mesh-delivery/modules/mesh-delivery-builder/buildrun/toolchaingate.go

package buildrun

import (
	"fmt"
	"maps"
	"slices"
	"sort"
	"sync"
	"time"

	"git.novox.be/novox/mesh-contracts/go/clients/meshmoduleassignerclient"
	"git.novox.be/novox/mesh-delivery/modules/mesh-delivery-contracts/go/clients/meshwantsclient"
	meshwants "git.novox.be/novox/mesh-delivery/modules/mesh-delivery-contracts/go/seats/meshwants"
	"git.novox.be/novox/mesh-delivery/modules/mesh-delivery-sdk/buildstep"
)

// A build module writes <node>.<language> of its toolchains state on every node it runs on, and nothing deletes the key
// when the module leaves a node: no part writes another module's state, and the module is gone. A report of a node the
// build module no longer runs on is therefore no builder's (issue 1940: mesh-build-node taken off ace on 2026-10-23 left
// ace.typescript at v26.12.0, and every Node fingerprint said "builders disagree").
//
// ToolchainGate stands between the toolchains buckets and what takes their entries in (the indexer's PutToolchain, a
// ToolchainSet): an entry is passed on only while the assignments (the seat mesh-module-assigner's state `assignments`, held by
// the assigner, its holder's state assignments, read by meshmoduleassignerclient.Placed; ADR 0567, ADR 0549 §3) places its build module on
// its node, and only once it was written since that assignments; it is passed on as deleted otherwise, said once, and
// passed on when its node writes it again. A report written before the assignments is one an earlier run of the module
// left (issue 1940, 2026-10-25 20:56: mesh-build-node put on ace again brought ace.typescript v26.12.0 of 2026-10-23
// back into the agreement before the module ran there). Until the assignments is settled, and while no node's assignments
// is held (before the assigner's adopt writes the first one, or where the state cannot be read), every entry is passed
// on. Builders that run and report different versions still disagree: that refusal is the guard of novox/hq ADR 0465,
// and it stays.
//
// This replaces the copies of the gate the indexer, the builder and the checker each kept over the controller's
// assignments (issue 1940's subtask, issue 774: those go with the controller's lease, ADR 0549 §6).
type ToolchainGate struct {
	mu      sync.Mutex
	modules map[string]bool                                           // the build modules whose toolchains pass the gate
	forward func(module, key string, value []byte, deleted bool) bool // answers whether it changed what it holds
	logf    func(format string, a ...any)
	placed  placementsSince
	known   bool                         // the assignments is settled
	raw     map[string]map[string][]byte // module -> toolchains key -> value, as the bucket holds it
	left    map[string]string            // <module>/<key> -> why it was said left out, so each reason is said once
	// OnPlaced is told, outside the gate's lock, once the assignments's history is read and after each revision of it
	// FollowAssignmentsIn hears from then on (the counted nodes, Nodes, may have changed). Nil tells nobody.
	OnPlaced func()
}

// placementsSince is where the gate reads which modules stand on which node, and since when: the assigner's state
// (meshmoduleassignerclient.Placed, fed by Assignments) or the assignment half (meshwantsclient.Placed, fed by Revision).
type placementsSince interface {
	Seeded() bool
	Since(node, module string) (time.Time, bool)
	ByNode() map[string][]string
}

// NewToolchainGate is a gate over the toolchains of modules, passing what stands on to forward and saying with logf what
// it leaves out.
func NewToolchainGate(modules []string, forward func(module, key string, value []byte, deleted bool) bool, logf func(string, ...any)) *ToolchainGate {
	g := &ToolchainGate{modules: map[string]bool{}, forward: forward, logf: logf, placed: meshmoduleassignerclient.NewPlaced(),
		raw: map[string]map[string][]byte{}, left: map[string]string{}}
	for _, m := range modules {
		g.modules[m] = true
	}
	return g
}

// UseModules replaces the build modules whose reports the gate passes on again when the assignments moves (a part that
// finds them from the seat list and the manifests, as the indexer does, gives them at each change), and passes every
// report it holds on again. It answers whether what it passed on changed.
func (g *ToolchainGate) UseModules(modules []string) bool {
	g.mu.Lock()
	defer g.mu.Unlock()
	g.modules = map[string]bool{}
	for _, m := range modules {
		g.modules[m] = true
	}
	return g.all()
}

// Toolchain takes one change of a build module's toolchains bucket in and answers whether what it passed on changed.
func (g *ToolchainGate) Toolchain(module, key string, value []byte, deleted bool) bool {
	g.mu.Lock()
	defer g.mu.Unlock()
	if g.raw[module] == nil {
		g.raw[module] = map[string][]byte{}
	}
	if deleted {
		delete(g.raw[module], key)
	} else {
		g.raw[module][key] = value
	}
	return g.pass(module, key)
}

