Rule 5 now treats internal http://192.168.68.116/v1 as non-canonical
but working (authenticated via nginx), producing a WARNING instead of a
FAILURE. The canonical internal path /litellm/v1 and the public host
https://litellm.sysloggh.net/v1 both PASS. Everything else FAILS.
Prose aligned: hermes-key-enforcement.prose.md now states the canonical
internal form, notes that internal /v1 still works but is flagged as
non-canonical (WARN not FAIL), and clarifies that the public host serves
/v1 ONLY (404 on /litellm/v1). Corrected the 'unauthenticated path'
wording at line 116, which was factually wrong.
Tests updated: BASE template uses canonical internal path; new test
cases prove canonical /litellm/v1 PASSES, wrong path FAILS, public host
PASSES, and internal /v1 WARNS (not FAILS). Fixed backwards comment in
test_old_rule5_check_would_fail_canonical.
Three states:
- healthy: passed, exit 0 (unchanged)
- busy (completion timed out after retry AND host /health answered):
⚠️ DEGRADED line, does NOT fail the run, exit 0
- host unreachable or real fault: ❌, exit 1 (unchanged)
Summary now reports degraded count:
- All pass, no degraded: '✅ All checks passed'
- All pass, 1+ degraded: '✅ All checks passed (1 degraded: gpu-dense)'
- Some failed: '❌ Some checks failed' or '❌ Some checks failed (1 degraded: ...)'
Host health mapping verified:
- gpu-dense -> 192.168.68.8:8080/health
- gpu-vision -> 192.168.68.110:8080/health
- strix-moe -> 192.168.68.15:8080/health
1. TIMEOUT KIND FIX: run_command returns (1, '', 'TIMEOUT') when its own
timeout fires. probe_http now checks for this before falling through to
'curl exit <rc>', so a 30s timeout reports 'timeout after 30s' not
'curl exit 1'.
2. BUSY/DEGRADED DETECTION: After both model probes fail, check the
model's host health endpoint (e.g. 192.168.68.8:8080/health for
gpu-dense). If the host answers 200, report 'busy (completion timed
out after retry; host healthy 200)' — do NOT fail the run on that
alone. If the host does not answer, that's a real FAIL.
3. RETRY TIMEOUT INCREASED: Single-host retry timeout raised from 45s to
90s. Worst-case prefill on a single-slot .8 host is ~76s (observed
83K-token prompt at 1078 tok/s), so 90s covers it.
New line shapes:
- Busy: 'gpu-dense: busy (completion timed out after retry; host healthy 200)'
- Real failure: 'probe-failed: gpu-dense timeout after 30s then timeout after 90s (2 attempts)'
Rule 5 in audit-hermes-config.py had an inverted check: it expected
base_url=http://192.168.68.116/v1, but the contract hermes-key-enforcement.prose.md
names http://192.168.68.116/litellm/v1 as CORRECT/CANONICAL in multiple places.
The audit script would FAIL a config using the contract's canonical internal path
and PASS one using a path the contract does not name.
Fix: Rule 5 now accepts the canonical internal base (http://192.168.68.116/litellm/v1)
AND the public base (https://litellm.sysloggh.net/v1), and FAILS anything else.
The internal nginx serves both /litellm/v1 and /v1; the public host serves /v1 only
(per 2026-09-19 probe from CT 116).
Tests aligned: BASE template updated to use the canonical internal path, and new
test cases added to prove the canonical internal path PASSES, a wrong path FAILS,
and the public host path PASSES.
The failure line now preserves both attempts' failure kinds instead of
hardcoding 'timeout after retry, 45s'. If both attempts fail, the report
shows: 'probe-failed: <model> <first kind> then <retry kind> (2 attempts)'.
This fixes the self-contradictory output when the first attempt timed out
but the retry failed with connection refused, and prevents the duration
from appearing twice when both attempts were timeouts.
