Improve hot-reload configs

[openminddev]

Now, we don't support to update a certain fields without restarting the whole system. It makes sense to load certain config fields (e.g., system_prompt_base) without restarting; watch for file changes.

[lau90eth]

Hi @openminddev

This PR implements a selective and safe hot-reload mechanism for runtime
configuration files, allowing specific fields (e.g. system_prompt_base)
to be updated at runtime without requiring a full system restart.

The solution introduces:

• A watchdog-based config file watcher
• An explicit whitelist of hot-reloadable fields
• Backward-compatible integration with the existing runtime lifecycle

The scope of the PR is intentionally limited to hot-reload improvements only.

Closes #984
Working on this bounty with PR #1031

[openminddev]

We are looking for a solution where

OM1/src/runtime/single_mode/cortex.py

Line 221 in 3ca608d

async def _check_config_changes(self) -> None:

can reload based on the changed fields.

[lau90eth]

Yes, that makes sense
I can add a runtime-level hook in CortexRuntime so that reload decisions
are made based on the set of changed fields (partial vs full reload).
I’ll follow up with a revised PR.

[Bara69chn]

Hi, Dev. I’d like to help on this bounty #984. I see PRs #1020/#1031 working on selective reload. I can help by testing the hot-reload behavior (which fields trigger partial vs full reload), adding edge-case handling (invalid JSON5, rapid edits), and adding unit tests/docs if needed. Please point me to the expected field whitelist or desired behavior

[Wanbogang]

Hi @openminddev
I've submitted a PR implementing the selective hot-reload feature with watchdog.

Key features:

• Watchdog-based monitoring (no more polling)
• Selective reload for system prompts and LLM params
• Critical fields (model, provider, api_key) marked as restart-required
• Backward compatible

PR: #1061

Looking forward to your feedback

[lau90eth]

Thanks @openminddev for the recent PRs and discussion around selective hot-reload.

Just to clarify alignment with the original maintainer feedback on this issue:
the key requirement was that reload decisions live explicitly in
src/runtime/single_mode/cortex.py, and that CortexRuntime itself decides
partial vs full reload based on the set of changed fields.

I’ve submitted a revised PR ( #1064 ) that focuses specifically on that point:

• config changes are diffed into a field-level patch
• the patch is passed to CortexRuntime.reload_from_patch(...)
• CortexRuntime owns the decision of safe in-place update vs full reload fallback

The watcher remains generic and does not encode reload policy.
This is intended to directly address the earlier feedback that led to closing
the previous attempt.

Happy to adapt or align if there’s a preferred direction going forward.