mesh-host/internal/agent/agent.go

// Package agent is the node agent (novox/hq ADR 0429 §4, ADR 0539): the node-engine's own process that follows the
// node's wants and applies one module at a time.
//
// **It watches, it never polls.** A want `want.<node>.<module>` names a fingerprint and the digest of the one composition
// the assigner allows the node to apply for it; the composition itself waits under `<node>.<module>.<fingerprint>` in
// the node's bucket `composed-<node>`. When both are there and the composition's digest is the want's, and what runs
// differs, the agent applies that module's composition and writes its have with the fingerprint and the digest. A want
// `removed` removes that module's own resources. Nothing is applied for a composition no want names, and a failed apply
// is retried by the next change of either key, never by a clock (research 063 §9.1) — except an apply that failed only
// because the bus did not answer for a moment (Result.Transient: no responders while the bus reloads its accounts, a
// reconnect): that one is retried with a bounded backoff (Env.Backoff), and at once when the engine is told the bus is
// back (RetryTransient), since no key will change for it.
package agent

import (
	"bytes"
	"context"
	"crypto/ed25519"
	"encoding/json"
	"fmt"
	"sort"
	"strings"
	"sync"
	"time"

	"git.novox.be/novox/mesh-sdk/go/catalogue"
	"git.novox.be/novox/mesh-sdk/go/delivery"
)

// Update is one change of a key in a watched bucket. Caught marks the end of the keys that were there when the watch
// began; every Update after it is a change.
type Update struct {
	Key      string
	Value    []byte
	Revision uint64
	Deleted  bool
	Caught   bool
}

// Watcher follows one bucket: the keys matching the pattern now, a Caught update, then every change, until ctx ends.
type Watcher interface {
	Watch(ctx context.Context, pattern string) (<-chan Update, error)
}

// NodeModule is the scope of the node's own composition, `<node>.node`.
const NodeModule = "node"

// Have is what a module's have key says, as the agent reads and writes it.
type Have struct {
	Module   string
	FP       string
	Composed string
	// Phase is "running", "failed", "held", "waiting-yes" or "removed"; empty reads as running.
	Phase  string
	Detail string
	// Judgement and Waits are, in phase waiting-yes, the judgement of the move and the digest of the composition that
	// waits for a person's yes (ADR 0547 §3).
	Judgement string
	Waits     string
	// Sealed is, on a node composing its own, each value the composition it names carries sealed to the node's key
	// (delivery.Have.Sealed): kept in the have so the composer reuses them after a restart (ADR 0547 §3).
	Sealed []string
}

// Haves is the node's have bucket as the agent uses it.
type Haves interface {
	// Read is what the node-engine last wrote for a module, and whether it wrote anything.
	Read(ctx context.Context, module string) (Have, bool)
	// Write writes a module's have where it differs from the bucket's.
	Write(ctx context.Context, h Have)
	// Removed writes the module's have with the phase removed, once, if a key of it exists.
	Removed(ctx context.Context, module string)
}

// Result is how one apply or removal came out.
type Result struct {
	Failed bool
	// Transient says the failure was the bus not answering for a moment (no responders, a timeout, a reconnect, a
	// JetStream API error while the bus reloads), not the module's: the apply is tried again without a new revision.
	Transient bool
	Detail    string
}

// Backup is the answer of the backup asked before a module moves.
type Backup struct {
	Done  bool
	Words string
}

// Env is everything the agent needs of the node.
type Env struct {
	Node     string
	Wants    Watcher
	Composed Watcher
	Haves    Haves
	// Apply applies one module's composition (scope NodeModule for the node's own), given the compositions the node's
	// other modules run so a path, unit or group they hold is not removed or taken. It is called with one module at a
	// time.
	Apply func(ctx context.Context, c delivery.Composed, others []delivery.Composed) Result
	// Remove removes the module's own resources and nothing of another's.
	Remove func(ctx context.Context, module string, others []delivery.Composed) Result
	// Backup backs a module up before a new composition of it is applied; nil backs up nothing.
	Backup func(ctx context.Context, module string) Backup
	// Kept is the resources the node last applied, by module, from the declaration it kept before the switch (ADR 0539
	// §9), as JSON; the node's own are under NodeModule. Nil when the node kept none. A module the agent does not hold yet
	// still runs from it.
	Kept func() map[string][]json.RawMessage
	// KeptFilled is the paths the kept declaration names as filled by the provisioner, by the module whose file is at
	// each (the node's own under NodeModule). A module that still runs from the kept declaration keeps its filled files
	// filled: without them the filling refused each one as "not named as filled" (issue 1945). Nil names none.
	KeptFilled func() map[string][]string
	// Trusted is the public keys a composition may be signed by: the controller's and the operator's (ADR 0539 §3). A
	// composition unsigned, signed by another key, or changed after it was signed is refused and nothing of it is
	// applied. Nil trusts no key, so nothing is applied.
	Trusted func() []ed25519.PublicKey
	// Recorded says the engine holds a record of every resource of the composition, so it was applied and not removed
	// since; nil says yes. A composition the kept declaration or a have says runs is applied again when it says no.
	Recorded func(ctx context.Context, c delivery.Composed) bool
	// Revisions are a composition key's revisions the bucket keeps, oldest first; nil reads none. The composition a want
	// names is taken from them when a newer one replaced it under the same key (ADR 0539 clause 1).
	Revisions func(ctx context.Context, key string) ([]Update, error)
	// Settled is told a module's apply or removal that succeeded, once what runs holds it (Everything), outside the
	// agent's lock: the declaration the node applies now names what that composition names, its filled files among
	// them. The node-engine's filling tells the provisioner's client then to fill again what it refused (mesh-host
	// internal/filling Knock; ace, 2026-10-25 21:17). Nil tells nothing.
	Settled func(module string)
	// Backoff is how long to wait before the try-th retry (1, 2, …) of an apply that failed transiently. Nil retries a
	// transient failure only on RetryTransient or a change of either key.
	Backoff func(try int) time.Duration
	Say     func(string)
	// SelfComposed says the compositions are the node's own (novox/hq ADR 0547 §2, issue 1126): Composed hands what the
	// node composed itself, unsigned, and a want names the build alone. The composition composed for the want's build is
	// the one applied, when it differs from what runs.
	SelfComposed bool
	// Policy is a module's upgrade policy as its manifest derives it (catalogue.DerivedUpgrade); with SelfComposed, a
	// move of a module of policy record that recreates a container or moves its data waits for a person's yes in its
	// want (ADR 0547 §3). Nil judges no move.
	Policy func(module string) string
}

