fleetd #669 Unit C: add the COLLABORATOR role and its authorization row
CI / shell-tests (pull_request) Failing after 7s
CI / contract (pull_request) Successful in 53s
CI / build (pull_request) Failing after 1m48s

Adds Role.COLLABORATOR, Principal.collaborator(), and the Authz.permits grant:
READ/METRICS open, REPLY/ASK via ownership, SEND limited to a configured lead
or collaborator via a new classifier parameter, everything else denied.
isSpawnedMember() stays WORKER || ARCHITECT. fleet_whoami reports role and
collaborator name with no leader key.

Per a same-day ticket correction, the existing 3-argument Authz.permits is
kept (fail-closed default via the new NO_KNOWN_LEAD_OR_COLLABORATOR
classifier) rather than deleted, so the 47 pre-existing call sites in
AuthzTest/CallerResolverTest are untouched; both production gates
(FleetMcp#denyFor, FleetApp#allow via the new permitsFor seam) call the
4-argument form explicitly with the shared constant.

Mutation-verified: removing the classifier conjunct from the SEND arm kills
exactly 3 tests (AuthzTest, FleetMcpAuthzTest, FleetAppAuthTest), one per
gate, no cascade.

Full mvn clean install at this commit: 1974 tests, 0 failures (baseline at
f0ff252 was 1964; +10 are the new collaborator-matrix tests). Independently
counted from target/surefire-reports/*.xml after rm -rf: 173 report files,
aggregate tests=1974 failures=0 errors=0.
This commit is contained in:
Dai Ha
2026-10-04 00:22:27 +02:00
parent f0ff25221e
commit bb29b001e4
9 changed files with 256 additions and 29 deletions
@@ -1,5 +1,7 @@
package dev.ltms.fleet.auth;
import java.util.function.Predicate;
/**
* The authorization table (CB-505), stated once and enforced on both entry paths.
*
@@ -60,58 +62,92 @@ public final class Authz {
}
/**
* Whether {@code caller} may perform {@code action} against {@code targetSession}.
*
* @param targetSession the session id in the request path; only consulted for the worker-scoped
* actions ({@code REPLY}, {@code ASK}), ignored otherwise, may be
* {@code null}
* Stands in for the terminal-to-tab registry a collaborator's {@code SEND} is checked
* against, until one exists: answers no for every target, so a collaborator reaches nothing
* today. Both production gates ({@code FleetMcp#denyFor}, {@code FleetApp#allow}) pass this
* exact instance, so the classifier is defined once and replacing it is a one-line change in
* each.
*/
public static final Predicate<String> NO_KNOWN_LEAD_OR_COLLABORATOR = target -> false;
/**
* Convenience form for a caller with no classifier to supply. Fails closed: a collaborator's
* {@code SEND} is refused, as if no terminal were a configured lead or collaborator — the
* same decision {@link #NO_KNOWN_LEAD_OR_COLLABORATOR} gives explicitly. Every other action's
* result is identical to the four-argument form's, since none of them consult the classifier.
*/
public static boolean permits(Principal caller, Action action, String targetSession) {
return permits(caller, action, targetSession, NO_KNOWN_LEAD_OR_COLLABORATOR);
}
/**
* Whether {@code caller} may perform {@code action} against {@code targetSession}.
*
* @param targetSession the session id in the request path; only consulted for the
* worker-scoped actions ({@code REPLY}, {@code ASK}) and for a
* collaborator's {@code SEND}, ignored otherwise, may be
* {@code null}
* @param knownLeadOrCollaborator whether a terminal is a configured lead or collaborator —
* consulted only for a collaborator's {@code SEND}, to confine
* it to another named peer and never a spawned member's
* terminal
*/
public static boolean permits(Principal caller, Action action, String targetSession,
Predicate<String> knownLeadOrCollaborator) {
if (caller == null || caller.isAnonymous()) {
return false; // authenticated as nothing ⇒ authorized for nothing
}
return switch (action) {
// Fleet lifecycle is the primary's alone — spawn, stop, drain. An architect
// deliberately does NOT get these (CB-548), so it cannot tear down or stand up workers
// even though it coordinates them; and a worker driving any of these would be a worker
// escalating into the orchestrator role.
