---
headline: jq Manual (development version)

body: |

  A jq program is a "filter": it takes an input, and produces an
  output. There are a lot of builtin filters for extracting a
  particular field of an object, or converting a number to a string,
  or various other standard tasks.

  Filters can be combined in various ways - you can pipe the output of
  one filter into another filter, or collect the output of a filter
  into an array.

  Some filters produce multiple results, for instance there's one that
  produces all the elements of its input array. Piping that filter
  into a second runs the second filter for each element of the
  array. Generally, things that would be done with loops and iteration
  in other languages are just done by gluing filters together in jq.

  It's important to remember that every filter has an input and an
  output. Even literals like "hello" or 42 are filters - they take an
  input but always produce the same literal as output. Operations that
  combine two filters, like addition, generally feed the same input to
  both and combine the results. So, you can implement an averaging
  filter as `add / length` - feeding the input array both to the `add`
  filter and the `length` filter and then performing the division.

  But that's getting ahead of ourselves. :) Let's start with something
  simpler:

manpage_intro: |
  jq(1) -- Command-line JSON processor
  ====================================

  ## SYNOPSIS

  `jq` [<options>...] <filter> [<files>...]

  `jq` can transform JSON in various ways, by selecting, iterating,
  reducing and otherwise mangling JSON documents. For instance,
  running the command `jq 'map(.price) | add'` will take an array of
  JSON objects as input and return the sum of their "price" fields.

  `jq` can accept text input as well, but by default, `jq` reads a
  stream of JSON entities (including numbers and other literals) from
  `stdin`. Whitespace is only needed to separate entities such as 1
  and 2, and true and false.  One or more <files> may be specified, in
  which case `jq` will read input from those instead.

  The <options> are described in the [INVOKING JQ] section; they
  mostly concern input and output formatting. The <filter> is written
  in the jq language and specifies how to transform the input
  file or document.

  ## FILTERS

manpage_epilogue: |
  ## BUGS

  Presumably. Report them or discuss them at:

      https://github.com/jqlang/jq/issues

  ## AUTHOR

  Stephen Dolan `<mu@netsoc.tcd.ie>`

sections:
  - title: Invoking jq
    body: |

      jq filters run on a stream of JSON data. The input to jq is
      parsed as a sequence of whitespace-separated JSON values which
      are passed through the provided filter one at a time. The
      output(s) of the filter are written to standard output, as a
      sequence of newline-separated JSON data.

      The simplest and most common filter (or jq program) is `.`,
      which is the identity operator, copying the inputs of the jq
      processor to the output stream.  Because the default behavior of
      the jq processor is to read JSON texts from the input stream,
      and to pretty-print outputs, the `.` program's main use is to
      validate and pretty-print the inputs.  The jq programming
      language is quite rich and allows for much more than just
      validation and pretty-printing.

      Note: it is important to mind the shell's quoting rules.  As a
      general rule it's best to always quote (with single-quote
      characters on Unix shells) the jq program, as too many characters with special
      meaning to jq are also shell meta-characters.  For example, `jq
      "foo"` will fail on most Unix shells because that will be the same
      as `jq foo`, which will generally fail because `foo is not
      defined`.  When using the Windows command shell (cmd.exe) it's
      best to use double quotes around your jq program when given on the
      command-line (instead of the `-f program-file` option), but then
      double-quotes in the jq program need backslash escaping. When using
      the Powershell (`powershell.exe`) or the Powershell Core
      (`pwsh`/`pwsh.exe`), use single-quote characters around the jq
      program and backslash-escaped double-quotes (`\"`) inside the jq
      program.

      * Unix shells: `jq '.["foo"]'`
      * Powershell: `jq '.[\"foo\"]'`
      * Windows command shell: `jq ".[\"foo\"]"`

      Note: jq allows user-defined functions, but every jq program
      must have a top-level expression.

      You can affect how jq reads and writes its input and output
      using some command-line options:

      * `--null-input` / `-n`:

        Don't read any input at all. Instead, the filter is run once
        using `null` as the input. This is useful when using jq as a
        simple calculator or to construct JSON data from scratch.

      * `--raw-input` / `-R`:

        Don't parse the input as JSON. Instead, each line of text is
        passed to the filter as a string. If combined with `--slurp`,
        then the entire input is passed to the filter as a single long
        string.

      * `--slurp` / `-s`:

        Instead of running the filter for each JSON object in the
        input, read the entire input stream into a large array and run
        the filter just once.

      * `--compact-output` / `-c`:

        By default, jq pretty-prints JSON output. Using this option
        will result in more compact output by instead putting each
        JSON object on a single line.

      * `--raw-output` / `-r`:

        With this option, if the filter's result is a string then it
        will be written directly to standard output rather than being
        formatted as a JSON string with quotes. This can be useful for
        making jq filters talk to non-JSON-based systems.

      * `--raw-output0`:

        Like `-r` but jq will print NUL instead of newline after each output.
        This can be useful when the values being output can contain newlines.
        When the output value contains NUL, jq exits with non-zero code.

      * `--join-output` / `-j`:

        Like `-r` but jq won't print a newline after each output.

      * `--ascii-output` / `-a`:

        jq usually outputs non-ASCII Unicode codepoints as UTF-8, even
        if the input specified them as escape sequences (like
        "\u03bc"). Using this option, you can force jq to produce pure
        ASCII output with every non-ASCII character replaced with the
        equivalent escape sequence.

      * `--sort-keys` / `-S`:

        Output the fields of each object with the keys in sorted order.

      * `--color-output` / `-C` and `--monochrome-output` / `-M`:

        By default, jq outputs colored JSON if writing to a
        terminal. You can force it to produce color even if writing to
        a pipe or a file using `-C`, and disable color with `-M`.
        When the `NO_COLOR` environment variable is not empty, jq disables
        colored output by default, but you can enable it by `-C`.

        Colors can be configured with the `JQ_COLORS` environment
        variable (see below).

      * `--tab`:

        Use a tab for each indentation level instead of two spaces.

      * `--indent n`:

        Use the given number of spaces (no more than 7) for indentation.

      * `--unbuffered`:

        Flush the output after each JSON object is printed (useful if
        you're piping a slow data source into jq and piping jq's
        output elsewhere).

      * `--stream`:

        Parse the input in streaming fashion, outputting arrays of path
        and leaf values (scalars and empty arrays or empty objects).
        For example, `"a"` becomes `[[],"a"]`, and `[[],"a",["b"]]`
        becomes `[[0],[]]`, `[[1],"a"]`, and `[[2,0],"b"]`.

        This is useful for processing very large inputs.  Use this in
        conjunction with filtering and the `reduce` and `foreach` syntax
        to reduce large inputs incrementally.

      * `--stream-errors`:

        Like `--stream`, but invalid JSON inputs yield array values
        where the first element is the error and the second is a path.
        For example, `["a",n]` produces `["Invalid literal at line 1,
        column 7",[1]]`.

        Implies `--stream`.  Invalid JSON inputs produce no error values
        when `--stream` without `--stream-errors`.

      * `--seq`:

        Use the `application/json-seq` MIME type scheme for separating
        JSON texts in jq's input and output.  This means that an ASCII
        RS (record separator) character is printed before each value on
        output and an ASCII LF (line feed) is printed after every
        output.  Input JSON texts that fail to parse are ignored (but
        warned about), discarding all subsequent input until the next
        RS.  This mode also parses the output of jq without the `--seq`
        option.

      * `-f` / `--from-file`:

        Read the filter from a file rather than from a command line,
        like awk's -f option. This changes the filter argument to be
        interpreted as a filename, instead of the source of a program.

      * `-L directory` / `--library-path directory`:

        Prepend `directory` to the search list for modules.  If this
        option is used then no builtin search list is used.  See the
        section on modules below.

      * `--arg name value`:

        This option passes a value to the jq program as a predefined
        variable. If you run jq with `--arg foo bar`, then `$foo` is
        available in the program and has the value `"bar"`. Note that
        `value` will be treated as a string, so `--arg foo 123` will
        bind `$foo` to `"123"`.

        Named arguments are also available to the jq program as
        `$ARGS.named`. When the name is not a valid identifier, this is
        the only way to access it.

      * `--argjson name JSON-text`:

        This option passes a JSON-encoded value to the jq program as a
        predefined variable. If you run jq with `--argjson foo 123`, then
        `$foo` is available in the program and has the value `123`.

      * `--slurpfile variable-name filename`:

        This option reads all the JSON texts in the named file and binds
        an array of the parsed JSON values to the given global variable.
        If you run jq with `--slurpfile foo bar`, then `$foo` is available
        in the program and has an array whose elements correspond to the
        texts in the file named `bar`.

      * `--rawfile variable-name filename`:

        This option reads in the named file and binds its content to the given
        global variable.  If you run jq with `--rawfile foo bar`, then `$foo` is
        available in the program and has a string whose content is set to the text
        in the file named `bar`.

      * `--args`:

        Remaining arguments are positional string arguments.  These are
        available to the jq program as `$ARGS.positional[]`.

      * `--jsonargs`:

        Remaining arguments are positional JSON text arguments.  These
        are available to the jq program as `$ARGS.positional[]`.

      * `--exit-status` / `-e`:

        Sets the exit status of jq to 0 if the last output value was
        neither `false` nor `null`, 1 if the last output value was
        either `false` or `null`, or 4 if no valid result was ever
        produced.  Normally jq exits with 2 if there was any usage
        problem or system error, 3 if there was a jq program compile
        error, or 0 if the jq program ran.

        Another way to set the exit status is with the `halt_error`
        builtin function.

      * `--binary` / `-b`:

        Windows users using WSL, MSYS2, or Cygwin, should use this option
        when using a native jq.exe, otherwise jq will turn newlines (LFs)
        into carriage-return-then-newline (CRLF).

      * `--version` / `-V`:

        Output the jq version and exit with zero.

      * `--build-configuration`:

        Output the build configuration of jq and exit with zero.
        This output has no supported format or structure and may change
        without notice in future releases.

      * `--help` / `-h`:

        Output the jq help and exit with zero.

      * `--`:

        Terminates argument processing.  Remaining arguments are not
        interpreted as options.

      * `--run-tests [filename]`:

        Runs the tests in the given file or standard input.  This must
        be the last option given and does not honor all preceding
        options.  The input consists of comment lines, empty lines, and
        program lines followed by one input line, as many lines of
        output as are expected (one per output), and a terminating empty
        line.  Compilation failure tests start with a line containing
        only `%%FAIL`, then a line containing the program to compile,
        then a line containing an error message to compare to the
        actual.

        Be warned that this option can change backwards-incompatibly.

  - title: Basic filters
    entries:
      - title: "Identity: `.`"
        body: |

          The absolute simplest filter is `.` .  This filter takes its
          input and produces the same value as output.  That is, this
          is the identity operator.

          Since jq by default pretty-prints all output, a trivial
          program consisting of nothing but `.` can be used to format
          JSON output from, say, `curl`.

          Although the identity filter never modifies the value of its
          input, jq processing can sometimes make it appear as though
          it does.  For example, using the current implementation of
          jq, we would see that the expression:

              1E1234567890 | .

          produces `1.7976931348623157e+308` on at least one platform.
          This is because, in the process of parsing the number, this
          particular version of jq has converted it to an IEEE754
          double-precision representation, losing precision.

          The way in which jq handles numbers has changed over time
          and further changes are likely within the parameters set by
          the relevant JSON standards.  Moreover, build configuration
          options can alter how jq processes numbers.

          The following remarks are therefore offered with the
          understanding that they are intended to be descriptive of the
          current version of jq and should not be interpreted as being
          prescriptive:

          (1) Any arithmetic operation on a number that has not
          already been converted to an IEEE754 double precision
          representation will trigger a conversion to the IEEE754
          representation.

          (2) jq will attempt to maintain the original decimal
          precision of number literals (if the `--disable-decnum`
          build configuration option was not used), but in expressions
          such `1E1234567890`, precision will be lost if the exponent
          is too large.

          (3) Comparisons are carried out using the untruncated
          big decimal representation of numbers if available, as
          illustrated in one of the following examples.

          The examples below use the builtin function `have_decnum` in
          order to demonstrate the expected effects of using / not
          using the `--disable-decnum` build configuration option, and
          also to allow automated tests derived from these examples to
          pass regardless of whether that option is used.

        examples:
          - program: '.'
            input: '"Hello, world!"'
            output: ['"Hello, world!"']

          - program: '.'
            input: '0.12345678901234567890123456789'
            output: ['0.12345678901234567890123456789']

          - program: '[., tojson] == if have_decnum then [12345678909876543212345,"12345678909876543212345"] else [12345678909876543000000,"12345678909876543000000"] end'
            input: '12345678909876543212345'
            output: ['true']

          - program: '[1234567890987654321,-1234567890987654321 | tojson] == if have_decnum then ["1234567890987654321","-1234567890987654321"] else ["1234567890987654400","-1234567890987654400"] end'
            input: 'null'
            output: ['true']

          - program: '. < 0.12345678901234567890123456788'
            input: '0.12345678901234567890123456789'
            output: ['false']

          - program: 'map([., . == 1]) | tojson == if have_decnum then "[[1,true],[1.000,true],[1.0,true],[1.00,true]]" else "[[1,true],[1,true],[1,true],[1,true]]" end'
            input: '[1, 1.000, 1.0, 100e-2]'
            output: ['true']

          - program: '. as $big | [$big, $big + 1] | map(. > 10000000000000000000000000000000) | . == if have_decnum then [true, false] else [false, false] end'
            input: '10000000000000000000000000000001'
            output: ['true']

      - title: "Object Identifier-Index: `.foo`, `.foo.bar`"
        body: |

          The simplest *useful* filter has the form `.foo`. When given a
          JSON object (aka dictionary or hash) as input, `.foo` produces
          the value at the key "foo" if the key is present, or null otherwise.

          A filter of the form `.foo.bar` is equivalent to `.foo | .bar`.

          The `.foo` syntax only works for simple, identifier-like keys, that
          is, keys that are all made of alphanumeric characters and
          underscore, and which do not start with a digit.

          If the key contains special characters or starts with a digit,
          you need to surround it with double quotes like this:
          `."foo$"`, or else `.["foo$"]`.

          For example `.["foo::bar"]` and `.["foo.bar"]` work while
          `.foo::bar` does not.

        examples:
          - program: '.foo'
            input: '{"foo": 42, "bar": "less interesting data"}'
            output: ['42']

          - program: '.foo'
            input: '{"notfoo": true, "alsonotfoo": false}'
            output: ['null']

          - program: '.["foo"]'
            input: '{"foo": 42}'
            output: ['42']

      - title: "Optional Object Identifier-Index: `.foo?`"
        body: |

          Just like `.foo`, but does not output an error when `.` is not an
          object.

        examples:
          - program: '.foo?'
            input: '{"foo": 42, "bar": "less interesting data"}'
            output: ['42']
          - program: '.foo?'
            input: '{"notfoo": true, "alsonotfoo": false}'
            output: ['null']
          - program: '.["foo"]?'
            input: '{"foo": 42}'
            output: ['42']
          - program: '[.foo?]'
            input: '[1,2]'
            output: ['[]']

      - title: "Object Index: `.[<string>]`"
        body: |

          You can also look up fields of an object using syntax like
          `.["foo"]` (`.foo` above is a shorthand version of this, but
          only for identifier-like strings).

      - title: "Array Index: `.[<number>]`"
        body: |

          When the index value is an integer, `.[<number>]` can index
          arrays.  Arrays are zero-based, so `.[2]` returns the third
          element.