// transientFor is how many transient retries say the phase applying before the have says failed: past them the bus has
// not answered for long enough that it is no longer a moment, and the failure is loud while the retries go on.
const transientFor = 4

// Agent follows the wants.
type Agent struct {
	env Env

	mu      sync.Mutex
	wants   map[string]wantNow // by module
	comps   map[string]compNow // by composed key
	running map[string]delivery.Have
	// applied is the composition each module was last applied from: a revision a want named from the key's history is
	// not the key's latest value, and what it holds is still held out of the others' applies.
	applied map[string]compNow
	// attempt is the last apply of a module that failed or is held, with the revisions it was made from.
	attempt map[string]attempt
	checked map[string]bool
	// prior is a module the node ran before this agent started, so it has something to back up.
	prior map[string]bool
	said  map[string]string
	// notComposed is the modules whose have says the node could not compose them: why, and the have before (NotComposed).
	notComposed map[string]notComposedMark
	retry       chan []string
	// tries counts the transient failures in a row of each module; again carries a retry whose wait ran out, and
	// transient asks for every transiently failed module at once. stopped is closed when Run returns.
	tries     map[string]int
	gen       uint64
	again     chan againAsk
	transient chan struct{}
	stopped   chan struct{}
}

type againAsk struct {
	module string
	gen    uint64
}

type wantNow struct {
	Want    delivery.Want
	Node    bool // the node's own want
	Present bool
	Rev     uint64
}

type compNow struct {
	Raw []byte
	C   delivery.Composed
	Dig string
	Rev uint64
	Err error
}

type attempt struct {
	FP, Composed     string
	WantRev, CompRev uint64
	Held             bool
	AnsweredRev      uint64
	// Transient marks a failure of the bus, retried by the clock; Gen names the retry that is due for it.
	Transient bool
	Gen       uint64
}

// New is an agent for a node.
func New(env Env) *Agent {
	if env.Say == nil {
		env.Say = func(string) {}
	}
	return &Agent{env: env, wants: map[string]wantNow{}, comps: map[string]compNow{}, running: map[string]delivery.Have{}, applied: map[string]compNow{},
		attempt: map[string]attempt{}, checked: map[string]bool{}, prior: map[string]bool{}, said: map[string]string{}, notComposed: map[string]notComposedMark{},
		retry: make(chan []string, 8), tries: map[string]int{}, again: make(chan againAsk, 64),
		transient: make(chan struct{}, 1), stopped: make(chan struct{})}
}

// Run follows the wants until ctx ends. The keys there when it starts are taken first, the node's own composition
// before any module's, then each change as it comes.
func (a *Agent) Run(ctx context.Context) error {
	defer close(a.stopped)
	wants, err := a.env.Wants.Watch(ctx, "want."+a.env.Node+".>")
	if err != nil {
		return fmt.Errorf("the agent cannot watch its wants: %w", err)
	}
	comps, err := a.env.Composed.Watch(ctx, a.env.Node+".>")
	if err != nil {
		return fmt.Errorf("the agent cannot watch its compositions: %w", err)
	}
	var wantsCaught, compsCaught bool
	dirty := map[string]bool{}
	for {
		select {
		case <-ctx.Done():
			return nil
		case modules := <-a.retry:
			// Applied again though the keys did not change: a container that waited because it said busy (ADR 0360
			// §11) is replaced once it is idle.
			a.mu.Lock()
			for _, m := range modules {
				delete(a.running, m)
				delete(a.attempt, m)
				dirty[m] = true
			}
			a.mu.Unlock()
		case ask := <-a.again:
			// The wait of a transiently failed apply ran out: tried again from the keys it has, if nothing tried it since.
			a.mu.Lock()
			if att, ok := a.attempt[ask.module]; ok && att.Transient && att.Gen == ask.gen {
				delete(a.attempt, ask.module)
				dirty[ask.module] = true
			}
			a.mu.Unlock()
		case <-a.transient:
			a.mu.Lock()
			for m, att := range a.attempt {
				if att.Transient {
					delete(a.attempt, m)
					dirty[m] = true
				}
			}
			a.mu.Unlock()
		case u, ok := <-wants:
			if !ok {
				return fmt.Errorf("the watch of the wants ended")
			}
			if u.Caught {
				wantsCaught = true
			} else if m := a.takeWant(u); m != "" {
				dirty[m] = true
			}
		case u, ok := <-comps:
			if !ok {
				return fmt.Errorf("the watch of the compositions ended")
			}
			if u.Caught {
				compsCaught = true
			} else if m := a.takeComposition(u); m != "" {
				dirty[m] = true
				// The node's own may wait for this one to read (unreadableModules): looked at again with it.
				dirty[NodeModule] = true
			}
		}
		// Nothing is applied until what was there at the start is all in: a module's composition may arrive after its
		// want, and the node's own composition comes before any module's.
		if !(wantsCaught && compsCaught) || len(dirty) == 0 {
			continue
		}
		a.settle(ctx, dirty)
		dirty = map[string]bool{}
	}
}