// Fleet lifecycle is the primary's alone — spawn, stop, drain. An architect and a
// collaborator deliberately do NOT get these, so neither can tear down or stand up
// workers even though one of them coordinates them; and a worker driving any of these
// would be a worker escalating into the orchestrator role.
case SPAWN, STOP, DRAIN, HANDOVER -> caller.isPrimary();
// Delivering a turn to a local session is open to the primary and the architect: an
// architect delegates to workers (that is the role's point) but still has no lifecycle
// rights. A worker is excluded — sending would be it escalating.
case SEND -> caller.isPrimary() || caller.isArchitect();
// Delivering a turn to a local session is open to the primary, the architect, and a
// collaborator whose target is itself a configured lead or collaborator: the architect
// delegates to workers (that is the role's point); a collaborator may reach only
// another named peer, never a spawned member's terminal. A worker is excluded —
// sending would be it escalating.
case SEND -> caller.isPrimary() || caller.isArchitect()
|| (caller.isCollaborator() && knownLeadOrCollaborator.test(targetSession));
// Same grant as SEND. Resolving a worker's blocked question is part of delegating to
// it, not a separate capability.
// Resolving a worker's blocked question is part of delegating to it, open to the same
// two roles that may stand up that delegation in the first place. Not a collaborator:
// resuming another session's turn is lifecycle-adjacent, not peer messaging.
case ANSWER -> caller.isPrimary() || caller.isArchitect();
// Same grant as SEND. This leaves the daemon over the coordination broker rather than
// addressing a local session, but the caller who may do one may do the other.
// Leaves the daemon over the coordination broker rather than addressing a local
// session, open to the same two roles as ANSWER. Not a collaborator: it is a
// local-tab peer with no cross-host route.
case COORD_SEND -> caller.isPrimary() || caller.isArchitect();
// The load-bearing rule: a caller acts only as the pane it occupies. CB-532 widened who
// that can be — a lead answering another lead is replying for its OWN terminal, which
// this already permits — while the rule itself is unchanged, and is what stops anyone
// forging a reply for a rendezvous someone else is waiting on. An architect's own pane
// passes through the same check, so it can answer a funnel that delegated to it. An
// unnamed primary (token/loopback, no pane) owns nothing and is still excluded.
// forging a reply for a rendezvous someone else is waiting on. An architect's or a
// collaborator's own pane passes through the same check, so each can answer a funnel
// that delegated to it. An unnamed primary (token/loopback, no pane) owns nothing and
// is still excluded.
case REPLY, ASK -> caller.ownsSession(targetSession);
// READ is roster, profile, and identity observation — fleet_list, fleet_profiles, and
// fleet_whoami — and carries no secrets: no ticket reply, no pending question, and no
// other session's turn state. Those live under TASK_READ. METRICS is the separate
// Prometheus scrape. Both stay open to every authenticated role.
case READ, METRICS -> caller.isPrimary() || caller.isWorker() || caller.isArchitect();
// Prometheus scrape. Both are open to every authenticated role, including a
// collaborator.
case READ, METRICS -> caller.isPrimary() || caller.isWorker() || caller.isArchitect()
|| caller.isCollaborator();
// Ticket polling and session status, open to every authenticated role the same as READ.
// Unlike READ, a holder may poll a ticket it did not create, or read another session's
// pending question and the turnId that answers it.
// Ticket polling and session status, open to every role READ is open to except a
// collaborator: ticket ids are a sequential counter with no owner check, so a holder
// could walk every ticket and read another session's delegation reply.
case TASK_READ -> caller.isPrimary() || caller.isWorker() || caller.isArchitect();
// fleetd #421: reading held lead-to-lead mail is the primary's alone. An architect
// holds READ today (CB-548), so "not primary" must mean not-architect here too — this
// is coordination between leads, not observation of the roster.
// is coordination between leads, not observation of the roster. The same reasoning
// excludes a collaborator.
case COORD_READ -> caller.isPrimary();
};
}
@@ -11,7 +11,9 @@ package dev.ltms.fleet.auth;
* @param pid the connecting process id, or {@code -1} when not resolvable (audit context)
* @param name for a lead resolved from the CB-530 {@code leaders:} registry, which lead it is;
* for an architect resolved from the CB-548 {@code architects:} registry, which
* slot it occupies; {@code null} for every other caller, including an unnamed primary
* slot it occupies; for a collaborator resolved from the {@code collaborators:}
* registry, which collaborator it is; {@code null} for every other caller,
* including an unnamed primary
*/
public record Principal(Role role, String terminal, long pid, String name) {
@@ -73,6 +75,19 @@ public record Principal(Role role, String terminal, long pid, String name) {
return new Principal(Role.ARCHITECT, terminal, pid, slotName);
}
/**
* A collaborator: a human-opened tab recognised by its exact label in the
* {@code collaborators:} registry.
