Each primary's allocator (seqjob/parjob my_alloc) sets ctx->alloc to a
fresh job allocation covering only its own hosts, so a multi-primary
COMPOUND ("1*{A} + 2*{B}") dispatched only the last primary's hosts --
leaking the earlier allocations while every primary's hosts had been
debited in ctx.uncomitted.
sched_make_compound_alloc_ now folds each primary's allocation into one
job allocation via the new libxce merge_job_alloc(): a host placed by
several primaries becomes one entry with per-resource values added
(matched on resname, instid and reserve, so distinct instances are never
summed); new hosts and shared allocations are appended/merged the same
way. Partial-failure rollback remains bug-1766.
Tests: "1*{..} + 2*{..}" on 2x2 slots now allocates 3 slots across two
distinct hosts; "1*{..} + 1*{..}" yields one host entry with 2 slots
(both kept only the last primary before). Scheduler suite 87/87, lib 1/1.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
seqjob_prepare_allocation() and parjob_prepare_allocation() returned 0
whenever ctx->schedobj was not PRIMARY, so a primary nested in a COMPOUND
(or ALTER) job never had its slot count or rusage checked: a 5-task
compound dispatched onto 4 slots, and an ALTER job was "placed" by its
first alternative with an empty (0-slot) allocation. Both hooks now
compute the want from ctx->primary -- the primary being allocated, set by
sched_ctx_set_primary() for nested primaries and identical to
obj->u.primary->resreq for PRIMARY jobs.
sched_ctx_set_primary() also clears ctx->worker, so the next primary of a
compound computes its job-level want rather than a per-host want left
over from the previous primary's host (PRIMARY jobs set it once, with
worker already NULL).
Tests (real ingest + live allocator): a 5-task COMPOUND on 2x2 slots now
PENDs; "{5*{..}} || {1*{..}}" falls back to the second alternative (1
slot). Both failed before. Scheduler suite 87/87.
Follow-up logged: bug-1863 (multi-primary COMPOUND keeps only the last
primary's allocation).
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
bug-1769: every builder stores xce_compound_schedobj_t* (24 bytes) in
obj->u.alters, but sched_make_alter_allocation() read them as
xce_sched_obj_t* (176 bytes) and ran the compound allocation on them -- a
heap-buffer-overflow read of ->u.compound and write of ->alloc for any real
`{..} || {..}` resreq. The compound allocation now takes its resreq
explicitly (sched_make_compound_alloc_), and each ALTER alternative is
tried on behalf of the ALTER obj itself; a failed alternative's partial
allocation is dropped before the next (rollback itself is bug-1766).
sched_ctx_init() rejected ALTER objs ("bad sched obj type"), so ALTER jobs
could never allocate at all -- it now accepts them like COMPOUND (all
downstream checks only distinguish PRIMARY).
bug-1771: plugin_commit_decision() fans every DISPATCH out to the legacy
"xce_sched_fcfs" plugin even when it never tracked the job; fcfs_del_job()
(and fcfs's job-add hook) dereferenced job->queue->name with no NULL
check, SEGV'ing the allocator for a job whose queue was never registered.
Tests (test_parjob_e2e_multislot.c, real ingest + live allocator thread):
test_alter_resreq_job_dispatches_via_real_allocator (ASAN
heap-buffer-overflow before, DISPATCH after) and
test_dispatch_of_job_with_unregistered_queue_does_not_crash (SEGV at
fcfs.c:318 before). Scheduler suite 87/87.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
All 16 subtasks were already done; check the 6 parent tasks and set
phase implementation-complete. Verified 2026-09-24/25: scheduler suite
87/87 (ASAN), live multi-host e2e PASS, HA soak sequential + chaotic PASS.
Out-of-scope follow-ups stay open: bug-1766, bug-1768, bug-1769,
bug-1771, bug-1860.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The main loop adds and reaps worker-list entries (add_worker,
cleanup_worker_list -> remove_worker) while controller request handlers on
vtask threads walk the same list with no lock. A rejoin resync overlapping
a job reap read a freed node (ASAN heap-use-after-free in
find_worker_by_jobid), killing the worker mid chaotic HA soak.
- xcecd_worker_list_lock/unlock: add_worker links a fully built entry under
it; remove_worker unlinks under it and frees after.
- handle_jobstatequery holds it from reading the head through the query
page (the report send is a non-blocking xce_request); job control and
remote task start copy what they need under it instead of keeping the
entry pointer.
- The finished-job retention store (appended by the main loop, read by
resync) gets its own mutex around its public entry points.
Test: test_build_report_safe_against_concurrent_list_mutation -- UAF 3/3
with no-op lock stubs, clean 5/5 with the fix; worker suite 10/10; chaotic
HA soak 2/2 PASS, 6/6 resync jobs reconciled, no ASAN reports.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The finish notify is fire-and-forget, so the retained outcome is the only
way a controller learns a finish lost in a failover. Its 60s window ran
from the finish time, so an 84s outage expired it before the rejoin
resync could report it; the controller got "unrecognized" and requeued a
completed job (then stuck: scheduler kept RUN, see bug-1860).
Expiry is now measured from max(finish_time, last reconnect)
(xcecd_on_reconnect), and nothing expires while checking_connection's
hello probe reports the connection unhealthy. The reported finish_time_ms
is unchanged.
Verified: worker suite 10/10 (two new retention tests, red first); chaotic
HA soak 3/3 PASS (was 1 in 3 failing).
