Add Principal.ownerKey(): role-prefixed, keyed on name for a named lead and
a collaborator (survives a handover's terminal change), on terminal for a
worker, architect and observer, and on a distinct "anonymous" value for an
unauthenticated caller so that case no longer relies on Authz refusing it
first. The unnamed primary keeps a null key, preserving its primary-wide
ticket rule.
Thread that key through Task.creatorOwner, Rendezvous.Owner, poll,
pendingAsk and answer in place of a raw terminal, in both the MCP and REST
surfaces, so a named lead whose terminal changes can still poll and answer
its own delegations while a different lead is refused both.
Mutation evidence (each one-line change killed a named test, then reverted
to green):
- Principal.ownerKey() PRIMARY case made unconditional (dropped the
null-name guard) -> ownerKeyCoversEveryRole dies:
"expected: <null> but was: <leader:null>"
- OBSERVER case changed to use the "worker" prefix -> ownerKeyCoversEveryRole
dies: "expected: <observer:term_observer> but was: <worker:term_observer>"
- prefixed() changed to drop the role prefix entirely -> both
ownerKeyCoversEveryRole and rolePrefixesKeepLeadAndArchitectKeysDistinct
die on a lead/architect key collision: "expected: <leader:opus> but was:
<opus>"
- PRIMARY case changed to key on terminal instead of name ->
rolePrefixesKeepLeadAndArchitectKeysDistinct and ownerKeyCoversEveryRole
die: "expected: <leader:opus> but was: <leader:term_lead>"
- ARCHITECT case changed to key on the slot name instead of terminal ->
same two tests die: "expected: <architect:opus> but was: <architect:design>"
All five mutations were caught by the existing test suite; no test needed
adding.
SessionManager.releaseRemoved() tore down a member's registry row and pane
but never cleared it from MemberPresence, so a terminal stayed marked
"present" for the daemon's lifetime after release/idle-reap/shutdown drain.
Clear it in the method's unconditional finally block, alongside the other
must-always-run teardown step, so every release path (explicit release,
the idle reaper's releaseIfCurrent, and a shutdown drain) forgets it the
same way, and a throw from the dirty-worktree check does not skip it.
MemberPresence.forget(null) throws NullPointerException (verified empirically:
ConcurrentHashMap.remove(null) NPEs on key.hashCode()), so the new call guards
on a non-null, non-blank terminal id rather than relying on forget to no-op.
Unit 2 replaces the /clear continuation with a real process restart, so every
paragraph in the skill that describes /clear goes false the moment the new jar
is deployed. The code is not merged yet, and a merge is not a deployment, so
rewriting those paragraphs now would hand a lead doing a handover tonight a
document that does not match the daemon it is talking to.
Add a dated note instead. It states that the /clear text stays accurate while
the old jar runs, and gives a test a lead can apply with no shell: the
fleet_handover tool description is served by the running daemon, so if it still
says "clear your pane", the old behaviour is live. It also names the two things
that change, including the one that doubles as a second indicator -- the
"never observed as WORKING after 8 consecutive IDLE/DONE polls" warning cannot
appear once the wait that logs it is deleted. The note names the condition for
deleting itself.
Re-measure the roll evidence while here. The skill recorded four
"lead-rollover: rolled" lines from 2026-09-22; the log now holds 20, against a
control of 86 "lead-rollover:" lines, and "Unknown command" still returns 0.
Add the elapsed spread (median 16507 ms, max 48261 ms, two above 45000 ms) with
the caveat that it times the whole roll and is dominated by the wait for the
calling turn to end, so a slow roll is not a failed one.
Markdown only, no code touched, so no build was run.
The comment said ticket ids are a sequential counter with no owner check, so a
holder could walk every ticket and read another session's reply. PRs #712 and
#716 added that owner check: MessageService.ownsTicket compares a ticket's
creatorTerminal to the caller on every read.
The rule is still right, so only the reason changes. This matters now because
fleetd #737 is deciding ticket ownership across a lead handover, and a reader
who believed the old text could delete the TASK_READ restriction on the grounds
that its stated reason no longer applies.