// Retry asks for the modules to be applied again from the keys they have, though nothing changed: what ADR 0360 §11
// needs for a container that waits because it said busy. It never blocks; a retry asked while eight wait is dropped, and
// asked again by its caller's next look.
func (a *Agent) Retry(modules ...string) {
	select {
	case a.retry <- modules:
	default:
	}
}

// RetryTransient asks for every module whose last apply failed transiently to be applied again now: the engine says it
// when it knows the bus answers again (after it wrote the bus's user list and the bus reloaded its accounts). It never
// blocks.
func (a *Agent) RetryTransient() {
	select {
	case a.transient <- struct{}{}:
	default:
	}
}

// notComposedMark is why a module's have says the node could not compose it, and the have it said before.
type notComposedMark struct {
	why   string
	prior Have
}

// NotComposed says the node composing its own could not compose its modules (ADR 0547 §2: a stand-in refused, an
// input that cannot be read): each module with a want gets its have failed, naming why, while it keeps running what it
// ran, so the failure is seen where every have is read and not only in the journal. The have keeps what runs: the
// fingerprint, the digest and the sealed values the agent runs, or, before it ran anything since its start, those its
// last have said (a refusal at the first compose after a start). Called with "" once the node composes again: each
// module so marked has its have back as it was, unless an apply wrote it since.
func (a *Agent) NotComposed(ctx context.Context, why string) {
	a.mu.Lock()
	var modules []string
	if why != "" {
		for module, w := range a.wants {
			if w.Present && !w.Want.Removed && module != NodeModule && a.notComposed[module].why != why {
				modules = append(modules, module)
			}
		}
	} else {
		for module := range a.notComposed {
			modules = append(modules, module)
		}
	}
	a.mu.Unlock()
	sort.Strings(modules)
	for _, module := range modules {
		a.mu.Lock()
		mark, marked := a.notComposed[module]
		r, ran := a.running[module]
		a.mu.Unlock()
		if why == "" {
			a.mu.Lock()
			delete(a.notComposed, module)
			a.mu.Unlock()
			back := mark.prior
			if ran {
				back = Have{Module: module, FP: r.FP, Composed: r.Composed, Phase: string(delivery.PhaseRunning)}
				if r.Phase != "" {
					back.Phase = string(r.Phase)
				}
			}
			if back.Composed == "" {
				continue // nothing said what ran: nothing to say back
			}
			a.writeHave(ctx, back)
			continue
		}
		if !marked {
			if ran {
				mark.prior = Have{Module: module, FP: r.FP, Composed: r.Composed, Phase: string(r.Phase)}
				if mark.prior.Phase == "" {
					mark.prior.Phase = string(delivery.PhaseRunning)
				}
			} else if h, ok := a.env.Haves.Read(ctx, module); ok {
				mark.prior = h // what its last have said runs: this agent has run nothing of it since its start
			}
		}
		mark.why = why
		a.mu.Lock()
		a.notComposed[module] = mark
		a.mu.Unlock()
		a.writeHave(ctx, Have{Module: module, FP: mark.prior.FP, Composed: mark.prior.Composed, Sealed: mark.prior.Sealed,
			Phase: string(delivery.PhaseFailed), Detail: "the node cannot compose it, and keeps running what it ran: " + why})
	}
}

// writeHave writes a module's have; on a node composing its own, with the sealed values of the composition it names
// when the caller gave none (Have.Sealed): those of the composition the agent applied or holds at that digest, else
// those the module's last have said at that digest.
//
// A writer that names no composition (one written from the agent's running entry before this agent ran anything of the
// module since its start: held by a failed backup, waiting for a yes, not composed) keeps what the module's last have
// said runs, its fingerprint, digest and sealed values, rather than a have naming nothing: what runs is not lost, and
// the boxes the composer reuses after a restart stay (review of L2, rounds 2 and 3).
func (a *Agent) writeHave(ctx context.Context, h Have) {
	if h.Composed == "" && h.Module != NodeModule {
		if last, ok := a.env.Haves.Read(ctx, h.Module); ok && last.Composed != "" {
			h.FP, h.Composed = last.FP, last.Composed
			if h.Sealed == nil {
				h.Sealed = last.Sealed
			}
		}
	}
	if a.env.SelfComposed && h.Sealed == nil && h.Module != NodeModule && h.Composed != "" {
		h.Sealed = a.sealedOf(ctx, h.Module, h.FP, h.Composed)
	}
	a.env.Haves.Write(ctx, h)
}

func (a *Agent) sealedOf(ctx context.Context, module, fp, composed string) []string {
	a.mu.Lock()
	if c, ok := a.applied[module]; ok && c.Err == nil && c.Dig == composed {
		a.mu.Unlock()
		return SealedIn(c.C)
	}
	if c, ok := a.compositionFor(module, fp); ok && c.Err == nil && c.Dig == composed {
		a.mu.Unlock()
		return SealedIn(c.C)
	}
	a.mu.Unlock()
	if h, ok := a.env.Haves.Read(ctx, module); ok && h.Composed == composed {
		return h.Sealed
	}
	return nil
}

