echo6-docs/vault/projects/meshai-config-hot-apply.md

title	type	tags	aliases	related	updated	status
meshai Config Hot-Apply — Kill the Restart-Required GUI Friction	project	mesh		meshai-region-routing-plan meshai-prod-compose-override meshai SESSION-HANDOFF-meshai-test meshcore-transport	2026-07-13	proposed

meshai Config Hot-Apply (proposed — backlog)

Goal (Matt, 2026-07-07): make GUI Save just apply instead of prompting a container restart. The restart-required system is high-friction and Matt loathes it. This doc captures the why + the fix so it can be picked up later. Surfaced while building meshai-region-routing-plan (coverage/region edits require a restart to take effect).

Why saves don't apply today

Split by who reads the config, and when:

• Live sections (already hot): notifications, rules, toggles, timezone, etc. Their consumer reads config.<section> fresh on every use (e.g. the dispatcher reads config.notifications per event), so the save endpoint's setattr(app.state.config, section, new_value) (dashboard/api/config_routes.py:176) is enough — instant.

• Restart-required sections snapshot config at BOOT and never re-read it. RESTART_REQUIRED_SECTIONS (config_routes.py:29-40) = connection, llm, mesh_sources, meshmonitor, dashboard, environmental, coverage, generic_sources. Examples:

  • CoverageFilter is built ONCE at pipeline start from areas_from_config(config.coverage) (notifications/pipeline/__init__.py:120-126) — frozen copy of the areas.
  • Adapters get their fetch bbox computed ONCE at boot via enclosing_bbox(coverage.areas) (main.py:413).
  • The radio connector opens the link ONCE with connection settings.

  A naive live mutate would leave these running objects on the OLD config while new data flows under new assumptions → a half-applied, inconsistent "transient AND-mode" (the exact thing the guard prevents). So the flag saves-to-disk-but-diverges-in-memory until restart. Crude but not arbitrary.

The fix: per-section reload hooks + a hot-apply framework

"Save just works" = on save, rebuild the component that snapshotted the config, atomically. Proposed framework: each section registers an optional apply_live(new_config) handler; the save endpoint calls it and only falls back to "restart required" when a section has no safe handler.

Tiers:

Tier	Sections	Effort
Easy wins	coverage, llm, environmental thresholds	Low — rebuild the filter / swap the LLM client / re-read next tick
Moderate	generic_sources, mesh_sources, environmental adapters	Medium — re-instantiate the adapter with new config
Genuinely disruptive	connection, meshmonitor, dashboard (port)	Replace full restart with an explicit "Reconnect" / "Rebind" action — a live half-swap of an open radio link is unsafe

Coverage is Tier-1: on save, rebuild CoverageFilter's area list + refresh the adapters' bbox → region edits apply fully live (both routing columns AND engine re-scoping). NOTE: the routing-column half is already solved in meshai-region-routing-plan (regions endpoint reads on-disk coverage, so columns are dynamic without restart); the remaining half is the engine re-scoping, which this work covers.

Two ways to scope

1. Narrow: just make coverage hot-apply (rebuild filter + adapter bbox on save) — bounded, testable, closes the region-routing loop end-to-end.
2. Broad: design the hot-apply framework and convert the whole GUI so Save just saves everywhere; restart becomes a last-resort scoped reconnect only. (Candidate for a local ultraplan.)

Payoff: kills ~90% of restart prompts (everything edited day-to-day) while keeping a safe, scoped reconnect for the handful of things that genuinely disrupt a live connection.