ModelRouter/internal/lua/vm.go

// Package lua implements the adapter runtime: bundles/loads *.lua adapter
// scripts, exposes json/string helpers, and lets Go call the protocol
// transform functions plus a build_headers signature hook.
//
// The runtime uses LuaJIT (github.com/aarzilli/golua) so every adapter runs on
// an independent interpreter state. Because a lua.State is NOT goroutine-safe,
// each adapter owns a pool of workers (states) and a worker is checked out for
// the duration of a single hook call.
//
// Pools are ELASTIC: the sum of max_concurrent over the sources using an adapter
// is a ceiling, not a preallocation. States are booted on demand in batches
// sized by that ceiling, and reclaimed by a single background janitor at a rate
// scaled inversely to the adapter's live connection count. An idle gateway
// therefore holds close to zero Lua states regardless of how much aggregate
// concurrency is configured. See adapterPool for the sizing rules.
package lua

import (
	"crypto/hmac"
	"crypto/sha256"
	"embed"
	"encoding/base64"
	"encoding/hex"
	"encoding/json"
	"errors"
	"fmt"
	"os"
	"path/filepath"
	"sort"
	"strconv"
	"strings"
	"sync"
	"time"

	// golua exposes the LuaJIT C API. Build with -tags luajit so the cgo
	// LDFLAGS resolve to -lluajit-5.1 (see golua's lua.go).
	golua "github.com/aarzilli/golua/lua"
)

//go:embed adapters/*.lua plugins/*.lua
var bundledAdapters embed.FS

// adapterGlobal is the reserved global holding the adapter table after the
// worker boots the script. A reserved name avoids clashing with user code.
const adapterGlobal = "__llmsproxy_adapter"

type APIAdapter struct {
	Name    string `json:"name"`
	Version string `json:"version"`
}

// worker wraps one LuaJIT interpreter state. Not goroutine-safe; each worker
// is used by exactly one goroutine at a time, checked out from its pool.
type worker struct {
	L *golua.State
}

func (w *worker) close() {
	if w.L != nil {
		w.L.Close()
	}
}

// staticInfo caches the immutable fields of an adapter script, extracted once
// at load time. Endpoint() and the static headers fallback read from here so
// they never churn a pooled worker.
type staticInfo struct {
	name     string
	version  string
	endpoint string
	// endpointStream is the streaming counterpart of endpoint. It exists so the
	// KERNEL never has to know a vendor's streaming path: Gemini streams over the
	// same path with a different verb (:generateContent -> :streamGenerateContent),
	// and the verb is part of the path, so the two cannot both be static.
	//
	// Before this field the kernel carried `strings.Replace(ep,
	// ":generateContent", ":streamGenerateContent", 1)` — vendor protocol
	// knowledge inside the dispatcher. It now lives here, where it belongs:
	// the adapter declares both endpoints and the kernel just picks one.
	// An adapter that omits endpoint_stream uses endpoint for both (most do).
	endpointStream string
	headers        map[string]string
	// registeredAs is the key the adapter was LOADED under, which can differ
	// from name when a script declares its own name. Kept so a reload replaces
	// the same pool entry rather than silently registering a second one.
	registeredAs string
	// ingress is the client-facing endpoint path this adapter CLAIMS to parse,
	// e.g. "/v1/messages". It is the INBOUND dual of serialize_* (which is
	// outbound): when an adapter declares it, the kernel YIELDS parsing of that
	// endpoint's request body to the adapter's deserialize_request hook, instead
	// of its own built-in parser. The kernel keeps a built-in parser as the
	// fallback (R5), so an unconfigured deployment still serves /v1/messages —
	// but once an adapter claims the endpoint, protocol iteration lives in the
	// adapter, not the kernel. Empty means the adapter claims no ingress.
	ingress string
}

// adapterPool owns the worker states for one adapter. Idle workers are held in
// a slice guarded by a cond; workers are created lazily on demand and reclaimed
// by the VM janitor once demand drops.
//
// Sizing is elastic rather than fixed (plan 阶段 2). Three inputs drive it:
//
//   - maxW — the adapter's maximum concurrency (Σ max_concurrent of every source
//     using it). It is the hard ceiling AND sets the growth step, so a hot,
//     high-concurrency adapter ramps in batches while a low-concurrency one
//     creeps up one state at a time.
//   - created / len(idle) — how many states already exist, i.e. how much room is
//     left to grow and how much is available to reclaim.
//   - inUse — how many states are checked out right now, i.e. the live
//     connection count for this adapter. It sets the shrink step: the fewer
//     connections, the more aggressively idle states are released.
type adapterPool struct {
	name   string
	script string
	static staticInfo

	mu      sync.Mutex
	cond    *sync.Cond
	idle    []*worker
	created int // live states: len(idle) + inUse (+ states being booted)
	inUse   int // states checked out right now == live connections
	waiting int // goroutines blocked waiting for a free state
	maxW    int // hard ceiling = Σ max_concurrent of the sources using this adapter
	used    bool
	closed  bool

	lastGrow   time.Time
	idleRounds int // consecutive janitor rounds that saw reclaimable slack
	peakInUse  int // high-water mark of inUse, for observability
	// revision counts how many times this adapter's SOURCE was replaced (or the
	// adapter was removed and re-added). Callers that cache a decision derived
	// from the script's declared surface read it via AdapterRevision, so a
	// runtime rewrite (a plugin installing a new version of its own adapter)
	// cannot leave them acting on a stale answer.
	revision uint64
	// pluginMode makes boot() store the returned table under pluginGlobal
	// instead of adapterGlobal. Everything else (elastic sizing, reclaim) is
	// identical, which is why plugins reuse this pool rather than getting a
	// second implementation.
	pluginMode bool
}

const (
	// growDivisor turns the adapter's max concurrency into the growth step CAP:
	// step <= clamp(ceil(maxW / growDivisor), 1, growStepCap). The actual batch
	// is additionally bounded by how many callers are queued (see growPlanLocked),
	// so a wide adapter may ramp in big jumps but never warms states nobody is
	// waiting for.
	growDivisor = 8
	growStepCap = 8
	// growCooldown keeps a burst of misses from batching repeatedly while the
	// previous batch is still booting.
	growCooldown = time.Second
	// maxPluginStates caps how many concurrent VM states ONE plugin may occupy.
	// A plugin is third-party code on the request path, so its ceiling is much
	// lower than an adapter's (which is sized from the sources'
	// max_concurrent): a plugin hook is a short synchronous call, so a handful
	// of states is already far more parallelism than any real hook needs, and a
	// runaway plugin cannot balloon memory the way a per-source adapter pool can.
	maxPluginStates = 4
	// residentWorkers is how many states an adapter keeps warm once it has
	// served at least one request. Booting is milliseconds, but keeping one warm
	// removes that from the critical path of the next request. Adapters that
	// were never used keep nothing.
	residentWorkers = 1
	// idleHeadroom is the spare idle state kept above the live connection count,
	// so the next concurrent request does not have to boot.
	idleHeadroom = 1
	// shrinkInterval is how often the VM janitor reclaims idle states.
	shrinkInterval = 30 * time.Second
	// shrinkGraceRounds is how many consecutive janitor rounds must see slack
	// before anything is released. It guards against tearing a pool down between
	// two bursts — NOT against ordinary traffic: a pool serving one request at a
	// time still has reclaimable slack, and requiring "no traffic at all" would
	// pin a burst's leftover states forever on any busy gateway. The counter is
	// therefore reset by GROWTH (real contention), not by a mere checkout.
	shrinkGraceRounds = 2
)

