mesh-build-go asks a run's input of the seat verb the ask names, and reads no checker state (issue 1921, B0 review round 1)

The operator's answer: reading mesh-delivery-checker.inputs broke ADR 0596 §2. The manifest drops that read and invokes
seat:mesh-delivery-checker.input; the run step asks the verb through the runtime (stdio.Ask) and takes only the bytes
the ask names (delivery.RunInput.Read).

Claude-Session: https://claude.ai/code/session_01NcbgfFSnn24M3GcqttoU7i

modules/mesh-build-go/README.md

@@ -49,6 +49,10 @@ already gone is no error, one a running step holds is refused); its caller calls
 failed fetch removes its own workspace. A workspace no caller released is removed two hours after its last use (ADR
 0509), and the removal is said in the log as the caller's fault.
 
+The programs this module builds from a commit and keeps in its cache (composers of compose asks, checking programs of run
+asks) are read by `build_go_kept` (kind, key, size, last use) and removed by `build_go_kept_drop {kind, key}`; one no ask used
+for 30 days is swept with the workspaces (issue 1921).
+
 A pin on a pull request's head (issue 1148): when `go mod download` fails on a `git.novox.be/novox/*` pseudo-version, the
 fetch looks for that commit among the forge's `refs/pull/*/head`, places the repository at it under the workspace's
 `pinned/beside/` as a plain directory, and hands it to Go (the fetch's commands, and the build with `GOPROXY=off`) through a

modules/mesh-build-go/buildstep/isolate.go

@@ -27,16 +27,19 @@ const Bubblewrap = "bwrap"
 //   - the system's programs and libraries read-only, and of /etc only what a program needs to run (users, libraries,
 //     certificates, time zone), with a hosts file naming only the loopback;
 //   - the Workspace and Writable directories read-write, the ReadOnly ones (the toolchain, the modules the fetch step
-//     downloaded) read-only;
+//     downloaded) read-only, and the ReadOnlyInside ones (inside the Workspace) read-only over it;
 //   - the shared warm Caches read-only under a throwaway writable layer (S18), or read-write for a trunk build alone
 //     (WriteCaches), so a candidate's code never changes what a later build reads.
 type Isolation struct {
-	Workspace   string
-	Writable    []string
-	ReadOnly    []string
-	Caches      []string
-	WriteCaches bool
-	Env         []string
+	Workspace string
+	Writable  []string
+	ReadOnly  []string
+	// ReadOnlyInside are directories inside the Workspace or a Writable one bound read-only after them, so they stay
+	// read-only (a later bind of a path wins; a ReadOnly one inside the Workspace would be covered by it).
+	ReadOnlyInside []string
+	Caches         []string
+	WriteCaches    bool
+	Env            []string
 	// Hosts is the file bound as /etc/hosts; Isolated writes one naming only the loopback.
 	Hosts string
 }
@@ -91,7 +94,7 @@ func (iso Isolation) command(env []string, name string, args ...string) ([]strin
 	for _, group := range []struct {
 		what string
 		dirs []string
-	}{{"writable directory", iso.Writable}, {"read-only directory", iso.ReadOnly}, {"cache", iso.Caches}} {
+	}{{"writable directory", iso.Writable}, {"read-only directory", iso.ReadOnly}, {"read-only directory", iso.ReadOnlyInside}, {"cache", iso.Caches}} {
 		for _, d := range group.dirs {
 			if err := checkDir(group.what, d); err != nil {
 				return nil, err
@@ -118,6 +121,9 @@ func (iso Isolation) command(env []string, name string, args ...string) ([]strin
 	for _, d := range iso.Writable {
 		argv = append(argv, "--bind", d, d)
 	}
+	for _, d := range iso.ReadOnlyInside {
+		argv = append(argv, "--ro-bind", d, d)
+	}
 	if iso.Hosts != "" {
 		argv = append(argv, "--ro-bind", iso.Hosts, "/etc/hosts")
 	}

modules/mesh-build-go/buildstep/pull.go

@@ -561,6 +561,36 @@ func (p *Puller) declineRun(msg *Pulled, ask delivery.StepAsk, key string, step
 	p.mu.Unlock()
 }
 
+// refuseUnread ends an ask this module cannot read: when its data holds an id that keys a step, the step is written
+// refused with the reason, so the asker learns at once that no node began it; otherwise (no id, not JSON) the ask is only
+// dropped and logged. The message is dropped either way.
+func (p *Puller) refuseUnread(msg *Pulled, why error) {
+	var probe struct {
+		ID   string `json:"id"`
+		Kind string `json:"kind"`
+	}
+	if json.Unmarshal(msg.Data, &probe) == nil && probe.ID != "" {
+		if key, kerr := delivery.AskStepKey(p.Node, probe.ID); kerr == nil {
+			tool := "ask"
+			if probe.Kind != "" {
+				tool = probe.Kind + " ask"
+			}
+			at := now().UTC()
+			step := delivery.Step{ID: probe.ID, Tool: tool, Node: p.Node, State: delivery.StepRefused, Started: at, Ended: at,
+				Reason: fmt.Sprintf("this build module on %s cannot read the ask: %v", p.Node, why), Owner: sdkstep.Owner}
+			if werr := writeEnd(p.Put, key, step); werr != nil {
+				p.logf("the unread ask %s cannot be answered on %s: %v", probe.ID, p.Node, werr)
+			} else {
+				p.logf("an ask refused: %v", why)
+				_ = msg.Term()
+				return
+			}
+		}
+	}
+	p.logf("an ask dropped: %v", why)
+	_ = msg.Term()
+}
+
 func withKind(msg *Pulled, kind string) *Pulled {
 	if msg != nil {
 		msg.Kind = kind
@@ -576,9 +606,10 @@ func (p *Puller) take(ctx context.Context, msg *Pulled) {
 		err = fmt.Errorf("the ask %s is a %s ask of %s, pulled from the %s %s worker", ask.ID, ask.Kind, ask.Language, p.Language, kind)
 	}
 	if err != nil {
-		// An ask that is not one runs nowhere: delivering it again would only fail again.
-		p.logf("an ask dropped: %v", err)
-		_ = msg.Term()
+		// An ask that is not one runs nowhere: delivering it again would only fail again. But an ask that names an id
+		// is answered on its step key as refused, naming why (a job this build module does not know, a version it does
+		// not read): a dropped ask must always be answered, or its asker waits to its bound for nothing.
+		p.refuseUnread(msg, err)
 		return
 	}
 	if lacks := delivery.Lacks(ask.Needs, p.Has); len(lacks) > 0 {

