Driver fixed 2026-07-16 (nvidia-driver-595-open from distro repo, old
graphics-drivers PPA for jammy disabled) — RTX 4070 Ti SUPER 16GB, CUDA
13.2, nvidia-container-toolkit already present from the 07-15 cutover
prerequisite install. Uncomments deploy.resources.reservations (nvidia
gpu) in services/ollama/docker-compose.yml, restoring clean formatting.
Docs close out the loose ends from the 07-15 declarative cutover:
- ollama-solaria-cutover runbook gets a "Wykonanie" section documenting
what actually happened (mv instead of rsync for the model store, the
missing nvidia-container-toolkit prerequisite, the driver-missing
discovery, CPU-only cutover, then the 07-16 GPU fix) plus a note on
the container disappearing after the 07-15 evening reboot (one-off,
boots fine now, root cause not established).
- hosts/solaria/services.yaml: ollama role comment now reflects actual
GPU-backed state instead of the previously-aspirational wording.
- hosts/solaria/README.md: drop stale Open WebUI mention (not in repo).
- docs/backlog.md: close the NVIDIA driver item; leave two follow-ups
(Ollama call batching before the mail phase, UNIQUE(envelope_id,
chunk_index) schema change for multi-model embeddings at phase 3).
- jobs/documents-ingest/README.md: timing section gets a GPU placeholder
line next to the existing 0.79s/chunk CPU baseline, to be filled in
after the live GPU benchmark.
Live recreate + GPU-vs-CPU embedding benchmark deliberately left for the
operator to run from the main checkout after merge, per worktree-aware
discipline — this worktree only owns the declarative fix and the docs.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
An "unknown"/empty node_name silently produced evt-unknown-* event files. One
landed in events/piha/ and, being lexically greater than every evt-piha-* name,
pinned the observer's (then lexical) checkpoint and blocked PIHA for ~34 days.
Defense in depth, two layers, both refuse rather than emit poison:
- config.py: node_name field_validator rejects ""/"unknown" (case-insensitive,
trimmed); validate_default=True so the "unknown" default itself is rejected
when NODE_NAME never reaches the process. main() catches ValidationError and
exits 1 with a clear FATAL message instead of a raw traceback.
- event_emitter.py: EventEmitter.__init__ guards node_name at the exact poison
site (node_name is embedded in the filename), so no future call path can
reintroduce evt-unknown-*.
Precedence unchanged and correct: Settings.load() uses os.environ.setdefault, so
env NODE_NAME wins over YAML; YAML only supplies node_name when env is absent.
api.py never writes event files (only /health, /trigger), so its _node_name
global cannot produce poison — verified, left as-is.
Tests (services/ha-diag-agent/tests/): new test_config.py (env load, env>YAML
precedence, fail-fast on unknown/empty/whitespace/default/YAML-unknown) and
EventEmitter guard tests. Full unit suite: 139 passed, 0 regressions.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Live cutover 2026-07-15 odkrył: nvidia-smi not found, brak libnvidia-ml
w ldcache — host nigdy nie miał sterownika, natywna Ollama mieliła na CPU.
Sekcja deploy.resources (nvidia) zakomentowana do czasu instalacji sterownika
(backlog). nvidia-container-toolkit doinstalowany na hoście (prerequisite,
brakował w runbooku). Kontener zweryfikowany: 5/5 modeli z bind-mounta,
embed bge-m3 dim=1024, reachable z PIHA przez tailnet.
Ollama's service.yaml/docker-compose.yml declared owner_node: solaria but
was never added to hosts/solaria/services.yaml, the manifest deploy-node.sh
actually reads — so it stayed running manually/natively instead of via the
declarative pipeline. Add the missing entry.
Also switch the port publish from 0.0.0.0 to loopback + TAILSCALE_BIND_IP
(same convention as llm-gateway@PIHA), since this is a private service and
llm-gateway@PIHA is the only consumer beyond the host itself.
