Infrahub Importing Data

CSV/TSV Import to Infrahub Objects

Overview

Expert guidance for turning user-provided CSV or TSV inputs — a single file, a folder of files, or any combination — into Infrahub object YAML that infrahubctl object load accepts. The skill is the bridge between "I have a spreadsheet" and "the objects exist on a reviewable branch."

The skill is strictly a consumer of the live schema. If a CSV column has no schema home, it stops, lists the offending columns, and routes the user to infrahub-managing-schemas to make the schema decision separately. It never proposes attribute additions, dropdown choices, or any other schema edit, and it never writes directly to the default branch.

Project Context

Existing schema files: !find . -name "*.yml" -path "*/schemas/*" -o -name "*schema*" -name "*.yml" 2>/dev/null | head -10

Existing object files: !find . -name "*.yml" -path "*/objects/*" 2>/dev/null | head -10

.infrahub.yml if present: !find . -maxdepth 2 -name ".infrahub.yml" 2>/dev/null | head -1

Candidate CSV/TSV inputs in the working tree: !find . -maxdepth 3 \( -name "*.csv" -o -name "*.tsv" \) 2>/dev/null | head -10

If invoked with arguments (e.g., /infrahub:importing-data inventory.csv), treat the first argument as the input file or folder; treat remaining arguments as additional paths to include.

When to Use

• A user hands you one or more CSV or TSV files and wants the data in Infrahub.
• The input is denormalized — a single sheet conflating multiple kinds — and you need to split it across the right schema nodes with correct load order.
• The CSV has columns that look like dropdown labels (e.g., Status: Active) and you need to emit the choice name (e.g., active).
• A parent kind's rows are repeated per child in the CSV, and the children belong inline as component children of the parent.
• Interface-shaped names (eth0..eth47) appear in sequence and you want range-collapsed emission.
• The user wants every imported value stamped with a lineage tag.

When NOT to Use

• The input is an LDJSON dump from infrahubctl export dump. Use infrahubctl import load; it reads a different format and is the wrong tool for CSV. The skills disambiguate so you don't conflate them.
• The user needs schema changes. This skill is read-only against the schema. Hand off to infrahub-managing-schemas.
• The user needs to export data from Infrahub to CSV. Single-kind export already works via infrahubctl object get -o csv; a future companion skill will cover multi-kind export.
• The input is JSON, XLSX, Parquet, or any other non-CSV format. v1 covers CSV + TSV only.

Rule Categories

Schema Features This Skill Depends On

The skill emits object YAML that conforms to the target schema's actual shape. If the schema is missing a feature the CSV requires, the load fails — which is why the skill checks these upfront and escalates rather than silently working around them.

If any of these is missing for the data the user wants to load, that's a schema migration, not an import fix. The skill stops and says so.

Workflow

Follow these 14 steps in order. The numbering is load-bearing — earlier steps gate later ones.

1. Normalize input to a flat file list. Accept a single path, a directory (recursive scan for *.csv / *.tsv), or a list, and resolve to a flat list before profiling. For each file, detect the CSV dialect (delimiter, encoding, BOM, line endings) before reading rows. Read rules/inputs-file-folder-list.md and rules/inputs-csv-dialect.md.

2. Discover the schema — read-only. Try sources in priority order and stop at the first one that returns the kinds the CSV needs:

   1. MCP server — the same surface infrahub-analyzing-data uses; sees deployed state including branch-specific dropdown choices and HFIDs.
   2. infrahubctl schema export --branch <name> — CLI export against the configured server when MCP is not connected.
   3. /api/schema?branch=<name> REST endpoint — direct HTTP fetch when infrahubctl is unavailable or pointing at a different server.
   4. Local schemas/*.yml in the repo — last resort; may lag the deployed state, so the step 11 server validate is required to catch divergence.

   Record which source you used so the user knows how authoritative the mapping is. Never propose a schema edit. Read rules/workflow-introspect-first.md.

3. Profile each file. Read header + sample rows (default 20; more if needed for dropdown, range, or denormalization detection). Note column data shapes, per-column distinct value counts, sample values, and sequence patterns. Read rules/workflow-profile-sample.md.