*
* <p>Carries {@link Role#COLLABORATOR}. {@code name} is reporting only — it lets
* {@code fleet_whoami} say which collaborator is asking. Identity is the {@code terminal}:
* like a worker's it comes from the connection, so {@code ownsSession} works exactly as it
* does for a worker — a collaborator acts as its own pane and no other.
*/
public static Principal collaborator(String name, String terminal, long pid) {
return new Principal(Role.COLLABORATOR, terminal, pid, name);
}
public boolean isPrimary() {
return role == Role.PRIMARY;
}
@@ -81,6 +96,10 @@ public record Principal(Role role, String terminal, long pid, String name) {
return role == Role.ARCHITECT;
}
public boolean isCollaborator() {
return role == Role.COLLABORATOR;
}
public boolean isWorker() {
return role == Role.WORKER;
}
@@ -120,6 +139,7 @@ public record Principal(Role role, String terminal, long pid, String name) {
return switch (role) {
case WORKER -> "worker:" + terminal;
case ARCHITECT -> "architect:" + name;
case COLLABORATOR -> "collaborator:" + name;
case PRIMARY -> name == null ? "primary" : "leader:" + name;
case ANONYMOUS -> "anonymous";
};
@@ -34,6 +34,17 @@ public enum Role {
*/
ARCHITECT,
/**
* A config-declared, human-opened tab recognised by its exact label (the {@code
* fleet.collaborators.<name>.tab} registry). Never spawned — identity comes from the
* connection, never a request argument, exactly like {@link #WORKER} and {@link #ARCHITECT}.
* May {@code SEND} only to a configured lead or collaborator, {@code REPLY}/{@code ASK} only
* as its own pane, and {@code READ}/{@code METRICS}; may not {@code SPAWN}/{@code STOP}/
* {@code DRAIN}/{@code HANDOVER}, poll a ticket ({@code TASK_READ}), or reach the
* coordination broker ({@code COORD_SEND}/{@code COORD_READ}).
*/
COLLABORATOR,
/** Authenticated as nothing. Authorized for nothing but {@code /healthz}. */
ANONYMOUS
}
@@ -689,7 +689,7 @@ public final class FleetMcp {
if (!authorizationEnforced) {
return null; // AuthorizationMode.UNENFORCED: authorization not enforced (fleetd #518)
}
if (Authz.permits(caller, action, target)) {
if (Authz.permits(caller, action, target, Authz.NO_KNOWN_LEAD_OR_COLLABORATOR)) {
if (action != Authz.Action.READ && action != Authz.Action.TASK_READ) {
AuditLog.allowed(caller, action, target); // reads would drown the trail
}
@@ -1309,6 +1309,18 @@ public final class FleetMcp {
}
return text(json(m));
}
if (caller.isCollaborator()) {
// A collaborator's name is its slot in the collaborators: registry; sessionId is its
// pane so a peer knows where to reach it. No leader key: a collaborator is not a
// primary for authorization, unlike a lead.
if (caller.name() != null) {
m.put("collaborator", caller.name());
}
if (caller.terminal() != null) {
m.put("sessionId", caller.terminal());
}
return text(json(m));
}
if (!caller.isWorker()) {
// CB-530: which lead, once more than one pane is configured as one. `role` deliberately
// still reads "primary" — the fallback ladder in CLAUDE.md keys on it, and a lead IS a
@@ -266,6 +266,17 @@ public final class FleetApp {
return app;
}
/**
* The authorization decision behind {@link #allow}, taking the caller directly rather than
* pulling it from a servlet {@link Context} — unit-testable without fabricating a live
* request, the same reason {@code FleetMcp#denyFor} is split from {@code FleetMcp#deny}. The
* classifier is the same shared instance {@link #allow} would use, so a test calling this
* exercises the real production gate, not a test-supplied stand-in.
*/
static boolean permitsFor(Principal caller, Authz.Action action, String target) {
return Authz.permits(caller, action, target, Authz.NO_KNOWN_LEAD_OR_COLLABORATOR);
}
/**
* Gate a handler on the CB-505 authorization table. Returns {@code true} when the request may
* proceed; otherwise writes the error response and returns {@code false}.
