One forge user, nox-agent, for every machine's agents' SSH pushes, each machine's key titled by machine (issue 991, ADR 0410)
mesh/merge-gate head 3a2560dc51fc merged into trunk 82df3bd58913 — pass: builds gitea → novox; no bus step; every machine composes with the change as it di…
mesh/repo-check head 3a2560dc51fc merged into trunk 82df3bd58913 — pass: its merge-check.sh passed, each part in its toolchain: merge-check.sh (go), merge-…
mesh/delivery delivered

This commit is contained in:
2026-10-17 00:46:43 +00:00
parent 82df3bd589
commit 3a2560dc51
9 changed files with 313 additions and 64 deletions
+1 -1
View File
@@ -195,7 +195,7 @@ func (m *Minter) Answer(ask Ask) (Answer, error) {
skip(h, "the forge names no origin it is reached at from that machine")
continue
}
account := forgeaccount.ForgeUser(h.Node, "machine")
account := forgeaccount.NodeAccount(h.Node)
if err := m.accounts().EnsureAccount(ctx, account); err != nil {
return Answer{}, err
}
@@ -314,12 +314,12 @@ func TestTheProviderWholeAnswersAMachineThatReportedItsKey(t *testing.T) {
e, ok = rt.boundEntry("g14.ssh-client.forge-account")
return ok
})
if e.Values["user"] != "agent-g14" || e.Values["ssh_host"] != "novox.internal" ||
if e.Values["user"] != "nox-agent" || e.Values["ssh_host"] != "novox.internal" ||
e.Values["known_host"] != "[novox.internal]:222 ssh-ed25519 AAAAC3HOSTKEY" || e.Values["host_key"] != "ssh-ed25519 AAAAC3HOSTKEY" || e.Values["ssh_alias"] != "mesh-forge" ||
e.Values["ssh_base"] != "ssh://git@novox.internal:222/" || e.Values["https_base"] != "https://git.novox.be/" {
t.Fatalf("values = %v", e.Values)
}
if got := forge.titles("agent-g14"); len(got) != 1 || got[0] != "mesh:g14.nox" {
if got := forge.titles("nox-agent"); len(got) != 1 || got[0] != "mesh:g14.nox" {
t.Fatalf("registered keys = %v", got)
}
@@ -329,7 +329,7 @@ func TestTheProviderWholeAnswersAMachineThatReportedItsKey(t *testing.T) {
eventually(t, "the new key registered and the old one gone", func() bool {
forge.mu.Lock()
defer forge.mu.Unlock()
k := forge.held["agent-g14"]
k := forge.held["nox-agent"]
return len(k) == 1 && strings.HasPrefix(k[0]["key"].(string), pubB)
})
}
@@ -2,7 +2,7 @@
// the node's runtime launches and speaks MCP to over stdio, beside the module's other bundles.
//
// For each machine whose ssh-client binds the provision and has reported its public key (ssh-client's state `keys`),
// it makes the machine's agent a forge account of its own (agent-<node>, in the agents team), registers the key for
// it makes the machine's agent the one forge user nox-agent (in the agents team; ADR 0410), registers the key for
// it, and serves where and as whom to reach the forge over ssh: the user, the forge machine's mesh name and port, its
// host key and the two url roots. A public key is no secret, so the provision has no credential.
//
@@ -69,6 +69,8 @@ type Service struct {
bound provider.Bus // the bus the entries are written to (set by Provider)
letGo map[string]bool // nodes whose per-node user of before has been said to be let go
making sync.Mutex // one Make at a time: they register keys, and two at once would race each other
}
@@ -81,7 +83,7 @@ func NewService(cfg Config) *Service {
if cfg.Log == nil {
cfg.Log = func(f string, a ...any) { fmt.Fprintf(os.Stderr, "[gitea forge-account] "+f+"\n", a...) }
}
return &Service{cfg: cfg, keys: map[string]machineKey{}, Pause: time.Sleep,
return &Service{cfg: cfg, keys: map[string]machineKey{}, letGo: map[string]bool{}, Pause: time.Sleep,
accounts: &forgeaccount.Accounts{Forge: cfg.Forge, Log: cfg.Log}}
}
@@ -127,9 +129,12 @@ func (s *Service) Make(ctx context.Context, c provider.Consumer) error {
}
have := s.keysOf(node)
if len(have) == 0 {
if err := s.releasePerNodeUser(ctx, node); err != nil {
return err
}
// A key the machine no longer reports is no longer wanted (ADR 0400 §2): the mesh's keys go, the operator's
// stay, and the entry that served the machine goes with them.