Example outputs:
- timeout then timeout: 'probe-failed: gpu-dense timeout after 30s then timeout after 45s (2 attempts)'
- timeout then refused: 'probe-failed: gpu-dense timeout after 30s then connection refused (2 attempts)'
- refused then refused: 'probe-failed: gpu-dense connection refused then connection refused (2 attempts)'
Single-host models (gpu-dense, gpu-vision, strix-moe) now retry once at
45s on initial 30s timeout failure before declaring probe-failed. This
prevents a single transient timeout (cold prefill ~13s or concurrent
generation hold) from failing the entire health digest.
Evidence: 2026-09-19 ~06:55Z digest failed gpu-dense at 30s; 06:56Z
direct probe 200 in 1.04s.
The failed-probe-fails-the-run property is preserved: if both attempts
fail, the script still exits non-zero with the target and duration named.
Closes: daily-health-digest false negative on single transient timeout
The leg comments were wrong:
- Line 207: Docker Stats showed :9323 (dockerd port) but should be :9324
- Line 215: PVE Exporter showed :9324 (docker-stats port) but should be :9221
These were the exact pairing this PR exists to correct.
Read back the changed lines to verify:
scripts/infra-monitoring.sh:207 shows Docker Stats (CT 116 :9324, 127.0.0.1 via SSH)
scripts/infra-monitoring.sh:215 shows PVE Exporter (CT 116 :9221, 127.0.0.1 via SSH)
Branch: fix/infra-monitoring-probe-ports-20260919
F1: Fixed leg comments to match the actual ports
- Line 207: Docker Stats now shows :9324 (was :9323)
- Line 215: PVE Exporter now shows :9221 (was :9324)
These were the exact pairing this PR exists to correct.
F2: Added per-leg assertions that prove which leg owns which port
The new assertions verify:
1. Docker Stats leg uses $DOCKER_STATS_PORT constant
2. PVE Exporter leg uses $PVE_EXPORTER_PORT constant
3. DOCKER_STATS_PORT constant is set to 9324
4. PVE_EXPORTER_PORT constant is set to 9221
Proof the new assertions bite:
Under the both-constants-swapped mutation (DOCKER_STATS_PORT=9221,
PVE_EXPORTER_PORT=9324), the suite fails with 25 passed / 2 failed
(failing exactly the two constant-value assertions). This proves the
per-leg assertions pin which leg owns which port, not just that both
ports appear somewhere in the SSH log.
Branch: fix/infra-monitoring-probe-ports-20260919
The Liveness Check section now documents ALL SIX verdict shapes exactly
as emitted by proxmox-monitor.sh:
1. ✅ PBS GC: healthy (last run Nh ago, pending-bytes: N B)
2. 🔴 PBS GC: stale (last run Nh ago, pending-bytes: N B)
3. 🔴 PBS GC: probe-failed: storepve:192.168.68.6 (expected JSON, got 000)
4. 🔴 PBS GC: probe-failed: storepve:192.168.68.6 (unparseable JSON)
5. 🔴 PBS GC: never-run (storepve-datastore not found in GC list)
6. 🔴 PBS GC: never-run (storepve-datastore has no last-run-endtime)
Fixed the quoted healthy example (line ~114) to include the pending-bytes
suffix the code now appends. Previously the prose only documented the
probe-failed shape, missing the PR's own headline cases (never-run).
Proof: grep -n 'never-run' proxmox-monitor.prose.md now returns two lines
(lines 106 and 108), documenting both never-run variants.
Branch: fix/pbs-gc-liveness-signal-20260919
Previously probed wrong ports:
- Docker Stats was at 9323 (dockerd metrics) but should be 9324
(harness-docker-stats, docker_container_* metrics)
- PVE Exporter was at 9324 (harness-docker-stats) but should be 9221
(harness-pve-exporter, 5 pve_* metrics)
Both exporters bind to 127.0.0.1 on CT 116 and must be probed via SSH.
Updated infrastructure-monitoring.prose.md to document the correct ports.
Added test assertions verifying the exact ports are probed.
