Admin Ops Console Design

Date: 2026-02-13 Approach: Enhanced Django Admin with Strategic Packages (Approach A) Audience: Solo dev / operator Dependencies: django-object-actions (per-object buttons); everything else is built-in Django

1. Custom Admin Dashboard Index

Replace the default Django admin index with a custom template showing system health at a glance.

Panels:

• Active Incidents — count + severity breakdown (critical/warning/info), links to filtered list
• Pipeline Health (24h) — total runs, success/fail/retrying counts, in-flight count
• Recent Check Runs — last 10 check runs with status badges, grouped by checker name
• Failed Pipelines — last 5 failed pipeline runs with error type and link to detail

Implementation:

• Subclass AdminSite, override index() to inject dashboard-specific query data (avoids adding these queries to every admin page via each_context())
• Override index_template to render the dashboard
• All queries use existing indexed fields (status, created_at, severity)
• Use select_related('incident') where needed

2. End-to-End Pipeline Tracing

Make it trivial to follow an alert through the full pipeline from any entry point.

AlertAdmin:

• Add “Pipeline Trace” readonly field showing: Alert -> Incident -> PipelineRun(s) -> StageExecutions
• Each element is a clickable admin link
• Add trace_id to search fields

IncidentAdmin:

• Enhance existing PipelineRunInline with stage execution count and last error preview per run
• Add “View Full Pipeline” link on each inline row

PipelineRunAdmin:

• Add “Pipeline Flow” readonly field at top of detail page
• Horizontal stage display: INGEST -> CHECK -> ANALYZE -> NOTIFY with colored status indicators
• Uses format_html(), no JS

CheckRunAdmin / AnalysisRunAdmin:

• Add “View Pipeline Run” link field navigating to parent PipelineRun via trace_id

Implementation: All custom readonly fields using format_html() and reverse() URL lookups. trace_id is already indexed everywhere.

3. Operational Actions

Per-Object Actions (django-object-actions)

Model	Action	Type	Behavior
Incident	Acknowledge	Real	Calls incident.acknowledge(), sets acknowledged_at
Incident	Resolve	Real	Calls incident.resolve(), sets resolved_at
Incident	Close	Real	Calls incident.close(), sets closed_at
PipelineRun	Mark for Retry	State-only	Sets status=RETRYING, increments total_attempts
PipelineRun	Mark Failed	State-only	Sets status=FAILED with reason

Bulk Actions (built-in Django)

Model	Action	Behavior
Incident	Acknowledge selected	Bulk acknowledge() on queryset
Incident	Resolve selected	Bulk resolve() on queryset
Alert	Resolve selected alerts	Bulk set status=RESOLVED
PipelineRun	Mark selected for retry	Bulk set status=RETRYING

Safety: Dangerous actions (retry, mark failed) show confirmation intermediate page. State-only actions don’t trigger pipeline execution — they mark records for the retry mechanism to pick up.

4. Aggregation Views

Aggregation panels rendered on the dashboard (no separate pages).

Panel	Query	Display
Pipeline success rate (24h)	PipelineRun grouped by status	“87% success (142/163)” with colored bar
Top failing checkers (7d)	CheckRun WARNING/CRITICAL grouped by checker_name, top 5	Ranked table
Error types (7d)	PipelineRun FAILED grouped by last_error_type, top 5	Ranked table
Intelligence provider usage (7d)	AnalysisRun grouped by provider with Count/Sum(total_tokens)	Table with token counts

Implementation: Computed in AdminSite.each_context(), rendered in dashboard template. Pure ORM aggregation on indexed fields. Styled HTML tables + percentage bar via format_html(). No charting library (Chart.js can be layered on later if needed).

5. Performance & Query Optimization

get_queryset() Overrides

ModelAdmin	select_related	prefetch_related
AlertAdmin	incident	—
IncidentAdmin	—	alerts, pipeline_runs
PipelineRunAdmin	incident	stage_executions
StageExecutionAdmin	pipeline_run	—
AnalysisRunAdmin	incident	—

Dashboard Queries

• Use .only() to fetch minimal columns
• Aggregations use DB-level Count/Sum, not Python-side
• No N+1 queries: every list_display method accessing related objects uses pre-fetched data