          Negative indices are allowed, with -1 referring to the last
          element, -2 referring to the next to last element, and so on.

        examples:
          - program: '.[0]'
            input: '[{"name":"JSON", "good":true}, {"name":"XML", "good":false}]'
            output: ['{"name":"JSON", "good":true}']

          - program: '.[2]'
            input: '[{"name":"JSON", "good":true}, {"name":"XML", "good":false}]'
            output: ['null']

          - program: '.[-2]'
            input: '[1,2,3]'
            output: ['2']

      - title: "Array/String Slice: `.[<number>:<number>]`"
        body: |

          The `.[<number>:<number>]` syntax can be used to return a
          subarray of an array or substring of a string. The array
          returned by `.[10:15]` will be of length 5, containing the
          elements from index 10 (inclusive) to index 15 (exclusive).
          Either index may be negative (in which case it counts
          backwards from the end of the array), or omitted (in which
          case it refers to the start or end of the array).
          Indices are zero-based.

        examples:
          - program: '.[2:4]'
            input: '["a","b","c","d","e"]'
            output: ['["c", "d"]']

          - program: '.[2:4]'
            input: '"abcdefghi"'
            output: ['"cd"']

          - program: '.[:3]'
            input: '["a","b","c","d","e"]'
            output: ['["a", "b", "c"]']

          - program: '.[-2:]'
            input: '["a","b","c","d","e"]'
            output: ['["d", "e"]']

      - title: "Array/Object Value Iterator: `.[]`"
        body: |

          If you use the `.[index]` syntax, but omit the index
          entirely, it will return *all* of the elements of an
          array. Running `.[]` with the input `[1,2,3]` will produce the
          numbers as three separate results, rather than as a single
          array. A filter of the form `.foo[]` is equivalent to
          `.foo | .[]`.

          You can also use this on an object, and it will return all
          the values of the object.

          Note that the iterator operator is a generator of values.

        examples:
          - program: '.[]'
            input: '[{"name":"JSON", "good":true}, {"name":"XML", "good":false}]'
            output:
              - '{"name":"JSON", "good":true}'
              - '{"name":"XML", "good":false}'

          - program: '.[]'
            input: '[]'
            output: []

          - program: '.foo[]'
            input: '{"foo":[1,2,3]}'
            output: ['1','2','3']

          - program: '.[]'
            input: '{"a": 1, "b": 1}'
            output: ['1', '1']

      - title: "`.[]?`"
        body: |

          Like `.[]`, but no errors will be output if . is not an array
          or object. A filter of the form `.foo[]?` is equivalent to
          `.foo | .[]?`.

      - title: "Comma: `,`"
        body: |

          If two filters are separated by a comma, then the
          same input will be fed into both and the two filters' output
          value streams will be concatenated in order: first, all of the
          outputs produced by the left expression, and then all of the
          outputs produced by the right. For instance, filter `.foo,
          .bar`, produces both the "foo" fields and "bar" fields as
          separate outputs.

          The `,` operator is one way to construct generators.

        examples:
          - program: '.foo, .bar'
            input: '{"foo": 42, "bar": "something else", "baz": true}'
            output: ['42', '"something else"']

          - program: ".user, .projects[]"
            input: '{"user":"stedolan", "projects": ["jq", "wikiflow"]}'
            output: ['"stedolan"', '"jq"', '"wikiflow"']

          - program: '.[4,2]'
            input: '["a","b","c","d","e"]'
            output: ['"e"', '"c"']

      - title: "Pipe: `|`"
        body: |

          The | operator combines two filters by feeding the output(s) of
          the one on the left into the input of the one on the right. It's
          similar to the Unix shell's pipe, if you're used to that.

          If the one on the left produces multiple results, the one on
          the right will be run for each of those results. So, the
          expression `.[] | .foo` retrieves the "foo" field of each
          element of the input array.  This is a cartesian product,
          which can be surprising.

          Note that `.a.b.c` is the same as `.a | .b | .c`.

          Note too that `.` is the input value at the particular stage
          in a "pipeline", specifically: where the `.` expression appears.
          Thus `.a | . | .b` is the same as `.a.b`, as the `.` in the
          middle refers to whatever value `.a` produced.

        examples:
          - program: '.[] | .name'
            input: '[{"name":"JSON", "good":true}, {"name":"XML", "good":false}]'
            output: ['"JSON"', '"XML"']

      - title: "Parenthesis"
        body: |

          Parenthesis work as a grouping operator just as in any typical
          programming language.

        examples:
          - program: '(. + 2) * 5'
            input: '1'
            output: ['15']

  - title: Types and Values
    body: |

      jq supports the same set of datatypes as JSON - numbers,
      strings, booleans, arrays, objects (which in JSON-speak are
      hashes with only string keys), and "null".

      Booleans, null, strings and numbers are written the same way as
      in JSON. Just like everything else in jq, these simple
      values take an input and produce an output - `42` is a valid jq
      expression that takes an input, ignores it, and returns 42
      instead.

      Numbers in jq are internally represented by their IEEE754 double
      precision approximation. Any arithmetic operation with numbers,
      whether they are literals or results of previous filters, will
      produce a double precision floating point result.

      However, when parsing a literal jq will store the original literal
      string. If no mutation is applied to this value then it will make
      to the output in its original form, even if conversion to double
      would result in a loss.

    entries:
      - title: "Array construction: `[]`"
        body: |

          As in JSON, `[]` is used to construct arrays, as in
          `[1,2,3]`. The elements of the arrays can be any jq
          expression, including a pipeline. All of the results produced
          by all of the expressions are collected into one big array.
          You can use it to construct an array out of a known quantity
          of values (as in `[.foo, .bar, .baz]`) or to "collect" all the
          results of a filter into an array (as in `[.items[].name]`)

          Once you understand the "," operator, you can look at jq's array
          syntax in a different light: the expression `[1,2,3]` is not using a
          built-in syntax for comma-separated arrays, but is instead applying
          the `[]` operator (collect results) to the expression 1,2,3 (which
          produces three different results).

          If you have a filter `X` that produces four results,
          then the expression `[X]` will produce a single result, an
          array of four elements.

        examples:
          - program: "[.user, .projects[]]"
            input: '{"user":"stedolan", "projects": ["jq", "wikiflow"]}'
            output: ['["stedolan", "jq", "wikiflow"]']
          - program: "[ .[] | . * 2]"
            input: '[1, 2, 3]'
            output: ['[2, 4, 6]']

      - title: "Object Construction: `{}`"
        body: |

          Like JSON, `{}` is for constructing objects (aka
          dictionaries or hashes), as in: `{"a": 42, "b": 17}`.

          If the keys are "identifier-like", then the quotes can be left
          off, as in `{a:42, b:17}`.  Variable references as key
          expressions use the value of the variable as the key.  Key
          expressions other than constant literals, identifiers, or
          variable references, need to be parenthesized, e.g.,
          `{("a"+"b"):59}`.

          The value can be any expression (although you may need to wrap
          it in parentheses if, for example, it contains colons), which
          gets applied to the {} expression's input (remember, all
          filters have an input and an output).

              {foo: .bar}

          will produce the JSON object `{"foo": 42}` if given the JSON
          object `{"bar":42, "baz":43}` as its input. You can use this
          to select particular fields of an object: if the input is an
          object with "user", "title", "id", and "content" fields and
          you just want "user" and "title", you can write

              {user: .user, title: .title}

          Because that is so common, there's a shortcut syntax for it:
          `{user, title}`.

          If one of the expressions produces multiple results,
          multiple dictionaries will be produced. If the input's

              {"user":"stedolan","titles":["JQ Primer", "More JQ"]}

          then the expression

              {user, title: .titles[]}

          will produce two outputs:

              {"user":"stedolan", "title": "JQ Primer"}
              {"user":"stedolan", "title": "More JQ"}

          Putting parentheses around the key means it will be evaluated as an
          expression. With the same input as above,

              {(.user): .titles}

          produces

              {"stedolan": ["JQ Primer", "More JQ"]}

          Variable references as keys use the value of the variable as
          the key.  Without a value then the variable's name becomes the
          key and its value becomes the value,

              "f o o" as $foo | "b a r" as $bar | {$foo, $bar:$foo}

          produces

              {"foo":"f o o","b a r":"f o o"}

        examples:
          - program: '{user, title: .titles[]}'
            input: '{"user":"stedolan","titles":["JQ Primer", "More JQ"]}'
            output:
              - '{"user":"stedolan", "title": "JQ Primer"}'
              - '{"user":"stedolan", "title": "More JQ"}'
          - program: '{(.user): .titles}'
            input: '{"user":"stedolan","titles":["JQ Primer", "More JQ"]}'
            output: ['{"stedolan": ["JQ Primer", "More JQ"]}']

      - title: "Recursive Descent: `..`"
        body: |

          Recursively descends `.`, producing every value.  This is the
          same as the zero-argument `recurse` builtin (see below).  This
          is intended to resemble the XPath `//` operator.  Note that
          `..a` does not work; use `.. | .a` instead.  In the example
          below we use `.. | .a?` to find all the values of object keys
          "a" in any object found "below" `.`.

          This is particularly useful in conjunction with `path(EXP)`
          (also see below) and the `?` operator.

        examples:
          - program: '.. | .a?'
            input: '[[{"a":1}]]'
            output: ['1']

  - title: Builtin operators and functions
    body: |

      Some jq operators (for instance, `+`) do different things
      depending on the type of their arguments (arrays, numbers,
      etc.). However, jq never does implicit type conversions. If you
      try to add a string to an object you'll get an error message and
      no result.

      Please note that all numbers are converted to IEEE754 double precision
      floating point representation. Arithmetic and logical operators are working
      with these converted doubles. Results of all such operations are also limited
      to the double precision.

      The only exception to this behaviour of number is a snapshot of original number
      literal. When a number which originally was provided as a literal is never
      mutated until the end of the program then it is printed to the output in its
      original literal form. This also includes cases when the original literal
      would be truncated when converted to the IEEE754 double precision floating point
      number.

    entries:
      - title: "Addition: `+`"
        body: |

          The operator `+` takes two filters, applies them both
          to the same input, and adds the results together. What
          "adding" means depends on the types involved:

          - **Numbers** are added by normal arithmetic.

          - **Arrays** are added by being concatenated into a larger array.

          - **Strings** are added by being joined into a larger string.

          - **Objects** are added by merging, that is, inserting all
            the key-value pairs from both objects into a single
            combined object. If both objects contain a value for the
            same key, the object on the right of the `+` wins. (For
            recursive merge use the `*` operator.)

          `null` can be added to any value, and returns the other
          value unchanged.

        examples:
          - program: '.a + 1'
            input: '{"a": 7}'
            output: ['8']
          - program: '.a + .b'
            input: '{"a": [1,2], "b": [3,4]}'
            output: ['[1,2,3,4]']
          - program: '.a + null'
            input: '{"a": 1}'
            output: ['1']
          - program: '.a + 1'
            input: '{}'
            output: ['1']
          - program: '{a: 1} + {b: 2} + {c: 3} + {a: 42}'
            input: 'null'
            output: ['{"a": 42, "b": 2, "c": 3}']

      - title: "Subtraction: `-`"
        body: |

          As well as normal arithmetic subtraction on numbers, the `-`
          operator can be used on arrays to remove all occurrences of
          the second array's elements from the first array.

        examples:
          - program: '4 - .a'
            input: '{"a":3}'
            output: ['1']
          - program: . - ["xml", "yaml"]
            input: '["xml", "yaml", "json"]'
            output: ['["json"]']

      - title: "Multiplication, division, modulo: `*`, `/`, `%`"
        body: |

          These infix operators behave as expected when given two numbers.
          Division by zero raises an error. `x % y` computes x modulo y.

          Multiplying a string by a number produces the concatenation of
          that string that many times. `"x" * 0` produces `""`.

          Dividing a string by another splits the first using the second
          as separators.

          Multiplying two objects will merge them recursively: this works
          like addition but if both objects contain a value for the
          same key, and the values are objects, the two are merged with
          the same strategy.

        examples:
          - program: '10 / . * 3'
            input: '5'
            output: ['6']
          - program: '. / ", "'
            input: '"a, b,c,d, e"'
            output: ['["a","b,c,d","e"]']
          - program: '{"k": {"a": 1, "b": 2}} * {"k": {"a": 0,"c": 3}}'
            input: 'null'
            output: ['{"k": {"a": 0, "b": 2, "c": 3}}']
          - program: '.[] | (1 / .)?'
            input: '[1,0,-1]'
            output: ['1', '-1']

      - title: "`abs`"
        body: |

          The builtin function `abs` is defined naively as: `if . < 0 then - . else . end`.

          For numeric input, this is the absolute value.  See the
          section on the identity filter for the implications of this
          definition for numeric input.

          To compute the absolute value of a number as a floating point number, you may wish use `fabs`.

        examples:
          - program: 'map(abs)'
            input: '[-10, -1.1, -1e-1]'
            output: ['[10,1.1,1e-1]']

      - title: "`length`"
        body: |

          The builtin function `length` gets the length of various
          different types of value:

          - The length of a **string** is the number of Unicode
            codepoints it contains (which will be the same as its
            JSON-encoded length in bytes if it's pure ASCII).

          - The length of a **number** is its absolute value.

          - The length of an **array** is the number of elements.

          - The length of an **object** is the number of key-value pairs.

          - The length of **null** is zero.

          - It is an error to use `length` on a **boolean**.

        examples:
          - program: '.[] | length'
            input: '[[1,2], "string", {"a":2}, null, -5]'
            output: ['2', '6', '1', '0', '5']


      - title: "`utf8bytelength`"
        body: |

          The builtin function `utf8bytelength` outputs the number of
          bytes used to encode a string in UTF-8.

        examples:
          - program: 'utf8bytelength'
            input: '"\u03bc"'
            output: ['2']

      - title: "`keys`, `keys_unsorted`"
        body: |

          The builtin function `keys`, when given an object, returns
          its keys in an array.

          The keys are sorted "alphabetically", by unicode codepoint
          order. This is not an order that makes particular sense in
          any particular language, but you can count on it being the
          same for any two objects with the same set of keys,
          regardless of locale settings.

          When `keys` is given an array, it returns the valid indices
          for that array: the integers from 0 to length-1.

          The `keys_unsorted` function is just like `keys`, but if
          the input is an object then the keys will not be sorted,
          instead the keys will roughly be in insertion order.

        examples:
          - program: 'keys'
            input: '{"abc": 1, "abcd": 2, "Foo": 3}'
            output: ['["Foo", "abc", "abcd"]']
          - program: 'keys'
            input: '[42,3,35]'
            output: ['[0,1,2]']

      - title: "`has(key)`"
        body: |

          The builtin function `has` returns whether the input object
          has the given key, or the input array has an element at the
          given index.

          `has($key)` has the same effect as checking whether `$key`
          is a member of the array returned by `keys`, although `has`
          will be faster.

        examples:
          - program: 'map(has("foo"))'
            input: '[{"foo": 42}, {}]'
            output: ['[true, false]']
          - program: 'map(has(2))'
            input: '[[0,1], ["a","b","c"]]'
            output: ['[false, true]']

      - title: "`in`"
        body: |

          The builtin function `in` returns whether or not the input key is in the
          given object, or the input index corresponds to an element
          in the given array. It is, essentially, an inversed version
          of `has`.

        examples:
          - program: '.[] | in({"foo": 42})'
            input: '["foo", "bar"]'
            output: ['true', 'false']
          - program: 'map(in([0,1]))'
            input: '[2, 0]'
            output: ['[false, true]']

      - title: "`map(f)`, `map_values(f)`"
        body: |

          For any filter `f`, `map(f)` and `map_values(f)` apply `f`
          to each of the values in the input array or object, that is,
          to the values of `.[]`.

