Agent Skills: Metrics Collector - Observability Metrics Instrumentation

Metrics Collector - Observability Metrics Instrumentation

Standardised patterns for collecting, storing, and querying application metrics. Codifies the project's existing database-backed approach (Supabase + PostgreSQL) and defines conventions for metric naming, aggregation, and display. Designed for Vercel/serverless — no Prometheus scraping required.

Description

Codifies database-backed metrics instrumentation for NodeJS-Starter-V1 using Supabase/PostgreSQL, covering standardised metric types (counters, gauges, histograms), naming conventions, time-series aggregation queries, and optional OpenTelemetry export for serverless-compatible observability.

When to Apply

Positive Triggers

• Adding new metrics or KPIs to the application
• Creating dashboard data sources or analytics endpoints
• Instrumenting API routes, agent executions, or background jobs
• Querying time-series metric data for trends and reports
• Setting up alerting thresholds based on metric values
• User mentions: "metrics", "KPI", "instrumentation", "analytics", "monitoring", "dashboard data"

Negative Triggers

• Adding log statements to code (use structured-logging instead)
• Designing dashboard UI components (use dashboard-patterns when available)
• Formatting error responses (use error-taxonomy instead)
• Scheduling periodic metric aggregation (use cron-scheduler instead)

Core Directives

The Three Laws of Metrics

1. Name consistently: Every metric follows {domain}_{entity}_{measurement} convention
2. Store durably: Metrics persist in PostgreSQL, not ephemeral in-memory
3. Aggregate lazily: Store raw events, compute aggregations at query time (or via cron)

Existing Project Metrics Infrastructure

Backend

| Component | Location | Purpose | |-----------|----------|---------| | AgentMetrics class | apps/backend/src/monitoring/agent_metrics.py | Task execution tracking, health reports | | TaskMetrics model | apps/backend/src/monitoring/agent_metrics.py | Per-task metric schema | | AgentHealthReport model | apps/backend/src/monitoring/agent_metrics.py | Agent health aggregation | | Analytics routes | apps/backend/src/api/routes/analytics.py | /metrics/overview, /metrics/agents, /metrics/costs | | Dashboard routes | apps/backend/src/api/routes/agent_dashboard.py | /stats, /list, /{id}/health, /performance/trends | | Health routes | apps/backend/src/api/routes/health.py | /health, /ready |

Database Tables

| Table | Stores | Key Columns | |-------|--------|-------------| | agent_runs | Agent execution records | id, agent_type, status, metadata, started_at, completed_at | | api_usage | LLM API call costs | model, cost_usd, input_tokens, output_tokens, created_at | | tool_usage_events | Tool invocation records | agent_run_id, tool details, timestamps |

Frontend

| Component | Location | Purpose | |-----------|----------|---------| | MetricTile | apps/web/components/status-command-centre/ | Stat tile with trend indicator | | Analytics dashboard | apps/web/app/dashboard-analytics/page.tsx | Overview metrics display | | Proxy route | apps/web/app/api/analytics/metrics/overview/route.ts | Backend proxy |

Metric Types

| Type | Behaviour | Storage | Examples | |------|-----------|---------|----------| | Counter | Monotonically increasing | INSERT per event, COUNT(*) for total | agent_run_total, llm_token_total | | Gauge | Point-in-time value (up or down) | UPSERT — latest value wins | agent_run_active, queue_depth | | Histogram | Value distribution | INSERT per observation, percentile queries | api_request_duration_ms |

All three types are stored in PostgreSQL (no in-memory counters — they vanish between serverless invocations). Counters and histograms go to metrics_events; gauges go to metrics_gauges.