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A failing TEST_ASSERT longjmps out and skips the fixture teardown, leaving
the commit engine's sweep thread and persist's fsync_worker asleep on a
static struct. The next test re-inits it underneath them, corrupting the
condvar, and a later teardown blocks in pthread_cond_destroy/join until
the 300s meson timeout. The failure line itself was lost because stdout
was fully buffered.
- controller_test: line-buffer stdout; tearDown() reaps what a failed test
leaked via new XCE_TESTING registries (persist_test_reap_unstopped, then
commit_test_reap_unfinalized_engines -- persist_stop before engine fini).
Clean run reaps 0; an injected failure now fails in 2s instead of hanging.
- persist: flusher_flush_once reports only its own durable advance, so
fsync_worker no longer re-fires the durable callback for an advance a
concurrent persist_flush made (flaky "must not re-fire" test). Regression
test test_flusher_pass_does_not_claim_concurrent_flush_advance.
- tests: jap observer is atomic and waits (bounded) for a continuation that
may resolve on the flusher thread; pump catch-up tests wait for trailing
replies instead of a single scan.
- config_repl: upgrade decode temp file via mkstemp instead of a
pthread_self()-derived /tmp name (same value across processes, opened
without O_EXCL -- collisions and a symlink target).
Verified: controller_unit_test 1222/1222; 24 parallel runs (6x4), no
hangs or ASAN errors.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
q->resreq holds an interned makeString() value (pointer past the xstring
object header). core_upsert_queue() and sched_update_existing_queue()
released the old value with FREEUP()/free() when a queue update carried a
new resreq, an ASAN bad-free. Seen as HA soak replay-parity errors=1: the
sim replay of a snapshot-restored queue followed by a config push aborted
in core_upsert_queue.
Both sites now dereference() the old value (NULL-guarded). Regression
tests cover each site; the existing queue-update test had skipped the
resreq branch.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Since bug-1790's fix (9caf1424) counts a job's full task total against
global limits, the repo's limit1 (all users except rshen share 3 slots)
pends xce_test's root-submitted 33-task job forever, so the multihost
e2e fixture failed even though the allocator is correct.
deploy-ha.sh gains an optional XCE_DEPLOY_CONFIG_DIR (default unchanged:
src/configs). run-multihost-verify.sh copies src/configs into its deploy
dir, replaces limits.yaml with an empty list, and deploys with that
override.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
- buglog: gitea_issue / gitea_synced_status for the 20 bugs created and the
30 already-synced entries (#25/#35 reopened as partial)
- .kiro/specs/*/spec.json: "gitea" milestone/issue mapping for all 66 specs
(65 milestones + 195 issues created, 158 closed at creation)
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
gitea-sync.py fixes found on the first live run (bug-1855 follow-up):
- 204/empty-body replies (label DELETE) no longer crash in json.loads
- every request has a 60s timeout, so a stalled call fails instead of hanging
- spec.json sync state is inserted as a trailing "gitea" block instead of
re-serializing the file (15 hand-formatted spec.json files would reflow)
- implementation issue body labels generated vs approved correctly; the
validation body uses the same done-phase set as the close rules
Project memory now routes all buglog writes through the buglog skill:
CLAUDE.md key rule, a cerebrum Do-Not-Repeat entry (with the older
duplicate-id / json.dump / tags-vs-fix lessons tagged as superseded), and
xce-audit-invariants step 5.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The buglog had no status field, so "is this bug open?" meant reading free
text, and gitea-sync closed any issue whose `fix` text was non-empty --
including "NOT FIXED"/"PENDING"/"PARTIALLY FIXED" ones (bug-1855).
- Backfill `status` (+ status_updated) on all 1,193 entries. The 85 entries
whose text said not-fixed/pending/out-of-scope were checked against HEAD
and git history; the rest are `fixed`. Result: 36 open, 10 partial,
2 investigating, 5 cannot-reproduce, 4 wontfix, 1 duplicate.
Insert-only diff: existing lines are untouched.
- .claude/rules/buglog-status.md defines the status values (kept out of
openwolf.md, which `openwolf update` overwrites).
- scripts/buglog.py + buglog skill: add (next free id), status, note,
recur, list, show, triage. Edits only the touched entry's lines; a full
json.dumps rewrite churns ~1,200 lines of mixed \u-escaping.
- gitea-sync (script + skill): state and a single bug:<status> label come
from `status`; reopen supported; issues shared by several entries follow
the most-open one; only status labels are swapped; per-entry write-back
with gitea_synced_status so unchanged entries cost no API calls; list-typed
`fix` no longer crashes; spec issues close on done phases (spec.json has
no implementation approval); --since no longer floods Gitea after the
backfill.
- Disable OpenWolf buglog auto-detection: its hook numbers new entries
len(bugs)+1, which collides with existing ids (bug-1856, upstream).
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
- .wolf/anatomy*, memory.md: hook-maintained index and session log
through the bug-1851..1854 work.
- .kiro/specs/libxce-alloc-preason-equality/: requirements, design,
research and tasks generated; none approved yet (spec.json phase
tasks-generated).
