Related

• delegate_task ignores subagent model config — always inherits parent model #11999 (closed) — earlier delegation config bug; different code path, but same general concern about runtime config plumbing
• [Bug]: /config terminal.timeout value does not match value in config.yaml nor TERMINAL_TIMEOUT env variable #16723 (open) — /config display reads stale source; symptom-adjacent (display side) but distinct from this (write-side staleness)
• [UX] /config is read-only in TUI — cannot modify settings mid-session #4073 (open) — /config is read-only in TUI mid-session — implicitly acknowledges runtime mutation isn't well-supported, but doesn't cover the silent-failure shape of this bug