Imported from richlander/dotnet-inspect (skills/query/SKILL.md). Install upstream with npx skills add richlander/dotnet-inspect --skill query. Copyright stays with the author.

dotnet-inspect: query and output system

The query system is like Go templates, without a DSL: inspection commands emit structured sections, with the broadest shared query surface on type, member, package, and library. project supports -D and -S but not general field/column projection. find supports -D discovery and field/column projection but not -S selection. diff supports -D and -S but not field/column projection. timeline supports section selection and projection but not -D discovery. workspace supports output formats, --count, and --rows, but not discovery, section selection, or field projection. depends supports -D, -S, categories, row windows, count, and field/column projection across its dependency graph and evidence sections. Positional depends <type> also has a separate complete-service --envelope path described below. Other relationship commands may still expose fixed output. Discover the shape first where available, then select and project.

dnx dotnet-inspect -y -- <command>

Output formats

Default output is Markdown. Pick a machine or compact shape when you need one:

• --table — compact aligned rows.
• --tsv — stable snake_case headers, no embedded tabs/newlines.
• --jsonl — one JSON object per row.
• --json-array — one JSON array for projected rows (--urls, --paths, --value, --print).
• --json — structured documents.
• --bare — one undecorated payload or URL list.
• --count — a bare row count.
• --value / --urls / --paths — project one selected section to scalar, URL, or path payloads.
• --print — print one document behind a selected section row; use --row N|first|last when the section renders multiple rows.
• --tree — a standalone tree for graph sections that support tree lowering.
• --mermaid — a standalone diagram; combine it with --markdown to embed the diagram in a Markdown document.

Positional depends <type> and single-Library API diff support presence-only --envelope. It implies JSON and emits the complete service value with schema_version, result_kind, content, share, and diagnostics. For dependencies, content is semantically identical to the owner-issued camelCase TypeDependencySectionResult selected by unprojected depends <type> --json; whitespace and property order may differ. --compact, --depth, and semantic relationship row selection remain available. Presentation formats, Discover/schema/effective modes, -S, explicit -v, Count, field/column or scalar projection, decoration, and rendered-line clipping are incompatible. For Library API Diff, unprojected --json and envelope content both serialize the complete LibraryApiDiffOutcome using result kind library-api-diff. --all and --compact remain admitted; Type/classification filters, sections, explicit verbosity, row/line controls, and non-API modes are incompatible with envelope output. Explicitly projected Diff JSON retains its presentation schema. Load skill compatibility for outcome and scope details. Asset-mode depends, other commands, and --evidence-envelope remain unadopted.

On find, plain --json retains the typed root result array. Adding --columns or --fields requests projected JSON instead: the result is a JSON document containing the same selected rows and snake_case fields as the --tsv and --jsonl formats.

For member -S "Call Graph", default Markdown is an edge table. Choose the view for the task without changing the graph or its ordered edge rows:

dnx dotnet-inspect -y -- member Type -m Method:1 -S "Call Graph"
dnx dotnet-inspect -y -- member Type -m Method:1 -S "Call Graph" --tree
dnx dotnet-inspect -y -- member Type -m Method:1 -S "Call Graph" --mermaid
dnx dotnet-inspect -y -- member Type -m Method:1 -S "Call Graph" --markdown --mermaid
dnx dotnet-inspect -y -- member Type -m Method:1 -S "Call Graph" --tsv

Use the Markdown table when edge evidence belongs in a document, --tree when call paths are the natural reading order, Mermaid for a diagram, and --tsv/--jsonl for one machine-readable edge row per relationship. from and to are always present; from_group, to_group, and label appear only when the whole graph uses them. A row window can therefore retain an optional field even when its selected values are empty. --tree and standalone --mermaid do not mix with another explicitly selected output format.

Discover and select sections

-D, -S, and -Q are the uppercase cross-command query namespace. Use -D to discover sections and fields, -S to select exact names, categories, compatible aliases, or wildcards, -Q to discover query facets and operators, and --columns/--fields to project values. Discover first instead of guessing names.

dnx dotnet-inspect -y -- member JsonSerializer --platform System.Text.Json -D --tsv
dnx dotnet-inspect -y -- member JsonSerializer --platform System.Text.Json -m Serialize -D "Member Index" --tsv
dnx dotnet-inspect -y -- member JsonSerializer --platform System.Text.Json -m Serialize -S "Member Index" --columns "Selector;Stable;Canonical Signature" --tsv

Structural discovery describes authored membership without running producers; effective discovery probes for data. Package and library differ:

On library, -D --effective runs full probes and remains scoped to base evidence unless a category is named.

