- Add validateAIFindings() that cross-checks AI findings against actual metrics
- Filter out low-confidence findings (CPU <50%, memory <60%, disk <70%)
- Always allow critical findings, backup issues, and reliability findings through
- Update AI system prompt with stricter thresholds and explicit noise examples
- Add 'before creating a finding' checklist for AI (the 3am test)
- Update AI.md docs with clear value proposition and expectations
- Add comprehensive tests for the validation layer
This ensures paying users get immediate value without noise.
PBS datastores are now displayed in the Storage overview alongside PVE
storage. Each PBS datastore is converted to a Storage entry with:
- type: 'pbs'
- content: 'backup'
- shared: true
- active: based on PBS instance status
This provides a complete picture of all storage resources in one view
while keeping detailed PBS info in the dedicated PBS section.
Closes#869
- Fix PVE nodes: buildNodeUrl in ProxmoxNodesSection.tsx now prioritizes
guestURL over host (was ignoring guestURL entirely)
- Add PBS support: GuestURL field added to PBSInstance config, model,
and API handlers
- Add PMG support: GuestURL field added to PMGInstance config, model,
and API handlers
- Update NodeSummaryTable to use guestURL for PBS nodes
- Frontend types updated for PBS/PMG guestURL support
The Guest URL setting in node configuration now works correctly across
all node types. When set, it takes priority over the Host URL when
clicking on node names to navigate to the Proxmox/PBS/PMG web UI.
Closes#870
- Add IsMockResource() helper to detect mock data resources by naming patterns
- Filter out heuristic findings from mock resources when PULSE_MOCK_MODE is off
- Mock findings still appear when demo mode is enabled
- Ensures users don't see mock data findings in production
The AI demo/mock findings were using a non-existent MOCK_ENABLED
env var. Changed to PULSE_MOCK_MODE which is the actual env var
used by the mock data system.
The staging images in preflight are intentionally amd64-only for speed,
but the publish workflow was just copying them instead of building
multi-arch. Now builds linux/amd64,linux/arm64 from source at publish.
Related to #868
- Add AnalyzeKubernetes to PatrolConfig and enable by default
- Add analyzeKubernetesCluster() for K8s heuristic analysis
- Detect offline clusters, nodes not ready, CrashLoopBackOff pods
- Detect failed/pending pods and unavailable deployments
- Add K8s clusters to LLM context in buildInfrastructureSummary()
- Add kubernetes_cluster as valid resource type in AI prompt
- Enhance analyzeDockerHost() for Docker/Podman
- Podman-aware messaging based on runtime
- Add unhealthy container detection (health checks)
- Add exited container with error detection
- Add high CPU detection (>90%)
- Add stale host detection (10+ minutes)
- Tiered severity for restarts (>10 = critical)
- Add 20 new tests for K8s and Docker analysis
New commands:
pulse mock enable - Enable mock mode
pulse mock disable - Disable mock mode
pulse mock status - Show current status
Makes it easy to toggle between mock and real data without
manually editing config files.
When MOCK_ENABLED=true, Pulse now injects realistic AI patrol
findings to showcase the AI features without requiring actual
LLM API calls. This enables the demo instance to demonstrate:
- Critical/warning/info findings with realistic content
- Patrol run history
- Actionable recommendations
Also includes refinements to dismissal logic from earlier work:
- Only 'not_an_issue' creates permanent suppression
- 'expected_behavior' and 'will_fix_later' just acknowledge
When a Docker host has a custom URL configured (e.g., Portainer link),
the host name in the container table's group headers is now a clickable
link that opens the URL in a new tab. A link icon appears next to the
host name to indicate it's clickable.
Related to #860
1. Fixed TestNewConfigPersistenceFailsWhenEncryptedDataPresentWithoutKey
- Test was picking up real encryption key from /etc/pulse during migration
- Now temporarily moves system key during test for proper isolation
- Uses t.Cleanup to ensure key is restored even on failure
2. Cleaned up console.log statements in production code
- Dashboard.tsx: replaced console.log with logger.debug for metadata events
- CompleteStep.tsx: removed verbose agent detection debug logs
These changes reduce log noise in production while maintaining debug
capability in development mode.