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
xcecli_retry_next's "xcecli failover: attempt N -> host:port" line now
carries "(after status=<rc>)". Without it a log cannot tell a dead
connection from a healthy primary's commit-timeout or admission-backpressure
reply -- the distinction bug-1850 (the live scheduler re-joining every
~100s) needs to find its triggering rc. Diagnostic only; no behavior change.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A chaotic round's walcheck compared all three members' on-disk WALs with
boundary=auto. The just-killed term-4 primary still held its NACKed,
never-committed lsn=58, which differed from the new term-5 leader's
own-term noop at lsn=58, and was counted as class=committed.
walcheck_count now takes an optional "midloop" mode, used only by the
per-round chaotic check: the boundary is the serving primary's own
committedLsn (live_committed_boundary() needs 3-of-3 convergence, which a
chaotic round never reaches), and only members guaranteed to hold the
committed prefix are compared -- the primary plus live standbys whose
appliedLsn has reached the boundary. Every other caller, including the
settled final sweep, still compares all three. Log lines name members=.
Also logs bug-1854 (scheduler re-adds every configured resource on each
rejoin config push; ERROR noise, changed definitions ignored) for
follow-up.
10-round chaotic soak: PASS (divergence 0, job loss 0, SIMVERIFY 9 files,
0 mismatches).
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
When the live worker + scheduler pair is up, the soak now starts the
scheduler with --record-dir, records (xceadmin scheduler record start)
from each scheduler (re)join until just before the next primary kill,
stops it cleanly so the .replay is flushed, and at the end replays every
clean recording through `xce_scheduler --sim --sim-verify`. Any decision
MISMATCH, a failed sim run, or zero clean recordings FAILs the soak:
[xce-test-ha] SIMVERIFY files=N mismatches=M errors=E unclean=U
Recordings that could not be stopped cleanly are counted as unclean and
never verified (their unflushed tail would give false mismatches), except
a stop whose reply was lost mid-failover: when the same scheduler pid
answers "not recording" and logged no recording abort, the first stop
landed and the file is kept. Mismatches are counted from both the sim's
log file and stderr (the logger creates its file lazily).
XCE_HA_SCHED_REPLAY_VERIFY=0 disables the gate.
First runs caught bug-1851 and bug-1852.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The recorder encoded each controller event batch after
xce_core_push_events() had applied it, and applying moves payload fields
out of the events: a WORKERS_PUSH's resvals (core_apply_worker_resvals
MOVE_PTR), a JOBS_PUSH's allocs, a JOB_RUSAGE's rusages, and a
CONFIG_PUSH's shared resvals (sched_apply_pending_config recorded after
sched_init_config). Recordings silently lost exactly those fields -- a
worker's resval names were written with NULL values, so --sim-verify
replayed "no slots" and turned 3 live DISPATCHes into PENDs.
driver.c now records the batch before the push; the periodic in-stream
SnapshotEntry moves out of sched_record_write_events() into new
sched_record_checkpoint_after_apply(), called after the push so a
checkpoint still reflects the events written ahead of it. The config
push is likewise recorded before sched_init_config().
Regression test: test_recorded_workers_push_keeps_resval_values (RED:
recorded value NULL; GREEN after). Periodic-checkpoint tests updated to
the write-then-checkpoint contract. meson test 109/109; 25 live soak
recordings replay with 0 mismatches.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
update_val_res() memcpy'd the xce_val_t into w->host.resvals, so the
"type"/"model" string resources aliased the interned xstrings owned by
w->host.attrs. rejoin_worker() drops the old attrs strings, then
apply_join_state() -> resync_arm_round() resolves the previous join's
still-outstanding round, firing the worker hooks before
add_hostattrs_resvals() re-seeds resvals; the apiserver hook's
collect_resval() strdup()ed the freed hostmodel. Live: a newly promoted
primary died ~8s after election during an HA soak.
A string resval now stores a private strdup (taken before the old copy is
freed), and the owning arrays free their entries: host.resvals gets
kill_resval (replacing the FIXME NULL killer) and c->sharedinst's
resval_obj_killer (previously a no-op) frees the string and the value --
both also leaked every xce_val_t before. Also closes the same alias for
worker-reported string resvals and config values.
Regression test: test_rejoin_worker_string_resvals_survive_outstanding_round
(RED reproduced the live ASAN stack; GREEN after). meson test 109/109.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Follow-ups from the bug-1846 investigation.
bug-1847: config pushes never reached a .replay recording -- handler.c
applied them synchronously, bypassing the pending-update queue the
recorder writes from. With bug-1846's deferral they're now applied at a
round boundary on the driver thread, so sched_apply_pending_config()
writes a CONFIG_PUSH event via the existing sched_record_write_events()
(no-op when not recording), ahead of the round's other recorded events.
Test: test_config_push_deferred.c's test_applied_config_push_is_recorded
(RED: 0 events; GREEN: 1). Fidelity limit, pre-existing: replay applies
configPush through core_events.c's CONFIG_PUSH path, which doesn't re-run
groupmgr_load_configs()/plugin_config_update(), so a limits/groups change
still won't take effect in sim.
bug-1848: xce_sched_obj_t.hlimits was created with a NULL zarray killer
at all three constructors, so reset_sched_obj() and schedobj teardown
leaked the per-worker xlist_t match_job_host_limit_() callocs. New
kill_hlimits_list() (schedobj.c) frees the list, never its vres entries
(owned by the limit plugin); core_events.c and snapshot_load.c use
identical static local copies, since both avoid sharing helpers with
schedobj.c and several test binaries link them without it. Test:
test_limit_host_dedup.c's test_hlimits_zarray_releases_its_lists (RED:
killer NULL; GREEN).
bug-1849: every --sim run warned "failed to resolve job picker plugin ...
picker rebuild will be skipped". Benign: the snapshot is loaded before
the sim registers its picker, and sim_bootstrap_picker() replays the same
queue/job walk afterward. Downgraded to TRACE with an explanation.
