Merge pull request 'Hold the Approve ask until a review names the head (issue 1803, ADR 0496)' (#432) from fix/1803-hold-ask-without-review into main

Twice an unreviewed pull request was asked and merged on a warning line alone.

health-undeclared

@@ -1,4 +1,4 @@
-59
+58
 The long-running resources of this catalogue that do not say how they are ready (`health`, novox/hq ADR 0240
 rule 8), as `module check` counts them. It may only go down: merge-check.sh fails a change that raises it, and
 one that lowers it writes the new number on the first line. From 2026-11-18, or once it is 0, a long-running

modules/disk-load/README.md

@@ -24,7 +24,7 @@ machine's resources a seat, these tools become its verbs.
 
 ## Tools
 
-All five only read.
+All six only read.
 
 | tool | what | replaces |
 |---|---|---|
@@ -32,6 +32,7 @@ All five only read.
 | `disk_top` | Over a window. `by=unit` (the default) ranks systemd units and containers by bytes moved, from their cgroups. `by=container` ranks containers by name, and `by=process` ranks processes | `iotop` |
 | `disk_filesystems` | Each filesystem that stores data, said once however many places it is mounted: size, used and free GiB, used % as `df` counts it, inodes used % | `df -h`, `df -i` |
 | `disk_devices` | Each disk without sampling: model, size, whether it spins, scheduler, queue limit, what it is built on, and since boot the GiB read and written, the busy share and the wait per read and per write | `lsblk`, `cat /proc/diskstats` |
+| `disk_folders` | Which folders fill a disk, like `du -x`: the largest folders under `path` (default `/`) to `depth` levels (default 2, at most 3), `limit` of them (default 20, at most 100), with bytes and GiB. Sizes are allocated bytes, a hard link once. It stays on the path's one filesystem: a mount below it is named, never entered. It reads every folder the tool runner can read (docker's included) and no content; folders it cannot read are counted and named. After 25 s (the tool runner gives a call 30 s) it answers what it has measured, marked `partial`, with `stopped_at` (novox/hq issue 1804) | `du -x`, `ncdu` |
 | `cpu_top` | Samples a window (`seconds`, default 3, at most 10). It answers the machine's CPU in use (% of all CPUs and as CPUs busy, from `/proc/stat`), then `by=unit` (the default) ranks systemd units and containers by CPU, from their cgroups' cpu.stat, busiest first (`limit`, default 15), with the busy time no unit accounts for; `by=container` ranks containers by name | `top`, `systemd-cgtop` |
 
 ### How each number is read
@@ -87,4 +88,5 @@ cover:
 - processes ranked, one that started inside the window left out, unreadable ones counted;
 - filesystems said once, mount escapes undone, a statfs that fails said;
 - the window's bounds;
+- folders ranked biggest first to the depth asked, a mount never entered, unreadable folders named, a walk past its bound answered as partial with where it stopped, a hard link counted once;
 - the manifest: tools listed equal tools served, all read-only, no resources, no machine named.

modules/disk-load/cmd/disk-load/folders.go

@@ -0,0 +1,197 @@
+package main
+
+import (
+	"fmt"
+	"io/fs"
+	"math"
+	"os"
+	"path/filepath"
+	"sort"
+	"strings"
+	"syscall"
+	"time"
+)
+
+// FolderBudget is how long a folder walk may take before it answers what it has measured.
+const FolderBudget = 25 * time.Second
+
+// namedUnreadable is how many unreadable folders are named; the rest are only counted.
+const namedUnreadable = 50
+
+// Folder is one folder's size: everything under it on the same filesystem.
+type Folder struct {
+	Path  string  `json:"path"`
+	Depth int     `json:"depth"`
+	Bytes int64   `json:"bytes"`
+	GiB   float64 `json:"gib"`
+}
+
+// FolderReport is the answer of disk_folders.
+type FolderReport struct {
+	Path              string   `json:"path"`
+	Bytes             int64    `json:"bytes"`
+	GiB               float64  `json:"gib"`
+	Largest           []Folder `json:"largest"`
+	Partial           bool     `json:"partial"`
+	StoppedAt         string   `json:"stopped_at,omitempty"`
+	Unreadable        int      `json:"unreadable"`
+	UnreadableFolders []string `json:"unreadable_folders,omitempty"`
+	Skipped           []string `json:"other_filesystems_skipped,omitempty"`
+	Said              []string `json:"said"`
+}
+
+func (m *Machine) readDir(real string) ([]os.DirEntry, error) {
+	if m.ReadDir != nil {
+		return m.ReadDir(real)
+	}
+	return os.ReadDir(real)
+}
+
+func (m *Machine) now() time.Time {
+	if m.Clock != nil {
+		return m.Clock()
+	}
+	return time.Now()
+}
+
+func devOf(path string, info fs.FileInfo) uint64 {
+	if st, ok := info.Sys().(*syscall.Stat_t); ok {
+		return uint64(st.Dev)
+	}
+	return 0
+}
+
+type inodeKey struct{ dev, ino uint64 }
+
+// walk is one folder measurement.
+type walk struct {
+	m        *Machine
+	maxDepth int
+	dev      uint64
+	deadline time.Time
+	seen     map[inodeKey]bool
+	folders  []Folder
+	bad      []string
+	skipped  []string
+	stopped  string
+}
+
+// Folders measures path like du -x: allocated bytes, a hard link once, never entering another filesystem.
+func (m *Machine) Folders(path string, depth, limit int) (FolderReport, error) {
+	if !filepath.IsAbs(path) {
+		return FolderReport{}, fmt.Errorf("path must be absolute, not %q", path)
+	}
+	path = filepath.Clean(path)
+	info, err := os.Stat(m.path(path))
+	if err != nil {
+		return FolderReport{}, fmt.Errorf("path %s: %w", path, err)
+	}
+	if !info.IsDir() {
+		return FolderReport{}, fmt.Errorf("path %s is not a folder", path)
+	}
+	budget := m.Budget
+	if budget == 0 {
+		budget = FolderBudget
+	}
+	w := &walk{m: m, maxDepth: depth, deadline: m.now().Add(budget), seen: map[inodeKey]bool{}}
+	w.dev = w.device(path, info)
+	total := w.dir(path, 0)
+
+	sort.SliceStable(w.folders, func(i, j int) bool {
+		if w.folders[i].Bytes != w.folders[j].Bytes {
+			return w.folders[i].Bytes > w.folders[j].Bytes
+		}
+		return w.folders[i].Path < w.folders[j].Path
+	})
+	if len(w.folders) > limit {
+		w.folders = w.folders[:limit]
+	}
+	sort.Strings(w.bad)
+	named := w.bad
+	if len(named) > namedUnreadable {
+		named = named[:namedUnreadable]
+	}
+	r := FolderReport{
+		Path: path, Bytes: total, GiB: gibOf(total), Largest: w.folders,
+		Partial: w.stopped != "", StoppedAt: w.stopped,
+		Unreadable: len(w.bad), UnreadableFolders: named, Skipped: w.skipped,
+	}
+	if r.Largest == nil {
+		r.Largest = []Folder{}
+	}
+	var said []string
+	if r.Partial {
+		said = append(said, fmt.Sprintf("PARTIAL: the walk stopped after %s at %s; sizes are what was measured so far", budget, r.StoppedAt))
+	}
+	said = append(said, fmt.Sprintf("%s holds %.1f GiB on its own filesystem", path, r.GiB))
+	if r.Unreadable > 0 {
+		said = append(said, fmt.Sprintf("%d folders could not be read (%s), so their size is missing", r.Unreadable, strings.Join(named, ", ")))
+	}
+	if len(r.Skipped) > 0 {
+		said = append(said, "other filesystems not entered: "+strings.Join(r.Skipped, ", "))
+	}
+	r.Said = said
+	return r, nil
+}
+
+func (w *walk) device(path string, info fs.FileInfo) uint64 {
+	if w.m.DevOf != nil {
+		return w.m.DevOf(path, info)
+	}
+	return devOf(path, info)
+}
+
+// dir is the allocated bytes under a folder; it records the folder when it is within the depth.
+func (w *walk) dir(path string, depth int) int64 {
+	if w.stopped != "" {
+		return 0
+	}
+	if w.m.now().After(w.deadline) {
+		w.stopped = path
+		return 0
+	}
+	entries, err := w.m.readDir(w.m.path(path))
+	if err != nil {
+		w.bad = append(w.bad, path)
+		return 0
+	}
+	var total int64
+	for _, e := range entries {
+		child := filepath.Join(path, e.Name())
+		info, err := os.Lstat(w.m.path(child))
+		if err != nil {
+			continue // gone since it was listed
+		}
+		if info.IsDir() {
+			if w.device(child, info) != w.dev {
+				w.skipped = append(w.skipped, child)
+				continue
+			}
+			total += w.allocated(info, false) + w.dir(child, depth+1)
+			continue
+		}
+		total += w.allocated(info, true)
+	}
+	if depth >= 1 && depth <= w.maxDepth {
+		w.folders = append(w.folders, Folder{Path: path, Depth: depth, Bytes: total, GiB: gibOf(total)})
+	}
+	return total
+}
+
+// allocated is the bytes an entry occupies on disk, as du counts them; a file with several links once.
+func (w *walk) allocated(info fs.FileInfo, file bool) int64 {
+	st, ok := info.Sys().(*syscall.Stat_t)
+	if !ok {
+		return info.Size()
+	}
+	if file && st.Nlink > 1 {
+		k := inodeKey{uint64(st.Dev), uint64(st.Ino)}
+		if w.seen[k] {
+			return 0
+		}
+		w.seen[k] = true
+	}
+	return st.Blocks * 512
+}
+
+func gibOf(b int64) float64 { return math.Round(float64(b)/gib*100) / 100 }

modules/disk-load/cmd/disk-load/folders_test.go

@@ -0,0 +1,170 @@
+package main
+
+import (
+	"fmt"
+	"io/fs"
+	"os"
+	"strings"
+	"testing"
+	"time"
+)
+
+const mb = 1 << 20
+
+// megabytes writes a file of n MiB under the fake root.
+func (f *fake) megabytes(path string, n int) { f.file(path, strings.Repeat("x", n*mb)) }
+
+func aFullDisk(t *testing.T) *fake {
+	f := newFake(t)
+	f.megabytes("var/lib/docker/overlay2/a/layer", 8)
+	f.megabytes("var/lib/docker/overlay2/b/layer", 4)
+	f.megabytes("var/log/big.log", 3)
+	f.megabytes("home/jo/video.mkv", 6)
+	f.megabytes("home/jo/deep/er/est/file", 1)
+	f.dir("empty")
+	return f
+}
+
+func folders(t *testing.T, m *Machine, args map[string]any) FolderReport {
+	t.Helper()
+	out, err := tool(t, m, "disk_folders").Run(args)
+	if err != nil {
+		t.Fatal(err)
+	}
+	return out.(FolderReport)
+}
+
+func paths(r FolderReport) string {
+	var p []string
+	for _, f := range r.Largest {
+		p = append(p, f.Path)
+	}
+	return strings.Join(p, ",")
+}
+
+// 1804/AC1: given a disk with folders of known sizes, when disk_folders runs with the defaults, then it answers
+// the largest folders to depth 2, biggest first, with bytes and GiB, and the total of the path.
+func TestFoldersAreRankedBiggestFirstToDepthTwoByDefault(t *testing.T) {
+	r := folders(t, aFullDisk(t).machine(), nil)
+	if got := paths(r); !strings.HasPrefix(got, "/var,/var/lib,") || !strings.Contains(got, "/home,/home/jo") {
+		t.Fatalf("ranked %s", got)
+	}
+	if r.Path != "/" || r.Partial || r.Unreadable != 0 {
+		t.Errorf("%+v", r)
+	}
+	top := r.Largest[0]
+	if top.Path != "/var" || top.Depth != 1 || top.Bytes < 15*mb || top.Bytes > 16*mb || top.GiB <= 0 {
+		t.Errorf("top %+v", top)
+	}
+	for _, f := range r.Largest {
+		if f.Depth > 2 || f.Depth < 1 {
+			t.Errorf("%+v is outside depth 2", f)
+		}
+		if strings.Contains(f.Path, "overlay2") {
+			t.Errorf("%s is deeper than 2", f.Path)
+		}
+	}
+	if r.Bytes < 22*mb || r.Bytes > 24*mb {
+		t.Errorf("total %d", r.Bytes)
+	}
+}
+
+// 1804/AC2: given a path, a depth and a limit, then only that folder is measured, depth is held at 3 and limit at 100.
+func TestFoldersTakeAPathADepthAndALimitHeldAtThreeAndAHundred(t *testing.T) {
+	f := aFullDisk(t)
+	r := folders(t, f.machine(), map[string]any{"path": "/var/lib/docker", "depth": 1.0})
+	if r.Path != "/var/lib/docker" || paths(r) != "/var/lib/docker/overlay2" {
+		t.Errorf("%s: %s", r.Path, paths(r))
+	}
+	r = folders(t, f.machine(), map[string]any{"path": "/home", "depth": 9.0, "limit": 2.0})
+	if len(r.Largest) != 2 {
+		t.Errorf("limit 2 gave %d", len(r.Largest))
+	}
+	r = folders(t, f.machine(), map[string]any{"path": "/home", "depth": 9.0})
+	if got := paths(r); !strings.Contains(got, "/home/jo/deep/er") || strings.Contains(got, "/home/jo/deep/er/est") {
+		t.Errorf("depth 3 under /home is /home/jo/deep/er: %s", got)
+	}
+	tl := tool(t, f.machine(), "disk_folders")
+	for _, bad := range []map[string]any{{"depth": 0.0}, {"limit": 0.0}, {"depth": "deep"}, {"path": "relative/dir"}, {"path": "/nope"}, {"path": "/var/log/big.log"}} {
+		if _, err := tl.Run(bad); err == nil {
+			t.Errorf("%v was accepted", bad)
+		}
+	}
+	for i := 0; i < 120; i++ {
+		f.file(fmt.Sprintf("many/d%03d/f", i), "x")
+	}
+	if r := folders(t, f.machine(), map[string]any{"path": "/many", "limit": 500.0}); len(r.Largest) != 100 {
+		t.Errorf("limit held at %d, not 100", len(r.Largest))
+	}
+}
+
+// 1804/AC3: given a mount below the path, then it is never entered or counted, and is named.
+func TestFoldersNeverCrossIntoAnotherFilesystem(t *testing.T) {
+	f := aFullDisk(t)
+	m := f.machine()
+	m.DevOf = func(path string, info fs.FileInfo) uint64 {
+		if path == "/var/lib/docker" {
+			return 7
+		}
+		return 1
+	}
+	r := folders(t, m, nil)
+	if strings.Contains(paths(r), "docker") || len(r.Skipped) != 1 || r.Skipped[0] != "/var/lib/docker" {
+		t.Errorf("%s, skipped %v", paths(r), r.Skipped)
+	}
+	if r.Bytes > 14*mb {
+		t.Errorf("total %d includes the mount", r.Bytes)
+	}
+}
+
+// 1804/AC4: given folders it cannot read, then they are counted and named and the rest is still answered.
+func TestUnreadableFoldersAreCountedAndNamed(t *testing.T) {
+	f := aFullDisk(t)
+	m := f.machine()
+	real := os.ReadDir
+	m.ReadDir = func(p string) ([]os.DirEntry, error) {
+		if strings.HasSuffix(p, "/var/log") || strings.HasSuffix(p, "/overlay2/b") {
+			return nil, fmt.Errorf("open %s: permission denied", p)
+		}
+		return real(p)
+	}
+	r := folders(t, m, map[string]any{"depth": 3.0})
+	if r.Unreadable != 2 || strings.Join(r.UnreadableFolders, ",") != "/var/lib/docker/overlay2/b,/var/log" {
+		t.Errorf("%d: %v", r.Unreadable, r.UnreadableFolders)
+	}
+	if r.Partial || !strings.Contains(strings.Join(r.Said, "\n"), "2 folders could not be read") {
+		t.Errorf("%+v", r)
+	}
+}
+
+// 1804/AC5: given a walk that outlasts its bound, then it answers what it measured, partial, with where it stopped.
+func TestALongWalkAnswersPartialAndSaysWhereItStopped(t *testing.T) {
+	f := aFullDisk(t)
+	m := f.machine()
+	now := time.Unix(0, 0)
+	m.Clock = func() time.Time { now = now.Add(10 * time.Second); return now }
+	r := folders(t, m, nil)
+	if !r.Partial || r.StoppedAt == "" || len(r.Said) < 2 || !strings.HasPrefix(r.Said[0], "PARTIAL") || !strings.Contains(r.Said[0], r.StoppedAt) {
+		t.Errorf("%+v", r)
+	}
+}
+
+// 1804/AC6: a hard link is counted once, and the tool changes nothing.
+func TestAHardLinkIsCountedOnce(t *testing.T) {
+	f := newFake(t)
+	f.megabytes("a/big", 5)
+	if err := os.Link(f.root+"/a/big", f.root+"/a/again"); err != nil {
+		t.Skip(err)
+	}
+	r := folders(t, f.machine(), nil)
+	if r.Bytes > 6*mb {
+		t.Errorf("counted twice: %d", r.Bytes)
+	}
+}
+
+// 1804/AC7: the bound sits below the 30 s the tool runner gives a call, so a partial answer reaches its caller.
+func TestTheWalkBoundIsBelowTheToolRunnersCallLimit(t *testing.T) {
+	if FolderBudget >= 30*time.Second {
+		t.Errorf("the budget %s is not below the 30 s a call gets", FolderBudget)
+	}
+}

modules/disk-load/cmd/disk-load/machine.go

@@ -4,6 +4,7 @@ import (
 	"bytes"
 	"context"
 	"fmt"
+	"io/fs"
 	"os"
 	"os/exec"
 	"path/filepath"
@@ -53,6 +54,13 @@ type Machine struct {
 	Run    Runner
 	Wait   func(time.Duration)
 	Statfs func(path string) (FSStat, error)
+
+	// The folder walk's seams, all optional: the folder reader, the filesystem a path is on, the clock and
+	// the time the walk may take (60 s when zero).
+	ReadDir func(path string) ([]os.DirEntry, error)
+	DevOf   func(path string, info fs.FileInfo) uint64
+	Clock   func() time.Time
+	Budget  time.Duration
 }
 
 // Here is the machine this process runs on.

modules/disk-load/cmd/disk-load/tools.go

@@ -136,6 +136,34 @@ func Tools(m *Machine) []stdio.Tool {
 				}, nil
 			},
 		},
+		{
+			Name: "disk_folders",
+			Description: "Which folders fill a disk, like du -x: the largest folders under a path, measured on disk (allocated bytes, a hard link " +
+				"once), staying on the path's one filesystem (a mount below it is named, never entered). path defaults to /, depth to 2 (at most 3), " +
+				"limit to 20 (at most 100). Sizes in bytes and GiB. It reads every folder the tool runner can read, none of the content; folders " +
+				"it cannot read are counted and named. It stops after 25 s and answers what it measured so far, marked partial, with where it stopped. " +
+				"Replaces du. (r)",
+			Input: map[string]any{
+				"path":  map[string]any{"type": "string", "description": "the absolute folder to measure (default /)"},
+				"depth": map[string]any{"type": "integer", "description": "how many levels below path to rank (default 2, at most 3)"},
+				"limit": map[string]any{"type": "integer", "description": "how many folders to answer (default 20, at most 100)"},
+			},
+			Run: func(args map[string]any) (any, error) {
+				depth, err := count(args, "depth", 2, 1, 3)
+				if err != nil {
+					return nil, err
+				}
+				limit, err := count(args, "limit", 20, 1, 100)
+				if err != nil {
+					return nil, err
+				}
+				path := text(args, "path")
+				if path == "" {
+					path = "/"
+				}
+				return m.Folders(path, depth, limit)
+			},
+		},
 		{
 			Name: "disk_filesystems",
 			Description: "The space of every filesystem that stores data (not proc, tmpfs or overlay layers): size, used and free in GiB, " +

modules/disk-load/module.json

@@ -5,6 +5,7 @@
     "disk_load",
     "disk_top",
     "disk_filesystems",
+    "disk_folders",
     "disk_devices",
     "cpu_top"
   ],

modules/docker/README.md

@@ -160,6 +160,7 @@ A failure is an error naming how it failed, never an empty answer.
 | `docker_disk_usage` | r | `docker system df -v`: total, active and reclaimable per kind, with the largest of each |
 | `docker_networks` | r | networks, subnets, and the containers on each |
 | `docker_volumes` | r | volumes, who mounts each, whether the mesh holds one of them, anonymous or not, and sizes if asked |
+| `docker_volumes_remove_anonymous` | a | leftover anonymous volumes no container mounts in any state (so none the mesh holds), optionally created before a time; each other volume answered with why it stays; none can be named. **A dry run unless `dry_run` is false, which needs `why`. Never forced** (issue 1796). Sizes only when `sizes` is true (docker is slow on very many volumes); at most `limit` (default 200) per call |
 | `docker_events` | r | the runtime's events over a window ending now (default 60 min, at most 24 h), without exec noise unless asked; **an exec's command line is shown with any secret it carried as `[redacted: <what it was>]`** |
 | `docker_daemon_config` | r | `daemon.json` as on disk, `docker info`'s essentials, and keys the daemon has not taken yet |
 | `docker_unlabelled` | r | the containers the mesh does not hold: the cleanup list |

modules/docker/cmd/docker-tools/docker.go

@@ -944,16 +944,9 @@ func (c *Client) Volumes(ctx context.Context, unmountedOnly, sizes bool) (map[st
 	}
 	size := map[string]string{}
 	if sizes {
-		out, err := c.docker(ctx, "system", "df", "--verbose", "--format", "{{json .}}")
-		if err != nil {
+		if size, err = c.volumeSizes(ctx); err != nil {
 			return nil, err
 		}
-		var v struct{ Volumes []map[string]any }
-		if err := json.Unmarshal([]byte(out), &v); err == nil {
-			for _, vol := range v.Volumes {
-				size[fmt.Sprint(vol["Name"])] = fmt.Sprint(vol["Size"])
-			}
-		}
 	}
 	vols := []map[string]any{}
 	if len(lines(names)) > 0 {
@@ -1215,3 +1208,134 @@ func (c *Client) DaemonConfig(ctx context.Context) (map[string]any, error) {
 	answer["note"] = "keys read only at start take effect at the daemon's next restart; with live-restore on, a restart keeps every container running"
 	return answer, nil
 }
+
+// volumeSizes is every volume's size by name, from docker system df -v, which walks every volume:
+// on a machine with a thousand of them it does not answer within a call's bound.
+func (c *Client) volumeSizes(ctx context.Context) (map[string]string, error) {
+	size := map[string]string{}
+	out, err := c.docker(ctx, "system", "df", "--verbose", "--format", "{{json .}}")
+	if err != nil {
+		return nil, err
+	}
+	var v struct{ Volumes []map[string]any }
+	if err := json.Unmarshal([]byte(out), &v); err == nil {
+		for _, vol := range v.Volumes {
+			size[fmt.Sprint(vol["Name"])] = fmt.Sprint(vol["Size"])
+		}
+	}
+	return size, nil
+}
+
+// RemoveVolumesAsk is what docker_volumes_remove_anonymous is asked to take.
+type RemoveVolumesAsk struct {
+	DryRun        bool
+	Why           string
+	CreatedBefore *time.Time
+	Sizes         bool
+	Limit         int // how many to remove in one call; 0 is 200
+}
+
+// RemoveAnonymousVolumes removes the leftover anonymous volumes (novox/hq issue 1796): those docker
+// labels anonymous, which no container mounts in any state, so none a mesh-held container holds. The
+// classification is Volumes' own, not a copy. Every other volume is answered with why it stays. Each
+// is removed by its own `docker volume rm`, never forced and never `volume prune`, so one the daemon
+// refuses is reported and the rest still go.
+func (c *Client) RemoveAnonymousVolumes(ctx context.Context, a RemoveVolumesAsk) (map[string]any, error) {
+	if !a.DryRun && strings.TrimSpace(a.Why) == "" {
+		return nil, fmt.Errorf("why is required when dry_run is false")
+	}
+	all, err := c.Volumes(ctx, false, false)
+	if err != nil {
+		return nil, err
+	}
+	candidates := []string{}
+	skipped := []map[string]any{}
+	skip := func(name, why string) { skipped = append(skipped, map[string]any{"name": name, "why": why}) }
+	for _, v := range all["volumes"].([]map[string]any) {
+		name := fmt.Sprint(v["name"])
+		switch by := v["mounted_by"].([]map[string]any); {
+		case v["anonymous"] != true:
+			skip(name, "not anonymous: a named volume is data")
+		case v["mesh_held"] == true:
+			skip(name, "mounted by a container the mesh holds")
+		case len(by) > 0:
+			skip(name, fmt.Sprintf("mounted by %v (%v)", by[0]["container"], by[0]["state"]))
+		case a.CreatedBefore != nil && !volumeOlder(fmt.Sprint(v["created"]), *a.CreatedBefore):
+			skip(name, "not created before "+a.CreatedBefore.Format(time.RFC3339)+" (created "+fmt.Sprint(v["created"])+")")
+		default:
+			candidates = append(candidates, name)
+		}
+	}
+	answer := map[string]any{"dry_run": a.DryRun, "candidates": len(candidates)}
+	// Sizes are asked for, never needed: the listing walks every volume and may not answer, and
+	// then the call answers without them. Nothing below depends on it.
+	sizes := map[string]string{}
+	if a.Sizes {
+		if got, err := c.volumeSizes(ctx); err != nil {
+			answer["sizes_unknown"] = err.Error()
+		} else {
+			sizes = got
+			answer["sizes"] = sizes
+		}
+	}
+	sum := func(names []string) int64 {
+		var t int64
+		for _, n := range names {
+			t += max(Bytes(sizes[n]), 0)
+		}
+		return t
+	}
+	if a.DryRun {
+		answer["skipped"] = skipped
+		answer["would_remove"] = candidates
+		if len(sizes) > 0 {
+			answer["would_free_bytes"] = sum(candidates)
+		}
+		answer["note"] = "nothing was removed: call again with dry_run false and why to remove them (limit volumes at most per call)"
+		return answer, nil
+	}
+	limit := a.Limit
+	if limit <= 0 {
+		limit = 200
+	}
+	take := candidates
+	if len(take) > limit {
+		take = candidates[:limit]
+	}
+	removed := []string{}
+	// Chunks of a hundred in one `docker volume rm`, so a thousand volumes are ten calls; a chunk
+	// docker refuses any of is done again one by one, so what it refuses is named.
+	for i := 0; i < len(take); i += 100 {
+		chunk := take[i:min(i+100, len(take))]
+		if _, err := c.docker(ctx, append([]string{"volume", "rm"}, chunk...)...); err == nil {
+			removed = append(removed, chunk...)
+			continue
+		}
+		for _, name := range chunk {
+			if _, err := c.docker(ctx, "volume", "rm", name); err != nil {
+				// The batch took it before it refused another (or before its bound ran out):
+				// a volume that is no longer there is a removed one.
+				if strings.Contains(strings.ToLower(err.Error()), "no such volume") {
+					removed = append(removed, name)
+					continue
+				}
+				skip(name, "docker refused: "+err.Error())
+				continue
+			}
+			removed = append(removed, name)
+		}
+	}
+	answer["skipped"] = skipped
+	answer["removed"], answer["count"], answer["why"] = removed, len(removed), a.Why
+	answer["remaining"] = len(candidates) - len(take)
+	if len(sizes) > 0 {
+		answer["freed_bytes"] = sum(removed)
+	}
+	return answer, nil
+}
+
+// volumeOlder is whether a volume's CreatedAt is before the cutoff; one that cannot be read is not.
+func volumeOlder(created string, cutoff time.Time) bool {
+	t, err := time.Parse(time.RFC3339Nano, created)
+	return err == nil && t.Before(cutoff)
+}

