ModelRouter/internal/lua/plugins.go

package lua

// Plugin runtime.
//
// A plugin is a single .lua file, loaded from its own directory, that extends
// the gateway in two ways:
//
//	1. Hooks: it registers callbacks on the request pipeline's stages
//	   (request_start, response_end, …). A hook receives a JSON table and
//	   returns either nil (no opinion) or a JSON object.
//	2. UI: at boot it returns HTML/CSS/JS fragments that the kernel injects
//	   into the WebUI — either as a whole new page, or as an extra element on
//	   an existing page.
//
// Design notes that are load-bearing (each one cost something to learn):
//
//   - Plugins run in a SEPARATE VM from adapters, and each plugin gets its own
//     elastic pool, exactly like an adapter. Sharing one state would let a
//     plugin's globals corrupt an adapter's protocol translation (or vice
//     versa), and a plugin is third-party code while an adapter is core.
//
//   - A plugin that errors must NEVER break request forwarding. Hook calls are
//     wrapped so a plugin error is logged and the original value is returned
//     unchanged. A broken plugin is a missing feature, not an outage — the same
//     reason transform_stream_chunk swallows errors today.
//
//   - The billing plugin therefore cannot be trusted to be the source of truth
//     for anything the gateway must enforce. It reads usage off the hook
//     payload and accumulates in its own state; the gateway's own quota
//     accounting (internal/gateway/stats.go) stays authoritative for limits.
//     Two accounting paths that disagree is worse than one that is slightly
//     less featureful, so the split is explicit and documented.

import (
	"encoding/json"
	"fmt"
	"log"
	"os"
	"path/filepath"
	"sort"
	"strings"
	"sync"
	"sync/atomic"
	"time"

	golua "github.com/aarzilli/golua/lua"
)

// Stage identifies a point in the request pipeline. Plugins may register a
// function for any stage; unknown stages are ignored at call time so an old
// plugin survives a gateway that grew new stages.
type Stage string

// The pipeline stages. The order here is the order they fire in; it is the
// contract plugins are written against.
const (
	// StageRequestStart fires after the gateway has parsed and authorized a
	// request but BEFORE any upstream slot is chosen. payload:
	//   stage, type ("chat"|"stream"|"image"), model (as requested),
	//   key (masked gateway key id), role, source (empty), stream (bool),
	//   messages_count, tools_count, ts (unix seconds).
	StageRequestStart Stage = "request_start"

	// StageRouted fires once a (source, model) slot has been chosen and before
	// the upstream call. payload adds: source, model (the resolved one),
	// tier (AUTO tier, -1 for the direct path), stream.
	StageRouted Stage = "routed"

	// StageChainStep fires ONCE PER STEP of an AUTO chain walk, and only on the
	// AUTO path (a direct request has no chain and therefore emits nothing).
	//
	// This exists because StageRouted cannot express degradation: it fires once,
	// after the walk, with the slot that finally won. "tier 1 was cooling so we
	// dropped to tier 3" and "tier 1 served it" were indistinguishable. That
	// distinction is the whole point of a priority chain, and it is what an
	// operator debugging "why did my expensive model not get used" needs.
	//
	// payload:
	//   kind      "tier_skip" | "slot_fail" | "tier_busy" | "selected"
	//   tier      the AUTO tier this step belongs to (1 = highest priority)
	//   source / model  set for slot_fail and selected
	//   reason    human-readable cause, for tier_skip and tier_busy
	//   error     the underlying error text, for slot_fail
	//   attempt   1-based slot attempt within this walk
	//
	// Ordering: every step precedes StageRouted, and the "selected" step is the
	// last one. A plugin accumulating the walk therefore has the full picture
	// by the time request_end arrives.
	//
	// These events are OBSERVATION ONLY — see the accounting note in
	// docs/plugins.md: nothing here feeds back into routing, cooldown or quota.
	StageChainStep Stage = "chain_step"

	// StageRequestEnd fires exactly once per request, after the client response
	// has been produced (or after a failure was recorded). payload adds:
	// source, model, ok, status, latency_ms, first_byte_ms, prompt_tokens,
	// completion_tokens, cache_hit_tokens, cache_miss_tokens, image_count,
	// error ("" when ok).
	//
	// This is the stage a billing plugin should read: it carries the final
	// accounting for the request, including the upstream's own usage numbers.
	StageRequestEnd Stage = "request_end"
)

// AllStages is the firing order, used by the docs and by the hook listing.
var AllStages = []Stage{
	StageRequestStart,
	StageChainStep,
	StageRouted,
	StageRequestEnd,
}

// UIExtension is what a plugin contributes to the WebUI at boot.
type UIExtension struct {
	// Page is a whole new sidebar entry + pane. Requires PageID and Title.
	// The kernel renders Page's HTML into a pane whose id is "tab-"+PageID and
	// adds a sidebar button with data-tab="<PageID>".
	Page *UIPage `json:"page,omitempty"`
	// Pages is the multi-page form of the same thing. A plugin with several
	// distinct screens (billing totals vs. the price rules that produced
	// them) would otherwise have to cram both into one pane behind tabs, or
	// smuggle the second one in as a hidden element. Both make the sidebar
	// lie about what the plugin contributes.
	//
	// Page and Pages merge: a plugin may use either or both.
	Pages []*UIPage `json:"pages,omitempty"`
	// Elements are snippets injected into EXISTING pages, keyed by target page
	// id (e.g. "status", "keys"). Order within a target is plugin load order.
	Elements []UIElement `json:"elements,omitempty"`
	// Visibility is the plugin-wide default audience, applied to every page and
	// element this plugin contributes. A per-page Audience narrows it further.
	Visibility string `json:"visibility,omitempty"`
	// StateScope splits the plugin's published state into global and per-caller
	// buckets, for a plugin that is readable by more than one role but must not
	// hand one caller another's data.
	StateScope *StateScope `json:"state_scope,omitempty"`
}

// UIPage is a plugin-provided page.
type UIPage struct {
	PageID string `json:"page_id"` // kebab-case; becomes data-tab and #tab-<id>
	Title  string `json:"title"`   // sidebar label
	Icon   string `json:"icon"`    // optional inline SVG or short glyph
	Order  int    `json:"order"`   // sidebar sort key (default 100)
	// Audience limits which API-key roles may see this page. "" and "all" mean
	// every authenticated role; "admin" hides it from user keys; a comma list
	// ("admin,user") names the roles allowed.
	//
	// This exists because a plugin page is not automatically public data. The
	// billing page shows GLOBAL cost across every source, and the quota page
	// shows upstream balances — neither belongs to whoever happens to hold a
	// user key, yet both were served to them because /api/ui-inject applied no
	// role filter at all.
	//
	// The plugin declares its own default and the gateway may override it (see
	// SetPluginVisibility), because the plugin author cannot know the operator's
	// policy but the operator must be able to impose one.
	Audience string `json:"audience,omitempty"`
	// Mount is the page body. It may contain <script> and <style>; the kernel
	// executes scripts AFTER injecting the HTML so the DOM exists, and exposes
	// `pluginAPI` to them (see docs/plugins.md).
	Mount string `json:"mount"`
}

// UIElement is a snippet injected into an existing page.
type UIElement struct {
	// Target is the id of the pane to inject into: "status", "chat", "keys",
	// "sort", "sources" or "adapters".
	Target string `json:"target"`
	// Audience is the same role filter as UIPage.Audience. An injected tile can
	// leak exactly what a whole page would — the billing status tile shows total
	// cost — so the field exists here too rather than only on pages.
	Audience string `json:"audience,omitempty"`
	// Anchor is where in the target pane: "top", "bottom" or "before:<sel>" /
	// "after:<sel>" for a CSS selector. Empty = "bottom".
	Anchor string `json:"anchor,omitempty"`
	Order  int    `json:"order,omitempty"` // sort key within the target
	Mount  string `json:"mount"`
}

// PluginInfo is the manifest a plugin declares.
type PluginInfo struct {
	Name        string `json:"name"`
	Version     string `json:"version"`
	Description string `json:"description,omitempty"`
	Author      string `json:"author,omitempty"`
}

// Plugin is one loaded plugin.
type Plugin struct {
	Info  PluginInfo
	Hooks map[Stage]string // stage -> exported function name
	UI    *UIExtension
	// LoadError is non-empty when the plugin failed to compile or register. Such
	// a plugin is listed in the UI with its error but is never called.
	LoadError string
	// Disabled marks a plugin the operator switched off. It stays on disk and
	// keeps its name, hooks and UI declared (so the UI can show it greyed out
	// and report what it WOULD contribute), but Fire never calls it and its
	// UI extension is excluded from the inject payload.
	//
	// Disabling is deliberately NOT deleting: a plugin that breaks a live
	// gateway is often one line away from being fixed, and an operator needs a
	// way to take it out of the request path without losing it. That is the same
	// reasoning as a systemd unit being masked rather than removed.
	Disabled bool
	dir      string
	// http lists endpoints this plugin mounts under /api/plugins/http/. Read at
	// load time, never mutated afterwards.
	http []HTTPHandler
	// exports lists the functions this plugin allows the WebUI to call through
	// POST /api/plugins/{name}/call/{fn}. It is read once at load from
	// plugin.exports and never changes, so Call can check it without taking a
	// Lua read on the hot path.
	exports []string
	// script is the plugin's source, kept so a reload can rebuild its state.
	script string
	// Builtin marks a plugin that shipped with the gateway. The UI shows it as
	// such so an operator can tell "example I can read" from "mine", and a
	// delete of a builtin is allowed but re-seeds on a fresh plugin dir.
	Builtin bool

	// state is the plugin's SINGLE authoritative Lua state, guarded by mu.
	//
	// A plugin deliberately does NOT use the adapter's elastic pool. Adapters
	// are per-worker stateless transforms, so N independent states are correct
	// (and necessary) for them. A plugin's hook, however, typically ACCUMULATES
	// into `plugin.state` — the billing plugin's totals live there — so two
	// states would mean two divergent sets of totals, and whichever worker a
	// hook happened to get would see a different number. That was a real bug
	// found by a test: SetState wrote prices to one worker, the hook then ran on
	// another and priced everything at zero.
	//
	// A single state is sufficient because every entry point (Fire, State,
	// SetState) serializes on mu, and a plugin hook is a short synchronous call.
	// A hook that blocks for seconds would stall every other plugin's hook,
	// which is the real cost — documented as a rule in docs/plugins.md.
	mu    sync.Mutex
	state *worker

	// persisted marks this plugin's state as having been restored from disk.
	// It gates the first save: without it, a freshly loaded plugin whose state
	// is still the compiled-in default would immediately overwrite the file it
	// was supposed to inherit from.
	persisted bool

	// ExtraHosts lists outbound hosts this plugin declared in its own source
	// (`plugin.hosts`) beyond the operator-configured sources. A REQUEST, not a
	// grant: the operator approves each one in the settings page.
	ExtraHosts []string

	// KernelDeclared lists the kernel actions this plugin asks for in its own
	// source (`plugin.kernel = {...}`). It is a REQUEST, shown to the operator
	// beside what was actually granted; it grants nothing by itself.
	KernelDeclared []string

	// Schedule is the periodic work this plugin declared, and Runtime is the
	// scheduler that runs it on THIS plugin's own VM (p.state) under p.mu — the
	// same state and the same lock a request hook uses. That is what makes a
	// token rotated by a task immediately visible to inference traffic.
	//
	// Runtime is nil when the plugin declared nothing — the common case stays
	// exactly as cheap as before: one VM, no extra goroutine. The point of the
	// schedule is auto check-in / token rotation: work that must happen with no
	// browser open and nobody clicking anything.
	Schedule []ScheduledTask
	Runtime  *PluginRuntime
}