// SealedIn is each value a composition carries sealed to its node's key: a file's `sealed` and a vault box
// (catalogue.VaultSealedField), as their JSON text, in the order of its resources.
func SealedIn(c delivery.Composed) []string {
	var out []string
	for _, raw := range c.Resources {
		var r map[string]json.RawMessage
		if json.Unmarshal(raw, &r) != nil {
			continue
		}
		for _, field := range []string{"sealed", catalogue.VaultSealedField} {
			f, ok := r[field]
			if !ok {
				continue
			}
			var text string
			if json.Unmarshal(f, &text) != nil {
				text = string(f) // a vault box is carried as its JSON
			}
			if text != "" {
				out = append(out, text)
			}
		}
	}
	return out
}

// Running is the compositions the node runs now: each module's that the agent applied or found already running, and
// the node's own. What the node judges its health by and says of itself is built from them.
func (a *Agent) Running() []delivery.Composed {
	a.mu.Lock()
	defer a.mu.Unlock()
	mods := make([]string, 0, len(a.running))
	for m := range a.running {
		mods = append(mods, m)
	}
	sort.Strings(mods)
	var out []delivery.Composed
	for _, m := range mods {
		h := a.running[m]
		if h.Phase == delivery.PhaseRemoved {
			continue
		}
		if c, ok := a.compositionFor(m, h.FP); ok && c.Err == nil && c.Dig == h.Composed {
			out = append(out, c.C)
		} else if c, ok := a.applied[m]; ok && c.Dig == h.Composed {
			out = append(out, c.C)
		}
	}
	return out
}

// takeWant takes a change of a want key and returns the module it concerns, or "" for a key that is not a want.
func (a *Agent) takeWant(u Update) string {
	a.mu.Lock()
	defer a.mu.Unlock()
	if _, err := delivery.ParseWantNodeKey(u.Key); err == nil {
		w := wantNow{Node: true, Rev: u.Revision}
		if !u.Deleted {
			nw, err := delivery.DecodeNodeWant(u.Value)
			if err != nil {
				a.sayOnce("want:node", fmt.Sprintf("the node's own want is not one: %v", err))
				return ""
			}
			w.Present, w.Want = true, delivery.Want{FP: "", Composed: nw.Composed}
		}
		a.wants[NodeModule] = w
		return NodeModule
	}
	node, module, err := delivery.ParseWantKey(u.Key)
	if err != nil || node != a.env.Node {
		return ""
	}
	w := wantNow{Rev: u.Revision}
	if !u.Deleted {
		want, err := delivery.DecodeWant(u.Value)
		if err != nil {
			a.sayOnce("want:"+module, fmt.Sprintf("the want of %s is not one: %v", module, err))
			return ""
		}
		w.Present, w.Want = true, want
	}
	a.wants[module] = w
	return module
}

// takeComposition takes a change of a composition key and returns the module it concerns.
func (a *Agent) takeComposition(u Update) string {
	node, module, _, isNode, err := delivery.ParseComposedKey(u.Key)
	if err != nil || node != a.env.Node {
		return ""
	}
	a.mu.Lock()
	defer a.mu.Unlock()
	if isNode {
		module = NodeModule
	}
	if u.Deleted {
		delete(a.comps, u.Key)
		return module
	}
	now := compNow{Raw: u.Value, Rev: u.Revision}
	c, err := delivery.DecodeComposed(u.Value)
	if err == nil && a.env.SelfComposed {
		// The node composed it itself: nothing crossed the bus, so nothing is signed (ADR 0547 §2).
		now.C = c
		now.Dig, err = delivery.ComposedDigestOf(c)
		if err != nil {
			now.C = delivery.Composed{}
		}
	} else if err == nil {
		var trusted []ed25519.PublicKey
		if a.env.Trusted != nil {
			trusted = a.env.Trusted()
		}
		// The signature before anything of the composition is used: its digest is what the want names.
		now.C = c
		now.Dig, err = delivery.VerifyComposed(c, trusted)
		if err != nil {
			now.C = delivery.Composed{}
			a.sayOnce("verify:"+u.Key, fmt.Sprintf("the composition %s is refused and nothing of it is applied: %v", u.Key, err))
			err = fmt.Errorf("refused: %w", err)
		}
	} else {
		a.sayOnce("decode:"+u.Key, fmt.Sprintf("the composition %s cannot be read, and nothing of it is applied: %v", u.Key, err))
	}
	now.Err = err
	a.comps[u.Key] = now
	return module
}

func (a *Agent) sayOnce(key, line string) {
	if a.said[key] == line {
		return
	}
	a.said[key] = line
	a.env.Say(line)
}

// settle applies what the dirty modules' wants allow: the node's own first, then each module in name order.
func (a *Agent) settle(ctx context.Context, dirty map[string]bool) {
	order := make([]string, 0, len(dirty))
	for m := range dirty {
		order = append(order, m)
	}
	sort.Slice(order, func(i, j int) bool {
		if (order[i] == NodeModule) != (order[j] == NodeModule) {
			return order[i] == NodeModule
		}
		return order[i] < order[j]
	})
	for _, m := range order {
		if ctx.Err() != nil {
			return
		}
		a.settleOne(ctx, m)
	}
}

func (a *Agent) compositionFor(module, fp string) (compNow, bool) {
	var key string
	var err error
	if module == NodeModule {
		key, err = delivery.ComposedNodeKey(a.env.Node)
	} else {
		key, err = delivery.ComposedKey(a.env.Node, module, fp)
	}
	if err != nil {
		return compNow{}, false
	}
	c, ok := a.comps[key]
	return c, ok
}