Live cutover (relocating the existing native model store, disabling the
systemd unit, bringing up the container, verifying bge-m3 embeddings +
GPU use) is documented in docs/infra/ollama-solaria-cutover-2026-07-15.md
but not executed here — no SSH access to SOLARIA from this worktree.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
check_containers() classified only exited/dead, running+unhealthy and running.
Docker's "restarting", "paused", "removing" and any future state matched no
branch, so a crash-looping container (restart policy + continuous crash) emitted
ZERO events and world-state kept showing it "healthy" (observed live:
lustro/pi-watchtower-1). Every Docker state is now handled:
- restarting + RestartCount >= CRASH_LOOP_RESTART_THRESHOLD (default 3):
reuse containers_not_running (high, crash_loop=true) — parity with exited/dead,
rides the existing supervisor-wired remediation path.
- restarting below threshold: new observational container_restarting (low),
visible but non-actionable so benign post-deploy restarts don't alarm.
- paused / unknown-or-future state: new diagnostic container_state_unexpected
(medium) — no more silent fall-through; new Docker states become visible.
- removing: conscious documented skip (ephemeral teardown).
- created: unchanged skip (compose tracking artifact).
RestartCount (top-level inspect field) distinguishes a crash-loop from a one-off
restart. Adds services/node-agent/tests/test_check_containers.py pinning the full
state table.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Per-node checkpoint now stores the last-processed event TIMESTAMP (int epoch)
instead of a file path compared lexically. A file is "new" iff its timestamp
(parsed from evt-<node>-<unixts>-<type>-<svc>.json, mtime fallback) exceeds the
node's checkpoint; processing is ordered by timestamp, not path.
Root cause (PIHA dead ~34d, 2026-07-12): a stray evt-unknown-<ts>-… file landed
in events/piha/, lexically greater than every evt-piha-… name. The lexical
checkpoint pinned there, so every genuinely newer piha event sorted "before" it
and was skipped forever. Event backlog grew to 7344 files, last_seen frozen,
shadow-read logged false SHADOW_LIVENESS_MISMATCH event=dead prom=up.
- _event_ts_from_path: filename epoch, mtime fallback; NEVER returns 0 for an
existing file (0 == "older than checkpoint" == the poison).
- _checkpoint_ts_from_value: graceful migration of pre-fix path-string
checkpoints (and the older last_processed_file format) to int epochs;
unparseable → 0 (reprocess all — safe, process_event is idempotent on
last_seen/world_state; bias to reprocess, never to skip).
- Preserved: quarantine of bad events, observer-source re-ingest guard.
- Regression tests (test_incident_lifecycle.py section 9): lexically-smaller-
but-newer processed, unparseable name falls back to mtime (not wedged),
ts-not-path ordering, both checkpoint-format migrations, helper units.
Separate bug filed in backlog (not fixed here): ha-diag-agent emits node=
"unknown" events (config.py node_name default) into another node's dir when
NODE_NAME reaches the compose volume path but not the app env — the source of
the poison file.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Krok 1 planu docs/kb/modules/05-faza2-plan.md §3/§6 (chunk-level embeddings,
1:N do envelope). Addytywna — 001_envelope.sql nietknięta (zweryfikowane
\d envelope po migracji: identyczny schemat + FK jako "Referenced by").
Schemat wg rekomendacji recon (§2 decyzja 1+2): osobna tabela (nie kolumna
w envelope, bo N-wartościowy chunking jest obowiązkowy przy dokumentach
>8k tokenów), embedding VECTOR(1024) pod bge-m3 (dense), HNSW cosine index,
kolumna `model` do trywialnego re-indexu przy zmianie modelu (kb-00 zasada
#1: indeks odtwarzalny). Idempotentna (CREATE TABLE/INDEX IF NOT EXISTS,
zweryfikowane podwójnym uruchomieniem na kb-postgres@PIHA — drugi run same
NOTICE "already exists, skipping").