Buglog also records bug-1850 (open): the ~100s reconnect churn behind
bug-1846 part 2 -- xcecli_retry_next() rotates/reconnects on every retry
(ETIMEDOUT after the 10s reply deadline, E_XCE_NOT_READY, redirects), so
a slow reply from a healthy primary forces a rejoin + full re-push.
Needs the live scheduler.log and a libxce spec.
Full scheduler suite 87/87 (asan). Replay of the captured file: same 184
mismatches, all attributable to the recording's pre-fix binary.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
sched_publish_retry_loop() calls sched_check_pending_rejoin() after every
successful publish -- inside the driver's decision-drain loop -- so after a
client-library reconnect the scheduler re-joins mid-round. The controller
answers every join (on_scheduler_join(), controller/sched.c) by re-pushing
config, workers and jobs. handler.c applied the config push synchronously
on the handler thread: sched_init_config() -> sched_init_queues() ->
builtin_fcfs on_queue_update(), which nulls the picker cursor when it
reinserts the queue the cursor rests on. next_job() then returned NULL and
the round ended with the rest of the queue never considered.
Evidence (scheduler-replay-20260923-085356.replay, temporary event-stream
diagnostics, reverted): 9 recorded rounds end after 11-12 decisions; every
one is followed by a workersPush + full 50-job jobsPush (the join's resync,
reporting just-dispatched jobs back as PEND), every full-length dispatching
round has none, and the config push never appears because it bypassed the
pending-update queue the recorder writes from -- which is also why replay
could not reproduce the truncation.
Fix (defer to the round edge, same direction as bug-1813): handler.c now
calls new sched_stash_config(), parking the Conf4Sched in new
xce_scheduler_t.pending_config under pending_config_lock (an older
unapplied push is freed). main.c's loop calls new
sched_apply_pending_config() before each round, running the unchanged
sched_init_config() strictly between sessions. sched_fini() frees a
never-applied push.
Tests:
- test_handlers.c: test_config_push_is_stashed_not_applied_on_handler_thread
-- RED (sched_init_config called on the handler thread), GREEN after.
- new test_config_push_deferred.c (sched_config_push_deferred_test): real
scheduler_req_handler() + real builtin_fcfs mid-session -- RED (next_job()
returned NULL right after the push), GREEN (session walks the remaining
jobs; configver unchanged until applied once between rounds).
- test_admin_request.c / test_job_array_expand.c / test_job_array_wire_e2e.c
#include handler.c directly and get a link stub for sched_stash_config().
Full scheduler suite 87/87 (asan).
Not addressed: config pushes are still not written to a recording (a
replay spanning a config change won't see it); the ~100s reconnect churn
itself (see bug-1812/1813).
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
match_job_host_limit_() (plugins/limit.c) runs on every per-host
allocation attempt via limit_prepare_allocation(), and unconditionally
xlist_store()d the matched vres into obj->hlimits[worker->idx]. That list
is only rebuilt by limit_job_election() on a configver bump, so a job that
kept pending on the same host gained one more copy per attempt.
sched_fill_limit_requests() then staged every copy into
ctx->alloc_req.host, and seqjob's make_host_alloc() built a host alloc
carrying all of them.
Found via temporary per-decision --sim-verify diagnostics (reverted, not
committed) against scheduler-replay-20260923-085356.replay: pc13v host
allocs held mem + slots + ~70 identical limit entries (72) instead of 3.
After the fix the replay side is always 3 entries, and the LIVE recording
itself still shows the inflation (72/81/96/114/138/177 entries, resetting
only on a config push) -- so this is a production bug, not sim-only.
make_job_account() creates one res_account per entry and
sched_update_res_accounts() charges each, so the host limit was charged
(and later released) N-fold.
Fix: hlimits_contains_() pointer-identity check (vres is interned per name
in privdata->instances); a matched vres is stored once per (job, worker).
Test: test_limit_host_dedup.c (sched_limit_host_dedup_test) -- five
plugin_prepare_allocation() attempts for one job on one worker must leave
hlimits[worker]->count == 1 and one staged copy of the vres in
alloc_req.host. RED before the fix (Expected 1 Was 2), GREEN after; full
scheduler suite 86/86 (asan).
Still open (bug-1846 part 2, logged in buglog): 9 live-recorded rounds end
after 11-12 decisions while replay processes the whole candidate list; not
root-caused, needs the live scheduler.log or a fresh capture.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
/xce-audit-invariants scheduler dimension-1/4 audit, prompted by a fresh
user-captured .replay file that still showed massive --sim-verify mismatch
volume after bug-1843's fix. Root-caused via temporary diagnostic
instrumentation against the real capture (reverted, never committed) rather
than trusting the surface symptom.
bug-1844: sim_verify_alloc_eq() (sim_verify.c) compared `instid`, a bare
per-process monotonic counter (scheduler.c's sched_get_id(), core->nextid)
stamped at allocation-build time -- not semantic content. The original
recording's counter and a fresh replay's counter have no reason to reach
the same value for "the same" dispatch. Live-verified: every DISPATCH
mismatch showed identical value/duration/resname/reserve, differing only
in instid (e.g. 826 vs 1). Dropped from the comparison, same treatment as
bug-1842's `from` exclusion.
