agentleFS
Sign inStart free
BrendenKennedyBrendenKennedy/claude-for-ai-platforms/.claude/skills/monitoring/SKILL.md

monitoring

Skill1 starsChanged 3 months agoMITView on GitHub

Production model monitoring — PROCESS.md P7 made concrete: a deployed model degrades by default; this makes it visible and actionable. Carries: what to log at inference (sampled inputs + predictions + model/data versions, schema-versioned), input drift detection (PSI / KS per feature; embedding-distance drift for images), concept drift and the delayed-ground-truth loop (proxy metrics until labels arrive), live performance on the P1 metric with slices, reference windows + alert thresholds, retraining triggers (schedule / drift / upstream event), and shadow-eval before a registry alias moves. Load when deploying a model, wiring drift detection, or asking "is the model stale". Triggers: model monitoring, drift, data drift, concept drift, PSI, KS test, production model, model decay, stale model, retrain trigger, shadow deployment, ground truth delay, feedback loop, prediction logging. **Model** decay only — systems telemetry (traces, metrics, logs, OpenTelemetry) is `observability`, SLOs and alerting policy are `reliability-sre`, and agent/LLM quality in production is `agent-evaluation`.

We split this skill into the rules it gives an agent

1 rule. 0 are common, 1 is distinctive, and 0 are very specific to this project.

  • 0

    Common

    Most of the 12 edited versions we read keep these.

  • 1

    Distinctive

    This author's own choices. Worth a look.

  • 0

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Feedback loops: when the model's output influences its future inputs (flagged items get inspected, so confirmed-defect labels skew toward …

    line 68

“Common” is measured: kept by at least 6 in 10 of the 12 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

monitoring — making model decay visible

On-demand: load this at deployment time (P7). Everything upstream still owns its part: the P1 metric is what you monitor, evaluation is how you score it, the registry (tracking-mlflow) is how a replacement ships, and retrain/promote decisions are governed (model-governance + the decision log). This skill owns the loop that notices decay. Tool-agnostic: the concepts map onto evidently/whylogs/custom jobs equally.

Scope boundary — this is model monitoring, not systems observability. Emitting traces, metrics, and logs is observability; SLIs, SLOs, error budgets, and alerting policy are reliability-sre; scoring an agent's output quality in production is agent-evaluation. The two meet at one point: a quality SLI (reliability-sre R1) is fed by the loop this skill describes.

First: log enough to monitor at all

At inference time, durably log (sampled if volume demands): input features (or image embeddings), the prediction + score, model version (registry alias/version), data/schema version, and timestamp — schema-versioned, because the monitoring jobs are consumers of this log and silent schema drift kills them first. No logging → every question below is unanswerable. (What may be logged is governed — PII rules in data-governance/security.)

The three questions, in escalating difficulty

  1. Input drift — are inputs still like training data? Needs no labels. Per-feature PSI or KS statistic against a reference window (the training set or a healthy production week). Rough PSI reading: <0.1 stable · 0.1–0.25 watch · >0.25 investigate. For images: distance between embedding distributions (a frozen feature extractor over samples), plus cheap proxies that catch camera/pipeline breakage — brightness, blur, resolution mix.
  2. Score drift — is the model behaving differently? Prediction/score distribution vs the reference window; a classifier whose positive rate doubles is telling you something no input stat caught.
  3. Performance — is it still right? Needs ground truth, which usually arrives late (returns, QC results, user corrections). Build the join deliberately: predictions matched to eventual labels → the P1 metric, computed on a rolling window, sliced like the eval harness slices (segment, site, class) — aggregates hide exactly the slice that broke. Until labels arrive, 1+2 are the early-warning system, not a substitute.

Thresholds, triggers, and the response

  • Set alert thresholds from the reference window's natural variation (its own week-to-week PSI), not from vibes — otherwise the pager teaches everyone to ignore it.
  • Retraining triggers, written down (the P7 gate demands this): schedule-based (every N weeks), drift-based (sustained PSI/performance breach), or upstream-event (new camera, new product line, patch — the concept-drift events you can see coming). Test the trigger path once before trusting it.
  • A retrained model does not go straight to production. Shadow it (score live traffic, compare, don't act) or canary it; promotion is the registry alias move, gated on the same eval evidence as any release, recorded in the decision log.
  • Name an owner and an alerting path. An unowned dashboard is decoration.

Gotchas

  • Feedback loops: when the model's output influences its future inputs (flagged items get inspected, so confirmed-defect labels skew toward what the model already flags), naive retraining amplifies the bias — sample some ground truth independent of the model's decisions.
  • Seasonality reads as drift — compare against the same season's reference or a rolling year, or every December pages you.
  • The silent failure mode is upstream, not the model: a renamed field, a changed unit, a camera swap. Schema checks + input-range assertions on the inference path catch in minutes what drift stats take weeks to surface.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

Check your setup, free