Zastosowana na żywej bazie: ssh piha docker exec kb-postgres psql, po
potwierdzeniu SQL przez Oskara. \dt + \d document_chunk + \d envelope
zweryfikowane po migracji.
Testy: dopisane sanity-testy 002 do packages/kb-mail/tests/test_migration.py
(wzorzec 001 — statyczne assercje na treści SQL, bez DB), 13/13 zielone.
Co NIE jest częścią tego kroku (§3 planu, odłożone): entity/entity_link
(graf encji) — szkic na przyszłość, nie blokuje domknięcia modułu 5.
Co dalej (plan §6, poza zakresem tego kroku): ollama pull bge-m3 na SOLARII,
token API Paperless, jobs/gmail-header-backfill/, adapter Paperless→koperta,
chunking+embed job.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Two config bugs found on the already-deployed split-host OCR worker
(module 3): (1) `command: celery ...` was routed through manage.py by
the image entrypoint because it didn't start with "/" — fixed with an
absolute gosu+celery path. (2) SCRATCH_DIR (/tmp/paperless) was not
shared over NFS like data/media/consume, so tasks picked up by
worker@SOLARIA instead of worker@PIHA failed with "File not found" —
fixed by adding a paperless_scratch NFS volume/bind mount on both
sides.
Verified live on PIHA + SOLARIA: test PDFs dropped into consume/ were
split across both workers, the SOLARIA-picked task completed OCR with
zero File not found errors, test documents cleaned up afterward.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Nextcloud zostaje prywatny (mesh/kapala.org); Gokapi to osobny publiczny
serwis do wysylania linkow do plikow na zewnatrz (Firefox Send alt).
owner_node=vps, storage lokalny dysk (nie S3), E2E encryption ON, port
53842 bindowany tylko na TAILSCALE_BIND_IP (hairpin NAT przez npm@VPS,
nigdy 0.0.0.0). Cutover checklist w README (DNS, wildcard *.okit.pl na
npm@VPS, setup wizard) — nie zdeployowane w tym commicie.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Co zrobione:
- Nextcloud host = PIHA (always-on dla aktywnego uzycia), owner_node +
LAN_BIND_IP/TRUSTED_PROXIES w .env, README zaktualizowane
- Redis brokera Paperlessa: requirepass, PAPERLESS_REDIS_PASSWORD w .env
po obu stronach (PIHA + worker@SOLARIA), healthchecki z auth
- Domeny potwierdzone: paper.kapala.org, cloud.kapala.org (Cloudflare
DNS-only -> Tailscale PIHA, wildcard cert juz pokrywa) — udokumentowane,
nic nie utworzone
- Backup Paperlessa zatwierdzony: document_exporter + rsync/borg -> SOLARIA,
retencja 7/4/6, offsite jako future-note
- Nextcloud pin: 34-apache (zweryfikowany aktualny stable, endoflife.date)
- Whoosh fallback-worker: zaakceptowane bez zmian
- Porty/wylaczenie local login/sizing OCR-workera: przeniesione z "decyzji"
na "TODO przy deployu"
- DECYZJE-do-podjecia.md zaktualizowane: wszystko poza portami/loginem/
sizingiem przeniesione do "Rozstrzygniete"
Tylko edycja configow w repo — nic nie zdeployowane, zadne kontenery nie
byly ruszane, DNS/vhosty nie utworzone.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Observer now optionally (PROM_SHADOW_URL) queries Prometheus up{} once per
cycle and LOGS SHADOW_LIVENESS_MISMATCH when its event-driven liveness
disagrees. Parallel-run only: compute_liveness and _emit_node_transition are
untouched; authoritative liveness stays 100% event-driven. Fail-open on any
Prometheus error (down/timeout/bad JSON -> {}). 9 new tests, incl. proof that
shadow-read does not change node_info liveness/status.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Kod zyl tylko na dysku PIHA w /opt/llm-gateway (bez gita). Przeniesiony do
services/llm-gateway/ pod pelny wzorzec homelaba:
- app/main.py: kod 1:1 z PIHA + OLLAMA_URL/CHAT_MODEL/CODE_MODEL
nadpisywalne przez env (defaulty bez zmian)
- docker-compose.yml: bind TYLKO do Tailscale IP (${TAILSCALE_BIND_IP},
wzorzec fleet-prometheus), nie 0.0.0.0 jak w starym compose
- service.yaml, env.example (bez sekretow), healthcheck.sh, README, testy
- hosts/piha/runtime/llm-gateway: mem_limit 256m (PIHA jest RAM-bound)
- rejestracja w hosts/piha/services.yaml i inventory/topology.yaml
Zepsuty /opt/llm-gateway/docker-compose.yml (zduplikowany klucz ports)
celowo NIE przeniesiony.