// revisionNamed is the revision of a module's composition key whose digest is the one given, signed by a trusted key,
// from the revisions the bucket keeps; false when none is.
func (a *Agent) revisionNamed(ctx context.Context, module, fp, digest string) (compNow, bool) {
	if a.env.Revisions == nil {
		return compNow{}, false
	}
	var key string
	var err error
	if module == NodeModule {
		key, err = delivery.ComposedNodeKey(a.env.Node)
	} else {
		key, err = delivery.ComposedKey(a.env.Node, module, fp)
	}
	if err != nil {
		return compNow{}, false
	}
	revs, err := a.env.Revisions(ctx, key)
	if err != nil {
		a.mu.Lock()
		a.sayOnce("history:"+key, fmt.Sprintf("the revisions of %s cannot be read: %v", key, err))
		a.mu.Unlock()
		return compNow{}, false
	}
	var trusted []ed25519.PublicKey
	if a.env.Trusted != nil {
		trusted = a.env.Trusted()
	}
	for i := len(revs) - 1; i >= 0; i-- {
		u := revs[i]
		if u.Deleted || u.Key != key {
			continue
		}
		c, err := delivery.DecodeComposed(u.Value)
		if err != nil {
			continue
		}
		dig, err := delivery.VerifyComposed(c, trusted)
		if err != nil || dig != digest {
			continue
		}
		return compNow{Raw: u.Value, C: c, Dig: dig, Rev: u.Revision}, true
	}
	return compNow{}, false
}

// others are what the node's other modules run now, so what they hold is not taken: the compositions the agent runs and,
// for every module the agent does not hold yet, the resources the kept declaration has for it (ADR 0429 §4).
func (a *Agent) others(except string) []delivery.Composed {
	var out []delivery.Composed
	mods := make([]string, 0, len(a.running))
	for m := range a.running {
		if m != except {
			mods = append(mods, m)
		}
	}
	sort.Strings(mods)
	for _, m := range mods {
		h := a.running[m]
		if h.Phase == delivery.PhaseRemoved {
			continue
		}
		if c, ok := a.compositionFor(m, h.FP); ok && c.Err == nil && c.Dig == h.Composed {
			out = append(out, c.C)
		} else if c, ok := a.applied[m]; ok && c.Dig == h.Composed {
			out = append(out, c.C)
		}
	}
	if a.env.Kept != nil {
		kept := a.env.Kept()
		names := make([]string, 0, len(kept))
		for m := range kept {
			if _, held := a.running[m]; !held && m != except {
				names = append(names, m)
			}
		}
		sort.Strings(names)
		for _, m := range names {
			c := delivery.Composed{Node: a.env.Node, Module: m, Resources: kept[m]}
			if a.env.KeptFilled != nil {
				c.Filled = a.env.KeptFilled()[m]
			}
			if m == NodeModule {
				c.Module = ""
			}
			out = append(out, c)
		}
	}
	return out
}

// unreadableModules are the keys of the module compositions in the node's bucket that cannot be read or are refused, of a
// module the assigner has not removed: for each module only the composition its want names, else its newest, since a
// stale one nobody applies (written before a rule the SDK now holds) must not hold the node's own back. Called holding
// a.mu.
func (a *Agent) unreadableModules() []string {
	type pick struct {
		key string
		c   compNow
	}
	current := map[string]pick{}
	for key, c := range a.comps {
		_, m, fp, isNode, err := delivery.ParseComposedKey(key)
		if err != nil || isNode {
			continue
		}
		w := a.wants[m]
		if w.Present && w.Want.Removed {
			continue
		}
		if w.Present && w.Want.FP != "" {
			if w.Want.FP == fp {
				current[m] = pick{key, c}
			}
			continue
		}
		if prev, ok := current[m]; !ok || c.Rev > prev.c.Rev {
			current[m] = pick{key, c}
		}
	}
	var out []string
	for _, p := range current {
		if p.c.Err != nil {
			out = append(out, p.key)
		}
	}
	sort.Strings(out)
	return out
}

// heldOut is what an apply or a removal holds out: others, and every module whose own composition the node's bucket holds
// and the agent does not run yet. A module that left the node's own composition for one of its own keeps its resources
// until that one is applied, so the node's own apply never removes them in between (issue 1945, live on 2026-10-25:
// shanks lost systemd-resolved's guard while its composition waited for a credential). The composition its want names,
// else its newest. Not part of Everything: what has not been applied is not what runs.
func (a *Agent) heldOut(except string) []delivery.Composed {
	out := a.others(except)
	held := map[string]bool{}
	for _, c := range out {
		held[c.Module] = true
	}
	waiting := map[string]compNow{}
	for key, c := range a.comps {
		_, m, fp, isNode, err := delivery.ParseComposedKey(key)
		if err != nil || isNode || c.Err != nil || m == except || held[m] {
			continue
		}
		if h, ok := a.running[m]; ok && h.Phase != delivery.PhaseRemoved {
			continue
		}
		if w := a.wants[m]; w.Present && w.Want.Removed {
			continue // a module the assigner removed holds nothing back
		}
		if w := a.wants[m]; w.Present && !w.Want.Removed && w.Want.FP == fp {
			waiting[m] = c
			continue
		}
		if w := a.wants[m]; w.Present && !w.Want.Removed && waiting[m].Rev != 0 && waiting[m].C.FP == w.Want.FP {
			continue
		}
		if prev, ok := waiting[m]; !ok || c.Rev > prev.Rev {
			waiting[m] = c
		}
	}
	names := make([]string, 0, len(waiting))
	for m := range waiting {
		names = append(names, m)
	}
	sort.Strings(names)
	for _, m := range names {
		out = append(out, waiting[m].C)
	}
	return out
}