if err := s.accounts.RemoveKeysTitled(ctx, user, forgeaccount.TitlePrefix+node+"."); err != nil {
if _, err := s.accounts.RemoveKeysTitled(ctx, user, forgeaccount.MachineScope(node)); err != nil {
return err
}
if s.bound != nil {
@@ -156,7 +161,13 @@ func (s *Service) Make(ctx context.Context, c provider.Consumer) error {
for _, k := range have {
wanted = append(wanted, forgeaccount.WantedKey{Title: forgeaccount.KeyTitle(k.Node, k.Account), Key: k.Public})
}
if err := s.accounts.EnsureKeys(ctx, user, wanted); err != nil {
// The forge holds one public key on one user only, so the key leaves agent-<node> before it joins nox-agent. That
// is done here, after everything that can fail before it, so the window in which the machine's key is on no user
// is only the span between these two calls; if the add fails, the next pass (at most ten minutes) puts it back.
if err := s.releasePerNodeUser(ctx, node); err != nil {
return err
}
if err := s.accounts.EnsureKeys(ctx, user, forgeaccount.MachineScope(node), wanted); err != nil {
return err
}
s.mu.Lock()
@@ -175,7 +186,37 @@ func (s *Service) Gone(ctx context.Context, e provider.Entry) error {
if user == "" {
return nil
}
return s.accounts.RemoveKeysTitled(ctx, user, forgeaccount.TitlePrefix+e.Consumer+".")
if err := s.releasePerNodeUser(ctx, e.Consumer); err != nil {
return err
}
_, err := s.accounts.RemoveKeysTitled(ctx, user, forgeaccount.MachineScope(e.Consumer))
return err
}
// releasePerNodeUser lets go the per-node user (agent-<node>) a machine had before one user held every machine's keys
// (ADR 0410 §3): the mesh keys registered on it are removed. The user is never
// deleted and stays in the agents team while agent-token mints it a token, and the keys of the operator's own stay. It is said once per machine.
func (s *Service) releasePerNodeUser(ctx context.Context, node string) error {
old := forgeaccount.NodeAccount(node)
if !forgeaccount.NodeName.MatchString(node) || old == forgeaccount.AgentUser {
return nil
}
removed, err := s.accounts.RemoveKeysTitled(ctx, old, forgeaccount.MachineScope(node))
if err != nil && !errors.Is(err, forgehttp.ErrNotFound) {
return err
}
if removed == 0 {
return nil // never existed, or holds no mesh key of the machine's: nothing was let go, nothing is said
}
s.mu.Lock()
said := s.letGo[node]
s.letGo[node] = true
s.mu.Unlock()
if !said {
s.cfg.Log("let go the forge user %s of before: its mesh keys are removed; the user "+
"is not deleted and stays in the agents team while its token is minted (the agents of every machine are now %s)", old, forgeaccount.AgentUser)
}
return nil
}
// KeyChanged takes in one machine key as the bus keeps it, and tells OnKeys its node.
@@ -127,6 +127,14 @@ func (f *fakeForge) ServeHTTP(w http.ResponseWriter, r *http.Request) {
case http.MethodGet:
reply(200, f.held[m[1]])
case http.MethodPost:
for user, ks := range f.held {
for _, k := range ks {
if user != m[1] && fmt.Sprint(k["key"]) != "" && strings.Fields(fmt.Sprint(k["key"]))[1] == strings.Fields(fmt.Sprint(in["key"]))[1] {
reply(422, map[string]any{"message": "Key content has been used as non-deploy key"})
return
}
}
}
f.next++
f.held[m[1]] = append(f.held[m[1]], map[string]any{"id": f.next, "title": in["title"], "key": in["key"]})
reply(201, map[string]any{"id": f.next})
@@ -157,7 +165,24 @@ func (f *fakeForge) ServeHTTP(w http.ResponseWriter, r *http.Request) {
case p == "/api/v1/teams/5" && r.Method == http.MethodPatch:
reply(200, map[string]any{})
case strings.HasPrefix(p, "/api/v1/teams/5/members/") && r.Method == http.MethodPut:
f.members = append(f.members, strings.TrimPrefix(p, "/api/v1/teams/5/members/"))
in := strings.TrimPrefix(p, "/api/v1/teams/5/members/")
for _, m := range f.members {
if m == in {