          In the absence of errors, `map(f)` always outputs an array
          whereas `map_values(f)` outputs an array if given an array,
          or an object if given an object.

          When the input to `map_values(f)` is an object, the output
          object has the same keys as the input object except for
          those keys whose values when piped to `f` produce no values
          at all.

          The key difference between `map(f)` and `map_values(f)` is
          that the former simply forms an array from all the values of
          `($x|f)` for each value, `$x`, in the input array or object,
          but `map_values(f)` only uses `first($x|f)`.

          Specifically, for object inputs, `map_values(f)` constructs
          the output object by examining in turn the value of
          `first(.[$k]|f)` for each key, `$k`, of the input.  If this
          expression produces no values, then the corresponding key
          will be dropped; otherwise, the output object will have that
          value at the key, `$k`.

          Here are some examples to clarify the behavior of `map` and
          `map_values` when applied to arrays. These examples assume the
          input is `[1]` in all cases:

              map(.+1)          #=>  [2]
              map(., .)         #=>  [1,1]
              map(empty)        #=>  []

              map_values(.+1)   #=>  [2]
              map_values(., .)  #=>  [1]
              map_values(empty) #=>  []

          `map(f)` is equivalent to `[.[] | f]` and
          `map_values(f)` is equivalent to `.[] |= f`.

          In fact, these are their implementations.


        examples:
          - program: 'map(.+1)'
            input: '[1,2,3]'
            output: ['[2,3,4]']

          - program: 'map_values(.+1)'
            input: '{"a": 1, "b": 2, "c": 3}'
            output: ['{"a": 2, "b": 3, "c": 4}']

          - program: 'map(., .)'
            input: '[1,2]'
            output: ['[1,1,2,2]']

          - program: 'map_values(. // empty)'
            input: '{"a": null, "b": true, "c": false}'
            output: ['{"b":true}']


      - title: "`pick(pathexps)`"
        body: |

          Emit the projection of the input object or array defined by the
          specified sequence of path expressions, such that if `p` is any
          one of these specifications, then `(. | p)` will evaluate to the
          same value as `(. | pick(pathexps) | p)`. For arrays, negative
          indices and `.[m:n]` specifications should not be used.

        examples:
          - program: 'pick(.a, .b.c, .x)'
            input: '{"a": 1, "b": {"c": 2, "d": 3}, "e": 4}'
            output: ['{"a":1,"b":{"c":2},"x":null}']

          - program: 'pick(.[2], .[0], .[0])'
            input: '[1,2,3,4]'
            output: ['[1,null,3]']


      - title: "`path(path_expression)`"
        body: |

          Outputs array representations of the given path expression
          in `.`.  The outputs are arrays of strings (object keys)
          and/or numbers (array indices).

          Path expressions are jq expressions like `.a`, but also `.[]`.
          There are two types of path expressions: ones that can match
          exactly, and ones that cannot.  For example, `.a.b.c` is an
          exact match path expression, while `.a[].b` is not.

          `path(exact_path_expression)` will produce the array
          representation of the path expression even if it does not
          exist in `.`, if `.` is `null` or an array or an object.

          `path(pattern)` will produce array representations of the
          paths matching `pattern` if the paths exist in `.`.

          Note that the path expressions are not different from normal
          expressions.  The expression
          `path(..|select(type=="boolean"))` outputs all the paths to
          boolean values in `.`, and only those paths.

        examples:
          - program: 'path(.a[0].b)'
            input: 'null'
            output: ['["a",0,"b"]']
          - program: '[path(..)]'
            input: '{"a":[{"b":1}]}'
            output: ['[[],["a"],["a",0],["a",0,"b"]]']

      - title: "`del(path_expression)`"
        body: |

          The builtin function `del` removes a key and its corresponding
          value from an object.

        examples:
          - program: 'del(.foo)'
            input: '{"foo": 42, "bar": 9001, "baz": 42}'
            output: ['{"bar": 9001, "baz": 42}']
          - program: 'del(.[1, 2])'
            input: '["foo", "bar", "baz"]'
            output: ['["foo"]']

      - title: "`getpath(PATHS)`"
        body: |

          The builtin function `getpath` outputs the values in `.` found
          at each path in `PATHS`.

        examples:
          - program: 'getpath(["a","b"])'
            input: 'null'
            output: ['null']
          - program: '[getpath(["a","b"], ["a","c"])]'
            input: '{"a":{"b":0, "c":1}}'
            output: ['[0, 1]']

      - title: "`setpath(PATHS; VALUE)`"
        body: |

          The builtin function `setpath` sets the `PATHS` in `.` to `VALUE`.

        examples:
          - program: 'setpath(["a","b"]; 1)'
            input: 'null'
            output: ['{"a": {"b": 1}}']
          - program: 'setpath(["a","b"]; 1)'
            input: '{"a":{"b":0}}'
            output: ['{"a": {"b": 1}}']
          - program: 'setpath([0,"a"]; 1)'
            input: 'null'
            output: ['[{"a":1}]']

      - title: "`delpaths(PATHS)`"
        body: |

          The builtin function `delpaths` deletes the `PATHS` in `.`.
          `PATHS` must be an array of paths, where each path is an array
          of strings and numbers.

        examples:
          - program: 'delpaths([["a","b"]])'
            input: '{"a":{"b":1},"x":{"y":2}}'
            output: ['{"a":{},"x":{"y":2}}']

      - title: "`to_entries`, `from_entries`, `with_entries(f)`"
        body: |

          These functions convert between an object and an array of
          key-value pairs. If `to_entries` is passed an object, then
          for each `k: v` entry in the input, the output array
          includes `{"key": k, "value": v}`.

          `from_entries` does the opposite conversion, and `with_entries(f)`
          is a shorthand for `to_entries | map(f) | from_entries`, useful for
          doing some operation to all keys and values of an object.
          `from_entries` accepts `"key"`, `"Key"`, `"name"`, `"Name"`,
          `"value"`, and `"Value"` as keys.

        examples:
          - program: 'to_entries'
            input: '{"a": 1, "b": 2}'
            output: ['[{"key":"a", "value":1}, {"key":"b", "value":2}]']
          - program: 'from_entries'
            input: '[{"key":"a", "value":1}, {"key":"b", "value":2}]'
            output: ['{"a": 1, "b": 2}']
          - program: 'with_entries(.key |= "KEY_" + .)'
            input: '{"a": 1, "b": 2}'
            output: ['{"KEY_a": 1, "KEY_b": 2}']


      - title: "`select(boolean_expression)`"
        body: |

          The function `select(f)` produces its input unchanged if
          `f` returns true for that input, and produces no output
          otherwise.

          It's useful for filtering lists: `[1,2,3] | map(select(. >= 2))`
          will give you `[2,3]`.

        examples:
          - program: 'map(select(. >= 2))'
            input: '[1,5,3,0,7]'
            output: ['[5,3,7]']
          - program: '.[] | select(.id == "second")'
            input: '[{"id": "first", "val": 1}, {"id": "second", "val": 2}]'
            output: ['{"id": "second", "val": 2}']


      - title: "`arrays`, `objects`, `iterables`, `booleans`, `numbers`, `normals`, `finites`, `strings`, `nulls`, `values`, `scalars`"
        body: |

          These built-ins select only inputs that are arrays, objects,
          iterables (arrays or objects), booleans, numbers, normal
          numbers, finite numbers, strings, null, non-null values, and
          non-iterables, respectively.

        examples:
          - program: '.[]|numbers'
            input: '[[],{},1,"foo",null,true,false]'
            output: ['1']

      - title: "`empty`"
        body: |

          `empty` returns no results. None at all. Not even `null`.

          It's useful on occasion. You'll know if you need it :)

        examples:
          - program: '1, empty, 2'
            input: 'null'
            output: ['1', '2']
          - program: '[1,2,empty,3]'
            input: 'null'
            output: ['[1,2,3]']

      - title: "`error`, `error(message)`"
        body: |

          Produces an error with the input value, or with the message
          given as the argument. Errors can be caught with try/catch;
          see below.

        examples:
          - program: 'try error catch .'
            input: '"error message"'
            output: ['"error message"']

          - program: 'try error("invalid value: \(.)") catch .'
            input: '42'
            output: ['"invalid value: 42"']

      - title: "`halt`"
        body: |

          Stops the jq program with no further outputs.  jq will exit
          with exit status `0`.

      - title: "`halt_error`, `halt_error(exit_code)`"
        body: |

          Stops the jq program with no further outputs.  The input will
          be printed on `stderr` as raw output (i.e., strings will not
          have double quotes) with no decoration, not even a newline.

          The given `exit_code` (defaulting to `5`) will be jq's exit
          status.

          For example, `"Error: something went wrong\n"|halt_error(1)`.

      - title: "`$__loc__`"
        body: |

          Produces an object with a "file" key and a "line" key, with
          the filename and line number where `$__loc__` occurs, as
          values.

        examples:
          - program: 'try error("\($__loc__)") catch .'
            input: 'null'
            output: ['"{\"file\":\"<top-level>\",\"line\":1}"']

      - title: "`paths`, `paths(node_filter)`"
        body: |

          `paths` outputs the paths to all the elements in its input
          (except it does not output the empty list, representing .
          itself).

          `paths(f)` outputs the paths to any values for which `f` is `true`.
          That is, `paths(type == "number")` outputs the paths to all numeric
          values.

        examples:
          - program: '[paths]'
            input: '[1,[[],{"a":2}]]'
            output: ['[[0],[1],[1,0],[1,1],[1,1,"a"]]']
          - program: '[paths(type == "number")]'
            input: '[1,[[],{"a":2}]]'
            output: ['[[0],[1,1,"a"]]']

      - title: "`add`, `add(generator)`"
        body: |

          The filter `add` takes as input an array, and produces as
          output the elements of the array added together. This might
          mean summed, concatenated or merged depending on the types
          of the elements of the input array - the rules are the same
          as those for the `+` operator (described above).

          If the input is an empty array, `add` returns `null`.

          `add(generator)` operates on the given generator rather than
          the input.

        examples:
          - program: add
            input: '["a","b","c"]'
            output: ['"abc"']
          - program: add
            input: '[1, 2, 3]'
            output: ['6']
          - program: add
            input: '[]'
            output: ["null"]
          - program: add(.[].a)
            input: '[{"a":3}, {"a":5}, {"b":6}]'
            output: ['8']

      - title: "`any`, `any(condition)`, `any(generator; condition)`"
        body: |

          The filter `any` takes as input an array of boolean values,
          and produces `true` as output if any of the elements of
          the array are `true`.

          If the input is an empty array, `any` returns `false`.

          The `any(condition)` form applies the given condition to the
          elements of the input array.

          The `any(generator; condition)` form applies the given
          condition to all the outputs of the given generator.

        examples:
          - program: any
            input: '[true, false]'
            output: ["true"]
          - program: any
            input: '[false, false]'
            output: ["false"]
          - program: any
            input: '[]'
            output: ["false"]

      - title: "`all`, `all(condition)`, `all(generator; condition)`"
        body: |

          The filter `all` takes as input an array of boolean values,
          and produces `true` as output if all of the elements of
          the array are `true`.

          The `all(condition)` form applies the given condition to the
          elements of the input array.

          The `all(generator; condition)` form applies the given
          condition to all the outputs of the given generator.

          If the input is an empty array, `all` returns `true`.

        examples:
          - program: all
            input: '[true, false]'
            output: ["false"]
          - program: all
            input: '[true, true]'
            output: ["true"]
          - program: all
            input: '[]'
            output: ["true"]

      - title: "`flatten`, `flatten(depth)`"
        body: |

          The filter `flatten` takes as input an array of nested arrays,
          and produces a flat array in which all arrays inside the original
          array have been recursively replaced by their values. You can pass
          an argument to it to specify how many levels of nesting to flatten.

          `flatten(2)` is like `flatten`, but going only up to two
          levels deep.

        examples:
          - program: flatten
            input: '[1, [2], [[3]]]'
            output: ["[1, 2, 3]"]
          - program: flatten(1)
            input: '[1, [2], [[3]]]'
            output: ["[1, 2, [3]]"]
          - program: flatten
            input: '[[]]'
            output: ["[]"]
          - program: flatten
            input: '[{"foo": "bar"}, [{"foo": "baz"}]]'
            output: ['[{"foo": "bar"}, {"foo": "baz"}]']

      - title: "`range(upto)`, `range(from; upto)`, `range(from; upto; by)`"
        body: |

          The `range` function produces a range of numbers. `range(4; 10)`
          produces 6 numbers, from 4 (inclusive) to 10 (exclusive). The numbers
          are produced as separate outputs. Use `[range(4; 10)]` to get a range as
          an array.

          The one argument form generates numbers from 0 to the given
          number, with an increment of 1.

          The two argument form generates numbers from `from` to `upto`
          with an increment of 1.

          The three argument form generates numbers `from` to `upto`
          with an increment of `by`.

        examples:
          - program: 'range(2; 4)'
            input: 'null'
            output: ['2', '3']
          - program: '[range(2; 4)]'
            input: 'null'
            output: ['[2,3]']
          - program: '[range(4)]'
            input: 'null'
            output: ['[0,1,2,3]']
          - program: '[range(0; 10; 3)]'
            input: 'null'
            output: ['[0,3,6,9]']
          - program: '[range(0; 10; -1)]'
            input: 'null'
            output: ['[]']
          - program: '[range(0; -5; -1)]'
            input: 'null'
            output: ['[0,-1,-2,-3,-4]']

      - title: "`floor`"
        body: |

          The `floor` function returns the floor of its numeric input.

        examples:
          - program: 'floor'
            input: '3.14159'
            output: ['3']

      - title: "`sqrt`"
        body: |

          The `sqrt` function returns the square root of its numeric input.

        examples:
          - program: 'sqrt'
            input: '9'
            output: ['3']

      - title: "`tonumber`"
        body: |

          The `tonumber` function parses its input as a number. It
          will convert correctly-formatted strings to their numeric
          equivalent, leave numbers alone, and give an error on all other input.

        examples:
          - program: '.[] | tonumber'
            input: '[1, "1"]'
            output: ['1', '1']

      - title: "`toboolean`"
        body: |

          The `toboolean` function parses its input as a boolean. It
          will convert correctly-formatted strings to their boolean
          equivalent, leave booleans alone, and give an error on all other input.

        examples:
          - program: '.[] | toboolean'
            input: '["true", "false", true, false]'
            output: ['true', 'false', 'true', 'false']

      - title: "`tostring`"
        body: |

          The `tostring` function prints its input as a
          string. Strings are left unchanged, and all other values are
          JSON-encoded.

        examples:
          - program: '.[] | tostring'
            input: '[1, "1", [1]]'
            output: ['"1"', '"1"', '"[1]"']

      - title: "`type`"
        body: |

          The `type` function returns the type of its argument as a
          string, which is one of null, boolean, number, string, array
          or object.

        examples:
          - program: 'map(type)'
            input: '[0, false, [], {}, null, "hello"]'
            output: ['["number", "boolean", "array", "object", "null", "string"]']

      - title: "`infinite`, `nan`, `isinfinite`, `isnan`, `isfinite`, `isnormal`"
        body: |

          Some arithmetic operations can yield infinities and "not a
          number" (NaN) values.  The `isinfinite` builtin returns `true`
          if its input is infinite.  The `isnan` builtin returns `true`
          if its input is a NaN.  The `infinite` builtin returns a
          positive infinite value.  The `nan` builtin returns a NaN.
          The `isnormal` builtin returns true if its input is a normal
          number.