@@ -279,7 +290,7 @@ public final class FleetApp {
return true; // legacy: authorization not enforced
}
Principal caller = ctx.attribute(CALLER);
if (Authz.permits(caller, action, target)) {
if (permitsFor(caller, action, target)) {
if (action != Authz.Action.READ && action != Authz.Action.METRICS
&& action != Authz.Action.TASK_READ) {
AuditLog.allowed(caller, action, target); // reads would drown the trail
@@ -14,6 +14,7 @@ class AuthzTest {
private static final Principal ANON = Principal.anonymous();
private static final Principal ARCH_DESIGN = Principal.architect("lead-designer", "term_design", 400);
private static final Principal ARCH_OTHER = Principal.architect("reviewer", "term_review", 500);
private static final Principal COLLABORATOR = Principal.collaborator("ops", "term_collab", 600);
@Test
void anonymousIsAuthorizedForNothing() {
@@ -166,4 +167,87 @@ class AuthzTest {
assertFalse(Authz.isUnauthenticated(WORKER_A));
assertFalse(Authz.isUnauthenticated(PRIMARY));
}
// ── the collaborator matrix ─────────────────────────────────────────────────────────────────
/**
* {@code SEND} for a collaborator is the one grant that is conditional rather than fixed:
* flipping only the classifier's answer for the target flips only this outcome.
*/
@Test
void aCollaboratorMaySendOnlyWhenTheClassifierAcceptsTheTarget() {
assertTrue(Authz.permits(COLLABORATOR, SEND, "term_lead", target -> true),
"the classifier accepting the target must grant SEND");
assertFalse(Authz.permits(COLLABORATOR, SEND, "term_lead", target -> false),
"the classifier refusing the target must deny SEND");
assertFalse(Authz.permits(COLLABORATOR, SEND, "term_lead"),
"the real production classifier recognises no terminal yet, so SEND is refused today");
}
/**
* Control for the test above: every other action's result for a collaborator does not move
* when the classifier does. Only {@code SEND} is wired to it.
*/
@Test
void theClassifierMovesOnlySendForACollaborator() {
for (Authz.Action a : Authz.Action.values()) {
if (a == SEND) {
continue;
}
assertEquals(
Authz.permits(COLLABORATOR, a, "term_lead"),
Authz.permits(COLLABORATOR, a, "term_lead", target -> true),
a + " must not depend on the classifier at all");
}
}
@Test
void aCollaboratorMayReadAndScrapeMetrics() {
assertTrue(Authz.permits(COLLABORATOR, READ, null));
assertTrue(Authz.permits(COLLABORATOR, METRICS, null));
}
@Test
void aCollaboratorMayReplyAndAskOnlyAsItsOwnPane() {
assertTrue(Authz.permits(COLLABORATOR, REPLY, "term_collab"),
"its own pane is its own");
assertTrue(Authz.permits(COLLABORATOR, ASK, "term_collab"));
assertFalse(Authz.permits(COLLABORATOR, REPLY, "term_design"),
"a collaborator must not reply on another pane");
assertFalse(Authz.permits(COLLABORATOR, REPLY, null),
"an absent target must not pass the own-session rule");
}
/**
* Every action denied to a collaborator, asserted denied even when the classifier would
* accept any target — proving none of these is actually gated on the classifier at all.
*/
@Test
void aCollaboratorIsDeniedLifecycleCoordinationAndTicketPolling() {
for (Authz.Action a : new Authz.Action[]{SPAWN, STOP, DRAIN, HANDOVER, ANSWER, COORD_SEND,
COORD_READ, TASK_READ}) {
assertFalse(Authz.permits(COLLABORATOR, a, "term_lead", target -> true),
"a collaborator must not " + a + " even when the classifier accepts every target");
}
}
@Test
void aCollaboratorIsNotCountedAsPrimaryWorkerOrArchitect() {
assertFalse(COLLABORATOR.isPrimary());
assertFalse(COLLABORATOR.isWorker());
assertFalse(COLLABORATOR.isArchitect());
assertTrue(COLLABORATOR.isCollaborator());
}
/**
* A collaborator is never spawned, so it must not be enrolled in the presence map as an
* available member. Control: both a worker and an architect — which ARE spawned — still are.