// Everything is what the node runs: the compositions the agent runs, and the kept declaration's resources for each module
// the agent does not hold yet. What the node judges its health by, and what the filling, the take act and the judgement
// of directories read, is built from it.
func (a *Agent) Everything() []delivery.Composed {
	a.mu.Lock()
	defer a.mu.Unlock()
	return a.others("\x00")
}

// HoldsNode says whether the agent runs the node's own composition, so the node's parts are its and not the kept
// declaration's.
func (a *Agent) HoldsNode() bool {
	a.mu.Lock()
	defer a.mu.Unlock()
	h, ok := a.running[NodeModule]
	return ok && h.Phase != delivery.PhaseRemoved
}

func (a *Agent) settleOne(ctx context.Context, module string) {
	a.mu.Lock()
	w := a.wants[module]
	a.mu.Unlock()
	if !w.Present {
		return
	}
	if w.Want.Removed {
		a.remove(ctx, module, w)
		return
	}
	a.mu.Lock()
	comp, have := a.compositionFor(module, w.Want.FP)
	a.mu.Unlock()
	if a.env.SelfComposed {
		if !have || comp.Err != nil {
			return // waits for the node's own composition of the want's build
		}
		// The composition the node composed for the want's build is the one it applies (ADR 0547 §2).
		w.Want.Composed = comp.Dig
		if a.waitsForYes(ctx, module, w, comp) {
			return
		}
	} else if (!have || comp.Err != nil || comp.Dig != w.Want.Composed) && w.Want.Composed != "" {
		// The composition the want names may be an older revision of the key, replaced by a newer one at the same
		// fingerprint (a want held there keeps its digest): taken from the revisions the bucket keeps (ADR 0539 clause 1).
		if old, found := a.revisionNamed(ctx, module, w.Want.FP, w.Want.Composed); found {
			comp, have = old, true
		}
	}
	if !have || comp.Err != nil {
		return // waits for the composition; a change of its key settles the module again
	}
	if comp.Dig != w.Want.Composed {
		// A composition whose digest no want names is not applied.
		a.mu.Lock()
		a.sayOnce("digest:"+module, fmt.Sprintf("the composition of %s at %s has digest %s, and its want names %s; nothing is applied until they agree",
			module, w.Want.FP, comp.Dig, w.Want.Composed))
		a.mu.Unlock()
		return
	}
	a.mu.Lock()
	if !a.checked[module] {
		a.checked[module] = true
		a.mu.Unlock()
		if a.baseline(ctx, module, w, comp) {
			return
		}
	} else {
		a.mu.Unlock()
	}
	a.mu.Lock()
	if r, ok := a.running[module]; ok && r.FP == w.Want.FP && r.Composed == w.Want.Composed && r.Phase != delivery.PhaseRemoved {
		a.mu.Unlock()
		return
	}
	att, tried := a.attempt[module]
	same := tried && att.FP == w.Want.FP && att.Composed == w.Want.Composed
	if same && !att.Held && att.WantRev == w.Rev && att.CompRev == comp.Rev {
		a.mu.Unlock()
		return // failed already for exactly these keys: retried by the next change of either
	}
	if same && att.Held && att.AnsweredRev >= w.Rev && att.CompRev == comp.Rev {
		a.mu.Unlock()
		return // held for a failed backup, and no answer since
	}
	if module == NodeModule {
		// Not while a module's composition cannot be read or is refused: what it carries may be what left the node's own,
		// and the node's own apply would remove it, held out by nothing (issue 1945, the review's gap 2).
		if bad := a.unreadableModules(); len(bad) > 0 {
			a.sayOnce("node-waits:"+strings.Join(bad, ","), fmt.Sprintf("the node's own composition is not applied while "+
				"%s cannot be read or is refused: what it carries would be removed", strings.Join(bad, ", ")))
			a.mu.Unlock()
			return
		}
	}
	others := a.heldOut(module)
	a.mu.Unlock()
	a.apply(ctx, module, w, comp, others)
}