// PoolStats is a snapshot of one adapter pool's elastic sizing, surfaced by the
// status API so the algorithm is observable instead of a black box.
type PoolStats struct {
	Name       string `json:"name"`
	Created    int    `json:"created"`
	Idle       int    `json:"idle"`
	InUse      int    `json:"in_use"`
	Waiting    int    `json:"waiting"`
	Max        int    `json:"max"`
	Resident   int    `json:"resident"`
	GrowStep   int    `json:"grow_step"`
	ShrinkStep int    `json:"shrink_step"`
	PeakInUse  int    `json:"peak_in_use"`
}

func newAdapterPool(name, script string, static staticInfo) *adapterPool {
	p := &adapterPool{name: name, script: script, static: static, maxW: 1, revision: 1}
	p.cond = sync.NewCond(&p.mu)
	return p
}

// setMax updates the pool's hard ceiling (Σ max_concurrent of its sources).
// Lowering it does not kill live states: the janitor reclaims the excess.
func (p *adapterPool) setMax(n int) {
	if n < 1 {
		n = 1
	}
	p.mu.Lock()
	p.maxW = n
	p.cond.Broadcast()
	p.mu.Unlock()
}

func ceilDiv(a, b int) int {
	if b <= 0 {
		return a
	}
	return (a + b - 1) / b
}

// growStepLocked is how many states to warm on a miss, derived from the
// adapter's maximum concurrency. Caller holds p.mu.
func (p *adapterPool) growStepLocked() int {
	step := ceilDiv(p.maxW, growDivisor)
	if step < 1 {
		return 1
	}
	if step > growStepCap {
		return growStepCap
	}
	return step
}

// residentLocked is the floor the janitor will not shrink below. Caller holds p.mu.
func (p *adapterPool) residentLocked() int {
	if !p.used {
		return 0
	}
	if residentWorkers > p.maxW {
		return p.maxW
	}
	return residentWorkers
}

// shrinkStepLocked is how many idle states to release this round. The step is
// inversely proportional to the live connection count: with no connections the
// slack collapses in one round, while a busy adapter gives up one state at a
// time so the hot path keeps its warm states. Caller holds p.mu.
func (p *adapterPool) shrinkStepLocked() int {
	excess := p.created - p.keepLocked()
	if excess <= 0 {
		return 0
	}
	step := ceilDiv(excess, 1+p.inUse)
	if step > excess {
		step = excess
	}
	if step > len(p.idle) {
		step = len(p.idle)
	}
	if step < 0 {
		return 0
	}
	return step
}

// keepLocked is the number of live states this pool should retain right now.
// Caller holds p.mu.
func (p *adapterPool) keepLocked() int {
	keep := p.residentLocked()
	if h := p.inUse + idleHeadroom; h > keep {
		keep = h
	}
	if keep > p.maxW {
		keep = p.maxW
	}
	return keep
}

func (p *adapterPool) stats() PoolStats {
	p.mu.Lock()
	defer p.mu.Unlock()
	return PoolStats{
		Name:       p.name,
		Created:    p.created,
		Idle:       len(p.idle),
		InUse:      p.inUse,
		Waiting:    p.waiting,
		Max:        p.maxW,
		Resident:   p.residentLocked(),
		GrowStep:   p.growStepLocked(),
		ShrinkStep: p.shrinkStepLocked(),
		PeakInUse:  p.peakInUse,
	}
}

// boot creates a fresh LuaJIT state, registers the shared globals and executes
// the adapter script, storing the returned adapter table in the state.
func (p *adapterPool) boot() (*worker, error) {
	L := golua.NewState()
	L.OpenLibs()
	setupGlobals(L)
	if err := L.DoString(p.script); err != nil {
		L.Close()
		return nil, fmt.Errorf("compile adapter %s: %w", p.name, err)
	}
	if L.Type(-1) != golua.LUA_TTABLE {
		L.Close()
		return nil, fmt.Errorf("adapter %s must return a table", p.name)
	}
	// SetGlobal POPS the value off the stack, so it can only be called once per
	// boot. Plugins therefore store under pluginGlobal only, and NOT under
	// adapterGlobal: a second SetGlobal on the now-empty stack would assign nil.
	if p.pluginMode {
		L.SetGlobal(pluginGlobal)
	} else {
		L.SetGlobal(adapterGlobal)
	}
	L.SetTop(0)
	return &worker{L: L}, nil
}

// growPlanLocked reserves capacity for a batch prewarm and returns how many
// EXTRA states to boot in the background (the caller already reserved one for
// itself).
//
// Two independent limits apply, and BOTH matter:
//
//   - the adapter's max concurrency caps the step (growStepLocked), so a
//     high-concurrency adapter is allowed to ramp in bigger jumps than a
//     narrow one;
//   - the number of goroutines actually BLOCKED waiting for a state bounds it
//     to real demand.
//
// Sizing on the ceiling alone over-provisions badly: an adapter with
// max_concurrent=76 has a step of 8, so two concurrent requests would warm 8
// states, and the janitor would throw 7 of them away a minute later — boot,
// discard, repeat, with RSS oscillating for no benefit. Prewarming only for
// goroutines that are genuinely queued keeps the ramp cheap without the churn.
//
// Caller holds p.mu.
func (p *adapterPool) growPlanLocked(contended bool) int {
	if !contended || p.waiting <= 0 {
		return 0
	}
	if time.Since(p.lastGrow) < growCooldown {
		return 0
	}
	extra := p.growStepLocked() - 1
	if extra > p.waiting {
		extra = p.waiting // never warm more than the queue needs
	}
	if extra <= 0 {
		return 0
	}
	if room := p.maxW - p.created; extra > room {
		extra = room
	}
	if extra <= 0 {
		return 0
	}
	// Reserve immediately so concurrent acquires respect the ceiling while the
	// batch is still booting.
	p.created += extra
	p.lastGrow = time.Now()
	return extra
}

// prewarm boots n reserved states and publishes them as idle. A boot failure
// releases its reservation; it is not fatal, since the synchronous path
// surfaces adapter errors already.
func (p *adapterPool) prewarm(n int) {
	for i := 0; i < n; i++ {
		p.mu.Lock()
		closed := p.closed
		p.mu.Unlock()
		if closed {
			p.releaseReservation(n - i)
			return
		}
		w, err := p.boot()
		if err != nil {
			p.releaseReservation(n - i)
			return
		}
		p.mu.Lock()
		if p.closed {
			p.created--
			p.mu.Unlock()
			w.close()
			p.releaseReservation(n - i - 1)
			return
		}
		p.idle = append(p.idle, w)
		p.cond.Signal()
		p.mu.Unlock()
	}
}

func (p *adapterPool) releaseReservation(n int) {
	if n <= 0 {
		return
	}
	p.mu.Lock()
	p.created -= n
	if p.created < 0 {
		p.created = 0
	}
	p.cond.Broadcast()
	p.mu.Unlock()
}