modules/docker/cmd/docker-tools/docker_test.go

@@ -431,3 +431,145 @@ func TestEventsCoveringTheWholeWindowSayItIsCompleteAndFilterTheTypeHere(t *test
 		t.Fatalf("a full buffer reaching past the window: %v", got)
 	}
 }
+
+// 1796/AC1: only anonymous volumes nothing mounts go, each answered; a named, a mounted and a too young one stay with why.
+func TestRemoveAnonymousVolumesTakesOnlyTheUnmountedAnonymousOnes(t *testing.T) {
+	f := machine().
+		on("docker volume ls --quiet", Ran{Stdout: "webdata\nanonmounted\nold\nyoung\nnamed\n"}).
+		on("docker volume inspect", Ran{Stdout: `[
+{"Name":"webdata","Driver":"local","CreatedAt":"2026-01-01T00:00:00Z","Labels":{"com.docker.volume.anonymous":""}},
+{"Name":"old","Driver":"local","CreatedAt":"2026-09-01T00:00:00Z","Labels":{"com.docker.volume.anonymous":""}},
+{"Name":"young","Driver":"local","CreatedAt":"2026-10-04T11:00:00Z","Labels":{"com.docker.volume.anonymous":""}},
+{"Name":"named","Driver":"local","CreatedAt":"2026-01-01T00:00:00Z"}]`}).
+		on("docker system df --verbose", Ran{Stdout: `{"Volumes":[{"Name":"old","Size":"1.5MB"},{"Name":"young","Size":"1kB"}]}`}).
+		on("docker volume rm old", Ran{Stdout: "old\n"})
+	c := client(f, 1000)
+	before := time.Date(2026, 10, 3, 0, 0, 0, 0, time.UTC)
+	got, err := c.RemoveAnonymousVolumes(context.Background(), RemoveVolumesAsk{DryRun: true, Sizes: true, CreatedBefore: &before})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if f.ran("docker volume rm") || !reflect.DeepEqual(got["would_remove"], []string{"old"}) {
+		t.Fatalf("dry run: %v %+v", got, f.calls)
+	}
+	if got["would_free_bytes"] != int64(1500000) || got["sizes"].(map[string]string)["old"] != "1.5MB" {
+		t.Fatalf("dry run size: %v", got)
+	}
+	skipped := got["skipped"].([]map[string]any)
+	why := map[string]string{}
+	for _, s := range skipped {
+		why[s["name"].(string)] = s["why"].(string)
+	}
+	if len(why) != 3 || !strings.Contains(why["webdata"], "mesh holds") || why["young"] == "" || why["named"] == "" {
+		t.Fatalf("skipped: %v", skipped)
+	}
+	got, err = c.RemoveAnonymousVolumes(context.Background(), RemoveVolumesAsk{Why: "leftovers", Sizes: true, CreatedBefore: &before})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !f.ran("docker volume rm old") || got["count"] != 1 || !reflect.DeepEqual(got["removed"], []string{"old"}) {
+		t.Fatalf("removed: %v %+v", got, f.calls)
+	}
+	if got["freed_bytes"] != int64(1500000) {
+		t.Fatalf("freed: %v", got)
+	}
+	for _, c := range f.calls {
+		if len(c.args) > 1 && c.args[0] == "volume" && c.args[1] == "rm" && (len(c.args) != 3 || c.args[2] != "old") {
+			t.Errorf("reached too far: %v", c.args)
+		}
+	}
+}
+
+// 1796/AC2: a volume that is mounted by a mesh-held container is never removed; the tool takes no name.
+func TestRemoveAnonymousVolumesRefusesANamedVolumeAndKeepsWhatTheMeshHolds(t *testing.T) {
+	for _, tool := range tools(client(&fake{}, 1000)) {
+		if tool.Name != "docker_volumes_remove_anonymous" {
+			continue
+		}
+		for _, k := range []string{"name", "names", "volume", "volumes"} {
+			if _, err := tool.Run(map[string]any{k: "x", "dry_run": false, "why": "w"}); err == nil {
+				t.Errorf("%s accepted", k)
+			}
+		}
+		if _, err := tool.Run(map[string]any{"dry_run": false}); err == nil {
+			t.Error("a removal without why was accepted")
+		}
+		if _, err := tool.Run(map[string]any{"created_before": "yesterday"}); err == nil {
+			t.Error("a bad time was accepted")
+		}
+		return
+	}
+	t.Fatal("tool not served")
+}
+
+// 1796/AC3: on a machine with a very large set of volumes docker's size listing does not answer in
+// time; the tool still answers, and removal never depends on it.
+func TestRemoveAnonymousVolumesAnswersWhenTheSizeListingDoesNot(t *testing.T) {
+	f := (&fake{}).
+		on("docker ps --all --quiet --no-trunc", Ran{}).
+		on("docker volume ls --quiet", Ran{Stdout: "a1\na2\na3\n"}).
+		on("docker volume inspect", Ran{Stdout: `[{"Name":"a1","Labels":{"com.docker.volume.anonymous":""}},{"Name":"a2","Labels":{"com.docker.volume.anonymous":""}},{"Name":"a3","Labels":{"com.docker.volume.anonymous":""}}]`}).
+		on("docker system df", Ran{Status: 124, Err: "no answer within 20 s"}).
+		on("docker volume rm", Ran{Stdout: "ok\n"})
+	c := client(f, 1000)
+	got, err := c.RemoveAnonymousVolumes(context.Background(), RemoveVolumesAsk{DryRun: true})
+	if err != nil || f.ran("docker system df") || !reflect.DeepEqual(got["would_remove"], []string{"a1", "a2", "a3"}) {
+		t.Fatalf("default dry run: %v %v %+v", got, err, f.calls)
+	}
+	got, err = c.RemoveAnonymousVolumes(context.Background(), RemoveVolumesAsk{DryRun: true, Sizes: true})
+	if err != nil || !reflect.DeepEqual(got["would_remove"], []string{"a1", "a2", "a3"}) || got["sizes_unknown"] == nil {
+		t.Fatalf("sizes asked, df silent: %v %v", got, err)
+	}
+	got, err = c.RemoveAnonymousVolumes(context.Background(), RemoveVolumesAsk{Why: "w", Sizes: true, Limit: 2})
+	if err != nil || got["count"] != 2 || got["remaining"] != 1 || !f.ran("docker volume rm a1 a2") {
+		t.Fatalf("removal: %v %v %+v", got, err, f.calls)
+	}
+}
+
+// 1796/AC4: a batch docker half-did is settled by name: what is already gone counts as removed, what is in use stays with why.
+func TestRemoveAnonymousVolumesCountsWhatAFailedBatchAlreadyRemoved(t *testing.T) {
+	f := (&fake{}).
+		on("docker ps --all --quiet --no-trunc", Ran{}).
+		on("docker volume ls --quiet", Ran{Stdout: "a1\na2\na3\n"}).
+		on("docker volume inspect", Ran{Stdout: `[{"Name":"a1","Labels":{"com.docker.volume.anonymous":""}},{"Name":"a2","Labels":{"com.docker.volume.anonymous":""}},{"Name":"a3","Labels":{"com.docker.volume.anonymous":""}}]`}).
+		on("docker volume rm a1 a2 a3", Ran{Status: 1, Stdout: "a1\n", Stderr: "Error response from daemon: remove a2: volume is in use - [c1]\n"}).
+		on("docker volume rm a1", Ran{Status: 1, Stderr: "Error response from daemon: get a1: no such volume\n"}).
+		on("docker volume rm a2", Ran{Status: 1, Stderr: "Error response from daemon: remove a2: volume is in use - [c1]\n"}).
+		on("docker volume rm a3", Ran{Stdout: "a3\n"})
+	got, err := client(f, 1000).RemoveAnonymousVolumes(context.Background(), RemoveVolumesAsk{Why: "w"})
+	if err != nil {
+		t.Fatal(err)
+	}
+	sk := got["skipped"].([]map[string]any)
+	if !reflect.DeepEqual(got["removed"], []string{"a1", "a3"}) || got["count"] != 2 || len(sk) != 1 || sk[0]["name"] != "a2" || !strings.Contains(sk[0]["why"].(string), "in use") {
+		t.Fatalf("%v", got)
+	}
+}
+
+// 1796/AC5: removal goes in batches of a hundred names.
+func TestRemoveAnonymousVolumesRemovesInBatchesOfAHundred(t *testing.T) {
+	names, entries := []string{}, []string{}
+	for i := 0; i < 250; i++ {
+		n := fmt.Sprintf("v%03d", i)
+		names = append(names, n)
+		entries = append(entries, `{"Name":"`+n+`","Labels":{"com.docker.volume.anonymous":""}}`)
+	}
+	f := (&fake{}).
+		on("docker ps --all --quiet --no-trunc", Ran{}).
+		on("docker volume ls --quiet", Ran{Stdout: strings.Join(names, "\n")}).
+		on("docker volume inspect", Ran{Stdout: "[" + strings.Join(entries, ",") + "]"}).
+		on("docker volume rm", Ran{Stdout: "ok\n"})
+	got, err := client(f, 1000).RemoveAnonymousVolumes(context.Background(), RemoveVolumesAsk{Why: "w", Limit: 250})
+	if err != nil || got["count"] != 250 {
+		t.Fatalf("%v %v", got, err)
+	}
+	sizes := []int{}
+	for _, c := range f.calls {
+		if len(c.args) > 1 && c.args[0] == "volume" && c.args[1] == "rm" {
+			sizes = append(sizes, len(c.args)-2)
+		}
+	}
+	if !reflect.DeepEqual(sizes, []int{100, 100, 50}) {
+		t.Fatalf("batches %v", sizes)
+	}
+}

modules/docker/cmd/docker-tools/main.go

@@ -11,6 +11,7 @@ import (
 	"math"
 	"os"
 	"strings"
+	"time"
 
 	stdio "git.novox.be/novox/mesh-sdk/go"
 )
@@ -286,6 +287,40 @@ func tools(c *Client) []stdio.Tool {
 				return c.Volumes(ctx, flag(args, "unmounted", false), flag(args, "sizes", false))
 			},
 		},
+		{
+			Name: "docker_volumes_remove_anonymous",
+			Description: "Remove the leftover anonymous volumes (novox/hq issue 1796): only a volume docker labels anonymous (as docker_volumes says), which no container mounts in any state " +
+				"and so none the mesh holds, and with created_before only one created before it. Every other volume is answered with why it stays; a named volume is data and is never taken, " +
+				"and no volume can be named. Each is removed on its own, never forced; the answer has the count and names removed. With sizes true the answer has each volume's size and the bytes freed; docker's size listing is slow on many volumes, so it is asked for, and when it does not answer the call answers without it. At most limit are removed per call, in batches. " +
+				"A dry run by default; dry_run false needs why. Replaces docker volume rm and docker volume prune. (a)",
+			Input: map[string]any{
+				"dry_run":        map[string]any{"type": "boolean", "description": "list only (default true)"},
+				"why":            map[string]any{"type": "string", "description": "why: required when dry_run is false"},
+				"created_before": map[string]any{"type": "string", "description": "RFC 3339 time: only volumes created before it (default: any age)"},
+				"sizes":          map[string]any{"type": "boolean", "description": "also read each volume's size and the bytes freed (default false: docker walks every volume for it and on a machine with very many may not answer; the call then answers without sizes)"},
+				"limit":          map[string]any{"type": "integer", "description": "how many to remove in one call (default 200, at most 2000); the answer says how many remain, call again for them"},
+			},
+			Run: func(args map[string]any) (any, error) {
+				for _, k := range []string{"name", "names", "volume", "volumes"} {
+					if _, ok := args[k]; ok {
+						return nil, fmt.Errorf("%s: this tool takes no volume by name; it removes only unmounted anonymous volumes", k)
+					}
+				}
+				limit, err := bounded(args, "limit", 200, 2000)
+				if err != nil {
+					return nil, err
+				}
+				ask := RemoveVolumesAsk{DryRun: flag(args, "dry_run", true), Why: optional(args, "why"), Sizes: flag(args, "sizes", false), Limit: limit}
+				if s := optional(args, "created_before"); s != "" {
+					t, err := time.Parse(time.RFC3339, s)
+					if err != nil {
+						return nil, fmt.Errorf("created_before must be an RFC 3339 time such as 2026-10-01T00:00:00Z")
+					}
+					ask.CreatedBefore = &t
+				}
+				return c.RemoveAnonymousVolumes(ctx, ask)
+			},
+		},
 		{
 			Name: "docker_events",
 			Description: "What the runtime did in a window ending now (default the last 60 minutes, at most 24 hours): containers created, started, died, health changes, images pulled — with mesh_held. Exec events are left out unless asked; " +

modules/docker/module.json

@@ -33,6 +33,7 @@
     "docker_disk_usage",
     "docker_networks",
     "docker_volumes",
+    "docker_volumes_remove_anonymous",
     "docker_events",
     "docker_daemon_config",
     "docker_unlabelled",

modules/gitea/bind.ts

@@ -0,0 +1,44 @@
+// **A consumer that cannot be bound yet is bound again, never fatal** (novox/hq issue 1793). When the tool runtime
+// starts this bundle, the runtime's answer to "bind this module's consumer" can be "no responders" for a few seconds (the
+// runtime restarted by a delivery, the bus not yet serving the stream). That rejection, left unhandled at the top of
+// index.ts, ended the bundle: nothing heard `mesh-controller.checked` again, and a verdict that landed later never
+// reached the forge's statuses. A bind that fails is said and asked again with a growing wait until it holds.
+
+/** How long to wait before the nth new attempt (1-based): 2 s, 5 s, 10 s, then every 30 s. */
+export function bindWait(attempt: number): number {
+  return [2_000, 5_000, 10_000][attempt - 1] ?? 30_000;
+}
+
+/**
+ * Binds with `bind`, asking again after each failure, until it holds. `sleep` is the wait; `say` how a failure is said.
+ * Never rejects; resolves once the bind holds.
+ */
+export async function bindUntilHeld(what: string, bind: () => Promise<unknown>,
+  sleep: (ms: number) => Promise<void> = (ms) => new Promise((r) => setTimeout(r, ms)),
+  say: (line: string) => void = (line) => console.error(line)): Promise<void> {
+  for (let attempt = 1; ; attempt++) {
+    try {
+      await bind();
+      if (attempt > 1) say(`[gitea] ${what} is bound after ${attempt} attempts`);
+      return;
+    } catch (err) {
+      const wait = bindWait(attempt);
+      say(`[gitea] ${what} could not be bound (${err instanceof Error ? err.message : String(err)}); asking again in ${wait / 1000}s`);
+      await sleep(wait);
+    }
+  }
+}
+
+/**
+ * **Starts the binds without making the bundle wait for them** (novox/hq issue 1793). The launcher imports the bundle
+ * and then answers the runtime's handshake, which the runtime gives 10 s and then kills the bundle without starting it
+ * again; a bind awaited at the top of the module that stays refused longer ends the bundle all the same. So this
+ * returns at once: the binds run in the order given, each until it holds, and `done` is called after the last.
+ */
+export function bindInBackground(binds: [what: string, bind: () => Promise<unknown>][], done: () => void,
+  sleep?: (ms: number) => Promise<void>, say?: (line: string) => void): void {
+  void (async () => {
+    for (const [what, bind] of binds) await bindUntilHeld(what, bind, sleep, say);
+    done();
+  })();
+}

modules/gitea/index.ts

@@ -47,6 +47,7 @@ import { CHECK_CONTEXT, REPO_CHECK_CONTEXT, commentFor, composedGate, groupGateS
 import { GateWriter, clonedAs, type BuildStarted } from "./gate.js";
 import { HOOK_PATH, HOOK_PORT, Pace, claimHead, ensureSystemHook, eventsTool, keptSecret, receiver, type ForgeEvent } from "./webhook.js";
 import { keepSiblings, noteComposed, noteVerdict, readSiblings, siblingsFile, siblingsOf, weighSiblings } from "./siblings.js";
+import { bindInBackground } from "./bind.js";
 import { announcementOf, keepTook, lateVerdictComment, outOfDateWords, sameCommit, tookFile, weighTrunks } from "./trunk.js";
 
 // Without a way to a token — configured, or mintable with the admin account (token.ts) — there is
@@ -590,15 +591,18 @@ if (gitea) {
   tick(() => pollRepos(client), 60_000);
   tick(() => pollMerged(client), () => pace.every());
   tick(() => pollPulls(client), () => pace.every());
-  await on("mesh-controller.checked", (event) => {
-    const c = (event.body ?? {}) as Checked;
-    // A head's own verdict takes its turn after every write of that head's gate already asked for (gate.ts).
-    // A verdict of a check asked before the one now pending is an older run's, and dropped (novox/hq issue 456).
-    if (!c.group && c.owner && c.repo && c.commit) {
-      return gates.verdict(c.owner, c.repo, c.commit, () => setVerdict(client, event), c.id);
-    }
-    return setVerdict(client, event);
-  });
-  await on("node-build-agent.started", buildStarted);
-  console.log("[gitea] watching for new repositories and merged pull requests");
+  // Bound again until they hold, and not awaited (bind.ts, issue 1793): the runtime kills a bundle that has not answered its
+  // handshake in 10 s, so the import must end while a refused bind is still asked again.
+  bindInBackground([
+    ["mesh-controller.checked", () => on("mesh-controller.checked", (event) => {
+      const c = (event.body ?? {}) as Checked;
+      // A head's own verdict takes its turn after every write of that head's gate already asked for (gate.ts).
+      // A verdict of a check asked before the one now pending is an older run's, and dropped (novox/hq issue 456).
+      if (!c.group && c.owner && c.repo && c.commit) {
+        return gates.verdict(c.owner, c.repo, c.commit, () => setVerdict(client, event), c.id);
+      }
+      return setVerdict(client, event);
+    })],
+    ["node-build-agent.started", () => on("node-build-agent.started", buildStarted)],
+  ], () => console.log("[gitea] watching for new repositories and merged pull requests"));
 }

modules/gitea/merge.ts

@@ -12,7 +12,7 @@ import { mergeableUnder, newestByContext } from "./checks.js";
 import { MERGE_METHODS, Refused, checkName } from "./guard.js";
 import { CHECK_CONTEXT, REPO_CHECK_CONTEXT } from "./pulls.js";
 import { approvedHeadsOf, groupRefusal, memberKey, namesSiblings, notApproved, readSiblings, siblingOutOfDate, siblingsFile } from "./siblings.js";
-import { announcementOf, askedWords, outOfDateWords, sameCommit, tookFile, tookWords, twelve, verdictAtTrunk } from "./trunk.js";
+import { announcementOf, askedWords, outOfDateWords, sameCommit, tookFile, tookWords, twelve, verdictAtTrunk, waitsForComposed } from "./trunk.js";
 
 /** A head named for a merge: at least 12 hex characters of its sha (novox/hq issue 381), up to a sha-256 commit's 64. */
 const REVIEWED_HEAD = /^[0-9a-f]{12,64}$/;
@@ -117,6 +117,14 @@ export function mergeTool(gitea: GiteaClient, given: MergeDeps): ToolDefinition
           : at.judged ? `was judged against trunk ${twelve(at.judged)}` : "names no trunk commit it was judged against";
         const took = tookWords(tookFile(deps.stateDir), `${owner}/${repo}`);
         const head = twelve(pull.head_sha), now = twelve(trunkHead);
+        // **A member that waits for its group's composed check is not announced again** (novox/hq issue 1106): its own
+        // check can only say "composed" once more, so an ask here brought a verdict naming no trunk and the next merge was
+        // refused the same way for ever. The composed check is mesh-delivery's to ask; this says whose verdict is awaited.
+        if (waitsForComposed(checks[CHECK_CONTEXT])) {
+          throw new Refused(`${owner}/${repo}#${number} is not merged: its merge gate ${checks[CHECK_CONTEXT]?.description}; the ` +
+            `verdict that names its trunk commit is its group's composed check's, and mesh-delivery asks it, this tool asks no check ` +
+            `(main is now ${now}). Ask again once the group's composed check has passed. Nothing was merged`);
+        }
         if (at.onItsWay) {
           throw new Refused(`${owner}/${repo}#${number} is not merged: its verdict ${judged}, and ${pull.base} is now ` +
             `${now}. Its check is already asked and not yet answered (${checks[CHECK_CONTEXT]?.description}), so it is not ` +

modules/gitea/pulls.ts

@@ -226,6 +226,10 @@ function repoWrong(repo: Layer | undefined, facts: RepoCheckFacts): boolean {
   return !!repo && repoStatus(repo, facts).state !== "success";
 }
 
+/** What a group member's gate says while the group's composed check is the one that judges it: the member's own check can say
+ *  nothing more of it, so asking that check again brings no verdict that names a trunk (novox/hq issue 1106). */
+export const WAITS_FOR_COMPOSED = "waits for its group's composed check";
+
 /** The forge's status for the gate. */
 export function statusFor(c: Checked): CommitStatus {
   const gate = c.gate ?? { verdict: c.verdict, summary: c.summary };
@@ -260,7 +264,7 @@ export function statusesFor(c: Checked, facts: RepoCheckFacts = {}, gateNow?: st
     // set. The gate stays pending, saying it waits for the group, unless a composed verdict already stands on it.
     const out: CommitStatus[] = [];
     if (!(gateNow ?? "").includes(" composed: ")) {
-      out.push({ state: "pending", context: CHECK_CONTEXT, description: "waits for its group's composed check" });
+      out.push({ state: "pending", context: CHECK_CONTEXT, description: WAITS_FOR_COMPOSED });
     }
     const repo = c["repo-check"];
     if (repo) out.push(judgedFirst(c, repoStatus(repo, facts)));

modules/gitea/test/bind.test.ts

@@ -0,0 +1,58 @@
+// **A consumer that cannot be bound yet is bound again** (novox/hq issue 1793): the runtime's "no responders" at the
+// bundle's start, seen 2026-10-22 00:28:37 on mesh-catalog#427's verdict, ended the bundle and the verdict was never
+// written on the forge.
+
+import { test } from "node:test";
+import assert from "node:assert/strict";
+
+import { bindInBackground, bindUntilHeld, bindWait } from "../dist/bind.js";
+
+test("a bind refused with no responders is asked again until it holds, and says so", async () => {
+  let calls = 0;
+  const waits: number[] = [];
+  const said: string[] = [];
+  await bindUntilHeld("mesh-controller.checked", async () => {
+    calls++;
+    if (calls < 4) throw new Error("binding gitea's consumer novox_gitea: nats: no responders available for request");
+  }, async (ms) => { waits.push(ms); }, (l) => said.push(l));
+  assert.equal(calls, 4);
+  assert.deepEqual(waits, [2_000, 5_000, 10_000]);
+  assert.equal(said.length, 4);
+  assert.match(said[0], /no responders available for request.*asking again in 2s/);
+  assert.match(said[3], /is bound after 4 attempts/);
+});
+
+test("a bind that holds the first time says nothing and waits for nothing", async () => {
+  const said: string[] = [];
+  await bindUntilHeld("x", async () => {}, async () => assert.fail("waited"), (l) => said.push(l));
+  assert.deepEqual(said, []);
+});
+
+test("the wait grows to 30 s and stays", () => {
+  assert.deepEqual([1, 2, 3, 4, 50].map(bindWait), [2_000, 5_000, 10_000, 30_000, 30_000]);
+});
+
+test("the binds start without being waited for: the caller returns while a bind keeps refusing, and they go in order", async () => {
+  const order: string[] = [];
+  let finished = false;
+  let refusals = 0;
+  const gate = new Promise<void>((r) => setTimeout(r, 0));
+  bindInBackground([
+    ["first", async () => { refusals++; if (refusals < 3) throw new Error("nats: no responders available for request"); order.push("first"); }],
+    ["second", async () => { order.push("second"); }],
+  ], () => { finished = true; }, async () => { await gate; }, () => {});
+  // Returned at once, with the first bind refused and nothing done.
+  assert.equal(finished, false);
+  assert.deepEqual(order, []);
+  await new Promise((r) => setTimeout(r, 20));
+  assert.deepEqual(order, ["first", "second"]);
+  assert.equal(finished, true);
+});
+
+test("a bind that never holds does not stop the caller from returning", () => {
+  let returned = false;
+  bindInBackground([["x", async () => { throw new Error("no responders"); }]], () => assert.fail("done"),
+    () => new Promise(() => {}), () => {});
+  returned = true;
+  assert.equal(returned, true);
+});

modules/gitea/test/bundle-import.test.ts

@@ -0,0 +1,38 @@
+// **The bundle's import ends while a bind is refused** (novox/hq issue 1793). The launcher imports index.js and then answers
+// node-tools' handshake, which has 10 s before the bundle is killed and never started again. A bind awaited at the top
+// of index.ts, refused for longer, ended the bundle at 00:28:37 on 2026-10-22. Run in a child: the polls' timers keep a
+// process alive, and the child says "imported" and exits as soon as the import resolves.
+
+import { test } from "node:test";
+import assert from "node:assert/strict";
+import { spawn } from "node:child_process";
+import { dirname, join } from "node:path";
+import { fileURLToPath, pathToFileURL } from "node:url";
+
+const here = dirname(fileURLToPath(import.meta.url));
+
+test("importing the bundle resolves while the broker keeps refusing the binds", async () => {
+  const index = pathToFileURL(join(here, "..", "dist", "index.js")).href;
+  const script = `
+    const { useBroker } = await import("@novox/mesh-sdk/messaging");
+    useBroker(() => ({
+      subscribe: async () => { throw new Error("binding gitea's consumer novox_gitea: nats: no responders available for request"); },
+      publish: async () => {}, request: async () => { throw new Error("no responders"); },
+      handle: async () => () => {}, close: async () => {},
+    }));
+    await import(${JSON.stringify(index)});
+    console.log("imported");
+    process.exit(0);`;
+  const env = { ...process.env, MESH_GITEA_TOKEN: "x", MESH_GITEA_URL: "http://127.0.0.1:1" };
+  for (const k of ["MESH_GITEA_STATE_DIR", "MESH_GITEA_ADMIN_USER", "MESH_GITEA_ADMIN_PASSWORD", "MESH_SERVED_MODULE"]) delete env[k];
+  const child = spawn(process.execPath, ["--input-type=module", "-e", script], { cwd: join(here, ".."), env, stdio: ["ignore", "pipe", "pipe"] });
+  let out = "";
+  let err = "";
+  child.stdout.on("data", (d) => (out += d));
+  child.stderr.on("data", (d) => (err += d));
+  const ended = new Promise<void>((r) => child.on("exit", () => r()));
+  const timer = setTimeout(() => child.kill("SIGKILL"), 3000);
+  await ended;
+  clearTimeout(timer);
+  assert.match(out, /imported/, `the import did not resolve within 3 s while a bind was refused; stderr: ${err.slice(0, 600)}`);
+});

modules/gitea/test/guard.test.ts

@@ -339,6 +339,30 @@ test("gitea_merge_pull_request refuses a verdict judged against an earlier trunk
   assert.equal(state.emitted.length, 4, "a merge asks no check");
 });
 
+// **A group member whose gate waits for its group's composed check is not announced again by the merge tool** (novox/hq issue
+// 1106): the member's own check can only say "composed" again, so each refusal's ask brought a verdict that names no trunk and
+// the next merge was refused the same way, for ever. The tool says whose verdict is awaited and asks nothing.
+test("gitea_merge_pull_request asks no check of a member that waits for its group's composed check", async () => {
+  const merge = () => tool("gitea_merge_pull_request").run({ owner: "novox", repo: "mesh-catalog", number: 7, head: REVIEWED });
+  state.trunk = TRUNK;
+  const before = state.emitted.length;
+  state.statuses = [{ context: "mesh/merge-gate", state: "pending", description: "waits for its group's composed check" },
+    { context: "mesh/repo-check", state: "success", description: "pass: y" }];
+  for (let i = 0; i < 3; i++) {
+    await assert.rejects(merge(), (err: Error) => {
+      assert.match(err.message, /waits for its group's composed check/);
+      assert.match(err.message, /mesh-delivery asks it, this tool asks no check/);
+      assert.match(err.message, /Nothing was merged/);
+      return true;
+    });
+  }
+  assert.equal(state.emitted.length, before, "no check is announced again and again");
+  assert.equal(state.merged, false);
+  // Once the composed verdict names the member's unmoved trunk, it merges.
+  state.statuses = judgedAt(TRUNK);
+  assert.equal(((await merge()) as Record<string, unknown>).merged, true);
+});
+
 // A merge lands only the head that was reviewed (novox/hq issue 381): mesh-controller#197 was reviewed at one head,
 // force-pushed, checked green at the new one, and merged there, because the tool merged whatever the head was.
 test("gitea_merge_pull_request merges only the head it is named, and refuses one that is not the head", async () => {

modules/gitea/trunk.ts

@@ -17,7 +17,7 @@
 import { existsSync, mkdirSync, readFileSync, renameSync, writeFileSync } from "node:fs";
 import { dirname } from "node:path";
 import type { CommitStatus, GiteaClient, GiteaPull, HeadCheck } from "./client.js";
-import { CHECK_CONTEXT, REPO_CHECK_CONTEXT } from "./pulls.js";
+import { CHECK_CONTEXT, REPO_CHECK_CONTEXT, WAITS_FOR_COMPOSED } from "./pulls.js";
 
 /** A commit as a verdict names it: its first twelve characters. */
 export function twelve(sha: string | undefined): string {
@@ -58,6 +58,11 @@ export function checkOnItsWay(gate: HeadCheck | undefined): boolean {
   return d.startsWith("waiting in the build queue") || d.startsWith("building on ") || d.startsWith("checking again");
 }
 
+/** Whether a gate says it waits for its group's composed check. */
+export function waitsForComposed(gate: HeadCheck | undefined): boolean {
+  return !!gate && gate.state === "pending" && (gate.description ?? "").startsWith(WAITS_FOR_COMPOSED);
+}
+
 /**
  * The statuses a trunk's move puts back to pending on one open pull request's head, from what the forge holds on it
  * now: none when it has no merge check yet, when a check is pending (on its way, or already out of date — its

modules/mailu/README.md

@@ -16,6 +16,18 @@ installation's are settings on the assignment, `settings set mailu <file>`:
 header from — the address the proxy forwards with. The front's own hostname is the name of its
 `web` route, told to it by the mesh.
 