modules/mesh-build-go/buildstep/pull_test.go

@@ -519,3 +519,39 @@ func TestAPoisonedAskIsSaidAndRunNoMore(t *testing.T) {
 		t.Errorf("acked %v naked %v, want it termed alone", acked, naked)
 	}
 }
+
+// An ask whose job this build module does not know is answered on its step key as refused, naming why, never only dropped
+// (a dropped ask leaves its asker waiting to its bound); one with no id is still dropped.
+func TestPullAnswersAnAskItCannotRead(t *testing.T) {
+	src, st := NewFakeAsks(), newSteps()
+	good := testAsk(delivery.AskCheck)
+	raw, err := delivery.EncodeStepAsk(good)
+	if err != nil {
+		t.Fatal(err)
+	}
+	var m map[string]any
+	_ = json.Unmarshal(raw, &m)
+	delete(m, "check")
+	m["frobnicate"] = map[string]any{"program": "x"}
+	odd, _ := json.Marshal(m)
+	// The job is not one this reader holds, so the ask carries none it knows: decode refuses it.
+	if _, err := delivery.DecodeStepAsk(odd); err == nil {
+		t.Fatal("the fixture ask decodes")
+	}
+	src.Publish(delivery.AskCheck, odd)
+	src.Publish(delivery.AskCheck, []byte(`{"v":1}`))
+	p := &Puller{Source: src, Language: delivery.AskGo, Node: "ace", Put: st.put, Capacity: func() int { return 1 },
+		PullWait: 10 * time.Millisecond, Run: func(ctx context.Context, a delivery.StepAsk) (any, error) {
+			t.Error("an unread ask ran")
+			return nil, nil
+		}}
+	ctx, cancel := context.WithCancel(context.Background())
+	defer cancel()
+	go p.Pull(ctx)
+	waitFor(t, "both termed", func() bool { _, tm, _ := src.Fates(); return len(tm) == 2 })
+	key, _ := delivery.AskStepKey("ace", good.ID)
+	s, ok := st.get(key)
+	if !ok || s.State != delivery.StepRefused || !strings.Contains(s.Reason, "cannot read the ask") || !strings.Contains(s.Reason, "exactly one") {
+		t.Fatalf("the unread ask's step = %+v (%v)", s, ok)
+	}
+}

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

@@ -121,6 +121,9 @@ func (m *Module) Puller(steps []StepTool, source buildstep.AskSource) *buildstep
 			if a.Compose != nil {
 				return m.compose(ctx, ask.Fetch, a)
 			}
+			if a.Run != nil {
+				return m.run(ctx, ask.Fetch, a)
+			}
 			return buildstep.RunAsk(ctx, ask, a)
 		}}
 }

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

@@ -38,7 +38,7 @@ var composerLocks sync.Map
 
 // composerPath is where the composer of key is kept: the cache directory, beside Go's own caches.
 func (m *Module) composerPath(key string) string {
-	return filepath.Join(m.Cache, "composers", key, "mesh-compose")
+	return m.programPath("composers", key)
 }
 
 // compose runs a compose ask: the composer of its commit, built once, over its request with the manifests filled.
@@ -98,17 +98,25 @@ func (m *Module) manifestOf(digest string) ([]byte, error) {
 
 // composerOf is the composer kept under key, or, when none is, the one built now from the fetched commit (built true).
 func (m *Module) composerOf(ctx context.Context, fetch buildstep.StepFunc, job delivery.ComposeJob, key string) (string, bool, error) {
-	exe := m.composerPath(key)
-	mu, _ := composerLocks.LoadOrStore(exe, &sync.Mutex{})
-	mu.(*sync.Mutex).Lock()
-	defer mu.(*sync.Mutex).Unlock()
+	return m.programOf(ctx, fetch, "composers", "composer", key, job.Fetch, job.Dir, job.Package)
+}
+
+// programOf is the program kept under key in the cache directory kind, or, when none is, the one built now: the commit
+// fetch names is fetched (the fetch step: the forge key and the network, none of the change's code), and package in the
+// Go module of dir is built in a run step shut off with nothing fetched. what names it in a failure.
+func (m *Module) programOf(ctx context.Context, fetch buildstep.StepFunc, kind, what, key string, fetchArgs map[string]any, dir, pkg string) (string, bool, error) {
+	exe := m.programPath(kind, key)
+	mu := lockOf(exe)
+	mu.Lock()
+	defer mu.Unlock()
 	if st, err := os.Stat(exe); err == nil && st.Mode().IsRegular() {
+		touch(exe, m.clock())
 		return exe, false, nil
 	}
 	if fetch == nil {
 		return "", false, buildstep.Refuse("this build module fetches nothing")
 	}
-	out, err := fetch(ctx, job.Fetch)
+	out, err := fetch(ctx, fetchArgs)
 	if err != nil {
 		return "", false, err
 	}
@@ -125,11 +133,11 @@ func (m *Module) composerOf(ctx context.Context, fetch buildstep.StepFunc, job d
 	defer m.releaseSaid(f.Workspace)
 	done := m.spaces().Hold(f.Workspace)
 	defer done()
-	dir := filepath.Join(m.Work, f.Workspace, "tree", job.Dir)
-	if _, err := os.Stat(filepath.Join(dir, "go.mod")); err != nil {
-		return "", false, fmt.Errorf("%s at %s holds no go.mod in %s, so no composer can be built there", job.Fetch["repository"], job.Fetch["commit"], job.Dir)
+	tree := filepath.Join(m.Work, f.Workspace, "tree", dir)
+	if _, err := os.Stat(filepath.Join(tree, "go.mod")); err != nil {
+		return "", false, fmt.Errorf("%s at %s holds no go.mod in %s, so no %s can be built there", fetchArgs["repository"], fetchArgs["commit"], dir, what)
 	}
-	into := filepath.Join(m.Work, f.Workspace, "composer", "mesh-compose")
+	into := filepath.Join(m.Work, f.Workspace, what, "built")
 	if err := os.MkdirAll(filepath.Dir(into), 0o700); err != nil {
 		return "", false, err
 	}
@@ -137,15 +145,15 @@ func (m *Module) composerOf(ctx context.Context, fetch buildstep.StepFunc, job d
 	if err != nil {
 		return "", false, err
 	}
-	env := append(m.goEnvFor(dir), "GOPROXY=off", "GOTOOLCHAIN=local", "CGO_ENABLED=0")
-	r := run(ctx, env, "go", "-C", dir, "build", "-trimpath", "-o", into, job.Package)
+	env := append(m.goEnvFor(tree), "GOPROXY=off", "GOTOOLCHAIN=local", "CGO_ENABLED=0")
+	r := run(ctx, env, "go", "-C", tree, "build", "-trimpath", "-o", into, pkg)
 	if r.Status != 0 || r.Err != "" {
 		// The words Go says when the module holds no such package (a mesh-sdk from before the composer) are kept whole,
 		// so the asker tells that apart from a fault.
-		return "", false, fmt.Errorf("the composer %s does not build at %s: %s", job.Package, job.Fetch["commit"], compileError(r.Err+"\n"+r.Stderr+"\n"+r.Stdout))
+		return "", false, fmt.Errorf("the %s %s does not build at %s: %s", what, pkg, fetchArgs["commit"], compileError(r.Err+"\n"+r.Stderr+"\n"+r.Stdout))
 	}
 	if err := copyExecutable(into, exe); err != nil {
-		return "", false, fmt.Errorf("the composer could not be kept: %w", err)
+		return "", false, fmt.Errorf("the %s could not be kept: %w", what, err)
 	}
 	return exe, true, nil
 }