bug-1845: sim_verify_job_preason_eq() required PEND decisions' full
host/resource reason-detail lists to match. core.h documents "simulation
isolation" as deliberate design: sim/replay never registers
core->unsatisfiable_lookup_cb, so it can never hit the unsatisfiable-cache
main.c's normal-mode branch registers. A live cache hit
(sched_pend_from_cached_reason_()) records only main_reason with zero
detail; replay always falls through to the full per-host computation and
always produces detail. Same semantic outcome (PEND, same main_reason) at
two architecturally-legitimate detail levels. Relaxed the comparator to
main_reason only. Separately verified the cache itself is sound (key is
safe within a round since capacity only decreases and the cache clears
every round) -- ruling it out as the cause of a second, still-open
divergence (bug-1846: one job genuinely dispatches one round early in
replay vs. the original recording -- a suspected resource-accounting
parity gap, not yet root-caused, logged for follow-up).
bug-1772 (previously deferred, incorrectly tagged "fixed"): re-found
independently by a dimension-4 snapshot-field-completeness audit subagent.
snapshot_build_primary_schedobj() (snapshot_load.c) never computed
resreq_fp, unlike its declared model core_build_primary_schedobj()
(core_events.c) -- every snapshot-restored job got a zeroed fingerprint,
which alloc_cache.c's cache key folds in directly alongside
user/queue/project, so two distinct restored jobs sharing those three
fields could silently share an incorrect unsatisfiable-cache verdict.
Added the same 4-field computation the live path uses.
TDD throughout: each fix has a dedicated regression test, and each was
verified via genuine red (revert fix, confirm the new test fails) then
green (restore, full 85/85 scheduler suite). Also ran three parallel
dimension audits (core I/O isolation, picker/allocator jbuf rendezvous) --
both clean, no violations found.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
sched_record_session_add_decision() (recording_session.c) grows its
per-round decisions array via realloc(), doubling from 8. On a failed
growth it silently dropped the current decision and every one after it
that round, without marking the round incomplete -- finalize() then
wrote the short array as an ordinary, apparently-complete ReplaySession.
--sim-verify's replay comparison had no way to distinguish that from a
genuinely smaller live round, producing spurious "expected: <none>"
mismatches for jobs that were, in reality, correctly decided live and
simply never recorded. Live scheduling itself was unaffected either
way -- only the .replay file's own record of that one round lost its
tail.
This refutes the leading theory from the prior investigation (replay
seeing a job's first appearance "one round early" via coarser event
batching in sim_loop.c) -- confirmed via targeted diagnostic
instrumentation against the original captured .replay file: replay
deterministically reproduces the full candidate set every round;
the recording itself has the one-round gap.
Fix: track truncation on the session ctx; finalize() now refuses to
write an incomplete session and fails loudly through the same
sched_record_abort_on_error() path every other recording failure in
this subsystem already uses.
TDD: added a regression test forcing a real, deterministic realloc()
failure via -Wl,--wrap=realloc (scoped to this one test binary --
no prior malloc/realloc fault-injection harness existed in this
codebase). Verified genuine red -> green: the test fails without the
fix, passes with it. Full 85/85 scheduler suite green.
Note: this prevents future recordings from losing decisions this way;
it cannot retroactively repair an already-captured .replay file.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
- .wolf/anatomy-index.json, anatomy.md, memory.md: OpenWolf hook-maintained
project index and session log, refreshed by this session's activity.
- .kiro/specs/controller-peers-env-fallback/: spec scaffold (brief,
requirements, design, research) from prior spec-init work.
- issues/new_requirements.md: scheduler simulator enhancement backlog
(info-command improvements, pend-reason detail, decision tracing).
Excluded as non-persistent: issues/Issues.md (empty placeholder),
src/lib/testsuite/test_xcectx_reconnect_race.c (self-documented as an
ASAN stress reproduction only, not a committed regression test, and not
wired into meson.build), scheduler-replay-20260922-095207.replay
(4MB binary debug capture, no .replay files are tracked in this repo).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
sim_verify_alloc_eq() compared xce_alloc_t.from, which make_alloc()
(xcealloc.c) stamps with the real OS wall clock (time(NULL)), not
core.c's virtualized sim clock. The recorded (expected) value is
whatever real second the original run happened to execute in; the
freshly-rebuilt (actual) value is whatever real second the replay
executes in -- two unrelated real-world timestamps, so every
otherwise-correct DISPATCH decision failed verification whenever a
replay ran even one wall-clock second away from its own recording.
Drop `from` from the comparison -- it's reservation bookkeeping, not
part of a decision's semantic identity. Verified against a real
captured .replay file: 120/120 false DISPATCH mismatches -> 0.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Existing worker-rejoin-resync coverage (sequential/chaotic) only proves a
worker-dispatched job survives a real failover -- it never re-checks the
worker's own resource accounting afterward, which is exactly where
bug-1841 lived. This mode over-submits jobs to empirically discover the
worker's real concurrency ceiling (self-verifying via a PEND overflow
probe, not assumed from resval), fails over the primary while jobs are
RUN, waits for genuine rejoin + DONE, then hard-asserts the probe job
(still PEND from before the kill) now reaches RUN -- proving the
scheduler's resInUse actually unfroze.