Comment-only. mvn -o clean install: Tests run: 2083, Failures: 0, Errors: 0,
BUILD SUCCESS. Flagged by the #705 option-1 worker as out of its scope, which
was the right call.
Adds Role.OBSERVER as the bottom rung CallerResolver falls to when a
herdr pane matches no live roster entry, lead, architect slot, or
collaborator tab. An observer may only READ/METRICS and REPLY/ASK on
its own pane. Widens the presence gate so an observer's MCP contact
still marks it deliverable, matching what already happens for a
worker or architect, so a pane that outlives a daemon restart is not
left permanently undeliverable.
Ships as defence in depth alongside the already-merged ticket-owner
check (#712/#716), which closed the reachable exploit this ticket
reported.
A member whose MCP contact lands between launcher.spawn() and registry.put()
had its presence marked, but the SPAWNING -> READY transition that markPresent
triggers found no registry entry yet and silently did nothing. The mark then
persisted while registration left the session in SPAWNING, with nothing to
retry the transition. That left the session undeliverable to reclaim/seat
accounting even though it was present and deliverable.
Add SessionManager.reconcilePresence, called right after registry.put in both
the plain-spawn and worktree-spawn paths, to retry the transition for a
terminal already marked present. One private helper serves both call sites.
Tests cover both orderings (contact-then-register and register-then-contact)
for both spawn paths, plus a terminal never marked present staying in
SPAWNING. The contact-then-register tests use a new PresenceRacingLauncher
test double that marks presence from inside spawn(), before acquire()'s own
registry.put runs.
confirm() takes the per-lead-terminal claim before handing the roll to
continuationRunner, and the only release path was runRollover's own
finally. If continuationRunner.accept itself throws, runRollover never
starts, so that finally never runs, and nothing else ever writes
rollingByTerminal — the claim is held forever and the terminal can never
be rolled again. This differs from fleetd #615, which covers a throw
INSIDE the continuation (runRollover already catches that and still
releases the claim) — this is a throw from the hand-off itself, which
is not reachable with today's virtual-thread runner but would be with
a bounded executor's RejectedExecutionException.
confirm() now catches that throw, releases the claim, and overwrites
the IN_PROGRESS outcome with a terminal FAILED one, matching how a
throw inside the continuation is already surfaced.
relaunch's javadoc said it reads the live config; it actually reads the
FleetConfig snapshot this launcher was constructed with (fleet.leaders
is the frozen half), so say that and note a profile/tab edit needs a
daemon restart.
relaunch's recognise-only refusal reused ensureLeads()'s log wording,
which claims the lead 'is not live' — true in ensureLeads()'s context
(reached only after a short live count), false in relaunch's (which
never counts liveness, by design). Dropped that clause.
Pulled the declared/creatable/profile-configured resolution shared by
ensureLeads() and relaunch() into one private resolveLaunchable(name)
helper (returns a new ResolvedLead(lead, profile) record, or null
having logged), so the three refusals and their wording live in one
place instead of two copies that can drift. Behaviour-preserving:
ensureLeads() keeps its own liveness-count logic around the shared
resolve, and the existing 37 LeadLauncherTest cases are unchanged and
still pass.
Two open() calls for the same lead terminal minted two tokens that both
passed confirm()'s ownership check, so both could reach the deferred
continuation and roll the same lead twice. confirm() now claims a
per-lead-terminal slot (an atomic put-if-absent) once every other gate has
passed, refusing a concurrent confirm with the new ROLL_ALREADY_RUNNING
reason; runRollover releases the claim in a finally, on both the success
and the thrown-exception path.
askSeq restarts at 0 on every daemon boot, so a turnId (session#n)
minted by one Rendezvous instance could be minted again by a later
instance and resolve to an unrelated ask. Fold a per-instance nonce
into the mint, the same way #719 fixed MessageService's ticket ids.
Adds LeadLauncher.relaunch(name), which starts exactly the named lead
from the live config, outside of ensureLeads()'s instances bookkeeping.