Addresses #866 - agents were logging 'WebSocket connection failed' warnings
even during normal reconnection scenarios (server restart, network blip, etc).
Changes:
- Normal close errors (1000, 1001, connection reset) now log at Debug level
- Only log Warning after 3+ consecutive failures
- Changed 'Connecting to Pulse' from Info to Debug to reduce noise
- Successful connections still log at Info level
The WebSocket is only used for AI command execution, not metrics, so
transient disconnections don't affect monitoring functionality.
User feedback fields (DismissedReason, UserNote, TimesRaised, Suppressed, Source)
were not being saved to disk, causing 'expected behavior' dismissals to be lost
after Pulse restarted.
- Add missing fields to AIFindingRecord in persistence.go
- Update FindingsPersistenceAdapter to save/load these fields
- Add comprehensive tests for dismissal persistence round-trip
Fixes issue where Frigate storage warning kept reappearing despite being
marked as expected behavior.
Since watch/info findings are filtered from the UI and never shown
to users, don't include them in the patrol run status summary.
This makes the summary consistent with what users actually see.
The LLM was confusing VMIDs because they weren't included in the
context. Now the formatted context shows:
### Container: ollama (VMID 200) on minipc
This prevents the AI from referencing the wrong VMID when generating
findings and recommendations.
When the service restarts, it now checks if a patrol ran within the
last hour. If so, it skips the initial patrol to avoid wasting API
tokens during development/maintenance when the service is restarted
frequently.
The scheduled patrol runs (every 6 hours) are not affected.
100 samples was causing 326k+ input tokens which is expensive.
24 samples (hourly resolution) still provides good pattern visibility
while significantly reducing token cost.
Estimated reduction: ~75% fewer metric tokens.
When AI patrol fails due to API issues like insufficient balance, invalid
API key, or rate limiting, we now create a finding that appears in the
AI Insights tab. This makes the issue visible to users rather than hidden
in logs.
The finding includes:
- Clear description of the issue (e.g., 'Insufficient API credits')
- Recommendation for how to fix it
- Evidence showing the actual error message
When a patrol run encounters errors (e.g., LLM call failed), don't
display 'All healthy' in the summary as that's misleading - the
analysis didn't complete properly.
Now shows 'Analysis incomplete (N errors)' instead, which correctly
explains why the status badge shows red/error.
Modern LLMs have 100k+ token contexts. 100 samples over 24h gives
~15 minute resolution while adding minimal token overhead.
This lets the LLM see fine-grained patterns, short spikes, and
accurately distinguish anomalies from normal behavior.
The in-memory MetricsHistory only retains 24 hours of data, not 7 days.
Changed computeGuestMetricSamples to use trendWindow24h instead of
trendWindow7d, and reduced sample count from 24 to 12 points.
This ensures the LLM actually receives metric samples in the context,
which wasn't happening before because the 7-day query returned empty data.
Shows a purple '⚡ Alert' badge on findings that were discovered through
alert-triggered analysis rather than scheduled patrol runs. This gives
users visibility into how findings were discovered without cluttering
the patrol run history table.
Bug Fixes:
- Fix boolean fields with 'omitempty' not persisting false values
- AlertTriggeredAnalysis, PatrolAnalyzeNodes/Guests/Docker/Storage
- omitempty causes Go to skip false (zero value) when marshaling JSON
- On reload, NewDefaultAIConfig() sets true, and missing field stays true
- Fix model dropdown losing selection after save (SolidJS reactivity issue)
- Added explicit 'selected' attribute to option elements
- Ensures browser maintains selection with optgroups during re-renders
Improvements:
- Change patrol type label from 'Quick' to 'Patrol' in history table
- Add chat_model and patrol_model to AI settings update log
- Add alert_triggered_analysis to AI config load log for debugging
Instead of relying on pre-computed trend heuristics (which can be misleading
for edge cases like step changes vs continuous growth), we now pass downsampled
raw data points to the LLM so it can interpret patterns directly.