Live-validated against a real 3-node cluster + xcecd_virtual worker:
settled at 3 RUN / 41 PEND (limits.yaml's limit1, not raw worker
capacity, was the actual ceiling -- exactly the kind of environment
detail this design avoids needing to assume), all 3 recovered to DONE via
genuine rejoin, probe reached RUN 219ms after recovery. Two earlier
design iterations are documented in the round's own comments: sequential
submit-then-wait let early jobs finish before late ones were even
submitted, and assuming resval's raw capacity as the concurrency ceiling
doesn't hold when queue/user limits bind first.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Worker-rejoin-resync sets XCE_WORKER_RECONCILING and pushes it to the
scheduler, which correctly freezes that worker's resInUse mid-reconciliation
(bug-1799/1801). But resync_resolve_round() only cleared the bit on the
controller's own worker_t, never notifying the scheduler -- so once a
rejoin round resolved, the scheduler's copy stayed RECONCILING forever,
permanently freezing that worker's resource accounting even after every
job on it finished. Live-observed via a replay capture: a 22-slot worker
stuck at used=22.00 for the rest of the run with zero RUN jobs left.
Added persist_notify_worker_status_updated() (mirrors bug-1679's job-side
persist_notify_job_updated_already_persisted()) and call it from
resync_resolve_round() so HOOK_DETAIL_WORKER_UPDATE fires without a
spurious WAL append. TDD: new spy-hook regression test, RED confirmed
against the unfixed code, GREEN after the fix. Full controller_unit_test
suite: 1220 Tests, 0 Failures, 0 Ignored.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Live 3-node HA soak crashed xce_controller: SIGSEGV in
peer_ctrl_msg_handler_inline (ha.c:7035) dereferencing a NULL
ep->data. gdb analysis of the actual core dump showed the endpoint
itself fully alive (refcnt=191964, tearing_down=0) while ep->data had
been synchronously freed and nulled by standby_ep_err_handler_primary,
the endpoint's err_handler, with zero coordination with the vtask pool
that was concurrently dispatching a peer-control message for the same
endpoint (g_drive_async_peerctrl_dispatch_enabled defaults ON).
The sibling handlers (vote_ep_err_handler, pre_vote_ep_err_handler)
already avoid this exact hazard, with an explicit comment: "Do NOT
free vctx here -- peer_ctrl_msg_handler may still reference ep->data."
standby_ep_err_handler_primary was never updated to the same
discipline when the vtask-pool dispatch machinery (bug-1197/bug-1298)
was introduced.
Fix: route standby_ep_err_handler_primary through the same
already-tested vote_ep_teardown_async() deferred, per-endpoint-FIFO-
serialized teardown that vote-endpoint demotion already uses, plus an
idempotency guard (endpoint_is_tearing_down) since the old "ep->data
== NULL" reentry check no longer applies once the free is deferred.
Also fixes a second-order double-unref this surfaced in
on_standby_connected's stale-duplicate eviction path, caught by ASAN.
TDD: two new tests prove the teardown is deferred until an in-flight
same-ep vtask completes, and that a repeated error callback on the
same endpoint is a safe no-op. Full controller suite: 1219/1219, zero
regressions.
Also logs bug-1840 (a separate scheduler heap-corruption crash from
the same soak run) as not-fixed pending an ASAN reproduction -- a
non-instrumented core dump cannot localize glibc's malloc_consolidate
abort back to the actual out-of-bounds write.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
sim_reconcile_unaccounted_jobs() (sim_loop.c) was calling
picker->picker_iface->on_dispatch() directly from the sim driver thread,
violating plugin.h's documented "allocator thread only" contract for
Post-allocation hooks. Dormant in practice only because builtin_fcfs.c
happens to lock internally around every hook -- a defensive measure the
contract text doesn't actually require of picker plugins.
Since scheduler-sim-live-allocation, sim mode spawns the same real
sched_job_picker()/sched_allocator() thread pair normal mode uses, so a
real allocator thread now exists to route through. Added a
reconcile_req/reconcile_ack cbuf pair on xce_sched_core_t and a reserved
XCE_SCHED_RECONCILE_WAKE sentinel pushed through jbuf to wake the
allocator thread without touching jbuf's pinned capacity-1/item-type
invariant (Req 5 AC 7). sched_allocator() services the request and calls
on_dispatch() itself; sim_reconcile_invoke_on_dispatch() blocks on the ack.
Falls back to a direct call when no allocator thread is live for a given
core (standalone unit tests), which is safe by construction.
TDD: new test spawns a stand-in allocator-relay thread and asserts
on_dispatch() fires there, not on the caller's thread -- confirmed RED
against the old direct-call code, GREEN after. Four other test files with
hand-rolled sched_allocator() stubs needed the same relay branch (found by
an actual hang at meson's timeout). Full scheduler suite: 85/85, zero
regressions.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
User-reported: replaying a captured session dispatched 44 single-slot jobs
onto a 22-slot host. Root cause traced through two layers:
- bug-1836: sim_reconcile_unaccounted_jobs() (the "Recorded-Decision
Reconciliation" backstop) force-applied any recorded DISPATCH decision
it found, with zero live capacity check. Added
sim_reconciled_alloc_fits_capacity(), reusing the same
sched_check_host_satisfy()/sched_check_shared_satisfy() admission checks
the live picker/allocator path already uses.