4. Build a mapping draft per file using the heuristics in reference.md:

   • Column → attribute: exact name match, then snake_case round-trip, then display-label fuzzy match, then unit-strip ((GB), (MHz)) — anything past that defers to the interview. Read rules/mapping-column-to-attribute.md.
   • Value coercion: Boolean / Number / DateTime / JSON attribute kinds require explicit type coercion before emission. Read rules/mapping-value-coercion.md.
   • Empty cells: decide omit vs fail by the schema's optional and default_value declarations. Read rules/mapping-empty-and-null.md.
   • Dropdown columns: build a label→name lookup from the schema and translate. Read rules/mapping-dropdown-label-to-name.md.
   • Reference columns: detect by name match against a relationship plus value shape against the target kind's HFID. Read rules/mapping-reference-and-hfid.md.
   • Range columns: collapse contiguous interface-style sequences. Read rules/range-detection.md.
   • Denormalized columns: detect repeated parent groups and split the input across kinds. Read rules/decomposition-denormalized-csv.md.
   • Same kind in multiple inputs: merge into one numbered output file, dedupe by HFID, surface conflicts in the interview. Read rules/inputs-merge-same-kind.md.

5. Fail-closed gate. If any column has no schema home, stop. Emit a structured report listing the unmapped columns and the kinds you checked, point the user at infrahub-managing-schemas, and exit. No partial writes. Read rules/workflow-fail-closed-on-unmapped-columns.md.

6. Up-front interview. Batch every remaining ambiguity into one round of multi-choice questions before any file is written. Read rules/workflow-up-front-interview.md.

7. Confirm and lock the mapping. Echo back the complete plan (target files, kinds, column bindings, branch name, lineage opt-in) before the first file is written.

8. Emit object YAML to a local working directory (default ./output_dir/). Each file conforms to the managing-objects envelope: apiVersion: infrahub.app/v1, kind: Object, spec.kind, spec.data. Files are numbered for load order (NN_<kind-plural>.yml). Each file carries a leading provenance comment naming the source CSV(s) + sha256 + emission timestamp. Cross-reference the format rules in ../infrahub-managing-objects/rules/format-structure.md, the file naming convention in ../infrahub-managing-objects/rules/organization-load-order.md, and the comment shape in rules/outputs-provenance-comment.md.

9. Self-check against managing-objects. Re-read the listed managing-objects rules and walk the emission against each one. Local only — no CLI, no server, no branch. Fix in place before moving on. Read rules/workflow-self-check-against-managing-objects.md.

10. Pre-flight reference closure. Walk every relationship reference in the emission and verify the target resolves either to a row in an upstream file or to a live object the introspection captured. Orphan references fail closed before any branch is touched. Read rules/workflow-pre-flight-closure.md.

11. Create the branch. Only after the self-check and pre-flight closure pass, run infrahubctl branch create <name> (default csv-import-YYYYMMDD-HHMM, user-overridable in the interview). Doing this last keeps throwaway branches out of the branch list. Read rules/workflow-branch-before-load.md.

12. Server validate on the branch: infrahubctl object validate ./output_dir/ --branch <name>. Catches schema-resolution errors the local checks can't (kind not in schema, attribute name miss, reference target missing). On error, translate the diagnostic back to the source CSV cell and ask the user to fix the input — don't hand-edit the emitted YAML. Read rules/workflow-validate-error-mapping.md.

13. Load on the branch: infrahubctl object load ./output_dir/ --branch <name>. Object load is not transactional across files — if file 17 of 20 fails, files 1–16 are already on the branch. The branch-first design means you discard the branch and re-run with a fresh branch name; you never clean up partial state by hand.

14. Hand off. Tell the user the branch is ready for review, the validate/load commands that were run, and how to open a proposed change in the UI. Don't auto-merge.

Supporting References

• examples.md — 8 worked CSV patterns: flat list, denormalized split, parent-with-children, dropdown label normalization, range collapse, lineage stamping, fail-closed report, merge + dedup
• reference.md — heuristic order, file envelope reminder, HFID decision matrix, file naming convention
• rules/ — individual rules by category prefix; start at rules/_sections.md
• ../infrahub-managing-objects/SKILL.md — the consumer-side rules every emitted file must conform to (envelope, value mapping, components, range, load order, branch-first)
• ../infrahub-analyzing-data/SKILL.md — MCP introspection pattern for live schema discovery
• ../infrahub-managing-schemas/SKILL.md — escape hatch when a column has no schema home
• ../infrahub-common/metadata-lineage.md — value metadata semantics (source is lineage only; locking needs owner + is_protected)
• ../infrahub-common/rules/workflow-branch-for-crud.md — shared branch-first rule that both schema and object writes inherit from
• ../infrahub-common/rules/connectivity-server-check.md — verify the server is reachable with infrahubctl info before any server-dependent command in the workflow
• ../infrahub-common/rules/connectivity-python-environment.md — detect the project's Python env prefix (uv run / poetry run) for all infrahubctl invocations
• ../infrahub-common/rules/workflow-information-priority.md -- Skill content first; how to consult docs.infrahub.app on a genuine gap (e.g. deleting nodes)