It retries the whole launch attempt (not just the agent_name_taken/
agent_pane_busy cases ResilientAgentLaunch already retries inside one
agents.start call) up to RELAUNCH_ATTEMPTS times.
launch() now returns the started Agent (null on failure) instead of a
boolean, so relaunch() and ensureLeads() share the same primitive.
Extract checkPaneCommandFits, the agent_pane_busy retry, and the agent_name_taken
retry out of HerdrPeerLauncher into a shared dev.ltms.fleet.herdr.ResilientAgentLaunch,
and route LeadLauncher.launch through the same seam instead of a bare agents.start
call. The lead's agent name now carries a per-process nonce and a per-start sequence
number (like a member's), so a stale agent_name_taken from an earlier crashed session
no longer blocks a legitimate relaunch outright.
other.sendAsync had never been called, so other's tasks map was empty and
poll(ticket) returned null regardless of whether the nonce existed — the
test passed against an empty map, not against a colliding id. Mint once on
other so it reaches the same sequence number as the first instance, making
the test exercise the actual collision the nonce guards against.
ticketSeq restarted at zero on every daemon boot with no persistence, so a
ticket id minted in one boot could be reused by a later boot and resolve to
an unrelated Task instead of failing to resolve at all. Mint each
MessageService instance's own short nonce once and fold it into every ticket
(task-<nonce>-<n>), so an id from one instance can never match another's id
space.
Adds a disjoint-id-space test and a foreign-instance-ticket test (with the
positive control) in MessageServiceTest.
The previous commit compared 4 configDir lines against 8 total profiles, but
four of those are opencode and never read CLAUDE_CONFIG_DIR. Every claude-code
profile does set one, so the original claim was right and this file said
otherwise.
The real defect is narrower: opus and sonnet name the operator's own config dir,
so for those members the store is shared, and ClaudeCodeLauncher's javadoc
already records that fleetd and the operator's session write that same file.
GET /profiles reports 8 live profiles and fleetd.yaml carries 4 configDir
lines, two of them pointing at the operator's own instance dir. The bullet
stated the blanket claim as a structural limit, so it would have been believed.
The conclusion it supported is unchanged: member-facing assets travel in the
worktree.
The javadoc said a losing releaseIfCurrent CAS is the call that arrives with no
known session. releaseIfCurrent is only called by the reaper, always with a
non-null expected, so it always has one; the null-known call is an overlapping
release that finds the registry entry already gone. The same wrong claim was in
a test's failure message.
Releasing, enter and leave drop private so SessionManagerTest binds them at
compile time. The six reflection helpers are gone, and a rename now breaks the
build instead of a test run.
fleetd #715 gates fleet_send{turnId} on the caller that created the delegation,
so the shipped block had to say so: the step-5 note already covered who can see
a pending question, not who can answer it.
Add a CallerResolverTest case proving a releasing architect resolves as
WORKER while still inside releaseRemoved's teardown window, with a
control resolve outside the window that must stay ARCHITECT so the
in-window assertion cannot pass against a slot that was never bound.
Split spawnedMemberRoleSurvivesAnOverlappingReleaseThatUnmarksEarly into
two SessionManagerTest cases, each reaching Releasing.enter/leave
directly through reflection so a mutation to one invariant (the depth
count in leave, or enter's prior-terminal preservation) can only fail
its own test.
Add a source-scrape test mirroring #718's MessageServicePollUsageTest
shape, for the same residual: a convenience overload that defaults the
turn's owner to null, left in place because deleting it would break 81
test-only call sites across 7 unrelated files.
The scanner is exercised against a file known to hold many real
one-argument rendezvous.open( calls before it is ever pointed at
production, using the identical matching logic for both. A pattern that
cannot find the known calls would also find none in production, and
that is exactly the failure mode a -based git grep regex hit earlier
on this ticket: git grep's -E engine does not treat \b as a word
boundary, so that pattern silently matched nothing anywhere, in clean
code and in the 81 real calls alike.