// Plugins is the loaded plugin set, owned by the VM.
type Plugins struct {
	mu      sync.RWMutex
	vm      *VM
	dir     string    // plugin directory; "" disables plugin loading entirely
	plugins []*Plugin // load order; the index is the stable plugin id
	// ui caches the merged UI extensions so the boot payload is computed once.
	ui atomic.Pointer[UIExtension]
	// stageFuncs is the precomputed stage -> []hookCall, in plugin load order.
	stageFuncs map[Stage][]hookCall
	// hookErr records per-stage plugin failures so a silently broken plugin is
	// visible in /api/status rather than merely missing.
	hookErr *hookErrors
	// logf, when set, receives persistence warnings. It is a field rather than a
	// direct log call so the plugin package stays free of a logging dependency
	// and tests can capture the warnings.
	logf func(format string, args ...interface{})
	// visibility holds the operator's per-plugin audience overrides, keyed by
	// plugin name. Consulted by UIForRole and the call endpoint's authorisation.
	visibility visibilityOverrides
	// adapterOwn records which plugin installed each adapter, so a plugin can
	// only update adapters it owns (see plugin_adapter_bridge.go for why this
	// is state and not a convention).
	adapterOwn     *adapterOwnership
	adapterOwnOnce sync.Once
	// routeWinners records the operator's arbitration for contested HTTP service
	// paths: a path two+ plugins declared, resolved to one winning plugin. Keyed
	// by normalized path. See route_conflicts.go — a contested path that is NOT
	// in here is unresolved and the dispatcher returns 409 rather than silently
	// picking a load-order winner. Persisted in .route-conflicts.json.
	routeWinners map[string]string
	// adapterWinners records the operator's arbitration for an adapter NAME two
	// things claimed (user requirement B): "incumbent" keeps the built-in / first
	// plugin, "candidate:<plugin>" lets that plugin's version win. Persisted in
	// .route-conflicts.json beside routeWinners. See route_conflicts.go.
	adapterWinners map[string]string
	// adapterConflicts holds adapter-name installs currently SUSPENDED pending an
	// operator decision: a plugin tried to install an adapter whose name an
	// incumbent already holds. The incumbent keeps serving; the candidate waits
	// here (with its code) until arbitrated. In-memory only — a restart replays
	// the install, which re-suspends unless a persisted winner says otherwise.
	adapterConflicts map[string]*AdapterConflict
	// shards runs exported functions on private VMs so blocking work can be
	// genuinely parallel. See plugin_shard.go for why the shared VM's lock makes
	// concurrency impossible and what a shard must therefore never touch.
	shards *shardPool
	// stateDirDisabled turns persistence off. Used by tests that assert state
	// starts empty, and by an embedder that has no writable plugin dir.
	stateDirDisabled bool
	// allowHTTP bounds plugin outbound requests to operator-declared hosts.
	// nil means deny-all: an unconfigured gateway must not silently grant a
	// plugin unrestricted network access.
	allowHTTP *httpAllowlist
	// kernelAPI is the injected kernel action surface (nil when unwired).
	// See kernel_api.go: a plugin calling __llmsproxy_kernel reaches the core
	// through this interface, never through a direct import (core depends on
	// lua, not the other way round).
	kernelAPI KernelAPI
	// secretBox seals the credential vault at rest. nil = write plaintext (tests
	// / embedders with no runtime dir). Injected by core via SetSecretBox: it is
	// the deployment's ONE master key, the same one that seals config.yaml, so a
	// lost master.key loses the plugin tokens too — loud, consistent, not a
	// second key to manage.
	secretBox secretSealer
	// kernelPerms records what the OPERATOR granted each plugin. Separate from
	// what a plugin DECLARES it wants (Plugin.KernelDeclared): a declaration is
	// a request, a grant is a privilege. Persisted in .access.json.
	kernelPerms kernelPermissionSet
	// sourceHosts remembers the source-derived allowlist so a rebuild can fold it
	// together with any granted login hosts.
	sourceHosts []string
	// extraHosts is the per-plugin set of APPROVED extra outbound hosts. Separate
	// from what a plugin declared (Plugin.ExtraHosts): a declaration is a request.
	extraHosts map[string][]string
	// save coalesces state writes into one background goroutine.
	//
	// A hook must NOT write synchronously: the billing plugin mutates state on
	// every single request, and serializing the whole state per request would put
	// a file write (plus a full JSON encode) on the hot path — measured at
	// 14.6us per hook call already, a write would dominate it. Instead hooks mark
	// the plugin dirty and this saver flushes, so N requests between two flushes
	// cost one write.
	save *stateSaver
}

// stateSaver coalesces state writes: one goroutine, a minimum interval between
// flushes, and a dirty set. Under load the number of writes is bounded by the
// timer rather than by the request rate.
type stateSaver struct {
	mu    sync.Mutex
	ps    *Plugins
	dirty map[string]bool
	wake  chan struct{}
	stop  chan struct{}
	done  chan struct{}
	// wakeDelay is how long a burst suppresses the next tick. A separate field
	// (not the constant 2s) so tests can shorten it: Close() waits for the saver
	// goroutine, so a test with the production interval pays that interval on
	// every teardown.
	wakeDelay time.Duration
	once      sync.Once
	// interval is the minimum gap between flushes.
	interval time.Duration
	// writes counts completed file writes; tests read it to assert that N
	// mutations did NOT become N writes.
	writes atomic.Int64
}

// setWakeDelayForTest shortens the debounce so tests do not pay it on teardown.
func (s *stateSaver) setWakeDelayForTest(d time.Duration) { s.wakeDelay = d }

func newStateSaver(ps *Plugins) *stateSaver {
	s := &stateSaver{
		ps:        ps,
		dirty:     map[string]bool{},
		wake:      make(chan struct{}, 1),
		stop:      make(chan struct{}),
		done:      make(chan struct{}),
		interval:  2 * time.Second,
		wakeDelay: 2 * time.Second,
	}
	go s.loop()
	return s
}

// markDirty records that a plugin's state changed. Never blocks: the channel is
// buffered, and a full buffer means a flush is already pending, which is exactly
// the state we want.
func (s *stateSaver) markDirty(name string) {
	if s == nil {
		return
	}
	s.mu.Lock()
	s.dirty[name] = true
	s.mu.Unlock()
	select {
	case s.wake <- struct{}{}:
	default:
	}
}

func (s *stateSaver) loop() {
	defer close(s.done)
	t := time.NewTicker(s.interval)
	defer t.Stop()
	for {
		select {
		case <-s.stop:
			// Final flush so a clean shutdown does not lose the tail — losing
			// the last interval of spend is the same bug in a smaller window.
			s.flush()
			return
		case <-t.C:
			s.flush()
		case <-s.wake:
			// Debounce a burst of marks into one write.
			time.Sleep(s.wakeDelay)
			s.flush()
		}
	}
}

// flush writes every dirty plugin's state.
func (s *stateSaver) flush() {
	if s == nil {
		return
	}
	s.mu.Lock()
	if len(s.dirty) == 0 {
		s.mu.Unlock()
		return
	}
	names := make([]string, 0, len(s.dirty))
	for n := range s.dirty {
		names = append(names, n)
	}
	s.dirty = map[string]bool{}
	s.mu.Unlock()

	// Snapshot + write here, off the request path. This is the ONLY place the
	// plugin's Lua tables are walked for persistence.
	for _, n := range names {
		s.ps.persistByName(n)
		s.writes.Add(1)
	}
}

// Close stops the saver AFTER its final flush completes.
//
// It must WAIT for that flush, not just signal it. Signalling and returning
// leaves the write to a goroutine that the caller is about to tear down
// (vm.Stop() frees the plugin states; the test process is exiting), so the
// final state is silently lost — which is the very defect persistence was added
// to fix. Close is on the shutdown path, where a few milliseconds is free.
func (s *stateSaver) Close() {
	if s == nil {
		return
	}
	s.once.Do(func() { close(s.stop) })
	<-s.done
}

type hookCall struct {
	pluginIdx int
	plugin    string
	fn        string
}

// hookErrors counts hook failures per stage, surfaced in /api/status so a
// silently broken plugin is visible instead of just missing.
type hookErrors struct {
	mu     sync.Mutex
	counts map[Stage]int
	last   map[Stage]string
}

func newHookErrors() *hookErrors {
	return &hookErrors{counts: map[Stage]int{}, last: map[Stage]string{}}
}

func (h *hookErrors) note(s Stage, msg string) {
	h.mu.Lock()
	h.counts[s]++
	if len(msg) > 200 {
		msg = msg[:200]
	}
	h.last[s] = msg
	h.mu.Unlock()
}

func (h *hookErrors) snapshot() map[string]map[string]interface{} {
	h.mu.Lock()
	defer h.mu.Unlock()
	out := map[string]map[string]interface{}{}
	for s, n := range h.counts {
		if n == 0 {
			continue
		}
		out[string(s)] = map[string]interface{}{"count": n, "last_error": h.last[s]}
	}
	return out
}

// pluginGlobal is the Lua global the plugin's returned table is stored under,
// mirroring adapterGlobal for adapters.
const pluginGlobal = "__llmsproxy_plugin"

// NewPlugins creates the plugin registry for a VM. dir is the plugin directory;
// a missing directory is not an error (plugins are optional).
func NewPlugins(vm *VM, dir string) *Plugins {
	ps := &Plugins{
		vm:         vm,
		stageFuncs: map[Stage][]hookCall{},
		hookErr:    newHookErrors(),
		dir:        dir,
		// Deny-by-default. NewPlugins is called during wiring, before any
		// source list is known; SetOutboundHosts replaces this with the real
		// set. A plugin loaded in that window simply cannot make requests.
		allowHTTP: newHTTPAllowlist(nil),
	}
	// The saver is started even with no dir: a gateway can be given a plugin dir
	// later, and a saver that only exists when dir != "" would silently never
	// flush. markDirtyLocked and flush both no-op when the dir is empty.
	ps.save = newStateSaver(ps)
	ps.shards = newShardPool(ps)
	return ps
}

// Close stops the background state saver, flushing once more first.
//
// The gateway MUST call this on shutdown. Without it the last flush interval's
// accumulation is lost, which is the same "restart loses the books" defect this
// persistence exists to fix, just scoped to a few seconds instead of the whole
// process lifetime.
// DisableStatePersistence turns persistence off (tests/benchmarks only).
func (ps *Plugins) DisableStatePersistence() { ps.stateDirDisabled = true }

// SetOutboundHosts declares which hosts plugins may contact via
// __llmsproxy_http. The gateway passes the host of every configured source:
// those are operator-declared upstreams, and reusing that set keeps a plugin's
// reach no wider than the deployment already trusts. An empty list means no
// outbound host is allowed at all — plugins then still load, they just cannot
// make requests.
func (ps *Plugins) SetOutboundHosts(hosts []string) {
	ps.mu.Lock()
	defer ps.mu.Unlock()
	// Remembered so a later extra-host grant can rebuild the allowlist as
	// sources ∪ granted-extra-hosts instead of losing the source half.
	ps.sourceHosts = append([]string(nil), hosts...)
	ps.rebuildOutboundAllowlistLocked()
}

// outboundAllowlist returns the current allowlist. bootPluginState reads it
// before any concurrent reload can swap it, and a nil result is treated as
// "deny everything" rather than "allow everything".
func (ps *Plugins) outboundAllowlist() *httpAllowlist {
	ps.mu.RLock()
	defer ps.mu.RUnlock()
	return ps.allowHTTP
}

func (ps *Plugins) Close() {
	if ps == nil {
		return
	}
	// Shard VMs are separate golua states; leaving them open leaks a VM per
	// shard on every reload.
	ps.shards.Shutdown()
	ps.save.Close()
}

