Orbit (GitLab Knowledge Graph) skill

Query the GitLab Knowledge Graph (product name Orbit) via the typed glab orbit remote CLI subcommands (shipped in glab v1.94.0+). The typed CLI handles the Content-Type header, response framing, and exit codes for you — always go through glab orbit remote.

Prerequisites

If glab orbit commands fail (command not found, auth errors, feature-flag exit codes), work through the first-run setup checklist in references/prerequisites.md.

Discovery

glab orbit remote --help and glab orbit remote query --help are the authoritative usage references. For entity properties, prefer the recipes in references/recipes.md over schema introspection — they already encode the columns and filters known to work.

If you must introspect, call glab orbit remote schema <Entity…> with explicit entity names — always pass the entity names you need rather than the unscoped form, which returns ~17 KB+ of output. Call schema at most once per session; schemas don't change mid-session. Use glab orbit remote dsl for the full DSL JSON Schema. Note that per-node outgoing_edges/incoming_edges are arrays of strings (edge type names), not objects — use --jq accordingly (e.g. schema Project --jq '.nodes[] | select(.name=="Project") | .properties').

Each glab orbit remote query has fixed per-call overhead. Prefer one aggregation query over N traversal queries for "how many X grouped by Y", and batch related lookups.

When editing Orbit docs or skills, fence executable query JSON as json orbit-query so docs smoke tests run it.

Running a query

Write the request body to a file and pass it to glab orbit remote query. Default output is llm (compact, agent-friendly); pass --format raw to pipe into jq. Endpoints are user-scoped — do not pass -R owner/repo.

Put the request body in /tmp/q.json:

{
  "query": {
    "query_type": "traversal",
    "nodes": [
      {"id": "p",  "entity": "Project",
       "filters": {"id": {"eq": 278964}}},
      {"id": "mr", "entity": "MergeRequest",
       "columns": ["iid", "title", "state"]}
    ],
    "relationships": [
      {"type": "IN_PROJECT", "from": "mr", "to": "p"}
    ],
    "order_by": "-mr.created_at",
    "limit": 5
  }
}

glab orbit remote query /tmp/q.json

filters is an object keyed by property name — not an array. Use either shorthand equality ({"state": "opened"}) or the operator form ({"iid": {"eq": 1216}}). Operators: eq, gt, lt, gte, lte, in, contains, starts_with, ends_with, is_null, is_not_null, plus text-token operators (token_match, all_tokens, any_tokens) for text-indexed properties — see query_language.md.

All queries declare node selectors in the nodes array — a 1-element array for neighbors and single-node traversal; multi-node traversal, aggregation, and path_finding add relationships.

• For multi-hop traversal edges, set relationships[].hops to an inclusive [min, max] pair ("hops": [1, 3]; [2, 2] for exactly 2). Omitted means [1, 1]. Max 3.
• For path_finding queries, set path.max_depth inside the required path sub-object. Max 3. hops does not apply to path_finding. When endpoints use filters, include path.rel_types to bound fan-out; path_finding follows edges only in their schema direction (see recipe pitfall).

Common pitfalls

Read references/recipes.md before constructing a query — the same question often has one canonical paste-ready shape and several wrong-looking-correct ones. Four traps recur:

• "Pipelines for a merge request" requires Pipeline.source = "merge_request_event". Both Pipeline.merge_request_id and the MergeRequest --TRIGGERED--> Pipeline edge return parent and downstream child pipelines (source = "parent_pipeline"). Apply the source = "merge_request_event" filter (or the canonical recipe) to match the MR Pipelines tab.
• Prefer single-node queries when you can bound the target entity directly. Adding nodes/relationships only to "anchor" a query (joining Project + MergeRequest + Pipeline when you already know merge_request_id) can change the row shape and skew aggregation counts. If recipes.md shows a single-node form, use it.
• HAS_LATEST_DIFF vs HAS_DIFF for file history. HAS_LATEST_DIFF points only at the most recent diff snapshot of an MR. "Every MR that ever touched this file" needs HAS_DIFF (all snapshots) — HAS_LATEST_DIFF here can substantially undercount long-lived files. See recipe.
• GitLab issues, epics, tasks, and incidents are the WorkItem entity, not Issue. Modern GitLab unifies these under work items, and Orbit follows the same model: there is no Issue node, so entity: "Issue" is rejected. Query WorkItem for any of them (see recipe).

Iteration budget

A single user question should resolve in at most 5 query attempts. Tweaking only limit/columns is not progress; changing entity, relationship type, or a filter is. Validation errors (HTTP 400) count toward the budget. If you exceed 5 without converging, give up loudly: report the shapes you tried, what failed, and the next step — do not keep iterating or inflate a partial answer. Full rules: references/troubleshooting.md.

Reporting results

Orbit answers are graph queries against ClickHouse, not an authoritative source of truth. Always surface known coverage gaps inline (e.g. HAS_LATEST_DIFF vs HAS_DIFF, time-bounded aggregates) and show the query body so the user can audit it. Do not add a "Methodology" header that implies rigor the data lacks. Full guidance and worked examples: references/reporting.md.

Repository map helpers

For code-structure orientation before planning a change, use a repo-map helper: the native local command glab orbit local repo-map for an uncommitted/branch-local checkout, or the bundled remote helper script (path relative to this skill root, not the user's current repo) for a project already indexed in Orbit Remote. See the repository-map rows in References below.

Local CLI (glab orbit local)

glab orbit local downloads and runs a managed Orbit CLI binary for indexing and querying a local copy of the Knowledge Graph (macOS/Linux only, x86_64/aarch64). Prefer it over glab orbit remote when indexing a local repository for offline analysis; use remote to query production. Install/run with glab orbit local (add --install or --update). Full config keys and pass-through args: references/local_cli.md.

References