Naming Convention

All metric names follow the pattern:

{domain}_{entity}_{measurement}[_{unit}]

| Segment | Examples | Rules | |---------|----------|-------| | domain | agent, api, cron, auth | snake_case, matches module | | entity | run, request, token, job | singular noun | | measurement | total, duration, rate, size | what is being measured | | unit (optional) | ms, bytes, usd, percent | SI or currency unit |

Standard Metrics

| Metric Name | Type | Labels | Description | |-------------|------|--------|-------------| | agent_run_total | Counter | agent_type, status | Total agent executions | | agent_run_duration_ms | Histogram | agent_type | Execution time | | agent_run_active | Gauge | agent_type | Currently running agents | | api_request_total | Counter | method, route, status_code | HTTP requests | | api_request_duration_ms | Histogram | method, route | Request latency | | llm_token_total | Counter | model, direction | Token usage (input/output) | | llm_cost_usd | Counter | model | LLM API cost | | cron_job_duration_ms | Histogram | job_name | Cron execution time | | cron_job_total | Counter | job_name, status | Cron executions | | auth_login_total | Counter | method, result | Login attempts |

Backend Patterns

MetricsRegistry

A thin wrapper around the existing AgentMetrics pattern, extended with standard metric types. Three methods: increment() (counter), observe() (histogram), set_gauge() (gauge).

from datetime import datetime, UTC
from src.state.supabase import SupabaseStateStore
from src.utils import get_logger

logger = get_logger(__name__)


class MetricsRegistry:
    """Centralised metrics collection and query interface."""

    def __init__(self) -> None:
        self.store = SupabaseStateStore()
        self.client = self.store.client

    async def increment(self, metric: str, value: int = 1, labels: dict[str, str] | None = None) -> None:
        """Increment a counter metric."""
        self.client.table("metrics_events").insert(
            {"metric_name": metric, "metric_type": "counter", "value": value,
             "labels": labels or {}, "recorded_at": datetime.now(UTC).isoformat()}
        ).execute()
        logger.debug("metric_recorded", metric=metric, type="counter", value=value)

    async def observe(self, metric: str, value: float, labels: dict[str, str] | None = None) -> None:
        """Record a histogram observation."""
        self.client.table("metrics_events").insert(
            {"metric_name": metric, "metric_type": "histogram", "value": value,
             "labels": labels or {}, "recorded_at": datetime.now(UTC).isoformat()}
        ).execute()

    async def set_gauge(self, metric: str, value: float, labels: dict[str, str] | None = None) -> None:
        """Set a gauge value (replaces previous)."""
        label_key = str(sorted((labels or {}).items()))
        self.client.table("metrics_gauges").upsert(
            {"metric_name": metric, "label_key": label_key, "labels": labels or {},
             "value": value, "recorded_at": datetime.now(UTC).isoformat()},
            on_conflict="metric_name,label_key",
        ).execute()

# Singleton instance
metrics = MetricsRegistry()

Instrumenting API Routes

Use BaseHTTPMiddleware to record api_request_total (counter) and api_request_duration_ms (histogram) on every request. Label with method, route, status_code. Example:

import time
from starlette.middleware.base import BaseHTTPMiddleware
from src.monitoring.metrics_registry import metrics

class MetricsMiddleware(BaseHTTPMiddleware):
    async def dispatch(self, request, call_next):
        start = time.monotonic()
        response = await call_next(request)
        duration_ms = (time.monotonic() - start) * 1000
        labels = {"method": request.method, "route": request.url.path, "status_code": str(response.status_code)}
        await metrics.increment("api_request_total", labels=labels)
        await metrics.observe("api_request_duration_ms", duration_ms, labels=labels)
        return response

Instrumenting Agent Executions

Extend AgentMetrics.track_task_execution to also record standard metrics:

await metrics.increment("agent_run_total", labels={"agent_type": agent_type, "status": status})
await metrics.observe("agent_run_duration_ms", task_metrics.duration_seconds * 1000, labels={"agent_type": agent_type})