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

@@ -150,11 +150,17 @@ func TestAComposeAskFailsLoudly(t *testing.T) {
 	}
 }
 
-// 1622: the module reads the delivery indexer's manifests, which a compose ask names by digest.
+// 1622: the module reads the delivery indexer's manifests, which a compose ask names by digest. 1921 (B0): a run's input
+// is asked of the checker seat's verb input, as the manifest invokes it, and no state of the checker's inputs is read
+// (ADR 0596 §2).
 func TestTheModuleReadsTheIndexersManifests(t *testing.T) {
-	if m := readManifest(t); !slices.Contains(m.Reads, ManifestsState) {
+	m := readManifest(t)
+	if !slices.Contains(m.Reads, ManifestsState) {
 		t.Errorf("reads = %v, lacks %s", m.Reads, ManifestsState)
 	}
+	if !slices.Contains(m.Invokes, runInputVerb) || slices.Contains(m.Reads, "mesh-delivery-checker.inputs") {
+		t.Errorf("invokes = %v, reads = %v: want %s invoked and no inputs state read", m.Invokes, m.Reads, runInputVerb)
+	}
 }
 
 // ADR 0593 §7 (C9): a compose ask naming the node-engine's repository builds its composer command in the Go module at the

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

@@ -348,6 +348,7 @@ func (m *Module) goEnvFor(dir string) []string {
 // net for a caller that never released one (issue 1839, ADR 0509).
 func (m *Module) CleanWorkspaces(now time.Time) {
 	m.sweepSaid(m.spaces().Sweep(now, buildstep.UnheldBound))
+	m.SweepKept(now, KeptBound)
 }
 
 // clock is the time now: Now when a test set it, else the wall clock.