// LoadDir loads every .lua file in dir as a plugin. Files are loaded in
// lexical order so a plugin's UI order is stable across restarts.
//
// A plugin that fails to compile or register is NOT fatal: it is kept with its
// LoadError so the UI can show it, and it is never called. This is the same
// posture as adapters, except adapters are core while plugins are not.
func (ps *Plugins) LoadDir() error {
	if ps.dir == "" {
		return nil
	}
	// Overrides first: they are keyed by plugin name and must be in place before
	// anything reads them, and a reload must not blank a restriction the
	// operator set earlier in this same process.
	if err := ps.LoadAccessOverrides(); err != nil {
		log.Printf("[core] plugin access overrides not loaded: %v", err)
	}
	// Kernel grants load beside them and for the same reasons: an in-memory map
	// would make "this plugin may reload config" look enforced right up to the
	// next restart, at which point the grant silently vanished.
	if err := ps.LoadKernelPermissions(); err != nil {
		log.Printf("[core] plugin kernel permissions not loaded: %v", err)
	}
	// Route-conflict arbitrations are operator policy too: who wins a service
	// path two plugins both declared. Must be in place before the first request
	// is dispatched, or a contested path would 409 until the operator re-chose.
	if err := ps.LoadRouteConflicts(); err != nil {
		log.Printf("[core] plugin route conflicts not loaded: %v", err)
	}
	if _, err := os.Stat(ps.dir); os.IsNotExist(err) {
		return nil
	}
	entries, err := os.ReadDir(ps.dir)
	if err != nil {
		return err
	}
	var names []string
	for _, e := range entries {
		if e.IsDir() || !strings.HasSuffix(e.Name(), ".lua") {
			continue
		}
		names = append(names, e.Name())
	}
	sort.Strings(names)
	for _, n := range names {
		path := filepath.Join(ps.dir, n)
		code, err := os.ReadFile(path)
		if err != nil {
			continue
		}
		if err := ps.LoadSource(strings.TrimSuffix(n, ".lua"), string(code)); err != nil {
			// LoadSource records the error on the plugin itself; keep going so
			// one bad plugin does not stop the others from loading.
			continue
		}
	}
	ps.rebuild()
	// Persisted host grants must be folded into the allowlist HERE, not only in
	// the WebUI grant path.
	//
	// Why (live defect): LoadKernelPermissions above loads `jp.Hosts` into
	// ps.extraHosts but deliberately does NOT fold it — its comment says "the
	// caller owns that fold". On the boot path nobody did. SetOutboundHosts had
	// already run (core.New, before LoadDir), so the allowlist held the source
	// hosts and nothing else, and every approved extra host (the plugin's whole
	// declared set: zcode.z.ai, copilot.tencent.com, …) was refused after a
	// restart with "host ... is not an allowed outbound destination".
	//
	// Symptom is the classic "restart loses the operator's decision": it clears
	// the moment an operator opens the plugin page and re-saves the grants. It
	// also broke the scheduled work — refresh_models fires 5s after boot, inside
	// that window, and reported the grant as missing for every platform.
	ps.mu.Lock()
	ps.rebuildOutboundAllowlistLocked()
	ps.mu.Unlock()
	return nil
}

// SeedBundled writes the plugins shipped with the gateway into dir when the
// directory does not exist yet, mirroring the adapter seeding rule: once the
// directory exists it is authoritative, so deleting or editing a shipped plugin
// is a real action that survives restarts.
// SeedBundled writes the plugins shipped with the gateway into dir when they are
// MISSING.
//
// This used to be "write everything only when the directory does not exist",
// which silently stopped delivering plugins added to the binary LATER: the
// directory had existed since the first install, so a newly embedded plugin
// (e.g. channels.lua) never landed on an upgraded deployment — it was in the
// binary and completely invisible in /api/plugins. The directory's mere
// existence is not evidence that its contents are current.
//
// The rule is now per-file: seed a bundled plugin only when a file of that name
// is absent. An OPERATOR'S file always wins, so customising or replacing a
// bundled plugin survives an upgrade — the same rule adapters already use.
func (ps *Plugins) SeedBundled() error {
	if ps.dir == "" {
		return nil
	}
	return writeBundledPlugins(ps.dir)
}

// bootPluginState compiles one plugin into a fresh Lua state and stores its
// returned table under pluginGlobal. It is the plugin counterpart of
// adapterPool.boot, minus the pooling: a plugin keeps exactly one state.
func bootPluginState(code, name string, allow *httpAllowlist, ps *Plugins) (*worker, error) {
	L := golua.NewState()
	L.OpenLibs()
	setupGlobals(L)
	// Only plugins get __llmsproxy_http. Adapters are operator-supplied config
	// and go through setupGlobals too, so installing the bridge here — not in
	// setupGlobals — is what keeps the capability scoped to installed .lua
	// files.
	if allow != nil {
		registerFn(L, httpGlobal, luaHTTPFn(L, allow))
	}
	// R4: plugins — and only plugins — can install/update/remove their own
	// adapters at runtime. Scoped here (not in setupGlobals) for the same reason
	// as __llmsproxy_http: adapters share setupGlobals and must not get it.
	if ps != nil {
		registerAdapterBridge(L, ps, name)
		// R4-d: the credential vault and the kernel action surface. Both are
		// plugin-only, for the same reason as the bridges above — an adapter
		// shares setupGlobals and must not be able to read credentials or drive
		// the kernel.
		registerSecretBridge(L, ps, name)
		registerKernelBridge(L, ps, name)
		// The plugin's own directory, as ground truth (see pluginDirGlobal).
		registerPluginDirBridge(L, ps.dir)
	}
	if err := L.DoString(code); err != nil {
		L.Close()
		return nil, fmt.Errorf("compile plugin %s: %w", name, err)
	}
	if L.Type(-1) != golua.LUA_TTABLE {
		L.Close()
		return nil, fmt.Errorf("plugin %s must return a table", name)
	}
	// SetGlobal POPS, so this is the only global it can set (see vm.go boot()).
	L.SetGlobal(pluginGlobal)
	L.SetTop(0)
	return &worker{L: L}, nil
}

// LoadSource loads one plugin from source text. name is the plugin id (the
// file's base name). It returns an error only for conditions the caller should
// see; a plugin that merely registers nothing is not an error.
func (ps *Plugins) LoadSource(name, code string) error {
	if name == "" {
		return fmt.Errorf("plugin name required")
	}
	if err := os.MkdirAll(ps.dir, 0755); err != nil && ps.dir != "" {
		return err
	}
	// Reloading must REPLACE, not append.
	//
	// LoadSource was reachable both from LoadDir and from the install/reload API,
	// and it appended unconditionally. Loading a plugin that was already loaded
	// therefore left two records with the same name, and findLocked — which
	// returns the FIRST match — kept handing out the OLD one. Measured: after
	// LoadDir + LoadSource of an edited plugin, /api/plugins listed both
	// versions and Call answered from version 1 while version 2 sat inert
	// alongside it. Editing a plugin and saving appeared to do nothing.
	//
	// Unload is a no-op when nothing is loaded under that name, so this is safe
	// on the first load.
	if ps.find(name) != nil {
		if err := ps.Unload(name); err != nil {
			return fmt.Errorf("replace %s: %w", name, err)
		}
	}
	p := &Plugin{
		Info:  PluginInfo{Name: name},
		Hooks: map[Stage]string{},
		dir:   ps.dir,
	}

	// A plugin runs in its OWN Lua state, separate from every adapter and every
	// other plugin, so an error in one cannot corrupt another. The returned
	// table is stored under pluginGlobal, exactly like adapterGlobal.
	//
	// One state, held for the plugin's lifetime (see Plugin.state): a hook that
	// accumulates into plugin.state would otherwise split its totals across
	// whichever worker it happened to run on.
	w, err := bootPluginState(code, name, ps.outboundAllowlist(), ps)
	if err != nil {
		p.LoadError = err.Error()
		ps.append(p)
		ps.rebuild()
		return err
	}
	p.script = code
	p.state = w
	// Builtin = the shipped source is byte-identical to what this file was
	// loaded from. That is stronger than "the name matches a bundled plugin": an
	// operator who EDITED billing.lua must not be told their copy is builtin,
	// or the UI would offer to overwrite their changes.
	if orig, err := ReadBundledPlugin(name); err == nil && orig == code {
		p.Builtin = true
	}

	// ---- manifest ----
	w.L.GetGlobal(pluginGlobal)
	if !w.L.IsNil(-1) {
		w.L.GetField(-1, "name")
		if s := w.L.ToString(-1); s != "" {
			p.Info.Name = s
		}
		w.L.SetTop(-2)
		w.L.GetField(-1, "version")
		if s := w.L.ToString(-1); s != "" {
			p.Info.Version = s
		}
		w.L.SetTop(-2)
		w.L.GetField(-1, "description")
		if s := w.L.ToString(-1); s != "" {
			p.Info.Description = s
		}
		w.L.SetTop(-2)
		w.L.GetField(-1, "author")
		if s := w.L.ToString(-1); s != "" {
			p.Info.Author = s
		}
		w.L.SetTop(-2)
	}
	w.L.SetTop(0)

	// ---- hooks ----
	for _, st := range AllStages {
		if fn, ok := pluginHookName(w.L, string(st)); ok {
			p.Hooks[st] = fn
		}
	}

	// ---- schedule ----
	//
	// A plugin that declares `schedule` gets a kernel-driven scheduler running
	// ON ITS OWN VM (the same p.state a hook uses), so its periodic work (auto
	// check-in, token rotation, quota polling) happens with no browser open and
	// no operator clicking anything. Without this the only periodic work a
	// plugin could do was a front-end poll, which is exactly the "nobody is
	// looking at the dashboard" case the feature exists for.
	//
	// The parse happens here (right after boot) so a malformed schedule is
	// visible at load time rather than silently ignored.
	p.Schedule = parseSchedule(w.L).tasks
	w.L.SetTop(0)
	if len(p.Schedule) > 0 {
		p.Runtime = ps.startRuntime(p, scheduleSpec{tasks: p.Schedule})
		if p.Runtime == nil {
			// The scheduler could not start. That must NOT silently disable the
			// plugin — hooks still work — but the operator needs to know their
			// scheduled work will never run.
			log.Printf("[scheduler] plugin %s: declared %d task(s) but no runtime could be started",
				p.Info.Name, len(p.Schedule))
		}
	}

	// ---- UI ----
	if ui, ok := readPluginUI(w.L, p.Info.Name); ok {
		p.UI = ui
	}

	// ---- declared kernel requests ----
	// A declaration is a REQUEST, not a grant (see kernel_api.go). Reading it at
	// load time lets the settings page show "this plugin wants add_source"
	// beside what was actually granted.
	p.KernelDeclared = readPluginKernelDeclared(w.L)
	w.L.SetTop(0)

	// ---- declared extra hosts ----
	// Some plugins need hosts that are not sources (account portals, claim
	// endpoints), which the outbound allowlist blocks by default. The plugin
	// declares what it needs; the operator approves each host. See plugin_outbound.go.
	p.ExtraHosts = readPluginExtraHosts(w.L)
	w.L.SetTop(0)

	// ---- http endpoints ----
	p.http = readPluginHTTP(w.L)

	// ---- exports ----
	// Read here, while the state is locked and before restore() runs: the list is
	// static for the life of the plugin, so Call never has to re-enter Lua just
	// to decide whether a call is allowed.
	p.exports = readPluginExports(w.L)

	// Restore persisted state AFTER the plugin compiled and registered its
	// hooks, so the restore overwrites the compiled-in defaults instead of being
	// overwritten by them. Doing it earlier would mean a restart resets the
	// totals back to whatever the .lua source initialises them to.
	ps.restore(p)

	ps.append(p)
	// Rebuild here rather than only in LoadDir: LoadSource is also the single-
	// plugin entry point (the WebUI upload path), and a caller that loads one
	// plugin and immediately fires a stage must not silently get nothing.
	ps.rebuild()
	return nil
}

// pluginHookName returns the exported function name a plugin registered for a
// stage. A plugin registers either `hooks = {request_end = "on_end"}` or a
// direct `request_end = function(...) end` on the returned table; both forms are
// accepted because the table form keeps the manifest tidy while the direct form
// is shorter for a single-hook plugin.
//
// For the anonymous-function forms the value is re-keyed onto the plugin table
// under a synthetic name so the hot path can fetch every hook by name.
//
// STACK DISCIPLINE (this binding aborts the PROCESS on a bad index — SIGABRT,
// not a Go panic, so nothing can recover it):
//
//   - GetField/SetField take the table by ABSOLUTE index, so the plugin table's
//     index must be re-read after every SetTop, since SetTop(0) invalidates it.
//   - Therefore each form re-pushes the plugin table and re-reads its index,
//     instead of caching one index across a reset. Getting this wrong was a
//     real crash found by running the test, not by reading the code.
func pluginHookName(L *golua.State, stage string) (string, bool) {
	L.SetTop(0)
	L.GetGlobal(pluginGlobal)
	if L.IsNil(-1) {
		L.SetTop(0)
		return "", false
	}
	plug := L.GetTop()