reply(204, nil)
return
}
}
f.members = append(f.members, in)
reply(204, nil)
case strings.HasPrefix(p, "/api/v1/teams/5/members/") && r.Method == http.MethodDelete:
gone := strings.TrimPrefix(p, "/api/v1/teams/5/members/")
var kept []string
for _, m := range f.members {
if m != gone {
kept = append(kept, m)
}
}
f.members = kept
reply(204, nil)
default:
http.Error(w, "unexpected "+r.Method+" "+p, 400)
@@ -239,7 +264,7 @@ func TestAConsumerWithAKeyGetsItsEntryAndItsKeyIsRegistered(t *testing.T) {
r.key("g14", "nox", pubA+" nox@g14 (mesh)")
r.pass()
e := r.bus.entry(t, "g14.ssh-client.forge-account")
want := map[string]any{"user": "agent-g14", "ssh_host": "novox.internal", "ssh_port": float64(222),
want := map[string]any{"user": "nox-agent", "ssh_host": "novox.internal", "ssh_port": float64(222),
"known_host": "[novox.internal]:222 ssh-ed25519 AAAAC3HOSTKEY", "ssh_alias": "mesh-forge",
"host_key": "ssh-ed25519 AAAAC3HOSTKEY", "ssh_base": "ssh://git@novox.internal:222/",
"https_base": "https://git.novox.be/"}
@@ -251,10 +276,10 @@ func TestAConsumerWithAKeyGetsItsEntryAndItsKeyIsRegistered(t *testing.T) {
if e.VaultName != "" || e.User != "" {
t.Errorf("a public key is no secret; the entry names a credential: %+v", e)
}
if got := r.forge.titles("agent-g14"); len(got) != 1 || got[0] != "mesh:g14.nox" {
if got := r.forge.titles("nox-agent"); len(got) != 1 || got[0] != "mesh:g14.nox" {
t.Errorf("registered keys = %v", got)
}
if !r.forge.users["agent-g14"] || len(r.forge.members) != 1 || r.forge.members[0] != "agent-g14" {
if !r.forge.users["nox-agent"] || len(r.forge.members) != 1 || r.forge.members[0] != "nox-agent" {
t.Errorf("account %v, team members %v", r.forge.users, r.forge.members)
}
if strings.Join(r.runner.args, " ") != "ssh-keyscan -t ed25519 -p 222 127.0.0.1" {
@@ -295,15 +320,15 @@ func TestAChangedKeyReplacesTheOldMeshKey(t *testing.T) {
r.key("g14", "nox", pubA+" nox@g14 (mesh)")
r.pass()
r.forge.mu.Lock()
r.forge.held["agent-g14"] = append(r.forge.held["agent-g14"], map[string]any{"id": int64(99), "title": "my own", "key": pubB})
r.forge.held["nox-agent"] = append(r.forge.held["nox-agent"], map[string]any{"id": int64(99), "title": "my own", "key": pubB})
r.forge.mu.Unlock()
r.key("g14", "nox", "ssh-ed25519 AAAAC3NzaC1lZDI1NTE5AAAAICCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCC nox@g14 (mesh)")
r.pass()
got := strings.Join(r.forge.titles("agent-g14"), ",")
got := strings.Join(r.forge.titles("nox-agent"), ",")
if got != "my own,mesh:g14.nox" {
t.Fatalf("keys = %s", got)
}
for _, k := range r.forge.held["agent-g14"] {
for _, k := range r.forge.held["nox-agent"] {
if k["title"] == "mesh:g14.nox" && !strings.Contains(k["key"].(string), "CCCC") {
t.Fatalf("the old key stayed: %v", k)
}
@@ -316,7 +341,7 @@ func TestAConsumerThatNoLongerBindsLosesItsMeshKeysAndKeepsItsAccount(t *testing
r.key("g14", "nox", pubA+" nox@g14 (mesh)")
r.pass()
r.forge.mu.Lock()
r.forge.held["agent-g14"] = append(r.forge.held["agent-g14"], map[string]any{"id": int64(99), "title": "my own", "key": pubB})
r.forge.held["nox-agent"] = append(r.forge.held["nox-agent"], map[string]any{"id": int64(99), "title": "my own", "key": pubB})
r.forge.mu.Unlock()
r.bus.mu.Lock()
delete(r.bus.buckets[provider.AssignmentsBucket], "g14.ssh-client")
@@ -325,10 +350,10 @@ func TestAConsumerThatNoLongerBindsLosesItsMeshKeysAndKeepsItsAccount(t *testing
if r.bus.has("g14.ssh-client.forge-account") {
t.Fatal("the entry stayed")
}
if got := r.forge.titles("agent-g14"); len(got) != 1 || got[0] != "my own" {
if got := r.forge.titles("nox-agent"); len(got) != 1 || got[0] != "my own" {
t.Fatalf("keys = %v", got)
}
if !r.forge.users["agent-g14"] {
if !r.forge.users["nox-agent"] {
t.Fatal("the account was removed")
}