Branch: fix/infra-monitoring-probe-ports-20260919
(a) Probe-failure detection: now treats empty OR unparseable JSON as
probe-failed, not never-run. This prevents 'command not found'
outputs from being rendered as service verdicts.
(b) pending-bytes: now extracted from JSON and reported in stale verdict.
(c) Null endtime: use .get() with explicit None check, not 0 fallback.
null values now correctly trigger never-run verdict instead of
arithmetic crash (set -u).
(d) Tests: Added 14-assertion stub-driven suite covering: healthy,
stale (>48h), probe-failed (empty and unparseable), null endtime,
datastore absent. Each test stubs ssh/curl to verify exact
behavior against the pre-fix head.
Branch: fix/pbs-gc-liveness-signal-20260919
Document the PBS GC schedule (00:00 UTC, not 20:00 UTC as PR #116 said),
what actually runs (pbs-gc.sh -> pct exec 107 -- proxmox-backup-manager),
datastore location (CT 107's /mnt/pbs-backup on /tank/pbs-backup, NOT
/media/easystore2), and the new liveness check (48h threshold, reports
age in hours, explicit healthy line).
Branch: fix/pbs-gc-liveness-signal-20260919
Add monitoring leg that checks storepve-datastore GC health:
- Reads GC state from CT 107 via pct exec
- FAILS if last-run-endtime is older than 48h
- Reports age in hours and pending-bytes status
- Uses JSON parsing for reliable data extraction
Test: All 5 legs OK, Exit 0.
Branch: fix/pbs-gc-liveness-signal-20260919
PR #118 finding F1 (low): The deployed LiteLLM on CT 116 uses
allowed_mcp_servers (193 occurrences in installed package), not
bare allowed_mcp. One-word doc fix.
Add the audit-hermes-config.py Rule 15 wording fix from PR #117:
- Violation message now reads 'URL is incorrect: <url> (expected: <expected>)'
- Detection logic unchanged
- Matches URL and not-in-known-list branches remain byte-identical
This consolidates relay #779 into a single PR (#118).
PR #117 follow-up (verify PASS-WITH-FINDINGS):
1. infrastructure-update.prose.md:
- Update access table: agent keys now have per-key MCP grants (2026-09-18)
- Strike-through old limitation: per-key grants now work
- Mark Migration Path as COMPLETED 2026-09-18
2. hermes-config-template.prose.md:
- Remove hedge ('may have been upgraded')
- State fact: per-key MCP grants verified 2026-09-18
This resolves the contradiction where one file asserted per-key
MCP access and the other denied it.
F1: Make kind classification real — append (<kind>) to every failure
line, assign kind=tls on curl exit 35/60, fix :112 where
kind=refused was set on successful retry. Update prose shape.
F3: Move credential placeholder skip from server probe to notify()
only — server is always probed (200 without auth verified live).
F4: notify() logs ALERT SUPPRESSED when credential unusable so
alerts from other legs are not silently dropped.
F7: Restore trailing newline in infra-monitoring.sh.
F5 (DO NOT CHANGE): Verified directly — ssh root@192.168.68.6
'grep -n keep-daily /etc/pve/jobs.cfg' returns five
prune-backups keep-daily=35 lines. Prose is CORRECT.
Test: 18 passed, 0 failed (bash scripts/test_infra_monitoring.sh)
When ZULIP_API_KEY is unset or contains 'placeholder'/'REDACTED', skip
the global Zulip server leg with a ⏭ marker instead of failing the whole
script. The pi/Tanko/kagentz legs do not need the Zulip API key and keep
their verdicts.
Tracked as: zulip-health-credential-placeholder-20260913 (captain-held)
This removes the repeated 'Action required' noise every cycle while
keeping the credential enforcement loud and visible.
Rewrote test to run the monitor with stubbed curl/ssh on PATH that
capture the exact argv of each probe call. The test now asserts the
URL+port of every leg actually requested, not source text or config
constants.