	// form 1: hooks = { request_end = "fn" }  (or = function)
	L.GetField(plug, "hooks")
	if L.Type(-1) != golua.LUA_TNIL {
		hooksIdx := L.GetTop()
		L.GetField(hooksIdx, stage)
		switch L.Type(-1) {
		case golua.LUA_TSTRING:
			name := L.ToString(-1)
			L.SetTop(0)
			if name != "" {
				return name, true
			}
			return "", false
		case golua.LUA_TFUNCTION:
			// An anonymous function: key it onto the plugin table under a stable
			// per-stage name so invoke() can fetch it like any other hook.
			name := "__hook_" + stage
			L.SetField(plug, name)
			L.SetTop(0)
			return name, true
		}
		L.SetTop(0)
	}
	// form 2: request_end = function(...) end directly on the table.
	// Re-push and re-read the index: SetTop(0) above invalidated `plug`.
	L.GetGlobal(pluginGlobal)
	if L.IsNil(-1) {
		L.SetTop(0)
		return "", false
	}
	plug = L.GetTop()
	L.GetField(plug, stage)
	if L.Type(-1) == golua.LUA_TFUNCTION {
		name := "__hook_" + stage
		L.SetField(plug, name)
		L.SetTop(0)
		return name, true
	}
	L.SetTop(0)
	return "", false
}

// readPluginUI reads the optional ui extension block.
// readPluginExports collects plugin.exports, the list of functions the WebUI is
// allowed to invoke through POST /api/plugins/{name}/call/{fn}.
//
// It accepts either an array of names or a map whose KEYS are the names — both
// shapes read naturally in Lua and authors reach for either. A missing or
// non-conforming table yields nil, which Call treats as "nothing is callable":
// a plugin that never declared an export simply has no callable surface.
func readPluginExports(L *golua.State) []string {
	// A plugin declares the functions the page may call as `exports = { "a", "b" }`.
	// Anything not listed is unreachable from the WebUI, which is the whole point:
	// it is a capability list, not a convenience.
	raw, ok := pluginFieldValue(L, "exports")
	if !ok {
		return nil
	}
	return decStringList(map[string]interface{}{"v": raw}, "v")
}

func readPluginUI(L *golua.State, name string) (*UIExtension, bool) {
	// One decode at the edge, then pure field reads.
	//
	// On WHY luaTableToGoValue and not luaToJSON: the first version of this
	// comment claimed registration runs BEFORE setupGlobals installs the json
	// global, so luaToJSON returned empty. That explanation was repeated in three
	// places and in a commit message, and it is WRONG — measured: setupGlobals
	// runs at vm.go:574, these readers at 787, and json IS present on the stack
	// when they run.
	//
	// What actually differs is INDEX SEMANTICS. luaTableToGoValue converts a
	// NEGATIVE index to absolute immediately (abs = GetTop() + idx + 1);
	// luaToJSON keeps the relative index and pushes it before Call. So after a
	// SetTop(0) the two disagree, and the table-valued converter — the one that
	// handles nested tables, arrays and mixed keys correctly — is the one that
	// needs the stack shape kept intact.
	//
	// Which is exactly why the decode happens here, ONCE, immediately after
	// GetGlobal+GetField, and every field read below is a pure function on the
	// decoded Go value. The old readers repeated the stack dance per field.
	m, ok := pluginFieldTable(L, "ui")
	if !ok {
		return nil, false
	}
	ctx := &decodeCtx{plugin: name}
	ui := decodeUIExtension(ctx, m)

	if ui.Page == nil && len(ui.Pages) == 0 && len(ui.Elements) == 0 {
		return nil, false
	}
	if len(ctx.problems) > 0 {
		log.Printf("[plugins] %s: %s", ctx.plugin, strings.Join(ctx.problems, "; "))
	}
	return &ui, true
}

// pluginFieldTable reads one top-level declaration table off the plugin global.
func pluginFieldTable(L *golua.State, field string) (map[string]interface{}, bool) {
	L.GetGlobal(pluginGlobal)
	if L.IsNil(-1) {
		L.SetTop(0)
		return nil, false
	}
	L.GetField(-1, field)
	if L.Type(-1) == golua.LUA_TNIL {
		L.SetTop(0)
		return nil, false
	}
	raw := luaTableToGoValue(L, -1)
	L.SetTop(0)
	m, ok := raw.(map[string]interface{})
	return m, ok
}

// decodeUIExtension reads a plugin's `ui` declaration.
//
// The field list below is the single source of truth for what the gateway
// understands. A declaration using anything else is REPORTED rather than
// silently dropped, because the old hand-written reader ignored unknown keys with
// no trace: `visibilty = "admin"` (one character off) restricted nothing and
// looked exactly like a plugin that simply chose not to declare an audience.
func decodeUIExtension(ctx *decodeCtx, m map[string]interface{}) UIExtension {
	ctx.strictKeys(m, "ui", "visibility", "state_scope", "page", "pages", "elements")

	var ui UIExtension
	ui.Visibility = decString(m, "visibility")

	if scope := decTable(m, "state_scope"); scope != nil {
		ctx.strictKeys(scope, "ui.state_scope", "global", "depth")
		sc := &StateScope{Global: decStringList(scope, "global")}
		if d := decIntMap(scope, "depth"); len(d) > 0 {
			sc.Depth = compileDepthMap(d)
		}
		ui.StateScope = sc
	}

	if pg := decTable(m, "page"); pg != nil {
		ui.Page = uipageFromValue(pg)
	}
	for _, pm := range decTableList(m, "pages") {
		ui.Pages = append(ui.Pages, uipageFromValue(pm))
	}
	for _, em := range decTableList(m, "elements") {
		ctx.strictKeys(em, "ui.elements[]", "target", "anchor", "mount", "audience", "order")
		el := UIElement{
			Target:   decString(em, "target"),
			Anchor:   decString(em, "anchor"),
			Mount:    decString(em, "mount"),
			Audience: decString(em, "audience"),
		}
		if n, ok := asFloat(em["order"]); ok {
			el.Order = int(n)
		}
		ui.Elements = append(ui.Elements, el)
	}
	return ui
}

func (ps *Plugins) append(p *Plugin) {
	ps.mu.Lock()
	ps.plugins = append(ps.plugins, p)
	ps.mu.Unlock()
}

// rebuild recomputes the stage dispatch table and the merged UI payload. It is
// called after any load so the hot path (Fire) is a slice walk with no map
// lookups or locking beyond one RLock.
func (ps *Plugins) rebuild() {
	ps.mu.Lock()
	defer ps.mu.Unlock()
	stageFuncs := map[Stage][]hookCall{}
	for i, p := range ps.plugins {
		// Disabled plugins are excluded from BOTH the dispatch table and the
		// merged UI below. Including them in the UI would render a page whose
		// refresh calls go to a plugin that is never consulted.
		if p.LoadError != "" || p.Disabled {
			continue
		}
		for _, st := range AllStages {
			if fn, ok := p.Hooks[st]; ok {
				stageFuncs[st] = append(stageFuncs[st], hookCall{pluginIdx: i, plugin: p.Info.Name, fn: fn})
			}
		}
	}
	ps.stageFuncs = stageFuncs

	merged := &UIExtension{}
	seenPage := map[string]bool{}
	for _, p := range ps.plugins {
		if p.LoadError != "" || p.Disabled || p.UI == nil {
			continue
		}
		// The single `page` field is folded into the same list as `pages`.
		// Assigning it to its own slot meant the LAST plugin to declare a page
		// silently replaced every earlier one — a second plugin contributing a
		// page erased the first from the sidebar with no error. One list, one
		// rule: first writer wins per page_id (a duplicate id would collide in
		// the DOM, and the plugin that loaded first is the better answer than
		// whichever happened to load last).
		var declared []*UIPage
		if p.UI.Page != nil {
			declared = append(declared, p.UI.Page)
		}
		declared = append(declared, p.UI.Pages...)
		for _, pg := range declared {
			if pg == nil || seenPage[pg.PageID] {
				continue
			}
			seenPage[pg.PageID] = true
			merged.Pages = append(merged.Pages, pg)
		}
		merged.Elements = append(merged.Elements, p.UI.Elements...)
	}
	// Stable order by declared Order then page id, so the sidebar does not
	// reshuffle when a plugin is reloaded.
	sort.SliceStable(merged.Pages, func(i, j int) bool {
		if merged.Pages[i].Order != merged.Pages[j].Order {
			return merged.Pages[i].Order < merged.Pages[j].Order
		}
		return merged.Pages[i].PageID < merged.Pages[j].PageID
	})
	ps.ui.Store(merged)
}

// DiskEntry is one .lua file in the plugin directory, whether or not it loaded.
// The management UI lists these rather than only the loaded set, so an operator
// can see (and fix) a plugin that failed to compile instead of finding it
// missing from the list.
type DiskEntry struct {
	Name     string `json:"name"`
	Path     string `json:"path"`
	Size     int64  `json:"size"`
	Loaded   bool   `json:"loaded"`
	Disabled bool   `json:"disabled"`
	Builtin  bool   `json:"builtin"`
	Error    string `json:"error,omitempty"`
	// Version/Description are read from the loaded plugin when available.
	Version     string `json:"version,omitempty"`
	Description string `json:"description,omitempty"`
	Hooks       int    `json:"hooks"`
	// Pages lists the UI page ids the plugin contributes. The desktop panel
	// shows it so an operator can tell at a glance that a plugin actually
	// renders something, without opening its source.
	Pages []string `json:"pages,omitempty"`
}

// pageNames returns the plugin's contributed UI page ids, de-duplicated.
// A plugin may declare `page` or `pages` or both, and the same id can appear
// twice; the loader's merge keeps the first, so this mirrors that order.
func pageNames(p *Plugin) []string {
	if p == nil || p.UI == nil {
		return nil
	}
	declared := make([]*UIPage, 0, len(p.UI.Pages)+1)
	if p.UI.Page != nil {
		declared = append(declared, p.UI.Page)
	}
	declared = append(declared, p.UI.Pages...)
	var out []string
	seen := map[string]bool{}
	for _, pg := range declared {
		if pg == nil || seen[pg.PageID] {
			continue
		}
		seen[pg.PageID] = true
		out = append(out, pg.PageID)
	}
	return out
}

// OnDisk lists every .lua file in the plugin directory with its load state.
func (ps *Plugins) OnDisk() []DiskEntry {
	out := []DiskEntry{}
	if ps.dir == "" {
		return out
	}
	entries, err := os.ReadDir(ps.dir)
	if err != nil {
		return out
	}
	for _, e := range entries {
		if e.IsDir() || !strings.HasSuffix(e.Name(), ".lua") {
			continue
		}
		name := strings.TrimSuffix(e.Name(), ".lua")
		info, _ := e.Info()
		de := DiskEntry{Name: name, Path: e.Name()}
		if info != nil {
			de.Size = info.Size()
		}
		ps.mu.RLock()
		if p := ps.findLocked(name); p != nil {
			de.Loaded = p.LoadError == ""
			de.Disabled = p.Disabled
			de.Builtin = p.Builtin
			de.Error = p.LoadError
			de.Version = p.Info.Version
			de.Description = p.Info.Description
			de.Hooks = len(p.Hooks)
			// A plugin may declare a single page or several; both forms end
			// up in Pages (see the merge at load time), so reporting that one
			// field covers either.
			de.Pages = pageNames(p)
		} else {
			// On disk but not in the running set: either it failed so badly
			// that LoadSource never produced a record, or the dir was written
			// after startup. Mark it by comparing with the bundled source.
			if code, err := os.ReadFile(filepath.Join(ps.dir, e.Name())); err == nil {
				if orig, err := ReadBundledPlugin(name); err == nil && orig == string(code) {
					de.Builtin = true
				}
			}
		}
		ps.mu.RUnlock()
		out = append(out, de)
	}
	sort.Slice(out, func(i, j int) bool {
		// builtins first, then alphabetical: the example plugin an operator
		// is most likely to want to read should not be buried under whatever
		// they installed most recently.
		if out[i].Builtin != out[j].Builtin {
			return out[i].Builtin
		}
		return out[i].Name < out[j].Name
	})
	return out
}