          Note that division by zero raises an error.

          Currently most arithmetic operations operating on infinities,
          NaNs, and sub-normals do not raise errors.

        examples:
          - program: '.[] | (infinite * .) < 0'
            input: '[-1, 1]'
            output: ['true', 'false']
          - program: 'infinite, nan | type'
            input: 'null'
            output: ['"number"', '"number"']

      - title: "`sort`, `sort_by(path_expression)`"
        body: |

          The `sort` functions sorts its input, which must be an
          array. Values are sorted in the following order:

          * `null`
          * `false`
          * `true`
          * numbers
          * strings, in alphabetical order (by unicode codepoint value)
          * arrays, in lexical order
          * objects

          The ordering for objects is a little complex: first they're
          compared by comparing their sets of keys (as arrays in
          sorted order), and if their keys are equal then the values
          are compared key by key.

          `sort_by` may be used to sort by a particular field of an
          object, or by applying any jq filter. `sort_by(f)` compares
          two elements by comparing the result of `f` on each element.
          When `f` produces multiple values, it firstly compares the
          first values, and the second values if the first values are
          equal, and so on.

        examples:
          - program: 'sort'
            input: '[8,3,null,6]'
            output: ['[null,3,6,8]']

          - program: 'sort_by(.foo)'
            input: '[{"foo":4, "bar":10}, {"foo":3, "bar":10}, {"foo":2, "bar":1}]'
            output: ['[{"foo":2, "bar":1}, {"foo":3, "bar":10}, {"foo":4, "bar":10}]']

          - program: 'sort_by(.foo, .bar)'
            input: '[{"foo":4, "bar":10}, {"foo":3, "bar":20}, {"foo":2, "bar":1}, {"foo":3, "bar":10}]'
            output: ['[{"foo":2, "bar":1}, {"foo":3, "bar":10}, {"foo":3, "bar":20}, {"foo":4, "bar":10}]']

      - title: "`group_by(path_expression)`"
        body: |

          `group_by(.foo)` takes as input an array, groups the
          elements having the same `.foo` field into separate arrays,
          and produces all of these arrays as elements of a larger
          array, sorted by the value of the `.foo` field.

          Any jq expression, not just a field access, may be used in
          place of `.foo`. The sorting order is the same as described
          in the `sort` function above.

        examples:
          - program: 'group_by(.foo)'
            input: '[{"foo":1, "bar":10}, {"foo":3, "bar":100}, {"foo":1, "bar":1}]'
            output: ['[[{"foo":1, "bar":10}, {"foo":1, "bar":1}], [{"foo":3, "bar":100}]]']

      - title: "`min`, `max`, `min_by(path_exp)`, `max_by(path_exp)`"
        body: |

          Find the minimum or maximum element of the input array.

          The `min_by(path_exp)` and `max_by(path_exp)` functions allow
          you to specify a particular field or property to examine, e.g.
          `min_by(.foo)` finds the object with the smallest `foo` field.

        examples:
          - program: 'min'
            input: '[5,4,2,7]'
            output: ['2']
          - program: 'max_by(.foo)'
            input: '[{"foo":1, "bar":14}, {"foo":2, "bar":3}]'
            output: ['{"foo":2, "bar":3}']

      - title: "`unique`, `unique_by(path_exp)`"
        body: |

          The `unique` function takes as input an array and produces
          an array of the same elements, in sorted order, with
          duplicates removed.

          The `unique_by(path_exp)` function will keep only one element
          for each value obtained by applying the argument. Think of it
          as making an array by taking one element out of every group
          produced by `group`.

        examples:
          - program: 'unique'
            input: '[1,2,5,3,5,3,1,3]'
            output: ['[1,2,3,5]']
          - program: 'unique_by(.foo)'
            input: '[{"foo": 1, "bar": 2}, {"foo": 1, "bar": 3}, {"foo": 4, "bar": 5}]'
            output: ['[{"foo": 1, "bar": 2}, {"foo": 4, "bar": 5}]']
          - program: 'unique_by(length)'
            input: '["chunky", "bacon", "kitten", "cicada", "asparagus"]'
            output: ['["bacon", "chunky", "asparagus"]']

      - title: "`reverse`"
        body: |

          This function reverses an array.

        examples:
          - program: 'reverse'
            input: '[1,2,3,4]'
            output: ['[4,3,2,1]']

      - title: "`contains(element)`"
        body: |

          The filter `contains(b)` will produce true if b is
          completely contained within the input. A string B is
          contained in a string A if B is a substring of A. An array B
          is contained in an array A if all elements in B are
          contained in any element in A. An object B is contained in
          object A if all of the values in B are contained in the
          value in A with the same key. All other types are assumed to
          be contained in each other if they are equal.

        examples:
          - program: 'contains("bar")'
            input: '"foobar"'
            output: ['true']
          - program: 'contains(["baz", "bar"])'
            input: '["foobar", "foobaz", "blarp"]'
            output: ['true']
          - program: 'contains(["bazzzzz", "bar"])'
            input: '["foobar", "foobaz", "blarp"]'
            output: ['false']
          - program: 'contains({foo: 12, bar: [{barp: 12}]})'
            input: '{"foo": 12, "bar":[1,2,{"barp":12, "blip":13}]}'
            output: ['true']
          - program: 'contains({foo: 12, bar: [{barp: 15}]})'
            input: '{"foo": 12, "bar":[1,2,{"barp":12, "blip":13}]}'
            output: ['false']

      - title: "`indices(s)`"
        body: |

          Outputs an array containing the indices in `.` where `s`
          occurs.  The input may be an array, in which case if `s` is an
          array then the indices output will be those where all elements
          in `.` match those of `s`.

        examples:
          - program: 'indices(", ")'
            input: '"a,b, cd, efg, hijk"'
            output: ['[3,7,12]']
          - program: 'indices(1)'
            input: '[0,1,2,1,3,1,4]'
            output: ['[1,3,5]']
          - program: 'indices([1,2])'
            input: '[0,1,2,3,1,4,2,5,1,2,6,7]'
            output: ['[1,8]']

      - title: "`index(s)`, `rindex(s)`"
        body: |

          Outputs the index of the first (`index`) or last (`rindex`)
          occurrence of `s` in the input.

        examples:
          - program: 'index(", ")'
            input: '"a,b, cd, efg, hijk"'
            output: ['3']
          - program: 'index(1)'
            input: '[0,1,2,1,3,1,4]'
            output: ['1']
          - program: 'index([1,2])'
            input: '[0,1,2,3,1,4,2,5,1,2,6,7]'
            output: ['1']
          - program: 'rindex(", ")'
            input: '"a,b, cd, efg, hijk"'
            output: ['12']
          - program: 'rindex(1)'
            input: '[0,1,2,1,3,1,4]'
            output: ['5']
          - program: 'rindex([1,2])'
            input: '[0,1,2,3,1,4,2,5,1,2,6,7]'
            output: ['8']

      - title: "`inside`"
        body: |

          The filter `inside(b)` will produce true if the input is
          completely contained within b. It is, essentially, an
          inversed version of `contains`.

        examples:
          - program: 'inside("foobar")'
            input: '"bar"'
            output: ['true']
          - program: 'inside(["foobar", "foobaz", "blarp"])'
            input: '["baz", "bar"]'
            output: ['true']
          - program: 'inside(["foobar", "foobaz", "blarp"])'
            input: '["bazzzzz", "bar"]'
            output: ['false']
          - program: 'inside({"foo": 12, "bar":[1,2,{"barp":12, "blip":13}]})'
            input: '{"foo": 12, "bar": [{"barp": 12}]}'
            output: ['true']
          - program: 'inside({"foo": 12, "bar":[1,2,{"barp":12, "blip":13}]})'
            input: '{"foo": 12, "bar": [{"barp": 15}]}'
            output: ['false']

      - title: "`startswith(str)`"
        body: |

          Outputs `true` if . starts with the given string argument.

        examples:
          - program: '[.[]|startswith("foo")]'
            input: '["fo", "foo", "barfoo", "foobar", "barfoob"]'
            output: ['[false, true, false, true, false]']

      - title: "`endswith(str)`"
        body: |

          Outputs `true` if . ends with the given string argument.

        examples:
          - program: '[.[]|endswith("foo")]'
            input: '["foobar", "barfoo"]'
            output: ['[false, true]']

      - title: "`combinations`, `combinations(n)`"
        body: |

          Outputs all combinations of the elements of the arrays in the
          input array. If given an argument `n`, it outputs all combinations
          of `n` repetitions of the input array.

        examples:
          - program: 'combinations'
            input: '[[1,2], [3, 4]]'
            output: ['[1, 3]', '[1, 4]', '[2, 3]', '[2, 4]']
          - program: 'combinations(2)'
            input: '[0, 1]'
            output: ['[0, 0]', '[0, 1]', '[1, 0]', '[1, 1]']

      - title: "`ltrimstr(str)`"
        body: |

          Outputs its input with the given prefix string removed, if it
          starts with it.

        examples:
          - program: '[.[]|ltrimstr("foo")]'
            input: '["fo", "foo", "barfoo", "foobar", "afoo"]'
            output: ['["fo","","barfoo","bar","afoo"]']

      - title: "`rtrimstr(str)`"
        body: |

          Outputs its input with the given suffix string removed, if it
          ends with it.

        examples:
          - program: '[.[]|rtrimstr("foo")]'
            input: '["fo", "foo", "barfoo", "foobar", "foob"]'
            output: ['["fo","","bar","foobar","foob"]']

      - title: "`trimstr(str)`"
        body: |

          Outputs its input with the given string removed at both ends, if it
          starts or ends with it.

        examples:
          - program: '[.[]|trimstr("foo")]'
            input: '["fo", "foo", "barfoo", "foobarfoo", "foob"]'
            output: ['["fo","","bar","bar","b"]']

      - title: "`trim`, `ltrim`, `rtrim`"
        body: |

          `trim` trims both leading and trailing whitespace.

          `ltrim` trims only leading (left side) whitespace.

          `rtrim` trims only trailing (right side) whitespace.

          Whitespace characters are the usual `" "`, `"\n"` `"\t"`, `"\r"`
          and also all characters in the Unicode character database with the
          whitespace property. Note that what considers whitespace might
          change in the future.

        examples:
          - program: 'trim, ltrim, rtrim'
            input: '" abc "'
            output: ['"abc"', '"abc "', '" abc"']

      - title: "`explode`"
        body: |

          Converts an input string into an array of the string's
          codepoint numbers.

        examples:
          - program: 'explode'
            input: '"foobar"'
            output: ['[102,111,111,98,97,114]']

      - title: "`implode`"
        body: |

          The inverse of explode.

        examples:
          - program: 'implode'
            input: '[65, 66, 67]'
            output: ['"ABC"']

      - title: "`split(str)`"
        body: |

          Splits an input string on the separator argument.

          `split` can also split on regex matches when called with
          two arguments (see the regular expressions section below).

        examples:
          - program: 'split(", ")'
            input: '"a, b,c,d, e, "'
            output: ['["a","b,c,d","e",""]']

      - title: "`join(str)`"
        body: |

          Joins the array of elements given as input, using the
          argument as separator. It is the inverse of `split`: that is,
          running `split("foo") | join("foo")` over any input string
          returns said input string.

          Numbers and booleans in the input are converted to strings.
          Null values are treated as empty strings. Arrays and objects
          in the input are not supported.

        examples:
          - program: 'join(", ")'
            input: '["a","b,c,d","e"]'
            output: ['"a, b,c,d, e"']
          - program: 'join(" ")'
            input: '["a",1,2.3,true,null,false]'
            output: ['"a 1 2.3 true  false"']

      - title: "`ascii_downcase`, `ascii_upcase`"
        body: |

          Emit a copy of the input string with its alphabetic characters (a-z and A-Z)
          converted to the specified case.

        examples:
          - program: 'ascii_upcase'
            input: '"useful but not for é"'
            output: ['"USEFUL BUT NOT FOR é"']

      - title: "`while(cond; update)`"
        body: |

          The `while(cond; update)` function allows you to repeatedly
          apply an update to `.` until `cond` is false.

          Note that `while(cond; update)` is internally defined as a
          recursive jq function.  Recursive calls within `while` will
          not consume additional memory if `update` produces at most one
          output for each input.  See advanced topics below.

        examples:
          - program: '[while(.<100; .*2)]'
            input: '1'
            output: ['[1,2,4,8,16,32,64]']

      - title: "`repeat(exp)`"
        body: |

          The `repeat(exp)` function allows you to repeatedly
          apply expression `exp` to `.` until an error is raised.

          Note that `repeat(exp)` is internally defined as a
          recursive jq function.  Recursive calls within `repeat` will
          not consume additional memory if `exp` produces at most one
          output for each input.  See advanced topics below.

        examples:
          - program: '[repeat(.*2, error)?]'
            input: '1'
            output: ['[2]']

      - title: "`until(cond; next)`"
        body: |

          The `until(cond; next)` function allows you to repeatedly
          apply the expression `next`, initially to `.` then to its own
          output, until `cond` is true.  For example, this can be used
          to implement a factorial function (see below).

          Note that `until(cond; next)` is internally defined as a
          recursive jq function.  Recursive calls within `until()` will
          not consume additional memory if `next` produces at most one
          output for each input.  See advanced topics below.

        examples:
          - program: '[.,1]|until(.[0] < 1; [.[0] - 1, .[1] * .[0]])|.[1]'
            input: '4'
            output: ['24']


      - title: "`recurse(f)`, `recurse`, `recurse(f; condition)`"
        body: |

          The `recurse(f)` function allows you to search through a
          recursive structure, and extract interesting data from all
          levels. Suppose your input represents a filesystem:

              {"name": "/", "children": [
                {"name": "/bin", "children": [
                  {"name": "/bin/ls", "children": []},
                  {"name": "/bin/sh", "children": []}]},
                {"name": "/home", "children": [
                  {"name": "/home/stephen", "children": [
                    {"name": "/home/stephen/jq", "children": []}]}]}]}

          Now suppose you want to extract all of the filenames
          present. You need to retrieve `.name`, `.children[].name`,
          `.children[].children[].name`, and so on. You can do this
          with:

              recurse(.children[]) | .name

          When called without an argument, `recurse` is equivalent to
          `recurse(.[]?)`.

          `recurse(f)` is identical to `recurse(f; true)` and can be
          used without concerns about recursion depth.

          `recurse(f; condition)` is a generator which begins by
          emitting . and then emits in turn .|f, .|f|f, .|f|f|f, ...  so long
          as the computed value satisfies the condition. For example,
          to generate all the integers, at least in principle, one
          could write `recurse(.+1; true)`.