// acquireAcrossSwap checks out a worker, re-reading the registry whenever the
// pool it was handed has already been retired.
//
// Retry (rather than one-shot) because the caller can lose the race repeatedly
// when several goroutines rotate the same adapter at once: each retry can land
// on a pool that the NEXT rotation is already closing. The bound is a small
// number of attempts, not unbounded — a permanently broken adapter must still
// surface an error instead of spinning.
func acquireAcrossSwap(v *VM, name string, p *adapterPool) (*worker, error) {
	// A rotation publishes the new pool BEFORE retiring the old one (see
	// LoadAdapterSource), so a lookup here can only ever find the newest one.
	// The wait is what makes this correct under sustained rotation: several
	// goroutines rewriting the same adapter keep retiring the pool this one
	// holds, and a caller that gave up at that moment would fail a request for
	// no reason other than landing in a microsecond-wide window.
	//
	// The bound is wall-clock, not attempts, so the behaviour does not change
	// with contention. It is generous enough to ride out a burst of rotations
	// and short enough that a genuinely gone adapter still surfaces promptly.
	deadline := time.Now().Add(2 * time.Second)
	for {
		w, err := p.acquire()
		if err == nil {
			return w, nil
		}
		if !errors.Is(err, ErrPoolClosed) {
			return nil, err
		}
		fresh := v.pool(name)
		if fresh == nil {
			return nil, fmt.Errorf("adapter %s not loaded", name)
		}
		p = fresh
		if time.Now().After(deadline) {
			return nil, fmt.Errorf("adapter %s: pool kept being replaced for 2s: %w", name, err)
		}
		// Yield briefly: the new pool needs a moment to publish, and spinning
		// here would burn the CPU that the rotation itself needs.
		time.Sleep(time.Millisecond)
	}
}

// ErrPoolClosed reports that the pool a caller was about to use from was
// retired between lookup and acquire — which happens on every adapter hot-swap
// (a plugin rotating a token rewrites the adapter, and LoadAdapterSource
// replaces its pool).
//
// It is a sentinel so callers can RETRY against the newly published pool
// instead of surfacing an error to the client. Without it, a token rotation
// under load produces a burst of "adapter X pool closed" failures — the
// rotation is the operator's intent, and the window is microseconds wide, so
// failing the request that happened to land there is strictly worse than
// re-reading the registry.
var ErrPoolClosed = errors.New("adapter pool closed")

func (p *adapterPool) acquire() (*worker, error) {
	p.mu.Lock()
	for {
		if p.closed {
			p.mu.Unlock()
			return nil, fmt.Errorf("adapter %s: %w", p.name, ErrPoolClosed)
		}
		if n := len(p.idle); n > 0 {
			w := p.idle[n-1]
			p.idle = p.idle[:n-1]
			p.checkoutLocked()
			p.mu.Unlock()
			return w, nil
		}
		if p.created < p.maxW {
			// Every existing state busy on a miss = real concurrency, not a
			// sequential caller reusing one warm state.
			contended := p.created > 0 && p.inUse >= p.created
			p.created++
			// Growing is the signal that capacity is genuinely short, so any
			// pending shrink decision is stale.
			p.idleRounds = 0
			extra := p.growPlanLocked(contended)
			p.mu.Unlock()
			if extra > 0 {
				go p.prewarm(extra)
			}
			w, err := p.boot()
			if err != nil {
				p.mu.Lock()
				p.created--
				p.cond.Signal()
				p.mu.Unlock()
				return nil, err
			}
			p.mu.Lock()
			p.checkoutLocked()
			p.mu.Unlock()
			return w, nil
		}
		p.waiting++
		p.cond.Wait()
		p.waiting--
	}
}

// checkoutLocked accounts one state as handed out. Caller holds p.mu.
func (p *adapterPool) checkoutLocked() {
	p.inUse++
	p.used = true
	if p.inUse > p.peakInUse {
		p.peakInUse = p.inUse
	}
}

func (p *adapterPool) release(w *worker) {
	w.L.SetTop(0)
	p.mu.Lock()
	if p.inUse > 0 {
		p.inUse--
	}
	if p.closed {
		p.created--
		p.mu.Unlock()
		w.close()
		return
	}
	// A lowered ceiling is honoured on the spot rather than waiting for the
	// janitor: returning a state the pool may no longer keep closes it.
	if p.created > p.maxW {
		p.created--
		p.cond.Signal()
		p.mu.Unlock()
		w.close()
		return
	}
	p.idle = append(p.idle, w)
	p.cond.Signal()
	p.mu.Unlock()
}

// reclaim releases idle states down to the current keep target, using a step
// scaled by the live connection count. It returns how many states were closed.
func (p *adapterPool) reclaim() int {
	p.mu.Lock()
	if p.closed {
		p.mu.Unlock()
		return 0
	}
	step := p.shrinkStepLocked()
	if step <= 0 {
		p.idleRounds = 0
		p.mu.Unlock()
		return 0
	}
	p.idleRounds++
	if p.idleRounds < shrinkGraceRounds {
		p.mu.Unlock()
		return 0
	}
	victims := append([]*worker(nil), p.idle[len(p.idle)-step:]...)
	p.idle = p.idle[:len(p.idle)-step]
	p.created -= step
	p.idleRounds = 0
	p.mu.Unlock()
	for _, w := range victims {
		w.close()
	}
	return step
}

func (p *adapterPool) shutdown() {
	p.mu.Lock()
	p.closed = true
	idle := p.idle
	p.idle = nil
	p.created -= len(idle)
	p.cond.Broadcast()
	p.mu.Unlock()
	for _, w := range idle {
		w.close()
	}
}

// VM aggregates per-adapter worker pools. All exported methods are safe for
// concurrent use.
type VM struct {
	mu    sync.RWMutex
	dir   string
	pools map[string]*adapterPool

	// janitor drives elastic shrinking. One goroutine serves every pool (not
	// one per adapter), so an idle gateway costs a single sleeping goroutine.
	janitorOnce sync.Once
	janitorStop chan struct{}
	janitorDone chan struct{}
}

func NewVM(dir string) *VM {
	return &VM{
		dir:         dir,
		pools:       map[string]*adapterPool{},
		janitorStop: make(chan struct{}),
		janitorDone: make(chan struct{}),
	}
}

// startJanitor launches the single shrink loop. Idempotent.
func (v *VM) startJanitor() {
	v.janitorOnce.Do(func() {
		go v.janitorLoop(shrinkInterval)
	})
}

func (v *VM) janitorLoop(every time.Duration) {
	defer close(v.janitorDone)
	t := time.NewTicker(every)
	defer t.Stop()
	for {
		select {
		case <-v.janitorStop:
			return
		case <-t.C:
			v.reclaimIdle()
		}
	}
}

// reclaimIdle runs one shrink round over every pool. Exported behaviour is
// tested through ReclaimIdleNow.
func (v *VM) reclaimIdle() int {
	v.mu.RLock()
	pools := make([]*adapterPool, 0, len(v.pools))
	for _, p := range v.pools {
		pools = append(pools, p)
	}
	v.mu.RUnlock()
	n := 0
	for _, p := range pools {
		n += p.reclaim()
	}
	return n
}

