ml-stack-tracking
Track and diagnose ML runs when a user asks for run identity, metrics, artifacts, dashboards, structured events, NaN or divergence alerts, stalls, privacy, or retrieval; use ledger semantics without c
SKILL.mdskills/ml-stack-tracking/SKILL.md
GitHubRun tracking and diagnostics
Require run ID, project, config, immutable input revisions, metric schema, artifact locations, privacy boundary, alert thresholds, and retention policy.
Workflow
- Create a provider-neutral run record and append config, metric, artifact, status, and error events using repository ledger semantics where available.
- Emit live-dashboard expectations and structured JSON that can be replayed; keep tracking private by default.
- Alert on NaN, divergence, stalls, threshold regressions, missing heartbeats, and budget violations with evidence and severity.
- Bound autonomous loops: no unapproved retries, parameter changes, promotion, publication, or destructive cleanup.
- Hand a retrieval-ready metric summary to
ml-stack-experiment,ml-stack-evaluation, orml-stack-audit; useml_stack.eventsto replay available ledger history.
Boundary and outputs
Do not claim Trackio or another adapter exists without a matching runtime operation. Abstain when event identity, timestamps, revisions, or artifact provenance cannot be reconstructed. Return run schema, event paths, alert diagnosis, operator action, and unknowns. Next owner: ml-stack-experiment for comparison, ml-stack-evaluation for quality, or ml-stack-audit for release evidence.
See run schema, alerts and diagnostics, tracking privacy, and metric handoff.