DoD: pytest 4 passed; docker build + smoke run OK (GET / -> gateway ok).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The test_run_once_* cases were flaky/order-dependent. Root cause: observer.observer
derives OBSERVER_STATE_FILE from STATE_DIR at import time. The helper patched
STATE_DIR but never OBSERVER_STATE_FILE, so run_once()/_save_checkpoint() wrote the
checkpoint to the real /opt/homelab/state/observer_checkpoint.json. Those node_checkpoints
(tmp paths tagged with a pytest run number) leaked across tests and across pytest runs;
run_once's `file_path > checkpoint` string compare then skipped/kept events based on
run-number ordering. The helper also never restored the module globals it overwrote.
Replace both ad-hoc helpers with an autouse monkeypatch fixture that redirects every
observer path — including OBSERVER_STATE_FILE — into the per-test tmp_path and reverts
them afterward. Tests no longer touch real disk and are deterministic.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Previously published "9090:9090" → Docker bound to 0.0.0.0 (all host interfaces,
including the public Hetzner IP), leaving tailscale-internal enforced only by the
VPS firewall. Now bind explicitly to the VPS Tailscale interface for
defense-in-depth: the port does not exist on the public IP at all.
- ports -> "${TAILSCALE_BIND_IP}:9090:9090"
- env.example: add TAILSCALE_BIND_IP (VPS Tailscale IP, verify via `tailscale ip -4`)
- README: deploy section — .env is mandatory; a missing .env makes Compose
silently bind 0.0.0.0 (warns, does not fail), so use --env-file and verify host_ip
Smoke: config with --env-file and with co-located .env both resolve
host_ip=100.95.58.48; with .env absent Compose warns and falls back to 0.0.0.0
(documented). .env is gitignored (global *.env rule).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Clean scaffold of a Prometheus instance dedicated to fleet liveness, deliberately
separate from the home `prom` on PIHA. Scrapes only itself + the VPS-local
node_exporter for now.
- image pinned to prom/prometheus:v3.5.0 (LTS) — no :latest anti-pattern
- TSDB retention 15d / 2GB on the tight VPS; --web.enable-lifecycle for reloads
- mem_limit 512m, oom_score_adj +200 (sacrificial OOM victim before control-plane)
- tailscale-internal exposure via plain published 9090, mirroring control-plane
- node_exporter scraped via host.docker.internal:9100 (it runs network_mode: host)
- secret-free prometheus.yml; alerting/rule_files left as commented placeholders
- in-container healthcheck via busybox wget (present in the image; curl is not)
Smoke: docker compose config OK; promtool check config SUCCESS; up -d ->
/-/healthy + /-/ready OK; both targets (prometheus, fleet-node) report up.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Fixes the "dead node shown NOMINAL" silent outage: node status was set only by
events and never expired, so a node that crashed/lost connectivity stayed
"online" forever (chelsty-infra was online for 16d, piha ~6d). The only thing
that flipped status to offline was a node_offline event, which an unreachable
node can never emit.