// ReclaimIdleNow forces one shrink round and reports how many worker states
// were closed. Intended for tests and for an explicit "release idle memory"
// action; normal operation relies on the janitor.
func (v *VM) ReclaimIdleNow() int { return v.reclaimIdle() }

// PoolStats returns the elastic sizing snapshot of every loaded adapter,
// sorted by name.
func (v *VM) PoolStats() []PoolStats {
	v.mu.RLock()
	pools := make([]*adapterPool, 0, len(v.pools))
	for _, p := range v.pools {
		pools = append(pools, p)
	}
	v.mu.RUnlock()
	out := make([]PoolStats, 0, len(pools))
	for _, p := range pools {
		out = append(out, p.stats())
	}
	sort.Slice(out, func(i, j int) bool { return out[i].Name < out[j].Name })
	return out
}

func (v *VM) Start() error {
	v.startJanitor()
	if v.dir == "" {
		return nil
	}
	// First-run seeding: a brand-new adapter dir is created and populated with
	// the bundled adapters. If the dir already exists it is treated as
	// authoritative and never rewritten — deleting a file there is a real delete.
	firstRun := false
	if _, err := os.Stat(v.dir); os.IsNotExist(err) {
		firstRun = true
		if err := os.MkdirAll(v.dir, 0755); err != nil {
			return fmt.Errorf("mkdir adapter dir: %w", err)
		}
	}
	if firstRun {
		if err := v.writeBundledAdapters(); err != nil {
			return err
		}
	}
	entries, err := os.ReadDir(v.dir)
	if err != nil {
		return err
	}
	for _, e := range entries {
		if filepath.Ext(e.Name()) != ".lua" {
			continue
		}
		if err := v.LoadAdapter(filepath.Join(v.dir, e.Name())); err != nil {
			fmt.Printf("[lua] preload %s: %v\n", e.Name(), err)
		}
	}
	return nil
}

// bundledPluginsDir is the directory inside the embedded FS holding the
// plugins shipped with the gateway. It is separate from adapters/ on purpose:
// the two are loaded by different machinery into different kinds of Lua state
// (a protocol transform vs. request-pipeline hooks), and keeping them apart
// makes it obvious that dropping a file in one does not affect the other.
const bundledPluginsDir = "plugins"

// ReadBundledPlugin returns the source of a plugin shipped with the gateway.
// It exists so a test (or an operator tool) can load a bundled plugin without
// depending on whether seeding has already run for this directory.
func ReadBundledPlugin(name string) (string, error) {
	data, err := bundledAdapters.ReadFile(bundledPluginsDir + "/" + name + ".lua")
	if err != nil {
		return "", fmt.Errorf("bundled plugin %s: %w", name, err)
	}
	return string(data), nil
}

// writeBundledPlugins seeds the plugin directory with the shipped plugins.
//
// It runs only when the directory does not exist yet (same rule as adapters):
// once the directory exists it is authoritative, so deleting a shipped plugin is
// a real delete and editing one survives restarts.
func writeBundledPlugins(dir string) error {
	if dir == "" {
		return nil
	}
	if err := os.MkdirAll(dir, 0755); err != nil {
		return fmt.Errorf("mkdir plugin dir: %w", err)
	}
	entries, err := bundledAdapters.ReadDir(bundledPluginsDir)
	if err != nil {
		return nil // nothing embedded; not an error
	}
	for _, e := range entries {
		if e.IsDir() || filepath.Ext(e.Name()) != ".lua" {
			continue
		}
		dst := filepath.Join(dir, e.Name())
		// Per-file seeding: a plugin already present is left ALONE. That is what
		// makes an operator's own copy (or a locally edited one) survive an
		// upgrade, and it is also why a newly embedded plugin still arrives —
		// absence of THAT file, not of the directory, is the trigger.
		if _, err := os.Stat(dst); err == nil {
			continue
		} else if !os.IsNotExist(err) {
			return err
		}
		data, err := bundledAdapters.ReadFile(bundledPluginsDir + "/" + e.Name())
		if err != nil {
			continue
		}
		if err := os.WriteFile(dst, data, 0644); err != nil {
			return err
		}
	}
	return nil
}

func (v *VM) Stop() {
	select {
	case <-v.janitorStop:
		// already stopped
	default:
		close(v.janitorStop)
	}
	v.mu.Lock()
	pools := v.pools
	v.pools = map[string]*adapterPool{}
	v.mu.Unlock()
	for _, p := range pools {
		p.shutdown()
	}
}

// ConfigureConcurrency sets each adapter's worker CEILING to the sum of
// max_concurrent of every source using it. It is a ceiling, not a preallocation:
// states are booted on demand and reclaimed when demand drops, so a config with
// a large aggregate concurrency no longer implies a large resident pool.
// Values below one are clamped to 1.
func (v *VM) ConfigureConcurrency(adapterConcurrency map[string]int) {
	v.mu.RLock()
	defer v.mu.RUnlock()
	for name, n := range adapterConcurrency {
		if p, ok := v.pools[name]; ok {
			p.setMax(n)
		}
	}
}

func (v *VM) ListAdapters() []APIAdapter {
	v.mu.RLock()
	defer v.mu.RUnlock()
	list := make([]APIAdapter, 0, len(v.pools))
	for name, p := range v.pools {
		list = append(list, APIAdapter{Name: name, Version: p.static.version})
	}
	sort.Slice(list, func(i, j int) bool { return list[i].Name < list[j].Name })
	return list
}

// AdapterForIngress returns the registered name of the adapter that CLAIMS the
// given client-facing endpoint path (its static.ingress), or "" when none does.
//
// This is the kernel-yields decision for an inbound endpoint (user requirement:
// "/v1/messages is the kernel's fallback endpoint; an adapter may declare it and
// the kernel actively yields"). The kernel keeps its own built-in parser as the
// fallback (R5) — AdapterForIngress returning "" means "no adapter claimed it,
// use the built-in", so an unconfigured deployment still serves the endpoint.
//
// Ties (two adapters claiming the same path) resolve to the lexically-first
// name so the choice is deterministic rather than map-iteration order. A real
// deployment should not have two adapters claiming one ingress; when it does,
// the plugin conflict-arbitration surface is the place to make it explicit, but
// a parse endpoint must still pick SOMETHING rather than vary per request.
func (v *VM) AdapterForIngress(path string) string {
	if path == "" {
		return ""
	}
	v.mu.RLock()
	defer v.mu.RUnlock()
	chosen := ""
	for name, p := range v.pools {
		p.mu.Lock()
		claims := p.static.ingress == path
		p.mu.Unlock()
		if claims && (chosen == "" || name < chosen) {
			chosen = name
		}
	}
	return chosen
}

// LoadAdapter compiles and registers an adapter script from a file.
//
// The path is validated for SAFETY before reading (R6): it must sit inside the
// VM's adapter dir and the derived name must be a legal adapter name (no path
// traversal, absolute paths, or whitespace). It does NOT reject built-in names
// here: an operator placing openai.lua in adapter_dir to OVERRIDE the bundled
// openai is the documented hot-override mechanism. Only PLUGIN installs are
// barred from shadowing a built-in (see InstallPluginAdapter), because a
// third-party plugin clobbering the kernel's shipped adapter is tampering,
// whereas the operator editing their own deployment is not.
func (v *VM) LoadAdapter(path string) error {
	name, err := ValidateAdapterPath(path, v.dir, nil)
	if err != nil {
		return err
	}
	data, err := os.ReadFile(path)
	if err != nil {
		return fmt.Errorf("read adapter: %w", err)
	}
	return v.LoadAdapterSource(name, string(data))
}