// baseline decides whether what runs already is what the want names, once per module at the agent's start: the node's
// own have (what this node-engine last applied), else the declaration it kept before the switch (ADR 0539 §9). True when
// it is, with the have written; nothing is applied.
func (a *Agent) baseline(ctx context.Context, module string, w wantNow, comp compNow) bool {
	// What a have or the kept declaration says runs is believed only when the engine still holds a record of each of its
	// resources: a resource removed since (a node's own apply took it, issue 1945) is made again by an apply.
	if a.env.Recorded != nil && !a.env.Recorded(ctx, comp.C) {
		if module != NodeModule {
			if _, ok := a.env.Haves.Read(ctx, module); ok {
				a.mu.Lock()
				a.prior[module] = true
				a.mu.Unlock()
			}
		}
		if a.env.Kept != nil {
			if _, ok := a.env.Kept()[module]; ok {
				a.mu.Lock()
				a.prior[module] = true
				a.mu.Unlock()
			}
		}
		a.mu.Lock()
		a.sayOnce("unrecorded:"+module+":"+comp.Dig, fmt.Sprintf("%s's composition %s is applied again: this engine holds "+
			"no record of some of its resources", module, short(comp.Dig)))
		a.mu.Unlock()
		return false
	}
	if module != NodeModule {
		if h, ok := a.env.Haves.Read(ctx, module); ok {
			a.mu.Lock()
			a.prior[module] = true
			a.mu.Unlock()
			if h.Composed != "" {
				// This node-engine wrote it: what it says is what runs, and the kept declaration says nothing more. A
				// failed apply or a different composition is applied again.
				if h.FP == w.Want.FP && h.Composed == w.Want.Composed &&
					(h.Phase == "" || h.Phase == string(delivery.PhaseRunning)) {
					a.mu.Lock()
					a.running[module] = delivery.Have{FP: h.FP, Composed: h.Composed, Phase: delivery.PhaseRunning}
					a.mu.Unlock()
					// Said again, so the session holds the module and its key carries the health this engine judges,
					// not the one an older engine left there (xorg's unknown on g14 and shanks, 2026-10-25).
					a.writeHave(ctx, Have{Module: module, FP: h.FP, Composed: h.Composed,
						Phase: string(delivery.PhaseRunning), Detail: h.Detail, Sealed: h.Sealed})
					return true
				}
				return false
			}
		}
	}
	if a.env.Kept == nil {
		return false
	}
	kept, ok := a.env.Kept()[module]
	if ok {
		a.mu.Lock()
		a.prior[module] = true
		a.mu.Unlock()
	}
	if !ok || !SameResources(kept, comp.C.Resources) {
		return false
	}
	a.mu.Lock()
	a.running[module] = delivery.Have{FP: w.Want.FP, Composed: w.Want.Composed, Phase: delivery.PhaseRunning}
	a.mu.Unlock()
	a.writeHave(ctx, Have{Module: module, FP: w.Want.FP, Composed: w.Want.Composed, Phase: string(delivery.PhaseRunning)})
	return true
}

func (a *Agent) apply(ctx context.Context, module string, w wantNow, comp compNow, others []delivery.Composed) {
	fp, dig := w.Want.FP, w.Want.Composed
	a.mu.Lock()
	_, ranBefore := a.running[module]
	ranBefore = ranBefore || a.prior[module]
	a.mu.Unlock()

	// A data module is backed up before a new composition of it moves (ADR 0429 §14), unless the want says to go on
	// without one (ADR 0539 §16).
	if comp.C.Data && ranBefore && a.env.Backup != nil && w.Want.Backup != delivery.BackupSkip {
		a.writeHave(ctx, Have{Module: module, FP: fp, Composed: dig, Phase: string(delivery.PhaseBackingUp)})
		if got := a.env.Backup(ctx, module); !got.Done {
			a.env.Say(fmt.Sprintf("%s is held: its backup before %s failed: %s", module, fp, got.Words))
			a.mu.Lock()
			a.attempt[module] = attempt{FP: fp, Composed: dig, WantRev: w.Rev, CompRev: comp.Rev, Held: true, AnsweredRev: w.Rev}
			prev := a.running[module]
			a.mu.Unlock()
			// It keeps running what it has: the have says what it holds and why it waits.
			a.writeHave(ctx, Have{Module: module, FP: prev.FP, Composed: prev.Composed, Phase: string(delivery.PhaseHeld),
				Detail: got.Words})
			return
		}
	}
	a.writeHave(ctx, Have{Module: module, FP: fp, Composed: dig, Phase: string(delivery.PhaseApplying)})
	res := a.env.Apply(ctx, comp.C, others)
	a.mu.Lock()
	what := fmt.Sprintf("%s at %s", module, fp)
	if module == NodeModule {
		what = "the node's own composition"
	}
	if res.Failed && res.Transient {
		a.tries[module]++
		try := a.tries[module]
		a.gen++
		gen := a.gen
		a.attempt[module] = attempt{FP: fp, Composed: dig, WantRev: w.Rev, CompRev: comp.Rev, Transient: true, Gen: gen}
		a.mu.Unlock()
		phase, when := delivery.PhaseApplying, "when the bus answers again"
		if a.env.Backoff != nil {
			wait := a.env.Backoff(try)
			when = "in " + wait.String()
			time.AfterFunc(wait, func() {
				select {
				case a.again <- againAsk{module, gen}:
				case <-a.stopped:
				case <-ctx.Done():
				}
			})
		}
		if try > transientFor {
			phase = delivery.PhaseFailed // still retried, and said as failed: no longer a moment
		}
		a.env.Say(fmt.Sprintf("applying %s failed because the bus did not answer (try %d), retried %s: %s", what, try, when, res.Detail))
		a.writeHave(ctx, Have{Module: module, FP: fp, Composed: dig, Phase: string(phase),
			Detail: fmt.Sprintf("retrying %s (try %d): %s", when, try, res.Detail)})
		return
	}
	delete(a.tries, module)
	if res.Failed {
		a.attempt[module] = attempt{FP: fp, Composed: dig, WantRev: w.Rev, CompRev: comp.Rev}
		a.mu.Unlock()
		a.env.Say(fmt.Sprintf("applying %s failed: %s", what, res.Detail))
		a.writeHave(ctx, Have{Module: module, FP: fp, Composed: dig, Phase: string(delivery.PhaseFailed), Detail: res.Detail})
		return
	}
	delete(a.attempt, module)
	delete(a.notComposed, module) // the apply says what it runs now
	a.running[module] = delivery.Have{FP: fp, Composed: dig, Phase: delivery.PhaseRunning}
	a.applied[module] = comp
	a.mu.Unlock()
	if a.env.Settled != nil {
		a.env.Settled(module)
	}
	a.writeHave(ctx, Have{Module: module, FP: fp, Composed: dig, Phase: string(delivery.PhaseRunning), Detail: res.Detail})
}