Changes:
- Add MetricSamples field to ResourceContext
- Add DownsampleMetrics() to reduce data points for LLM consumption
- Add formatMetricSamples() to format data compactly (e.g., 'Disk: 26→26→31%')
- Add computeGuestMetricSamples() to gather 7-day sampled history
- Populate MetricSamples for VMs and containers during context build
- Add History section to formatted context output
The LLM now sees actual patterns like 'stable for 6 days then jumped' rather
than just '45.8%/day growth rate' - allowing for much more nuanced interpretation.
This approach:
- Leverages LLM's pattern recognition instead of hard-coded heuristics
- Provides 7 days of data (~24 samples) for context on normal behavior
- Uses minimal tokens due to compact formatting with deduplication
- Is more future-proof as LLMs improve
Example output:
**History (7d sampled, oldest→newest)**: Disk: 26→26→26→26→26→31%
Refs: Frigate disk usage false positive investigation
- Remove unused correlations state and constants from AIOverviewTable
- Remove unused runbook-related imports, state, and functions from Alerts
- Add type annotation to Set() to fix type error
- Removes dead code left over from runbook UI removal
Filter out 'watch' and 'info' severity findings from the API response.
These lower-severity findings were mostly noise:
- 'watch': CPU is 35% instead of 11% (who cares)
- 'info': Stopped container exists (knew that)
Now only showing actionable findings:
- critical: Something is broken NOW
- warning: Something needs attention soon
Users prefer silence to noise.
Fix Receipts was showing 'No fixes logged' most of the time since:
- Runbooks were removed
- Remediation logging was inconsistent
Just adds visual clutter without value. Removed ~100 lines of UI code.
Runbooks were a half-built feature that provided no value:
- Only 3 runbooks existed
- AI dynamic remediation already covers the same ground
- Added UI complexity without benefit
Removed:
- runbooks.go and runbooks_test.go
- Handler functions in ai_handlers.go
- Routes in router.go
- Test cases in ai_handlers_test.go
- Auto-fix call in patrol.go
Kept (dead code but harmless):
- Frontend types/API calls (will 404)
- RecordIncidentRunbook function (unused)
Less code = easier to maintain.
Updated LLM prompt with explicit guidance on what NOT to report:
- Small baseline deviations (7% vs 4% is normal variance)
- Low utilization (under 50% CPU or 60% memory is fine)
- Stopped containers that aren't autostart
- 'Elevated' metrics still well under limits
Severity guidelines made more specific:
- CRITICAL: disk >95%, service down, data loss
- WARNING: disk >85%, memory >90%, failures
- WATCH: Only for trends projected to hit critical in <7 days
- INFO: Context/observations
Key message to LLM: 'Users prefer silence to noise'
Only flag things that require operator action.
Smarter anomaly detection to reduce false positives:
**Learning Window:** 7 days → 14 days
- Captures weekly patterns (weekday vs weekend)
**Metric-Specific Thresholds:**
CPU:
- Only report if usage >70% AND >2x baseline
- Low CPU variance (5% vs 10%) is not actionable
Memory:
- Report if >80% OR (>1.5x baseline AND >60%)
- Memory is more stable, lower threshold makes sense
Disk:
- Report if >85% usage OR +15 percentage points growth
- Disk problems are critical, use absolute thresholds
Other metrics:
- Use 2x threshold as default
This dramatically reduces 'noise' anomalies while catching
actual problems that need operator attention.
The AI Intelligence Summary was adding noise rather than value:
- Predictions duplicated patrol findings
- Correlations were not actionable
- 'Fixed' items were vague diagnostics
- Status changes were startup noise
The real value is in the patrol findings section which shows:
- Actual issues found (critical/warning/watch/info)
- Actionable recommendations
- Suppression rules
Keeping the patrol findings, removing the redundant summary.
More aggressive noise filtering:
1. Anomaly threshold raised from 1.5x to 2x
- 1.5x is too borderline to be actionable
- Now requires genuinely significant deviation
2. Filter out 'Ran diagnostic' and 'Executed command' fallback items
- These are generic summaries that provide no value
- Only show remediations with specific, meaningful descriptions
Goal: If something shows in AI Intelligence, it should demand attention.
rcourtman