// LoadAdapterSource registers an adapter from source code. Replacing an
// existing adapter shuts the old pool down and re-boots lazily, carrying the
// configured ceiling over so a hot-reload does not reset elastic sizing.
//
// The name is validated for path SAFETY only (no traversal/absolute/space).
// Built-in names are allowed: an operator override is legitimate. A plugin
// that must NOT shadow a built-in uses InstallPluginAdapter, which adds the
// reserved-name gate.
func (v *VM) LoadAdapterSource(name, code string) error {
	if err := ValidateAdapterName(name, nil); err != nil {
		return err
	}
	// 共享前导注入点 —— 只此一处。适配器源码有三个编译点
	//（worker boot / HasFn 探针 / inspectScript），它们全部源自这里存进
	// pool 的 script 文本，所以拼在这里就同时覆盖三条路径，且不改变 Lua
	// 作用域语义（运行时 SetGlobal 会在热重载/池重建时丢失）。
	// 见 adapter_prelude.go 的开篇：12 个适配器里有 12680 字节的逐字重复。
	code = withAdapterPrelude(code)
	static, err := inspectScript(name, code)
	if err != nil {
		return err
	}
	maxW := 1
	v.mu.RLock()
	if p, ok := v.pools[static.name]; ok {
		p.mu.Lock()
		maxW = p.maxW
		p.mu.Unlock()
	}
	v.mu.RUnlock()

	v.mu.Lock()
	// A reload must also retire an adapter that was loaded under a DIFFERENT
	// name than the one this script declares (a plugin can register under a key
	// it controls while shipping an adapter that declares the name sources
	// reference). Without this, updating that adapter would install a second
	// pool and leave the old one serving traffic.
	//
	// The key can live under EITHER name (this is the "reload replaces the same
	// pool entry rather than silently registering a second one" rule):
	//
	//   - pools[static.name]  — the pool filed under the DECLARED name (the
	//     common case, and the only case when a plugin installs under a key that
	//     matches its script's own name), or
	// The pool registry is keyed by the DECLARED name (static.name), always:
	// first install files pools[static.name], and a reload replaces that same
	// slot. registeredAs only RECORDS the registration key for reverse lookups
	// (EffectiveAdapterName scans for it); it is not a second key space.
	//
	// An earlier version redirected the key to prev.static.registeredAs here,
	// assuming pools were filed under the registration key — but first installs
	// file under the declared name, so that redirect created a SECOND pool and
	// left the old one serving whenever key≠declared (a real bug: a v2 install
	// via arbitration left v1 serving, and UpdateAdapter had the same hole).
	key := static.name
	old := v.pools[key]
	// The new pool MUST be filed under the SAME slot as the old one (key), and
	// named after that slot, so a later lookup by either the DECLARED name or
	// the REGISTRATION KEY finds the one live pool. Naming it after static.name
	// would move the pool to the declared-name slot when the incumbent lived
	// under the registration key, leaving the registration-key slot pointing at
	// the stale (old) pool that shutdown() then tears down — callers reading by
	// the registration key see "adapter missing", callers by declared name see
	// a pool that isn't the one serving. key is the stable identity; the
	// declared name is reachable through static.name.
	p := newAdapterPool(key, code, *static)
	p.setMax(maxW)
	// Carry the revision forward so the new source keeps counting ABOVE the old
	// one. A cache keyed on (vm, revision, name) must see a value different
	// from the previous generation's; restarting at 1 could collide with a
	// value an earlier generation already published, which would resurrect a
	// stale cache entry — exactly the bug this revision exists to prevent.
	//
	// The old pool is read while v.mu is HELD (a nested RLock here would
	// self-deadlock: sync.RWMutex is not reentrant) and while its own mu is
	// free, because shutdown() — which takes p.mu — happens after v.mu is
	// released below. No other path can mutate p.revision while we hold v.mu,
	// because every writer reaches the pool through v.mu first.
	if old != nil {
		old.mu.Lock()
		p.revision = old.revision + 1
		old.mu.Unlock()
	}
	v.pools[key] = p
	v.mu.Unlock()
	// shutdown() happens AFTER v.mu is released, and the new pool is already
	// published, so a request arriving in between sees the new code. Shutting
	// down first would open a window where the adapter is neither old nor new:
	// concurrent rotations then fail with "adapter X pool closed" — the exact
	// failure a token-rotation plugin would hit under load.
	if old != nil {
		old.shutdown()
	}
	v.startJanitor()
	return nil
}

// RemoveAdapter evicts an adapter and shuts its workers down.
func (v *VM) RemoveAdapter(name string) {
	v.mu.Lock()
	p := v.pools[name]
	delete(v.pools, name)
	v.mu.Unlock()
	if p != nil {
		p.shutdown()
	}
}

// poolGlobal returns the Lua global holding a pool's returned table: plugins
// store theirs under one global, adapters under another. Transform and friends
// must read the SAME one boot() wrote, or a plugin-mode pool answers
// "has no table" even though the state is perfectly healthy.
func poolGlobal(p *adapterPool) string {
	if p != nil && p.pluginMode {
		return pluginGlobal
	}
	return adapterGlobal
}

func (v *VM) pool(name string) *adapterPool {
	v.mu.RLock()
	defer v.mu.RUnlock()
	return v.pools[name]
}

// bundledNames returns the set of adapter names that ship inside the kernel.
// It is derived from the embedded FS (same source writeBundledAdapters uses),
// so a new built-in adapter is automatically reserved and an operator cannot
// clobber it by importing a same-named external adapter.
// BundledAdapterNames is the exported counterpart of bundledNames for callers
// outside the lua package (core.UploadAdapter validates against it). It returns
// the set of adapter names shipped inside the kernel so an external import that
// shadows one is rejected up front (R6).
func (v *VM) BundledAdapterNames() []string { return v.bundledNames() }

func (v *VM) bundledNames() []string {
	entries, err := bundledAdapters.ReadDir("adapters")
	if err != nil {
		return nil
	}
	out := make([]string, 0, len(entries))
	for _, e := range entries {
		if e.IsDir() {
			continue
		}
		out = append(out, strings.TrimSuffix(e.Name(), ".lua"))
	}
	return out
}

func (v *VM) writeBundledAdapters() error {
	// Derive the set from the embedded FS instead of a hardcoded name list so
	// adding an adapter never requires maintaining a second list.
	entries, err := bundledAdapters.ReadDir("adapters")
	if err != nil {
		return err
	}
	for _, e := range entries {
		if e.IsDir() || !strings.HasSuffix(e.Name(), ".lua") {
			continue
		}
		dst := filepath.Join(v.dir, e.Name())
		if _, err := os.Stat(dst); err == nil {
			continue
		}
		data, err := bundledAdapters.ReadFile("adapters/" + e.Name())
		if err != nil {
			continue
		}
		if err := os.WriteFile(dst, data, 0644); err != nil {
			return err
		}
	}
	return nil
}