@Package groups Package Info, Signals, Statistics, Target Frameworks, Signature, Dependencies, Vulnerabilities, Manifest, Runtime Dependencies, and the unbounded Package files listing. @Files groups the curated nuspec, README, and skill-file sections. The type listing's @Surface category groups API Info, public type-kind and type-forwarder inventories, and Inspection Failures. Member @Member follows the resolved view: member-kind summaries for a type, the matching overload inventory for a member name, or signature and local implementation evidence for one selected overload. Use its domain doors for audit, call, decompiler, performance, source, or SourceLink evidence. Diff @Diff composes Changes, Analysis Diff, and Implementation Diff; select the non-composable Finding Transitions section by exact name. Project @Project composes restored dependency Skills and Package README file inventories. Vocabulary @Vocabulary composes the complete product-owned vocabulary document; use @API or @Decompiler for the corresponding query family. Switches is a section. There are no user-facing @All, @Default, or @Hidden categories.

Library Unsafe Members is intentionally standalone rather than category owned. Select it directly with -S "Unsafe Members"; use -D "Unsafe Members" for its fields or -D --schema to find it in the complete static graph.

Bare -S returns high-value, fixed-length, network-free sections from the package or library base categories. Sections without evidence are omitted. -S --count returns the candidate count map, including zero rows. Explicit sections/categories override base scope and may authorize expensive work. Focused selection omits identity; include Package Info or Library Info when needed.

Some large families expose only their category door in the top-level catalog. Use library X -D @Performance or -D @Metadata; add --effective for populated members. Row formats require a concrete section or homogeneous family. Heterogeneous categories use Markdown/JSON; Performance:* flattens kinds and adds Kind when multiple kinds have rows.

Discover query capabilities

-Q (alias --query-help) is structural and does not acquire or inspect a target. It is available on library, type, member, package, package query, and find. Use it before constructing filters; displayed columns do not imply support for --where, --order-by, or --top.

dnx dotnet-inspect -y -- library -Q
dnx dotnet-inspect -y -- type -Q "Body Shapes"
dnx dotnet-inspect -y -- library -Q "Performance: Arrays" --json
dnx dotnet-inspect -y -- library -Q @Performance

Bare -Q lists query-capable sections. Named -Q lists exact facet keys, operators, comparisons, and values; -v:d adds examples. JSON retains typed arrays of operators and legal values. Named TSV/JSONL output requires one section. --columns, --fields, --rows, and --count shape the discovery rows, not inspected data.

Do not combine -Q with -S, -D, --where, --order-by, or --top. Each named description is a companion section called Query: <Section>; -S "Query: Body Shapes" selects it directly, but normal output and data wildcards omit companions. -D "Query: Body Shapes" describes one companion's columns; companion schema discovery requires one resolved section. A known section with no implemented query bindings says so; for example, find -Q Results does not advertise Package Query terms as API-search predicates.

package query -Q Packages exposes the product-owned Package Query term keys, operators, values, and examples admitted by the CLI:

dnx dotnet-inspect -y -- package query -Q Packages --json
dnx dotnet-inspect -y -- package query Azure.Mcp \
  --where "tool=true"
dnx dotnet-inspect -y -- package query 'Azure.Mcp*' \
  --where "tool-format=v2" --take 20 -n 5 --jsonl
dnx dotnet-inspect -y -- package query 'Polly.*' \
  --where "depends=System.Threading.Tasks.Extensions" \
  --where "dependency-target=netstandard2.0"

--where repeats select product terms, not arbitrary package-field expressions. Independent terms are ANDed; the broad tool=true term identifies the .NET tool package type from manifest evidence. Use tool-format=v1 or tool-format=v2 for settings-based format classification; those specific formats are ORed. Query rows represent individual packages, with exact versions and product-authored evidence. Dependency predicates inspect all nuspec groups by default; use dependency-target=<TFM> to select one compatible group, or dependency-target=all to spell the default explicitly. The query scope all remains distinct from a manifest's any group and does not request traversal. --take bounds candidate work, while -n and --rows select final matched-package rows. Without explicit --take, a simple -n N is pushed into execution: direct package rows use an effective candidate bound of N, while filtered queries scan until N matches or their default candidate bound. Pushdown is capped at 1,000 candidates; larger semantic heads remain valid and are applied after bounded execution. Selecting a package-content term is itself approval for archive acquisition and permits at most 20 candidates; use --nuspec-only to reject such a query. --count observes selected rows and succeeds only when completion or a satisfied finite row selection proves that count exact. Reached candidate bounds and failures remain visible. Package Query does not accept API-search scopes, source overrides, or ranking. Query-execution flags cannot be combined with -Q.

library -Q Integrations describes the ecosystem facet for the whole Integration family. All integrations are enabled by default; use library MyLibrary.dll -S Integrations --where "ecosystem=ecosystem.aspire" to narrow the ordinary result. The initial supported value is ecosystem.aspire. Use a concrete section such as Integration: Aspire for TSV/JSONL. This predicate does not combine with Body Shapes or Performance Triage filters/rankings.