func (a *Agent) remove(ctx context.Context, module string, w wantNow) {
	a.mu.Lock()
	att := a.attempt[module]
	r, ran := a.running[module]
	if att.FP == "removed" && att.WantRev == w.Rev {
		a.mu.Unlock()
		return
	}
	if ran && r.Phase == delivery.PhaseRemoved {
		a.mu.Unlock()
		return
	}
	others := a.heldOut(module)
	a.mu.Unlock()
	res := a.env.Remove(ctx, module, others)
	a.mu.Lock()
	if res.Failed {
		a.attempt[module] = attempt{FP: "removed", WantRev: w.Rev}
		a.mu.Unlock()
		a.env.Say(fmt.Sprintf("removing %s failed: %s", module, res.Detail))
		return
	}
	delete(a.attempt, module)
	a.running[module] = delivery.Have{Phase: delivery.PhaseRemoved}
	a.mu.Unlock()
	if a.env.Settled != nil {
		a.env.Settled(module)
	}
	a.env.Haves.Removed(ctx, module)
}

// SameResources says whether two lists of resources hold the same resources, each by its id and its canonical JSON.
func SameResources(a, b []json.RawMessage) bool {
	ma, ok := canonByID(a)
	if !ok {
		return false
	}
	mb, ok := canonByID(b)
	if !ok || len(ma) != len(mb) {
		return false
	}
	for id, x := range ma {
		if y, there := mb[id]; !there || x != y {
			return false
		}
	}
	return true
}

func canonByID(rs []json.RawMessage) (map[string]string, bool) {
	out := map[string]string{}
	for _, raw := range rs {
		var head struct {
			ID string `json:"id"`
		}
		if json.Unmarshal(raw, &head) != nil || head.ID == "" {
			return nil, false
		}
		canon, err := Canonical(raw)
		if err != nil {
			return nil, false
		}
		out[head.ID] = string(canon)
	}
	return out, true
}

// Canonical is raw with its object keys sorted and its spacing dropped, its numbers kept as written.
func Canonical(raw json.RawMessage) (json.RawMessage, error) {
	var v any
	d := json.NewDecoder(bytes.NewReader(raw))
	d.UseNumber()
	if err := d.Decode(&v); err != nil {
		return nil, err
	}
	return json.Marshal(v)
}

func short(s string) string {
	if i := strings.IndexByte(s, ':'); i >= 0 {
		s = s[i+1:]
	}
	if len(s) > 12 {
		s = s[:12]
	}
	return s
}

// waitsForYes judges a move the node composed for a module of policy record (ADR 0547 §3, ADR 0377 §3): one that
// recreates a container or moves the module's data, or that cannot be judged, is not applied while the want names no yes
// for this composition; the have says phase waiting-yes with the judgement, and the assigner asks a person. A composition
// equal to what runs is no move. True when the module waits.
func (a *Agent) waitsForYes(ctx context.Context, module string, w wantNow, comp compNow) bool {
	if module == NodeModule || a.env.Policy == nil || a.env.Policy(module) != catalogue.PolicyRecord || w.Want.Yes == comp.Dig {
		return false
	}
	a.mu.Lock()
	r, ran := a.running[module]
	if ran && r.Composed == comp.Dig && r.Phase != delivery.PhaseRemoved {
		a.mu.Unlock()
		return false
	}
	from, known := a.runningComposition(module)
	prior := ran || a.prior[module]
	a.mu.Unlock()
	if !ran && !known {
		// Nothing of it run since this agent's start, and no kept declaration: its last have says what runs (review of L2,
		// round 3). The same composition is no move; another one is a move whose source cannot be read, so it waits for
		// a yes rather than being taken as the module's first composition.
		if h, ok := a.env.Haves.Read(ctx, module); ok && h.Composed != "" {
			if h.Composed == comp.Dig {
				return false
			}
			prior = true
		}
	}
	var j delivery.Judged
	switch {
	case known:
		j = delivery.JudgeMove(&from, comp.C)
	case prior:
		j = delivery.Judged{Judgement: delivery.JudgedCannot, Words: "what " + module + " runs on this node cannot be read"}
	default:
		j = delivery.JudgeMove(nil, comp.C)
	}
	switch j.Judgement {
	case delivery.JudgedRecreates, delivery.JudgedMovesData, delivery.JudgedCannot:
	default:
		return false
	}
	a.mu.Lock()
	a.sayOnce("yes:"+module+":"+comp.Dig, fmt.Sprintf("%s at %s waits for a person's yes: %s", module, w.Want.FP, j.Words))
	prev := a.running[module]
	a.mu.Unlock()
	a.writeHave(ctx, Have{Module: module, FP: prev.FP, Composed: prev.Composed, Phase: string(delivery.PhaseWaitingYes),
		Detail: j.Words, Judgement: string(j.Judgement), Waits: comp.Dig})
	return true
}

// runningComposition is what the module runs now: the composition last applied, else its resources in the declaration
// the node kept from before the switch. Called with a.mu held.
func (a *Agent) runningComposition(module string) (delivery.Composed, bool) {
	if c, ok := a.applied[module]; ok && c.Err == nil {
		return c.C, true
	}
	if a.env.Kept != nil {
		if rs, ok := a.env.Kept()[module]; ok {
			return delivery.Composed{Node: a.env.Node, Module: module, Resources: rs}, true
		}
	}
	return delivery.Composed{}, false
}