Explain Migration Gaps

Explain migration gaps

Goal: for each panel or widget that did not migrate cleanly, explain why in plain language and give step-by-step guidance to rebuild it manually in Kibana. Read-only — read artifacts a completed migrate run already wrote; do not re-run migration or touch any cluster.

Which panels to explain

Filter panels[] in <output-dir>/dashboards/migration_manifest.json by panels[].status. Non-clean statuses differ by source:

• Grafana: migrated_with_warnings, requires_manual, not_feasible
• Datadog: warning, requires_manual, not_feasible, blocked

Skip panels whose status is clean (migrated on Grafana, ok on Datadog). When the user names a specific panel, match by title, source_panel_id, and dashboard title/id in the same manifest entry.

Exception — clean panels that fail numeric parity: a panel can be clean here yet still FAIL obs-migrate compare. If validate-side-by-side routed you here for a clean (migrated / ok) panel, do not skip it — use the Parity failures (from validate-side-by-side) section below.

Inputs (artifact table)

Assume the user installed the package (obs-migrate on PATH); prefix .venv/bin/ only for a repo checkout. Every artifact below is written by a normal migrate run — no source checkout required.

Workflow

1. Locate the output dir — the --output-dir from the user's migrate run (or ask which run they mean if several exist).
2. Read the manifest — open <output-dir>/dashboards/migration_manifest.json and scan panels[].
3. Filter non-clean panels — keep entries whose status is non-clean for the source (Grafana: migrated_with_warnings, requires_manual, not_feasible; Datadog: warning, requires_manual, not_feasible, blocked).
4. State why each panel failed or degraded — lead with panels[].reasons. Add Grafana panels[].notes or Datadog panels[].warnings / panels[].semantic_losses when they add detail the reasons omit.
5. Map to a Kibana rebuild path — use Grafana panels[].transformation_redesign_tasks[].kibana_alternative and .description when present; otherwise use panels[].recommended_target, panels[].target_candidates, and the matching packet's recommended_target / candidate_targets in verification_packets.json. Cross-check Grafana feature_gap_report.json when the gap is feature-level rather than query-level.
6. Produce guidance matched to status — branch on the panel's status; do not over-promise a tweak path where the engine marked a hard stop:
   • migrated_with_warnings (Grafana) / warning (Datadog) or requires_manual: give concrete finish/rebuild steps — target panel type, ES|QL or Lens sketch from the verification packet / query_ir, post-rebuild checks (fields, time range, group-by). Ground steps in Grafana transformation_redesign_tasks[].kibana_alternative when present, otherwise recommended_target / target_candidates.
   • not_feasible or blocked: do not walk through a step-by-step tweak — explain the redesign constraint (what semantic capability Kibana lacks or must be re-modeled) and why, citing reasons (and Datadog semantic_losses when relevant). Only mention an alternative target if recommended_target or target_candidates genuinely offers one; otherwise say a net-new design is required.
7. Prioritize when many panels need work — use migration_summary.md must-fix ordering; explain the highest-impact panels first unless the user names a specific one.

Parity failures (from validate-side-by-side)

When validate-side-by-side routed a panel here, the manifest status may still be clean (migrated / ok) — the panel translated but obs-migrate compare did not pass numeric parity. Read <output-dir>/dashboards/comparison_report.json (or the sibling .md table with the same columns) and locate the row by dashboard + panel title/id.

1. Read the row — start with verdict, reason, and max_relative_error (when present). A FAIL means bucket values diverged beyond the strict threshold; STRUCTURAL or ERROR on a panel you expected numeric proof means the oracle could not run or only a shape check ran.
2. Rule out data/window/step mismatch first — a FAIL is NOT automatically a translation defect. Re-run obs-migrate compare with --window-minutes and --step-seconds aligned to the source panel's time range and resolution. When live telemetry is sparse or mismatched, use obs-migrate seed-sample-data so both sides read the same synthetic data, compare again, then obs-migrate remove-sample-data --confirm to tear down seeder-owned streams.
3. Only after mismatch is ruled out, treat a persistent FAIL as a real translation defect — explain what the emitted ES|QL computes versus the source PromQL (from the verification packet / manifest query context) and give the same rebuild/redesign path as any other gap: transformation_redesign_tasks, recommended_target, target_candidates, and verification-packet sketches.
4. SKIP and ERROR are not passes — they mean the oracle could not verify the panel (unsupported construct or ES query error). Say explicitly that numeric parity was not proven; do not treat them as green.

Honest limits

• obs-migrate migrate does not populate the richer heuristic review_explanation. Those fields appear only when the migration was run with grafana-migrate --review-explanations (or the equivalent flag on a Grafana migrate invocation). Default migrate output still has reasons, notes, and transformation_redesign_tasks on Grafana — use those first.
• Datadog has no review_explanation or transformation_redesign_tasks equivalent. Datadog panels carry warnings and semantic_losses instead; reason rebuild steps from reasons, warnings, semantic_losses, and target_candidates.
• When rich fields are absent, derive rebuild guidance from panels[].reasons plus recommended_target / target_candidates and the source query in the verification packet — do not invent a migration path the engine did not suggest.
• Never invent a feasible path for a genuinely not_feasible panel. Say it needs a redesign and why (from reasons and semantic-loss notes). blocked on Datadog similarly means the engine could not proceed — treat as full manual rebuild, not a tweak.
• This skill does not prove panels render correctly — empty uploaded panels may be missing telemetry, not a translation bug. Overall coverage counts are report-migration-coverage; numerical proof for panels that did migrate is validate-side-by-side.

See also

• report-migration-coverage skill — shareable coverage summary and manual-effort buckets from summary counts.
• validate-side-by-side skill — prove migrated panels match source numerically.
• docs/command-contract.md — artifact paths and migrate flags for the installed version.