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

@@ -0,0 +1,130 @@
+package main
+
+// **Kept programs have tools and a sweep** (novox/hq issue 1921, the seams review's must 2): the programs this module builds
+// from a commit and keeps under its cache so each commit is built once on a node, the composers of compose asks and the
+// programs of run asks, are read by build_go_kept, removed one by one by build_go_kept_drop, and swept when unused for
+// KeptBound. A use touches the program's file, so its modification time is its last use.
+
+import (
+	"fmt"
+	"os"
+	"path/filepath"
+	"sort"
+	"sync"
+	"time"
+
+	stdio "git.novox.be/novox/mesh-sdk/go"
+)
+
+// The tools of the kept programs.
+const (
+	KeptTool     = "build_go_kept"
+	KeptDropTool = "build_go_kept_drop"
+)
+
+// KeptBound is how long a kept program that no ask used is kept: a month, longer than any group lives, so only the build
+// of a commit nobody asks about any more goes.
+const KeptBound = 30 * 24 * time.Hour
+
+// keptKinds are the cache directories of kept programs and the name of the program in each key's directory.
+var keptKinds = map[string]string{"composers": "mesh-compose", "programs": "program"}
+
+// KeptProgram is one kept program: its kind (the cache directory), its key, its size and when an ask last used it.
+type KeptProgram struct {
+	Kind     string    `json:"kind"`
+	Key      string    `json:"key"`
+	Bytes    int64     `json:"bytes"`
+	LastUsed time.Time `json:"last-used"`
+}
+
+// programPath is where the program of kind and key is kept.
+func (m *Module) programPath(kind, key string) string {
+	return filepath.Join(m.Cache, kind, key, keptKinds[kind])
+}
+
+// lockOf is the mutex that guards the program at path against a build or a use while it is dropped.
+func lockOf(path string) *sync.Mutex {
+	l, _ := composerLocks.LoadOrStore(path, &sync.Mutex{})
+	return l.(*sync.Mutex)
+}
+
+// touch records a use of the kept program at path.
+func touch(path string, at time.Time) { _ = os.Chtimes(path, at, at) }
+
+// Kept lists the kept programs, composers then programs, each kind by key.
+func (m *Module) Kept() []KeptProgram {
+	out := []KeptProgram{}
+	for _, kind := range []string{"composers", "programs"} {
+		entries, _ := os.ReadDir(filepath.Join(m.Cache, kind))
+		for _, e := range entries {
+			if !e.IsDir() {
+				continue
+			}
+			st, err := os.Stat(m.programPath(kind, e.Name()))
+			if err != nil || !st.Mode().IsRegular() {
+				continue
+			}
+			out = append(out, KeptProgram{Kind: kind, Key: e.Name(), Bytes: st.Size(), LastUsed: st.ModTime().UTC()})
+		}
+	}
+	return out
+}
+
+// DropKept removes the kept program of kind and key; one already gone is no error.
+func (m *Module) DropKept(kind, key string) error {
+	if _, ok := keptKinds[kind]; !ok {
+		return fmt.Errorf("a kept program is of kind composers or programs, not %q", kind)
+	}
+	if key == "" || key != filepath.Base(key) || key == "." || key == ".." {
+		return fmt.Errorf("%q is no kept program's key", key)
+	}
+	exe := m.programPath(kind, key)
+	l := lockOf(exe)
+	l.Lock()
+	defer l.Unlock()
+	return os.RemoveAll(filepath.Dir(exe))
+}
+
+// SweepKept removes the kept programs no ask used within bound before now, and answers what it removed.
+func (m *Module) SweepKept(now time.Time, bound time.Duration) []KeptProgram {
+	var gone []KeptProgram
+	for _, k := range m.Kept() {
+		if now.Sub(k.LastUsed) <= bound {
+			continue
+		}
+		if m.DropKept(k.Kind, k.Key) == nil {
+			gone = append(gone, k)
+		}
+	}
+	sort.Slice(gone, func(i, j int) bool { return gone[i].Key < gone[j].Key })
+	if len(gone) > 0 && m.Say != nil {
+		m.Say("removed %d kept programs no ask used within %s", len(gone), bound)
+	}
+	return gone
+}
+
+// keptTools are the tools that read and drop the kept programs: no guard and no floor, since they only free room.
+func (m *Module) keptTools() []stdio.Tool {
+	return []stdio.Tool{
+		{Name: KeptTool,
+			Description: "The programs this node keeps, built from a commit so each commit is built once: the composers of compose asks and " +
+				"the checking programs of run asks, each with its kind, key, size in bytes and when an ask last used it. One no ask used " +
+				"for 30 days is removed. Read-only.",
+			Run: func(map[string]any) (any, error) { return map[string]any{"kept": m.Kept()}, nil }},
+		{Name: KeptDropTool,
+			Description: "Remove one kept program by its kind (composers or programs) and key, as build_go_kept names them: the next ask " +
+				"for that commit builds it again. One already gone is no error.",
+			Input: map[string]any{"type": "object", "properties": map[string]any{
+				"kind": map[string]any{"type": "string", "enum": []string{"composers", "programs"}},
+				"key":  map[string]any{"type": "string", "description": "the key build_go_kept names"},
+			}, "required": []string{"kind", "key"}},
+			Run: func(args map[string]any) (any, error) {
+				kind, _ := args["kind"].(string)
+				key, _ := args["key"].(string)
+				if err := m.DropKept(kind, key); err != nil {
+					return nil, err
+				}
+				return map[string]any{"dropped": key, "kind": kind}, nil
+			}},
+	}
+}

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

@@ -0,0 +1,70 @@
+package main
+
+import (
+	"os"
+	"path/filepath"
+	"testing"
+	"time"
+)
+
+func keepProgram(t *testing.T, m *Module, kind, key string, size int, used time.Time) {
+	t.Helper()
+	exe := m.programPath(kind, key)
+	if err := os.MkdirAll(filepath.Dir(exe), 0o700); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.WriteFile(exe, make([]byte, size), 0o700); err != nil {
+		t.Fatal(err)
+	}
+	touch(exe, used)
+}
+
+func toolOf(t *testing.T, m *Module, name string) func(map[string]any) (any, error) {
+	t.Helper()
+	for _, tool := range m.Tools(m.StepTools()) {
+		if tool.Name == name {
+			return tool.Run
+		}
+	}
+	t.Fatalf("no tool %s", name)
+	return nil
+}
+
+// Seams review must 2 (issue 1921): the kept composers and programs are listed with their size and last use, dropped one
+// by one, and swept with the workspaces when unused past KeptBound; a use keeps a program.
+func TestKeptProgramsAreListedDroppedAndSwept(t *testing.T) {
+	m, _ := fetchModule(t, &fakeStep{})
+	now := time.Now().UTC().Truncate(time.Second)
+	keepProgram(t, m, "composers", "old", 10, now.Add(-KeptBound-time.Hour))
+	keepProgram(t, m, "programs", "recent", 20, now.Add(-time.Hour))
+	keepProgram(t, m, "programs", "gone", 30, now)
+
+	out, err := toolOf(t, m, KeptTool)(nil)
+	if err != nil {
+		t.Fatal(err)
+	}
+	list := out.(map[string]any)["kept"].([]KeptProgram)
+	if len(list) != 3 || list[0].Kind != "composers" || list[0].Key != "old" || list[0].Bytes != 10 || !list[0].LastUsed.Equal(now.Add(-KeptBound-time.Hour)) {
+		t.Fatalf("the kept programs = %+v", list)
+	}
+
+	for _, bad := range []map[string]any{{"kind": "nope", "key": "gone"}, {"kind": "programs", "key": "../composers"}, {"kind": "programs", "key": ""}} {
+		if _, err := toolOf(t, m, KeptDropTool)(bad); err == nil {
+			t.Errorf("dropped with %v", bad)
+		}
+	}
+	for i := 0; i < 2; i++ { // a second drop is no error
+		if _, err := toolOf(t, m, KeptDropTool)(map[string]any{"kind": "programs", "key": "gone"}); err != nil {
+			t.Fatalf("drop %d: %v", i, err)
+		}
+	}
+	if _, err := os.Stat(m.programPath("programs", "gone")); !os.IsNotExist(err) {
+		t.Error("the dropped program is still kept")
+	}
+
+	// The sweep runs with the workspaces' (CleanWorkspaces, on its timer).
+	m.CleanWorkspaces(now)
+	if got := m.Kept(); len(got) != 1 || got[0].Key != "recent" {
+		t.Fatalf("kept after the sweep = %+v", got)
+	}
+}

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

@@ -58,6 +58,7 @@ func main() {
 	m.Settings = m.Home + "/settings.json"
 	m.ReadKept = readOf(stdio.State(KeptState)).Get
 	m.ReadManifest = readOf(stdio.State(ManifestsState)).Get
+	m.AskInput = stdio.Ask
 	m.Probe = ProbeGo(m.Home, m.Cache)
 	m.Bubblewrap = ProbeBubblewrap
 	// Outside a launch by the runtime — a build, a check — there is no node and no state to write.