Now node status is derived from freshness (now - last_seen), recomputed every
observer cycle (incl. cycles with no new events):
- always-on: fresh <=180s, stale 180-600s, dead >600s (3x the 60s heartbeat)
- remote/LTE (chelsty-*): fresh <=900s, stale 900-3600s, dead >3600s
Thresholds + tier logic live in ONE shared helper, services/control-plane/src/
liveness.py, imported by the observer and both operator UIs (bind-mounted into
the agent-system webui image). No 3x copy.
Transitions are not silent: the observer emits node_stale / node_offline /
node_online (recovery) events tagged source=observer (skipped on re-ingest so
they never reset last_seen), routed by the supervisor to alert_only actions.
Read-time safety net: both UIs recompute liveness from last_seen at request
time, so a stalled observer still surfaces dead nodes. Services inherit their
node's liveness (cascade, variant B) without mutating services.json.
Replaces the earlier binary NODE_OFFLINE_TTL_SECS flip.
Tests: liveness unit tests, observer 3-state + transitions/recovery/baseline +
self-event skip, operator_ui read-time net + cascade, supervisor node-event
routing. 89 passed. docker compose config valid for both stacks.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Nodes that crash or lose connectivity without emitting node_offline
stay online in world state indefinitely. _prune_stale_world() now
flips any online node to offline if its last_seen is older than
NODE_OFFLINE_TTL_SECS (default 300 s = 5× the 60 s heartbeat interval).
Nodes with last_seen=None (never reported) and already-offline nodes
are left unchanged. Five new tests cover all branches.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Addresses review findings — all reverting to the live config that worked:
- config.yml mounts at /app/vikunja/config.yml (live path), not /etc/vikunja.
- Remove the vikunja container healthcheck: the image ships no wget/curl, so
an in-container HTTP check is always unhealthy. Health stays on db
(pg_isready) + host-side healthcheck.sh (curl). Live had no app healthcheck.
- Secrets injected exclusively via env_file (.env) on BOTH services; dropped
the ${VAR:?} parse-time interpolation that depended on a .env in cwd. db now
also has env_file. Non-secret env stays inline.
- Rename .env.example -> env.example to match repo convention (forgejo).
Verified: `docker compose -f services/vikunja/docker-compose.yml config` passes;
services/vikunja/.env is gitignored.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Wraps the existing manually-run Vikunja instance on PIHA in the standard
service layout. No deploy — definition only, for review.
- services/vikunja/: docker-compose.yml (vikunja + postgres:16-alpine),
service.yaml, README, healthcheck, config.yml (OIDC, secret-free),
.env.example. Real .env stays gitignored.
- Pins EXISTING named volumes (vikunja_vikunja_db, vikunja_vikunja_files)
so DB + uploaded files survive cutover.
- extra_hosts forgejo.okit.pl=192.168.31.5 (npm on PIHA) so OIDC discovery
resolves over LAN instead of flaky public DNS.
- OIDC client secret injected via env (VIKUNJA_AUTH_OPENID_PROVIDERS_
FORGEJO_CLIENTSECRET); config.yml keeps trailing-slash authurl/redirecturl.
- Registers vikunja in hosts/piha/services.yaml + inventory/topology.yaml,
plus hosts/piha/runtime/vikunja override (advisory mem_limits).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Recovery from bad merge of task/observer-poison-quarantine (c255a02)
which carried false deletes from a stale branch base. Re-applies only
the genuine observer changes on top of correct master state.
When an event file fails to parse (malformed JSON, truncated, corrupted),
the observer previously kept retrying on every cycle while the node's
checkpoint stayed pinned — all subsequent good events for that node lost.
Now: first parse failure -> atomic os.replace to STATE_DIR/observer_failed_events/<node>/
with collision handling. Checkpoint advances, downstream events flow.
Move failures are logged but don't crash the loop.
Complementary to the atomic_write_json fix on state files; this addresses
the same race-pattern on event files instead.