          The recursive calls in `recurse` will not consume additional
          memory whenever `f` produces at most a single output for each
          input.

        examples:
          - program: 'recurse(.foo[])'
            input: '{"foo":[{"foo": []}, {"foo":[{"foo":[]}]}]}'
            output:
              - '{"foo":[{"foo":[]},{"foo":[{"foo":[]}]}]}'
              - '{"foo":[]}'
              - '{"foo":[{"foo":[]}]}'
              - '{"foo":[]}'

          - program: 'recurse'
            input: '{"a":0,"b":[1]}'
            output:
              - '{"a":0,"b":[1]}'
              - '0'
              - '[1]'
              - '1'

          - program: 'recurse(. * .; . < 20)'
            input: '2'
            output: ['2', '4', '16']

      - title: "`walk(f)`"
        body: |

          The `walk(f)` function applies f recursively to every
          component of the input entity.  When an array is
          encountered, f is first applied to its elements and then to
          the array itself; when an object is encountered, f is first
          applied to all the values and then to the object.  In
          practice, f will usually test the type of its input, as
          illustrated in the following examples.  The first example
          highlights the usefulness of processing the elements of an
          array of arrays before processing the array itself.  The second
          example shows how all the keys of all the objects within the
          input can be considered for alteration.

        examples:
          - program: 'walk(if type == "array" then sort else . end)'
            input: '[[4, 1, 7], [8, 5, 2], [3, 6, 9]]'
            output:
              - '[[1,4,7],[2,5,8],[3,6,9]]'

          - program: 'walk( if type == "object" then with_entries( .key |= sub( "^_+"; "") ) else . end )'
            input: '[ { "_a": { "__b": 2 } } ]'
            output:
              - '[{"a":{"b":2}}]'

      - title: "`have_literal_numbers`"
        body: |

          This builtin returns true if jq's build configuration
          includes support for preservation of input number literals.

      - title: "`have_decnum`"
        body: |

          This builtin returns true if jq was built with "decnum",
          which is the current literal number preserving numeric
          backend implementation for jq.

      - title: "`$JQ_BUILD_CONFIGURATION`"
        body: |

          This builtin binding shows the jq executable's build
          configuration.  Its value has no particular format, but
          it can be expected to be at least the `./configure`
          command-line arguments, and may be enriched in the
          future to include the version strings for the build
          tooling used.

          Note that this can be overridden in the command-line
          with `--arg` and related options.

      - title: "`$ENV`, `env`"
        body: |

          `$ENV` is an object representing the environment variables as
          set when the jq program started.

          `env` outputs an object representing jq's current environment.

          At the moment there is no builtin for setting environment
          variables.

        examples:
          - program: '$ENV.PAGER'
            input: 'null'
            output: ['"less"']

          - program: 'env.PAGER'
            input: 'null'
            output: ['"less"']

      - title: "`transpose`"
        body: |

          Transpose a possibly jagged matrix (an array of arrays).
          Rows are padded with nulls so the result is always rectangular.

        examples:
          - program: 'transpose'
            input: '[[1], [2,3]]'
            output: ['[[1,2],[null,3]]']

      - title: "`bsearch(x)`"
        body: |

          `bsearch(x)` conducts a binary search for x in the input
          array.  If the input is sorted and contains x, then
          `bsearch(x)` will return its index in the array; otherwise, if
          the array is sorted, it will return (-1 - ix) where ix is an
          insertion point such that the array would still be sorted
          after the insertion of x at ix.  If the array is not sorted,
          `bsearch(x)` will return an integer that is probably of no
          interest.

        examples:
          - program: 'bsearch(0)'
            input: '[0,1]'
            output: ['0']
          - program: 'bsearch(0)'
            input: '[1,2,3]'
            output: ['-1']
          - program: 'bsearch(4) as $ix | if $ix < 0 then .[-(1+$ix)] = 4 else . end'
            input: '[1,2,3]'
            output: ['[1,2,3,4]']

      - title: "String interpolation: `\\(exp)`"
        body: |

          Inside a string, you can put an expression inside parens
          after a backslash. Whatever the expression returns will be
          interpolated into the string.

        examples:
          - program: '"The input was \(.), which is one less than \(.+1)"'
            input: '42'
            output: ['"The input was 42, which is one less than 43"']

      - title: "Convert to/from JSON"
        body: |

          The `tojson` and `fromjson` builtins dump values as JSON texts
          or parse JSON texts into values, respectively.  The `tojson`
          builtin differs from `tostring` in that `tostring` returns strings
          unmodified, while `tojson` encodes strings as JSON strings.

        examples:
          - program: '[.[]|tostring]'
            input: '[1, "foo", ["foo"]]'
            output: ['["1","foo","[\"foo\"]"]']
          - program: '[.[]|tojson]'
            input: '[1, "foo", ["foo"]]'
            output: ['["1","\"foo\"","[\"foo\"]"]']
          - program: '[.[]|tojson|fromjson]'
            input: '[1, "foo", ["foo"]]'
            output: ['[1,"foo",["foo"]]']

      - title: "Format strings and escaping"
        body: |

          The `@foo` syntax is used to format and escape strings,
          which is useful for building URLs, documents in a language
          like HTML or XML, and so forth. `@foo` can be used as a
          filter on its own, the possible escapings are:

          * `@text`:

            Calls `tostring`, see that function for details.

          * `@json`:

            Serializes the input as JSON.

          * `@html`:

            Applies HTML/XML escaping, by mapping the characters
            `<>&'"` to their entity equivalents `&lt;`, `&gt;`,
            `&amp;`, `&apos;`, `&quot;`.

          * `@uri`:

            Applies percent-encoding, by mapping all reserved URI
            characters to a `%XX` sequence.

          * `@urid`:

            The inverse of `@uri`, applies percent-decoding, by mapping
            all `%XX` sequences to their corresponding URI characters.

          * `@csv`:

            The input must be an array, and it is rendered as CSV
            with double quotes for strings, and quotes escaped by
            repetition.

          * `@tsv`:

            The input must be an array, and it is rendered as TSV
            (tab-separated values). Each input array will be printed as
            a single line. Fields are separated by a single
            tab (ascii `0x09`). Input characters line-feed (ascii `0x0a`),
            carriage-return (ascii `0x0d`), tab (ascii `0x09`) and
            backslash (ascii `0x5c`) will be output as escape sequences
            `\n`, `\r`, `\t`, `\\` respectively.

          * `@sh`:

            The input is escaped suitable for use in a command-line
            for a POSIX shell. If the input is an array, the output
            will be a series of space-separated strings.

          * `@base64`:

            The input is converted to base64 as specified by RFC 4648.

          * `@base64d`:

            The inverse of `@base64`, input is decoded as specified by RFC 4648.
            Note\: If the decoded string is not UTF-8, the results are undefined.

          This syntax can be combined with string interpolation in a
          useful way. You can follow a `@foo` token with a string
          literal. The contents of the string literal will *not* be
          escaped. However, all interpolations made inside that string
          literal will be escaped. For instance,

              @uri "https://www.google.com/search?q=\(.search)"

          will produce the following output for the input
          `{"search":"what is jq?"}`:

              "https://www.google.com/search?q=what%20is%20jq%3F"

          Note that the slashes, question mark, etc. in the URL are
          not escaped, as they were part of the string literal.

        examples:
          - program: '@html'
            input: '"This works if x < y"'
            output: ['"This works if x &lt; y"']

          - program: '@sh "echo \(.)"'
            input: "\"O'Hara's Ale\""
            output: ["\"echo 'O'\\\\''Hara'\\\\''s Ale'\""]

          - program: '@base64'
            input: '"This is a message"'
            output: ['"VGhpcyBpcyBhIG1lc3NhZ2U="']

          - program: '@base64d'
            input: '"VGhpcyBpcyBhIG1lc3NhZ2U="'
            output: ['"This is a message"']

      - title: "Dates"
        body: |

          jq provides some basic date handling functionality, with some
          high-level and low-level builtins.  In all cases these
          builtins deal exclusively with time in UTC.

          The `fromdateiso8601` builtin parses datetimes in the ISO 8601
          format to a number of seconds since the Unix epoch
          (1970-01-01T00:00:00Z).  The `todateiso8601` builtin does the
          inverse.

          The `fromdate` builtin parses datetime strings.  Currently
          `fromdate` only supports ISO 8601 datetime strings, but in the
          future it will attempt to parse datetime strings in more
          formats.

          The `todate` builtin is an alias for `todateiso8601`.

          The `now` builtin outputs the current time, in seconds since
          the Unix epoch.

          Low-level jq interfaces to the C-library time functions are
          also provided: `strptime`, `strftime`, `strflocaltime`,
          `mktime`, `gmtime`, and `localtime`.  Refer to your host
          operating system's documentation for the format strings used
          by `strptime` and `strftime`.  Note: these are not necessarily
          stable interfaces in jq, particularly as to their localization
          functionality.

          The `gmtime` builtin consumes a number of seconds since the
          Unix epoch and outputs a "broken down time" representation of
          Greenwich Mean Time as an array of numbers representing
          (in this order): the year, the month (zero-based), the day of
          the month (one-based), the hour of the day, the minute of the
          hour, the second of the minute, the day of the week, and the
          day of the year -- all one-based unless otherwise stated.  The
          day of the week number may be wrong on some systems for dates
          before March 1st 1900, or after December 31 2099.

          The `localtime` builtin works like the `gmtime` builtin, but
          using the local timezone setting.

          The `mktime` builtin consumes "broken down time"
          representations of time output by `gmtime` and `strptime`.

          The `strptime(fmt)` builtin parses input strings matching the
          `fmt` argument.  The output is in the "broken down time"
          representation consumed by `mktime` and output by `gmtime`.

          The `strftime(fmt)` builtin formats a time (GMT) with the
          given format.  The `strflocaltime` does the same, but using
          the local timezone setting.

          The format strings for `strptime` and `strftime` are described
          in typical C library documentation.  The format string for ISO
          8601 datetime is `"%Y-%m-%dT%H:%M:%SZ"`.

          jq may not support some or all of this date functionality on
          some systems. In particular, the `%u` and `%j` specifiers for
          `strptime(fmt)` are not supported on macOS.

        examples:
          - program: 'fromdate'
            input: '"2015-03-05T23:51:47Z"'
            output: ['1425599507']

          - program: 'strptime("%Y-%m-%dT%H:%M:%SZ")'
            input: '"2015-03-05T23:51:47Z"'
            output: ['[2015,2,5,23,51,47,4,63]']

          - program: 'strptime("%Y-%m-%dT%H:%M:%SZ")|mktime'
            input: '"2015-03-05T23:51:47Z"'
            output: ['1425599507']

      - title: "SQL-Style Operators"
        body: |

          jq provides a few SQL-style operators.

          * `INDEX(stream; index_expression)`:

            This builtin produces an object whose keys are computed by
            the given index expression applied to each value from the
            given stream.

          * `JOIN($idx; stream; idx_expr; join_expr)`:

            This builtin joins the values from the given stream to the
            given index.  The index's keys are computed by applying the
            given index expression to each value from the given stream.
            An array of the value in the stream and the corresponding
            value from the index is fed to the given join expression to
            produce each result.

          * `JOIN($idx; stream; idx_expr)`:

            Same as `JOIN($idx; stream; idx_expr; .)`.

          * `JOIN($idx; idx_expr)`:

            This builtin joins the input `.` to the given index, applying
            the given index expression to `.` to compute the index key.
            The join operation is as described above.

          * `IN(s)`:

            This builtin outputs `true` if `.` appears in the given
            stream, otherwise it outputs `false`.

          * `IN(source; s)`:

            This builtin outputs `true` if any value in the source stream
            appears in the second stream, otherwise it outputs `false`.

      - title: "`builtins`"
        body: |

          Returns a list of all builtin functions in the format `name/arity`.
          Since functions with the same name but different arities are considered
          separate functions, `all/0`, `all/1`, and `all/2` would all be present
          in the list.

  - title: Conditionals and Comparisons
    entries:
      - title: "`==`, `!=`"
        body: |

          The expression 'a == b' will produce 'true' if the results of evaluating
          a and b are equal (that is, if they represent equivalent JSON values) and
          'false' otherwise. In particular, strings are never considered equal
          to numbers.  In checking for the equality of JSON objects, the ordering of keys
          is irrelevant.  If you're coming from JavaScript, please note that jq's `==` is like
          JavaScript's `===`, the "strict equality" operator.

          != is "not equal", and 'a != b' returns the opposite value of 'a == b'

        examples:
          - program: '. == false'
            input: 'null'
            output: ['false']

          - program: '. == {"b": {"d": (4 + 1e-20), "c": 3}, "a":1}'
            input: '{"a":1, "b": {"c": 3, "d": 4}}'
            output: ['true']

          - program: '.[] == 1'
            input: '[1, 1.0, "1", "banana"]'
            output: ['true', 'true', 'false', 'false']

      - title: if-then-else-end
        body: |

          `if A then B else C end` will act the same as `B` if `A`
          produces a value other than false or null, but act the same
          as `C` otherwise.

          `if A then B end` is the same as `if A then B else .  end`.
          That is, the `else` branch is optional, and if absent is the
          same as `.`. This also applies to `elif` with absent ending `else` branch.

          Checking for false or null is a simpler notion of
          "truthiness" than is found in JavaScript or Python, but it
          means that you'll sometimes have to be more explicit about
          the condition you want.  You can't test whether, e.g. a
          string is empty using `if .name then A else B end`; you'll
          need something like `if .name == "" then A else B end` instead.

          If the condition `A` produces multiple results, then `B` is evaluated
          once for each result that is not false or null, and `C` is evaluated
          once for each false or null.

          More cases can be added to an if using `elif A then B` syntax.

        examples:
          - program: |-
              if . == 0 then
                "zero"
              elif . == 1 then
                "one"
              else
                "many"
              end
            input: '2'
            output: ['"many"']

      - title: "`>`, `>=`, `<=`, `<`"
        body: |

          The comparison operators `>`, `>=`, `<=`, `<` return whether
          their left argument is greater than, greater than or equal
          to, less than or equal to or less than their right argument
          (respectively).

          The ordering is the same as that described for `sort`, above.

        examples:
          - program: '. < 5'
            input: '2'
            output: ['true']

      - title: "`and`, `or`, `not`"
        body: |

          jq supports the normal Boolean operators `and`, `or`, `not`.
          They have the same standard of truth as if expressions -
          `false` and `null` are considered "false values", and
          anything else is a "true value".

          If an operand of one of these operators produces multiple
          results, the operator itself will produce a result for each input.

          `not` is in fact a builtin function rather than an operator,
          so it is called as a filter to which things can be piped
          rather than with special syntax, as in `.foo and .bar |
          not`.

          These three only produce the values `true` and `false`, and
          so are only useful for genuine Boolean operations, rather
          than the common Perl/Python/Ruby idiom of
          "value_that_may_be_null or default". If you want to use this
          form of "or", picking between two values rather than
          evaluating a condition, see the `//` operator below.

        examples:
          - program: '42 and "a string"'
            input: 'null'
            output: ['true']
          - program: '(true, false) or false'
            input: 'null'
            output: ['true', 'false']
          - program: '(true, true) and (true, false)'
            input: 'null'
            output: ['true', 'false', 'true', 'false']
          - program: '[true, false | not]'
            input: 'null'
            output: ['[false, true]']

      - title: "Alternative operator: `//`"
        body: |

          The `//` operator produces all the values of its left-hand
          side that are neither `false` nor `null`. If the
          left-hand side produces no values other than `false` or
          `null`, then `//` produces all the values of its right-hand
          side.

          A filter of the form `a // b` produces all the results of
          `a` that are not `false` or `null`.  If `a` produces no
          results, or no results other than `false` or `null`, then `a
          // b` produces the results of `b`.

          This is useful for providing defaults: `.foo // 1` will
          evaluate to `1` if there's no `.foo` element in the
          input. It's similar to how `or` is sometimes used in Python
          (jq's `or` operator is reserved for strictly Boolean
          operations).