Aggregation Patterns

Time-Series Bucketing

Query metrics grouped by time period for trend charts. Fetch events from metrics_events, group by truncated timestamp (minute/hour/day), and compute count, sum, avg, min, max per bucket. Use PostgreSQL date_trunc via Supabase RPC for server-side efficiency, or bucket in Python for small datasets:

async def get_metric_timeseries(
    self, metric_name: str, bucket: str = "hour", since_hours: int = 24,
) -> list[dict[str, Any]]:
    since = (datetime.now(UTC) - timedelta(hours=since_hours)).isoformat()
    result = self.client.table("metrics_events").select("value, recorded_at").eq(
        "metric_name", metric_name
    ).gte("recorded_at", since).order("recorded_at").execute()

    buckets: dict[str, list[float]] = {}
    for row in (result.data or []):
        ts = datetime.fromisoformat(row["recorded_at"])
        fmt = {"hour": "%Y-%m-%dT%H:00:00Z", "day": "%Y-%m-%dT00:00:00Z"}.get(bucket, "%Y-%m-%dT%H:%M:00Z")
        buckets.setdefault(ts.strftime(fmt), []).append(row["value"])

    return [{"timestamp": ts, "count": len(v), "sum": sum(v), "avg": sum(v) / len(v), "min": min(v), "max": max(v)} for ts, v in buckets.items()]

Percentile Calculation (Histogram Metrics)

For histogram metrics, compute p50/p90/p95/p99 using statistics.quantiles. Query metrics_events filtered by metric_name and time range, sort values, then calculate:

import statistics

async def get_percentiles(self, metric_name: str, since_hours: int = 24) -> dict[str, float]:
    since = (datetime.now(UTC) - timedelta(hours=since_hours)).isoformat()
    result = self.client.table("metrics_events").select("value").eq(
        "metric_name", metric_name
    ).gte("recorded_at", since).execute()
    values = sorted(r["value"] for r in (result.data or []))
    if len(values) < 2:
        return {"p50": 0, "p90": 0, "p95": 0, "p99": 0}
    q = statistics.quantiles(values, n=100)
    return {"p50": q[49], "p90": q[89], "p95": q[94], "p99": q[98]}

Summary Endpoint

Expose a /metrics/summary endpoint returning counters, gauges, and histogram percentiles:

@router.get("/metrics/summary")
async def get_metrics_summary(since_hours: int = Query(24, ge=1, le=720)) -> dict[str, Any]:
    registry = MetricsRegistry()
    return {
        "counters": await registry.get_counter_totals(since_hours),
        "gauges": await registry.get_current_gauges(),
        "histograms": {
            "api_request_duration_ms": await registry.get_percentiles("api_request_duration_ms", since_hours),
            "agent_run_duration_ms": await registry.get_percentiles("agent_run_duration_ms", since_hours),
        },
        "since_hours": since_hours,
        "generated_at": datetime.now(UTC).isoformat(),
    }

Frontend Patterns

Define MetricSummary TypeScript interface mirroring the backend response. Use the existing proxy route pattern (apps/web/app/api/analytics/) to forward requests. Poll at 30-second intervals (matching the existing analytics dashboard) or use Supabase Realtime for gauge updates. Display via the existing MetricTile component from status-command-centre/.

Database Schema

metrics_events Table (Counters + Histograms)

CREATE TABLE IF NOT EXISTS metrics_events (
  id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
  metric_name TEXT NOT NULL,
  metric_type TEXT NOT NULL CHECK (metric_type IN ('counter', 'histogram')),
  value DOUBLE PRECISION NOT NULL,
  labels JSONB NOT NULL DEFAULT '{}',
  recorded_at TIMESTAMPTZ NOT NULL DEFAULT NOW()
);
CREATE INDEX idx_metrics_events_name_time ON metrics_events (metric_name, recorded_at DESC);
CREATE INDEX idx_metrics_events_labels ON metrics_events USING GIN (labels);

metrics_gauges Table

CREATE TABLE IF NOT EXISTS metrics_gauges (
  metric_name TEXT NOT NULL,
  label_key TEXT NOT NULL DEFAULT '',
  labels JSONB NOT NULL DEFAULT '{}',
  value DOUBLE PRECISION NOT NULL,
  recorded_at TIMESTAMPTZ NOT NULL DEFAULT NOW(),
  PRIMARY KEY (metric_name, label_key)
);

Retention Policy

Schedule a cron job (/api/cron/metrics-cleanup) to DELETE FROM metrics_events WHERE recorded_at < NOW() - INTERVAL '90 days'.