Regression test asserts: bad event quarantined to failed_events dir,
removed from hot path, subsequent good event processed (node online),
checkpoint moves to good event.
Port 8087 is no longer mapped to the host. Operator verify commands
that used curl http://localhost:8087/health now use docker exec with
Python's urllib (the image is python:3.11-slim, no curl binary).
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Port 8087 conflicted with zigbee2mqtt on piha (8087:8080 mapping active
for 7+ days), preventing ha-diag-agent from starting.
Grep across the full repo confirms no external consumer (no nginx/npm
proxy, no Prometheus scrape, no control-plane reference) uses this port.
The Docker healthcheck runs inside the container network namespace and
does not require a host-side mapping. Internal FastAPI binding on 8087
is unchanged.
Removed: ports section from docker-compose.yml and service.yaml.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- main.py: rename event= to ha_event= in _log.warning() — structlog treats
'event' as a reserved positional arg; the old name caused TypeError when
any check returned unhealthy results (events were still emitted, but the
check was logged as check_error instead of check_unhealthy)
- tests/test_ha_client.py: replace aioresponses with unittest.mock — aioresponses
0.7.8 is incompatible with aiohttp >=3.12 (missing stream_writer kwarg)
- pyproject.toml: remove aioresponses from dev dependencies
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
stability-agent had no USER instruction and no user: in compose, running
as root and writing root-owned files to /opt/homelab bind-mount.
- Dockerfile: add useradd -m -u 1000 homelab + USER homelab
- docker-compose.yml: add user: "1000:1000" and group_add: ["999"]
(GID 999 = docker group on VPS) to retain docker.sock:ro access
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
node-agent had no USER instruction and no user: in compose, running
as root and writing root-owned files to /opt/homelab bind-mount.
- Dockerfile: add useradd -m -u 1000 homelab + USER homelab
- docker-compose.yml: add user: "1000:1000" and group_add: ["999"]
(GID 999 = docker group on VPS) to retain docker.sock access
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Executor was the only control-plane container running as root (uid=0),
writing root-owned files to /opt/homelab via bind-mount and triggering
false sudo on every deploy.
- Dockerfile: add USER homelab after useradd (useradd already present)
- docker-compose.yml: add user: "1000:1000" and group_add: ["999"]
(GID 999 = docker group on VPS) so executor retains docker.sock access
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
deploy-local.sh previously ran `sudo chown -R 1000:1000` and
`sudo chmod -R 775` unconditionally on every deploy, which blocked
non-TTY execution (CC/CI) on VPS where /opt/homelab is already 1000:1000.
Both steps are now conditional using `find ... -print -quit`:
- chown: runs only if any file/dir is NOT uid/gid 1000
- chmod: runs only if any directory is missing -775 permission bits
When everything is correct (steady state on VPS), both steps log
"already correct, skipping" and never invoke sudo. If a new directory
was created by root (e.g. a manual mkdir, volume mount, or restart artefact),
the remediation path triggers automatically — the self-heal property is preserved.
Smoke-tested in Docker (ubuntu:22.04):
Case 1 (1000:1000 + 775): chown skipped, chmod skipped ✓
Case 2 (root-owned subdir): chown triggered ✓
Case 3 (700 dir perms): chmod triggered ✓
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Two root causes for stale "active" incidents on the dashboard:
1. TypeError bug in _prune_stale_world: last_occurrence / resolved_at
can be an ISO-8601 string (stability-agent via events.py) or a Unix
int (node-agent). The previous session's auto-resolve did plain
`time.time() - last_occ` which raises TypeError for strings,
silently preventing _save_world() from being called and leaving
incidents perpetually "active" on disk.
Fix: add _parse_ts(ts) -> float that handles int, float, and
ISO-8601 strings uniformly. All timestamp arithmetic now goes through
it; returns 0.0 on None / garbage to keep comparisons safe.