          Note: `some_generator // defaults_here` is not the same
          as `some_generator | . // defaults_here`.  The latter will
          produce default values for all non-`false`, non-`null`
          values of the left-hand side, while the former will not.
          Precedence rules can make this confusing.  For example, in
          `false, 1 // 2` the left-hand side of `//` is `1`, not
          `false, 1` -- `false, 1 // 2` parses the same way as `false,
          (1 // 2)`.  In `(false, null, 1) | . // 42` the left-hand
          side of `//` is `.`, which always produces just one value,
          while in `(false, null, 1) // 42` the left-hand side is a
          generator of three values, and since it produces a
          value other `false` and `null`, the default `42` is not
          produced.

        examples:
          - program: 'empty // 42'
            input: 'null'
            output: ['42']
          - program: '.foo // 42'
            input: '{"foo": 19}'
            output: ['19']
          - program: '.foo // 42'
            input: '{}'
            output: ['42']
          - program: '(false, null, 1) // 42'
            input: 'null'
            output: ['1']
          - program: '(false, null, 1) | . // 42'
            input: 'null'
            output: ['42', '42', '1']

      - title: try-catch
        body: |

          Errors can be caught by using `try EXP catch EXP`.  The first
          expression is executed, and if it fails then the second is
          executed with the error message.  The output of the handler,
          if any, is output as if it had been the output of the
          expression to try.

          The `try EXP` form uses `empty` as the exception handler.

        examples:
          - program: 'try .a catch ". is not an object"'
            input: 'true'
            output: ['". is not an object"']
          - program: '[.[]|try .a]'
            input: '[{}, true, {"a":1}]'
            output: ['[null, 1]']
          - program: 'try error("some exception") catch .'
            input: 'true'
            output: ['"some exception"']

      - title: Breaking out of control structures
        body: |

          A convenient use of try/catch is to break out of control
          structures like `reduce`, `foreach`, `while`, and so on.

          For example:

              # Repeat an expression until it raises "break" as an
              # error, then stop repeating without re-raising the error.
              # But if the error caught is not "break" then re-raise it.
              try repeat(exp) catch if .=="break" then empty else error

          jq has a syntax for named lexical labels to "break" or "go (back) to":

              label $out | ... break $out ...

          The `break $label_name` expression will cause the program to
          act as though the nearest (to the left) `label $label_name`
          produced `empty`.

          The relationship between the `break` and corresponding `label`
          is lexical: the label has to be "visible" from the break.

          To break out of a `reduce`, for example:

              label $out | reduce .[] as $item (null; if .==false then break $out else ... end)

          The following jq program produces a syntax error:

              break $out

          because no label `$out` is visible.

      - title: "Error Suppression / Optional Operator: `?`"
        body: |

          The `?` operator, used as `EXP?`, is shorthand for `try EXP`.

        examples:
          - program: '[.[] | .a?]'
            input: '[{}, true, {"a":1}]'
            output: ['[null, 1]']
          - program: '[.[] | tonumber?]'
            input: '["1", "invalid", "3", 4]'
            output: ['[1, 3, 4]']

  - title: Regular expressions
    body: |

      jq uses the
      [Oniguruma regular expression library](https://github.com/kkos/oniguruma/blob/master/doc/RE),
      as do PHP, TextMate, Sublime Text, etc, so the
      description here will focus on jq specifics.

      Oniguruma supports several flavors of regular expression, so it is important to know
      that jq uses the ["Perl NG" (Perl with named groups)](https://github.com/kkos/oniguruma/blob/master/doc/SYNTAX.md) flavor.

      The jq regex filters are defined so that they can be used using
      one of these patterns:

          STRING | FILTER(REGEX)
          STRING | FILTER(REGEX; FLAGS)
          STRING | FILTER([REGEX])
          STRING | FILTER([REGEX, FLAGS])

      where:

      * STRING, REGEX, and FLAGS are jq strings and subject to jq string interpolation;
      * REGEX, after string interpolation, should be a valid regular expression;
      * FILTER is one of `test`, `match`, or `capture`, as described below.

      Since REGEX must evaluate to a JSON string, some characters that are needed
      to form a regular expression must be escaped. For example, the regular expression
      `\s` signifying a whitespace character would be written as `"\\s"`.

      FLAGS is a string consisting of one of more of the supported flags:

      * `g` - Global search (find all matches, not just the first)
      * `i` - Case insensitive search
      * `m` - Multi line mode (`.` will match newlines)
      * `n` - Ignore empty matches
      * `p` - Both s and m modes are enabled
      * `s` - Single line mode (`^` -> `\A`, `$` -> `\Z`)
      * `l` - Find longest possible matches
      * `x` - Extended regex format (ignore whitespace and comments)

      To match a whitespace with the `x` flag, use `\s`, e.g.

          jq -n '"a b" | test("a\\sb"; "x")'

      Note that certain flags may also be specified within REGEX, e.g.

          jq -n '("test", "TEst", "teST", "TEST") | test("(?i)te(?-i)st")'

      evaluates to: `true`, `true`, `false`, `false`.

    entries:
      - title: "`test(val)`, `test(regex; flags)`"
        body: |

          Like `match`, but does not return match objects, only `true` or `false`
          for whether or not the regex matches the input.

        examples:
          - program: 'test("foo")'
            input: '"foo"'
            output: ['true']
          - program: '.[] | test("a b c # spaces are ignored"; "ix")'
            input: '["xabcd", "ABC"]'
            output: ['true', 'true']

      - title: "`match(val)`, `match(regex; flags)`"
        body: |

          **match** outputs an object for each match it finds.  Matches have
          the following fields:

          * `offset` - offset in UTF-8 codepoints from the beginning of the input
          * `length` - length in UTF-8 codepoints of the match
          * `string` - the string that it matched
          * `captures` - an array of objects representing capturing groups.

          Capturing group objects have the following fields:

          * `offset` - offset in UTF-8 codepoints from the beginning of the input
          * `length` - length in UTF-8 codepoints of this capturing group
          * `string` - the string that was captured
          * `name` - the name of the capturing group (or `null` if it was unnamed)

          Capturing groups that did not match anything return an offset of -1

        examples:
          - program: 'match("(abc)+"; "g")'
            input: '"abc abc"'
            output:
             - '{"offset": 0, "length": 3, "string": "abc", "captures": [{"offset": 0, "length": 3, "string": "abc", "name": null}]}'
             - '{"offset": 4, "length": 3, "string": "abc", "captures": [{"offset": 4, "length": 3, "string": "abc", "name": null}]}'
          - program: 'match("foo")'
            input: '"foo bar foo"'
            output: ['{"offset": 0, "length": 3, "string": "foo", "captures": []}']
          - program: 'match(["foo", "ig"])'
            input: '"foo bar FOO"'
            output:
             - '{"offset": 0, "length": 3, "string": "foo", "captures": []}'
             - '{"offset": 8, "length": 3, "string": "FOO", "captures": []}'
          - program: 'match("foo (?<bar123>bar)? foo"; "ig")'
            input: '"foo bar foo foo  foo"'
            output:
             - '{"offset": 0, "length": 11, "string": "foo bar foo", "captures": [{"offset": 4, "length": 3, "string": "bar", "name": "bar123"}]}'
             - '{"offset": 12, "length": 8, "string": "foo  foo", "captures": [{"offset": -1, "length": 0, "string": null, "name": "bar123"}]}'

          - program: '[ match("."; "g")] | length'
            input: '"abc"'
            output: ['3']


      - title: "`capture(val)`, `capture(regex; flags)`"
        body: |

         Collects the named captures in a JSON object, with the name
         of each capture as the key, and the matched string as the
         corresponding value.

        examples:
          - program: 'capture("(?<a>[a-z]+)-(?<n>[0-9]+)")'
            input: '"xyzzy-14"'
            output: ['{ "a": "xyzzy", "n": "14" }']

      - title: "`scan(regex)`, `scan(regex; flags)`"
        body: |

          Emit a stream of the non-overlapping substrings of the input
          that match the regex in accordance with the flags, if any
          have been specified.  If there is no match, the stream is empty.
          To capture all the matches for each input string, use the idiom
          `[ expr ]`, e.g. `[ scan(regex) ]`.  If the regex contains capturing
          groups, the filter emits a stream of arrays, each of which contains
          the captured strings.

        examples:
          - program: 'scan("c")'
            input: '"abcdefabc"'
            output: ['"c"', '"c"']
          - program: 'scan("(a+)(b+)")'
            input: '"abaabbaaabbb"'
            output: ['["a","b"]', '["aa","bb"]', '["aaa","bbb"]']

      - title: "`split(regex; flags)`"
        body: |

          Splits an input string on each regex match.

          For backwards compatibility, when called with a single argument,
          `split` splits on a string, not a regex.

        examples:
          - program: 'split(", *"; null)'
            input: '"ab,cd, ef"'
            output: ['["ab","cd","ef"]']


      - title: "`splits(regex)`, `splits(regex; flags)`"
        body: |

          These provide the same results as their `split` counterparts,
          but as a stream instead of an array.

        examples:
          - program: 'splits(", *")'
            input: '"ab,cd,   ef, gh"'
            output: ['"ab"','"cd"','"ef"','"gh"']
          - program: 'splits(",? *"; "n")'
            input: '"ab,cd ef,  gh"'
            output: ['"ab"','"cd"','"ef"','"gh"']

      - title: "`sub(regex; tostring)`, `sub(regex; tostring; flags)`"
        body: |

          Emit the string obtained by replacing the first match of
          regex in the input string with `tostring`, after
          interpolation.  `tostring` should be a jq string or a stream
          of such strings, each of which may contain references to
          named captures. The named captures are, in effect, presented
          as a JSON object (as constructed by `capture`) to
          `tostring`, so a reference to a captured variable named "x"
          would take the form: `"\(.x)"`.

        examples:
          - program: 'sub("[^a-z]*(?<x>[a-z]+)"; "Z\(.x)"; "g")'
            input: '"123abc456def"'
            output: ['"ZabcZdef"']

          - program: '[sub("(?<a>.)"; "\(.a|ascii_upcase)", "\(.a|ascii_downcase)")]'
            input: '"aB"'
            output: ['["AB","aB"]']

      - title: "`gsub(regex; tostring)`, `gsub(regex; tostring; flags)`"
        body: |

          `gsub` is like `sub` but all the non-overlapping occurrences of the regex are
          replaced by `tostring`, after interpolation. If the second argument is a stream
          of jq strings, then `gsub` will produce a corresponding stream of JSON strings.

        examples:
          - program: 'gsub("(?<x>.)[^a]*"; "+\(.x)-")'
            input: '"Abcabc"'
            output: ['"+A-+a-"']

          - program: '[gsub("p"; "a", "b")]'
            input: '"p"'
            output: ['["a","b"]']


  - title: Advanced features
    body: |
      Variables are an absolute necessity in most programming languages, but
      they're relegated to an "advanced feature" in jq.

      In most languages, variables are the only means of passing around
      data. If you calculate a value, and you want to use it more than once,
      you'll need to store it in a variable. To pass a value to another part
      of the program, you'll need that part of the program to define a
      variable (as a function parameter, object member, or whatever) in
      which to place the data.

      It is also possible to define functions in jq, although this is
      is a feature whose biggest use is defining jq's standard library
      (many jq functions such as `map` and `select` are in fact written
      in jq).

      jq has reduction operators, which are very powerful but a bit
      tricky.  Again, these are mostly used internally, to define some
      useful bits of jq's standard library.

      It may not be obvious at first, but jq is all about generators
      (yes, as often found in other languages).  Some utilities are
      provided to help deal with generators.

      Some minimal I/O support (besides reading JSON from standard
      input, and writing JSON to standard output) is available.

      Finally, there is a module/library system.

    entries:
      - title: "Variable / Symbolic Binding Operator: `... as $identifier | ...`"
        body: |

          In jq, all filters have an input and an output, so manual
          plumbing is not necessary to pass a value from one part of a program
          to the next. Many expressions, for instance `a + b`, pass their input
          to two distinct subexpressions (here `a` and `b` are both passed the
          same input), so variables aren't usually necessary in order to use a
          value twice.

          For instance, calculating the average value of an array of numbers
          requires a few variables in most languages - at least one to hold the
          array, perhaps one for each element or for a loop counter. In jq, it's
          simply `add / length` - the `add` expression is given the array and
          produces its sum, and the `length` expression is given the array and
          produces its length.

          So, there's generally a cleaner way to solve most problems in jq than
          defining variables. Still, sometimes they do make things easier, so jq
          lets you define variables using `expression as $variable`. All
          variable names start with `$`. Here's a slightly uglier version of the
          array-averaging example:

              length as $array_length | add / $array_length

          We'll need a more complicated problem to find a situation where using
          variables actually makes our lives easier.


          Suppose we have an array of blog posts, with "author" and "title"
          fields, and another object which is used to map author usernames to
          real names. Our input looks like:

              {"posts": [{"title": "First post", "author": "anon"},
                         {"title": "A well-written article", "author": "person1"}],
               "realnames": {"anon": "Anonymous Coward",
                             "person1": "Person McPherson"}}

          We want to produce the posts with the author field containing a real
          name, as in:

              {"title": "First post", "author": "Anonymous Coward"}
              {"title": "A well-written article", "author": "Person McPherson"}

          We use a variable, `$names`, to store the realnames object, so that we
          can refer to it later when looking up author usernames:

              .realnames as $names | .posts[] | {title, author: $names[.author]}

          The expression `exp as $x | ...` means: for each value of expression
          `exp`, run the rest of the pipeline with the entire original input, and
          with `$x` set to that value.  Thus `as` functions as something of a
          foreach loop.

          Just as `{foo}` is a handy way of writing `{foo: .foo}`, so
          `{$foo}` is a handy way of writing `{foo: $foo}`.

          Multiple variables may be declared using a single `as` expression by
          providing a pattern that matches the structure of the input
          (this is known as "destructuring"):

              . as {realnames: $names, posts: [$first, $second]} | ...

          The variable declarations in array patterns (e.g., `. as
          [$first, $second]`) bind to the elements of the array in from
          the element at index zero on up, in order.  When there is no
          value at the index for an array pattern element, `null` is
          bound to that variable.

          Variables are scoped over the rest of the expression that defines
          them, so

              .realnames as $names | (.posts[] | {title, author: $names[.author]})

          will work, but

              (.realnames as $names | .posts[]) | {title, author: $names[.author]}

          won't.

          For programming language theorists, it's more accurate to
          say that jq variables are lexically-scoped bindings.  In
          particular there's no way to change the value of a binding;
          one can only setup a new binding with the same name, but which
          will not be visible where the old one was.

        examples:
          - program: '.bar as $x | .foo | . + $x'
            input: '{"foo":10, "bar":200}'
            output: ['210']
          - program: '. as $i|[(.*2|. as $i| $i), $i]'
            input: '5'
            output: ['[10,5]']
          - program: '. as [$a, $b, {c: $c}] | $a + $b + $c'
            input: '[2, 3, {"c": 4, "d": 5}]'
            output: ['9']
          - program: '.[] as [$a, $b] | {a: $a, b: $b}'
            input: '[[0], [0, 1], [2, 1, 0]]'
            output: ['{"a":0,"b":null}', '{"a":0,"b":1}', '{"a":2,"b":1}']

      - title: 'Destructuring Alternative Operator: `?//`'
        body: |

          The destructuring alternative operator provides a concise mechanism
          for destructuring an input that can take one of several forms.

