Giving each controller a Reload method worked for the notification pipes, which
only hold values, but not for the two bot channels: the NapCat client is stopped
through a sync.Once and the Telegram polling context is created with the
controller, so neither can be re-pointed once running. It also meant five
bespoke implementations of the same idea.
Replace the whole set instead. Manager.ReplaceAll swaps in a freshly built set
under the registry lock, stops the outgoing controllers outside it, and starts
the incoming ones off the calling goroutine so a slow handshake does not hold
the settings request open. ReplaceAll is now the only way controllers are
installed: Register is gone, because a lone registration would slip a controller
in without recording the settings it was built from, which is what NeedsRebuild
compares against.
The rebuild runs from the reload hook and only fires when the controllerMethod
section actually changed, so saving a message template or a webhook endpoint
leaves the channels alone. NeedsRebuild compares the whole section by value, and
a test pins that reloading a file reproduces it exactly — a default applied
inconsistently would make every save look like a controller change and tear down
every channel.
With controllers immutable after construction and replaced wholesale, the atomic
pointers and Reload methods have no writers left, so they are gone and the pipes
return to plain fields. notifyReload now serializes hook execution for the same
reason: a hook mutates process-wide state, and two overlapping settings updates
would otherwise each decide to rebuild from their own view of what was applied.
Kept from that work: applyEmailProxy restores gomail's default dialer when
networkUseProxy is turned off. The old constructor only ever installed the proxy
dialer, so switching the flag back left SMTP tunnelled through a proxy with no
way to undo it short of a restart.
Documents the resulting behaviour in the README under "What applies without a
restart".
A settings update replaced the in-memory configuration but nothing told the code
that derives state from it. The logger is the first such consumer: its debug
flag is captured once at startup by postLog.SetDebugMode, so toggling
system.debugMode at runtime changed what config.IsDebugMode() reported while
DEBUG lines stayed filtered by the value from boot.
Add a hook registry to the config package. OnReload registers a callback and
every writer of config.json runs the registered hooks with the configuration now
in effect. It lives in config so packages that already depend on it (the logger,
and later the controller manager) can react without config importing them back,
which would be an import cycle.
Hooks run after settingsLock is released, not inside it. They will rebuild
controllers once those support reloading, which can wait on a network call, and
settingsLock is also held by the node tracker's background save — running a hook
under the lock would stall node registration behind a settings edit.
runSettingsUpdate now owns that lock/notify sequence for all four writer paths
(UpdateSettings and the three webhook endpoint helpers).
A panicking hook is logged and skipped: by then the new configuration is on disk
and published, so reporting the write as failed would be a lie, and the hooks
registered after the broken one still need to run.
NanamiAdmin