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

@@ -93,7 +93,12 @@ type Module struct {
 	// ReadManifest reads a key of the delivery indexer's manifests state (ManifestsState), which a compose ask's request
 	// names its manifests by (compose.go); GoModCache, when set, is the Go module cache in place of Cache's (a test).
 	ReadManifest func(key string) (json.RawMessage, error)
-	GoModCache   string
+	// AskInput asks a seat verb the module invokes (module.json) as the runtime calls it (stdio.Ask): the verb a run ask's
+	// input names, answering the input by its digest (delivery.RunInput.Read).
+	AskInput func(key string, body any) (json.RawMessage, error)
+	// PlaceHead, when set, lays a run's head at a directory instead of fetching it (a test).
+	PlaceHead  func(ctx context.Context, h delivery.RunHead, dir string) error
+	GoModCache string
 
 	spacesOnce sync.Once
 	workspaces *buildstep.Workspaces

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

@@ -115,7 +115,7 @@ func TestAGoThatCannotBeRunRefusesEveryStepAndHealthIsUnhealthy(t *testing.T) {
 		{Name: "short_probe", Run: run},
 	}, m.StepTools()...)
 	for _, tool := range m.Tools(steps) {
-		if tool.Name == StatusTool || tool.Name == HealthTool || tool.Name == StepReadTool || tool.Name == StepsReadTool || tool.Name == ToolchainsReadTool || tool.Name == ReleaseTool || tool.Name == StepEndTool {
+		if tool.Name == StatusTool || tool.Name == HealthTool || tool.Name == StepReadTool || tool.Name == StepsReadTool || tool.Name == ToolchainsReadTool || tool.Name == ReleaseTool || tool.Name == KeptTool || tool.Name == KeptDropTool || tool.Name == StepEndTool {
 			continue
 		}
 		var reason string

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

@@ -0,0 +1,132 @@
+package main
+
+// **A run ask runs a program built from a change, over heads it reads** (novox/hq ADR 0644, ADR 0656; issue 1921): the
+// checking program of a group's check. The module builds the program at the commit the ask names, once per
+// delivery.RunKey, as a compose ask builds a composer (programOf); fetches every head the ask names and lays each read-only
+// under the directory MESH_RUN_HEADS names, at `<owner>/<name>`; asks the seat verb the ask's input names for it by its
+// digest (as module.json invokes it; never a read of the asker's state, ADR 0596 §2) and writes it to the file
+// MESH_RUN_INPUT names; and runs the program shut off with the ask's
+// arguments: no network, none of the account. The answer is delivery.RunAnswer: the program's exit, the last line of its
+// standard output and the first line of its standard error. A program that fails is an answer, not a fault of the step.
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"os"
+	"path/filepath"
+	"strings"
+
+	"git.novox.be/novox/mesh-catalog/modules/mesh-build-go/buildstep"
+	sdkstep "git.novox.be/novox/mesh-sdk/go/buildstep"
+	"git.novox.be/novox/mesh-sdk/go/delivery"
+)
+
+// run runs a run ask: the program of its commit, built once, over its heads and its input.
+func (m *Module) run(ctx context.Context, fetch buildstep.StepFunc, a delivery.StepAsk) (any, error) {
+	job := *a.Run
+	body, err := m.inputOf(job.Input)
+	if err != nil {
+		return nil, err
+	}
+	installed, err := m.Probe(ctx)
+	if err != nil {
+		return nil, err
+	}
+	key := delivery.RunKey(job.Program, delivery.ToolchainUsed{Version: installed, Recipe: m.Recipe.Version()}, a.Packages)
+	fetchArgs := sdkstep.FetchRequest{Repository: job.Program.Repository, Commit: job.Program.Commit, Packages: a.Packages}.Args()
+	exe, built, err := m.programOf(ctx, fetch, "programs", "checking program", key, fetchArgs, job.Program.Dir, job.Program.Package)
+	if err != nil {
+		return nil, err
+	}
+	ans, err := m.runProgram(ctx, exe, job, body)
+	if err != nil {
+		return nil, err
+	}
+	ans.Built, ans.Key = built, key
+	// The answer is written on the step key as one value: one over the bus's value limit (the input's bound) fails the
+	// step, naming its size, before it is written.
+	if raw, err := json.Marshal(ans); err != nil {
+		return nil, err
+	} else if len(raw) > delivery.RunInputBound {
+		return nil, fmt.Errorf("the checking program's answer is %d bytes, over the bound of %d bytes a step key holds", len(raw), delivery.RunInputBound)
+	}
+	return ans, nil
+}
+
+// inputOf is the input a run names, asked of the seat verb the ask names (module.json invokes it) and checked against
+// its digest and size: bytes that are not the input named refuse the ask.
+func (m *Module) inputOf(in delivery.RunInput) ([]byte, error) {
+	if m.AskInput == nil {
+		return nil, errors.New("this module asks no seat verb for a run's input")
+	}
+	raw, err := in.Read(m.AskInput)
+	if err != nil && raw != nil {
+		return nil, buildstep.Refuse("%v", err)
+	}
+	return raw, err
+}
+
+// placeHead lays one head's commit as a plain directory at dir, with no symlink in it; a test sets its own.
+func (m *Module) placeHead(ctx context.Context, h delivery.RunHead, dir string) error {
+	if m.PlaceHead != nil {
+		return m.PlaceHead(ctx, h, dir)
+	}
+	if err := os.MkdirAll(dir, 0o700); err != nil {
+		return err
+	}
+	if err := buildstep.FetchCommit(ctx, m.Run, dir, m.remote(h.Repository), h.Commit, "", m.fetchOptions(ctx, h.Repository, h.Commit, "")...); err != nil {
+		return err
+	}
+	if err := m.dissociate(ctx, dir); err != nil {
+		return err
+	}
+	if link := firstSymlink(dir); link != "" {
+		return fmt.Errorf("%s holds a symlink, %s", h.Repository, link)
+	}
+	return nil
+}
+
+// runProgram lays the heads and the input in a workspace of its own and runs exe shut off over them.
+func (m *Module) runProgram(ctx context.Context, exe string, job delivery.RunJob, input []byte) (delivery.RunAnswer, error) {
+	id := sdkstep.NewID()
+	done := m.spaces().Hold(id)
+	defer m.releaseSaid(id)
+	defer done()
+	ws := filepath.Join(m.Work, id)
+	heads := filepath.Join(ws, "heads")
+	for _, h := range job.Heads {
+		if err := m.placeHead(ctx, h, filepath.Join(heads, filepath.FromSlash(h.Repository))); err != nil {
+			return delivery.RunAnswer{}, fmt.Errorf("the head of %s at %s: %w", h.Repository, h.Commit, err)
+		}
+	}
+	if err := os.MkdirAll(heads, 0o700); err != nil {
+		return delivery.RunAnswer{}, err
+	}
+	file := filepath.Join(ws, "input")
+	if err := os.WriteFile(file, input, 0o600); err != nil {
+		return delivery.RunAnswer{}, err
+	}
+	run := m.Run
+	if !m.Inside {
+		iso := m.isolation(id, false)
+		iso.ReadOnly = append(iso.ReadOnly, exe)
+		iso.ReadOnlyInside = []string{heads} // bound after the workspace, so the program cannot change a head
+		run = buildstep.Isolated(m.Run, iso)
+	}
+	r := run(ctx, []string{delivery.RunHeadsEnv + "=" + heads, delivery.RunInputEnv + "=" + file}, exe, job.Args...)
+	if r.Err != "" && r.Status == 0 {
+		return delivery.RunAnswer{}, fmt.Errorf("the checking program could not run: %s", r.Err)
+	}
+	ans := delivery.RunAnswer{Exit: r.Status, Said: firstLine(r.Err + "\n" + r.Stderr)}
+	line := strings.TrimSpace(lastLine(r.Stdout))
+	switch {
+	case line == "":
+	case json.Valid([]byte(line)):
+		ans.Output = json.RawMessage(line)
+	default:
+		ans.Output, _ = json.Marshal(line)
+	}
+	return ans, nil
+}