OpenTelemetry Bridge (Optional)

For production with full observability infrastructure (Grafana, Datadog), optionally export metrics to an OTel collector by creating MeterProvider instruments that mirror the database metric names. Only enable when OTEL_EXPORTER_OTLP_ENDPOINT is configured. The database-backed approach works standalone for Vercel/serverless.

Anti-Patterns

| Anti-Pattern | Why It Fails | Correct Approach | |---|---|---| | In-memory counters on serverless | Lost between invocations, no persistence | Database-backed metrics | | Prometheus scrape endpoint on Vercel | No persistent process to scrape | Database storage + query endpoints | | Logging metrics as unstructured strings | Cannot aggregate or query | MetricsRegistry with typed methods | | Recording per-request without labels | Cannot filter by route or status | Always include method, route, status_code labels | | Querying raw events for dashboards | Slow on large datasets | Pre-aggregate via cron or use time-bucket queries | | Unbounded metrics_events growth | Storage costs, slow queries | 90-day retention cron job |

Checklist for New Metrics

Definition

• [ ] Metric name follows {domain}_{entity}_{measurement}[_{unit}] convention
• [ ] Type chosen correctly: counter (cumulative), gauge (current), histogram (distribution)
• [ ] Labels defined (minimum context for filtering)
• [ ] Added to Standard Metrics table in this document

Implementation

• [ ] Uses MetricsRegistry singleton (metrics.increment, metrics.observe, metrics.set_gauge)
• [ ] Integrated with structured-logging (debug-level metric recording logs)
• [ ] Error handling — metric recording failures must not break the request

Query

• [ ] Time-series aggregation tested for the new metric
• [ ] Summary endpoint includes the new metric (if high-value)
• [ ] Frontend display considered (MetricTile or chart)

Maintenance

• [ ] Covered by 90-day retention policy (or custom retention if needed)
• [ ] Documentation updated in Standard Metrics table

Response Format

[AGENT_ACTIVATED]: Metrics Collector
[PHASE]: {Design | Implementation | Review}
[STATUS]: {in_progress | complete}

{metrics analysis or implementation guidance}

[NEXT_ACTION]: {what to do next}

Integration Points

Structured Logging

• MetricsRegistry emits metric_recorded debug-level logs
• Shares correlation_id from request context for tracing

Error Taxonomy

• Metric recording failures: SYS_RUNTIME_METRICS (500) — logged but never thrown to caller
• Invalid metric names: DATA_VALIDATION_METRIC_NAME (422)

Cron Scheduler

• metrics-cleanup cron job deletes events older than 90 days
• Cron job metrics (cron_job_total, cron_job_duration_ms) recorded by cron handlers

API Contract

• /metrics/summary endpoint included in OpenAPI docs
• MetricSummary response model typed with Pydantic (backend) and TypeScript interface (frontend)

Dashboard Patterns

• MetricsRegistry provides the data layer for dashboard visualisations
• Time-series aggregation powers trend charts
• Gauge values power real-time status displays

Australian Localisation (en-AU)

• Timezone: AEST/AEDT — recorded_at stored as UTC, converted in display
• Currency: llm_cost_usd stored in USD (API billing currency); convert to AUD for display
• Spelling: serialise, analyse, optimise, behaviour, colour
• Date Format: ISO 8601 in storage; DD/MM/YYYY in UI display