A2: PVE node assertions now check the exact URL in the curl log
(https://192.168.68.9:8006/... must appear), so a wrong IP
(e.g. .99) fails the test.
A3: Grafana/Prometheus/LiteLLM assertions check the URL the call
actually builds, so a hardcoded wrong port in the CALL (while the
config variable stays correct) fails the test.
Mutation evidence:
A2: sed s/192.168.68.9/192.168.68.99/ in PVE_NODES -> suite FAILS
A3: sed s/"$GRAFANA_PORT"/"9999"/ in probe call -> suite FAILS
Results: 18 passed, 0 failed (baseline); 17/18 on each mutation
A1: Test -k assertion now checks use_k:+-k syntax (actual bash pattern)
A2: PVE_NODES assertions now count expected nodes and verify exact array size
A3: Test now asserts liveness behavior (PVE_API_LIVENESS=1) not source text
B1: disk-gc GC schedule corrected: cron runs pbs-gc.sh (not proxmox-backup-manager),
schedule is 20:00 LOCAL (00:00 UTC, not 20:00 UTC), host timezone America/New_York
B2: PROBE SHAPE now documents actual output shape including TLS flag notes
C1: TLS kind is now printed in PVE API failure output
C2: SSH retry logic clarified - retry is in probe_http function (not unreachable)
HOST-RED on media volumes says 'capacity decision — owner to decide' not
'immediate owner attention'. Media volumes are report-only at all levels; the
urgency language was misleading. PBS datastore and host-root get the immediate
attention wording.
Closes the welcome-back proposal: 'how the disk-gc check should classify a
media volume so HOST-RED stops meaning nothing.'
The 20:00 UTC PBS GC cron (proxmox-backup-manager datastore prune)
applies only to /tank/pbs-backup (pbs-datastore). It does NOT touch
media volumes (/media/*) which are report-only at all threat levels.
This clarification prevents the recurring confusion where a 96% media
volume triggers a GC expectation, when the GC schedule never applies
to it.
Closes the 2026-09-17 correction: 'the GC schedule is now 20:00 UTC,
protects the backup datastore, NOT the nearly-full media volume.'
The 2026-09-17 false-verdict incident (third recurrence) showed that prose
policy is not a control: the agent probed :9325/:9405 (nonexistent ports),
CT 116 for PVE API (should be real PVE nodes), and rendered TLS failures as
connection-refused. This moves the canonical probe set into
scripts/infra-monitoring.sh (executed verbatim by the contract) and adds
scripts/test_infra_monitoring.sh which asserts every probed port matches the
documented value.
- scripts/infra-monitoring.sh: one script per contract pattern; all targets,
ports, paths, and expected-status rules in code; -k for PVE self-signed
certs; non-zero exit naming every failed target; no OK summary on failure
- scripts/test_infra_monitoring.sh: 20 assertions covering port drift,
monitoring-host-as-PVE-node, and missing -k flag
- infrastructure-monitoring.prose.md: check-health section now references the
script as executable owner; paste its raw output verbatim
Proof: all 13 legs pass (exit 0); deliberately broken Grafana port (9325)
produces 'probe-failed: 192.168.68.116:9325 (expected 200)' and exit 1.
Implement Rule 15 automated enforcement for MCP servers:
- Validate MCP server URLs against known endpoints (ra-h-os, litellm)
- Check for authentication headers on MCP server configs
- Warn if header values look like env-vars instead of literal keys
- Warn if no auth header is present
This ensures the MCP URL/header invariants from the prose contract
are enforced at the earliest shared boundary (before config
application).