+## Mail client autoconfiguration
+
+`mailu-automx` answers what mail clients ask for their settings (Thunderbird, Outlook, Apple) for the domains in
+`AMX_MAIL_DOMAINS`, naming the IMAP and SMTP servers and ports of mailu's environment (`AMX_IMAP_ADDRESS`,
+`AMX_IMAP_PORT`, `AMX_SMTP_ADDRESS`, `AMX_SMTP_PORT`). It is `cmd/automx`, a Go program of the standard library
+only in this module's `go.mod` (ADR 0480 clauses 1, 2, 9). It runs in the mirrored alpine image the route proxy's
+`trust` step runs, from the `program-automx` archive mounted read-only at `/opt/mesh`; the container is restarted
+with the archive and with `mailu.env`, which the program reads at each start. It keeps no database. The file the
+Python automx kept in `data-automx` is left where it is on the machine, not deleted by the mesh (ADR 0480
+clause 10). The answers it gave before are the fixtures in `cmd/automx/testdata/captured`, with how they were
+taken (`CAPTURED`) and what the program answers differently (`DIFFERENCES`).
+
 ## Files it does not know
 
 Each directory or archive this module declares says how it treats a file it did not write (ADR 0422, ADR 0416).
@@ -23,7 +35,7 @@ Each directory or archive this module declares says how it treats a file it did
 - `${dir:mesh-state}`: everything in it is the program's; the mesh moves nothing out (ADR 0422 §2).
 - `${dir:state}`: everything in it is the program's; the mesh moves nothing out (ADR 0422 §2).
 - `${dir:grants}`: everything in it is the program's; the mesh moves nothing out (ADR 0422 §2).
-- `${dir:data-automx}`: everything in it is the program's; the mesh moves nothing out (ADR 0422 §2).
+- the archive `program-automx` (`${dir:state}/automx`): everything in it is the program's; the mesh moves nothing out (ADR 0422 §2).
 - `${dir:data-certs}`: everything in it is the program's; the mesh moves nothing out (ADR 0422 §2).
 - `${dir:data-data}`: everything in it is the program's; the mesh moves nothing out (ADR 0422 §2).
 - `${dir:data-dkim}`: everything in it is the program's; the mesh moves nothing out (ADR 0422 §2).

modules/mailu/automx.ts

@@ -1,87 +0,0 @@
-// What the autoconfiguration module (automx2) holds, as stored: its provider, server, domain and server_domain
-// rows, read from its SQLite file inside its container with the sqlite3 program the image carries, opened
-// read-only. Nothing is written and nothing is interpreted: the rows are the evidence for what automx answers
-// to mail clients (novox/hq issue 1630, ADR 0480), and the reason the answer may be empty.
-
-import type { Exec } from "./health.js";
-
-export const AUTOMX_DATABASE = "/data/db.sqlite";
-
-/** The tables read, each with its columns and its ordering. Fixed text: the tool takes no query from its caller. */
-export const AUTOMX_TABLES: Record<string, { columns: string[]; order: string }> = {
-  provider: { columns: ["id", "name", "short_name"], order: "id" },
-  server: { columns: ["id", "prio", "name", "port", "type", "socket_type", "user_name", "authentication"], order: "id" },
-  domain: { columns: ["id", "name", "provider_id", "ldapserver_id"], order: "id" },
-  server_domain: { columns: ["server_id", "domain_id"], order: "domain_id, server_id" },
-};
-
-/**
- * The one statement that reads a table. The sqlite3 in the mailu-automx image does not know `-json` (it answers
- * "sqlite3: Error: unknown option: -json", issue 1630), so each column is printed as `<type>:<hex of its bytes>`, which
- * the oldest sqlite3 prints in its default list mode: no quoting, no separator or newline inside a cell, and a NULL
- * (`null:`) apart from an empty text (`text:`).
- */
-export function selectOf(table: string): string {
-  const t = AUTOMX_TABLES[table];
-  return `SELECT ${t.columns.map((c) => `typeof(${c}) || ':' || hex(${c})`).join(", ")} FROM ${table} ORDER BY ${t.order}`;
-}
-
-export type Row = Record<string, string | number | null>;
-
-export type AutomxRows = {
-  database: string;
-  tables: Record<string, Row[]>;
-  /** servers whose name is empty: the client is told an empty host name for them */
-  empty_names: { id: number; type: string }[];
-  /** tables that could not be read, each with what sqlite3 or docker said on one line */
-  unreadable: Record<string, string>;
-};
-
-/** One cell as selectOf prints it: `<type>:<hex>`; text is the bytes read as UTF-8, a number is read from its text. */
-function cellOf(cell: string): string | number | null {
-  const at = cell.indexOf(":");
-  if (at < 0 || cell.length - at - 1 > 0 && !/^([0-9A-Fa-f]{2})+$/.test(cell.slice(at + 1))) throw new Error(`sqlite3 printed ${JSON.stringify(cell.slice(0, 40))}, not a cell`);
-  const type = cell.slice(0, at);
-  const bytes = Buffer.from(cell.slice(at + 1), "hex").toString("utf8");
-  switch (type) {
-    case "null": return null;
-    case "integer": case "real": return Number(bytes);
-    default: return bytes;
-  }
-}
-
-/** sqlite3 prints one line per row, cells apart by `|`, and nothing at all for a table without rows. */
-export function parseRows(table: string, output: string): Row[] {
-  const columns = AUTOMX_TABLES[table].columns;
-  const rows: Row[] = [];
-  for (const line of output.split("\n")) {
-    if (line.trim() === "") continue;
-    const cells = line.replace(/\r$/, "").split("|");
-    if (cells.length !== columns.length) throw new Error(`sqlite3 printed ${cells.length} cells for ${columns.length} columns`);
-    rows.push(Object.fromEntries(columns.map((c, i) => [c, cellOf(cells[i])])));
-  }
-  return rows;
-}
-
-/** The command that reads one table: read-only, one fixed statement. */
-export function readCommand(table: string): string[] {
-  return ["sqlite3", "-readonly", AUTOMX_DATABASE, selectOf(table)];
-}
-
-/** Read every table; one that fails is said in `unreadable` and does not hide the others. */
-export async function readAutomx(exec: Exec, container: string): Promise<AutomxRows> {
-  const answer: AutomxRows = { database: AUTOMX_DATABASE, tables: {}, empty_names: [], unreadable: {} };
-  for (const table of Object.keys(AUTOMX_TABLES)) {
-    try {
-      answer.tables[table] = parseRows(table, await exec(container, readCommand(table)));
-    } catch (e) {
-      const err = e as { stderr?: string; message?: string };
-      const text = (err?.stderr ?? "").trim() || (err?.message ?? String(e)).trim();
-      answer.unreadable[table] = text.split("\n").map((l) => l.trim()).filter(Boolean)[0] ?? "it failed and said nothing";
-    }
-  }
-  for (const s of answer.tables.server ?? []) {
-    if (String(s.name ?? "").trim() === "") answer.empty_names.push({ id: Number(s.id), type: String(s.type) });
-  }
-  return answer;
-}

modules/mailu/automx/Dockerfile

@@ -1,44 +0,0 @@
-# automx2 — the autoconfig/autodiscover responder, carried by the mailu module as its own
-# artifact: it is a config-baked sidecar of this mail server, not a standalone application
-# (novox/hq ADR 0015 draws that line at applications).
-#
-# The base is named rather than pinned (novox/hq issue 044): declared in module.json's
-# `build.on`. The build context is the module's own directory; every ADD says so.
-ARG PYTHON_BASE
-
-FROM ${PYTHON_BASE}
-RUN apk add --no-cache bash sqlite
-WORKDIR /automx2
-
-ADD automx/files/setupvenv.sh /automx2/setupvenv.sh
-ADD automx/files/start /automx2/start
-ADD automx/files/setup /automx2/setup
-ADD automx/files/setup-db /automx2/setup-db
-ADD automx/files/add-domains /automx2/add-domains
-RUN chmod u+x setupvenv.sh start add-domains setup setup-db
-
-RUN ./setupvenv.sh \
-   && . .venv/bin/activate \
-   && pip install automx2==2021.6
-
-# The launcher `start` expects. In the predecessor's image this wrapper appeared during a build
-# step that never made it into the files this module carries — the image worked and the recipe
-# could not reproduce it. Written here explicitly, verbatim from the proven image, so the build
-# is the whole truth about the image again.
-RUN mkdir -p .venv/scripts && printf '%s\n' \
-   '#!/usr/bin/env bash' \
-   'set -euo pipefail' \
-   '. .venv/bin/activate' \
-   "export FLASK_ENV='production'" \
-   "export FLASK_APP='automx2.server:app'" \
-   'flask "$@"' > .venv/scripts/flask.sh && chmod +x .venv/scripts/flask.sh
-
-ENV AUTOMX2_CONF=/etc/automx2.conf
-ADD automx/files/automx2.conf /etc/automx2.conf
-
-# VOLUME deliberately absent: the anonymous /data volume is exactly what lost db.sqlite on
-# every recreate (measured on novox 2026-08-10). The manifest binds a real directory instead.
-ENTRYPOINT ["/bin/sh"]
-CMD ["./start"]
-
-EXPOSE 4243

modules/mailu/automx/files/add-domains

@@ -1,49 +0,0 @@
-#!/usr/bin/env bash
-set -e
-
-echo "${MAIL_DOMAINS}"
-
-# Split domains into array
-IFS=', ' read -r -a array <<< "${AMX_MAIL_DOMAINS}"
-
-# User configurable section -- START
-PROVIDER_ID=001
-SQL_CMD="";
-
-# Iterate domains resulting from split on second arg
-for element in "${array[@]}"
-do
-  # Set vars
-  DOMAIN=$element
-  PROVIDER_NAME=$DOMAIN
-  PROVIDER_SHORTNAME=$DOMAIN
-
-  # Optional LDAP server
-  #LDAP_SERVER="ldap.${DOMAIN}"
-  # User configurable section -- END
-  s1_id=$((PROVIDER_ID + 1))
-  s2_id=$((PROVIDER_ID + 2))
-  s3_id=$((PROVIDER_ID + 3))
-  dom_id=$((PROVIDER_ID + 4))
-
-  s3_id='NULL'
-
-  SQL_CMD=$(cat <<EOT
-$SQL_CMD
-INSERT INTO provider(id, name, short_name) VALUES(${PROVIDER_ID}, '${PROVIDER_NAME}', '${PROVIDER_SHORTNAME}');
-INSERT INTO server(id, port, type, name, socket_type, user_name, authentication)
-VALUES(${s1_id}, ${AMX_IMAP_PORT}, 'imap', '${AMX_IMAP_SERVER}', 'STARTTLS', '%EMAILADDRESS%', 'password-cleartext');
-INSERT INTO server(id, port, type, name, socket_type, user_name, authentication)
-VALUES(${s2_id}, ${AMX_SMTP_PORT}, 'smtp', '${AMX_SMTP_ADDRESS}', 'STARTTLS', '%EMAILADDRESS%', 'password-cleartext');
-INSERT INTO domain(id, name, provider_id, ldapserver_id) VALUES(${dom_id}, '${DOMAIN}', ${PROVIDER_ID}, ${s3_id});
-INSERT INTO server_domain(server_id, domain_id) VALUES(${s1_id}, ${dom_id});
-INSERT INTO server_domain(server_id, domain_id) VALUES(${s2_id}, ${dom_id});
-EOT
-)
-
-  PROVIDER_ID=$((PROVIDER_ID+10))
-
-done
-
-echo -e ${SQL_CMD}
-echo -e ${SQL_CMD} | sqlite3 /data/db.sqlite

modules/mailu/automx/files/automx2.conf

@@ -1,21 +0,0 @@
-[automx2]
-# A typical production setup would use loglevel = WARNING
-loglevel = WARNING
-# Echo SQL commands into log? Used for debugging.
-db_echo = false
-
-
-# In-memory SQLite database
-# db_uri = sqlite:///:memory:
-
-# SQLite database in a UNIX-like file system
-db_uri = sqlite:////data/db.sqlite
-
-# MySQL database on a remote server. This example does not use an encrypted
-# connection and is therefore *not* recommended for production use.
-#db_uri = mysql://username:password@server.example.com/db
-
-# Number of proxy servers between automx2 and the client (default: 0).
-# If your logs only show 127.0.0.1 or ::1 as the source IP for incoming
-# connections, proxy_count probably needs to be changed.
-proxy_count = 1

modules/mailu/automx/files/setup

@@ -1,12 +0,0 @@
-#!/usr/bin/env bash
-set -e
-
-if [ ! -e /data/db.sqlite ]; then
-  # DB SETUP
-  echo "SETTING UP DB"
-  ./setup-db
-
-  echo "ADDING DOMAINS"
-  # Add the domains
-  ./add-domains
-fi

modules/mailu/automx/files/setup-db

@@ -1,84 +0,0 @@
-#!/usr/bin/env bash
-set -e
-
-# LDAP-Server
-LDAP=$(cat <<EOT
-CREATE TABLE ldapserver(
-  id INT PRIMARY KEY  NOT NULL,
-  name  TEXT NOT NULL,
-  port INT NOT NULL,
-  use_ssl INT NOT NULL,
-  search_base TEXT NOT NULL,
-  search_filter TEXT NOT NULL,
-  attr_uid TEXT NOT NULL,
-  attr_cn TEXT NOT NULL,
-  bind_password TEXT NOT NULL,
-  bind_user TEXT NOT NULL
-);
-EOT
-)
-
-# Provider
-PROVIDER=$(cat <<EOT
-CREATE TABLE provider(
-  id INT PRIMARY KEY  NOT NULL,
-  name  TEXT NOT NULL,
-  short_name TEXT NOT NULL
-);
-EOT
-)
-
-# Server
-SERVER=$(cat <<EOT
-CREATE TABLE server(
-  id INT PRIMARY KEY  NOT NULL,
-  prio INT NOT NULL DEFAULT 10,
-  name  TEXT NOT NULL,
-  port INT NOT NULL,
-  type TEXT NOT NULL,
-  socket_type TEXT NOT NULL,
-  user_name TEXT NOT NULL,
-  authentication TEXT NOT NULL
-);
-EOT
-)
-
-# Domain
-DOMAIN=$(cat <<EOT
-CREATE TABLE domain(
-  id INT PRIMARY KEY  NOT NULL,
-  name  TEXT NOT NULL,
-  provider_id INT NOT NULL,
-  ldapserver_id INT NULL,
-  FOREIGN KEY(ldapserver_id) REFERENCES ldapserver(id),
-  FOREIGN KEY(provider_id) REFERENCES provider(id)
-);
-CREATE UNIQUE INDEX domain_name ON domain(name);
-EOT
-)
-
-# Server-Domain
-SERVER_DOMAIN=$(cat <<EOT
-CREATE TABLE server_domain(
-  server_id INT NOT NULL,
-  domain_id INT NOT NULL,
-  FOREIGN KEY(server_id) REFERENCES server(id),
-  FOREIGN KEY(domain_id) REFERENCES domain(id)
-);
-EOT
-)
-
-## TODO Foreign keys
-
-SQL_CMD=$(cat <<EOT
-$LDAP
-$PROVIDER
-$SERVER
-$DOMAIN
-$SERVER_DOMAIN
-EOT
-)
-
-echo -e ${SQL_CMD}
-
-echo -e ${SQL_CMD} | sqlite3 /data/db.sqlite

modules/mailu/automx/files/setupvenv.sh

@@ -1,38 +0,0 @@
-#!/usr/bin/env bash
-# vim:ts=4:sw=4:noet
-#
-# Creates a Python 3 virtual environment. The target directory can be passed
-# as a parameter. The default path is '.venv' in the current directory.
-
-dir="${1:-.venv}"
-echo "Setup dir $dir"
-
-set -e
-if [ -d "${dir}" ]; then
-	echo >&2 "Directory '${dir}' already exists, exiting."
-	exit 1
-fi
-python3 -m venv "${dir}"
-source "${dir}/bin/activate"
-
-set +e
-pip install -U pip setuptools wheel || true
-
-#set -e
-## vim:tabstop=4:noexpandtab
-##
-## Creates a Python 3 virtual environment. The target directory can be passed
-## as a parameter. The default path is 'venv' in the current directory.
-#
-#dir="${1:-venv}"
-#
-#set -e
-#if [ -d "${dir}" ]; then
-#        echo "Directory '${dir}' already exists, exiting." >&2
-#        exit 1
-#fi
-#python3 -m venv "${dir}"
-#. "${dir}/bin/activate"
-#
-#set +e
-#pip install -U pip setuptools || true

modules/mailu/automx/files/start

@@ -1,8 +0,0 @@
-#!/usr/bin/env bash
-set -e
-
-# Setup
-./setup
-
-# Start
-./.venv/scripts/flask.sh run --host=0.0.0.0 --port=4243

modules/mailu/cmd/automx/automx_test.go

@@ -0,0 +1,260 @@
+package main
+
+import (
+	"bytes"
+	"io"
+	"net/http"
+	"net/http/httptest"
+	"net/url"
+	"os"
+	"path/filepath"
+	"sort"
+	"strconv"
+	"strings"
+	"testing"
+)
+
+// The environment mailu gives automx on novox when the fixtures were captured (files/mailu.env): the mail domain,
+// the route name for both servers, the ports.
+var capturedEnvironment = map[string]string{
+	"AMX_MAIL_DOMAINS": "novox.be",
+	"AMX_IMAP_ADDRESS": "mail.novox.be",
+	"AMX_IMAP_PORT":    "143",
+	"AMX_SMTP_ADDRESS": "mail.novox.be",
+	"AMX_SMTP_PORT":    "587",
+	"AUTOMX_PORT":      "4243",
+}
+
+func getenvOf(m map[string]string) func(string) string {
+	return func(k string) string { return m[k] }
+}
+
+func testServer(t *testing.T) *httptest.Server {
+	t.Helper()
+	cfg, err := configFrom(getenvOf(capturedEnvironment))
+	if err != nil {
+		t.Fatal(err)
+	}
+	s := httptest.NewServer(newHandler(cfg))
+	t.Cleanup(s.Close)
+	return s
+}
+
+const captured = "testdata/captured"
+
+// A captured request: its first line is the method and the URL the client used, then the headers sent, then after
+// a blank line the body sent (CAPTURED, how the files were written).
+type request struct {
+	method, url string
+	header      [][2]string
+	body        string
+}
+
+func readRequest(t *testing.T, stem string) request {
+	t.Helper()
+	raw, err := os.ReadFile(stem + ".request")
+	if err != nil {
+		t.Fatal(err)
+	}
+	head, body, hasBody := strings.Cut(string(raw), "\n\n")
+	if !hasBody {
+		head = strings.TrimSuffix(head, "\n")
+	}
+	lines := strings.Split(head, "\n")
+	method, target, ok := strings.Cut(lines[0], " ")
+	if !ok {
+		t.Fatalf("%s: first line %q is not METHOD URL", stem, lines[0])
+	}
+	r := request{method: method, url: target, body: body}
+	for _, l := range lines[1:] {
+		k, v, ok := strings.Cut(l, ": ")
+		if !ok {
+			t.Fatalf("%s: header line %q", stem, l)
+		}
+		r.header = append(r.header, [2]string{k, v})
+	}
+	return r
+}
+
+func readFile(t *testing.T, name string) string {
+	t.Helper()
+	b, err := os.ReadFile(name)
+	if err != nil {
+		t.Fatal(err)
+	}
+	return string(b)
+}
+
+// A fixture asks for its body to differ only where DIFFERENCES says so: a stem with a .corrected.body is answered
+// with that, and its Content-Length is the corrected body's.
+func expectedBody(t *testing.T, stem string) (body string, corrected bool) {
+	if _, err := os.Stat(stem + ".corrected.body"); err == nil {
+		return readFile(t, stem+".corrected.body"), true
+	}
+	return readFile(t, stem+".body"), false
+}
+
+func TestEveryCapturedRequestIsAnsweredAsAutomxAnswered(t *testing.T) {
+	stems, _ := filepath.Glob(filepath.Join(captured, "*.request"))
+	if len(stems) < 17 {
+		t.Fatalf("%d captured requests found, want the 17 of issue 1630", len(stems))
+	}
+	srv := testServer(t)
+	client := &http.Client{CheckRedirect: func(*http.Request, []*http.Request) error { return http.ErrUseLastResponse }}
+	for _, f := range stems {
+		stem := strings.TrimSuffix(f, ".request")
+		t.Run(filepath.Base(stem), func(t *testing.T) {
+			rq := readRequest(t, stem)
+			u, err := url.Parse(rq.url)
+			if err != nil {
+				t.Fatal(err)
+			}
+			// The mesh's route reaches the program with the client's name and scheme; the program is told them in
+			// the forwarding headers (the Location of fixture 14 holds them).
+			req, err := http.NewRequest(rq.method, srv.URL+u.RequestURI(), strings.NewReader(rq.body))
+			if err != nil {
+				t.Fatal(err)
+			}
+			for _, h := range rq.header {
+				req.Header.Set(h[0], h[1])
+			}
+			req.Host = u.Host
+			req.Header.Set("X-Forwarded-Proto", u.Scheme)
+			resp, err := client.Do(req)
+			if err != nil {
+				t.Fatal(err)
+			}
+			defer resp.Body.Close()
+			got, _ := io.ReadAll(resp.Body)
+
+			if want := strings.TrimRight(readFile(t, stem+".status"), "\n"); resp.Status != want {
+				t.Errorf("status %q, captured %q", resp.Status, want)
+			}
+			wantBody, corrected := expectedBody(t, stem)
+			if !bytes.Equal(got, []byte(wantBody)) {
+				t.Errorf("body\n%q\nwant\n%q", got, wantBody)
+			}
+			want := map[string]string{}
+			for _, l := range strings.Split(strings.TrimRight(readFile(t, stem+".headers"), "\n"), "\n") {
+				if k, v, ok := strings.Cut(l, ": "); ok {
+					want[k] = v
+				}
+			}
+			if corrected {
+				want["Content-Length"] = strconv.Itoa(len(wantBody))
+			}
+			// Server names the Python program that answered (DIFFERENCES); Date changes on every request.
+			delete(want, "Server")
+			var extra []string
+			for k := range resp.Header {
+				if k == "Date" || k == "Server" {
+					continue
+				}
+				if _, ok := want[k]; !ok {
+					extra = append(extra, k)
+				}
+			}
+			sort.Strings(extra)
+			if len(extra) > 0 {
+				t.Errorf("headers the capture does not have: %v", extra)
+			}
+			for k, v := range want {
+				if g := resp.Header.Get(k); g != v {
+					t.Errorf("header %s: %q, captured %q", k, g, v)
+				}
+			}
+		})
+	}
+}
+
+// DIFFERENCES: the one change clause 9 of ADR 0480 allows to an answer is the IMAP server's name, written from
+// AMX_IMAP_ADDRESS where automx wrote an empty one.
+func TestTheOnlyCorrectedAnswerIsTheImapServersName(t *testing.T) {
+	corrected, _ := filepath.Glob(filepath.Join(captured, "*.corrected.body"))
+	if len(corrected) != 3 {
+		t.Fatalf("%d corrected answers, want the three Thunderbird answers (01, 02, 03): %v", len(corrected), corrected)
+	}
+	for _, c := range corrected {
+		orig := readFile(t, strings.TrimSuffix(c, ".corrected.body")+".body")
+		want := strings.Replace(orig, "<hostname />", "<hostname>mail.novox.be</hostname>", 1)
+		if strings.Contains(want, "<hostname />") || want == orig {
+			t.Fatalf("%s: the capture has no single empty hostname to correct", c)
+		}
+		if got := readFile(t, c); got != want {
+			t.Errorf("%s differs from the capture by more than the IMAP server's name", c)
+		}
+	}
+}
+
+func TestTheImapServerIsNamedFromTheImapAddress(t *testing.T) {
+	env := map[string]string{}
+	for k, v := range capturedEnvironment {
+		env[k] = v
+	}
+	env["AMX_IMAP_ADDRESS"] = "imap.example.test"
+	env["AMX_IMAP_PORT"] = "993"
+	cfg, err := configFrom(getenvOf(env))
+	if err != nil {
+		t.Fatal(err)
+	}
+	rec := httptest.NewRecorder()
+	newHandler(cfg).ServeHTTP(rec, httptest.NewRequest("GET", "/mail/config-v1.1.xml?emailaddress=a@novox.be", nil))
+	body := rec.Body.String()
+	if !strings.Contains(body, `<incomingServer type="imap"><hostname>imap.example.test</hostname><port>993</port>`) {
+		t.Errorf("incoming server not named from AMX_IMAP_ADDRESS and AMX_IMAP_PORT: %s", body)
+	}
+	if !strings.Contains(body, `<outgoingServer type="smtp"><hostname>mail.novox.be</hostname><port>587</port>`) {
+		t.Errorf("outgoing server changed: %s", body)
+	}
+}
+
+func TestEveryDomainOfTheEnvironmentIsServed(t *testing.T) {
+	env := map[string]string{}
+	for k, v := range capturedEnvironment {
+		env[k] = v
+	}
+	env["AMX_MAIL_DOMAINS"] = "novox.be, example.test"
+	cfg, err := configFrom(getenvOf(env))
+	if err != nil {
+		t.Fatal(err)
+	}
+	h := newHandler(cfg)
+	for _, d := range []string{"novox.be", "example.test"} {
+		rec := httptest.NewRecorder()
+		h.ServeHTTP(rec, httptest.NewRequest("GET", "/mail/config-v1.1.xml?emailaddress=a@"+d, nil))
+		if rec.Code != 200 || !strings.Contains(rec.Body.String(), "<domain>"+d+"</domain>") {
+			t.Errorf("%s: %d %s", d, rec.Code, rec.Body.String())
+		}
+	}
+}
+
+// Failure is loud (issue 1631 AC2): a missing variable stops the program and its name is in the message.
+func TestAMissingVariableStopsTheProgramNamingIt(t *testing.T) {
+	for name := range capturedEnvironment {
+		t.Run(name, func(t *testing.T) {
+			env := map[string]string{}
+			for k, v := range capturedEnvironment {
+				env[k] = v
+			}
+			delete(env, name)
+			if _, err := configFrom(getenvOf(env)); err == nil || !strings.Contains(err.Error(), name) {
+				t.Fatalf("configFrom without %s: %v", name, err)
+			}
+			var out bytes.Buffer
+			if code := run(getenvOf(env), &out, func(*http.Server) error { t.Fatal("served without its configuration"); return nil }); code == 0 || !strings.Contains(out.String(), name) {
+				t.Errorf("run without %s: exit %d, said %q", name, code, out.String())
+			}
+		})
+	}
+}
+
+func TestAPortThatIsNotANumberIsRefusedNamingIt(t *testing.T) {
+	env := map[string]string{}
+	for k, v := range capturedEnvironment {
+		env[k] = v
+	}
+	env["AMX_SMTP_PORT"] = "submission"
+	if _, err := configFrom(getenvOf(env)); err == nil || !strings.Contains(err.Error(), "AMX_SMTP_PORT") {
+		t.Fatalf("got %v", err)
+	}
+}