// Assignments takes one revision of the seat mesh-module-assigner's assignments in (at is when the bus wrote it) and answers whether what it
// passed on changed. A value that does not read is said, and places nothing on its node.
func (g *ToolchainGate) Assignments(key string, value []byte, deleted bool, at time.Time) bool {
	g.mu.Lock()
	defer g.mu.Unlock()
	placed, ok := g.placed.(*meshmoduleassignerclient.Placed)
	if !ok {
		g.logf("the assignments: %s is not read, the gate takes the assignment half (design 47)", key)
		return false
	}
	was := placed.Seeded()
	node, changed, err := placed.Put(key, value, deleted, at)
	if err != nil {
		g.logf("the assignments: %v (issue 1940)", err)
	}
	if !g.known {
		return false
	}
	if was != g.placed.Seeded() {
		// The first assignments came, or the last went: every report's standing changed.
		return g.all()
	}
	return g.again(node, changed)
}

// Revision takes one assignment graph revision the part's view holds in (design 47 *Each part acts on its own entries*:
// the placements come from the seat mesh-wants' assignment half, never the assigner's state; assignmentview's Held hands
// each) and answers whether what it passed on changed. The first revision taken settles the gate, as the assignments'
// history read does; a later one passes on again the reports of each node whose build modules' standing changed. A gate
// fed by Revision reads no Assignments.
func (g *ToolchainGate) Revision(r meshwants.AssignmentRevision) bool {
	g.mu.Lock()
	defer g.mu.Unlock()
	placed, ok := g.placed.(*meshwantsclient.Placed)
	if !ok {
		placed = meshwantsclient.NewPlaced()
		g.placed = placed
	}
	was := placed.Seeded()
	changed := placed.Take(r)
	if changed == nil {
		return false
	}
	if !g.known || was != placed.Seeded() {
		g.known = true
		return g.all()
	}
	moved := false
	for _, node := range slices.Sorted(maps.Keys(changed)) {
		moved = g.again(node, changed[node]) || moved
	}
	return moved
}

// Nodes are the nodes the assignments places any module on, sorted: the nodes the mesh counts (a node leaves the count
// when the assignments no longer names it; ADR 0624 §1). None until the assignments is read.
func (g *ToolchainGate) Nodes() []string {
	var out []string
	for node, modules := range g.placed.ByNode() {
		if len(modules) > 0 {
			out = append(out, node)
		}
	}
	sort.Strings(out)
	return out
}

// Settled says the assignments's history is read: from now on only the reports of placed nodes pass, once a node's
// assignments is held. It answers whether what it passed on changed.
func (g *ToolchainGate) Settled() bool {
	g.mu.Lock()
	defer g.mu.Unlock()
	g.known = true
	return g.all()
}

// all passes on every report again. The caller holds mu.
func (g *ToolchainGate) all() bool {
	changed := false
	for _, module := range slices.Sorted(maps.Keys(g.raw)) {
		for _, k := range slices.Sorted(maps.Keys(g.raw[module])) {
			changed = g.pass(module, k) || changed
		}
	}
	return changed
}

// again passes on every report of node by a build module among modules. The caller holds mu.
func (g *ToolchainGate) again(node string, modules []string) bool {
	changed := false
	for _, module := range modules {
		if !g.modules[module] {
			continue
		}
		for _, k := range slices.Sorted(maps.Keys(g.raw[module])) {
			if n, _, err := buildstep.ParseToolchainKey(k); err == nil && n == node {
				changed = g.pass(module, k) || changed
			}
		}
	}
	return changed
}

// pass hands one key on as it stands: its value when its node is placed and wrote it since (or the assignments is not
// known), deleted otherwise. The caller holds mu.
func (g *ToolchainGate) pass(module, key string) bool {
	value, held := g.raw[module][key]
	if !held {
		return g.forward(module, key, nil, true)
	}
	node, _, err := buildstep.ParseToolchainKey(key)
	why := ""
	if err == nil && g.known && g.placed.Seeded() {
		why = g.leftOut(module, node, value)
	}
	if why == "" {
		delete(g.left, module+"/"+key)
		return g.forward(module, key, value, false)
	}
	if g.left[module+"/"+key] != why {
		g.left[module+"/"+key] = why
		g.logf("%s.%s is left out of the agreed toolchain: %s (issue 1940)", ToolchainsBucket(module), key, why)
	}
	return g.forward(module, key, nil, true)
}

// leftOut is why a node's report is no builder's, or empty when it counts: the catalogue does not place the module on its
// node, or the report was written before it did. A value that is no toolchain is passed on for its reader to refuse.
// The caller holds mu.
func (g *ToolchainGate) leftOut(module, node string, value []byte) string {
	since, placed := g.placed.Since(node, module)
	if !placed {
		return fmt.Sprintf("the catalogue does not place %s on %s, so it is no builder", module, node)
	}
	tc, err := buildstep.DecodeToolchain(value)
	if err == nil && tc.At.Before(since) {
		return fmt.Sprintf("it was written at %s, before %s was placed on %s at %s: it counts once %s writes it again",
			tc.At.UTC().Format(time.RFC3339), module, node, since.UTC().Format(time.RFC3339), node)
	}
	return ""
}