Addressed all 4 ask-user findings from the review:
f1: Qualified the MCP access verification claim - noted that it may
contradict infrastructure-update.prose.md and that LiteLLM version may
have been upgraded since that contract was written.
f2: Added key rotation note documenting that MCP headers use literal keys
and do NOT auto-rotate with the vault. Added TODO to consider adding
MCP header regeneration to the Key Update Procedure.
f3: Added MCP server checks to audit-hermes-config.py (Rule 15):
- Validate MCP server URLs against known endpoints
- Check for authentication headers
- Warn if header values look like env-vars instead of literal keys
f4: Updated Rule 15 verification instruction to include MCP initialize
handshake test, not just /v1/models check.
f5: Added NetBird dependency note documenting that 502 errors on MCP
requests may indicate NetBird outage, not auth failure.
- Documented MCP endpoint verification (2026-08-07): tested with real key,
confirmed initialize handshake works and virtual keys have MCP access
- Added note about Accept header requirement (handled by MCP client library)
- Clarified that the Accept header is NOT part of the config template
- Added litellm MCP server entry to mcp_servers section with correct URL
(https://litellm.sysloggh.net/mcp) and header format
- Updated Rule 15 to be more specific about endpoint validation and
header requirements (REAL keys, not env-var references)
- Added MCP Server Configuration section with implementation details
- Documented the 2026-08-07 Tanko incident where ra-h-os was pointing
to litellm endpoint with env header causing 401 floods
- Updated frontmatter to reflect the changes
Fixes: #keyless-mcp-incident-20260807
Refs: Rule 15 (MCP Endpoint and Header Validation)
Scripts that had hardcoded credentials:
- scripts/zulip-monitor.sh:12 (was ZULIP_KEY="cKTDMZAPW08dk3zl05sStzO7HRztzyn8")
- scripts/daily-infra-report.py:25 (was ZULIP_KEY="cKTDMZAPW08dk3zl05sStzO7HRztzyn8")
Both now read from environment variable ZULIP_API_KEY (set by vault-backed start script)
with loud failure if not present.
Other credentials in scripts/:
- capture-dsh-token.sh: uses TOKEN variable with fallback (not a secret)
- pm2-self-heal.sh: reads TELEGRAM_BOT_TOKEN from /root/.pi/agent/extensions/telegram/.env (acceptable)
- prose-ai-review.sh: uses GITEA_TOKEN from .env file with LITELLM_KEY fallback (not secrets)
No other hardcoded credentials found.
Proof of behavior:
With ZULIP_API_KEY set:
bash scripts/zulip-monitor.sh → Server: HTTP 200 (authenticated)
python3 scripts/daily-infra-report.py --json → Collecting infrastructure data...
Without ZULIP_API_KEY set:
bash scripts/zulip-monitor.sh → "ZULIP_API_KEY not set — refusing to run with no credential"
python3 scripts/daily-infra-report.py --json → "ZULIP_API_KEY not set — refusing to run with no credential"
Cred source: environment variable ZULIP_API_KEY (set by vault-backed start script)
No key rotation (that is a separate decision).
Fix 1: GPU leg degradation is not just SSH probe failure — it also covers
gpu-no-port and gpu-ghost conditions. Rewrite to match check_gpu_ports reality.
Fix 2: Add skipped and partial exemplars for all four legs (LiteLLM keys,
GPU ports, CTs, Vault secrets) so the template covers the rule rather than
only the happy path.
Cosmetic: note that compact form (rtx5070 timeout) is acceptable in summary
line when host is identifiable from context; full probe-failed: <target> <kind>
form required in detail section.
Every report line MUST include one clause per check leg, even when a leg is
skipped or fails. Missing leg must never look the same as healthy leg.
Required legs:
- LiteLLM keys: N/M (names) status
- GPU ports: N/M (rtx3090, rtx5070, strixhalo) status — or SKIPPED (reason)
- CTs: N/M running (names)
- Vault secrets: status
GPU leg is never skipped by configuration; only SSH probe failure causes
degraded status.
The key-count line in litellm-health-check now reports self-describing
output: '18 total (10 on page 1)' instead of bare '18' or '4'. Uses
total_count from the paginated API response and names what was
counted. Previous bare numbers could not reconcile changes between
runs; now a reader sees both the total and the page 1 sample.
root
abiba-bot
mumuni-bot