modules/mailu/cmd/automx/main.go

@@ -0,0 +1,290 @@
+// automx answers mail clients' autoconfiguration requests for the domains mailu serves: Thunderbird's
+// /mail/config-v1.1.xml, Outlook's /autodiscover/autodiscover.xml and Apple's /mobileconfig. It replaces the
+// Python automx2 image (novox/hq ADR 0480 clauses 1, 9, 10): standard library only, no database, every answer
+// made from mailu's environment at the start, so a change of the domains, servers or ports reaches clients when
+// the program is started again.
+//
+// The answers are those automx2 2021.6 gave, captured in testdata/captured (CAPTURED, issue 1630), byte for byte,
+// but for the one thing testdata/captured/DIFFERENCES lists: the IMAP server is named, from AMX_IMAP_ADDRESS.
+package main
+
+import (
+	"encoding/xml"
+	"errors"
+	"fmt"
+	"html"
+	"io"
+	"net/http"
+	"os"
+	"strconv"
+	"strings"
+	"time"
+)
+
+// The environment, all required: a program that cannot say which servers clients use must not run (failure is
+// loud). The names are those of mailu's files/mailu.env.
+const (
+	envDomains  = "AMX_MAIL_DOMAINS"
+	envImapHost = "AMX_IMAP_ADDRESS"
+	envImapPort = "AMX_IMAP_PORT"
+	envSmtpHost = "AMX_SMTP_ADDRESS"
+	envSmtpPort = "AMX_SMTP_PORT"
+	envListen   = "AUTOMX_PORT"
+)
+
+type config struct {
+	domains            []string
+	imapHost, smtpHost string
+	imapPort, smtpPort int
+	listen             int
+}
+
+// configFrom reads the environment; it names every variable that is missing or wrong.
+func configFrom(getenv func(string) string) (config, error) {
+	var c config
+	var problems []string
+	text := func(name string) string {
+		v := strings.TrimSpace(getenv(name))
+		if v == "" {
+			problems = append(problems, name+" is not set")
+		}
+		return v
+	}
+	port := func(name string) int {
+		v := text(name)
+		if v == "" {
+			return 0
+		}
+		n, err := strconv.Atoi(v)
+		if err != nil || n < 1 || n > 65535 {
+			problems = append(problems, fmt.Sprintf("%s is %q, not a port number", name, v))
+			return 0
+		}
+		return n
+	}
+	// automx2's add-domains split the list at commas and spaces.
+	for _, d := range strings.FieldsFunc(text(envDomains), func(r rune) bool { return r == ',' || r == ' ' }) {
+		c.domains = append(c.domains, strings.ToLower(d))
+	}
+	c.imapHost, c.imapPort = text(envImapHost), port(envImapPort)
+	c.smtpHost, c.smtpPort = text(envSmtpHost), port(envSmtpPort)
+	c.listen = port(envListen)
+	if len(problems) > 0 {
+		return config{}, errors.New(strings.Join(problems, "; "))
+	}
+	return c, nil
+}
+
+// What automx2 sent (Werkzeug), copied from the capture.
+const (
+	htmlType = "text/html; charset=utf-8"
+	xmlType  = "application/xml"
+
+	notFound = `<!doctype html>
+<html lang=en>
+<title>404 Not Found</title>
+<h1>Not Found</h1>
+<p>The requested URL was not found on the server. If you entered the URL manually please check your spelling and try again.</p>
+`
+	notAllowed = `<!doctype html>
+<html lang=en>
+<title>405 Method Not Allowed</title>
+<h1>Method Not Allowed</h1>
+<p>The method is not allowed for the requested URL.</p>
+`
+	internalError = `<!doctype html>
+<html lang=en>
+<title>500 Internal Server Error</title>
+<h1>Internal Server Error</h1>
+<p>The server encountered an internal error and was unable to complete your request. Either the server is overloaded or there is an error in the application.</p>
+`
+	badRequest = `<!doctype html>
+<html lang=en>
+<title>400 Bad Request</title>
+<h1>Bad Request</h1>
+<p>The browser (or proxy) sent a request that this server could not understand.</p>
+`
+	index = `<html><body>Show Thunderbird-style <a href="/mail/config-v1.1.xml?emailaddress=abc@example.com">XML configuration</a> for abc@example.com.</body></html>`
+)
+
+func reply(w http.ResponseWriter, status int, contentType, body string) {
+	w.Header().Set("Content-Type", contentType)
+	if status != http.StatusNoContent {
+		w.Header().Set("Content-Length", strconv.Itoa(len(body)))
+	}
+	w.WriteHeader(status)
+	if status != http.StatusNoContent {
+		io.WriteString(w, body)
+	}
+}
+
+func missing(w http.ResponseWriter, argument string) {
+	reply(w, http.StatusBadRequest, htmlType, fmt.Sprintf("Missing request argument %q", argument))
+}
+
+type handler struct{ config }
+
+func newHandler(c config) http.Handler { return handler{c} }
+
+func (h handler) ServeHTTP(w http.ResponseWriter, r *http.Request) {
+	var allow string
+	switch r.URL.Path {
+	case "/", "/mail/config-v1.1.xml", "/mobileconfig/":
+		allow = "GET, HEAD, OPTIONS"
+	case "/autodiscover/autodiscover.xml":
+		allow = "POST, OPTIONS"
+	case "/mobileconfig":
+		h.redirectToSlash(w, r)
+		return
+	default:
+		reply(w, http.StatusNotFound, htmlType, notFound)
+		return
+	}
+	if r.Method == http.MethodOptions {
+		w.Header().Set("Allow", allow)
+		reply(w, http.StatusOK, htmlType, "")
+		return
+	}
+	if !strings.Contains(", "+allow+",", ", "+r.Method+",") {
+		w.Header().Set("Allow", allow)
+		reply(w, http.StatusMethodNotAllowed, htmlType, notAllowed)
+		return
+	}
+	switch r.URL.Path {
+	case "/":
+		reply(w, http.StatusOK, htmlType, index)
+	case "/mail/config-v1.1.xml":
+		h.byAddress(w, r.URL.Query().Get("emailaddress"), r.URL.Query().Has("emailaddress"), "emailaddress", h.thunderbird)
+	case "/mobileconfig/":
+		h.byAddress(w, r.URL.Query().Get("emailaddress"), r.URL.Query().Has("emailaddress"), "emailaddress", h.apple)
+	case "/autodiscover/autodiscover.xml":
+		h.outlook(w, r)
+	}
+}
+
+// byAddress is the part the three protocols share: the address is required, and a domain that is not served is
+// answered with no content.
+func (h handler) byAddress(w http.ResponseWriter, address string, present bool, argument string, answer func(http.ResponseWriter, int, string)) {
+	if !present {
+		missing(w, argument)
+		return
+	}
+	domain := strings.ToLower(address[strings.LastIndex(address, "@")+1:])
+	for i, d := range h.domains {
+		if strings.Contains(address, "@") && d == domain {
+			answer(w, i, d)
+			return
+		}
+	}
+	reply(w, http.StatusNoContent, htmlType, "")
+}
+
+// thunderbird is Mozilla's autoconfiguration, config-v1.1. The provider id is automx2's: the provider rows were
+// numbered 1, 11, 21 ... by add-domains.
+func (h handler) thunderbird(w http.ResponseWriter, i int, domain string) {
+	d := html.EscapeString(domain)
+	server := func(kind, host string, port int) string {
+		tag := map[string]string{"imap": "incomingServer", "smtp": "outgoingServer"}[kind]
+		return fmt.Sprintf(`<%s type="%s"><hostname>%s</hostname><port>%d</port><socketType>STARTTLS</socketType>`+
+			`<username>%%EMAILADDRESS%%</username><authentication>password-cleartext</authentication></%s>`,
+			tag, kind, html.EscapeString(host), port, tag)
+	}
+	body := fmt.Sprintf(`<clientConfig version="1.1"><emailProvider id="automx2-%d"><identity /><domain>%s</domain>`+
+		`<displayName>%s</displayName><displayShortName>%s</displayShortName>%s%s</emailProvider></clientConfig>`,
+		1+10*i, d, d, d, server("imap", h.imapHost, h.imapPort), server("smtp", h.smtpHost, h.smtpPort))
+	reply(w, http.StatusOK, xmlType, body)
+}
+
+// apple and outlook answer a served domain as automx2 did: with an error (DIFFERENCES, issue 1691). They are
+// kept as captured here, so the move to a Go program changes the answer only where clause 9 says.
+func (h handler) apple(w http.ResponseWriter, _ int, _ string) {
+	reply(w, http.StatusBadRequest, htmlType, badRequest)
+}
+
+// redirectToSlash is Werkzeug's strict-slash redirect, to the name the client used.
+func (h handler) redirectToSlash(w http.ResponseWriter, r *http.Request) {
+	scheme := "http"
+	if p := strings.TrimSpace(strings.Split(r.Header.Get("X-Forwarded-Proto"), ",")[0]); p != "" {
+		scheme = p
+	}
+	host := r.Host
+	if f := strings.TrimSpace(strings.Split(r.Header.Get("X-Forwarded-Host"), ",")[0]); f != "" {
+		host = f
+	}
+	target := scheme + "://" + host + "/mobileconfig/"
+	if r.URL.RawQuery != "" {
+		target += "?" + r.URL.RawQuery
+	}
+	w.Header().Set("Location", target)
+	e := html.EscapeString(target)
+	reply(w, http.StatusPermanentRedirect, htmlType, fmt.Sprintf(`<!doctype html>
+<html lang=en>
+<title>Redirecting...</title>
+<h1>Redirecting...</h1>
+<p>You should be redirected automatically to the target URL: <a href="%s">%s</a>. If not, click the link.
+`, e, e))
+}
+
+// outlook reads the request's EMailAddress anywhere in the document, whatever the namespace. A body that is not
+// one XML document is automx2's 500.
+func (h handler) outlook(w http.ResponseWriter, r *http.Request) {
+	dec := xml.NewDecoder(io.LimitReader(r.Body, 1<<20))
+	var address string
+	var found, elements, inAddress = false, 0, false
+	for {
+		tok, err := dec.Token()
+		if err == io.EOF {
+			break
+		}
+		if err != nil {
+			reply(w, http.StatusInternalServerError, htmlType, internalError)
+			return
+		}
+		switch t := tok.(type) {
+		case xml.StartElement:
+			elements++
+			if t.Name.Local == "EMailAddress" && !found {
+				inAddress = true
+			}
+		case xml.CharData:
+			if inAddress {
+				address += string(t)
+			}
+		case xml.EndElement:
+			if inAddress && t.Name.Local == "EMailAddress" {
+				inAddress, found = false, true
+			}
+		}
+	}
+	if elements == 0 {
+		reply(w, http.StatusInternalServerError, htmlType, internalError)
+		return
+	}
+	h.byAddress(w, strings.TrimSpace(address), found, "EMailAddress", func(w http.ResponseWriter, _ int, _ string) {
+		reply(w, http.StatusInternalServerError, htmlType, internalError)
+	})
+}
+
+// run is the program: configuration from the environment, then serving. A configuration that cannot be read is
+// said, with the variables' names, and the exit code is not zero.
+func run(getenv func(string) string, out io.Writer, serve func(*http.Server) error) int {
+	c, err := configFrom(getenv)
+	if err != nil {
+		fmt.Fprintf(out, "automx: %v\n", err)
+		return 1
+	}
+	srv := &http.Server{
+		Addr:              ":" + strconv.Itoa(c.listen),
+		Handler:           newHandler(c),
+		ReadHeaderTimeout: 10 * time.Second,
+	}
+	if err := serve(srv); err != nil {
+		fmt.Fprintf(out, "automx: %v\n", err)
+		return 1
+	}
+	return 0
+}
+
+func main() {
+	os.Exit(run(os.Getenv, os.Stderr, func(s *http.Server) error { return s.ListenAndServe() }))
+}

modules/mailu/cmd/automx/manifest_test.go

@@ -0,0 +1,202 @@
+package main
+
+import (
+	"encoding/json"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+)
+
+// The digest route-proxy's `trust` artifact runs: one alpine image is held for both (ADR 0480 clause 3). It was
+// read from mesh-controller's modules/route-proxy/module.json and is the one hello-web runs.
+const alpineDigest = "alpine@sha256:28bd5fe8b56d1bd048e5babf5b10710ebe0bae67db86916198a6eec434943f8b"
+
+type manifest struct {
+	Tools     []string         `json:"tools"`
+	Resources []map[string]any `json:"resources"`
+	Listens   []struct {
+		Name string `json:"name"`
+		Port int    `json:"port"`
+	} `json:"listens"`
+	Data struct {
+		Own []struct {
+			ID   string `json:"id"`
+			Path string `json:"path"`
+		} `json:"own"`
+	} `json:"data"`
+	Build map[string]any `json:"build"`
+}
+
+func readManifest(t *testing.T) manifest {
+	t.Helper()
+	var m manifest
+	if err := json.Unmarshal([]byte(readFile(t, "../../module.json")), &m); err != nil {
+		t.Fatal(err)
+	}
+	return m
+}
+
+func (m manifest) resource(id string) map[string]any {
+	for _, r := range m.Resources {
+		if r["id"] == id {
+			return r
+		}
+	}
+	return nil
+}
+
+func (m manifest) artifact(name string) map[string]any {
+	arts, _ := m.Build["artifacts"].([]any)
+	for _, a := range arts {
+		if a, _ := a.(map[string]any); a["name"] == name {
+			return a
+		}
+	}
+	return nil
+}
+
+func strs(v any) []string {
+	var out []string
+	l, _ := v.([]any)
+	for _, e := range l {
+		out = append(out, e.(string))
+	}
+	return out
+}
+
+// Clause 2: a Go bundle from a package of the module's own go.mod, stamped for alpine; an archive resource unpacks
+// it; the container mounts that read-only at /opt/mesh, calls the program there and names the archive in restart-on.
+func TestAutomxIsAGoBundleForAlpineMountedAtOptMeshAndRestartedWithIt(t *testing.T) {
+	m := readManifest(t)
+	b := m.artifact("automx")
+	if b == nil || b["kind"] != "bundle" || b["language"] != "go" || b["system"] != "alpine" || b["from"] != "cmd/automx" || b["binary"] != "automx" {
+		t.Fatalf("artifact automx is %v, want a Go bundle, system alpine, from cmd/automx, binary automx", b)
+	}
+	if _, loads := b["loads"]; loads {
+		t.Errorf("a program run in a container is not loaded by the tool runner: %v", b["loads"])
+	}
+	goMod := readFile(t, "../../go.mod")
+	if !strings.HasPrefix(goMod, "module mailu\n") {
+		t.Errorf("the module's go.mod is %q; the bundle is a package of it", goMod)
+	}
+	if _, err := os.Stat(filepath.Join("..", "..", filepath.FromSlash(b["from"].(string)), "main.go")); err != nil {
+		t.Errorf("the bundle's package is not in the module's directory: %v", err)
+	}
+
+	unpack := m.resource("program-automx")
+	if unpack == nil || unpack["type"] != "archive" || unpack["artifact"] != "automx" || unpack["unknown"] != "program" {
+		t.Fatalf("resource program-automx is %v, want an archive of the bundle", unpack)
+	}
+	where, _ := unpack["path"].(string)
+	c := m.resource("automx")
+	if c == nil || c["type"] != "container" {
+		t.Fatalf("no container automx: %v", c)
+	}
+	if got := strs(c["volumes"]); len(got) != 1 || got[0] != where+":/opt/mesh:ro" {
+		t.Errorf("volumes %v, want the archive's place %q mounted read-only at /opt/mesh", got, where)
+	}
+	if got := strs(c["args"]); len(got) != 1 || got[0] != "/opt/mesh/automx" {
+		t.Errorf("args %v, want the program at /opt/mesh/automx", got)
+	}
+	if !contains(strs(c["restart-on"]), "program-automx") {
+		t.Errorf("restart-on %v does not name the archive: a new program would leave the old one running", c["restart-on"])
+	}
+	if !contains(strs(c["restart-on"]), "config-env") {
+		t.Errorf("restart-on %v does not name the environment file the program reads at its start", c["restart-on"])
+	}
+}
+
+// Clause 3: an upstream artifact pinned by digest, the one route-proxy's trust artifact runs; never an image
+// built from a Dockerfile (clause 11), and no Dockerfile or base to build one on.
+func TestAutomxRunsTheMirroredAlpineImageAndBuildsNoImage(t *testing.T) {
+	m := readManifest(t)
+	c := m.resource("automx")
+	base := m.artifact(c["artifact"].(string))
+	if base == nil || base["kind"] != "upstream" || base["from"] != alpineDigest {
+		t.Errorf("the container runs artifact %v, want an upstream artifact %s", base, alpineDigest)
+	}
+	if _, has := c["image"]; has {
+		t.Errorf("the container names an image %v, fetched from a registry on the machine", c["image"])
+	}
+	for _, a := range m.Build["artifacts"].([]any) {
+		if a.(map[string]any)["kind"] == "image" {
+			t.Errorf("artifact %v is an image built from a Dockerfile (ADR 0452)", a)
+		}
+	}
+	if _, has := m.Build["on"]; has {
+		t.Errorf("build.on %v: nothing is built on a base image any more", m.Build["on"])
+	}
+	for _, gone := range []string{"../../automx/Dockerfile", "../../automx"} {
+		if _, err := os.Stat(gone); err == nil {
+			t.Errorf("%s is still there", gone)
+		}
+	}
+}
+
+// Clause 1: the container stays as it was for what watches it.
+func TestAutomxKeepsItsContainerNameNetworkPortAndEnvironment(t *testing.T) {
+	m := readManifest(t)
+	c := m.resource("automx")
+	if c["name"] != "mailu-automx" || c["network"] != "mailu" {
+		t.Errorf("container %v on network %v, want mailu-automx on mailu", c["name"], c["network"])
+	}
+	if got := strs(c["ports"]); len(got) != 1 || got[0] != "4243" {
+		t.Errorf("ports %v, want 4243", got)
+	}
+	if got := strs(c["env-file"]); len(got) != 1 || got[0] != "${dir:state}/mailu.env" {
+		t.Errorf("env-file %v, want mailu's environment only", got)
+	}
+	if !strings.Contains("\n"+readFile(t, "../../files/mailu.env"), "\nAUTOMX_PORT=4243\n") {
+		t.Errorf("mailu.env does not say AUTOMX_PORT=4243")
+	}
+	for _, l := range m.Listens {
+		if l.Name == "autoconfig" && l.Port == 4243 {
+			return
+		}
+	}
+	t.Errorf("the autoconfig endpoint is not on 4243: %v", m.Listens)
+}
+
+// Clause 10: a database nobody reads is gone from the declaration (the file stays on disk), and the container says
+// whether it answers with one captured request and its captured answer.
+func TestAutomxKeepsNoDatabaseAndSaysWhetherItAnswers(t *testing.T) {
+	m := readManifest(t)
+	if m.resource("data-automx") != nil {
+		t.Errorf("the data-automx directory is still declared")
+	}
+	for _, it := range m.Data.Own {
+		if strings.Contains(it.ID, "automx") || strings.Contains(it.Path, "data-automx") {
+			t.Errorf("data item %v is still declared", it)
+		}
+	}
+	for _, tool := range m.Tools {
+		if tool == "automx_rows" {
+			t.Errorf("the tool automx_rows reads a database that no longer exists (issue 1700)")
+		}
+	}
+	h, _ := m.resource("automx")["health"].(map[string]any)
+	if h["kind"] != "http" || h["endpoint"] != "autoconfig" {
+		t.Fatalf("health %v, want an http check of the autoconfig endpoint", h)
+	}
+	// The request is captured fixture 06 and the answer its status and body.
+	rq := readRequest(t, filepath.Join(captured, "06-autoconfig-root"))
+	if u := strings.SplitN(rq.url, "/", 4); "/"+u[len(u)-1] != h["path"] || rq.method != "GET" {
+		t.Errorf("health path %v is not captured request %s %s", h["path"], rq.method, rq.url)
+	}
+	if st := readFile(t, filepath.Join(captured, "06-autoconfig-root.status")); !strings.HasPrefix(st, "200 ") || h["status"] != float64(200) {
+		t.Errorf("health status %v, captured %q", h["status"], st)
+	}
+	if body, _ := h["body"].(string); body == "" || !strings.Contains(readFile(t, filepath.Join(captured, "06-autoconfig-root.body")), body) {
+		t.Errorf("health body %q is not in the captured answer", h["body"])
+	}
+}
+
+func contains(l []string, s string) bool {
+	for _, e := range l {
+		if e == s {
+			return true
+		}
+	}
+	return false
+}

modules/mailu/cmd/automx/testdata/captured/01-thunderbird-autoconfig-ok.corrected.body

@@ -0,0 +1 @@
+<clientConfig version="1.1"><emailProvider id="automx2-1"><identity /><domain>novox.be</domain><displayName>novox.be</displayName><displayShortName>novox.be</displayShortName><incomingServer type="imap"><hostname>mail.novox.be</hostname><port>143</port><socketType>STARTTLS</socketType><username>%EMAILADDRESS%</username><authentication>password-cleartext</authentication></incomingServer><outgoingServer type="smtp"><hostname>mail.novox.be</hostname><port>587</port><socketType>STARTTLS</socketType><username>%EMAILADDRESS%</username><authentication>password-cleartext</authentication></outgoingServer></emailProvider></clientConfig>

modules/mailu/cmd/automx/testdata/captured/02-thunderbird-on-automx-name.corrected.body

@@ -0,0 +1 @@
+<clientConfig version="1.1"><emailProvider id="automx2-1"><identity /><domain>novox.be</domain><displayName>novox.be</displayName><displayShortName>novox.be</displayShortName><incomingServer type="imap"><hostname>mail.novox.be</hostname><port>143</port><socketType>STARTTLS</socketType><username>%EMAILADDRESS%</username><authentication>password-cleartext</authentication></incomingServer><outgoingServer type="smtp"><hostname>mail.novox.be</hostname><port>587</port><socketType>STARTTLS</socketType><username>%EMAILADDRESS%</username><authentication>password-cleartext</authentication></outgoingServer></emailProvider></clientConfig>

modules/mailu/cmd/automx/testdata/captured/03-thunderbird-on-autodiscover-name.corrected.body

@@ -0,0 +1 @@
+<clientConfig version="1.1"><emailProvider id="automx2-1"><identity /><domain>novox.be</domain><displayName>novox.be</displayName><displayShortName>novox.be</displayShortName><incomingServer type="imap"><hostname>mail.novox.be</hostname><port>143</port><socketType>STARTTLS</socketType><username>%EMAILADDRESS%</username><authentication>password-cleartext</authentication></incomingServer><outgoingServer type="smtp"><hostname>mail.novox.be</hostname><port>587</port><socketType>STARTTLS</socketType><username>%EMAILADDRESS%</username><authentication>password-cleartext</authentication></outgoingServer></emailProvider></clientConfig>

modules/mailu/cmd/automx/testdata/captured/CAPTURED

@@ -14,6 +14,13 @@ Each capture is four files with the same stem:
 The files were written from the tool's answers by a script (not by retyping each by hand); the tool returns JSON, so
 a later capture that wants exact bytes needs a tool that writes the answer to a file.
 