2. Orphaned active incidents: _resolve_incident clears service["incident_id"]
and marks the incident "resolved" in memory, but if incidents.json was
truncated mid-write (pre-atomic-write era), the observer loaded it at
next startup with status="active" and no service entry pointing to it.
No code ever touched these orphans again.
Fix: _prune_stale_world now runs two cleanup passes each cycle:
- Case 1 (healthy-linked): service.status=="healthy" AND incident_id
still set → resolve immediately (service cannot have active incident)
- Case 2 (orphaned): active incident with no service link AND
last_occurrence > 5 min ago → resolve (5-min guard for creation race)
Both cases are wrapped in try/except so a bug here never crashes the
observer loop or blocks _save_world.
Also fixes the 7-day stale-incident prune to use _parse_ts so
ISO-string resolved_at values are handled correctly.
3. Operator UI: current_incidents() now filters to status=="active" only.
Resolved incidents were previously included in the /incidents endpoint,
making the dashboard show a wall of historical records as if active.
Nocturnal job investigation: _cleanup_control_plane_fs in node-agent runs
every 60s on VPS (not midnight-specific); it reads observer_checkpoint.json
(now written atomically) and deletes old event files. No non-atomic writes
found. Midnight clustering was likely external (logrotate / OS flush);
the supervisor's resilient loader already handles such transient issues.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
11 new test cases in test_state_reliability.py covering:
- atomic_write_json: produces valid JSON, no .tmp left behind, overwrites,
works with nested structures
- _load_actual_state: returns False on empty / truncated file, returns True
on valid files, preserves last-known-good state across a parse failure
- reconcile: empty/truncated services.json or incidents.json generates zero
actions (skip-cycle semantics proven end-to-end)
- healthy service with valid world state generates no spurious action
All 32 tests (11 new + 21 existing) pass.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Two independent fixes for the false-alarm storm caused by race-condition
reads of truncated world state files:
1. Atomic writes: _atomic_write_json (write→fsync→os.replace) replaces
all bare open('w')+json.dump calls in supervisor and executor, so the
action-file pipeline is never visible in a half-written state.
2. Resilient loader: _load_actual_state now returns False when any world
state file fails to parse (empty or truncated mid-write). reconcile()
skips the entire drift check on False instead of treating {} as "all
services missing". actual_state retains its last-known-good values so
a single bad cycle does not wipe accumulated context.
Before: parse error → raw[key]={} → all desired services missing →
wall of redeploy actions → drift_resolved_auto churn on next cycle.
After: parse error → WARNING logged → cycle skipped → no actions.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
7 cases: package importable, fresh ok, stale, unreachable, HTTP error,
missing last_update field, unparseable timestamp. pytest.ini sets pythonpath=src
so tests run without PYTHONPATH set in the environment.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
WORKDIR is /app but the package lives under src/; without PYTHONPATH set
`python -m brain_watchdog.main` raised ModuleNotFoundError on startup.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Polls /summary on VPS over Tailscale every 60s; computes freshness
locally from last_update epoch (never trusts self-reported status).
Alerts via Telegram Bot API directly after 3 consecutive failures;
sends recovery message on heal. State (fail_count, alerted) persisted
to volume so debounce survives restarts.
- services/brain-watchdog/: Python service, no external deps (stdlib only)
- hosts/piha/runtime/brain-watchdog/: override with mem_limit 64m
- hosts/piha/services.yaml + inventory/topology.yaml: manifest entries
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
operator_ui.py called .replace() on last_update without checking type —
an integer value (written by the materializer) raised AttributeError and
silently fell back to os.path.getmtime(), which was stuck at 5/29 after a
deploy with preserved timestamps. web.py had the same class of bug but
worse: it unconditionally replaced last_update with mtime, ignoring the
JSON field entirely. Both now branch on isinstance(str) and cast numeric
values directly to float, with mtime only as a last-resort fallback.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Oskar Kapala