          Suppose we have an API that returns a list of resources and events
          associated with them, and we want to get the user_id and timestamp of
          the first event for each resource. The API (having been clumsily
          converted from XML) will only wrap the events in an array if the resource
          has multiple events:

              {"resources": [{"id": 1, "kind": "widget", "events": {"action": "create", "user_id": 1, "ts": 13}},
                             {"id": 2, "kind": "widget", "events": [{"action": "create", "user_id": 1, "ts": 14}, {"action": "destroy", "user_id": 1, "ts": 15}]}]}

          We can use the destructuring alternative operator to handle this structural change simply:

              .resources[] as {$id, $kind, events: {$user_id, $ts}} ?// {$id, $kind, events: [{$user_id, $ts}]} | {$user_id, $kind, $id, $ts}

          Or, if we aren't sure if the input is an array of values or an object:

              .[] as [$id, $kind, $user_id, $ts] ?// {$id, $kind, $user_id, $ts} | ...

          Each alternative need not define all of the same variables, but all named
          variables will be available to the subsequent expression. Variables not
          matched in the alternative that succeeded will be `null`:

              .resources[] as {$id, $kind, events: {$user_id, $ts}} ?// {$id, $kind, events: [{$first_user_id, $first_ts}]} | {$user_id, $first_user_id, $kind, $id, $ts, $first_ts}

          Additionally, if the subsequent expression returns an error, the
          alternative operator will attempt to try the next binding. Errors
          that occur during the final alternative are passed through.

              [[3]] | .[] as [$a] ?// [$b] | if $a != null then error("err: \($a)") else {$a,$b} end

        examples:
          - program: '.[] as {$a, $b, c: {$d, $e}} ?// {$a, $b, c: [{$d, $e}]} | {$a, $b, $d, $e}'
            input: '[{"a": 1, "b": 2, "c": {"d": 3, "e": 4}}, {"a": 1, "b": 2, "c": [{"d": 3, "e": 4}]}]'
            output: ['{"a":1,"b":2,"d":3,"e":4}', '{"a":1,"b":2,"d":3,"e":4}']
          - program: '.[] as {$a, $b, c: {$d}} ?// {$a, $b, c: [{$e}]} | {$a, $b, $d, $e}'
            input: '[{"a": 1, "b": 2, "c": {"d": 3, "e": 4}}, {"a": 1, "b": 2, "c": [{"d": 3, "e": 4}]}]'
            output: ['{"a":1,"b":2,"d":3,"e":null}', '{"a":1,"b":2,"d":null,"e":4}']
          - program: '.[] as [$a] ?// [$b] | if $a != null then error("err: \($a)") else {$a,$b} end'
            input: '[[3]]'
            output: ['{"a":null,"b":3}']

      - title: 'Defining Functions'
        body: |

          You can give a filter a name using "def" syntax:

              def increment: . + 1;

          From then on, `increment` is usable as a filter just like a
          builtin function (in fact, this is how many of the builtins
          are defined). A function may take arguments:

              def map(f): [.[] | f];

          Arguments are passed as _filters_ (functions with no
          arguments), _not_ as values. The same argument may be
          referenced multiple times with different inputs (here `f` is
          run for each element of the input array).  Arguments to a
          function work more like callbacks than like value arguments.
          This is important to understand.  Consider:

              def foo(f): f|f;
              5|foo(.*2)

          The result will be 20 because `f` is `.*2`, and during the
          first invocation of `f` `.` will be 5, and the second time it
          will be 10 (5 * 2), so the result will be 20.  Function
          arguments are filters, and filters expect an input when
          invoked.

          If you want the value-argument behaviour for defining simple
          functions, you can just use a variable:

              def addvalue(f): f as $f | map(. + $f);

          Or use the short-hand:

              def addvalue($f): ...;

          With either definition, `addvalue(.foo)` will add the current
          input's `.foo` field to each element of the array.  Do note
          that calling `addvalue(.[])` will cause the `map(. + $f)` part
          to be evaluated once per value in the value of `.` at the call
          site.

          Multiple definitions using the same function name are allowed.
          Each re-definition replaces the previous one for the same
          number of function arguments, but only for references from
          functions (or main program) subsequent to the re-definition.
          See also the section below on scoping.

        examples:
          - program: 'def addvalue(f): . + [f]; map(addvalue(.[0]))'
            input: '[[1,2],[10,20]]'
            output: ['[[1,2,1], [10,20,10]]']
          - program: 'def addvalue(f): f as $x | map(. + $x); addvalue(.[0])'
            input: '[[1,2],[10,20]]'
            output: ['[[1,2,1,2], [10,20,1,2]]']

      - title: 'Scoping'
        body: |

          There are two types of symbols in jq: value bindings (a.k.a.,
          "variables"), and functions.  Both are scoped lexically,
          with expressions being able to refer only to symbols that
          have been defined "to the left" of them.  The only exception
          to this rule is that functions can refer to themselves so as
          to be able to create recursive functions.

          For example, in the following expression there is a binding
          which is visible "to the right" of it, `... | .*3 as
          $times_three | [. + $times_three] | ...`, but not "to the
          left".  Consider this expression now, `... | (.*3 as
          $times_three | [. + $times_three]) | ...`: here the binding
          `$times_three` is _not_ visible past the closing parenthesis.

      - title: "`isempty(exp)`"
        body: |

          Returns true if `exp` produces no outputs, false otherwise.

        examples:
          - program: 'isempty(empty)'
            input: 'null'
            output: ['true']

          - program: 'isempty(.[])'
            input: '[]'
            output: ['true']

          - program: 'isempty(.[])'
            input: '[1,2,3]'
            output: ['false']

      - title: "`limit(n; expr)`"
        body: |

          The `limit` function extracts up to `n` outputs from `expr`.

        examples:
          - program: '[limit(3; .[])]'
            input: '[0,1,2,3,4,5,6,7,8,9]'
            output: ['[0,1,2]']

      - title: "`skip(n; expr)`"
        body: |

          The `skip` function skips the first `n` outputs from `expr`.

        examples:
          - program: '[skip(3; .[])]'
            input: '[0,1,2,3,4,5,6,7,8,9]'
            output: ['[3,4,5,6,7,8,9]']

      - title: "`first(expr)`, `last(expr)`, `nth(n; expr)`"
        body: |

          The `first(expr)` and `last(expr)` functions extract the first
          and last values from `expr`, respectively.

          The `nth(n; expr)` function extracts the nth value output by `expr`.
          Note that `nth(n; expr)` doesn't support negative values of `n`.

        examples:
          - program: '[first(range(.)), last(range(.)), nth(5; range(.))]'
            input: '10'
            output: ['[0,9,5]']
          - program: '[first(empty), last(empty), nth(5; empty)]'
            input: 'null'
            output: ['[]']

      - title: "`first`, `last`, `nth(n)`"
        body: |

          The `first` and `last` functions extract the first
          and last values from any array at `.`.

          The `nth(n)` function extracts the nth value of any array at `.`.

        examples:
          - program: '[range(.)]|[first, last, nth(5)]'
            input: '10'
            output: ['[0,9,5]']

      - title: "`reduce`"
        body: |

          The `reduce` syntax allows you to combine all of the results of
          an expression by accumulating them into a single answer.
          The form is `reduce EXP as $var (INIT; UPDATE)`.
          As an example, we'll pass `[1,2,3]` to this expression:

              reduce .[] as $item (0; . + $item)

          For each result that `.[]` produces, `. + $item` is run to
          accumulate a running total, starting from 0 as the input value.
          In this example, `.[]` produces the results `1`, `2`, and `3`,
          so the effect is similar to running something like this:

              0 | 1 as $item | . + $item |
                  2 as $item | . + $item |
                  3 as $item | . + $item

        examples:
          - program: 'reduce .[] as $item (0; . + $item)'
            input: '[1,2,3,4,5]'
            output: ['15']

          - program: 'reduce .[] as [$i,$j] (0; . + $i * $j)'
            input: '[[1,2],[3,4],[5,6]]'
            output: ['44']

          - program: 'reduce .[] as {$x,$y} (null; .x += $x | .y += [$y])'
            input: '[{"x":"a","y":1},{"x":"b","y":2},{"x":"c","y":3}]'
            output: ['{"x":"abc","y":[1,2,3]}']

      - title: "`foreach`"
        body: |

          The `foreach` syntax is similar to `reduce`, but intended to
          allow the construction of `limit` and reducers that produce
          intermediate results.

          The form is `foreach EXP as $var (INIT; UPDATE; EXTRACT)`.
          As an example, we'll pass `[1,2,3]` to this expression:

              foreach .[] as $item (0; . + $item; [$item, . * 2])

          Like the `reduce` syntax, `. + $item` is run for each result
          that `.[]` produces, but `[$item, . * 2]` is run for each
          intermediate values. In this example, since the intermediate
          values are `1`, `3`, and `6`, the `foreach` expression produces
          `[1,2]`, `[2,6]`, and `[3,12]`. So the effect is similar
          to running something like this:

              0 | 1 as $item | . + $item | [$item, . * 2],
                  2 as $item | . + $item | [$item, . * 2],
                  3 as $item | . + $item | [$item, . * 2]

          When `EXTRACT` is omitted, the identity filter is used.
          That is, it outputs the intermediate values as they are.

        examples:
          - program: 'foreach .[] as $item (0; . + $item)'
            input: '[1,2,3,4,5]'
            output: ['1','3','6','10','15']

          - program: 'foreach .[] as $item (0; . + $item; [$item, . * 2])'
            input: '[1,2,3,4,5]'
            output: ['[1,2]','[2,6]','[3,12]','[4,20]','[5,30]']

          - program: 'foreach .[] as $item (0; . + 1; {index: ., $item})'
            input: '["foo", "bar", "baz"]'
            output:
              - '{"index":1,"item":"foo"}'
              - '{"index":2,"item":"bar"}'
              - '{"index":3,"item":"baz"}'

      - title: Recursion
        body: |

          As described above, `recurse` uses recursion, and any jq
          function can be recursive.  The `while` builtin is also
          implemented in terms of recursion.

          Tail calls are optimized whenever the expression to the left of
          the recursive call outputs its last value.  In practice this
          means that the expression to the left of the recursive call
          should not produce more than one output for each input.

          For example:

              def recurse(f): def r: ., (f | select(. != null) | r); r;

              def while(cond; update):
                def _while:
                  if cond then ., (update | _while) else empty end;
                _while;

              def repeat(exp):
                def _repeat:
                  exp, _repeat;
                _repeat;

      - title: Generators and iterators
        body: |

            Some jq operators and functions are actually generators in
            that they can produce zero, one, or more values for each
            input, just as one might expect in other programming
            languages that have generators.  For example, `.[]`
            generates all the values in its input (which must be an
            array or an object), `range(0; 10)` generates the integers
            between 0 and 10, and so on.

            Even the comma operator is a generator, generating first
            the values generated by the expression to the left of the
            comma, then the values generated by the expression on the
            right of the comma.

            The `empty` builtin is the generator that produces zero
            outputs.  The `empty` builtin backtracks to the preceding
            generator expression.

            All jq functions can be generators just by using builtin
            generators.  It is also possible to construct new generators
            using only recursion and the comma operator.  If
            recursive calls are "in tail position" then the
            generator will be efficient.  In the example below the
            recursive call by `_range` to itself is in tail position.
            The example shows off three advanced topics: tail recursion,
            generator construction, and sub-functions.

        examples:
          - program: 'def range(init; upto; by):
                    def _range:
                        if (by > 0 and . < upto) or (by < 0 and . > upto)
                        then ., ((.+by)|_range)
                        else empty end;
                    if init == upto then empty elif by == 0 then init else init|_range end;
                range(0; 10; 3)'
            input: 'null'
            output: ['0', '3', '6', '9']
          - program: 'def while(cond; update):
                    def _while:
                        if cond then ., (update | _while) else empty end;
                    _while;
                [while(.<100; .*2)]'
            input: '1'
            output: ['[1,2,4,8,16,32,64]']

  - title: 'Math'
    body: |

      jq currently only has IEEE754 double-precision (64-bit) floating
      point number support.

      Besides simple arithmetic operators such as `+`, jq also has most
      standard math functions from the C math library.  C math functions
      that take a single input argument (e.g., `sin()`) are available as
      zero-argument jq functions.  C math functions that take two input
      arguments (e.g., `pow()`) are available as two-argument jq
      functions that ignore `.`.  C math functions that take three input
      arguments are available as three-argument jq functions that ignore
      `.`.

      Availability of standard math functions depends on the
      availability of the corresponding math functions in your operating
      system and C math library.  Unavailable math functions will be
      defined but will raise an error.

      One-input C math functions: `acos` `acosh` `asin` `asinh` `atan`
      `atanh` `cbrt` `ceil` `cos` `cosh` `erf` `erfc` `exp` `exp10`
      `exp2` `expm1` `fabs` `floor` `gamma` `j0` `j1` `lgamma` `log`
      `log10` `log1p` `log2` `logb` `nearbyint` `rint` `round`
      `significand` `sin` `sinh` `sqrt` `tan` `tanh` `tgamma` `trunc`
      `y0` `y1`.

      Two-input C math functions: `atan2` `copysign` `drem` `fdim`
      `fmax` `fmin` `fmod` `frexp` `hypot` `jn` `ldexp` `modf`
      `nextafter` `nexttoward` `pow` `remainder` `scalb` `scalbln` `yn`.

      Three-input C math functions: `fma`.

      See your system's manual for more information on each of these.

  - title: 'I/O'
    body: |

      At this time jq has minimal support for I/O, mostly in the
      form of control over when inputs are read.  Two builtins functions
      are provided for this, `input` and `inputs`, that read from the
      same sources (e.g., `stdin`, files named on the command-line) as
      jq itself.  These two builtins, and jq's own reading actions, can
      be interleaved with each other.  They are commonly used in combination
      with the null input option `-n` to prevent one input from being read
      implicitly.

      Two builtins provide minimal output capabilities, `debug`, and
      `stderr`.  (Recall that a jq program's output values are always
      output as JSON texts on `stdout`.) The `debug` builtin can have
      application-specific behavior, such as for executables that use
      the libjq C API but aren't the jq executable itself.  The `stderr`
      builtin outputs its input in raw mode to stderr with no additional
      decoration, not even a newline.

      Most jq builtins are referentially transparent, and yield constant
      and repeatable value streams when applied to constant inputs.
      This is not true of I/O builtins.

    entries:
      - title: "`input`"
        body: |

          Outputs one new input.

          Note that when using `input` it is generally necessary to
          invoke jq with the `-n` command-line option, otherwise
          the first entity will be lost.

              echo 1 2 3 4 | jq '[., input]' # [1,2] [3,4]

      - title: "`inputs`"
        body: |

          Outputs all remaining inputs, one by one.

          This is primarily useful for reductions over a program's
          inputs.  Note that when using `inputs` it is generally necessary
          to invoke jq with the `-n` command-line option, otherwise
          the first entity will be lost.

              echo 1 2 3 | jq -n 'reduce inputs as $i (0; . + $i)' # 6

      - title: "`debug`, `debug(msgs)`"
        body: |

          These two filters are like `.` but have as a side-effect the
          production of one or more messages on stderr.

          The message produced by the `debug` filter has the form

              ["DEBUG:",<input-value>]

          where `<input-value>` is a compact rendition of the input
          value.  This format may change in the future.

          The `debug(msgs)` filter is defined as `(msgs | debug | empty), .`
          thus allowing great flexibility in the content of the message,
          while also allowing multi-line debugging statements to be created.