// Count returns how many plugins loaded (including ones with LoadError).
func (ps *Plugins) Count() int {
	ps.mu.RLock()
	defer ps.mu.RUnlock()
	return len(ps.plugins)
}

// List returns a JSON-friendly view of every loaded plugin, for the status API.
func (ps *Plugins) List() []map[string]interface{} {
	ps.mu.RLock()
	defer ps.mu.RUnlock()
	out := make([]map[string]interface{}, 0, len(ps.plugins))
	for _, p := range ps.plugins {
		stages := make([]string, 0, len(p.Hooks))
		for _, st := range AllStages {
			if _, ok := p.Hooks[st]; ok {
				stages = append(stages, string(st))
			}
		}
		row := map[string]interface{}{
			"name":        p.Info.Name,
			"version":     p.Info.Version,
			"description": p.Info.Description,
			"author":      p.Info.Author,
			"hooks":       stages,
			"loaded":      p.LoadError == "",
			"disabled":    p.Disabled,
			"builtin":     p.Builtin,
		}
		if p.LoadError != "" {
			row["error"] = p.LoadError
		}
		if p.UI != nil {
			ui := map[string]interface{}{}
			// Every contributed page, listed under `pages`, INCLUDING the single
			// `page` one. Two keys for the same thing looks redundant until you
			// try to render the count: the management panel asked for `pages`,
			// got an empty list for a plugin that plainly has a page, and showed
			// "contributes 0 pages" next to a working screen. `page` stays as the
			// backwards-compatible single-page accessor.
			ids := make([]string, 0, len(p.UI.Pages)+1)
			if p.UI.Page != nil && p.UI.Page.PageID != "" {
				ids = append(ids, p.UI.Page.PageID)
			}
			for _, pg := range p.UI.Pages {
				if pg != nil && pg.PageID != "" {
					ids = append(ids, pg.PageID)
				}
			}
			if len(ids) > 0 {
				ui["pages"] = ids
				if p.UI.Page != nil {
					ui["page"] = p.UI.Page.PageID
				}
			}
			if len(p.UI.Elements) > 0 {
				ui["elements"] = len(p.UI.Elements)
			}
			row["ui"] = ui
		}
		out = append(out, row)
	}
	return out
}

// HookErrors returns per-stage hook failure counts (empty when all is well).
func (ps *Plugins) HookErrors() map[string]map[string]interface{} {
	return ps.hookErr.snapshot()
}

// UI returns the merged UI extensions to inject into the WebUI.
// audienceAllows reports whether a role is permitted by an audience string.
//
// The grammar is deliberately small, because it is authored in two places (a
// .lua file and config.yaml) and a richer language would only give someone a way
// to express a rule nobody audits:
//
//	""  or "all"   everyone
//	"admin"        admin role only
//	"user"         user role only
//	"admin,user"   either
//
// "*" is deliberately NOT a synonym for "all": an audience that looks
// restrictive in a config file while meaning everything is exactly the kind of
// setting that fails silently.
// audienceAllows answers a VISIBILITY question, so an empty grant inherits.
func audienceAllows(want, role string) bool {
	return ParseAccessRule(want).AllowsVisibility(role)
}

// visibilityOverrides holds the operator's per-plugin audience overrides.
//
// A plugin declares its own default and the operator may force one. Both sides
// need to be able to: the plugin author does not know the deployment's policy,
// and the operator must not have to edit a bundled .lua — which the next
// upgrade would overwrite anyway.
type visibilityOverrides map[string]string

// seedShardConfig copies the shared plugin's `prices` table onto a shard VM.
//
// `prices` is CONFIGURATION, as opposed to `state` (accumulated history). A shard
// must not write state — its copy is thrown away — but it must be able to READ
// the operator's configuration, or a plugin that branches on its configuration
// behaves one way through Plugins.Call and another through ShardCall.
//
// Found by the OpenDesign adapter, which reads prices.control_key: without this, a
// fanout probe saw an empty table, concluded "not configured", and reported a
// skip for every host — while the identical call through the shared VM worked.
// The page uses the fanout path, so the feature would have been dead in
// production and alive in every test that called the shared VM.
//
// Called on every checkout (boot and reuse), so a configuration change reaches
// pooled VMs too. The push is a copy: shards get the same values, never a live
// reference to the shared table.
func (ps *Plugins) seedShardConfig(name string, w *worker) {
	if ps == nil || w == nil || w.L == nil {
		return
	}
	ps.mu.RLock()
	p := ps.find(name)
	ps.mu.RUnlock()
	if p == nil {
		return
	}
	p.mu.Lock()
	if p.state == nil || p.state.L == nil || p.state.L == w.L {
		p.mu.Unlock()
		return
	}
	prices := snapshotField(p.state.L, "prices")
	p.mu.Unlock()
	if prices == nil {
		return
	}
	L := w.L
	L.SetTop(0)
	defer L.SetTop(0)
	L.GetGlobal(pluginGlobal)
	if L.IsNil(-1) {
		return
	}
	plug := L.GetTop()
	pushGoValue(L, prices)
	L.SetField(plug, "prices")
}

// SetPluginVisibility records an operator override for a plugin's audience. An
// empty value clears the override and restores the plugin's own default.
func (ps *Plugins) SetPluginVisibility(plugin, audience string) {
	ps.mu.Lock()
	defer ps.mu.Unlock()
	if ps.visibility == nil {
		ps.visibility = visibilityOverrides{}
	}
	plugin = strings.TrimSpace(plugin)
	audience = strings.TrimSpace(audience)
	if audience == "" {
		delete(ps.visibility, plugin)
	} else {
		ps.visibility[plugin] = audience
	}
	ps.saveAccessOverridesLocked()
}

// VisibilityOverride returns the operator's audience override for a plugin,
// or "" when none is set. Exposed so /api/plugins can show which visibility is
// imposed rather than only the one in effect.
func (ps *Plugins) VisibilityOverride(plugin string) string {
	ps.mu.RLock()
	defer ps.mu.RUnlock()
	return ps.visibility[plugin]
}

// uiAudience resolves the audience in force for a plugin: the operator's
// override if set, else the plugin's declared default.
func (ps *Plugins) uiAudience(p *Plugin, declared string) string {
	ps.mu.RLock()
	override := ps.visibility[p.Info.Name]
	ps.mu.RUnlock()
	if override != "" {
		return override
	}
	return declared
}

// UIForRole returns the UI a given role may see.
//
// Filtering happens HERE, not in the HTTP layer, so every consumer sees the same
// view — the WebUI boot payload, the Electron shell and the tests. Filtering at
// one call site means the others quietly keep serving everything.
//
// A plugin's page/element can narrow visibility further than its own default
// (a public summary page beside an admin-only one), so the role check runs on
// the resolved plugin audience AND on each item's own declaration.
func (ps *Plugins) UIForRole(role string) *UIExtension {
	full := ps.UI()
	if full == nil {
		return &UIExtension{}
	}
	// An admin sees everything, including anything a plugin hides from itself
	// by mistake: anything else makes an override unverifiable from the browser.
	if role == "admin" {
		return full
	}
	out := &UIExtension{}
	// The plugin-level default is per plugin, so pages are filtered by walking
	// plugins rather than pages: a page id alone does not say who owns it.
	for _, p := range ps.plugins {
		if p.UI == nil || p.LoadError != "" {
			continue
		}
		aud := ps.uiAudience(p, p.UI.Visibility)
		if !audienceAllows(aud, role) {
			continue
		}
		for _, pg := range p.UI.Pages {
			if pg == nil || !audienceAllows(pg.Audience, role) {
				continue
			}
			out.Pages = append(out.Pages, pg)
		}
		for _, el := range p.UI.Elements {
			if !audienceAllows(el.Audience, role) {
				continue
			}
			out.Elements = append(out.Elements, el)
		}
		// Page (singular) is the other half of the same declaration and both
		// merge in UI(). Handling only Pages made every plugin that used the
		// single-page form disappear for a user key — silently, with no error,
		// because "this plugin contributes nothing to you" is indistinguishable
		// from "this plugin is not installed".
		if p.UI.Page != nil && audienceAllows(p.UI.Page.Audience, role) {
			if out.Page == nil {
				out.Page = p.UI.Page
			}
			if !containsPage(out.Pages, p.UI.Page.PageID) {
				out.Pages = append(out.Pages, p.UI.Page)
			}
		}
	}
	return out
}

func (ps *Plugins) UI() *UIExtension {
	if u := ps.ui.Load(); u != nil {
		return u
	}
	return &UIExtension{}
}

// State returns a plugin's own published state. A plugin publishes it by
// setting `plugin.state = {...}` inside its hook; that is the only way a hook's
// numbers reach the WebUI, because request_end is the LAST pipeline stage and
// has no downstream consumer to hand a return value to.
//
// Returns nil when the plugin does not exist or has not published anything.
func (ps *Plugins) State(name string) interface{} {
	ps.mu.RLock()
	p := ps.find(name)
	ps.mu.RUnlock()
	if p == nil {
		return nil
	}
	p.mu.Lock()
	w := p.state
	if w == nil {
		p.mu.Unlock()
		return nil
	}
	defer p.mu.Unlock()
	L := w.L
	L.SetTop(0)
	defer L.SetTop(0)
	L.GetGlobal(pluginGlobal)
	if L.IsNil(-1) {
		return nil
	}
	plug := L.GetTop()
	L.GetField(plug, "state")
	if L.Type(-1) == golua.LUA_TNIL {
		return nil
	}
	var out interface{}
	if err := luaToJSON(L, -1, &out); err != nil {
		return nil
	}
	return out
}

// Prices returns the plugin's current price table (the `prices` field, kept
// separate from `state`).
//
// Exposed because the billing rules editor has to show what is actually in
// effect. Reconstructing it from config alone would be wrong the moment a
// price was injected directly through SetState or the two drifted, and an
// editor that displays a different table from the one billing uses is worse
// than no editor at all.
func (ps *Plugins) Prices(name string) interface{} {
	ps.mu.RLock()
	p := ps.find(name)
	ps.mu.RUnlock()
	if p == nil {
		return nil
	}
	p.mu.Lock()
	defer p.mu.Unlock()
	if p.state == nil || p.state.L == nil {
		return nil
	}
	return snapshotField(p.state.L, "prices")
}

// SetState replaces a plugin's published state (admin API). It is how a
// configuration change (a new price for a model) reaches the plugin without
// reloading it.
//
// CONTRACT: a `prices` key in the payload is applied to the plugin's SEPARATE
// `prices` field and STRIPPED from `state`. That split is deliberate: prices are
// configuration while state is accumulated history, and a single replaceable
// field would make a price update wipe the totals (or make the totals carry a
// stale price table). A plugin that keeps its config elsewhere can ignore the
// convention and read the whole payload from `state` instead.
func (ps *Plugins) SetState(name string, state interface{}) error {
	ps.mu.RLock()
	p := ps.find(name)
	ps.mu.RUnlock()
	if p == nil {
		return fmt.Errorf("plugin %s not loaded", name)
	}
	p.mu.Lock()
	w := p.state
	if w == nil {
		p.mu.Unlock()
		return fmt.Errorf("plugin %s has no state", name)
	}
	defer p.mu.Unlock()
	L := w.L
	L.SetTop(0)
	defer L.SetTop(0)
	L.GetGlobal(pluginGlobal)
	if L.IsNil(-1) {
		return fmt.Errorf("plugin %s has no table", name)
	}
	plug := L.GetTop()