// Transform runs adapter.Fn(raw) and returns the resulting string.
//
// One retry on ErrPoolClosed: the pool was swapped by a concurrent rewrite
// (plugin token rotation) between this lookup and the checkout. Re-reading the
// registry gets the new pool, which is exactly what the caller wants.
func (v *VM) Transform(name, fn, raw string) (string, error) {
	p := v.pool(name)
	if p == nil {
		return "", fmt.Errorf("adapter %s not loaded", name)
	}
	w, err := acquireAcrossSwap(v, name, p)
	if err != nil {
		return "", err
	}
	defer p.release(w)
	L := w.L

	L.SetTop(0)
	defer L.SetTop(0)

	L.GetGlobal(poolGlobal(p))
	if L.IsNil(-1) {
		return "", fmt.Errorf("adapter %s has no table", name)
	}
	L.GetField(-1, fn)
	if L.IsNil(-1) || !L.IsFunction(-1) {
		return "", fmt.Errorf("adapter %s missing %s", name, fn)
	}
	L.PushString(raw)
	if err := L.Call(1, 1); err != nil {
		return "", fmt.Errorf("%s: %w", fn, err)
	}
	return L.ToString(-1), nil
}

// BuildHeaders executes adapter.build_headers(meta). If the adapter defines no
// build_headers hook, it falls back to the static adapter.headers table.
func (v *VM) BuildHeaders(name string, meta map[string]interface{}) (map[string]string, error) {
	p := v.pool(name)
	if p == nil {
		return nil, fmt.Errorf("adapter %s not loaded", name)
	}

	w, err := p.acquire()
	if err != nil {
		return nil, err
	}
	defer p.release(w)
	L := w.L

	L.SetTop(0)
	defer L.SetTop(0)

	L.GetGlobal(adapterGlobal)
	if L.IsNil(-1) {
		return nil, fmt.Errorf("adapter %s has no table", name)
	}
	L.GetField(-1, "build_headers")
	if !L.IsFunction(-1) {
		// no hook: fall back to the static headers captured at load time
		p.mu.Lock()
		hdr := make(map[string]string, len(p.static.headers))
		for k, h := range p.static.headers {
			hdr[k] = h
		}
		p.mu.Unlock()
		return hdr, nil
	}

	L.SetTop(0)
	L.GetGlobal(adapterGlobal)
	L.GetField(-1, "build_headers")
	pushGoValue(L, meta)
	if err := L.Call(1, 1); err != nil {
		return nil, fmt.Errorf("build_headers: %w", err)
	}
	if L.Type(-1) != golua.LUA_TTABLE {
		return nil, fmt.Errorf("build_headers returned non-table")
	}
	headers := map[string]string{}
	luaTableToMap(L, -1, headers)
	return headers, nil
}

// TransformError executes the optional adapter.transform_error(status, body)
// hook: per-source condensing of an upstream error response into a short
// client-facing reason. ok=false means the adapter defines no hook (or it
// failed) — the caller falls back to the generic Go-side condenser.
func (v *VM) TransformError(name string, status int, body string) (reason string, ok bool, err error) {
	p := v.pool(name)
	if p == nil {
		return "", false, fmt.Errorf("adapter %s not loaded", name)
	}
	w, err := p.acquire()
	if err != nil {
		return "", false, err
	}
	defer p.release(w)
	L := w.L

	L.SetTop(0)
	defer L.SetTop(0)
	L.GetGlobal(adapterGlobal)
	if L.IsNil(-1) {
		return "", false, nil
	}
	L.GetField(-1, "transform_error")
	if !L.IsFunction(-1) {
		return "", false, nil
	}
	L.SetTop(0)
	L.GetGlobal(adapterGlobal)
	L.GetField(-1, "transform_error")
	L.PushInteger(int64(status))
	L.PushString(body)
	if callErr := L.Call(2, 1); callErr != nil {
		return "", false, fmt.Errorf("transform_error: %w", callErr)
	}
	if L.Type(-1) != golua.LUA_TSTRING {
		return "", false, nil
	}
	return L.ToString(-1), true, nil
}

func (v *VM) Endpoint(name string) string {
	p := v.pool(name)
	if p == nil {
		return ""
	}
	p.mu.Lock()
	defer p.mu.Unlock()
	return p.static.endpoint
}

// StreamEndpoint returns the adapter's streaming endpoint, or "" when it
// declares none (in which case the caller keeps using Endpoint).
//
// This is the whole mechanism that keeps vendor streaming verbs out of the
// kernel: an adapter whose stream path differs declares endpoint_stream, and
// the dispatcher only ever compares two opaque strings.
func (v *VM) StreamEndpoint(name string) string {
	p := v.pool(name)
	if p == nil {
		return ""
	}
	p.mu.Lock()
	defer p.mu.Unlock()
	return p.static.endpointStream
}

// HasFn reports whether the named adapter declares a callable field fn,
// WITHOUT acquiring a worker and WITHOUT running the script.
//
// It exists for capability probing that runs on the request path: the
// out-bound serializer asks "does this adapter declare serialize_chunk?" once
// per adapter, not once per chunk. Probing by calling Transform would be
// wrong twice over — it costs a pool checkout, and a MISSING hook comes back
// as an error string rather than a boolean.
//
// The probe runs on a throwaway state built from the SAME script text, so the
// answer reflects what the script declares. A script whose table is built
// conditionally on its environment could differ per state; that is already
// true of every other staticInfo field (endpoint, headers), so the pool does
// not get a special exemption here.
func (v *VM) HasFn(name, fn string) bool {
	p := v.pool(name)
	if p == nil {
		return false
	}
	p.mu.Lock()
	script := p.script
	p.mu.Unlock()
	if script == "" {
		return false
	}
	L := golua.NewState()
	defer L.Close()
	L.OpenLibs()
	setupGlobals(L)
	if err := L.DoString(script); err != nil {
		return false
	}
	if L.Type(-1) != golua.LUA_TTABLE {
		return false
	}
	L.GetField(-1, fn)
	defer L.Pop(1)
	return !L.IsNil(-1) && L.IsFunction(-1)
}

// ---- static inspection (compile-once at load time) ----

func inspectScript(name, code string) (*staticInfo, error) {
	L := golua.NewState()
	L.OpenLibs()
	setupGlobals(L)
	if err := L.DoString(code); err != nil {
		L.Close()
		return nil, fmt.Errorf("compile adapter: %w", err)
	}
	if L.Type(-1) != golua.LUA_TTABLE {
		L.Close()
		return nil, fmt.Errorf("adapter script must return a table")
	}
	// The registration key is the default identity; a script MAY declare its own
	// name, in which case the declared one wins. That case matters for plugins:
	// a plugin registers under a key it controls, but the adapter it installs
	// may declare the name sources actually reference — so the bridge reports
	// the effective name back instead of guessing.
	info := &staticInfo{name: name, headers: map[string]string{}}
	if n := readString(L, -1, "name"); n != "" {
		info.name = n
	}
	info.registeredAs = name
	info.version = readString(L, -1, "version")
	info.endpoint = readString(L, -1, "endpoint")
	info.endpointStream = readString(L, -1, "endpoint_stream")
	info.ingress = readString(L, -1, "ingress")
	if getFieldBatch(L, "headers") {
		luaTableToMap(L, -1, info.headers)
		L.Pop(1)
	} else {
		L.Pop(1)
	}
	L.Close()
	return info, nil
}

