relay-785 / self-heal-contract-stale-script-pointers-20260924. gpu/ had been
silent since 2026-09-14T18:02:03Z (12 days) and pm2/ had never posted a run.
FINDING - the gpu-self-heal executor was never lost, only its schedule was.
The brief concluded the mechanism was gone. It is not: on CT 116
/opt/inference-harness/scripts/gpu-self-heal.py exists (mtime 2026-09-21 11:43),
is self-posting via gitea-logger.sh, and runs cleanly - I ran it and it posted
health-logs/gpu/gpu-self-heal-20260926-144618.json. What was missing was the
cron entry, dropped during a CT 116 /etc/cron.d rework on 2026-09-21. So this is
a schedule restoration, not a resurrection.
DECISIONS
* gpu-self-heal -> RESTORE. Schedule re-added as /etc/cron.d/gpu-self-heal
on CT 116, matching the observed historical cadence (:02 past
0/6/12/18). Proven with a REAL unattended cron fire (temporary */1 entry,
removed after): gpu-self-heal-20260926-144802.json committed 14:48:03Z.
The contract now names the executor, schedule, log and posting, which it
previously did not.
* pm2 health-logs posting -> RETIRE the claim. pm2-self-heal.prose.md required
appending to health-logs/pm2/ as a 'hard rule', but scripts/pm2-self-heal.sh
contains no Gitea or push code and never did; the directory has held only its
init commit since 2026-07-28. Corrected to point at the contract-runner's
durable per-run logs and failure note instead of adding a second, redundant
posting path.
DEAD-MAN'S-SWITCH - scripts/health-log-freshness.py + contract. Absence of logs
must raise an alarm, and the producer cannot raise it, so this runs on CT 100,
a different host from the producers, and fails when the newest health-logs/gpu
entry is older than 12h (litellm 18h). Verified it would have caught the real
gap: evaluated at 2026-09-20 the newest gpu/ entry was 126h old against a 12h
limit -> STALE.
Bugs found and fixed while testing, each caught by a test that bit:
* the documented HEALTH_LOG_MAX_AGE_* override was never implemented;
* a Gitea password held under GITEA_TOKEN was sent as an API token -> HTTP 401;
now every candidate auth is tried and the first that works is used;
* the ~/.git-credentials fallback filtered on GITEA_URL's host, which missed
whenever GITEA_URL pointed at the internal IP -> zero candidates.
Exit codes: 0 fresh, 1 stale-or-unreadable, 2 cannot run. A directory that
cannot be read is a failure, never a skip.
Evidence: live PASS; stale override names the directory and producer; no
credential -> exit 2; whole thing runs green through contract-run.sh.
prose-lint: PASSED.
- Remove retired names (qwen3.6-27B-code, qwen3.6-35B-udq4) from live alias claims
- Update Strix Halo model to Carnice-Qwen3.6-MoE-35B-A3B-Q4_K_M.gguf (strix-moe, 256K ctx)
- Fix litellm-health step 7 probe to gpu-vision (monitor key scoped)
- Move qwen3.6-27B-code/35B-udq4 from raw-but-live to non-resolving in audit
- Fold in pm2-self-heal: remove spoton-service (live PM2 set is 4/4)
- Update hermes templates, key enforcement, timeout tables to live names
audit-hermes-config.py Rule 8 required auxiliary.vision.model and
auxiliary.web_extract.model to equal the retired 'gpu-light', so a config
adopting the live canonical 'gpu-vision' FAILED our own audit - the audit was
enforcing a dead alias (400 Invalid model name). Rule 8 now requires
gpu-vision; retired names gpu-light/crew-auto join the raw-name rejection set;
the guidance message names the live aliases.
Sweep of the remaining references: gpu-self-heal stops canonicalizing
gpu-light; hermes-config-template, hermes-agent-baseline, hermes-key-enforcement,
inference-optimization, litellm-client-timeouts and gpu-fleet now use the live
gpu-vision alias. Where a file restated model/rpm/weight/fallback state it now
points at CT 116 /opt/inference-harness/litellm_config.yaml instead of
duplicating it. koby's .129 config is report-only and recorded, not edited.
Adds tests/test_audit_hermes_config_alias.py: executes the audit CLI and asserts
gpu-vision passes while gpu-light and gemma-4-12b fail.
- pm2-self-heal.sh: restart abiba-telegram when TEL_RESTARTS > 1000 even if 'online'
(catches quiet crash-loops like the 10k-restarts spoton incident); alert includes count.
- pm2-self-heal.prose.md: document the crash-loop guard under Rule 2.
- gpu-fleet.prose.md / gpu-self-heal.prose.md: gpu-dense is Qwen3.8-27B-Uncensored-Q4_K_M
(~16.8GB, alias qwen3.6-27B-code for LiteLLM routing), not SmartCode-Fable-5 (verified live
on .8:8080 Aug 16).
Logs (LITELLM-HEALTH, GPU-SELF-HEAL, PM2) now pushed to SyslogSolution/health-logs
instead of creating orphan nodes in the shared knowledge graph.
- litellm-self-heal: Phase 4 now calls gitea-logger instead of kg-logger
- gpu-self-heal: Reporting section updated to Gitea path
- pm2-self-heal: Log step redirected to Gitea
- 271 existing orphan nodes remain in graph (no delete tool available)
root
mumuni-bot