	// Pull `prices` out of the payload before storing the rest as state.
	body := state
	if m, ok := state.(map[string]interface{}); ok {
		if prices, has := m["prices"]; has {
			pushGoValue(L, prices)
			L.SetField(plug, "prices")
			rest := make(map[string]interface{}, len(m))
			for k, v := range m {
				if k != "prices" {
					rest[k] = v
				}
			}
			if len(rest) == 0 {
				// A prices-ONLY payload is a configuration change, not a state
				// reset. Leaving `state` untouched is what makes repricing safe:
				// replacing it with an empty table would silently erase every
				// accumulated total, so the next request would start from zero
				// and the dashboard would show a sudden drop in spend.
				//
				// The prices still have to be persisted here: returning before
				// the write below would leave the new price table in memory
				// only, and a restart would silently revert to the old prices
				// while the operator believed the change took effect.
				var curState, curPrices interface{}
				curState = snapshotField(L, "state")
				curPrices = snapshotField(L, "prices")
				if ps.dir != "" && !ps.stateDirDisabled && curState != nil {
					writeJSONAtomic(ps.stateFile(name), map[string]interface{}{
						"version": 1,
						"state":   curState,
						"prices":  curPrices,
					})
				}
				return nil
			}
			body = rest
		}
	}

	pushGoValue(L, body)
	L.SetField(plug, "state")
	L.SetTop(0)
	// A state replacement is exactly the kind of thing an operator restarts the
	// gateway for, so it must survive the restart. The snapshot is taken from
	// the values just written — the caller already holds p.mu, so calling
	// persistNow here would deadlock on that same non-reentrant mutex.
	writeJSONAtomic(ps.stateFile(name), map[string]interface{}{
		"version": 1,
		"state":   body,
		"prices":  readPricesFromPayload(state),
	})
	return nil
}

// readPricesFromPayload recovers the prices table a caller passed to SetState,
// which SetState stores on the plugin (not in state). Used only for the
// persistence record, so that a restart restores configuration and history to
// the two fields they belong in rather than collapsing them.
func readPricesFromPayload(state interface{}) interface{} {
	if m, ok := state.(map[string]interface{}); ok {
		return m["prices"]
	}
	return nil
}

// markDirty snapshots a plugin's state and hands it to the saver.
//
// It is called from the hook path, so it must not block on file I/O — that is
// the saver's job. It DOES walk the Lua tables, because that has to happen
// while the caller still holds p.mu and the VM is guaranteed alive; the flush
// (Snapshotting in the flush goroutine instead was a use-after-free: vm.Stop()
// frees the Lua states. See persistByName for why the walk is safe here.)
//
// Cost is one JSON conversion per request, which the billing plugin would pay
// anyway inside its own hook.
func (ps *Plugins) markDirtyLocked(p *Plugin) {
	if ps == nil || p == nil || ps.stateDirDisabled || p.state == nil || ps.dir == "" {
		return
	}
	// NOTE: this only records THAT the state changed. It deliberately does not
	// snapshot it.
	//
	// Snapshotting here cost 264us per request — the snapshot walks the plugin's
	// Lua tables through luaValueToGo + json.Marshal + json.Unmarshal (three
	// conversions), and the billing plugin does it on every single request. That
	// turned a 14.6us hook into a 385us one and put the gateway's hot path
	// behind a bookkeeping step. The saver pulls the snapshot on its own schedule
	// instead, so N requests between flushes cost ONE snapshot.
	ps.save.markDirty(p.Info.Name)
}

// persistByName snapshots one plugin and writes its state file.
//
// It is called from the saver goroutine, which means it walks a Lua state that
// vm.Stop() may be about to free. That was a real crash: the first version had
// the saver read the tables itself and the process died with SIGSEGV inside
// golua once shutdown raced a flush.
//
// Two things make it safe now:
//   - Plugins.Close() stops the saver AND waits for its final flush BEFORE the
//     caller stops the VM (see Plugins.Close), so no flush can be in flight when
//     the Lua states go away.
//   - p.mu is held while the tables are read, so a hook cannot be mutating them
//     underneath. A hook that arrives after the VM stopped cannot run either,
//     because the request path is already torn down by then.
func (ps *Plugins) persistByName(name string) {
	if ps == nil || ps.stateDirDisabled || ps.dir == "" {
		return
	}
	ps.mu.RLock()
	p := ps.find(name)
	ps.mu.RUnlock()
	if p == nil || p.state == nil {
		return
	}
	ps.persistNow(p)
}

// ---------- state persistence ----------
//
// A plugin's state lives in the Lua VM, which dies with the process. For the
// billing plugin that means a restart silently zeroes every total — measured on
// this gateway: 218,241 requests and 26.2M prompt tokens gone after one
// `systemctl restart`. For a REPORTING plugin whose whole purpose is the number
// it accumulates, that is not a rounding error, it is the feature not working.
//
// Three rules shape this:
//
//  1. `prices` (configuration) and `state` (accumulated history) are persisted
//     to SEPARATE files. Restoring them together would let a price edit look
//     like a state reset, or a state restore resurrect stale prices — SetState
//     already keeps them apart in memory, and disk has to agree.
//  2. Writes are atomic (temp file + rename). A crash mid-write must leave the
//     previous state readable, not a truncated JSON file that fails to parse on
//     the next boot and loses the total anyway.
//  3. A corrupt or unreadable state file is a WARNING, never a startup error.
//     Forwarding must not depend on a plugin's bookkeeping surviving.

// stateFile is the per-plugin state path, kept beside the plugin source so an
// operator can find (and delete) it next to the plugin it belongs to.
func (ps *Plugins) stateFile(name string) string {
	return filepath.Join(ps.dir, "."+name+".state.json")
}

// persistNow writes a plugin's state and prices to disk synchronously. Callers
// on the hot path must use ps.markDirtyLocked instead; this is the flush worker and
// the admin-state path, where durability matters more than latency.
//
// It is called after every
// state mutation, so it must be cheap enough not to matter: the billing plugin
// mutates on every request, and writing the whole state per request would put a
// file write on the hot path.
func (ps *Plugins) persistNow(p *Plugin) {
	if ps == nil || p == nil || ps.stateDirDisabled || p.state == nil || ps.dir == "" {
		return
	}
	p.mu.Lock()
	state, prices := readStateAndPrices(p.state.L)
	p.mu.Unlock()
	if state == nil {
		return
	}
	writeJSONAtomic(ps.stateFile(p.Info.Name), map[string]interface{}{
		"version": 1,
		"state":   state,
		"prices":  prices,
	})
}

// readStateAndPrices pulls both tables out of the Lua state. The caller holds
// p.mu. Returns nils when the plugin table is missing, which happens for a
// plugin that failed to compile.
//
// Each field is read by REBUILDING the stack from the global, because the
// conversion helper (luaToJSON) ends with L.SetTop(0) on every path — it treats
// the whole stack as its own. Reusing a saved index across two conversions
// addresses a slot that no longer exists, and this binding aborts the process
// (SIGABRT) instead of reporting a bad index. The first version did exactly
// that and crashed inside the Lua C layer on every hook call.
func readStateAndPrices(L *golua.State) (state, prices interface{}) {
	state = snapshotField(L, "state")
	prices = snapshotField(L, "prices")
	L.SetTop(0)
	return state, prices
}

// snapshotField converts plugin.<field> into Go values, leaving the stack clean.
// Caller holds p.mu; caller is responsible for the Lua state being alive.
func snapshotField(L *golua.State, field string) interface{} {
	L.SetTop(0)
	defer L.SetTop(0)
	L.GetGlobal(pluginGlobal)
	if L.IsNil(-1) {
		return nil
	}
	L.GetField(L.GetTop(), field)
	if L.Type(-1) != golua.LUA_TTABLE {
		return nil
	}
	var out interface{}
	if err := luaToJSON(L, -1, &out); err != nil {
		return nil
	}
	return out
}

// writeJSONAtomic writes v to path via a temp file + rename, so a reader (or a
// crash) never observes a half-written file.
func writeJSONAtomic(path string, v interface{}) {
	b, err := json.MarshalIndent(v, "", "  ")
	if err != nil {
		return
	}
	dir := filepath.Dir(path)
	if err := os.MkdirAll(dir, 0755); err != nil {
		return
	}
	tmp, err := os.CreateTemp(dir, filepath.Base(path)+".tmp*")
	if err != nil {
		return
	}
	tmpName := tmp.Name()
	if _, err := tmp.Write(b); err != nil {
		tmp.Close()
		os.Remove(tmpName)
		return
	}
	if err := tmp.Close(); err != nil {
		os.Remove(tmpName)
		return
	}
	if err := os.Rename(tmpName, path); err != nil {
		os.Remove(tmpName)
	}
}

// restore re-applies a persisted state (and prices) onto a freshly compiled
// plugin. Called once per plugin after it compiles and registers its hooks.
//
// Ordering matters: this runs AFTER compilation so the plugin's own defaults
// exist, and it OVERWRITES them, so a restart continues from the saved totals
// rather than from the values baked into the .lua source.
func (ps *Plugins) restore(p *Plugin) {
	if ps == nil || p == nil || p.state == nil || ps.stateDirDisabled {
		return
	}
	b, err := os.ReadFile(ps.stateFile(p.Info.Name))
	if err != nil {
		return // no saved state yet: the compiled-in default stands
	}
	var saved struct {
		Version int         `json:"version"`
		State   interface{} `json:"state"`
		Prices  interface{} `json:"prices"`
	}
	if err := json.Unmarshal(b, &saved); err != nil {
		// A corrupt file must not stop the gateway: the plugin keeps its
		// compiled-in defaults and the operator sees a warning in the log.
		ps.logf("plugin %s: ignoring unreadable state file %s: %v",
			p.Info.Name, ps.stateFile(p.Info.Name), err)
		return
	}
	p.mu.Lock()
	defer p.mu.Unlock()
	L := p.state.L
	L.SetTop(0)
	defer L.SetTop(0)
	L.GetGlobal(pluginGlobal)
	if L.IsNil(-1) {
		return
	}
	plug := L.GetTop()
	if saved.Prices != nil {
		pushGoValue(L, saved.Prices)
		L.SetField(plug, "prices")
	}
	if saved.State != nil {
		pushGoValue(L, saved.State)
		L.SetField(plug, "state")
	}
	p.persisted = true
}

// SetEnabled turns a plugin's dispatch on or off without touching its file.
//
// The state is on the Plugin record (not derived from disk) so a disable survives
// as long as the process lives and is trivially re-enabled; it deliberately does
// NOT persist across restarts, because a "disable" that silently outlives the
// operator's intent is its own surprise. An operator who wants it permanent
// moves the file out of the plugin dir.
func (ps *Plugins) SetEnabled(name string, enabled bool) error {
	ps.mu.Lock()
	p := ps.findLocked(name)
	ps.mu.Unlock()
	if p == nil {
		return fmt.Errorf("plugin %s not loaded", name)
	}
	if p.LoadError != "" {
		return fmt.Errorf("plugin %s failed to load (%s); fix the file before enabling it", name, p.LoadError)
	}
	p.mu.Lock()
	p.Disabled = !enabled
	p.mu.Unlock()
	ps.rebuild()
	return nil
}

// Enabled reports whether a plugin is currently dispatching.
func (ps *Plugins) Enabled(name string) bool {
	ps.mu.RLock()
	p := ps.findLocked(name)
	ps.mu.RUnlock()
	return p != nil && p.LoadError == "" && !p.Disabled
}