// getFieldBatch fetches field of the table at the top of the stack and returns
// true when the fetched value (now at the top) is a table.
func getFieldBatch(L *golua.State, field string) bool {
	L.GetField(-1, field)
	return L.Type(-1) == golua.LUA_TTABLE
}

func readString(L *golua.State, tableIdx int, field string) string {
	L.GetField(tableIdx, field)
	defer L.Pop(1)
	if L.Type(-1) == golua.LUA_TNIL || L.Type(-1) == golua.LUA_TNONE {
		return ""
	}
	return normalizeString(L, -1)
}

func normalizeString(L *golua.State, idx int) string {
	if L.Type(idx) == golua.LUA_TNUMBER {
		n := L.ToNumber(idx)
		return strconv.FormatFloat(n, 'f', -1, 64)
	}
	return L.ToString(idx)
}

// ---- shared globals installed into every worker state ----

// restorePcall puts the real pcall/xpcall back in place. golua's OpenLibs
// replaces them with unsafe_ variants (it captures errors via its own panic
// handler); the built-in adapters rely on pcall for soft error handling.
func restorePcall(L *golua.State) {
	for _, name := range []string{"pcall", "xpcall"} {
		L.GetGlobal("unsafe_" + name)
		if !L.IsNil(-1) {
			L.SetGlobal(name)
		} else {
			L.Pop(1)
		}
	}
}

// hookGuardName is the Lua global that wraps every plugin hook call.
const hookGuardName = "__llmsproxy_call_hook"

// hookGuardSrc defines the wrapper. It exists because of a hard constraint in
// the binding: golua's callEx calls L.StackTrace() on ANY pcall error, and
// StackTrace() calls lua_getinfo, which SIGSEGVs in this LuaJIT build once the
// stack is deep enough (a request_end payload with the AUTO chain trace does
// it). Go cannot recover from a C-level signal, so a single Lua mistake inside
// one plugin killed the whole gateway — the production outage this was written
// for. Repeated crashes proved the boundary is not theoretical.
//
// pcall INSIDE Lua catches the error before golua ever sees a non-zero
// pcall status, so the C stack-trace path is never entered. The failure comes
// back as an ordinary (nil, message) pair, which the Go side records in
// hook_errors and moves on from — the documented contract that "a broken plugin
// must not affect request forwarding" finally holds for script errors too, not
// just for Go panics.
//
//	returns: (result, errorMessage) — both nil/"" on success.
const hookGuardSrc = `
function ` + hookGuardName + `(fn, payload)
  local ok, res = pcall(fn, payload)
  if not ok then
    return nil, tostring(res)
  end
  return res, nil
end`

func registerHookGuard(L *golua.State) {
	if err := L.DoString(hookGuardSrc); err != nil {
		// Nothing useful to do here beyond leaving the global absent: invoke()
		// checks for it and falls back to a direct call, which still works for
		// correct plugins (only their ERRORS stop being survivable).
		L.Pop(1)
	}
}

func setupGlobals(L *golua.State) {
	restorePcall(L)
	buildJSONTable(L)
	registerHookGuard(L)

	registerFn(L, "hmac_sha256_hex", func(L *golua.State) int {
		key := L.ToString(1)
		data := L.ToString(2)
		m := hmac.New(sha256.New, []byte(key))
		m.Write([]byte(data))
		L.PushString(hex.EncodeToString(m.Sum(nil)))
		return 1
	})
	registerFn(L, "sha256_hex", func(L *golua.State) int {
		h := sha256.Sum256([]byte(L.ToString(1)))
		L.PushString(hex.EncodeToString(h[:]))
		return 1
	})
	registerFn(L, "base64_encode", func(L *golua.State) int {
		L.PushString(base64.StdEncoding.EncodeToString([]byte(L.ToString(1))))
		return 1
	})
	registerFn(L, "tohex", func(L *golua.State) int {
		L.PushString(hex.EncodeToString([]byte(L.ToString(1))))
		return 1
	})
	registerFn(L, "log", func(L *golua.State) int {
		fmt.Printf("[adapter/%s] %s\n", L.ToString(1), L.ToString(2))
		return 0
	})
}

func buildJSONTable(L *golua.State) {
	L.NewTable()
	for _, name := range []string{"encode", "decode"} {
		var fn golua.LuaGoFunction
		switch name {
		case "encode":
			fn = func(L *golua.State) int {
				b, err := jsonEncode(luaValueToGo(L, 1))
				if err != nil {
					L.PushString("null")
					return 1
				}
				L.PushString(string(b))
				return 1
			}
		case "decode":
			fn = func(L *golua.State) int {
				v, err := jsonDecode(L.ToString(1))
				if err != nil {
					L.PushNil()
					return 1
				}
				pushGoValue(L, v)
				return 1
			}
		}
		L.PushGoClosure(fn)
		L.SetField(-2, name)
	}
	L.SetGlobal("json")
}

func registerFn(L *golua.State, name string, fn golua.LuaGoFunction) {
	L.PushGoClosure(fn)
	L.SetGlobal(name)
}

// ---- value conversions ----

func luaValueToGo(L *golua.State, idx int) interface{} {
	switch L.Type(idx) {
	case golua.LUA_TSTRING:
		return L.ToString(idx)
	case golua.LUA_TNUMBER:
		return L.ToNumber(idx)
	case golua.LUA_TBOOLEAN:
		return L.ToBoolean(idx)
	case golua.LUA_TTABLE:
		return luaTableToGoValue(L, idx)
	default:
		return nil
	}
}

// luaTableToGo value converts a Lua table to either []interface{} (array 1..n)
// or map[string]interface{}.
func luaTableToGoValue(L *golua.State, idx int) interface{} {
	abs := idx
	if idx < 0 {
		abs = L.GetTop() + idx + 1
	}
	m := map[string]interface{}{}
	allInt, n, maxK := true, 0, 0
	L.PushNil()
	for L.Next(abs) != 0 {
		kt := L.Type(-2)
		if kt == golua.LUA_TNUMBER {
			k := int(L.ToNumber(-2))
			if k > maxK {
				maxK = k
			}
		} else {
			allInt = false
		}
		m[keyToString(L, -2)] = luaValueToGo(L, -1)
		L.Pop(1)
		n++
	}
	if allInt && maxK > 0 && n == maxK {
		arr := make([]interface{}, maxK)
		for i := 1; i <= maxK; i++ {
			arr[i-1] = m[strconv.Itoa(i)]
		}
		return arr
	}
	return m
}

func luaTableToMap(L *golua.State, idx int, dst map[string]string) {
	abs := idx
	if idx < 0 {
		abs = L.GetTop() + idx + 1
	}
	L.PushNil()
	for L.Next(abs) != 0 {
		dst[keyToString(L, -2)] = normalizeString(L, -1)
		L.Pop(1)
	}
}