}
@@ -428,11 +453,11 @@ func TestADeletedKeyEntryRemovesTheMeshKeyAndTheBoundEntryButNotTheOperatorsKey(
r.key("g14", "nox", pubA+" nox@g14 (mesh)")
r.pass()
r.forge.mu.Lock()
r.forge.held["agent-g14"] = append(r.forge.held["agent-g14"], map[string]any{"id": int64(99), "title": "my own", "key": pubB})
r.forge.held["nox-agent"] = append(r.forge.held["nox-agent"], map[string]any{"id": int64(99), "title": "my own", "key": pubB})
r.forge.mu.Unlock()
r.svc.KeyGone("g14.nox")
r.pass()
if got := r.forge.titles("agent-g14"); len(got) != 1 || got[0] != "my own" {
if got := r.forge.titles("nox-agent"); len(got) != 1 || got[0] != "my own" {
t.Fatalf("keys = %v", got)
}
if r.bus.has("g14.ssh-client.forge-account") {
@@ -446,7 +471,7 @@ func TestAPassBeforeTheKeysAreLoadedRemovesAndDeletesNothing(t *testing.T) {
r.key("g14", "nox", pubA+" nox@g14 (mesh)")
r.pass()
r.forge.mu.Lock()
r.forge.held["agent-g14"] = append(r.forge.held["agent-g14"], map[string]any{"id": int64(99), "title": "my own", "key": pubB})
r.forge.held["nox-agent"] = append(r.forge.held["nox-agent"], map[string]any{"id": int64(99), "title": "my own", "key": pubB})
r.forge.mu.Unlock()
// A restart: the keys are not read yet, so a node with none is not a node that reported none.
r.svc.mu.Lock()
@@ -454,7 +479,7 @@ func TestAPassBeforeTheKeysAreLoadedRemovesAndDeletesNothing(t *testing.T) {
r.svc.loaded = false
r.svc.mu.Unlock()
out := r.pass()
if got := strings.Join(r.forge.titles("agent-g14"), ","); got != "mesh:g14.nox,my own" {
if got := strings.Join(r.forge.titles("nox-agent"), ","); got != "mesh:g14.nox,my own" {
t.Fatalf("keys = %s", got)
}
if !r.bus.has("g14.ssh-client.forge-account") {
@@ -468,7 +493,7 @@ func TestAPassBeforeTheKeysAreLoadedRemovesAndDeletesNothing(t *testing.T) {
r.svc.loaded = true
r.svc.mu.Unlock()
r.pass()
if got := strings.Join(r.forge.titles("agent-g14"), ","); got != "my own" || r.bus.has("g14.ssh-client.forge-account") {
if got := strings.Join(r.forge.titles("nox-agent"), ","); got != "my own" || r.bus.has("g14.ssh-client.forge-account") {
t.Fatalf("keys = %s", got)
}
}
@@ -541,3 +566,133 @@ func TestKeepRunningBacksOffToTwoMinutes(t *testing.T) {
t.Fatalf("the backoff is %v", pauses)
}
}
// The per-node users made before (ADR 0410 §3) are let go: their mesh keys and their team seat go, the user and the
// operator's own keys stay, and it is said once.
func TestThePerNodeUserMadeBeforeIsLetGoButNeverDeleted(t *testing.T) {
r := newRig(t)
var said []string
r.svc.cfg.Log = func(f string, a ...any) { said = append(said, fmt.Sprintf(f, a...)) }
r.bus.bindClient("g14")
r.key("g14", "nox", pubA+" nox@g14 (mesh)")
r.forge.mu.Lock()
r.forge.users["agent-g14"] = true
r.forge.members = []string{"agent-g14", "agent-ace"}
r.forge.held["agent-g14"] = []map[string]any{
{"id": int64(90), "title": "mesh:g14.nox", "key": pubA}, {"id": int64(91), "title": "my own", "key": pubB}}
r.forge.held["agent-ace"] = []map[string]any{{"id": int64(92), "title": "mesh:ace.nox", "key": pubB}}
r.forge.mu.Unlock()
r.pass()
r.pass()
if got := r.forge.titles("agent-g14"); len(got) != 1 || got[0] != "my own" {
t.Fatalf("the old user's keys = %v", got)
}
if got := r.forge.titles("agent-ace"); len(got) != 1 {
t.Fatalf("another machine's old user was touched: %v", got)
}
if got := strings.Join(r.forge.members, ","); got != "agent-g14,agent-ace,nox-agent" {
t.Fatalf("team = %s", got)
}
if !r.forge.users["agent-g14"] {
t.Fatal("the old user was deleted")
}
n := 0
for _, l := range said {
if strings.Contains(l, "agent-g14") && strings.Contains(l, "let go") {
n++
}
}
if n != 1 {
t.Fatalf("said %d times: %v", n, said)
}
}
// A key on nox-agent that is not titled `mesh:` survives every pass: key kept, key changed, key gone (ADR 0410 §2).