+A fixture is not the bytes on the wire in four ways (issue 1701):
+  - Date is left out of .headers (it changes on every request).
+  - .headers are sorted by name.
+  - .status holds the tool's status line without the protocol (the tool reported "HTTP/1.1" apart).
+  - Header names are as the tool's Go HTTP client gives them (canonical case, after the mesh's route), not as
+    automx wrote them.
+
 What the answers show (read, not guessed):
   - GET /mail/config-v1.1.xml?emailaddress=... answers 200 application/xml on all three names; same body on each.
   - The incoming (IMAP) server has an EMPTY hostname (<hostname />), port 143, STARTTLS. The outgoing server is
@@ -27,4 +34,8 @@ What the answers show (read, not guessed):
   - GET /mobileconfig?emailaddress=... redirects (308) to /mobileconfig/?emailaddress=...; for the served domain that
     answers 400 with no profile; unknown domain 204; no emailaddress 400.
   - GET / on autoconfig answers 200 with a small HTML page linking to an example address.
-  - Not captured: the database rows (no mesh tool reads /data/db.sqlite; issue 1630 AC4).
+  - The database rows behind the answers were read once with the tool automx_rows (mailu, issue 1630 AC4), which
+    ran sqlite3 read-only inside mailu-automx. The tool is retired with the database (ADR 0480 clause 10, issue
+    1700), so nothing reads the rows any more; the answers above are what they gave.
+  - The Go program that replaced automx2 answers these requests as they are captured here, but for what
+    DIFFERENCES lists.

modules/mailu/cmd/automx/testdata/captured/DIFFERENCES

@@ -0,0 +1,25 @@
+How the Go program (modules/mailu/cmd/automx) answers differently from the automx2 2021.6 answers captured here
+(issue 1631, ADR 0480 clause 9). automx_test.go answers every .request and compares with the capture; this is the
+whole list of where the two differ.
+
+1. The IMAP server's name (the one correction clause 9 allows). automx2 answered <hostname /> for the incoming
+   server because its setup script read AMX_IMAP_SERVER, which mailu's environment does not set (it sets
+   AMX_IMAP_ADDRESS). The program names it from AMX_IMAP_ADDRESS (and the port from AMX_IMAP_PORT).
+   Fixtures 01, 02 and 03 each have a <stem>.corrected.body: the capture with <hostname /> replaced by
+   <hostname>mail.novox.be</hostname> and nothing else (a test checks that). Their Content-Length is that file's
+   size, 634, not the captured 612.
+
+2. The Server header. The capture has "Server: Werkzeug/3.1.9 Python/3.9.5". The program sends no Server header, and
+   Date, which the captured .headers leave out, is sent as net/http sends it.
+
+Not changed, though issue 1691 calls them faults; they are kept as captured because clause 9 allows only the first
+correction, and their cause is not known from the capture:
+  - Outlook's POST for a served domain (fixture 08) answers 500.
+  - Apple's /mobileconfig/ for a served domain (fixture 15) answers 400 with no profile.
+  Whether automx2 failed there because of the empty IMAP name is a guess; it can be tested only against the
+  running program after the delivery. Issue 1691 stays open for them.
+
+Assumptions the fixtures do not settle (answers to requests nobody captured): a domain is compared in lower case;
+an address without "@" is an unknown domain (204); HEAD is answered as GET and OPTIONS with the Allow header and
+no body; the redirect of /mobileconfig takes its scheme and host from X-Forwarded-Proto and X-Forwarded-Host when
+present (the route sends the client's, as fixture 14's Location shows) and from the request otherwise.

modules/mailu/go.mod

@@ -0,0 +1,3 @@
+module mailu
+
+go 1.22

modules/mailu/module.json

@@ -71,8 +71,7 @@
     "mailu_list_domains",
     "mailu_read_mail",
     "mailu_search_mail",
-    "mailu_health",
-    "automx_rows"
+    "mailu_health"
   ],
   "emits": [
     "user.created",
@@ -216,12 +215,6 @@
         "class": "rebuildable",
         "why": "certificates, issued again"
       },
-      {
-        "id": "automx",
-        "path": "${dir:data-automx}",
-        "class": "rebuildable",
-        "why": "client autoconfiguration, written again"
-      },
       {
         "id": "fetchmail",
         "path": "${dir:data-fetchmail}",
@@ -308,12 +301,6 @@
       "mode": "0700",
       "unknown": "program"
     },
-    {
-      "id": "data-automx",
-      "type": "directory",
-      "mode": "0700",
-      "unknown": "program"
-    },
     {
       "id": "config-env",
       "type": "file",
@@ -681,11 +668,25 @@
       "takes-config": ["domain", "proxy-address", "sitename", "website", "url", "container"],
       "merge": "json"
     },
+    {
+      "id": "program-automx",
+      "type": "archive",
+      "path": "${dir:state}/automx",
+      "artifact": "automx",
+      "unknown": "program"
+    },
     {
       "id": "automx",
       "type": "container",
       "name": "mailu-automx",
-      "artifact": "automx",
+      "artifact": "alpine",
+      "health": {
+        "kind": "http",
+        "endpoint": "autoconfig",
+        "path": "/",
+        "status": 200,
+        "body": "Show Thunderbird-style"
+      },
       "network": "mailu",
       "env-file": [
         "${dir:state}/mailu.env"
@@ -694,17 +695,18 @@
         "4243"
       ],
       "volumes": [
-        "${dir:data-automx}:/data"
+        "${dir:state}/automx:/opt/mesh:ro"
+      ],
+      "args": [
+        "/opt/mesh/automx"
+      ],
+      "restart-on": [
+        "program-automx",
+        "config-env"
       ]
     }
   ],
   "build": {
-    "on": [
-      {
-        "arg": "PYTHON_BASE",
-        "image": "python@sha256:25f3cfeaceca14921366af4d1240b56457ef46273bdb508c7b0e8f469f6fd228"
-      }
-    ],
     "artifacts": [
       {
         "name": "code",
@@ -731,8 +733,16 @@
       },
       {
         "name": "automx",
-        "kind": "image",
-        "from": "automx/Dockerfile"
+        "kind": "bundle",
+        "language": "go",
+        "system": "alpine",
+        "from": "cmd/automx",
+        "binary": "automx"
+      },
+      {
+        "name": "alpine",
+        "kind": "upstream",
+        "from": "alpine@sha256:28bd5fe8b56d1bd048e5babf5b10710ebe0bae67db86916198a6eec434943f8b"
       },
       {
         "name": "files",

modules/mailu/test/automx-rows.test.ts

@@ -1,103 +0,0 @@
-// What automx_rows reads (novox/hq issue 1630): the rows of automx2's database as stored, through a read-only
-// sqlite3 inside its container. The database here is a real SQLite file made by the test with the tables
-// automx/files/setup-db creates and the rows add-domains writes — with the IMAP server's name empty, the case
-// the live instance is in. The fake container runs each command's statement against that database and prints
-// it the way the image's sqlite3 does in list mode (one line per row, cells apart by "|", nothing for no rows); that sqlite3
-// refuses `-json` ("sqlite3: Error: unknown option: -json", captured from novox/mailu.automx_rows, issue 1630). Run against the compiled module.
-
-import { test } from "node:test";
-import assert from "node:assert/strict";
-import { DatabaseSync } from "node:sqlite";
-
-import { readAutomx, readCommand, selectOf, parseRows, AUTOMX_DATABASE } from "../dist/automx.js";
-
-function database(imapName: string): DatabaseSync {
-  const db = new DatabaseSync(":memory:");
-  db.exec(`
-    CREATE TABLE provider(id INT PRIMARY KEY NOT NULL, name TEXT NOT NULL, short_name TEXT NOT NULL);
-    CREATE TABLE server(id INT PRIMARY KEY NOT NULL, prio INT NOT NULL DEFAULT 10, name TEXT NOT NULL, port INT NOT NULL,
-      type TEXT NOT NULL, socket_type TEXT NOT NULL, user_name TEXT NOT NULL, authentication TEXT NOT NULL);
-    CREATE TABLE domain(id INT PRIMARY KEY NOT NULL, name TEXT NOT NULL, provider_id INT NOT NULL, ldapserver_id INT NULL);
-    CREATE TABLE server_domain(server_id INT NOT NULL, domain_id INT NOT NULL);
-    INSERT INTO provider VALUES (1, 'example.org', 'example.org');
-    INSERT INTO server(id, port, type, name, socket_type, user_name, authentication)
-      VALUES (2, 143, 'imap', '${imapName}', 'STARTTLS', '%EMAILADDRESS%', 'password-cleartext');
-    INSERT INTO server(id, port, type, name, socket_type, user_name, authentication)
-      VALUES (3, 587, 'smtp', 'mail.example.org', 'STARTTLS', '%EMAILADDRESS%', 'password-cleartext');
-    INSERT INTO domain VALUES (5, 'example.org', 1, NULL);
-    INSERT INTO server_domain VALUES (2, 5), (3, 5);
-  `);
-  return db;
-}
-
-function container(db: DatabaseSync, asked: string[][] = []) {
-  return async (_name: string, command: string[]) => {
-    asked.push(command);
-    if (command.includes("-json")) throw Object.assign(new Error("Command failed: docker exec"), { stderr: "sqlite3: Error: unknown option: -json\n" });
-    const rows = db.prepare(command[command.length - 1]).all() as unknown as Record<string, string>[];
-    return rows.length === 0 ? "" : rows.map((r) => Object.values(r).join("|")).join("\n") + "\n";
-  };
-}
-
-test("the rows come back as stored, and an empty server name is said", async () => {
-  const answer = await readAutomx(container(database("")), "mailu-automx");
-  assert.equal(answer.database, AUTOMX_DATABASE);
-  assert.deepEqual(answer.unreadable, {});
-  assert.deepEqual(answer.tables.provider, [{ id: 1, name: "example.org", short_name: "example.org" }]);
-  assert.deepEqual(answer.tables.domain, [{ id: 5, name: "example.org", provider_id: 1, ldapserver_id: null }]);
-  assert.deepEqual(answer.tables.server_domain, [{ server_id: 2, domain_id: 5 }, { server_id: 3, domain_id: 5 }]);
-  assert.deepEqual(
-    answer.tables.server.map((s) => [s.id, s.type, s.name, s.port]),
-    [[2, "imap", "", 143], [3, "smtp", "mail.example.org", 587]],
-  );
-  assert.deepEqual(answer.empty_names, [{ id: 2, type: "imap" }]);
-});
-
-test("a named IMAP server is not reported empty", async () => {
-  const answer = await readAutomx(container(database("mail.example.org")), "mailu-automx");
-  assert.deepEqual(answer.empty_names, []);
-});
-
-test("every command opens the file read-only and runs one fixed SELECT", async () => {
-  const asked: string[][] = [];
-  await readAutomx(container(database("x"), asked), "mailu-automx");
-  assert.equal(asked.length, 4);
-  for (const c of asked) {
-    assert.deepEqual(c.slice(0, 3), ["sqlite3", "-readonly", AUTOMX_DATABASE]);
-    assert.match(c[3], /^SELECT /);
-    assert.equal(c.length, 4);
-    assert.ok(!c.includes("-json"));
-  }
-  assert.deepEqual(readCommand("provider"), ["sqlite3", "-readonly", AUTOMX_DATABASE, selectOf("provider")]);
-});
-
-test("a table that cannot be read is said and does not hide the others", async () => {
-  const db = database("x");
-  const exec = async (name: string, command: string[]) => {
-    if (command[3].includes("FROM domain ")) throw Object.assign(new Error("Command failed: docker exec"), { stderr: "Error: unable to open database file\nmore" });
-    return container(db)(name, command);
-  };
-  const answer = await readAutomx(exec, "mailu-automx");
-  assert.deepEqual(answer.unreadable, { domain: "Error: unable to open database file" });
-  assert.equal(answer.tables.provider.length, 1);
-  assert.equal(answer.tables.domain, undefined);
-});
-
-test("a NULL, an empty text, a number and a text holding a bar or a newline are read apart", () => {
-  const hex = (t: string) => Buffer.from(t).toString("hex").toUpperCase();
-  const line = [`integer:${hex("5")}`, `text:${hex("a|b\nc")}`, "text:", "null:"].join("|");
-  const rows = parseRows("domain", line + "\n");
-  assert.deepEqual(rows, [{ id: 5, name: "a|b\nc", provider_id: "", ldapserver_id: null }]);
-});
-
-test("output that is not cells is an unreadable table, not a guess", async () => {
-  const exec = async () => "Error: something else\n";
-  const answer = await readAutomx(exec, "mailu-automx");
-  assert.equal(Object.keys(answer.unreadable).length, 4);
-});
-
-test("the image's refusal of -json is said as the table's reason", async () => {
-  const exec = async (_n: string, command: string[]) => container(database("x"))("", ["sqlite3", "-json", ...command.slice(1)]);
-  const answer = await readAutomx(exec, "mailu-automx");
-  assert.deepEqual(answer.unreadable.domain, "sqlite3: Error: unknown option: -json");
-});

modules/mailu/test/mail-flows.test.ts

@@ -135,7 +135,6 @@ test("the health tool is served without the admin API, on the smtp container it
   await import("../dist/tools/index.js");
   const mailu = collectTools({ MESH_MAILU_SMTP_CONTAINER: "mailu-smtp" }).find((m) => m.module === "mailu");
   const names = (mailu?.tools ?? []).map((t) => t.name);
-  assert.deepEqual(names, ["mailu_health", "automx_rows"]);
+  assert.deepEqual(names, ["mailu_health"]);
   assert.match(mailu!.tools[0].description, /inside mailu-smtp/);
-  assert.match(mailu!.tools[1].description, /read-only, inside mailu-automx/);
 });

modules/mailu/tools/index.ts

@@ -4,7 +4,6 @@
 
 import { registerModuleTools, type ToolDefinition } from "@novox/mesh-sdk/tools";
 import { MailuClient } from "../client.js";
-import { readAutomx } from "../automx.js";
 import { dockerExec, mailFlows, DEFERRED_BOUND_SECONDS, PUBLIC_NAME, SPREAD_DOMAINS } from "../health.js";
 
 /**
@@ -27,23 +26,6 @@ export function getHealthTool(env: NodeJS.ProcessEnv): ToolDefinition {
   };
 }
 
-/**
- * What automx holds, as stored (novox/hq issue 1630): the rows behind the answers it gives mail clients. It needs
- * no admin API, only the automx container, so like the health tool it is served without the API.
- */
-export function getAutomxTool(env: NodeJS.ProcessEnv): ToolDefinition {
-  const container = env.MESH_MAILU_AUTOMX_CONTAINER || "mailu-automx";
-  return {
-    name: "automx_rows",
-    description:
-      `Read the autoconfiguration module's database, as stored, read-only, inside ${container}: its provider, ` +
-      "server, domain and server_domain rows, and the servers whose name is empty (a mail client is then told " +
-      "an empty host name). Writes nothing.",
-    input: {},
-    run: async () => readAutomx(dockerExec(), container),
-  };
-}
-
 export function getMailuTools(mailu: MailuClient): ToolDefinition[] {
   return [
     {
@@ -191,5 +173,5 @@ registerModuleTools("mailu", (env) => {
   } catch {
     tools = [];
   }
-  return [...tools, getHealthTool(env), getAutomxTool(env)];
+  return [...tools, getHealthTool(env)];
 });

modules/mailu/tsconfig.json

@@ -8,5 +8,5 @@
     "skipLibCheck": true,
     "noEmit": true
   },
-  "include": ["client.ts", "health.ts", "automx.ts", "index.ts", "tools/index.ts"]
+  "include": ["client.ts", "health.ts", "index.ts", "tools/index.ts"]
 }

modules/mesh-build-go/cmd/mesh-build-go/fetch.go

@@ -145,7 +145,7 @@ func (m *Module) goModDirs(ctx context.Context, tree string) ([]string, error) {
 	}
 	dirs := []string{}
 	for _, l := range strings.Split(r.Stdout, "\n") {
-		if l = strings.TrimSpace(l); l != "" {
+		if l = strings.TrimSpace(l); l != "" && !inTestdata(l) {
 			dirs = append(dirs, filepath.ToSlash(filepath.Dir(l)))
 		}
 	}
@@ -153,6 +153,17 @@ func (m *Module) goModDirs(ctx context.Context, tree string) ([]string, error) {
 	return dirs, nil
 }
 
+// inTestdata says a path lies under a directory named testdata, whose go.mod files are fixtures Go ignores, not modules
+// of the tree (issue 1810).
+func inTestdata(path string) bool {
+	for _, seg := range strings.Split(filepath.ToSlash(path), "/") {
+		if seg == "testdata" {
+			return true
+		}
+	}
+	return false
+}
+
 // goEnv is what a go command of a step is given beyond the base environment: the installed go alone (GOTOOLCHAIN=local, ADR 0477 §1: no toolchain is ever downloaded),
 // modules resolved by their go.mod and go.sum, the module's own caches, the mesh's own modules fetched directly (by
 // git, through the account's rewrite of the forge's address to ssh), and nothing that asks a question. For a directory

modules/mesh-build-go/cmd/mesh-build-go/gomoddirs_test.go

@@ -0,0 +1,27 @@
+package main
+
+import (
+	"context"
+	"reflect"
+	"testing"
+
+	"git.novox.be/novox/mesh-sdk/go/forgekey"
+)
+
+// 1810/AC1: the go.mod files under a testdata directory are fixtures Go itself ignores, not modules of the tree, so
+// the build modules neither test nor digest them (a repository holding this module held one whose test reads
+// MESH_CHECK_BESIDE, which failed every check of it).
+func TestGoModDirsLeaveOutTheFixturesUnderTestdata(t *testing.T) {
+	m := &Module{}
+	m.Run = func(ctx context.Context, env []string, name string, args ...string) forgekey.Ran {
+		return forgekey.Ran{Stdout: "checks/deps/go.mod\ngo.mod\nmodules/x/cmd/x/testdata/repos/beside/go.mod\nmodules/x/go.mod\ntestdata/go.mod\n"}
+	}
+	got, err := m.goModDirs(context.Background(), t.TempDir())
+	if err != nil {
+		t.Fatal(err)
+	}
+	want := []string{".", "checks/deps", "modules/x"}
+	if !reflect.DeepEqual(got, want) {
+		t.Fatalf("go.mod directories = %v, want %v", got, want)
+	}
+}

modules/mesh-build-go/cmd/mesh-build-go/recipe.go

@@ -20,7 +20,7 @@ const PostgresMajor = 18
 
 // RecipeSteps is the recipe's own version: changed whenever the steps of fetch, test or build change in a way
 // that can change what they produce (ADR 0405 §7). It is one of the three values the recipe version hashes.
-const RecipeSteps = "1"
+const RecipeSteps = "2"
 
 // natsModule is the library the integration level's bus is started from (ADR 0470 §2).
 const natsModule = "github.com/nats-io/nats-server/v2"

modules/mesh-delivery-gate/cmd/mesh-delivery-gate/approval.go

@@ -380,18 +380,19 @@ func (h *Holder) approval(d *Delivery, now time.Time) (string, func()) {
 		// A group is not asked while a member's required status is not a success at its head (issue 1275): the group
 		// says which, and is asked once every member may merge.
 		var holdMember, holdSays string
+		var holdOwn bool
 		var readErr error
 		for _, m := range reads {
 			if gid == "" {
 				break
 			}
-			says, _, err := h.notASuccess(m)
+			says, ownOnly, err := h.notASuccess(m)
 			if err != nil {
 				readErr = err
 				break
 			}
 			if says != "" {
-				holdMember, holdSays = m.id, says
+				holdMember, holdSays, holdOwn = m.id, says, ownOnly
 				break
 			}
 		}
@@ -421,7 +422,18 @@ func (h *Holder) approval(d *Delivery, now time.Time) (string, func()) {
 		if readErr != nil || holdSays != "" {
 			stale()
 			h.mu.Lock()
-			defer h.mu.Unlock()
+			var own func()
+			// A member's own repository check pending or missing (a sibling's new head put it out of date, issue 1792) is
+			// asked again here, once for the head: nothing else asks it, and the composed check sets no repository check.
+			if d := h.deliveries[holdMember]; holdOwn && d != nil && d.Ready != nil {
+				own = h.askOwnAfterRefusal(d)
+			}
+			defer func() {
+				h.mu.Unlock()
+				if own != nil {
+					own()
+				}
+			}()
 			undo()
 			if g := h.groups[gid]; g != nil && holdSays != "" {
 				if g.Hold == nil || g.Hold.Kind != "unasked" || g.Hold.Says != holdSays || g.Hold.Member != holdMember {
@@ -634,6 +646,7 @@ func (h *Holder) answered(key string, ans sdk.Answer) []func() {
 	for _, d := range covered {
 		r := d.Ready
 		r.Warrant = ans.Words
+		r.Repeats = 0      // the operator chose: whatever the tool refused before is asked afresh (issue 1106)
 		h.mergeAgainNow(d) // the operator chose: the merge is asked at once
 		switch ans.Option {
 		case OptionApprove:

modules/mesh-delivery-gate/cmd/mesh-delivery-gate/bus_test.go

@@ -519,30 +519,29 @@ func TestAPassingVerdictAsksTheApproveAndItsApproveMergesWithinSecondsOverARealB
 	b.publish(t, "mesh.mod.gitea.event.pull.updated", announced)
 	passed := b.publish(t, "mesh.seat."+ControllerSeat+".event.checked", judgedAt("novox/app", 7, head, "pass", "t1t1t1t1t1t1"))
 
-	// The ask is written into the gate's bucket asks within seconds of the verdict's event, with no label, comment or mark
-	// written by anyone, and nothing is asked of a router.
+	// No review names the head, so nothing is asked (ADR 0496, issue 1803): the delivery waits for a review.
+	rt.until("the delivery to wait for a review", 5*time.Second, func() bool {
+		e, err := b.kvs["pulls"].Get(ctx, "novox.app.7")
+		var k PullKey
+		return err == nil && json.Unmarshal(e.Value(), &k) == nil && k.Verdict == "pass"
+	})
+	time.Sleep(time.Second)
+	if key := onlyAsk(b); key != "" {
+		t.Fatalf("a head with no review was asked: %s", key)
+	}
+	// The description is edited to add the review record (issue 1079): the forge module's edit event carries it, and the
+	// ask is written into the gate's bucket asks within seconds, with no label, comment or mark written by anyone, and
+	// nothing is asked of a router.
+	edit := time.Now()
+	b.publish(t, "mesh.mod.gitea.event.forge.pull-request.edited", map[string]any{"owner": "novox", "repo": "app", "number": 7,
+		"title": announced.Title, "head_sha": head, "base": "main", "body": "Reviewed-by: lead@shanks at " + head})
+	_ = passed
 	var key string
 	rt.until("the ask to be written", 5*time.Second, func() bool { key = onlyAsk(b); return key != "" })
 	ask := askAt(t, b, key)
-	t.Logf("the ask was written %s after the verdict's event", time.Since(passed))
-	if !strings.Contains(ask.Ask.Explanation, "not reviewed at this head") || ask.Ask.ID != sdk.AskID(key) {
-		t.Fatalf("the first ask does not say the record is missing: %s", ask.Ask.Explanation)
-	}
-	// The description is edited to add the review record (issue 1079): the forge module's edit event carries it. The unanswered
-	// ask is taken away and written again, once, with the record read (the heads are the same, so is the key).
-	b.publish(t, "mesh.mod.gitea.event.forge.pull-request.edited", map[string]any{"owner": "novox", "repo": "app", "number": 7,
-		"title": announced.Title, "head_sha": head, "base": "main", "body": "Reviewed-by: lead@shanks at " + head})
-	rt.until("the ask to be written again with the record", 5*time.Second, func() bool {
-		a, ok := askNow(b, key) // the key may be absent for an instant while the ask is taken away and written again
-		return ok && strings.Contains(a.Ask.Explanation, "reviewed by lead on shanks")
-	})
-	// The ask is overwritten, never taken away: the key is there at every revision, and the second asking has an id of its own
-	// (review round 1, G2), which the router would take as new.
-	if h, err := b.kvs["asks"].History(ctx, key); err != nil || len(h) != 2 || h[1].Operation() != jetstream.KeyValuePut {
-		t.Fatalf("the ask's history is %d revisions (%v); wanted written and written again", len(h), err)
-	}
-	if again := askAt(t, b, key); again.Asking != 2 || again.Ask.ID != sdk.AskingID(key, 2) || again.Ask.ID == ask.Ask.ID {
-		t.Fatalf("the second asking is %d under %s, the first was %s", again.Asking, again.Ask.ID, ask.Ask.ID)
+	t.Logf("the ask was written %s after the edit's event", time.Since(edit))
+	if !strings.Contains(ask.Ask.Explanation, "reviewed by lead on shanks") || ask.Ask.ID != sdk.AskID(key) {
+		t.Fatalf("the ask does not say the record: %s", ask.Ask.Explanation)
 	}
 	if len(rt.asksOf("gitea.gitea_merge_pull_request")) != 0 {
 		t.Fatalf("merged before the Approve")

modules/mesh-delivery-gate/cmd/mesh-delivery-gate/checks.go

@@ -190,7 +190,7 @@ func (h *Holder) deliveryOf(repository string, number int, commit string) *check
 		c.Check = &v
 	}
 	if g := h.groups[found.Group]; g != nil {
-		state, why := GroupState(g, h.membersOf(g))
+		state, why, _ := h.groupState(g, h.membersOf(g))
 		c.groupSaid = state + ": " + why
 		if g.Check != nil && g.Check.Verdict != "" {
 			c.groupSaid += "; composed check " + g.Check.Verdict + " — " + g.Check.Summary

modules/mesh-delivery-gate/cmd/mesh-delivery-gate/effects.go

@@ -131,6 +131,9 @@ func statusOf(d *Delivery) (string, string) {
 	case Proposed, Checked:
 		return "pending", "checking"
 	case Ready:
+		if d.holdsForReview() {
+			return "pending", clip("ready, "+d.Ready.Waits, 140) // novox/hq ADR 0496: no Approve is asked before a review
+		}
 		return "success", "ready: it delivers once merged"
 	case Rejected:
 		return "failure", clip("rejected: "+why, 140)
@@ -154,6 +157,9 @@ func statusOf(d *Delivery) (string, string) {
 		}
 		return "failure", clip("failed: "+why, 140)
 	case Superseded:
+		if d.coveredStop() {
+			return "success", clip(d.coveredStopSaid(), 140) // its walk's taker delivered it (issue 1786)
+		}
 		return "warning", clip("superseded: "+why, 140)
 	case Stopped:
 		if d.coveredStop() {

modules/mesh-delivery-gate/cmd/mesh-delivery-gate/events_test.go

@@ -158,11 +158,11 @@ func (*askedAgainError) Error() string {
 	return "the forge refused gitea_merge_pull_request: novox/app#7 is not merged: its verdict was judged against trunk aaaaaaaaaaaa, and main is now bbbbbbbbbbbb. The check of head cccccccccccc merged into trunk bbbbbbbbbbbb was asked; took 3s. Ask again at cccccccccccc once both statuses pass. Nothing was merged"
 }
 
-// Issue 1022 (ADR 0421 §5): the review record is read and put in the ask; it is never a gate.
-func TestTheAskSaysFirstWhenThePullRequestIsNotReviewedAtThisHead(t *testing.T) {
-	for _, c := range []struct{ name, body, want string }{
-		{"no line", "Fixes issue 7.", "not reviewed at this head"},
-		{"another head", "Reviewed-by: lead@shanks at " + pushed, "the Reviewed-by line names " + pushed[:12]},
+// Issue 1022 (ADR 0421 §5), as ADR 0496 changed it (issue 1803): the review record is read; a head not reviewed is not asked.
+func TestAHeadNotReviewedAtThisHeadIsNotAsked(t *testing.T) {
+	for _, c := range []struct{ name, body string }{
+		{"no line", "Fixes issue 7."},
+		{"another head", "Reviewed-by: lead@shanks at " + pushed},
 	} {
 		t.Run(c.name, func(t *testing.T) {
 			w := newWorld(t)
@@ -171,13 +171,8 @@ func TestTheAskSaysFirstWhenThePullRequestIsNotReviewedAtThisHead(t *testing.T)
 			p.Body = c.body
 			w.h.PullUpdated(p)
 			hear(t, w, ControllerSeat+".checked", judgedAt("novox/app", 7, readyHead, "pass", "t1t1t1t1t1t1"))
-			if w.op.count() != 1 {
-				t.Fatalf("an unreviewed pull request must still be asked: %d", w.op.count())
-			}
-			a := w.op.ask(w.op.open()[0])
-			first := strings.SplitN(a.Explanation, "\n", 2)[0]
-			if !strings.Contains(first, "not reviewed at this head") || !strings.Contains(a.Explanation, c.want) {
-				t.Fatalf("the ask does not say first that it is not reviewed at this head:\n%s", a.Explanation)
+			if w.op.count() != 0 {
+				t.Fatalf("an unreviewed pull request was asked: %d", w.op.count())
 			}
 		})
 	}

modules/mesh-delivery-gate/cmd/mesh-delivery-gate/fakes_test.go

@@ -601,10 +601,11 @@ type world struct {
 	source *fakeSource
 	owed   *fakeOwed
 	// asked is the key of the ask a test answered.
-	asked string
-	now   time.Time
-	said  []map[string]any
-	mu    sync.Mutex
+	asked  string
+	now    time.Time
+	logbuf strings.Builder
+	said   []map[string]any
+	mu     sync.Mutex
 }
 
 func newWorld(t *testing.T) *world {
@@ -615,7 +616,7 @@ func newWorldOver(t *testing.T, store *memStore, ctl *fakeController, forge *fak
 	w := &world{t: t, store: store, ctl: ctl, forge: forge, now: now}
 	w.h = &Holder{Store: store, Controller: ctl, Forge: forge,
 		Now:  func() time.Time { w.mu.Lock(); defer w.mu.Unlock(); return w.now },
-		Logf: func(format string, a ...any) { t.Logf(format, a...) },
+		Logf: func(format string, a ...any) { t.Logf(format, a...); fmt.Fprintf(&w.logbuf, format+"\n", a...) },
 		Emit: func(event string, body any) error {
 			w.mu.Lock()
 			defer w.mu.Unlock()
@@ -682,7 +683,9 @@ func (w *world) state(id string) State { return w.delivery(id).State }
 func pr(repo string, number int, sha, branch string, paths ...string) PullEvent {
 	owner, name, _ := strings.Cut(repo, "/")
 	return PullEvent{Owner: owner, Repo: name, Number: number, Base: "main", Head: branch, HeadSHA: sha,
-		HTMLURL: fmt.Sprintf("https://forge.invalid/%s/pulls/%d", repo, number), Paths: paths}
+		HTMLURL: fmt.Sprintf("https://forge.invalid/%s/pulls/%d", repo, number), Paths: paths,
+		// Reviewed at its head, as the Approve is asked only for a reviewed head (novox/hq ADR 0496, issue 1803).
+		Body: "Reviewed-by: tester@shanks at " + sha}
 }
 