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

@@ -0,0 +1,141 @@
+//go:build integration
+
+package main
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"net"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/nats-io/nats.go"
+	"github.com/nats-io/nats.go/jetstream"
+
+	"git.novox.be/novox/mesh-catalog/modules/mesh-build-go/buildstep"
+	"git.novox.be/novox/mesh-sdk/go/delivery"
+	"git.novox.be/novox/mesh-sdk/go/testbus"
+)
+
+// 1921/C3: a run ask put on the Go build seat's check stream is pulled by the module started whole on the real bus and
+// answered on its step key <node>.<ask id>: the program, run shut off over its heads and its input (asked by digest of the
+// seat verb the ask names, which a fake checker answers on the bus as the checker's verb input does), answers its last
+// line and exit; a program cannot write under its heads, and one
+// that dials the machine's network is shut off, and the answer says it failed.
+func TestARunAskIsAnsweredOnTheBusAndItsProgramIsShutOff(t *testing.T) {
+	needTools(t, "the shut-off run", "bash", "bwrap")
+	nc, err := nats.Connect(testbus.URL(t))
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer nc.Close()
+	js, err := jetstream.New(nc)
+	if err != nil {
+		t.Fatal(err)
+	}
+	ctx, stop := context.WithCancel(context.Background())
+	defer stop()
+	if err := buildstep.DeclareAskWorkers(ctx, js, delivery.AskGo, 2*time.Minute); err != nil {
+		t.Fatal(err)
+	}
+	t.Cleanup(func() { _ = js.DeleteStream(context.Background(), "SEAT_NODE_BUILD_GO") })
+	tcp, err := net.Listen("tcp", "127.0.0.1:0")
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer tcp.Close()
+
+	states := newStates()
+	m := moduleWith(states, "go1.26.0")
+	m.Work, m.Cache, m.Home = t.TempDir(), t.TempDir(), t.TempDir()
+	m.Capacity = func() int { return 1 }
+	m.Run = StepRunner(m.Home)
+	for _, c := range m.caches() {
+		if err := os.MkdirAll(c, 0o700); err != nil {
+			t.Fatal(err)
+		}
+	}
+	m.PlaceHead = func(_ context.Context, h delivery.RunHead, dir string) error { return os.MkdirAll(dir, 0o700) }
+	body := `{"v":1}`
+	// The checker's verb input on the bus, and the module asking it as the runtime calls a seat verb it invokes.
+	asked := make(chan string, 4)
+	sub, err := nc.Subscribe("mesh.seat.mesh-delivery-checker.tool.input", func(msg *nats.Msg) {
+		var ask delivery.RunInputAsk
+		_ = json.Unmarshal(msg.Data, &ask)
+		asked <- ask.Digest
+		reply, _ := json.Marshal(map[string]any{"result": delivery.RunInputAnswer{Digest: ask.Digest, Input: []byte(body)}})
+		_ = msg.Respond(reply)
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer sub.Unsubscribe()
+	m.AskInput = func(key string, b any) (json.RawMessage, error) {
+		if key != runInputVerb {
+			return nil, errors.New("asked " + key)
+		}
+		data, _ := json.Marshal(b)
+		msg, err := nc.Request("mesh.seat.mesh-delivery-checker.tool.input", data, 5*time.Second)
+		if err != nil {
+			return nil, err
+		}
+		var r struct {
+			Result json.RawMessage `json:"result"`
+		}
+		return r.Result, json.Unmarshal(msg.Data, &r)
+	}
+
+	a := runAskOf(t, body)
+	key := delivery.RunKey(a.Run.Program, delivery.ToolchainUsed{Version: "go1.26.0", Recipe: m.Recipe.Version()}, nil)
+	exe := m.programPath("programs", key)
+	os.MkdirAll(filepath.Dir(exe), 0o700)
+	script := "#!/bin/bash\n" + `if touch "$MESH_RUN_HEADS/novox/mesh-sdk/x" 2>/dev/null; then echo '{"wrote":true}'; exit 9; fi; echo "start"; exec 3<>/dev/tcp/127.0.0.1/` + strings.Split(tcp.Addr().String(), ":")[1] + " || exit 7\n" + `echo '{"dialed":true}'` + "\n"
+	if err := os.WriteFile(exe, []byte(script), 0o700); err != nil {
+		t.Fatal(err)
+	}
+	p := m.Puller(fakeSteps(new([]string)), &buildstep.JetStreamAsks{JS: js, Language: delivery.AskGo})
+	p.PullWait = 200 * time.Millisecond
+	go p.Pull(ctx)
+
+	raw, err := delivery.EncodeStepAsk(a)
+	if err != nil {
+		t.Fatal(err)
+	}
+	subject, _ := delivery.AskSubject(delivery.AskGo, delivery.AskCheck)
+	if _, err := js.Publish(ctx, subject, raw); err != nil {
+		t.Fatal(err)
+	}
+	stepKey, _ := delivery.AskStepKey("ace", a.ID)
+	var step delivery.Step
+	for deadline := time.Now().Add(60 * time.Second); time.Now().Before(deadline); time.Sleep(50 * time.Millisecond) {
+		states.mu.Lock()
+		v := states.steps[stepKey]
+		states.mu.Unlock()
+		if len(v) > 0 && v[len(v)-1].State != delivery.StepRunning {
+			step = v[len(v)-1]
+			break
+		}
+	}
+	if step.State != delivery.StepPassed {
+		t.Fatalf("the run ask ended %q: %s", step.State, step.Reason)
+	}
+	var ans delivery.RunAnswer
+	if err := json.Unmarshal(step.Result, &ans); err != nil {
+		t.Fatal(err)
+	}
+	if ans.Key != key || ans.Exit != 7 || strings.Contains(string(ans.Output), "dialed") || string(ans.Output) != `"start"` {
+		t.Fatalf("the answer of a program that dials out = %+v output %s", ans, ans.Output)
+	}
+	select {
+	case d := <-asked:
+		if d != a.Run.Input.Digest {
+			t.Errorf("the verb input was asked %s, want %s", d, a.Run.Input.Digest)
+		}
+	default:
+		t.Error("the run's input was not asked of the checker's verb input")
+	}
+}