Correlate one member's Findings

Select Finding Census by its exact name for one body-backed method or accessor. It returns one indivisible envelope containing the census receipt, raw Facts, annotated-source document, and document-local fact-to-instance sidecar. The receipt scopes every instance key so display-identical Findings remain distinct.

dnx dotnet-inspect -y -- member JsonSerializer \
  --package System.Text.Json Serialize:1 \
  -S "Finding Census" --json

The section is explicit-only: categories, broad wildcards, and non-exact selectors omit or reject it. Markdown and exact singleton JSON preserve the envelope. Table, TSV, JSONL, count, row-window, field, and column projections fail because they cannot preserve the correlation document.

Query rendered body shapes

At library scope, select exact rendered C# syntax occurrences with the stable IDs from the C# Body Kinds vocabulary. A Kind=... predicate auto-selects the explicit-only Body Shapes section when no -S selection is present:

dnx dotnet-inspect -y -- vocabulary -S "C# Body Kinds"
dnx dotnet-inspect -y -- library MyLib.dll \
  --where "Kind=ObjectCreationExpression" --jsonl
dnx dotnet-inspect -y -- library MyLib.dll \
  --where "Kind=InvocationExpression" \
  --where "Finding=analysis.call-site" \
  --where "Shape=sync-call-in-async" \
  --where "Confidence>=medium" --jsonl
dnx dotnet-inspect -y -- member Widget Render:1 --library MyLib.dll \
  --where "Kind=InvocationExpression" --jsonl
dnx dotnet-inspect -y -- type Widget --library MyLib.dll \
  --where "Kind=InvocationExpression" --jsonl

At library scope, repeated Performance Triage predicates are ANDed before decompilation. The matching opportunities are mapped through their typed source owner identities and only those MethodDef bodies are searched for Kind. Body Shapes is the default occurrence section. Explicitly select -S "Body Shape Summary" for exact Kind/Match groups with a Count column; --columns "Match;Count" hides the already-known kind. Summary windows select groups without reducing their occurrence counts. --count counts the surviving rows in the selected view, and hiding columns never aggregates. Occurrence Member/Token and start/end coordinates locate matches in rendered C# method bodies, not original source files or IL.

Select a Performance section separately when the canonical candidate/evidence/IL rows are also needed. Performance --top and --order-by do not compose with Body Shapes; use --rows to limit rendered matches.

Type scope requires one exact type and searches only its MethodDef and accessor bodies. Member scope requires one exact member name or stable selector and decompiles only the selected MethodDef body. An unambiguous method or single-accessor member is auto-selected; overloaded names require Name:N or Name~digest. A property or event with multiple body accessors requires an accessor selector; use Name~digest:1/Name~digest:2 when the owner is overloaded. Every body query requires exactly one case-sensitive Kind=... predicate. Type and member scope do not yet compose it with Performance Triage predicates.

Filter and order performance rows

On type/member Performance Triage or one concrete library Performance: <Kind> section, use --where with a discovered field name and repeat it to combine predicates. Use --order-by "Field desc,Other asc" before applying output limits.

dnx dotnet-inspect -y -- library MyLib.dll -S "Performance: Arrays" \
  --where "Finding=analysis.allocation" --order-by "RootReach desc" --jsonl

Performance:* orders within each Kind group before flattening, while --rows caps the flattened sequence. Do not combine those flags expecting a global field-ranked prefix; use --top N for the curated global rank, or select one concrete kind when a specific field controls the order.

Limit output

Prefer built-in limits to shell pipes:

• -n N and numeric shorthand like -6 cap output lines on commands that have not adopted semantic rows, like head.
• --tail takes the same count from the end, like tail.
• --rows N takes the first N data rows per table on commands that retain the legacy row window, preserving headings and headers; add --tail for the last N. On adopted semantic-row surfaces, use -n N instead.
• On commands retaining the legacy row window, --rows 2..10 is an absolute 1-based inclusive range (nine rows), 2+10 means ten rows starting at row 2, and 10.. runs from row 10 to the end. These legacy ranges reject --head/--tail, and all legacy --rows forms reject -n.
• --row is not a window. With --print, --value, --urls, or --paths, it selects one displayed row, not a compacted projection position. first/last mean rendered endpoints; missing payloads fail instead of sliding. -n N may still limit the result.
• --count counts rows in one selected table.

find, package query, package --versions / --versions-with-feed, and demo list use semantic rows. -n N selects complete items. Package version listings and demo list also accept -n N --lines to clip rendered lines. --rows on those surfaces accepts only A..B, A.., and ..B; -n and --rows compose as stages in argv order. --head and --tail modify -n, not the range. On package query, --take N separately bounds package work before semantic row selection.