 // verdict is the controller's `checked` for a head.
@@ -719,3 +722,5 @@ func (f *fakeForge) Comment(owner, repo string, number int, body string) error {
 	f.calls = append(f.calls, "comment "+body)
 	return nil
 }
+
+func (w *world) logs() string { return w.logbuf.String() }

modules/mesh-delivery-gate/cmd/mesh-delivery-gate/holder.go

@@ -467,7 +467,7 @@ func (h *Holder) fillFromAnnouncement(d *Delivery, p PullEvent) {
 // PullEdited takes a pull request's edit (novox/hq issue 1079): its description is read again for the review record, the
 // `after:` lines and the issues it names, exactly as an announcement reads them. An Approve ask still unanswered whose words
 // the new record changes is withdrawn and asked again with it read, without a word on the pull request: nothing was
-// approved, so no Approve ended. An answered ask is left as it is. The record is never a gate (ADR 0421 §5). An edit with no
+// approved, so no Approve ended. An answered ask is left as it is. A record naming the head lifts the hold at once (ADR 0496 §4). An edit with no
 // description, or of a pull request with no open delivery at that head, changes nothing.
 func (h *Holder) PullEdited(p PullEvent) {
 	if p.HeadSHA == "" || p.Number == 0 {
@@ -995,6 +995,12 @@ func (h *Holder) Checked(c CheckedEvent) {
 		}
 		h.moved(d, []Transition{t})
 	}
+	if d.Ready != nil && !d.Ready.OwnAskedAt.IsZero() {
+		// The own check asked for this head answered (issue 1792): one own ask is in flight at a time, so a status still
+		// pending or missing after its answer is asked again.
+		d.Ready.OwnAskedAt = time.Time{}
+		h.keep(d)
+	}
 	if d.State == Rejected && c.Trunk != "" && d.Check != nil && !sameCommit(d.Check.Trunk, c.Trunk) {
 		// Judged again against another trunk commit — the merge tool's ask of a verdict out of date, or a check asked
 		// elsewhere (novox/hq ADR 0323 §5, §7): taken as the recheck it is.
@@ -1390,10 +1396,31 @@ func (h *Holder) plan() []func() {
 	// Following a walk that was taken over: the walk that took it over, read as soon as it is heard of, asked for
 	// by its id when it is not.
 	for _, d := range h.deliveries {
-		if d.State.Final() || d.Walk == nil || d.Walk.TakenOverBy == "" {
+		if d.Walk == nil || d.Walk.TakenOverBy == "" {
+			continue
+		}
+		if d.State.Final() {
+			// Ended superseded by its walk before the walk that took it over was heard to answer (issue 1786): the
+			// state stays final, as a covered superseded member's does (issue 442), and the end of that walk is
+			// kept as DeliveredBy, which the group and the views read. Never logged, never changed otherwise.
+			if d.supersededByItsWalk() && d.DeliveredBy == "" {
+				if s, known := h.walks[d.Walk.TakenOverBy]; known {
+					if _, ok := walkAnswers(s.walk, d); ok && s.walk.State == walkDone {
+						d.DeliveredBy = fmt.Sprintf("the walk %s, which took its walk %s over, ended done", s.walk.ID, d.Walk.ID)
+						h.tellCovered(d, now)
+					}
+				} else if a := h.askWalk(d.Walk.TakenOverBy, "which took a delivery's walk over", now); a != nil {
+					asks = append(asks, a)
+				}
+			}
 			continue
 		}
 		if s, known := h.walks[d.Walk.TakenOverBy]; known {
+			if s.walk.IsBatch() {
+				// A batch is no walk yet (ADR 0276): the walk it is cut into answers the merge. Waited for, never
+				// read as a walk that does not answer it (issue 1786).
+				continue
+			}
 			if c, ok := walkAnswers(s.walk, d); ok {
 				h.linkWalk(d, s.walk, c)
 			} else {
@@ -2175,7 +2202,7 @@ func (h *Holder) planGroup(g *Group, members []*Delivery, now time.Time) []func(
 	// moves — and its group counts it as reached (novox/hq issue 442).
 	if g.Closed {
 		for _, d := range members {
-			if d.State != Superseded || d.CoveredBy != "" {
+			if d.State != Superseded || d.CoveredBy != "" || d.DeliveredBy != "" {
 				continue
 			}
 			var ask []string
@@ -2190,15 +2217,22 @@ func (h *Holder) planGroup(g *Group, members []*Delivery, now time.Time) []func(
 		}
 	}
 	// Its state, said when it changes.
-	state, why := GroupState(g, members)
+	state, why, held := h.groupState(g, members)
 	status := groupStatusState(state)
+	if held {
+		status = "pending" // no Approve exists while a member waits for its review (novox/hq ADR 0496 §2)
+	}
 	if state == "delivering" && !(g.Check != nil && (g.Check.Verdict == "pass" || g.Check.Verdict == "warning") &&
 		sameMembers(g.Check.Commits, commitsOf(members)) && g.Check.OutOfDate == "") {
 		// Success there lets the merge tool merge its members: only for a composed check that passed for these heads.
 		status = "pending"
 	}
-	if state != g.Said || status != g.SaidStatus {
-		g.Said, g.SaidStatus = state, status
+	reviewSaid := ""
+	if held {
+		reviewSaid = why
+	}
+	if state != g.Said || status != g.SaidStatus || reviewSaid != g.SaidReview {
+		g.Said, g.SaidStatus, g.SaidReview = state, status, reviewSaid
 		for _, d := range members {
 			g.Owed = append(g.Owed, Effect{Kind: EffectStatus, Commit: d.Commit, Context: "mesh/delivery-group",
 				State: status, Line: clip(fmt.Sprintf("group %s %s: %s", g.ID, state, why), 140),
@@ -2692,6 +2726,29 @@ func GroupState(g *Group, members []*Delivery) (string, string) {
 	return "checking", fmt.Sprintf("%d of %d member(s) ready", count[Ready], n)
 }
 
+// groupState is GroupState, but a group whose members all passed and compose waits for a review while any member the
+// Approve would cover has none at its head: its why names the members (novox/hq ADR 0496 §2, issue 1803), and held
+// says its mesh/delivery-group is pending, not a success. Under the lock.
+func (h *Holder) groupState(g *Group, members []*Delivery) (state, why string, held bool) {
+	state, why = GroupState(g, members)
+	if state != "ready" || g.Closed {
+		return state, why, false
+	}
+	var waiting []string
+	for _, m := range members {
+		if m.MergedAs != "" || m.State.Final() || h.covering(m) != nil {
+			continue
+		}
+		if _, ok := reviewWords(m); !ok {
+			waiting = append(waiting, "at "+shortOf(m.Commit, 12)+" ("+pullSaid(m)+")")
+		}
+	}
+	if len(waiting) == 0 {
+		return state, why, false
+	}
+	return state, waitsForReview + strings.Join(waiting, " and ")[len("at "):], true
+}
+
 // supersededWhy begins why a group whose members all ended is stopped. It is kept short: the forge's status line,
 // "group <id> stopped: <why>", is clipped at 140 characters, and the member's id must fit in it whole.
 const supersededWhy = "superseded, not carried by a later delivery: "

modules/mesh-delivery-gate/cmd/mesh-delivery-gate/holder_test.go

@@ -19,8 +19,12 @@ const (
 
 func TestAPullRequestsHeadIsProposedCheckedAndReady(t *testing.T) {
 	w := newWorld(t)
-	w.h.PullUpdated(pr("novox/app", 7, head, "feat/x", "src/a.go"))
+	p := pr("novox/app", 7, head, "feat/x", "src/a.go")
+	p.Body = "" // an empty description is never read again, so the review set below stays
+	w.h.PullUpdated(p)
 	id := IDOf("novox/app", head)
+	// This head is not a full commit id, so its description line cannot name it: its review is set as read.
+	w.h.deliveries[id].ReviewedBy, w.h.deliveries[id].ReviewedAt = "tester@shanks", head
 	if w.state(id) != Proposed {
 		t.Fatalf("a head is %s", w.state(id))
 	}
@@ -359,6 +363,9 @@ func TestAGroupIsCheckedTogetherAndLetGoOnceEveryMemberIsMerged(t *testing.T) {
 	ctl, cat := IDOf("novox/mesh-controller", ctlHead), IDOf("novox/mesh-catalog", catHead)
 	w.h.PullUpdated(pr("novox/mesh-controller", 1, ctlHead, "feat/x", "internal/a.go"))
 	w.h.PullUpdated(pr("novox/mesh-catalog", 2, catHead, "feat/x", "modules/app/module.json"))
+	// These heads are not full commit ids, so their description lines cannot name them: the reviews are set as read.
+	w.h.deliveries[ctl].ReviewedBy, w.h.deliveries[ctl].ReviewedAt = "tester@shanks", ctlHead
+	w.h.deliveries[cat].ReviewedBy, w.h.deliveries[cat].ReviewedAt = "tester@shanks", catHead
 	if w.delivery(ctl).Group != "feat/x" || w.delivery(cat).Group != "feat/x" {
 		t.Fatalf("one branch name in two repositories made no group: %q %q", w.delivery(ctl).Group, w.delivery(cat).Group)
 	}

modules/mesh-delivery-gate/cmd/mesh-delivery-gate/issue1079_test.go

@@ -10,18 +10,19 @@ import (
 // this head". The forge's edit event is followed; an Approve ask not yet answered is withdrawn and asked again with the
 // record read (the operator agreed, 2026-10-18). The review record is never a gate (ADR 0421 §5).
 
-// askedUnreviewed is a green pull request whose Approve is asked while its description holds no Reviewed-by line.
-func askedUnreviewed(t *testing.T) *world {
+// askedReviewed is a green pull request whose Approve is asked, its review line naming the head by one reviewer.
+func askedReviewed(t *testing.T) *world {
 	w := newWorld(t)
 	w.op.auto = false
 	p := pr("novox/app", 7, readyHead, "feat/x", "src/a.go")
 	p.Title = "Fix the apostrophe (issue 948)"
+	p.Body = "Reviewed-by: first@shanks at " + readyHead
 	w.h.PullUpdated(p)
 	w.h.Checked(judgedAt("novox/app", 7, readyHead, "pass", "t1t1t1t1t1t1"))
 	w.forge.open = []OpenPull{openPull("novox/app", 7, readyHead)}
 	w.h.SeeOpenPulls()
 	w.settleAll()
-	if open := w.op.open(); len(open) != 1 || !strings.Contains(w.op.ask(open[0]).Explanation, "not reviewed at this head") {
+	if open := w.op.open(); len(open) != 1 || !strings.Contains(w.op.ask(open[0]).Explanation, "reviewed by first on shanks") {
 		t.Fatalf("setup: the ask is %v", w.op.open())
 	}
 	return w
@@ -33,10 +34,10 @@ func edited(body string) PullEvent {
 	return p
 }
 
-func TestAReviewedByLineWrittenAfterTheAskWithdrawsTheUnansweredAskAndAsksAgain(t *testing.T) {
-	w := askedUnreviewed(t)
+func TestAReviewedByLineChangedAfterTheAskWithdrawsTheUnansweredAskAndAsksAgain(t *testing.T) {
+	w := askedReviewed(t)
 	first := w.op.open()[0]
-	w.h.PullEdited(edited("Reviewed-by: reviewer@shanks at " + readyHead))
+	w.h.PullEdited(edited("Reviewed-by: second@shanks at " + readyHead))
 	w.settleAll()
 	open := w.op.open()
 	// The heads are the same, so the key is: the ask is taken away and written again with the record in its words, which the
@@ -47,8 +48,7 @@ func TestAReviewedByLineWrittenAfterTheAskWithdrawsTheUnansweredAskAndAsksAgain(
 	if len(w.op.cancelled) != 0 || w.op.byKey[first].Asking != 2 {
 		t.Fatalf("the first ask was taken away (%v) or not written again as asking 2 (%d)", w.op.cancelled, w.op.byKey[first].Asking)
 	}
-	if e := w.op.ask(open[0]).Explanation; !strings.Contains(e, "reviewed by reviewer on shanks at "+shortOf(readyHead, 12)) ||
-		strings.Contains(e, "not reviewed") {
+	if e := w.op.ask(open[0]).Explanation; !strings.Contains(e, "reviewed by second on shanks at "+shortOf(readyHead, 12)) {
 		t.Fatalf("the new ask does not say the record: %s", e)
 	}
 	if len(w.forge.comments) != 0 {
@@ -63,35 +63,29 @@ func TestAReviewedByLineWrittenAfterTheAskWithdrawsTheUnansweredAskAndAsksAgain(
 }
 
 func TestAnEditThatChangesNothingTheAskSaysAsksNothingAndNoEditChangesAnAnsweredAsk(t *testing.T) {
-	w := askedUnreviewed(t)
-	// The same words: no description with a line, then a description without one.
-	w.h.PullEdited(edited("A new sentence in the description."))
+	w := askedReviewed(t)
+	// The same words: a description with the same line and a new sentence.
+	w.h.PullEdited(edited("A new sentence.\nReviewed-by: first@shanks at " + readyHead))
 	w.settleAll()
 	if w.op.count() != 1 || len(w.op.cancelled) != 0 {
 		t.Fatalf("an edit that changes no word of the ask asked again: %d asks, cancelled %v", w.op.count(), w.op.cancelled)
 	}
-	// An edit that brings a line for another head changes the words: the ask is made again, still saying not reviewed.
-	w.h.PullEdited(edited("Reviewed-by: reviewer@shanks at " + strings.Repeat("a", 40)))
-	w.settleAll()
-	if w.op.count() != 2 || len(w.op.open()) != 1 {
-		t.Fatalf("a line for another head: %d asks, open %v", w.op.count(), w.op.open())
-	}
 	// Answered: a later edit changes nothing.
 	w.op.answer(w.op.open()[0], OptionApprove)
 	w.settleAll()
 	asked := w.op.count()
-	w.h.PullEdited(edited("Reviewed-by: reviewer@shanks at " + readyHead))
+	w.h.PullEdited(edited("Reviewed-by: second@shanks at " + readyHead))
 	w.settleAll()
 	if w.op.count() != asked || len(w.op.open()) != 0 {
 		t.Fatalf("an edit after the Approve asked again: %d asks, open %v", w.op.count(), w.op.open())
 	}
-	if d := w.delivery(IDOf("novox/app", readyHead)); d.ReviewedBy != "reviewer@shanks" {
+	if d := w.delivery(IDOf("novox/app", readyHead)); d.ReviewedBy != "second@shanks" {
 		t.Fatalf("the record is still read: %q", d.ReviewedBy)
 	}
 }
 
 func TestAnEditOfAnUnknownOrMergedOrEmptyPullRequestChangesNothing(t *testing.T) {
-	w := askedUnreviewed(t)
+	w := askedReviewed(t)
 	other := pr("novox/other", 3, readyHead, "feat/y")
 	other.Body = "Reviewed-by: a@b at " + readyHead
 	w.h.PullEdited(other) // no delivery: nothing made

modules/mesh-delivery-gate/cmd/mesh-delivery-gate/issue1348_test.go

@@ -173,3 +173,43 @@ func TestAGroupMergedWholeMovesSourceMainOnceAndAnnouncesWithTheTeller(t *testin
 		t.Fatalf("source.main moved again: %d", len(w.source.moves))
 	}
 }
+
+// 1106: the same refusal again and again with the trunk not moved is not asked on without end. Three refusals in a row with
+// the same cause end the asking by a verdict, a status or a composed check, and say it urgent; a trunk's move, or the
+// operator's Approve, starts it again.
+func TestARefusalThatRepeatsWithNoTrunkMoveIsNotAskedAgainWithoutEnd(t *testing.T) {
+	w := answeredFirst(t)
+	d := w.h.deliveries[IDOf(gCat, gHk)]
+	refuse := func(took string) {
+		w.h.ForgeRefused(sdk.Forge{Kind: sdk.ForgeRefused, Repository: gCat, Number: 2, Head: gHk, Act: "merge",
+			Words: "novox/mesh-catalog#2 is not merged: its verdict names no trunk commit it was judged against, and main is now f70a99dbf485. " +
+				"The check of head d2d2d2d2d2d2 merged into trunk f70a99dbf485 was asked; the last check took " + took + ". Ask again once both statuses pass"})
+	}
+	tried := w.now
+	for i, took := range []string{"4m1s", "4m9s"} {
+		refuse(took)
+		d.Ready.LastTry = tried
+		w.h.mergeAgainNow(d)
+		if !d.Ready.LastTry.IsZero() {
+			t.Fatalf("refusal %d: the merge was not asked again at once on a verdict, as after a first refusal", i+1)
+		}
+	}
+	refuse("4m5s") // the third in a row, with only the check's duration changed
+	if d.Ready.Repeats != 3 {
+		t.Fatalf("%d refusals counted for the same cause, wanted 3", d.Ready.Repeats)
+	}
+	d.Ready.LastTry = tried
+	w.h.mergeAgainNow(d) // the composed check or a verdict answers: nothing is asked
+	if d.Ready.LastTry.IsZero() {
+		t.Fatal("the merge was asked again after the same refusal three times with the trunk not moved")
+	}
+	if lines := urgentLines(w); len(lines) != 1 || !strings.Contains(lines[0].Says, "names no trunk commit") {
+		t.Fatalf("the repeated refusal is not said urgent: %+v", lines)
+	}
+	// The trunk moves: the merge is asked again.
+	d.TrunkNow = "a1a1a1a1a1a1a1a1"
+	w.h.mergeAgainNow(d)
+	if !d.Ready.LastTry.IsZero() {
+		t.Fatal("a trunk's move did not let the merge be asked again")
+	}
+}

modules/mesh-delivery-gate/cmd/mesh-delivery-gate/issue1786_test.go

@@ -0,0 +1,126 @@
+package main
+
+import (
+	"strings"
+	"testing"
+	"time"
+)
+
+// novox/hq issue 1786: two merges whose walks the controller folded, before they were cut, into a batch that was
+// still queued (not yet cut into a walk) ended superseded: the batch is no walk that answers a merge, so the follow
+// loop took the walk's taken_over_by off ("the walk that took its own over does not answer it") and the table's
+// supersession then held ("names no walk that took it over"). Delivered by the batch's walk, they showed
+// superseded for good. A batch that took a walk over is waited for; and one already ended so follows its taker.
+
+const (
+	head1786  = "1786178617861786178617861786178617861786"
+	merge1786 = "a1786a1786a1786a1786a1786a1786a1786a17860"
+)
+
+// folded1786 is a delivery whose walk was folded into a batch that is still queued.
+func folded1786(t *testing.T) (*world, string) {
+	t.Helper()
+	w := newWorld(t)
+	id := IDOf("novox/app", head1786)
+	w.h.PullUpdated(pr("novox/app", 7, head1786, "feat/x", "src/a.go"))
+	w.h.Checked(verdict("novox/app", 7, head1786, "pass"))
+	w.h.PullMerged(merged("novox/app", 7, head1786, merge1786))
+	w.h.WalkMoved(w.ctl.put(aWalk("plan-own", "novox/app", merge1786, true)))
+	w.settleAll()
+	if s := w.state(id); s != Delivering {
+		t.Fatalf("setup: %s, want delivering", s)
+	}
+	closes := w.now.Add(time.Minute)
+	batch := Walk{ID: "plan-batch", Repository: "novox/other", Branch: "main", Commit: "ffff", Created: w.now, State: walkQueued,
+		Delivery: &WalkDelivery{Awaits: "mesh-delivery", Batch: &WalkBatch{ClosesAt: &closes},
+			Merges: []WalkMerge{{Repository: "novox/app", Commit: merge1786}}}}
+	w.h.WalkMoved(w.ctl.put(batch))
+	folded := w.ctl.walk("plan-own")
+	folded.State = walkSuperseded
+	folded.Note = "folded into plan-batch before it was cut"
+	folded.Delivery.TakenOverBy = "plan-batch"
+	w.h.WalkMoved(w.ctl.put(folded))
+	w.settleAll()
+	return w, id
+}
+
+func cut1786(w *world, state string) {
+	b := w.ctl.walk("plan-batch")
+	b.State = state
+	b.Delivery.Batch = nil
+	b.Tiers, b.Modules = [][]string{{"app"}}, map[string]*WalkModule{"app": {}}
+	w.h.WalkMoved(w.ctl.put(b))
+	w.settleAll()
+}
+
+func TestAWalkFoldedIntoAQueuedBatchIsNotSupersededAndFollowsTheBatchWhenItIsCut(t *testing.T) {
+	w, id := folded1786(t)
+	if d := w.delivery(id); d.State.Final() || d.Walk.TakenOverBy != "plan-batch" {
+		t.Fatalf("folded into a queued batch it is %s, taken over by %q; wanted it still following plan-batch", d.State, d.Walk.TakenOverBy)
+	}
+	cut1786(w, walkBuilding)
+	cut1786(w, walkDone)
+	if s := w.state(id); s != Delivered {
+		t.Fatalf("the batch's walk is done and it is %s; wanted delivered", s)
+	}
+}
+
+func TestADeliverySupersededWhileItsTakerWasABatchIsHealedWhenTheTakerEndsDone(t *testing.T) {
+	w, id := folded1786(t)
+	d := w.h.deliveries[id]
+	// The state kept by the fault: superseded, its walk record naming the walk that took it over.
+	d.State = Superseded
+	d.Transitions = append(d.Transitions, Transition{At: w.now, From: Delivering, To: Superseded, Event: EvSuperseded})
+	if err := w.h.Store.PutDelivery(d); err != nil {
+		t.Fatal(err)
+	}
+	cut1786(w, walkBuilding)
+	cut1786(w, walkDone)
+	w.later(time30s)
+	w.settleAll()
+	d = w.delivery(id)
+	// A final state stays final (ADR 0239 §5): the taker's end is kept as DeliveredBy, as issue 442 and 747 do.
+	if d.State != Superseded || d.DeliveredBy == "" || !memberReached(d) || !d.coveredStop() {
+		t.Fatalf("superseded by the fault and its taker done, it is %s, delivered by %q, reached %v; wanted superseded, delivered by the taker, reached",
+			d.State, d.DeliveredBy, memberReached(d))
+	}
+	if st, _ := statusOf(d); st != "success" {
+		t.Fatalf("its status is %s; wanted success", st)
+	}
+}
+
+func TestAFinalDeliveryWhoseTakerDoesNotAnswerItIsNeitherHealedNorLoggedEveryPass(t *testing.T) {
+	w, id := folded1786(t)
+	d := w.h.deliveries[id]
+	d.State = Superseded
+	d.Transitions = append(d.Transitions, Transition{At: w.now, From: Delivering, To: Superseded, Event: EvSuperseded})
+	other := aWalk("plan-batch", "novox/other", "ffff", false)
+	other.State = walkDone
+	w.h.WalkMoved(w.ctl.put(other))
+	w.settleAll()
+	before := w.logs()
+	w.later(time30s)
+	w.settleAll()
+	if d := w.delivery(id); d.DeliveredBy != "" || d.State != Superseded {
+		t.Fatalf("healed by a walk that does not answer it: %s %q", d.State, d.DeliveredBy)
+	}
+	if n := strings.Count(w.logs(), "does not answer its merge") - strings.Count(before, "does not answer its merge"); n != 0 {
+		t.Fatalf("logged %d times on a pass over a final record", n)
+	}
+}
+
+func TestASupersededDeliveryWithNoWalkThatTookItOverStaysSuperseded(t *testing.T) {
+	w := newWorld(t)
+	id := IDOf("novox/app", head1786)
+	w.h.PullUpdated(pr("novox/app", 7, head1786, "feat/x", "src/a.go"))
+	w.h.PullUpdated(pr("novox/app", 7, "b"+head1786[1:], "feat/x", "src/a.go"))
+	w.settleAll()
+	if s := w.state(id); s != Superseded {
+		t.Fatalf("a head superseded by a newer one is %s", s)
+	}
+	w.later(time30s)
+	w.settleAll()
+	if s := w.state(id); s != Superseded {
+		t.Fatalf("it moved: %s", s)
+	}
+}