// keyToString reads the stack value at idx as a string without modifying the
// original (golua's ToString mutates numeric stack values in place, which
// would break the next lua_next call), then pops the temporary copy.
func keyToString(L *golua.State, idx int) string {
	L.PushValue(idx)
	defer L.Pop(1)
	return normalizeString(L, -1)
}

func pushGoValue(L *golua.State, v interface{}) {
	switch x := v.(type) {
	case nil:
		L.PushNil()
	case string:
		L.PushString(x)
	case bool:
		L.PushBoolean(x)
	case int:
		L.PushNumber(float64(x))
	case int64:
		L.PushNumber(float64(x))
	case float64:
		L.PushNumber(x)
	case []interface{}:
		L.NewTable()
		for i, it := range x {
			pushGoValue(L, it)
			L.RawSeti(-2, i+1)
		}
	case map[string]interface{}:
		L.NewTable()
		for k, it := range x {
			pushGoValue(L, it)
			L.SetField(-2, k)
		}
	default:
		if b, err := json.Marshal(x); err == nil {
			var iv interface{}
			if json.Unmarshal(b, &iv) == nil {
				pushGoValue(L, iv)
				return
			}
		}
		L.PushNil()
	}
}

func jsonEncode(v interface{}) ([]byte, error) { return json.Marshal(v) }

// luaToJSON converts the Lua value at idx into a Go value via json.encode, then
// unmarshals it into out. It is the bridge used by the plugin manifest/UI
// reader: the plugin returns a plain Lua table, and Go wants a typed struct.
//
// It goes through JSON rather than walking the Lua stack directly because the
// adapter/plugin boundary already speaks JSON everywhere else (transform_request
// gets a JSON string, hooks get a JSON string), so this keeps one representation
// instead of two.
func luaToJSON(L *golua.State, idx int, out interface{}) error {
	if L.GetTop() < 1 {
		return fmt.Errorf("empty stack")
	}
	abs := idx
	if abs < 0 {
		abs = L.GetTop() + 1 + abs
	}
	if abs < 1 || abs > L.GetTop() {
		return fmt.Errorf("index %d out of range (top=%d)", idx, L.GetTop())
	}
	// Absolute indices throughout: this binding aborts the process (SIGABRT)
	// on a bad index rather than panicking, so the stack is captured before any
	// push instead of being addressed relative to a shifting top.
	//
	// json.encode is pushed onto the stack and the value is pushed AFTER it, so
	// Call(1, 1) invokes it (Call takes no function index — it calls whatever
	// sits below the nargs values).
	L.GetGlobal("json")
	if L.IsNil(-1) {
		L.SetTop(0)
		return fmt.Errorf("json global missing")
	}
	L.GetField(-1, "encode")
	if L.Type(-1) != golua.LUA_TFUNCTION {
		L.SetTop(0)
		return fmt.Errorf("json.encode missing")
	}
	L.PushValue(abs)
	if err := L.Call(1, 1); err != nil {
		L.SetTop(0)
		return err
	}
	if L.GetTop() < 1 || L.Type(-1) != golua.LUA_TSTRING {
		L.SetTop(0)
		return fmt.Errorf("json.encode did not return a string")
	}
	s := L.ToString(-1)
	L.SetTop(0)
	return json.Unmarshal([]byte(s), out)
}
func jsonDecode(s string) (interface{}, error) {
	var v interface{}
	if err := json.Unmarshal([]byte(s), &v); err != nil {
		return nil, err
	}
	return v, nil
}

// AdapterRevision returns a monotonically increasing revision for the named
// adapter's CURRENT source. Every LoadAdapterSource / RemoveAdapter bumps it.
//
// Purpose: callers that cache a decision derived from an adapter's declared
// surface (e.g. "does this adapter declare serialize_chunk?") store the
// revision alongside the cached value and treat a mismatch as invalid. That
// removes the need for those callers to be told when to invalidate — which
// they previously were not, so a plugin rewriting an adapter at runtime could
// leave a stale decision in place forever.
//
// The revision is per-adapter rather than global so an unrelated adapter's
// reload does not discard a still-valid cache entry.
func (v *VM) AdapterRevision(name string) uint64 {
	p := v.pool(name)
	if p == nil {
		return 0
	}
	p.mu.Lock()
	defer p.mu.Unlock()
	return p.revision
}

// InstallPluginAdapter registers an adapter ON BEHALF OF A PLUGIN, with the
// one extra rule plugins (and only plugins) are held to: the name must not
// shadow a built-in adapter.
//
// Why plugins are stricter than operators: an operator dropping openai.lua in
// adapter_dir is overriding their own deployment (documented, intended). A
// third-party plugin registering an adapter named "openai" would silently
// swap out the kernel's shipped protocol translation for code the operator
// never reviewed — tampering with the shipped artifact. R6 requires that to
// fail loudly with the conflict named, not succeed quietly.
//
// This is the entry point Phase 3's plugin adapter-install bridge calls.
func (v *VM) InstallPluginAdapter(name, code string) error {
	if err := ValidateAdapterName(name, v.bundledNames()); err != nil {
		return err
	}
	return v.LoadAdapterSource(name, code)
}

// AdapterVersion returns the version string an adapter declares, or "" when the
// adapter is not loaded.
//
// It exists so an operator (or a plugin's own UI) can SEE that a runtime
// rewrite happened: after a token rotation the version bumps, and /api/status
// reports it. Without that, a rotation is invisible from outside the process.
func (v *VM) AdapterVersion(name string) string {
	p := v.pool(name)
	if p == nil {
		return ""
	}
	p.mu.Lock()
	defer p.mu.Unlock()
	return p.static.version
}

// AdapterIngress returns the client-facing endpoint path an adapter CLAIMS
// (its static.ingress), or "" when it claims none. Exported so operator-facing
// surfaces (and tests) can see which adapter, if any, has taken over parsing of
// an inbound endpoint — the same declaration AdapterForIngress consults.
func (v *VM) AdapterIngress(name string) string {
	p := v.pool(name)
	if p == nil {
		return ""
	}
	p.mu.Lock()
	defer p.mu.Unlock()
	return p.static.ingress
}

// EffectiveAdapterName resolves the name an adapter is actually registered
// under, given the key it was loaded with.
//
// A script may declare its own `name`, and the declared one wins at load time.
// Callers that keep per-adapter state (the plugin ownership table, the
// serializer cache) must key it on the EFFECTIVE name, or the state attaches to
// a key nothing will ever look up again.
func (v *VM) EffectiveAdapterName(registeredAs string) string {
	if p := v.pool(registeredAs); p != nil {
		p.mu.Lock()
		if p.static.name != "" {
			n := p.static.name
			p.mu.Unlock()
			return n
		}
		p.mu.Unlock()
		return registeredAs
	}
	// The script declared a name different from the key it was loaded under, so
	// the pool is filed under the DECLARED name. Find it that way, or every
	// caller would conclude "not loaded" and re-register a duplicate.
	v.mu.RLock()
	defer v.mu.RUnlock()
	for _, p := range v.pools {
		if p.static.registeredAs == registeredAs {
			p.mu.Lock()
			n := p.static.name
			p.mu.Unlock()
			if n != "" {
				return n
			}
			return registeredAs
		}
	}
	return registeredAs
}