- bug-1838: even with that gate, reconciliation still preempted every
freshly-PEND job before the live picker/allocator got a turn each round
-- unnecessary now that scheduler-sim-live-allocation gave sim mode the
same real, capacity-checked picker/allocator threads normal mode uses.
Narrowed reconciliation's trigger back to JOB_IS_RUN-only (undoing a
2026-09-11 widening whose own justification no longer holds), so a
freshly-pushed job now always flows through the same live, verified
decision path as normal mode. Reconciliation now only ever fires for a
job already RUN from before the replay's own observation window --
the one case structurally invisible to the live picker.
Also logged (bug-1837, not fixed here, needs its own design decision):
sim_reconcile_unaccounted_jobs() calls a picker's on_dispatch() hook from
the driver thread instead of the allocator thread plugin.h documents it
belonging to -- currently inert since the only real picker plugin happens
to lock internally, but not contract-safe for a future lock-free one.
Fixing bug-1838 required re-examining every test exercising the old
PEND-inclusive trigger; several needed their fixtures corrected to
register a real worker (harmless before any capacity check existed), and
others needed their actual intent re-derived and their assertions
inverted or restated now that a PEND job is never eligible for
reconciliation. Dated correction notes added in place to the
scheduler-sim-eager-reconciliation spec, matching its own existing
2026-09-11 correction-note precedent.
Verified: full scheduler suite 85/85, independently re-confirmed by
running the four most affected binaries directly.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
CtrlHelloRequest.key was declared `Data` in controller.capnp. This
project's capnp-c backend silently no-ops encoding of Data-typed fields
without zero-initializing the caller's local struct, leaving an
uninitialized capn_ptr that write_CtrlHelloRequest unconditionally wrote
to the wire -- SIGSEGV in capn_setp/copy_ptr. Hit live: ctrl-1 won its
Raft election and crashed calling ha_send_ctrlhello_to_all_standbys().
Same bug class as bug-1504/bug-1505, both already fixed by switching the
field from Data to List(UInt8); bug-1505's own writeup had flagged
CtrlHelloRequest.key as a known-but-unfixed instance. Applied the
identical fix here: Data -> List(UInt8) + keyLen companion in the schema,
uint8_t*/key_len in typemap.h. The field is unused in production (always
NULL), so no consumer code changed. Added a round-trip regression test
covering both the NULL-key production shape and real key content.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
update_limit_res_value() computed a percent-based limit's capacity from
the resource's own live, fluctuating value (e.g. "mem"), letting the
enforced cap drift with unrelated memory pressure instead of staying
pinned to a stable fraction of total capacity. instantiate_limit() now
resolves a static capacity-source resource once at instantiation time
(mem->maxmem, swap->maxswap, table-driven), stored on a new
xce_vres_t.percent_res field; a missing counterpart falls back to
today's behavior rather than failing instantiation.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01RLSNJhNoULvVcUhXEFafQC
Adds an if/then JSON Schema (draft 2020-12) conditional constraint to
limits.schema.json's Limit definition: when percent is true, value
must be a number in [0.0, 1.0]. This is the class of authoring mistake
bug-1831/bug-1832 fixed manually (night_mem_pct: value: 50 instead of
0.5) -- now caught automatically at real `xceadmin config build` time
(the actual cluster.bin generation path deploy-ha.sh uses), not merely
by a unit test.
Empirically verified against the real jsonschema library this repo
uses: a violation surfaces with the exact same file/json_pointer/
constraint contract validate_yaml()'s existing ValidationError already
provides, so zero changes were needed to its error-reporting code.
While implementing this, the new constraint caught an identical
pre-existing defect in xceadmin's own checked-in "full" golden test
fixture (fixtures/full/configs/limits.yaml, night_mem_pct: value: 50)
-- fixed the same way, and regenerated the checked-in expected.bin
golden artifact to match the corrected input.
Added 3 tests to test_validation.py: reject an out-of-range percent
value, accept an in-range one, and confirm a non-percent value is
never range-checked regardless of magnitude.
Build gotcha discovered and documented (cerebrum.md): this repo has
two copies of every bundled schema -- src/configs/schema/ (canonical)
and src/xceadmin/src/xcecommon/schema_files/json/ (a gitignored local
mirror meson-build.sh always overwrites from the canonical copy on
every build). Only the canonical copy needed a source-controlled edit.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01RLSNJhNoULvVcUhXEFafQC
night_mem_pct (resname: mem, percent: true) had value: 50, but
xce_limit_val_t.val is a 0-1 fraction convention when percent is set,
not a 0-100 number (see bug-1831 for the full pipeline trace). As
configured, its actual capacity computed as 50x (5000%) the resource's
current value -- a cap that could never realistically be exceeded by
any actual demand, making this limit functionally dead in any real
deployment.
night_mem_pct was authored directly in YAML for a docs/example config
with no C-literal ancestry to cross-check against, unlike limits2
(migrated verbatim from a hand-written C literal .val = 0.2 for a
documented "20% mem cap") -- it used the human-intuitive but incorrect
0-100 scale instead.
Changed value: 50 to 0.5. This is a real behavior change, not just a
display fix: night_mem_pct will now actually constrain the night
queue's mem usage at 50% for the first time in any deployment using
this config.