// Unload removes a plugin from the running set. Its states are closed so the
// memory goes back; a subsequent LoadSource with the same name works again.
func (ps *Plugins) Unload(name string) error {
	ps.mu.Lock()
	idx := -1
	for i, p := range ps.plugins {
		if p.Info.Name == name || strings.HasSuffix(filepath.Base(p.dir), name+".lua") {
			idx = i
			break
		}
	}
	if idx < 0 {
		ps.mu.Unlock()
		return fmt.Errorf("plugin %s not loaded", name)
	}
	p := ps.plugins[idx]
	ps.plugins = append(ps.plugins[:idx], ps.plugins[idx+1:]...)
	ps.mu.Unlock()
	// Stop the scheduler BEFORE touching the state. Stop() closes the loop and
	// waits for it, so after it returns no NEW task can be dispatched — the set
	// of in-flight tasks is now frozen, which is what makes the check below
	// meaningful rather than racy.
	inFlight := false
	if p.Runtime != nil {
		p.Runtime.Stop()
		inFlight = p.Runtime.InFlight()
		p.Runtime = nil
	}
	// A task holds p.mu for the whole duration of its Lua call. If one is still
	// in flight we must NOT take p.mu (a wedged task never releases it, so Lock
	// would hang the unload forever) and must NOT close the state underneath it
	// (that is a use-after-free in the Lua C layer, which aborts the process
	// instead of returning an error). Leaking this one state is the bounded,
	// non-crashing outcome; the operator's action is to remove the file and
	// restart, which is what a plugin wedged by its own task requires anyway.
	if inFlight {
		log.Printf("[scheduler] plugin %s: unloaded while a task was still running; "+
			"its Lua state is left open to avoid a use-after-free", name)
		ps.rebuild()
		return nil
	}
	// Lifecycle hook: if the plugin exported on_unload, call it BEFORE closing
	// the state. This is generic plugin lifecycle (part of "调度插件"), not any
	// feature: a plugin that spawned OS processes or holds other external
	// resources is the one that knows how to release them, and Unload is the
	// only guaranteed point where it still can. Best-effort by design — a
	// hook that errors must not block an unload the operator explicitly asked
	// for (a wedged plugin has its own path above).
	p.mu.Lock()
	if p.state != nil && p.state.L != nil {
		p.state.L.GetGlobal(pluginGlobal)
		if L := p.state.L; L.Type(-1) == golua.LUA_TTABLE {
			L.GetField(-1, "on_unload")
			if L.Type(-1) == golua.LUA_TFUNCTION {
				if err := callFnOn(L, "on_unload"); err != nil {
					log.Printf("[plugins] %s: on_unload: %v", name, err)
				}
			}
			L.Pop(2) // adapter/plugin table + field remnant
		} else {
			L.Pop(1)
		}
		p.state.L.Close()
		p.state = nil
	}
	p.mu.Unlock()
	ps.rebuild()
	return nil
}

// find returns a loaded plugin by declared name, taking the read lock.
func (ps *Plugins) find(name string) *Plugin {
	ps.mu.RLock()
	defer ps.mu.RUnlock()
	return ps.findLocked(name)
}

// findLocked returns a loaded plugin by declared name. The caller holds ps.mu.
func (ps *Plugins) findLocked(name string) *Plugin {
	for _, p := range ps.plugins {
		if p.Info.Name == name {
			return p
		}
	}
	return nil
}

// Fire calls every plugin registered for a stage, in plugin load order.
//
// A plugin may MUTATE the payload by returning a JSON object: any keys it
// returns are merged into the payload for the next hook and returned to the
// caller. Returning nil or an empty table means "no opinion". This lets a
// plugin add fields (the billing plugin adds `cost_usd`) without the gateway
// having to know about them.
//
// Errors are contained: a plugin that throws is logged against the stage and
// skipped. Forwarding never depends on plugin health.
func (ps *Plugins) Fire(stage Stage, payload map[string]interface{}) map[string]interface{} {
	ps.mu.RLock()
	calls := ps.stageFuncs[stage]
	ps.mu.RUnlock()
	if len(calls) == 0 {
		return payload
	}
	// Resolve the plugins and drop the lock BEFORE running any hook.
	//
	// A hook can install, disable or remove a plugin (the admin API is reachable
	// from a hook that has the key), and those paths take ps.mu for write. Taking
	// a per-iteration RLock like the single-plugin case did works but serializes
	// on a shared cache line under load; a snapshot plus one lookup is cheaper and
	// — more importantly — it means the parallel section below never holds ps.mu,
	// so a hook cannot deadlock against a concurrent reload.
	type target struct {
		p  *Plugin
		hc hookCall
	}
	targets := make([]target, 0, len(calls))
	for _, hc := range calls {
		ps.mu.RLock()
		p := ps.plugins[hc.pluginIdx]
		ps.mu.RUnlock()
		if p == nil || p.LoadError != "" {
			continue
		}
		// A plugin disarmed by an overrunning task still holds its own lock (the
		// task never returned), so dispatching to it would block THIS request
		// behind a lock that is never released. Skip it and record why: the
		// operator needs to see that inference traffic is being dropped and which
		// plugin is responsible, not watch requests hang.
		if p.Runtime != nil && p.Runtime.Disarmed() {
			ps.hookErr.note(stage, p.Info.Name+": "+errPluginWedged.Error())
			continue
		}
		targets = append(targets, target{p: p, hc: hc})
	}
	if len(targets) == 0 {
		return payload
	}

	// One plugin is the overwhelmingly common case (a gateway with the bundled
	// billing plugin has exactly one). Paying for goroutines there would be pure
	// overhead, so it takes the direct path.
	if len(targets) == 1 {
		if out := ps.runHook(stage, targets[0].p, targets[0].hc.fn, payload); len(out) > 0 {
			for k, v := range out {
				payload[k] = v
			}
		}
		return payload
	}

	// Parallel across plugins. Each hook receives the SAME payload snapshot, and
	// the returned tables are merged afterwards IN PLUGIN LOAD ORDER, so the
	// documented contract ("a returned table's keys are merged into the payload")
	// still holds deterministically.
	//
	// What changes: a hook no longer sees the keys another hook just added. The
	// sequential behaviour made that possible, and docs/plugins.md described it
	// ("the payload is passed to the next plugin unchanged"). It was never used —
	// the bundled billing plugin returns nil on every stage, with a comment saying
	// nobody downstream reads it — but it IS a contract change and is called out
	// there rather than left as a surprise.
	//
	// Why this is safe for the per-plugin lock: each plugin has its own Lua state
	// and its own p.mu, so two plugins never touch the same state. What is shared
	// is the payload, and it is only READ here — merging happens after every hook
	// has returned, on the caller's goroutine.
	outs := make([]map[string]interface{}, len(targets))
	var wg sync.WaitGroup
	for i := range targets {
		wg.Add(1)
		go func(i int) {
			defer wg.Done()
			// A hook that panics would otherwise take the whole process with it,
			// and in the sequential version a panic could not escape Fire either.
			// recover() here restores that: the plugin is skipped and noted.
			defer func() {
				if rec := recover(); rec != nil {
					ps.hookErr.note(stage, fmt.Sprintf("%s: panic in hook: %v", targets[i].hc.plugin, rec))
				}
			}()
			outs[i] = ps.runHook(stage, targets[i].p, targets[i].hc.fn, payload)
		}(i)
	}
	wg.Wait()

	// Merge in load order so the result does not depend on goroutine scheduling.
	// A later plugin's value wins on a key collision, exactly as it did when the
	// hooks ran one after another.
	for _, out := range outs {
		for k, v := range out {
			payload[k] = v
		}
	}
	return payload
}

// runHook invokes one hook and folds its returned table into payload. It
// contains the plugin's error handling so both the sequential and the parallel
// path behave identically on failure.
func (ps *Plugins) runHook(stage Stage, p *Plugin, fn string, payload map[string]interface{}) map[string]interface{} {
	out, err := ps.invoke(p, fn, payload)
	if err != nil {
		ps.hookErr.note(stage, p.Info.Name+": "+err.Error())
		return nil
	}
	return out
}

// invoke runs one plugin hook on that plugin's own state, under its pool's
// concurrency cap. The plugin's returned table is re-fetched each call because
// the pool is elastic: a plugin may have several states, and the hook function
// lives in each.
func (ps *Plugins) invoke(p *Plugin, fn string, payload map[string]interface{}) (map[string]interface{}, error) {
	p.mu.Lock()
	w := p.state
	if w == nil {
		p.mu.Unlock()
		return nil, fmt.Errorf("no state for plugin %s", p.Info.Name)
	}
	defer p.mu.Unlock()
	L := w.L
	L.SetTop(0)
	defer L.SetTop(0)

	L.GetGlobal(pluginGlobal)
	if L.IsNil(-1) {
		return nil, fmt.Errorf("plugin table missing")
	}
	plug := L.GetTop() // absolute, so nothing below shifts
	// The hook is called THROUGH a pcall guard (see hookGuardSrc) so a script
	// error returns as values instead of raising into golua, whose error path
	// SIGSEGVs the process. Desired stack before the payload push:
	// [pluginGlobal, guard, hookfn]. Build it in that order — GetGlobal(guard)
	// then GetField(hookfn) lands the function exactly above the guard, with no
	// Remove/Insert juggling. (The first version reordered with Remove/Insert
	// and ended up calling pluginGlobal as if it were the hook, producing
	// "attempt to call a table value" and silently zeroed every total.)
	L.GetGlobal(hookGuardName)
	if L.IsNil(-1) {
		// Guard absent (a state built before this existed): drop the nil so the
		// stack is [pluginGlobal, hookfn] and the hook is called directly.
		// Correct plugins still work; only their errors stop being survivable.
		L.Pop(1)
	}
	L.GetField(plug, fn)
	if !L.IsFunction(-1) {
		L.SetTop(0)
		return nil, fmt.Errorf("hook %s missing", fn)
	}
	// The hook is called with a DECODED table, not the raw JSON string: the
	// adapter protocol passes JSON text to its transforms (they decode it
	// themselves), but a plugin hook receives a table so it can read
	// payload.model directly. Passing the string made every hook fail with
	// "attempt to index local 'payload' (a string value)".
	//
	// Decoding is done WITHOUT JSON. json.Marshal + json.Unmarshal here cost
	// 9.6us of the 14.6us hook — 67% of the call spent re-deriving a tree that
	// pushGoValue walks natively anyway. plainForLua does the same conversion
	// by type-switching, and falls back to JSON only for a type it does not
	// model, so an exotic payload still arrives instead of vanishing.
	pushGoValue(L, plainForLua(payload))
	if err := L.Call(2, 2); err != nil {
		return nil, err
	}
	// The guard returns TWO values: the hook's value, then an error string
	// (nil on success). Drop the error slot unconditionally so the value is
	// left on top. Peeking instead of popping was the first bug here: on a
	// SUCCESSFUL call the top is nil, so "the error is non-nil" is false, the
	// pop was skipped, and the code then read the (always-present) nil error as
	// if it were the hook's answer. Every plugin's returned table silently
	// became "no opinion" while the tests that only checked totals kept
	// passing.
	errMsg := ""
	if L.GetTop() >= 2 {
		if !L.IsNil(-1) {
			errMsg = L.ToString(-1)
		}
		L.Pop(1) // the error slot, nil or not
	}
	if errMsg != "" {
		// A hook that raised partway may already have mutated plugin.state, so
		// mark it dirty before bailing: the accounting it managed to do is
		// still real and should be persisted.
		ps.markDirtyLocked(p)
		return nil, fmt.Errorf("plugin %s: %s", p.Info.Name, errMsg)
	}
	// Read the return value FIRST, then snapshot state for persistence.
	//
	// The order is load-bearing. markDirtyLocked walks the Lua tables and resets
	// the stack, so calling it before the return value is read destroyed the
	// hook's answer — every plugin that returned a table silently became a
	// plugin that "had no opinion". Reading first costs nothing and keeps the
	// documented merge contract intact.
	var out map[string]interface{}
	if L.GetTop() < 1 || L.IsNil(-1) {
		ps.markDirtyLocked(p)
		return nil, nil
	}
	if err := luaToJSON(L, -1, &out); err != nil {
		ps.markDirtyLocked(p)
		return nil, nil // not a table: treat as "no opinion"
	}
	// A hook that ran at all may have mutated plugin.state, whether or not it
	// returned anything. Marking here (not only on a returned table) is what
	// makes an accumulating plugin like billing durable: its totals change on
	// every call and it returns nil every time.
	ps.markDirtyLocked(p)
	return out, nil
}