func TestAKeyOnTheAgentUserTitledOtherwiseSurvivesEveryPass(t *testing.T) {
r := newRig(t)
r.bus.bindClient("g14")
r.key("g14", "nox", pubA+" nox@g14 (mesh)")
r.pass()
r.forge.mu.Lock()
r.forge.held["nox-agent"] = append(r.forge.held["nox-agent"], map[string]any{"id": int64(99), "title": "by hand", "key": pubB})
r.forge.mu.Unlock()
r.pass()
r.key("g14", "nox", pubB+" nox@g14 (mesh)") // the machine's key changed to the one held by hand
r.pass()
r.svc.KeyGone("g14.nox")
r.pass()
got := r.forge.titles("nox-agent")
if len(got) != 1 || got[0] != "by hand" {
t.Fatalf("keys = %v", got)
}
}
// Letting a per-node user go is said once per user across passes (ADR 0410 §3).
func TestLettingEachPerNodeUserGoIsSaidOnceAcrossPasses(t *testing.T) {
r := newRig(t)
var said []string
r.svc.cfg.Log = func(f string, a ...any) { said = append(said, fmt.Sprintf(f, a...)) }
for n, pub := range map[string]string{"g14": pubA, "ace": pubB} {
r.bus.bindClient(n)
r.key(n, "nox", pub+" nox@"+n+" (mesh)")
r.forge.held["agent-"+n] = []map[string]any{{"id": int64(80 + len(n)), "title": "mesh:" + n + ".nox", "key": pub}}
}
r.pass()
r.pass()
r.pass()
count := map[string]int{}
for _, l := range said {
for _, u := range []string{"agent-g14", "agent-ace"} {
if strings.Contains(l, u) && strings.Contains(l, "let go") {
count[u]++
}
}
}
if count["agent-g14"] != 1 || count["agent-ace"] != 1 {
t.Fatalf("said %v: %v", count, said)
}
}
// Gitea holds one public key on one user only (ADR 0410 §3): the key leaves agent-<node> before it joins nox-agent.
func TestAMachinesKeyMovesFromItsOldUserToTheAgentUser(t *testing.T) {
r := newRig(t)
r.bus.bindClient("g14")
r.key("g14", "nox", pubA+" nox@g14 (mesh)")
r.forge.held["agent-g14"] = []map[string]any{{"id": int64(90), "title": "mesh:g14.nox", "key": pubA}}
out := r.pass()
if got := r.forge.titles("agent-g14"); len(got) != 0 {
t.Fatalf("old user keys = %v (pass: %v)", got, out)
}
if got := r.forge.titles("nox-agent"); len(got) != 1 || got[0] != "mesh:g14.nox" {
t.Fatalf("nox-agent keys = %v (pass: %v)", got, out)
}
}
// The old user lets the key go only just before nox-agent takes it: a pass that fails earlier leaves the key where it
// is, so the machine is never on no user (ADR 0410 §3).
func TestAPassThatFailsBeforeTheKeyIsAddedLeavesTheOldUsersKey(t *testing.T) {
r := newRig(t)
r.runner.err = errors.New("sshd down")
r.bus.bindClient("g14")
r.key("g14", "nox", pubA+" nox@g14 (mesh)")
r.forge.held["agent-g14"] = []map[string]any{{"id": int64(90), "title": "mesh:g14.nox", "key": pubA}}
r.pass()
if got := r.forge.titles("agent-g14"); len(got) != 1 {
t.Fatalf("old user keys = %v", got)
}
}
// Nothing is said for a user that never existed or held no mesh key of the machine's.
func TestNothingIsSaidForAPerNodeUserWithoutMeshKeys(t *testing.T) {
r := newRig(t)
var said []string
r.svc.cfg.Log = func(f string, a ...any) { said = append(said, fmt.Sprintf(f, a...)) }
r.bus.bindClient("g14")
r.key("g14", "nox", pubA+" nox@g14 (mesh)")
r.pass()
for _, l := range said {
if strings.Contains(l, "let go") {
t.Fatalf("said %q", l)
}
}
}