Discovered during live manual verification of scheduler-sim-diagnostics
Requirement 6 against a real deployed cluster.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01RLSNJhNoULvVcUhXEFafQC
fmt_limit()/fmt_limit_units() formatted a percent-based limit's value
as a bare snprintf(..., "%f%%", limit->val.val) -- echoing the raw
stored value with a literal "%" appended. But xce_limit_val_t.val is
genuinely a 0-1 fraction when percent is set (confirmed end-to-end:
no /100 or *100 scaling exists anywhere in the config->capnp->runtime
pipeline; confirmed authoritative by the original hand-written C
literal this codebase's oldest percent-based limit was migrated from,
.val = 0.2 for a documented "20% mem cap"). A correctly-configured 20%
limit therefore displayed as "0.200000%" instead of "20.000000%" --
misleading, though technically an honest echo of the stored value.
Fixed by multiplying by 100.0 before formatting. Display-only --
update_limit_res_value()'s actual capacity computation was already
correct and is untouched. fmt_limit() (a thin XCE_UNITS_RAW wrapper)
inherits this fix for its own percent-branch output too, a deliberate,
intentional change from its pre-fix behavior for percent-based limits
specifically (its 5 pre-existing direct tests only assert rc > 0,
never the exact percent string, so this is safe).
Updated test_limit.c's percent-mode test fixture from an unrealistic
.val=80.0 to a realistic .val=0.8 (assertion text unchanged, since
0.8*100=80 either way) and test_repl_limit_table.c's own percent test
to expect its shared fixture's arbitrary .val=50.0 constant correctly
rendering as "5000.000000%". RED confirmed via the raw Unity testlog's
:FAIL: lines (the aggregate meson summary masks single-test failures,
a known harness quirk in this project); GREEN confirmed after, both
lib and scheduler suites, 0 Unity FAILs.
Discovered during live manual verification of scheduler-sim-diagnostics
Requirement 6 against a real deployed cluster.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01RLSNJhNoULvVcUhXEFafQC
User's own independent retest against a real deployed cluster produced
a fully populated, correctly human-formatted mem-based limit usage
line (limits2: accounted=18022400.000000 (17.19 MB)
capacity=4635344384.000000 (4.32 GB)) -- directly contradicting the
original entry's "never populates" framing, which was based on a
single short-lived agent-run live test.
Downgraded from a confirmed defect to an unreproduced, timing/
session-shape-dependent observation. The original single-run finding
was real and remains unexplained, but a categorical "mem-based limits
don't populate usage via replay" claim is now known to be false.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01RLSNJhNoULvVcUhXEFafQC
Discovered during live manual verification of scheduler-sim-diagnostics
Requirement 6 (human-readable info limit formatting) against a real
deployed cluster with a real captured replay. A mem-based limit
(limits2/ug1_global_mem/night_mem_pct) never shows a live usage
instance even when its matching job is genuinely dispatched with an
intact rusage[mem=...] reservation -- while a slots-based limit
(limit1) in the same replay correctly populates one. Suspected timing/
ordering gap between WORKERS_PUSH replay-event application and
match_job_global_limit_()/match_job_host_limit_()'s resmgr lookup, not
confirmed. Out of Requirement 6's boundary (display-formatting only,
proven correct independently) and out of bug-1826's already-closed
scope (matching invocation itself, not this resource-type-specific
timing gap). Needs its own dedicated investigation.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01RLSNJhNoULvVcUhXEFafQC
Final validation task for the Human-Readable Units increment (tasks
7-10.1). Adds three tests closing gaps no prior task's own tests
covered:
- Cross-surface consistency (Requirement 6.5): a single info limit
request against a byte-valued limit with both a configured setting
and a live usage instance shows the SAME units convention in both
lines, across mode switches -- neither task 8.2's nor task 9's own
tests combined both surfaces in one capture.
- Non-byte-valued exclusion (6.3): a new "slots"-resourced fixture
(the first non-"mem" fixture in this file) proves the entire
combined display is byte-for-byte identical across all three units
modes.
- Percentage exclusion (6.3) at the REPL-dispatch level: a percent
limit's own percentage value survives real repl_dispatch()/
set-units switching unchanged, complementing test_limit.c's
existing library-level proof of the same invariant.
Full regression sweep: lib suite 368/368, scheduler suite 85/85,
including test_repl_limit_table.c (35/35), test_repl_resource_table.c,
test_repl_resource_formatting.c, and test_limit_plugin.c all
unaffected. No production code changed. This closes out Requirement 6
(ACs 6.1-6.5) and reconfirms Requirement 5 (ACs 5.1-5.3) end-to-end.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01RLSNJhNoULvVcUhXEFafQC
Six pre-existing assertions in test_repl_limit_table.c (across 4 test
functions) used bare strstr() substring checks against whole-number
fixture values on "mem" -- e.g. strstr(captured, "accounted=4.00").
Task 9's fix changed the actual output to "accounted=4.000000 (4.00
B)", but the old substring is a literal prefix of the new one, so
these assertions kept passing by coincidence of formatting, providing
zero actual regression protection for the fix.
Strengthened each to assert the full corrected "<raw> (<human>)"
string. Proved via TDD: reverting task 9's fix now makes all 6
strengthened assertions (plus task 9's own 2 genuine tests) fail;
restoring the fix returns the suite to 85/85 green.
Confirmed via full-file scan that no other loose assertion of this
shape remains, and that info worker/info resource's own regression
suites were never vulnerable to this class of bug (they build expected
strings dynamically rather than hardcoding literal substrings).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01RLSNJhNoULvVcUhXEFafQC