// Call invokes a plugin function that the plugin declared in plugin.exports.
//
// The gate is the declaration, not the caller: a plugin lists the handful of
// entry points its own UI is allowed to reach, and anything absent is refused.
// Without that list this endpoint would be a remote Lua evaluator — it could
// call the plugin's private helpers, hand them crafted arguments, and read
// whatever the plugin can see, which for a plugin holding an API key means
// handing that key to anyone who can POST to the gateway.
//
// The return value is JSON-shaped rather than restricted to a map, because a
// quota report is a nested structure with numbers and booleans that must not be
// flattened. Hook results still go through invoke, whose map-typed contract the
// state merge depends on.
// Call invokes an exported function on a plugin's SHARED VM, serialised by the
// plugin's mutex.
//
// The lock is what makes one Lua VM safe to share, and it is held for the whole
// call — including any network wait inside it. That is correct for a hook and
// wrong for a function that blocks: twenty concurrent quota probes serialise
// into the sum of their timeouts (measured: 29s on a 20-source gateway, against
// 1.4–3.3s for a single probe with no contention).
//
// callOn below is the same work without the shared-VM assumption, so a caller
// that owns a private VM can use it directly.
func (ps *Plugins) Call(name, fn string, args map[string]interface{}) (interface{}, error) {
	ps.mu.RLock()
	var p *Plugin
	for _, c := range ps.plugins {
		if c.Info.Name == name {
			p = c
			break
		}
	}
	ps.mu.RUnlock()
	if p == nil {
		return nil, fmt.Errorf("plugin %s not found", name)
	}
	if !pluginExportAllowed(p, fn) {
		return nil, fmt.Errorf("plugin %s does not export %q", name, fn)
	}
	p.mu.Lock()
	defer p.mu.Unlock()
	w := p.state
	if w == nil {
		return nil, fmt.Errorf("no state for plugin %s", name)
	}
	out, err := callOn(w, fn, args)
	// An on-demand call mutates state just as a hook does — quota writes its
	// host→adapter memo here, and that is the whole point of the function — but
	// markDirty used to fire only from hooks and SetState. quota declares no
	// hooks, so every probe it ran through this path lived in memory and the memo
	// file was only rewritten when something ELSE dirtied the plugin. The
	// staleness is what made the memo look inert from the outside: a negative
	// entry appeared only after the next configuration change happened to flush.
	//
	// Marking here is the same call the hook path already makes, so the cost is the
	// same (the saver snapshots on its own schedule, not per call).
	ps.markDirtyLocked(p)
	return out, err
}

// callOn runs one exported function on a specific VM.
//
// It does NOT check the export allowlist and does NOT persist state: both belong
// to the shared-VM path in Call. A shard VM has no plugin identity to check
// against and no state file of its own — the export allowlist in Call is the
// security boundary, and calling this with a function the plugin does not export
// is a programming error, not an attack.
//
// The caller must own w exclusively. No lock is taken here.
func callOn(w *worker, fn string, args map[string]interface{}) (interface{}, error) {
	if w == nil || w.L == nil {
		return nil, fmt.Errorf("no lua state")
	}
	L := w.L
	L.SetTop(0)
	defer L.SetTop(0)

	L.GetGlobal(pluginGlobal)
	if L.IsNil(-1) {
		return nil, fmt.Errorf("plugin table missing")
	}
	plug := L.GetTop()
	L.GetGlobal(hookGuardName)
	if L.IsNil(-1) {
		L.Pop(1)
	}
	L.GetField(plug, fn)
	if !L.IsFunction(-1) {
		return nil, fmt.Errorf("export %s missing", fn)
	}
	if args == nil {
		args = map[string]interface{}{}
	}
	pushGoValue(L, plainForLua(args))
	if err := L.Call(2, 2); err != nil {
		return nil, err
	}
	errMsg := ""
	if L.GetTop() >= 2 {
		if !L.IsNil(-1) {
			errMsg = L.ToString(-1)
		}
		L.Pop(1)
	}
	if errMsg != "" {
		return nil, fmt.Errorf("plugin: %s", errMsg)
	}
	if L.GetTop() < 1 || L.IsNil(-1) {
		return nil, nil
	}
	var out interface{}
	if err := luaToJSON(L, -1, &out); err != nil {
		return nil, err
	}
	return out, nil
}

// pluginExportAllowed reports whether fn is in the plugin's exports list.
// An absent list means nothing is callable — the safe reading, since silently
// allowing all would turn an undeclared plugin into a remote evaluator.
func pluginExportAllowed(p *Plugin, fn string) bool {
	for _, e := range p.exports {
		if e == fn {
			return true
		}
	}
	return false
}

// StateVisibleTo reports whether a role may read a plugin's published state.
//
// The audience is resolved exactly as for the UI: operator override first, then
// the plugin's declaration. A plugin with neither is readable by any
// authenticated role, which is the pre-existing behaviour and the reason a
// newly installed widget keeps working without an edit.

// HTTPHandler is one endpoint a plugin registers on the gateway's router.
//
// A plugin that only receives hooks (billing) has nothing to expose. A plugin
// that computes something the operator may want to query from OUTSIDE the
// browser — a health probe, a CI check, another tool — needs a route, and
// hard-coding one in the gateway would mean every such plugin needs a gateway
// change. So the plugin declares the path, the method and the role it requires,
// and the gateway mounts it.
//
// The handler is a Lua function taking (request) and returning (response). Both
// are plain tables: request carries method/path/query/headers/body, response
// carries status/headers/body/json. Nothing here can reach the Go side except
// through those two values, so the surface a plugin can express is bounded by
// what a table can say.
type HTTPHandler struct {
	// Path is the route suffix under /api/plugins/http/. It may contain ":name"
	// segments, captured into request.params.
	Path string `json:"path"`
	// Method is the HTTP method; "" or GET. Only GET/POST/PUT/DELETE/PATCH are
	// accepted — an arbitrary verb would need a switch in the router for no gain.
	Method string `json:"method,omitempty"`
	// Roles are the API-key roles allowed to call it. EMPTY MEANS EVERYONE, which
	// is the wrong default for an endpoint a plugin invented: an unauthenticated
	// internet scan should not be able to ask it questions. RequireRole exists
	// so a plugin that really wants public access has to say so.
	Roles []string `json:"roles,omitempty"`
	// Keys names SPECIFIC gateway keys allowed to call it, by their configured
	// name (the `name:` on a gateway_keys entry), never by the secret.
	//
	// Names rather than secrets for three reasons: a secret is rotated without
	// notice and a rule written against it would silently stop matching; putting
	// secrets in a plugin file scatters credentials across the install; and the
	// operator reads names. A key with no name cannot be named here.
	//
	// When Keys is non-empty it is the WHOLE rule, not an addition to Roles. An
	// OR would mean an operator writing a two-key allowlist had also granted
	// every user key, which is the opposite of what they meant.
	Keys []string `json:"keys,omitempty"`
	// RequireAuth defaults to true in practice: the gateway refuses a handler
	// with neither Roles nor an explicit opt-out.
	Public bool `json:"public,omitempty"`
	// Handler is the name of the plugin function to invoke.
	Handler string `json:"handler"`
	// Description is shown in /api/plugins so an operator can see what a plugin
	// has exposed without reading the .lua.
	Description string `json:"description,omitempty"`
	// Plugin is filled in by HTTPHandlers: which plugin registered this route.
	// The gateway dispatcher needs it to know which Lua state to enter, and
	// deriving it from the path would mean trusting the plugin's own prefix.
	Plugin string `json:"-"`
}

// routePath renders the handler's full mount path.
func (h HTTPHandler) routePath() string {
	return "/api/plugins/http/" + strings.Trim(h.Path, "/")
}

// allows reports whether a caller may invoke this handler.
//
// keyName is the caller key's configured name ("" when it has none), not the
// secret.
// allows answers a CAPABILITY question, so an empty grant denies. See
// AccessRule.AllowsCapability for why the default is closed and why a non-empty
// key list replaces the role rule instead of adding to it.
func (h HTTPHandler) allows(role, keyName string) bool {
	return accessRuleFor(h.Roles, h.Keys, h.Public).AllowsCapability(role, keyName)
}

func (ps *Plugins) StateVisibleTo(name, role string) bool {
	if role == "admin" {
		return true
	}
	ps.mu.RLock()
	defer ps.mu.RUnlock()
	for _, p := range ps.plugins {
		if p.Info.Name != name {
			continue
		}
		override := ps.visibility[name]
		if override != "" {
			return audienceAllows(override, role)
		}
		declared := ""
		if p.UI != nil {
			declared = p.UI.Visibility
		}
		return audienceAllows(declared, role)
	}
	// An unknown plugin name: no visibility can be established, and this endpoint
	// answers with an empty state rather than an error, so refusing here would
	// change behaviour for a name that may simply not be installed.
	return true
}

func containsPage(pages []*UIPage, id string) bool {
	for _, p := range pages {
		if p != nil && p.PageID == id {
			return true
		}
	}
	return false
}

// readPluginHTTP collects plugin.http, the endpoints the plugin asks the gateway
// to mount.
//
// A malformed entry is DROPPED rather than failing the plugin: a plugin with one
// bad route should still load its hooks and pages, and the bad route simply does
// not exist. Failing the whole plugin would take a working billing widget down
// over a typo in an experimental endpoint.
func readPluginHTTP(L *golua.State) []HTTPHandler {
	// `http` is a LIST of declarations, each naming the handler function by
	// STRING:
	//
	//	http = {
	//	  { path = "health", method = "GET", roles = { "admin" }, handler = "health" },
	//	}
	//
	// handler is a name, not a value. An inline function silently produced an
	// empty handler field and the endpoint was dropped without a word.
	raw, ok := pluginFieldValue(L, "http")
	if !ok {
		return nil
	}
	// A single declaration may be given as a bare table instead of a one-element
	// list; accepting both is friendlier than demanding one spelling.
	items, isList := raw.([]interface{})
	if !isList {
		if one, isTable := raw.(map[string]interface{}); isTable {
			items = []interface{}{one}
		} else {
			return nil
		}
	}
	ctx := &decodeCtx{plugin: "http"}
	var list []HTTPHandler
	for _, it := range items {
		h, ok := it.(map[string]interface{})
		if !ok {
			continue
		}
		ctx.strictKeys(h, "http[]", "path", "method", "handler", "description",
			"roles", "keys", "public")
		dh := HTTPHandler{
			Path:        decString(h, "path"),
			Method:      decString(h, "method"),
			Description: decString(h, "description"),
			Roles:       decStringList(h, "roles"),
			Keys:        decStringList(h, "keys"),
		}
		// handler must be a NAME, not a value. asString would happily turn a
		// function into its address ("function: 0x..."), which then failed much
		// later at dispatch with no connection to the declaration that caused it.
		// Rejecting it here names the field and the path.
		if hv, present := h["handler"]; present {
			name, isStr := hv.(string)
			if !isStr {
				ctx.warnf("http[%s]: handler must be the NAME of a function on the "+
					"plugin table, not the function itself", decString(h, "path"))
				continue
			}
			dh.Handler = name
		}
		if p, ok := h["public"].(bool); ok {
			dh.Public = p
		}
		// A malformed entry is dropped rather than fatal: one bad route must not
		// cost a plugin its hooks and pages. The reason is named so a plugin whose
		// endpoint silently vanished can find out why.
		method, mErr := normalizeHandlerMethod(dh.Method)
		switch {
		case mErr != nil:
			ctx.warnf("http[%s]: %v", dh.Path, mErr)
			continue
		case validateHandlerPath(dh.Path) != nil:
			ctx.warnf("http[%s]: invalid path", dh.Path)
			continue
		case dh.Handler == "":
			ctx.warnf("http[%s]: handler must name a function on the plugin table "+
				"(an inline function is not supported)", dh.Path)
			continue
		}
		dh.Method = method
		list = append(list, dh)
	}
	if len(ctx.problems) > 0 {
		log.Printf("[plugins] %s", strings.Join(ctx.problems, "; "))
	}
	return list
}

// uipageFromValue builds a UIPage from a decoded Lua table.
func uipageFromValue(m map[string]interface{}) *UIPage {
	pg := &UIPage{
		PageID:   asString(m["page_id"]),
		Title:    asString(m["title"]),
		Icon:     asString(m["icon"]),
		Mount:    asString(m["mount"]),
		Audience: asString(m["audience"]),
	}
	if n, ok := asFloat(m["order"]); ok {
		pg.Order = int(n)
	}
	return pg
}