Record the turn's owner on the forward rendezvous waiter (Rendezvous.Owner,
a three-state record: no record / unnamed primary / named terminal). A
fresh fleet_ask copies that owner onto the ask turn; a coalesced duplicate
ask keeps the first owner. answer() compares the answering caller against
the stored owner before taking the session lock or reopening the resumed
waiter, and a mismatch returns the new NOT_TURN_OWNER outcome instead of
STALE_TURN, with no rendezvous/task/question cleanup.
send() and answer() both drop their no-caller overloads; every call site
in MessageService, FleetMcp and FleetApp now threads an explicit caller
terminal through. AuthzTest pins that the unnamed-primary null allowance
is safe only because ANONYMOUS never reaches ANSWER.
Covers all four combinations of capture input x adapter: blocking-send and
async-send delegations, answered through both FleetMcp and FleetApp, each
with a hijack attempt refused and the real owner's answer succeeding as a
control.
A caller from a pane being torn down used to be briefly absent from the
session registry, so CallerResolver fell through to a lead/architect tab map
for the same terminal and could resolve the wrong role.
SessionManager now keeps a depth-counted "releasing" marker per pane, written
before the registry removal and cleared in a finally once release finishes.
A depth count (not a Set) is needed because two threads can race teardown of
the same pane; a Set-based unmark by the losing thread would reopen the
window while the winning thread is still mid-teardown. Both removal sites
(the unconditional release() and the idle reaper's CAS releaseIfCurrent())
go through one shared helper, so all four entry routes (release, the
context-cap release in completeTurn, reapIdle, drainSnapshot) are covered.
SessionManager.spawnedMemberRole is the one reader: it checks the live
registry first (via the existing no-copy findByTerminal), then the releasing
marker. FleetdAssembly now wires this method reference instead of its own
untested inline lambda, which also drops a roster() list copy + stream from
the per-request hot path. CallerResolver is unchanged — its contract already
fit.
It said "null if the primary", which reads as every primary. Only the unnamed
primary has no terminal; a named lead carries one, so a reader deciding what a
null means was given the wrong rule.
fleet_status (MCP) and GET /sessions/{id}/status (REST) handed any TASK_READ
holder another session's open fleet_ask question, its turnId and its ticket,
with no check that the caller created that delegation. MessageService.pendingAsk
now takes the caller's terminal and reuses the existing ownsTicket comparison;
FleetMcp.status and FleetApp.sessionStatus both thread the resolved caller
terminal through. The base status line and REST's ready field are unaffected.
MessageServicePollUsageTest's messages.poll( receiver anchor only
covers a MessageService reached through a variable, field, or
parameter named "messages". Adds a second assertion in the same
class that every such declaration under src/main/java uses that
name, with its own file-walk and declaration-count controls, so a
future declaration under a different name turns this check red
instead of leaving the original scan silently blind to it.
Also tightens the Authz.permits(Principal, Action, String) javadoc
sentence to read as a plain contract statement.
Adds a source-scrape test over src/main/java that fails if any caller
reaches the fail-open single-argument poll(String) overload instead of
poll(String, String). The scan anchors on the "messages.poll(" receiver
to avoid matching java.util.Queue.poll(), and balances parentheses to
avoid being fooled by a two-argument call whose first argument contains
nested parens.
Also documents Authz.permits(Principal, Action, String) as a test
convenience whose default classifier denies every collaborator.
Three parts. GET /tasks/{ticket} passed no caller, so it used the overload that skips the ownership
check and any worker could read any ticket; it now passes the caller resolved from the same CALLER
attribute the authorization gate reads. The wait:false send path recorded no creator terminal, so a
REST-created ticket matched no terminal-bearing caller and its own creator was refused; it now
records one. Both handlers gained a scrape guard with its own control assertion.
Resolved one conflict in FleetMcpAuthzTest by keeping both sides: PR #717 and this branch each
appended tests at the same point. The test count is the check on that resolution — 2014 + 3 + 1 =
2018, so no test was dropped.
Verified in a throwaway worktree off main: Tests run: 2018, Failures: 0, BUILD SUCCESS. Three
mutations, each confirmed live with mvn -o compile before the suite ran. FleetMcp:527 is the one
that survived before this work and now kills theFleetPollHandlerActuallyThreadsCallerTerminalIntoPoll.