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

@@ -0,0 +1,154 @@
+package main
+
+import (
+	"context"
+	"crypto/sha256"
+	"encoding/hex"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+	"time"
+
+	"git.novox.be/novox/mesh-catalog/modules/mesh-build-go/buildstep"
+	sdkstep "git.novox.be/novox/mesh-sdk/go/buildstep"
+	"git.novox.be/novox/mesh-sdk/go/composer"
+	"git.novox.be/novox/mesh-sdk/go/delivery"
+	"git.novox.be/novox/mesh-sdk/go/forgekey"
+)
+
+const runCommit = "5555555555555555555555555555555555555555"
+
+// runInputVerb is the seat verb a run ask's input is served on, as the checker names it in the ask.
+const runInputVerb = "seat:mesh-delivery-checker.input"
+
+// answersInput is a seat verb serving body by its digest, as the checker's verb input answers (delivery.RunInputAnswer).
+func answersInput(body string) func(key string, b any) (json.RawMessage, error) {
+	return func(key string, b any) (json.RawMessage, error) {
+		ask, _ := b.(delivery.RunInputAsk)
+		if sum := sha256.Sum256([]byte(body)); key != runInputVerb || ask.Digest != hex.EncodeToString(sum[:]) {
+			return nil, fmt.Errorf("the checker keeps no input %s", ask.Digest)
+		}
+		return json.Marshal(delivery.RunInputAnswer{Digest: ask.Digest, Input: []byte(body)})
+	}
+}
+
+func runAskOf(t *testing.T, body string) delivery.StepAsk {
+	t.Helper()
+	in, err := delivery.RunInputOf(runInputVerb, []byte(body))
+	if err != nil {
+		t.Fatal(err)
+	}
+	return delivery.StepAsk{ID: sdkstep.NewID(), Kind: delivery.AskCheck, Language: delivery.AskGo, Asker: "mesh-delivery-checker", At: time.Now(),
+		Run: &delivery.RunJob{Program: delivery.RunProgram{Repository: "novox/mesh-delivery", Commit: runCommit, Dir: ".", Package: "./cmd/prog"},
+			Heads: []delivery.RunHead{{Repository: "novox/mesh-sdk", Commit: runCommit}, {Repository: "novox/mesh-host", Commit: runCommit}},
+			Input: in, Args: []string{"-seats"}}}
+}
+
+// 1921/C1: a kept checking program is run over the input checked against its digest and the heads laid under
+// MESH_RUN_HEADS at <owner>/<name>, and the answer carries its exit and last line; nothing is fetched or built; a
+// program that fails, or says no JSON, is an answer; an input that is not the one named is refused; and the program was
+// kept under delivery.RunKey, so a second ask finds it.
+func TestAKeptProgramIsRunOverItsHeadsAndInput(t *testing.T) {
+	m := composeModule(t)
+	body := `{"v":1,"case":"x"}`
+	m.AskInput = answersInput(body)
+	m.PlaceHead = func(_ context.Context, h delivery.RunHead, dir string) error {
+		return os.MkdirAll(dir, 0o700)
+	}
+	a := runAskOf(t, body)
+	key := delivery.RunKey(a.Run.Program, delivery.ToolchainUsed{Version: "go1.26.0", Recipe: m.Recipe.Version()}, nil)
+	exe := m.programPath("programs", key)
+	os.MkdirAll(filepath.Dir(exe), 0o700)
+	script := "#!/bin/sh\n" +
+		`echo "noise"; echo "args $*" >&2; ` +
+		`printf '{"input":%s,"heads":"%s"}\n' "$(cat "$MESH_RUN_INPUT")" "$(cd "$MESH_RUN_HEADS" && ls -d */* | tr '\n' ' ')"; exit 3` + "\n"
+	if err := os.WriteFile(exe, []byte(script), 0o700); err != nil {
+		t.Fatal(err)
+	}
+	fetched := 0
+	fetch := func(context.Context, map[string]any) (any, error) { fetched++; return nil, nil }
+	out, err := m.run(context.Background(), fetch, a)
+	if err != nil {
+		t.Fatal(err)
+	}
+	ans := out.(delivery.RunAnswer)
+	if ans.Built || ans.Key != key || fetched != 0 || ans.Exit != 3 || ans.Said != "args -seats" ||
+		string(ans.Output) != `{"input":{"v":1,"case":"x"},"heads":"novox/mesh-host novox/mesh-sdk "}` {
+		t.Fatalf("answer = %+v (fetched %d) output %s", ans, fetched, ans.Output)
+	}
+	if entries, _ := os.ReadDir(m.Work); len(entries) != 0 {
+		t.Errorf("the run's workspace was left: %v", entries)
+	}
+	// Shut off, the heads are bound read-only after the writable workspace, so a write under MESH_RUN_HEADS fails: a
+	// read-only bind before the workspace's is covered by it.
+	m.Inside = false
+	var argv []string
+	m.Run = func(_ context.Context, _ []string, _ string, args ...string) forgekey.Ran {
+		argv = args
+		return forgekey.Ran{Status: 0, Stdout: "{}"}
+	}
+	if _, err := m.run(context.Background(), fetch, a); err != nil {