*/
@Test
void isSpawnedMemberIsFalseForACollaboratorButTrueForAWorkerAndAnArchitect() {
assertFalse(COLLABORATOR.isSpawnedMember());
assertTrue(WORKER_A.isSpawnedMember());
assertTrue(ARCH_DESIGN.isSpawnedMember());
}
}
@@ -97,6 +97,7 @@ class FleetMcpAuthzTest {
private static final Principal WORKER_A = Principal.worker("term_a", 200);
private static final Principal ANON = Principal.anonymous();
private static final Principal ARCH_DESIGN = Principal.architect("lead-designer", "term_design", 400);
private static final Principal COLLABORATOR = Principal.collaborator("ops", "term_collab", 600);
// --- the table, enforced on THIS path too ---------------------------------------------------
@@ -226,6 +227,19 @@ class FleetMcpAuthzTest {
"the caller IS authenticated — it is just not the right role");
}
/**
* {@code denyFor} passes the real production classifier, not a test-supplied one — no
* terminal is recognised as a configured lead or collaborator, so a collaborator's SEND is
* refused over MCP.
*/
@Test
void aCollaboratorMayNotSendOverMcpWithTheRealProductionClassifier() {
FleetMcp m = mcp(true);
McpSchema.CallToolResult denied = m.denyFor(COLLABORATOR, Authz.Action.SEND, "term_lead");
assertNotNull(denied, "no terminal is recognised as a lead or collaborator yet");
assertTrue(denied.isError());
}
@Test
void theLegacyConstructorLeavesTheGateOpen() {
// The 22 pre-existing FleetMcpTest cases rely on no authorization being enforced.
@@ -1837,6 +1837,32 @@ class FleetMcpTest {
assertTrue(out.contains("\"sessionId\":\"term_design\""), out);
}
/**
* A collaborator reports its own role and name, never the {@code leader} key a lead gets —
* {@code role} already reads {@code "collaborator"}, so a {@code leader} key alongside it
* would be self-contradicting. Control: the same call shape fed a named lead must still carry
* {@code leader}, so this is not passing because the key stopped being emitted for everyone.
*/
@Test
void whoamiReportsACollaboratorWithNoLeaderKeyButALeadStillGetsOne() {
FakeHerdr h = new FakeHerdr();
SessionManager sessions = sessionManager(h, "http://gx00.gw:8000", Set.of("gx00.gw"));
McpSchema.CallToolResult collabRes = FleetMcp.whoami(
Principal.collaborator("ops", "term_collab", 700), sessions);
assertNotEquals(Boolean.TRUE, collabRes.isError());
String collabOut = textOf(collabRes);
assertTrue(collabOut.contains("\"role\":\"collaborator\""), collabOut);
assertTrue(collabOut.contains("\"collaborator\":\"ops\""), collabOut);
assertTrue(collabOut.contains("\"sessionId\":\"term_collab\""), collabOut);
assertFalse(collabOut.contains("leader"), collabOut);
McpSchema.CallToolResult leadRes = FleetMcp.whoami(
Principal.leader("opus", "term_lead", 100), sessions);
String leadOut = textOf(leadRes);
assertTrue(leadOut.contains("\"leader\":\"opus\""), leadOut);
}
/**
* CB-548: an architect SEND delegates as its own pane (recording the per-target delegation) but
* must NEVER become the legacy singleton "primary" fallback — the per-target map does not cure
@@ -6,6 +6,7 @@ import com.fasterxml.jackson.databind.ObjectMapper;
import dev.ltms.fleet.auth.CallerResolver;
import dev.ltms.fleet.auth.Authz;
import dev.ltms.fleet.auth.MemberRegistry;
import dev.ltms.fleet.auth.Principal;
import dev.ltms.fleet.config.FleetConfig;
import dev.ltms.fleet.guard.SubscriptionGuard;
import dev.ltms.fleet.herdr.AgentControl;
@@ -202,6 +203,18 @@ class FleetAppAuthTest {
}
}
/**
* {@code permitsFor} is the exact decision {@link FleetApp#allow} makes, passing the real
* production classifier rather than a test-supplied one — no terminal is recognised as a
* configured lead or collaborator, so a collaborator's SEND is refused through the REST gate.
*/
@Test
void aCollaboratorMayNotSendOverRestWithTheRealProductionClassifier() {
Principal collaborator = Principal.collaborator("ops", "term_collab", 700);
assertFalse(FleetApp.permitsFor(collaborator, Authz.Action.SEND, "term_lead"),
"no terminal is recognised as a lead or collaborator yet");
}
// --- loopback-trust: the caller is the primary -------------------------------------------
@Test