          For example, the expression:

              1 as $x | 2 | debug("Entering function foo with $x == \($x)", .) | (.+1)

          would produce the value 3 but with the following two lines
          being written to stderr:

              ["DEBUG:","Entering function foo with $x == 1"]
              ["DEBUG:",2]

      - title: "`stderr`"
        body: |

          Prints its input in raw and compact mode to stderr with no
          additional decoration, not even a newline.

      - title: "`input_filename`"
        body: |

          Returns the name of the file whose input is currently being
          filtered.  Note that this will not work well unless jq is
          running in a UTF-8 locale.

      - title: "`input_line_number`"
        body: |

          Returns the line number of the input currently being filtered.

  - title: 'Streaming'
    body: |

      With the `--stream` option jq can parse input texts in a streaming
      fashion, allowing jq programs to start processing large JSON texts
      immediately rather than after the parse completes.  If you have a
      single JSON text that is 1GB in size, streaming it will allow you
      to process it much more quickly.

      However, streaming isn't easy to deal with as the jq program will
      have `[<path>, <leaf-value>]` (and a few other forms) as inputs.

      Several builtins are provided to make handling streams easier.

      The examples below use the streamed form of `["a",["b"]]`, which is
      `[[0],"a"],[[1,0],"b"],[[1,0]],[[1]]`.

      Streaming forms include `[<path>, <leaf-value>]` (to indicate any
      scalar value, empty array, or empty object), and `[<path>]` (to
      indicate the end of an array or object).  Future versions of jq
      run with `--stream` and `--seq` may output additional forms such
      as `["error message"]` when an input text fails to parse.

    entries:
      - title: "`truncate_stream(stream_expression)`"
        body: |

          Consumes a number as input and truncates the corresponding
          number of path elements from the left of the outputs of the
          given streaming expression.

        examples:
          - program: 'truncate_stream([[0],"a"],[[1,0],"b"],[[1,0]],[[1]])'
            input: '1'
            output: ['[[0],"b"]', '[[0]]']

      - title: "`fromstream(stream_expression)`"
        body: |

          Outputs values corresponding to the stream expression's
          outputs.

        examples:
          - program: 'fromstream(1|truncate_stream([[0],"a"],[[1,0],"b"],[[1,0]],[[1]]))'
            input: 'null'
            output: ['["b"]']

      - title: "`tostream`"
        body: |

          The `tostream` builtin outputs the streamed form of its input.

        examples:
          - program: '. as $dot|fromstream($dot|tostream)|.==$dot'
            input: '[0,[1,{"a":1},{"b":2}]]'
            output: ['true']

  - title: Assignment
    body: |
      Assignment works a little differently in jq than in most
      programming languages. jq doesn't distinguish between references
      to and copies of something - two objects or arrays are either
      equal or not equal, without any further notion of being "the
      same object" or "not the same object".

      If an object has two fields which are arrays, `.foo` and `.bar`,
      and you append something to `.foo`, then `.bar` will not get
      bigger, even if you've previously set `.bar = .foo`.  If you're
      used to programming in languages like Python, Java, Ruby,
      JavaScript, etc. then you can think of it as though jq does a full
      deep copy of every object before it does the assignment (for
      performance it doesn't actually do that, but that's the general
      idea).

      This means that it's impossible to build circular values in jq
      (such as an array whose first element is itself). This is quite
      intentional, and ensures that anything a jq program can produce
      can be represented in JSON.

      All the assignment operators in jq have path expressions on the
      left-hand side (LHS).  The right-hand side (RHS) provides values
      to set to the paths named by the LHS path expressions.

      Values in jq are always immutable.  Internally, assignment works
      by using a reduction to compute new, replacement values for `.` that
      have had all the desired assignments applied to `.`, then
      outputting the modified value.  This might be made clear by this
      example: `{a:{b:{c:1}}} | (.a.b|=3), .`.  This will output
      `{"a":{"b":3}}` and `{"a":{"b":{"c":1}}}` because the last
      sub-expression, `.`, sees the original value, not the modified
      value.

      Most users will want to use modification assignment operators,
      such as `|=` or `+=`, rather than `=`.

      Note that the LHS of assignment operators refers to a value in
      `.`.  Thus `$var.foo = 1` won't work as expected (`$var.foo` is
      not a valid or useful path expression in `.`); use `$var | .foo =
      1` instead.

      Note too that `.a,.b=0` does not set `.a` and `.b`, but
      `(.a,.b)=0` sets both.

    entries:
      - title: "Update-assignment: `|=`"
        body: |
          This is the "update" operator `|=`.  It takes a filter on the
          right-hand side and works out the new value for the property
          of `.` being assigned to by running the old value through this
          expression. For instance, `(.foo, .bar) |= .+1` will build an
          object with the `foo` field set to the input's `foo` plus 1,
          and the `bar` field set to the input's `bar` plus 1.

          The left-hand side can be any general path expression; see `path()`.

          Note that the left-hand side of `|=` refers to a value in `.`.
          Thus `$var.foo |= . + 1` won't work as expected (`$var.foo` is
          not a valid or useful path expression in `.`); use `$var |
          .foo |= . + 1` instead.

          If the right-hand side outputs no values (i.e., `empty`), then
          the left-hand side path will be deleted, as with `del(path)`.

          If the right-hand side outputs multiple values, only the first
          one will be used (COMPATIBILITY NOTE: in jq 1.5 and earlier
          releases, it used to be that only the last one was used).

        examples:
          - program: '(..|select(type=="boolean")) |= if . then 1 else 0 end'
            input: '[true,false,[5,true,[true,[false]],false]]'
            output: ['[1,0,[5,1,[1,[0]],0]]']

      - title: "Arithmetic update-assignment: `+=`, `-=`, `*=`, `/=`, `%=`, `//=`"
        body: |

          jq has a few operators of the form `a op= b`, which are all
          equivalent to `a |= . op b`. So, `+= 1` can be used to
          increment values, being the same as `|= . + 1`.

        examples:
          - program: .foo += 1
            input: '{"foo": 42}'
            output: ['{"foo": 43}']

      - title: "Plain assignment: `=`"
        body: |

          This is the plain assignment operator.  Unlike the others, the
          input to the right-hand side (RHS) is the same as the input to
          the left-hand side (LHS) rather than the value at the LHS
          path, and all values output by the RHS will be used (as shown
          below).

          If the RHS of `=` produces multiple values, then for each such
          value jq will set the paths on the left-hand side to the value
          and then it will output the modified `.`.  For example,
          `(.a,.b) = range(2)` outputs `{"a":0,"b":0}`, then
          `{"a":1,"b":1}`.  The "update" assignment forms (see above) do
          not do this.

          This example should show the difference between `=` and `|=`:

          Provide input `{"a": {"b": 10}, "b": 20}` to the programs

              .a = .b

          and

              .a |= .b

          The former will set the `a` field of the input to the `b`
          field of the input, and produce the output `{"a": 20, "b": 20}`.
          The latter will set the `a` field of the input to the `a`
          field's `b` field, producing `{"a": 10, "b": 20}`.

        examples:
          - program: .a = .b
            input: '{"a": {"b": 10}, "b": 20}'
            output: ['{"a":20,"b":20}']

          - program: .a |= .b
            input: '{"a": {"b": 10}, "b": 20}'
            output: ['{"a":10,"b":20}']

          - program: (.a, .b) = range(3)
            input: 'null'
            output:
              - '{"a":0,"b":0}'
              - '{"a":1,"b":1}'
              - '{"a":2,"b":2}'

          - program: (.a, .b) |= range(3)
            input: 'null'
            output: ['{"a":0,"b":0}']

      - title: Complex assignments
        body: |
          Lots more things are allowed on the left-hand side of a jq assignment
          than in most languages. We've already seen simple field accesses on
          the left hand side, and it's no surprise that array accesses work just
          as well:

              .posts[0].title = "JQ Manual"

          What may come as a surprise is that the expression on the left may
          produce multiple results, referring to different points in the input
          document:

              .posts[].comments |= . + ["this is great"]

          That example appends the string "this is great" to the "comments"
          array of each post in the input (where the input is an object with a
          field "posts" which is an array of posts).

          When jq encounters an assignment like 'a = b', it records the "path"
          taken to select a part of the input document while executing a. This
          path is then used to find which part of the input to change while
          executing the assignment. Any filter may be used on the
          left-hand side of an equals - whichever paths it selects from the
          input will be where the assignment is performed.

          This is a very powerful operation. Suppose we wanted to add a comment
          to blog posts, using the same "blog" input above. This time, we only
          want to comment on the posts written by "stedolan". We can find those
          posts using the "select" function described earlier:

              .posts[] | select(.author == "stedolan")

          The paths provided by this operation point to each of the posts that
          "stedolan" wrote, and we can comment on each of them in the same way
          that we did before:

              (.posts[] | select(.author == "stedolan") | .comments) |=
                  . + ["terrible."]

  - title: Comments

    body: |

      You can write comments in your jq filters using `#`.

      A `#` character (not part of a string) starts a comment.
      All characters from `#` to the end of the line are ignored.

      If the end of the line is preceded by an odd number of backslash
      characters, the following line is also considered part of the
      comment and is ignored.

      For example, the following code outputs `[1,3,4,7]`

          [
            1,
            # foo \
            2,
            # bar \\
            3,
            4, # baz \\\
            5, \
            6,
            7
            # comment \
              comment \
              comment
          ]

      Backslash continuing the comment on the next line can be useful
      when writing the "shebang" for a jq script:

          #!/bin/sh --
          # total - Output the sum of the given arguments (or stdin)
          # usage: total [numbers...]
          # \
          exec jq --args -MRnf -- "$0" "$@"

          $ARGS.positional |
          reduce (
            if . == []
              then inputs
              else .[]
            end |
            . as $dot |
            try tonumber catch false |
            if not or isnan then
              @json "total: Invalid number \($dot).\n" | halt_error(1)
            end
          ) as $n (0; . + $n)

      The `exec` line is considered a comment by jq, so it is ignored.
      But it is not ignored by `sh`, since in `sh` a backslash at the
      end of the line does not continue the comment.
      With this trick, when the script is invoked as `total 1 2`,
      `/bin/sh -- /path/to/total 1 2` will be run, and `sh` will then
      run `exec jq --args -MRnf -- /path/to/total 1 2` replacing itself
      with a `jq` interpreter invoked with the specified options (`-M`,
      `-R`, `-n`, `--args`), that evaluates the current file (`$0`),
      with the arguments (`$@`) that were passed to `sh`.

  - title: Modules
    body: |

      jq has a library/module system.  Modules are files whose names end
      in `.jq`.

      Modules imported by a program are searched for in a default search
      path (see below).  The `import` and `include` directives allow the
      importer to alter this path.

      Paths in the search path are subject to various substitutions.

      For paths starting with `~/`, the user's home directory is
      substituted for `~`.

      For paths starting with `$ORIGIN/`, the directory where the jq
      executable is located is substituted for `$ORIGIN`.

      For paths starting with `./` or paths that are `.`, the path of
      the including file is substituted for `.`.  For top-level programs
      given on the command-line, the current directory is used.

      Import directives can optionally specify a search path to which
      the default is appended.

      The default search path is the search path given to the `-L`
      command-line option, else `["~/.jq", "$ORIGIN/../lib/jq",
      "$ORIGIN/../lib"]`.

      Null and empty string path elements terminate search path
      processing.

      A dependency with relative path `foo/bar` would be searched for in
      `foo/bar.jq` and `foo/bar/bar.jq` in the given search path. This
      is intended to allow modules to be placed in a directory along
      with, for example, version control files, README files, and so on,
      but also to allow for single-file modules.

      Consecutive components with the same name are not allowed to avoid
      ambiguities (e.g., `foo/foo`).

      For example, with `-L$HOME/.jq` a module `foo` can be found in
      `$HOME/.jq/foo.jq` and `$HOME/.jq/foo/foo.jq`.

      If `.jq` exists in the user's home directory, and is a file (not a
      directory), it is automatically sourced into the main program.

    entries:
      - title: "`import RelativePathString as NAME [<metadata>];`"
        body: |

          Imports a module found at the given path relative to a
          directory in a search path.  A `.jq` suffix will be added to
          the relative path string.  The module's symbols are prefixed
          with `NAME::`.

          The optional metadata must be a constant jq expression.  It
          should be an object with keys like `homepage` and so on.  At
          this time jq only uses the `search` key/value of the metadata.
          The metadata is also made available to users via the
          `modulemeta` builtin.

          The `search` key in the metadata, if present, should have a
          string or array value (array of strings); this is the search
          path to be prefixed to the top-level search path.

      - title: "`include RelativePathString [<metadata>];`"
        body: |

          Imports a module found at the given path relative to a
          directory in a search path as if it were included in place.  A
          `.jq` suffix will be added to the relative path string.  The
          module's symbols are imported into the caller's namespace as
          if the module's content had been included directly.

          The optional metadata must be a constant jq expression.  It
          should be an object with keys like `homepage` and so on.  At
          this time jq only uses the `search` key/value of the metadata.
          The metadata is also made available to users via the
          `modulemeta` builtin.

      - title: "`import RelativePathString as $NAME [<metadata>];`"
        body: |

          Imports a JSON file found at the given path relative to a
          directory in a search path.  A `.json` suffix will be added to
          the relative path string.  The file's data will be available
          as `$NAME::NAME`.

          The optional metadata must be a constant jq expression.  It
          should be an object with keys like `homepage` and so on.  At
          this time jq only uses the `search` key/value of the metadata.
          The metadata is also made available to users via the
          `modulemeta` builtin.

          The `search` key in the metadata, if present, should have a
          string or array value (array of strings); this is the search
          path to be prefixed to the top-level search path.

      - title: "`module <metadata>;`"
        body: |

          This directive is entirely optional.  It's not required for
          proper operation.  It serves only the purpose of providing
          metadata that can be read with the `modulemeta` builtin.

          The metadata must be a constant jq expression.  It should be
          an object with keys like `homepage`.  At this time jq doesn't
          use this metadata, but it is made available to users via the
          `modulemeta` builtin.

      - title: "`modulemeta`"
        body: |

          Takes a module name as input and outputs the module's metadata
          as an object, with the module's imports (including metadata)
          as an array value for the `deps` key and the module's defined
          functions as an array value for the `defs` key.

          Programs can use this to query a module's metadata, which they
          could then use to, for example, search for, download, and
          install missing dependencies.

  - title: Colors
    body: |

      To configure alternative colors just set the `JQ_COLORS`
      environment variable to colon-delimited list of partial terminal
      escape sequences like `"1;31"`, in this order:

        - color for `null`
        - color for `false`
        - color for `true`
        - color for numbers
        - color for strings
        - color for arrays
        - color for objects
        - color for object keys

      The default color scheme is the same as setting
      `JQ_COLORS="0;90:0;39:0;39:0;39:0;32:1;39:1;39:1;34"`.

      This is not a manual for VT100/ANSI escapes.  However, each of
      these color specifications should consist of two numbers separated
      by a semi-colon, where the first number is one of these:

        - 1 (bright)
        - 2 (dim)
        - 4 (underscore)
        - 5 (blink)
        - 7 (reverse)
        - 8 (hidden)

      and the second is one of these:

        - 30 (black)
        - 31 (red)
        - 32 (green)
        - 33 (yellow)
        - 34 (blue)
        - 35 (magenta)
        - 36 (cyan)
        - 37 (white)