+		t.Fatal(err)
+	}
+	line, ws, heads := " "+strings.Join(argv, " ")+" ", -1, -1
+	for i := 0; i+2 < len(argv); i++ {
+		switch {
+		case argv[i] == "--bind" && filepath.Dir(argv[i+1]) == m.Work:
+			ws = i
+		case argv[i] == "--ro-bind" && filepath.Base(argv[i+1]) == "heads" && argv[i+2] == argv[i+1]:
+			heads = i
+		}
+	}
+	if ws < 0 || heads < ws || !strings.Contains(line, " "+delivery.RunHeadsEnv+" "+argv[heads+1]+" ") {
+		t.Errorf("the heads are not bound read-only over the workspace: %s", line)
+	}
+	// An answer over the bus's value limit fails the step, naming its size.
+	m.Run = func(context.Context, []string, string, ...string) forgekey.Ran {
+		return forgekey.Ran{Status: 0, Stdout: `"` + strings.Repeat("x", delivery.RunInputBound) + `"`}
+	}
+	if _, err := m.run(context.Background(), fetch, a); err == nil || !strings.Contains(err.Error(), "over the bound") {
+		t.Errorf("an answer over the bound: %v", err)
+	}
+	m.Inside, m.Run = true, StepRunner(t.TempDir())
+	// An input other than the one named is refused, naming it; one the verb does not answer fails the step, saying why.
+	m.AskInput = func(_ string, b any) (json.RawMessage, error) {
+		return json.Marshal(delivery.RunInputAnswer{Digest: b.(delivery.RunInputAsk).Digest, Input: []byte(`{"other":1}`)})
+	}
+	if _, err := m.run(context.Background(), fetch, a); err == nil || !strings.Contains(err.Error(), "not the one the ask names") || !errors.As(err, new(*buildstep.Refusal)) {
+		t.Errorf("a wrong input: %v", err)
+	}
+	m.AskInput = answersInput(`{"v":2}`)
+	if _, err := m.run(context.Background(), fetch, a); err == nil || !strings.Contains(err.Error(), "keeps no input") {
+		t.Errorf("an input the verb does not answer: %v", err)
+	}
+}
+
+// 1921/C2 (F2): a compose ask's seats reach the composer: the module decodes and fills the request and writes it again,
+// and used to drop the seats.
+func TestAComposeAskKeepsItsSeats(t *testing.T) {
+	m := composeModule(t)
+	read, digests := indexerHolds(t, `{"module":"demo","version":"1"}`)
+	m.ReadManifest = read
+	a := composeAskOf(t, digests[0])
+	req := composer.Request{V: composer.Version, Seats: []json.RawMessage{json.RawMessage(`{"seat":"node-x"}`)},
+		Cases: []composer.Case{{Name: "change", Manifests: map[string]composer.Manifest{"demo": {Digest: digests[0]}}}}}
+	a.Compose.Request, _ = json.Marshal(req)
+	key := delivery.ComposerKey("novox/mesh-sdk", composeCommit, "go", composer.Package, delivery.ToolchainUsed{Version: "go1.26.0", Recipe: m.Recipe.Version()}, nil)
+	exe := m.composerPath(key)
+	os.MkdirAll(filepath.Dir(exe), 0o700)
+	script := "#!/bin/sh\n" + `if grep -q '"seats":\[{"seat":"node-x"}\]' "$2"; then S=true; else S=false; fi; ` +
+		`echo '{"v":1,"sdk":"stand-in","cases":[{"name":"change","manifests":{"demo":{"accepted":'$S'}}}]}'` + "\n"
+	os.WriteFile(exe, []byte(script), 0o700)
+	out, err := m.compose(context.Background(), nil, a)
+	if err != nil {
+		t.Fatal(err)
+	}
+	var v composer.Verdict
+	if err := json.Unmarshal(out.(delivery.ComposeAnswer).Verdict, &v); err != nil || !v.Cases[0].Manifests["demo"].Accepted {
+		t.Fatalf("the composer did not receive the seats: %+v %v", v, err)
+	}
+}

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

@@ -45,6 +45,7 @@ func (m *Module) Tools(steps []StepTool) []stdio.Tool {
 	tools = append(tools, m.ReadTools()...)
 	tools = append(tools, buildstep.StepEndTool(StepEndTool, m.Node, m.stepScan(), m.Held))
 	tools = append(tools, m.releaseTool())
+	tools = append(tools, m.keptTools()...)
 	return append(tools,
 		stdio.Tool{Name: StatusTool,
 			Description: "What the Go build module is on this machine: the Go version its recipe builds with and the one installed, " +

modules/mesh-build-go/module.json

@@ -23,6 +23,9 @@
     "mesh-delivery-checker.kept",
     "mesh-delivery-indexer.manifests"
   ],
+  "invokes": [
+    "seat:mesh-delivery-checker.input"
+  ],
   "config": {
     "capacity": {
       "type": "number",
@@ -50,6 +53,8 @@
     "build_go_build",
     "build_go_imports",
     "build_go_release",
+    "build_go_kept",
+    "build_go_kept_drop",
     "build_go_step",
     "build_go_steps",
     "build_go_step_end",