FleetApp:698 kills the new creator test. FleetApp:899 kills three, including the behavioural test.
Every file restored byte-identical.
A worker now gets neither array, and the key is absent rather than present-and-empty. leads stays
visible to the primary, an architect and a collaborator; members only to the primary and an
architect. That split follows the rule the collaborators array already states: you may list what
you could address. A collaborator may send to a lead, and leads is the only place the bridge gives
it that address, so hiding it would have left a shipped grant unusable.
Verified in a throwaway worktree off main: Tests run: 2014, Failures: 0, BUILD SUCCESS, against a
main baseline of 2008 that I measured myself. Dropping the collaborator clause from leadsVisibleTo
compiled green and then killed exactly two tests, the truth-table row and the behavioural test.
sendMessage's wait:false branch now records the resolved caller's own
terminal as the ticket's creatorTerminal, the same way taskStatus already
resolves its caller, so a REST-created ticket's own creator can still poll
it under the ownership check that now gates GET /tasks/{ticket}.
A collaborator may SEND to a lead, and fleet_list's leads array is the
only place this tool gives it a lead's sessionId -- its own
fleet_whoami carries no lead address. leadsVisibleTo now returns true
for caller.isCollaborator() as well as primary and architect, matching
the existing rule for the sibling collaborators array (every role that
may SEND to a named peer). membersVisibleTo is unchanged: a
collaborator may never SEND to a spawned member.
Updated the truth-table tests for both predicates, added a behavioural
test proving a collaborator's fleet_list output contains leads and not
members, and updated fleet_list's tool description.
The block is the instruction surface this repo ships, so a merged change that makes it false is an
incomplete change. Two had gone stale:
- the lead table said fleet_list does not report collaborators. #709 made it report them, visible
to the primary, an architect and another collaborator, never a worker.
- the collaborator section justified the ticket refusal by saying ticket ids are a plain counter
with no owner check. #712 added that owner check, so the stated reason no longer held. The role
gate is what refuses a collaborator; the recorded creator terminal is the second line.
wiki/7-Use-Cases.md carries the same edit and was pushed to its own remote, verified by ref. The
sync check prints True.
Omit the leads and members keys entirely (never an empty array) from
fleet_list's result for a worker, matching the existing
coordinatorVisibleTo/collaboratorsVisibleTo pattern: two new named
predicates (leadsVisibleTo, membersVisibleTo) are consulted before
assembling either array, so a worker holding only READ can no longer
read every session on the daemon through this tool. Primary and
architect callers are unaffected.
GET /tasks/{ticket} now resolves the caller the same way allow(...) does and
threads that terminal into MessageService.poll(ticket, callerTerminal) instead
of the no-check overload, so a worker can no longer read a ticket a different
session created over REST. Adds a source-scrape guard (with its own control
assertion) for both the fleet_poll MCP handler and this REST route, plus a
behavioural test driving GET /tasks/{ticket} with three differently-resolved
callers against one shared MessageService.
The roster is consulted ahead of every tab map, so a live registered member
is no longer read back as a lead. Both refusals still matter, for the narrower
case where the pane is alive and the roster holds no entry for it.
Behaviour unchanged. Verified in a throwaway worktree, not piped:
Tests run: 2008, FleetConfigTest 170, BUILD SUCCESS.
It returned architectTerminals, so its name contradicted its contents and
collided with fleet_list's members array. No production caller.
Verified in a throwaway worktree, not piped: Tests run: 2008, BUILD SUCCESS.
Closes the MCP door only. The REST door (GET /tasks/{ticket}) still calls the
no-check poll overload, so #705 stays open.
Verified in a throwaway worktree, not piped: Tests run: 2008, BUILD SUCCESS.
Both refusals said a member landing in a lead's or collaborator's
labelled tab would be read back as that identity. CallerResolver
consults the spawned-member roster ahead of every tab map, so a live
registered member is never misread this way. State the real condition
instead: the hazard applies only while the pane is alive and carries
no entry in the spawned-member roster.
Dai Ha