modules/mesh-delivery-gate/cmd/mesh-delivery-gate/issue1792_test.go

@@ -0,0 +1,120 @@
+package main
+
+import "testing"
+
+const gHk2 = "f4f4f4f4f4f4f4f4f4f4f4f4f4f4f4f4f4f4f4f4"
+
+// Issue 1792: a sibling's new head put a member's own mesh/repo-check back to pending ("out of date: judged beside ..."),
+// and nothing asked it again. The composed check passed, the group stayed held ("unasked") on the member's pending
+// status, and the operator's Approve was never asked. The group's hold on a member's own pending repository check asks
+// that check, once for the head, with no composed check (ADR 0382 §2: each member's own check stays required).
+func TestAGroupHeldBeforeItsAskOnAMembersOwnPendingRepoCheckAsksThatCheckOnce(t *testing.T) {
+	w := newWorld(t)
+	w.op.auto = false
+	w.h.PullUpdated(pr(gCtl, 1, gHc, "feat/g", "internal/a.go"))
+	w.h.PullUpdated(pr(gCat, 2, gHk, "feat/g", "modules/app/module.json"))
+	w.h.Checked(judged(gCtl, 1, gHc, "pass", gTrunk0))
+	w.h.Checked(judged(gCat, 2, gHk, "pass", gTrunk0))
+	w.settleAll()
+	g := w.h.groups["feat/g"]
+	setChecks(w, gCat, gHk, "success", "success")
+	setChecks(w, gCtl, gHc, "success", "pending")
+	asked := len(w.ctl.checks)
+	w.h.Checked(CheckedEvent{Group: "feat/g", ID: g.Check.Asked, Verdict: "pass", Summary: "composed",
+		Members: []struct {
+			Owner  string `json:"owner"`
+			Repo   string `json:"repo"`
+			Commit string `json:"commit"`
+			Trunk  string `json:"trunk,omitempty"`
+		}{{Owner: "novox", Repo: "mesh-controller", Commit: gHc, Trunk: gTrunk0[:12]},
+			{Owner: "novox", Repo: "mesh-catalog", Commit: gHk, Trunk: gTrunk0[:12]}}})
+	w.settleAll()
+	w.later(time30s)
+	w.settleAll()
+	own, composed := ownChecks(w, asked)
+	if len(composed) != 0 || len(own) != 1 || own[0] != ":"+IDOf(gCtl, gHc) {
+		t.Fatalf("own checks %v, composed %v; wanted the member's own check asked once", own, composed)
+	}
+	if w.op.count() != 0 {
+		t.Fatalf("the operator was asked with a required status pending: %d", w.op.count())
+	}
+}
+
+func composedPassAt(w *world, hc, hk string) {
+	g := w.h.groups["feat/g"]
+	w.h.Checked(CheckedEvent{Group: "feat/g", ID: g.Check.Asked, Verdict: "pass", Summary: "composed",
+		Members: []struct {
+			Owner  string `json:"owner"`
+			Repo   string `json:"repo"`
+			Commit string `json:"commit"`
+			Trunk  string `json:"trunk,omitempty"`
+		}{{Owner: "novox", Repo: "mesh-controller", Commit: hc, Trunk: gTrunk0[:12]},
+			{Owner: "novox", Repo: "mesh-catalog", Commit: hk, Trunk: gTrunk0[:12]}}})
+}
+
+// heldOnOwnPending is the group of issue 1792 after its first own ask: the member's own check asked, once.
+func heldOnOwnPending(t *testing.T) *world {
+	w := newWorld(t)
+	w.op.auto = false
+	w.h.PullUpdated(pr(gCtl, 1, gHc, "feat/g", "internal/a.go"))
+	w.h.PullUpdated(pr(gCat, 2, gHk, "feat/g", "modules/app/module.json"))
+	w.h.Checked(judged(gCtl, 1, gHc, "pass", gTrunk0))
+	w.h.Checked(judged(gCat, 2, gHk, "pass", gTrunk0))
+	w.settleAll()
+	setChecks(w, gCat, gHk, "success", "success")
+	setChecks(w, gCtl, gHc, "success", "pending")
+	asked := len(w.ctl.checks)
+	composedPassAt(w, gHc, gHk)
+	w.settleAll()
+	w.later(time30s)
+	w.settleAll()
+	if own, _ := ownChecks(w, asked); len(own) != 1 {
+		t.Fatalf("the first own ask: %v", own)
+	}
+	return w
+}
+
+// A second sibling push puts the member's repo-check out of date again after its own check answered: asked again.
+func TestAMembersOwnCheckIsAskedAgainWhenASecondSiblingHeadPutsItOutOfDate(t *testing.T) {
+	w := heldOnOwnPending(t)
+	w.h.Checked(judged(gCtl, 1, gHc, "pass", gTrunk0)) // its own check answered
+	setChecks(w, gCtl, gHc, "success", "success")
+	w.settleAll()
+	w.h.PullUpdated(pr(gCat, 2, gHk2, "feat/g", "modules/app/module.json"))
+	w.h.Checked(judged(gCat, 2, gHk2, "pass", gTrunk0))
+	w.settleAll()
+	setChecks(w, gCat, gHk2, "success", "success")
+	setChecks(w, gCtl, gHc, "success", "pending")
+	asked := len(w.ctl.checks)
+	composedPassAt(w, gHc, gHk2)
+	w.settleAll()
+	w.later(time30s)
+	w.settleAll()
+	own, composed := ownChecks(w, asked)
+	if len(composed) != 0 || len(own) != 1 || own[0] != ":"+IDOf(gCtl, gHc) {
+		t.Fatalf("own %v composed %v; wanted the member's own check asked again, once", own, composed)
+	}
+}
+
+// An own ask before the Approve, answered, and then a merge refused for the member's own pending status at the same head:
+// asked again (the one ask is per answer, not per head's life).
+func TestAMergeRefusedForOwnPendingAfterAnAnsweredOwnAskAsksItAgain(t *testing.T) {
+	w := heldOnOwnPending(t)
+	w.h.Checked(judged(gCtl, 1, gHc, "pass", gTrunk0))
+	setChecks(w, gCtl, gHc, "success", "success")
+	w.later(time30s)
+	w.settleAll()
+	if len(w.op.open()) != 1 {
+		t.Fatalf("the group is not asked once its statuses are green: %d", w.op.count())
+	}
+	w.forge.refuseRepo = map[string]error{gCtl: ownPending}
+	w.op.answer(w.op.open()[0], OptionApprove)
+	w.settleAll()
+	w.h.PullMerged(merged(gCat, 2, gHk, gT1))
+	asked := len(w.ctl.checks)
+	w.settleAll()
+	own, composed := ownChecks(w, asked)
+	if len(composed) != 0 || len(own) != 1 || own[0] != ":"+IDOf(gCtl, gHc) {
+		t.Fatalf("own %v composed %v; wanted the member's own check asked again after the refusal", own, composed)
+	}
+}

modules/mesh-delivery-gate/cmd/mesh-delivery-gate/issue1803_test.go

@@ -0,0 +1,223 @@
+package main
+
+import (
+	"strings"
+	"testing"
+)
+
+// novox/hq issue 1803, ADR 0496: the Approve is asked only for a head with a review recorded at that head. A missing
+// line, `Reviewed-by: pending` or a line naming another commit holds the ask; the delivery stays ready and says it waits
+// for a review at the head. An edit naming the head asks at once. A group asks when every member is reviewed. An ask
+// already raised is not touched.
+
+const stale1803 = "9f9f9f9f9f9f9f9f9f9f9f9f9f9f9f9f9f9f9f9f"
+
+// greenWithBody is a green pull request whose description is body, the operator not answering by himself.
+func greenWithBody(t *testing.T, body string) *world {
+	w := newWorld(t)
+	w.op.auto = false
+	p := pr("novox/app", 7, readyHead, "feat/x", "src/a.go")
+	p.Title = "Fix the apostrophe (issue 948)"
+	p.Body = body
+	w.h.PullUpdated(p)
+	w.h.Checked(judgedAt("novox/app", 7, readyHead, "pass", "t1t1t1t1t1t1"))
+	w.forge.open = []OpenPull{openPull("novox/app", 7, readyHead)}
+	w.h.SeeOpenPulls()
+	w.settleAll()
+	return w
+}
+
+func (w *world) wantHeld(t *testing.T, label string) {
+	t.Helper()
+	want := "waits for a review at " + shortOf(readyHead, 12)
+	d := w.delivery(IDOf("novox/app", readyHead))
+	if w.op.count() != 0 || len(w.op.open()) != 0 {
+		t.Fatalf("%s: the operator was asked (%d asks)", label, w.op.count())
+	}
+	if d.State != Ready || d.Ready == nil || !strings.Contains(d.Ready.Waits, want) {
+		t.Fatalf("%s: state %s, waits %+v; wanted ready and %q", label, d.State, d.Ready, want)
+	}
+	found := false
+	for k, v := range w.forge.statuses {
+		if strings.HasSuffix(strings.SplitN(k, " ", 2)[1], "mesh/delivery") && strings.Contains(v, want) {
+			found = true
+		}
+	}
+	if !found {
+		t.Fatalf("%s: no mesh/delivery status says %q: %v", label, want, w.forge.statuses)
+	}
+}
+
+func TestA1803PendingReviewLineHoldsTheAsk(t *testing.T) {
+	w := greenWithBody(t, "Reviewed-by: pending")
+	w.wantHeld(t, "1803/AC1 pending")
+	for range 3 {
+		w.later(2 * 60 * 1e9)
+		w.settleAll()
+	}
+	w.wantHeld(t, "1803/AC1 pending after ticks")
+}
+
+func TestA1803MissingReviewLineHoldsTheAsk(t *testing.T) {
+	w := greenWithBody(t, "No review line at all.")
+	w.wantHeld(t, "1803/AC1 missing")
+}
+
+func TestA1803ReviewLineOfAnotherCommitHoldsTheAsk(t *testing.T) {
+	w := greenWithBody(t, "Reviewed-by: reviewer@shanks at "+stale1803)
+	w.wantHeld(t, "1803/AC2 stale")
+}
+
+func TestA1803ReviewLineNamingTheHeadIsAsked(t *testing.T) {
+	w := greenWithBody(t, "Reviewed-by: reviewer@shanks at "+readyHead)
+	if len(w.op.open()) != 1 {
+		t.Fatalf("1803/AC3: a review at the head is not asked: %v", w.op.open())
+	}
+	if e := w.op.ask(w.op.open()[0]).Explanation; !strings.Contains(e, "reviewed by reviewer on shanks") {
+		t.Fatalf("the ask does not name the review: %s", e)
+	}
+}
+
+func TestA1803AnEditNamingTheHeadRaisesTheAskAtOnce(t *testing.T) {
+	w := greenWithBody(t, "Reviewed-by: pending")
+	w.wantHeld(t, "1803/AC4 before the edit")
+	e := pr("novox/app", 7, readyHead, "feat/x")
+	e.Body = "Reviewed-by: reviewer@shanks at " + readyHead
+	w.h.PullEdited(e)
+	w.settleAll()
+	if len(w.op.open()) != 1 {
+		t.Fatalf("1803/AC4: the edit naming the head raised no ask: %d asks", w.op.count())
+	}
+	// An unanswered ask whose review is taken back is withdrawn: the hold applies again (ADR 0496 §1).
+	e.Body = "Reviewed-by: pending"
+	w.h.PullEdited(e)
+	w.settleAll()
+	if len(w.op.open()) != 0 {
+		t.Fatalf("1803/AC4: an unanswered ask stayed after its review was taken back: open %v", w.op.open())
+	}
+}
+
+func TestA1803AnApproveAlreadyGivenIsNotAffected(t *testing.T) {
+	w := greenWithBody(t, "Reviewed-by: reviewer@shanks at "+readyHead)
+	w.op.answer(w.op.open()[0], OptionApprove)
+	w.settleAll()
+	if len(w.forge.merges) != 1 {
+		t.Fatalf("1803/AC6: the approved head was not merged: %v", w.forge.merges)
+	}
+	// A later edit to pending changes nothing of the Approve given.
+	e := pr("novox/app", 7, readyHead, "feat/x")
+	e.Body = "Reviewed-by: pending"
+	w.h.PullEdited(e)
+	w.settleAll()
+	if d := w.delivery(IDOf("novox/app", readyHead)); d.Ready == nil || d.Ready.Approved == nil {
+		t.Fatalf("1803/AC6: the Approve given was lost: %+v", d.Ready)
+	}
+}
+
+const gTools, gHt = "novox/mesh-tools", "e3e3e3e3e3e3e3e3e3e3e3e3e3e3e3e3e3e3e3e3"
+
+// groupOf is a green group of the given members (repository, number, head, review body), composed green, the operator not
+// answering by himself.
+func groupOf(t *testing.T, members [][4]string) *world {
+	w := newWorld(t)
+	w.op.auto = false
+	for i, m := range members {
+		p := pr(m[0], i+1, m[2], "feat/g", "src/a.go")
+		p.Body = m[3]
+		w.h.PullUpdated(p)
+	}
+	for i, m := range members {
+		w.h.Checked(judged(m[0], i+1, m[2], "pass", gTrunk0))
+	}
+	w.settleAll()
+	g := w.h.groups["feat/g"]
+	if g == nil || g.Check == nil {
+		t.Fatalf("no composed check was asked: %+v", g)
+	}
+	c := CheckedEvent{Group: "feat/g", ID: g.Check.Asked, Verdict: "pass", Summary: "composed"}
+	for _, m := range members {
+		owner, repo, _ := strings.Cut(m[0], "/")
+		c.Members = append(c.Members, struct {
+			Owner  string `json:"owner"`
+			Repo   string `json:"repo"`
+			Commit string `json:"commit"`
+			Trunk  string `json:"trunk,omitempty"`
+		}{Owner: owner, Repo: repo, Commit: m[2], Trunk: gTrunk0[:12]})
+	}
+	w.h.Checked(c)
+	w.settleAll()
+	return w
+}
+
+func reviewedAt(head string) string { return "Reviewed-by: reviewer@shanks at " + head }
+
+// wantGroupSays checks the group's state, its mesh/delivery-group status and the member's mesh/delivery status.
+func (w *world) wantGroupSays(t *testing.T, label, member, head string, want, notWant string) {
+	t.Helper()
+	g := w.h.groups["feat/g"]
+	state, why, held := w.h.groupState(g, w.h.membersOf(g))
+	if state != "ready" || !held || !strings.Contains(why, want) || (notWant != "" && strings.Contains(why, notWant)) {
+		t.Fatalf("%s: group state %s, why %q, held %v; wanted ready waiting with %q and without %q", label, state, why, held, want, notWant)
+	}
+	for _, c := range []string{"mesh/delivery-group", "mesh/delivery"} {
+		said := w.forge.statuses[member+"@"+head+" "+c]
+		if !strings.HasPrefix(said, "pending") || !strings.Contains(said, want) || (notWant != "" && strings.Contains(said, notWant)) {
+			t.Fatalf("%s: %s on %s says %q; wanted pending with %q and without %q", label, c, member, said, want, notWant)
+		}
+	}
+}
+
+func TestA1803AGroupIsAskedOnlyWhenEveryMemberIsReviewedAtItsHead(t *testing.T) {
+	w := groupOf(t, [][4]string{{gCtl, "", gHc, reviewedAt(gHc)}, {gCat, "", gHk, "Reviewed-by: pending"}})
+	if w.op.count() != 0 {
+		t.Fatalf("1803/AC5: a group with an unreviewed member was asked: %d asks", w.op.count())
+	}
+	w.wantGroupSays(t, "1803/AC5", gCtl, gHc, "mesh-catalog#2", "mesh-controller#1")
+	e := pr(gCat, 2, gHk, "feat/g")
+	e.Body = reviewedAt(gHk)
+	w.h.PullEdited(e)
+	w.settleAll()
+	if len(w.op.open()) != 1 {
+		t.Fatalf("1803/AC5: every member reviewed, yet %d asks", w.op.count())
+	}
+	if said := w.forge.statuses[gCtl+"@"+gHc+" mesh/delivery-group"]; !strings.HasPrefix(said, "success") {
+		t.Fatalf("1803/AC5: the group's status once asked is %q", said)
+	}
+}
+
+func TestA1803TheWaitingWordsFollowTheMembersStillWaiting(t *testing.T) {
+	w := groupOf(t, [][4]string{{gCtl, "", gHc, reviewedAt(gHc)}, {gCat, "", gHk, "Reviewed-by: pending"}, {gTools, "", gHt, "Reviewed-by: pending"}})
+	w.wantGroupSays(t, "1803/AC7 two waiting", gCtl, gHc, "mesh-catalog#2", "")
+	w.wantGroupSays(t, "1803/AC7 two waiting", gCtl, gHc, "mesh-tools#3", "")
+	e := pr(gCat, 2, gHk, "feat/g")
+	e.Body = reviewedAt(gHk)
+	w.h.PullEdited(e)
+	w.settleAll()
+	if w.op.count() != 0 {
+		t.Fatalf("1803/AC7: a member still waits, yet %d asks", w.op.count())
+	}
+	// Only the other member is named now, on the group and on every member's own status.
+	w.wantGroupSays(t, "1803/AC7 one waiting", gCtl, gHc, "mesh-tools#3", "mesh-catalog#2")
+	w.wantGroupSays(t, "1803/AC7 one waiting", gCat, gHk, "mesh-tools#3", "mesh-catalog#2")
+}
+
+func TestA1803APushAfterAnApprovingReviewHoldsTheNewHead(t *testing.T) {
+	w := greenWithBody(t, reviewedAt(readyHead))
+	if len(w.op.open()) != 1 {
+		t.Fatalf("setup: the reviewed head is not asked")
+	}
+	// A new head on the same pull request whose description still names the old commit.
+	const pushed2 = "7a7a7a7a7a7a7a7a7a7a7a7a7a7a7a7a7a7a7a7a"
+	p := pr("novox/app", 7, pushed2, "feat/x", "src/a.go")
+	p.Body = reviewedAt(readyHead)
+	w.h.PullUpdated(p)
+	w.h.Checked(judgedAt("novox/app", 7, pushed2, "pass", "t1t1t1t1t1t1"))
+	w.settleAll()
+	d := w.delivery(IDOf("novox/app", pushed2))
+	if d.Ready == nil || !strings.Contains(d.Ready.Waits, waitsForReview+shortOf(pushed2, 12)) {
+		t.Fatalf("1803/AC8: the new head waits %+v; wanted it to wait for a review at %s", d.Ready, shortOf(pushed2, 12))
+	}
+	if len(w.op.open()) != 0 {
+		t.Fatalf("1803/AC8: the new head was asked on the old review: %v", w.op.open())
+	}
+}

modules/mesh-delivery-gate/cmd/mesh-delivery-gate/merges.go

@@ -287,9 +287,20 @@ func (h *Holder) ForgeRefused(f sdk.Forge) {
 	if r.RefusedSince == nil {
 		r.RefusedSince = &now
 	}
+	r.noteRefusal(d, answer)
 	if !outOfDateRefusal(answer) {
 		r.Urgent = true // never a warning again for this head
 	}
+	if r.Repeats >= maxRepeatedRefusals {
+		// Loud, not endless (novox/hq issue 1106): the same refusal again with the trunk not moved is no trunk's move, so
+		// the merge is not asked again by a verdict or a status; the operator is told, urgent, and decides.
+		r.Urgent = true
+		if r.Repeats == maxRepeatedRefusals {
+			h.Logf("[mesh-delivery] %s: the merge tool refused it %d times in a row with the same words and the trunk not moved, so "+
+				"the merge is asked no more until the trunk moves, a new head is announced or the operator approves again: %s",
+				d.ID, r.Repeats, answer)
+		}
+	}
 	if answer != r.Answer {
 		h.Logf("[mesh-delivery] %s is approved, and the merge tool did not merge it (asked again at the next verdict, Approve or composed check): %s", d.ID,
 			answer)

modules/mesh-delivery-gate/cmd/mesh-delivery-gate/model.go

@@ -132,7 +132,7 @@ type Delivery struct {
 	Ready *ReadyRun `json:"ready,omitempty"`
 	// ReviewedBy and ReviewedAt are the review record its pull request's description holds: the line
 	// `Reviewed-by: <who> at <full sha>` (novox/hq ADR 0421 §5, issue 1022), as written. Read at every head; never a
-	// gate: the ask says first when the line is missing or names another head.
+	// hold: no Approve is asked until it names the head (novox/hq ADR 0496, issue 1803).
 	ReviewedBy string `json:"reviewed_by,omitempty"`
 	ReviewedAt string `json:"reviewed_at,omitempty"`
 	// FreshWanted is whether a passing verdict that arrived out of date is to have its check asked against the
@@ -351,7 +351,7 @@ func (d *Delivery) coveredStop() bool {
 	switch d.State {
 	case Stopped:
 		return d.Stranded != "" && d.CoveredBy != "" || d.DeliveredBy != ""
-	case Failed:
+	case Failed, Superseded:
 		return d.DeliveredBy != ""
 	}
 	return false
@@ -397,6 +397,9 @@ func (d *Delivery) strandedEndedSaid() string {
 
 // coveredStopSaid is how a covered stop or failure reads: on the forge's status, in its view and note, and in `show`.
 func (d *Delivery) coveredStopSaid() string {
+	if d.State == Superseded {
+		return "delivered after its walk was superseded: " + d.DeliveredBy
+	}
 	ended := "its stop"
 	if d.State == Failed {
 		ended = "its failure"
@@ -496,7 +499,7 @@ func ReviewIn(body string) (who, sha string) {
 }
 
 // reviewWords is what the delivery's review record says of its head, for the ask: reviewed by whom, or, first, that it
-// is not reviewed at this head (never a gate, ADR 0421 §5). ok is whether it is reviewed at this head.
+// is not reviewed at this head (the ask is held then, ADR 0496). ok is whether it is reviewed at this head.
 func reviewWords(d *Delivery) (words string, ok bool) {
 	switch {
 	case d.ReviewedBy == "":
@@ -1016,7 +1019,10 @@ type Group struct {
 	Hold *GroupHold `json:"hold,omitempty"`
 	// SaidStatus is the mesh/delivery-group status last written for it: a change of the status a state takes is
 	// written too, though the state did not change (novox/hq issue 902).
-	SaidStatus string   `json:"said_status,omitempty"`
+	SaidStatus string `json:"said_status,omitempty"`
+	// SaidReview is the wait for a review last said of the group while it was held for one (ADR 0496 §2): a change of
+	// the members still waiting is said too, though the state did not change.
+	SaidReview string   `json:"said_review,omitempty"`
 	Owed       []Effect `json:"owed,omitempty"`
 }
 
@@ -1071,3 +1077,13 @@ func mustJSON(v any) json.RawMessage {
 	raw, _ := json.Marshal(v)
 	return raw
 }
+
+// supersededByItsWalk is whether a delivery ended superseded because its walk was: its last move was a supersession
+// from published or delivering, not a newer head of its pull request's (issue 1786).
+func (d *Delivery) supersededByItsWalk() bool {
+	if d.State != Superseded || len(d.Transitions) == 0 {
+		return false
+	}
+	t := d.Transitions[len(d.Transitions)-1]
+	return t.Event == EvSuperseded && (t.From == Published || t.From == Delivering)
+}

modules/mesh-delivery-gate/cmd/mesh-delivery-gate/ready.go

@@ -85,6 +85,11 @@ type ReadyRun struct {
 	// Urgent is set once a refusal of this head was one that is urgent at once: its severity never steps down to a
 	// warning after (issue 1275, AC6).
 	Urgent bool `json:"urgent,omitempty"`
+	// Repeats is how many times in a row the merge tool refused with the same words while the trunk stayed where RepeatTrunk
+	// says: past maxRepeatedRefusals the merge is asked no more by a verdict, a status or a composed check, and the refusal
+	// is urgent (novox/hq issue 1106). A trunk's move, a new head or the operator's Approve starts the count again.
+	Repeats     int    `json:"repeats,omitempty"`
+	RepeatTrunk string `json:"repeat_trunk,omitempty"`
 	// RefusedSince is when the merge tool began refusing it without a break.
 	RefusedSince *time.Time `json:"refused_since,omitempty"`
 	// Merged is when the merge tool merged it.
@@ -214,7 +219,14 @@ func (h *Holder) readyAsks(now time.Time) []func() {
 		r.Failed = ""
 		waits := h.readyWaits(d)
 		if waits != r.Waits {
+			wasHeld := d.holdsForReview()
 			r.Waits = waits
+			if (wasHeld || d.holdsForReview()) && d.shownOnItsPullRequest() {
+				// The status and the view say what it waits for, again whenever the words change while held, and no more once asked.
+				state, desc := statusOf(d)
+				d.owe(Effect{Kind: EffectView, Since: now})
+				d.owe(Effect{Kind: EffectStatus, Commit: d.Commit, Context: "mesh/delivery", State: state, Line: desc, Since: now})
+			}
 			if waits != "" {
 				h.Logf("[mesh-delivery] %s waits: %s", d.ID, waits)
 			}
@@ -322,10 +334,43 @@ func (h *Holder) askFresh(d *Delivery, now time.Time) func() {
 // verdict, an Approve, a group's composed check), so the wait after a refusal does not hold it. Under the lock.
 func (h *Holder) mergeAgainNow(d *Delivery) {
 	if d != nil && d.Ready != nil && d.Ready.Merged == nil {
+		if d.Ready.repeatedRefusal(d) {
+			return // the same refusal again and again with the trunk where it was: asking more changes nothing (issue 1106)
+		}
 		d.Ready.LastTry = time.Time{}
 	}
 }
 
+// maxRepeatedRefusals is how many refusals in a row, with the same words and the trunk not moved, end the asking again: a
+// refusal that repeats with no trunk move is not a trunk's move, and the check asked after each only brings the same words.
+const maxRepeatedRefusals = 3
+
+// repeatedRefusal is whether the merge was refused maxRepeatedRefusals times with these words and the trunk where it is
+// now: it is then asked no more until something changes (a trunk's move, a new head, the operator's Approve).
+func (r *ReadyRun) repeatedRefusal(d *Delivery) bool {
+	return r.Repeats >= maxRepeatedRefusals && r.RepeatTrunk == d.TrunkNow
+}
+
+// refusalKey is a refusal's first sentence, up to the trunk it names ("... and main is now <t>"): what follows says how long the
+// last check took and the like, which differs each time and is no change of cause.
+func refusalKey(answer string) string {
+	if i := strings.Index(answer, " is now "); i >= 0 {
+		if j := strings.Index(answer[i:], ". "); j >= 0 {
+			return answer[:i+j]
+		}
+	}
+	return answer
+}
+
+// noteRefusal counts a refusal: the same words with the trunk where it was count on, anything else starts the count at one.
+func (r *ReadyRun) noteRefusal(d *Delivery, answer string) {
+	if refusalKey(answer) == refusalKey(r.Answer) && r.RepeatTrunk == d.TrunkNow && r.Repeats > 0 {
+		r.Repeats++
+		return
+	}
+	r.Repeats, r.RepeatTrunk = 1, d.TrunkNow
+}
+
 // readyWaits is what a delivery's merge waits on; empty when the Approve may be asked and the merge asked. A ready
 // delivery is asked. So is a proposed one whose verdict went out of date, when an Approve was asked for it already (it
 // stays through a trunk move, ADR 0421 §3) and its check against the trunk's head is not on its way; one without an
@@ -343,12 +388,48 @@ func (h *Holder) readyWaits(d *Delivery) string {
 	default:
 		return "its check has not passed yet (" + string(d.State) + ")"
 	}
+	// ADR 0496 §1, §2: the Approve is asked only for heads with a review recorded at them. An ask already raised or
+	// answered is not touched (covering).
+	if h.covering(d) == nil {
+		if w := h.reviewWaits(d); w != "" {
+			return w
+		}
+	}
 	if d.Group == "" {
 		return ""
 	}
 	return h.groupTurnToMerge(d)
 }
 
+// waitsForReview is how a hold for a review begins, as the delivery's state and its mesh/delivery status say it.
+const waitsForReview = "waits for a review at "
+
+// reviewWaits is why the Approve is not asked yet for want of a review (novox/hq ADR 0496, issue 1803): a delivery
+// whose description holds no review line, `pending` or a line naming another commit, and for a group any member so.
+// Empty when every delivery the ask covers is reviewed at its head. Under the lock.
+func (h *Holder) reviewWaits(d *Delivery) string {
+	members := h.askMembers(d)
+	var waiting []string
+	for _, m := range members {
+		if _, ok := reviewWords(m); ok {
+			continue
+		}
+		if len(members) == 1 {
+			return waitsForReview + shortOf(m.Commit, 12)
+		}
+		waiting = append(waiting, "at "+shortOf(m.Commit, 12)+" ("+pullSaid(m)+")")
+	}
+	if len(waiting) == 0 {
+		return ""
+	}
+	return waitsForReview + strings.Join(waiting, " and ")[len("at "):]
+}
+
+// holdsForReview is whether a delivery's wait is the hold for a review.
+func (d *Delivery) holdsForReview() bool {
+	return d.Ready != nil && strings.HasPrefix(d.Ready.Waits, waitsForReview)
+}
+
 // groupTurnToMerge is why a group member may not merge yet; empty when it may: the group's composed check passed for
 // its members' heads as they are and is not out of date. No member waits for another's merge or delivery (novox/hq
 // ADR 0418 §4): once the composed check passed and the Approve covers the heads, every member is merged in one pass.

modules/mesh-delivery-gate/cmd/mesh-delivery-gate/verbs.go

@@ -134,14 +134,14 @@ func (h *Holder) Show(id string) (any, error) {
 			out["reads"] = d.strandedEndedSaid() // stopped, and nothing can carry what it waited for (issue 1581)
 		}
 		if g := h.groups[d.Group]; g != nil {
-			state, why := GroupState(g, h.membersOf(g))
+			state, why, _ := h.groupState(g, h.membersOf(g))
 			out["group"] = map[string]any{"group": g, "state": state, "why": why}
 		}
 		return out, nil
 	}
 	if g := h.groups[id]; g != nil {
 		members := h.membersOf(g)
-		state, why := GroupState(g, members)
+		state, why, _ := h.groupState(g, members)
 		return map[string]any{"group": g, "state": state, "why": why, "members": members}, nil
 	}
 	return nil, fmt.Errorf("no delivery or group %q: `deliveries` lists them", id)
@@ -184,7 +184,7 @@ func (h *Holder) Groups(all bool) []GroupLine {
 		if len(members) == 0 {
 			continue
 		}
-		state, why := GroupState(g, members)
+		state, why, _ := h.groupState(g, members)
 		if !all && (state == "delivered" || state == "failed" || state == "stopped") {
 			continue
 		}

modules/mesh-provisioner-client/cmd/mesh-provisioner-client/client.go

@@ -119,6 +119,9 @@ type Client struct {
 	Templates *templateStore
 	// Now dates the reports; nil is time.Now.
 	Now func() time.Time
+	// SocketUp wakes the client when the node-engine's write socket is made again (watchSocket); nil never does.
+	// A module whose write failed for want of the socket is filled again then (issue 1165), and nothing else is.
+	SocketUp <-chan struct{}
 
 	// watchBound watches a provider's bucket of entries for this machine's keys, once per run (consumers.go); set by
 	// Run.
@@ -147,6 +150,8 @@ type state struct {
 	// of the consumers left out of one of its files.
 	providing map[string]bool
 	leftOut   map[string]string
+	// unplaced is a module whose last fill failed for want of the node-engine's write socket (issue 1165).
+	unplaced map[string]bool
 	// bound is each provider's bucket of entries this client watches, by key: this machine's consumers' entries alone;
 	// boundReady is a bucket whose current entries are all read, and boundUsers the modules filled from each.
 	bound      map[string]map[string][]byte
@@ -161,6 +166,8 @@ type update struct {
 	changed []byte
 	// asked is the tool runner asking for a module's value (asks.go). Set, entry is nil.
 	asked *nats.Msg
+	// socketUp is the node-engine's write socket made again (issue 1165). Set, entry is nil.
+	socketUp bool
 }
 
 // Run serves until ctx ends or a watch stops, which it returns: run again, it starts from the current values.
@@ -248,6 +255,23 @@ func (c *Client) Run(ctx context.Context) error {
 		return fmt.Errorf("listening for %s: %w", SecretChangedSubject, err)
 	}
 	defer func() { _ = said.Unsubscribe() }()
+	// The node-engine's write socket made again: what failed for want of it is filled again, in the loop below.
+	if c.SocketUp != nil {
+		go func() {
+			for {
+				select {
+				case <-ctx.Done():
+					return
+				case <-c.SocketUp:
+					select {
+					case updates <- update{socketUp: true}:
+					case <-ctx.Done():
+						return
+					}
+				}
+			}
+		}()
+	}
 	// This machine's tool runner asking for a module's config value (asks.go): answered from this run's state, in
 	// its loop, so an answer is never read from half an update.
 	asks, err := c.Conn.Subscribe(AskSubject(c.Node), func(m *nats.Msg) {
@@ -266,7 +290,7 @@ func (c *Client) Run(ctx context.Context) error {
 		reported: map[string][]byte{}, secrets: map[string]heldSecret{}, items: map[string]map[string]itemValue{},
 		recorded: map[string][]byte{}, assigned: map[string]bool{}, binds: map[string]binding{},
 		providing: map[string]bool{}, leftOut: map[string]string{}, bound: map[string]map[string][]byte{},
-		boundReady: map[string]bool{}, boundUsers: map[string]map[string]bool{}}
+		boundReady: map[string]bool{}, unplaced: map[string]bool{}, boundUsers: map[string]map[string]bool{}}
 	// Every current value first: nothing is filled from half of them.
 	current := map[string]bool{}
 	var early []*nats.Msg
@@ -282,6 +306,9 @@ func (c *Client) Run(ctx context.Context) error {
 				early = append(early, u.asked)
 				continue
 			}
+			if u.socketUp {
+				continue // the first fill below tries every file
+			}
 			if u.changed != nil {
 				// Nothing is held yet: every secret is asked for when it is first filled.
 				continue
@@ -326,6 +353,10 @@ func (c *Client) Run(ctx context.Context) error {
 				switch {
 				case u.asked != nil:
 					c.respond(s, u.asked)
+				case u.socketUp:
+					for m := range s.unplaced {
+						dirty[m] = true
+					}
 				case u.changed != nil:
 					for _, m := range c.secretChanged(s, u.changed) {
 						dirty[m] = true
@@ -550,6 +581,7 @@ func (c *Client) leftModules(s *state) []string {
 // leave removes a module's needs and its report: it is no longer the provisioner's on this node.
 func (c *Client) leave(ctx context.Context, kvs map[string]jetstream.KeyValue, s *state, module string) {
 	prefix := c.Node + "." + module + "."
+	delete(s.unplaced, module)
 	removed := 0
 	for _, key := range sortedKeys(s.needs) {
 		if strings.HasPrefix(key, prefix) {
@@ -660,6 +692,7 @@ func (c *Client) reconcile(ctx context.Context, kvs map[string]jetstream.KeyValu
 		c.forgetItems(ctx, kvs, s, module, keep)
 	}
 	moved := false
+	delete(s.unplaced, module)
 	for _, f := range targets {
 		// File by file (issue 1553, ADR 0349 clause 4): a target with a value that waits keeps the file there is.
 		if !allKnown(f.text, values) {
@@ -672,6 +705,9 @@ func (c *Client) reconcile(ctx context.Context, kvs map[string]jetstream.KeyValu
 			err = errFill(a)
 		}
 		if err != nil {
+			if socketAbsent(err) {
+				s.unplaced[module] = true
+			}
 			failed = append(failed, fmt.Sprintf("%s: %v", f.Path, err))
 			continue
 		}

modules/mesh-provisioner-client/cmd/mesh-provisioner-client/client_test.go

@@ -200,6 +200,14 @@ type engine struct {
 }
 
 func startEngine(t *testing.T, owners map[string]string) *engine {
+	t.Helper()
+	e := newEngine(t, owners)
+	e.listen(t)
+	return e
+}
+
+// newEngine is an engine whose socket is not made yet: a node-engine not started, or replacing itself.
+func newEngine(t *testing.T, owners map[string]string) *engine {
 	t.Helper()
 	// A unix socket's path is short: under /tmp, not the test's own directory.
 	dir, err := os.MkdirTemp("/tmp", "p5-")
@@ -209,6 +217,12 @@ func startEngine(t *testing.T, owners map[string]string) *engine {
 	t.Cleanup(func() { os.RemoveAll(dir) })
 	e := &engine{root: t.TempDir(), socket: filepath.Join(dir, "fill.sock"), owner: owners,
 		moved: map[string]int{}, asked: map[string]int{}, wrote: map[string]string{}}
+	return e
+}
+
+// listen makes the socket and answers on it until stop, which removes it as a node-engine stopping does.
+func (e *engine) listen(t *testing.T) (stop func()) {
+	t.Helper()
 	l, err := net.Listen("unix", e.socket)
 	if err != nil {
 		t.Fatal(err)
@@ -233,7 +247,7 @@ func startEngine(t *testing.T, owners map[string]string) *engine {
 			}()
 		}
 	}()
-	return e
+	return func() { l.Close(); os.Remove(e.socket) }
 }
 
 func (e *engine) at(path string) string { return filepath.Join(e.root, path) }
@@ -349,10 +363,14 @@ func aClient(t *testing.T, b *aBus, e *engine, state string) (*Client, func()) {
 		t.Fatal(err)
 	}
 	secrets.Within = 2 * time.Second
-	c := &Client{Node: node, JS: js, Conn: nc, Secrets: secrets,
+	ctx, cancel := context.WithCancel(context.Background())
+	up, err := watchSocket(ctx, e.socket)
+	if err != nil {
+		t.Fatal(err)
+	}
+	c := &Client{Node: node, JS: js, Conn: nc, Secrets: secrets, SocketUp: up,
 		Fill:      SocketFiller{Socket: e.socket, Busy: 10 * time.Millisecond, Tries: 3},
 		Templates: &templateStore{dir: state, fetch: HTTPFetch}}
-	ctx, cancel := context.WithCancel(context.Background())
 	done := make(chan error, 1)
 	go func() { done <- c.Run(ctx) }()
 	stopped := false
@@ -703,3 +721,23 @@ func TestNatsResticsFilesFillFromTheBackupPlacement(t *testing.T) {
 	b.assign(t, "restic", plan("# a\n/srv/a\n", "# a\nstate\n"))
 	eventually(t, "a module taken away refills backups.conf", func() bool { return e.read(t, backups) == head+"# a\n/srv/a\n" })
 }
+
+// defaults writes the defaults a module's manifest declares as mesh-config resolves them (source default), for a
+// moved module whose templates read config values (issue 1780).
+func (b *aBus) defaults(t *testing.T, module string, manifest []byte) {
+	t.Helper()
+	var m struct {
+		Config map[string]struct {
+			Default json.RawMessage `json:"default"`
+		} `json:"config"`
+	}
+	if err := json.Unmarshal(manifest, &m); err != nil {
+		t.Fatal(err)
+	}
+	for key, d := range m.Config {
+		if len(d.Default) > 0 && string(d.Default) != "null" {
+			b.put(t, ConfigBucket, node+"."+module+"."+key, fmt.Sprintf(`{"node":%q,"module":%q,"key":%q,"value":%s,`+
+				`"source":"default"}`, node, module, key, d.Default))
+		}
+	}
+}

modules/mesh-provisioner-client/cmd/mesh-provisioner-client/environment_test.go

@@ -158,6 +158,7 @@ func TestNatsTheEnvironmentOfTheCataloguesModulesIsWhatTheOldFormComposed(t *tes
 				if !manifest.Moved || len(manifest.Items) == 0 {
 					t.Fatalf("%s is not a moved contributor of items", m)
 				}
+				b.defaults(t, m, read(m+"/module.json"))
 				texts := map[string]string{}
 				var planned []PlannedItem
 				for _, it := range manifest.Items {

modules/mesh-provisioner-client/cmd/mesh-provisioner-client/main.go

@@ -133,7 +133,11 @@ func serve(ctx context.Context, credential, socket, stateDir string) error {
 	if err != nil {
 		return err
 	}
-	client := &Client{Node: c.Node, JS: js, Conn: nc, Secrets: secrets,
+	up, err := watchSocket(ctx, socket)
+	if err != nil {
+		return err
+	}
+	client := &Client{Node: c.Node, SocketUp: up, JS: js, Conn: nc, Secrets: secrets,
 		Fill:      SocketFiller{Socket: socket, Busy: time.Second, Tries: 10},
 		Templates: &templateStore{dir: filepath.Join(stateDir, "templates"), fetch: HTTPFetch}}
 	wait := time.Second

modules/mesh-provisioner-client/cmd/mesh-provisioner-client/socket_integration_test.go

@@ -0,0 +1,108 @@
+//go:build integration
+
+package main
+
+// Issue 1165: these tests make a real unix socket and watch its directory, so they run in the integration level.
+
+import (
+	"context"
+	"net"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+	"time"
+)
+
+// Issue 1165: a file whose write failed because the node-engine's socket was gone is written when the socket is made
+// again, with no timer; a file that was placed is not asked again by it.
+func TestNatsAFillThatFailedForWantOfTheSocketIsTriedWhenTheSocketReturns(t *testing.T) {
+	b := startBus(t)
+	heard := b.listen(t)
+	e := newEngine(t, map[string]string{dunstrc: "owned"})
+	b.config(t, "dunst", "font-size", "13")
+	b.assign(t, "dunst", dunst(t))
+	aClient(t, b, e, t.TempDir())
+
+	eventually(t, "the failure is said", func() bool {
+		r, _ := heard.last("dunst")
+		return r.Status == StatusFailed && strings.Contains(r.Reason, "fill.sock")
+	})
+	if got := e.read(t, dunstrc); got != "" {
+		t.Fatalf("written without a socket: %q", got)
+	}
+	e.listen(t)
+	eventually(t, "dunstrc is written once the socket is back", func() bool {
+		return strings.Contains(e.read(t, dunstrc), "Monospace 13")
+	})
+	eventually(t, "and the report says applied", func() bool {
+		r, _ := heard.last("dunst")
+		return r.Status == StatusApplied
+	})
+}
+
+func TestNatsAReturningSocketAsksNothingOfAModuleThatWasPlaced(t *testing.T) {
+	b := startBus(t)
+	e := newEngine(t, map[string]string{dunstrc: "owned"})
+	stop := e.listen(t)
+	b.config(t, "dunst", "font-size", "13")
+	b.assign(t, "dunst", dunst(t))
+	aClient(t, b, e, t.TempDir())
+	eventually(t, "dunstrc is filled", func() bool { return strings.Contains(e.read(t, dunstrc), "Monospace 13") })
+	_, before := e.counts(dunstrc)
+
+	// The node-engine replaces itself: the socket goes and returns, and a later change proves the loop ran past it.
+	stop()
+	e.listen(t)
+	b.config(t, "dunst", "font-size", "15")
+	eventually(t, "the change is written", func() bool { return strings.Contains(e.read(t, dunstrc), "Monospace 15") })
+	_, after := e.counts(dunstrc)
+	if after-before != 1 {
+		t.Errorf("dunstrc was asked %d times for one change", after-before)
+	}
+}
+
+func TestWatchSocketWakesWhenTheSocketIsMadeAndWhenItsDirectoryReturns(t *testing.T) {
+	base, err := os.MkdirTemp("/tmp", "ws-")
+	if err != nil {
+		t.Fatal(err)
+	}
+	t.Cleanup(func() { os.RemoveAll(base) })
+	dir := filepath.Join(base, "engine")
+	if err := os.Mkdir(dir, 0o755); err != nil {
+		t.Fatal(err)
+	}
+	socket := filepath.Join(dir, "fill.sock")
+	ctx, cancel := context.WithCancel(context.Background())
+	defer cancel()
+	up, err := watchSocket(ctx, socket)
+	if err != nil {
+		t.Fatal(err)
+	}
+	heard := func(what string) {
+		t.Helper()
+		select {
+		case <-up:
+		case <-time.After(5 * time.Second):
+			t.Fatalf("no wake for %s", what)
+		}
+	}
+	l, err := net.Listen("unix", socket)
+	if err != nil {
+		t.Fatal(err)
+	}
+	heard("the socket made")
+	l.Close()
+	// The whole directory goes, as with a runtime directory removed on stop, and returns with the socket.
+	if err := os.RemoveAll(dir); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.Mkdir(dir, 0o755); err != nil {
+		t.Fatal(err)
+	}
+	if l, err = net.Listen("unix", socket); err != nil {
+		t.Fatal(err)
+	}
+	defer l.Close()
+	heard("the directory and socket back")
+}

modules/mesh-provisioner-client/cmd/mesh-provisioner-client/socketwatch.go

@@ -0,0 +1,103 @@
+package main
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"os"
+	"path/filepath"
+	"syscall"
+	"unsafe"
+
+	"golang.org/x/sys/unix"
+)
+
+// socketAbsent says whether a write failed because nothing answers at the node-engine's write socket: its file is
+// gone, or it is there with no one listening. That is the node-engine replacing itself (issue 1165); the write is
+// tried again when the socket is back (watchSocket).
+func socketAbsent(err error) bool {
+	return errors.Is(err, syscall.ENOENT) || errors.Is(err, syscall.ECONNREFUSED)
+}
+
+// watchSocket tells, on the channel it returns, each time the file at socket is made, moved in or changed in its
+// attributes — a node-engine that starts makes its write socket and sets its mode — until ctx ends. It is the
+// operating system's own event (inotify), not a timer: the client is on the node-engine's machine. Events that
+// come while one is unread are one. When the socket's directory is gone it watches the directory above for its
+// return. The channel only wakes: what is owed a write is the client's to read.
+func watchSocket(ctx context.Context, socket string) (<-chan struct{}, error) {
+	fd, err := unix.InotifyInit1(unix.IN_CLOEXEC | unix.IN_NONBLOCK)
+	if err != nil {
+		return nil, fmt.Errorf("watching %s: %w", socket, err)
+	}
+	f := os.NewFile(uintptr(fd), "inotify")
+	out := make(chan struct{}, 1)
+	dir := filepath.Dir(socket)
+	// The directory's watch and, while the directory is absent, its parent's.
+	if err := addWatches(fd, dir); err != nil {
+		_ = f.Close()
+		return nil, fmt.Errorf("watching %s: %w", socket, err)
+	}
+	go func() { <-ctx.Done(); _ = f.Close() }()
+	go func() {
+		buf := make([]byte, 16*(unix.SizeofInotifyEvent+unix.NAME_MAX+1))
+		for {
+			n, err := f.Read(buf)
+			if err != nil {
+				return
+			}
+			wake, rewatch := false, false
+			for off := 0; off+unix.SizeofInotifyEvent <= n; {
+				ev := (*unix.InotifyEvent)(unsafe.Pointer(&buf[off]))
+				name := ""
+				if ev.Len > 0 {
+					raw := buf[off+unix.SizeofInotifyEvent : off+unix.SizeofInotifyEvent+int(ev.Len)]
+					for i, b := range raw {
+						if b == 0 {
+							raw = raw[:i]
+							break
+						}
+					}
+					name = string(raw)
+				}
+				off += unix.SizeofInotifyEvent + int(ev.Len)
+				switch {
+				case ev.Mask&(unix.IN_DELETE_SELF|unix.IN_MOVE_SELF|unix.IN_IGNORED) != 0:
+					rewatch = true
+				case name == filepath.Base(socket):
+					wake = true
+				case name == filepath.Base(dir):
+					// The directory came back: watched again, and the socket may already be in it.
+					rewatch, wake = true, true
+				}
+			}
+			if rewatch {
+				if err := addWatches(fd, dir); err != nil {
+					logf("[provisioner-client] watching %s again: %v", socket, err)
+				}
+				if _, err := os.Lstat(socket); err == nil {
+					wake = true
+				}
+			}
+			if wake {
+				select {
+				case out <- struct{}{}:
+				default:
+				}
+			}
+		}
+	}()
+	return out, nil
+}
+
+// addWatches watches dir for what is made in it, and its parent for dir's own return; a watch already there is kept.
+func addWatches(fd int, dir string) error {
+	if _, err := unix.InotifyAddWatch(fd, filepath.Dir(dir), unix.IN_CREATE|unix.IN_MOVED_TO); err != nil {
+		return err
+	}
+	_, err := unix.InotifyAddWatch(fd, dir, unix.IN_CREATE|unix.IN_MOVED_TO|unix.IN_ATTRIB|unix.IN_DELETE_SELF|
+		unix.IN_MOVE_SELF)
+	if errors.Is(err, syscall.ENOENT) {
+		return nil // absent: its parent's watch tells when it is made
+	}
+	return err
+}

modules/mesh-provisioner-client/cmd/mesh-provisioner-client/xorg_golden_test.go

@@ -63,6 +63,7 @@ func TestNatsTheSessionFilesFilledFromTheirContributorsItemsAreWhatTheOldFormCom
 				if !c.Moved || len(c.Shell) != 0 {
 					t.Fatalf("%s is not moved or still holds shell entries", m)
 				}
+				b.defaults(t, m, []byte(read(m+"/module.json")))
 				texts := map[string]string{}
 				var planned []PlannedItem
 				for _, it := range c.Items {

modules/mesh-provisioner-client/go.mod

@@ -6,6 +6,7 @@ require (
 	git.novox.be/novox/mesh-sdk/go v0.1.14-0.20261021114014-66e8fdaa62e2
 	github.com/nats-io/nats-server/v2 v2.11.17
 	github.com/nats-io/nats.go v1.54.0
+	golang.org/x/sys v0.48.0
 )
 
 require (
@@ -17,6 +18,5 @@ require (
 	github.com/nats-io/nkeys v0.4.16 // indirect
 	github.com/nats-io/nuid v1.0.1 // indirect
 	golang.org/x/crypto v0.57.0 // indirect
-	golang.org/x/sys v0.48.0 // indirect
 	golang.org/x/time v0.15.0 // indirect
 )

modules/minio/client.ts

@@ -256,11 +256,22 @@ export class MinioClient {
     if (argv.some((a) => a.includes(this.rootPassword))) {
       throw new Error("refused to run mc: its command line would carry the root password (novox/hq issue 282)");
     }
-    const { stdout } = await execFileAsync(this.mcBin, argv, {
-      timeout: 30_000,
-      env: { ...process.env, MC_HOST_mesh: mcHost(this.baseUrl, this.rootUser, this.rootPassword) },
-    });
-    return stdout.trim();
+    const mcHostValue = mcHost(this.baseUrl, this.rootUser, this.rootPassword);
+    try {
+      const { stdout } = await execFileAsync(this.mcBin, argv, {
+        timeout: 30_000,
+        env: { ...process.env, MC_HOST_mesh: mcHostValue },
+      });
+      return stdout.trim();
+    } catch (err) {
+      // A fresh error with only the redacted message: execFile's own carries the whole command line, and
+      // its cmd, stdout and stderr properties would carry it on (issue 1766).
+      const known = [mcHostValue, this.rootPassword, encodeURIComponent(this.rootPassword), ...secretArgs(argv)];
+      const e = err as { stderr?: unknown; message?: string };
+      const reason = typeof e.stderr === "string" && e.stderr.trim() !== "" ? e.stderr.trim() : (e.message ?? String(err));
+      const shown = argv.filter((a) => a !== this.mcConfigDir).join(" ");
+      throw new Error(redactSecrets(`mc ${shown} failed: ${reason}`, known));
+    }
   }
 
   // --- SigV4 request plumbing ---------------------------------------------
@@ -360,6 +371,26 @@ function uriEncode(str: string, encodeSlash: boolean): string {
   return out;
 }
 
+/** The values that follow --secret-key and --access-key in an argument list. */
+function secretArgs(argv: string[]): string[] {
+  const out: string[] = [];
+  argv.forEach((a, i) => { if ((a === "--secret-key" || a === "--access-key") && i + 1 < argv.length) out.push(argv[i + 1]); });
+  return out;
+}
+
+/**
+ * The one redaction every `mc` failure goes through (issue 1766): the token after --secret-key or
+ * --access-key (as a separate word or after "="), any MC_HOST_ value, and each known secret, wherever they
+ * appear in the text, become "[redacted]".
+ */
+export function redactSecrets(text: string, known: string[] = []): string {
+  let out = text
+    .replace(/(--(?:secret|access)-key(?:=|\s+))('[^']*'|"[^"]*"|\S+)/g, "$1[redacted]")
+    .replace(/(MC_HOST_\w+=)\S+/g, "$1[redacted]");
+  for (const k of known) if (k !== "") out = out.split(k).join("[redacted]");
+  return out;
+}
+
 /**
  * The MC_HOST_<alias> URL `mc` reads an alias from: the endpoint with the credentials in its userinfo,
  * each percent-encoded so a reserved character in either cannot end it early.

modules/minio/links.ts

@@ -36,6 +36,9 @@ export function linkPolicy(bucket: string): string {
   });
 }
 
+/** A link key's secret: 40 hex characters, minio's limit for a secret key (issue 1766). */
+const LINK_SECRET = /^[0-9a-f]{40}$/;
+
 export function linkSecretPath(dir: string, bucket: string): string {
   return join(dir, `${bucket}.secret`);
 }
@@ -47,9 +50,12 @@ export async function ensureLinkKey(admin: LinkAdmin, dir: string, bucket: strin
   const path = linkSecretPath(dir, bucket);
   mkdirSync(dir, { recursive: true, mode: 0o700 });
   let secret = existsSync(path) ? readFileSync(path, "utf8").trim() : "";
+  // A stored secret of the wrong shape (an earlier version made 64 characters; minio takes at most 40) is
+  // replaced, and its key made again (issue 1766).
+  if (!LINK_SECRET.test(secret)) secret = "";
   if (secret !== "" && (await admin.accessKeyExists(id))) return;
   if (secret === "") {
-    secret = randomBytes(32).toString("hex");
+    secret = randomBytes(20).toString("hex");
     writeFileSync(path, secret + "\n", { mode: 0o600 });
     chmodSync(path, 0o600);
   }

modules/minio/test/links.test.ts

@@ -115,3 +115,58 @@ test("the link key id is the one the object-store tool derives (cmd/object-store
   // 1516/AC1
   assert.equal(linkKeyId("brother"), "link0b4bd77cec705cf5");
 });
+
+// Issue 1766: a link key's secret is minio's length, and no mc failure carries an argument's secret.
+
+import { chmodSync, writeFileSync, mkdirSync } from "node:fs";
+import { MinioClient } from "../client.ts";
+
+test("a link key's secret is 40 characters, and a stored 64-character secret is replaced", async () => {
+  // 1766/AC1
+  const { minio, dir } = setup();
+  await ensureLinkKey(minio, dir, "fresh");
+  assert.equal(minio.keys.get(linkKeyId("fresh"))!.secret.length, 40);
+  assert.match(minio.keys.get(linkKeyId("fresh"))!.secret, /^[0-9a-f]{40}$/);
+
+  mkdirSync(dir, { recursive: true });
+  const old = "ab".repeat(32);
+  writeFileSync(linkSecretPath(dir, "old"), old + "\n");
+  await ensureLinkKey(minio, dir, "old");
+  const made = minio.keys.get(linkKeyId("old"))!.secret;
+  assert.equal(made.length, 40);
+  assert.notEqual(made, old);
+  assert.equal(readFileSync(linkSecretPath(dir, "old"), "utf8").trim(), made);
+});
+
+function failingMc(): { client: MinioClient; root: string } {
+  const d = mkdtempSync(join(process.env.TMPDIR ?? tmpdir(), "minio-mc-"));
+  const bin = join(d, "mc");
+  // Says its arguments and its alias on stderr, then fails, the way a refused command does.
+  writeFileSync(bin, '#!/bin/sh\necho "args: $*" >&2\necho "env: MC_HOST_mesh=$MC_HOST_mesh" >&2\nexit 1\n');
+  chmodSync(bin, 0o755);
+  const root = "RootPass-s3cret/ok";
+  const client = new MinioClient({ endpoint: "http://127.0.0.1:9000", rootUser: "root", rootPassword: root, region: "us-east-1", mcBin: bin, mcConfigDir: d });
+  return { client, root };
+}
+
+test("an mc failure's message holds no --secret-key value and no MC_HOST value", async () => {
+  // 1766/AC2
+  const { client, root } = failingMc();
+  const secret = "SuperSecretValue0123456789";
+  const err: any = await client.createAccessKey("b", "ACCESSKEYID1234567", secret).then(() => null, (e) => e);
+  assert.ok(err, "the failure is raised");
+  for (const text of [String(err.message), String(err.stack ?? ""), JSON.stringify(Object.getOwnPropertyNames(err).map((k) => err[k]))]) {
+    assert.ok(!text.includes(secret), "no secret key");
+    assert.ok(!text.includes("ACCESSKEYID1234567"), "no access key");
+    assert.ok(!text.includes(root) && !text.includes(encodeURIComponent(root)), "no root password");
+    assert.ok(!text.includes("root:"), "no MC_HOST value");
+  }
+  assert.match(err.message, /\[redacted\]/);
+});
+
+test("a consumer key's create passes the minted secret unchanged", async () => {
+  // 1766/AC3
+  const s = setup();
+  await s.h.create(provision);
+  assert.equal(s.minio.keys.get("brother-key")!.secret, "consumer-secret");
+});