[1.23.1]

  * Restore toJSONFilter instance for pure `a -> [a]`.
    This went missing after my ill-considered revision to #105,
    commit 183af9d9f1066be974ac55fd23a4c985999d3ce8 . See jgm/pandoc#8976.

  * Generalize ToJSONFilter instance. Previously a pure function `a -> a`
    could only be promoted to a filter in IO. Now we allow it to work with
    any instance of MonadIO.  (This adds to #105.)

  * Allow bytestring 0.12.

[1.23.0.1]

  * Allow aeson 2.2.

  * Remove dependency on string-qq (recbecca skinner).

[1.23]

  * Remove Null constructor from Block (#91) [API change].

  * ToJSONFilter: Add instance for MonadIO (#105, Willem Van Onsem)
    [API change].

  * Add `Figure` block constructor (Albert Krewinkel, Aner Lucero,
    and Christian Despres) [API change]. The new Figure block represents
    a figure with attributes, caption, and arbitrary block content.

[1.22.2.1]

  * Allow aeson 2.1.* and criterion 1.6.

[1.22.2]

  * Use StrictData in Text.Pandoc.Definition.

  * Add Walkable Meta(Value) Pandoc instances (Travis Cardwell).

[1.22.1]

  * Text.Pandoc.Builder: add simpleFigure, simpleFigureWith,
    and the SimpleFigure bidirectional pattern synonym
    (Aner Lucero) [API change].

  * Allow bytestring 0.11 (Alexander Batischev).

  * Update stack resolver to lts-18.10

  * Allow aeson 2+.  Tested with aeson 2.0.1.0.

  * Allow transformers 0.6.

  * Fix incorrect table ColWidth documentation (#85, Nils Carlson).
    The documentation stated that the ColWidth represented the width
    of the column as a fraction of the table width when in represents
    a percentage of the text width.

[1.22]

  * Deprecate isNull from Builder:  null can serve just as well (#67).
    Use null instead of isNull in Arbitrary (Christian Despres, #84).

  * Use untagged JSON encoding for single-constructor types (#75, #76,
    Christian Despres).  All of the single constructor types related
    to Table are now represented in JSON either as arrays (for
    multi-argument constructors) or as the representation of the
    inner type (for single argument constructors). This behaviour
    for newtype-defined and multi-argument non-record types is now
    consistent across the entire JSON interface, with the exception
    of Pandoc itself (which is represented as a JSON object with
    additional metadata). Multi-argument records (of which Citation
    is the only example) are still represented as objects with the
    record accessors as keys.

  * The Meta and Citation types now use derived JSON serialization (newtype
    and generic, respectively). The format remains the same as before
    (Christian Despres).

  * New serialization tests now test that Meta and the Table types are
    encoded properly in JSON (Christian Despres).

  * Use TH To/FromJSON instances (Christian Despres).

  * Remove unused Legacy modules (#80, Despres).
    They are not exported, and are not used internally.

  * Change the table builder to permit looser intermediate table heads (#77,
    Christian Despres).

    The table builder (and the normalizeTableBody function) now permit
    cells in the intermediate head of a TableBody to extend past the
    RowHeadColumns. This allows for intermediate tables to have
    subheadings that extend across the entire table.

    Formerly the table builder would treat the intermediate head like the
    intermediate body, and clip or drop cells that extended past the row
    head.

  * Update QuickCheck lower bound.

  * Fix redundant pattern match.

[1.21]

  * Add Underline constructor (#68, Vaibhav Sagar).

  * Improve table types to allow col, rowspans and more (#65, Christian
    Despres).  The additions include modification of the Block type, some
    newtypes related to tables, and changes to the table builders.  The table
    builder is now aware of the new Table constructor, and normalizes the
    input table appropriately, so that when laid onto a grid the resulting
    table has no empty spaces, overlapping cells, or cells that extend beyond
    their section boundary.

    Three properties of normalization are checked:

    - Normalization is idempotent.
    - Each row of a normalized table is an initial segment of the
      corresponding row in the unnormalized table, modulo changed cell
      dimensions, dropped cells, and padding with empty cells. This is only
      checked for the first row of the TableBody, however, due to row head
      difficulties.
    - The sum of the cell lengths in the first row of every
      normalized table section is always equal to the total table width.

    `simpleTable` has been changed so that a null header list becomes a
    TableHead with a null body, not a TableHead with a single empty row.

  * Bump QuickCheck upper bound.

  * Change lower bound for QuickCheck to 2.10 (needed for `liftShrink2`).

  * Small code quality improvements (Joseph C. Sible, #69).

  * Allow aeson 1.5 (#72, Felix Yan).

  * Fixed documentation typo (Merlin Göttlinger).

  * Add COMPLETE pragmas to the pattern definitions (Christian Despres).

[1.20]

  * Change all uses of String in type definitions to strict Text
    (Christian Despres) [API change].  The MetaValue instances using String
    have been kept, and parallel ones using Text were added.

  * Remove the Arbitrary Text orphan instance (Christian Despres).
    This instance should not have been in the Text.Pandoc.Arbitrary, since
    it would have been exported with the rest of the instances in that
    module. Instead, more shrink* functions were added to compensate for
    the absence of this instance.

  * Add Text.Pandoc.Legacy.Definition (Christian Despres).
    To ease the transition to Text, this module provides an interface
    compatible with the String one, so that any unqualified imports of
    Text.Pandoc.Definition in other packages can be replaced by
    Text.Pandoc.Legacy.Definition without other code changes. This is done
    with PatternSynonyms.

    Some of the constructors of the types Meta, MetaValue, Block, Inline,
    Format, and Citation required PatternSynonym handling. The Attr and
    Target types had to be redefined, and certain functions had to be
    rewritten to handle String or the old Attr and Target types in this
    module. This module otherwise exports the definitions in
    Text.Pandoc.Definition unchanged.

    This is not a perfect drop-in replacement, since some imports like
    Inline(..) will no longer work. This may also cause incomplete pattern
    warnings when used, since the coverage checker does not seem to be
    aware of PatternSynonyms.

  * Add Text.Pandoc.Legacy.Builder (Christian Despres).
    Like Text.Pandoc.Legacy.Definition, this modules provides a
    compatibility interface while the transition to Text takes
    place. Unlike that module, this module only requires redefining the
    ToMetaValue and HasMeta classes and a few functions so that they use
    the old types. No PatternSynonyms are required.

  * Change Semigroup/Monoid instance for Meta.
    Previously `<>` was left-biased, so if meta1 and meta2 both
    contained a field 'foo', the value from meta1 would be retained
    in `meta1 <> meta2`, and the value from meta2 ignored.
    This is counterintuitive and doesn't work well with pandoc;
    for example, we want to be able to override a value in an
    earlier `--metadata-file` with a later one on the command line.

    It also makes the behavior of metadata more like other
    things (such as reference links, where later definitions
    take precedence over earlier ones).

    Note that this change may break some current workflows,
    if one is relying on metadata fields that occur later in
    a document to be overridden by those occurring earlier.

[1.17.6.1]

  * Relax version bound for string-qq.

[1.17.6]

  * Walk: export walk and query helpers (Albert Krewinkel) [API change].
    The `walk*M` and `query*` functions are helpful when defining new
    `Walkable` instances.
  * Allow QuickCheck 2.13.
  * Document meaning of Int in ListAttributes (#45).
  * Update copyright year spans to include 2019 (Albert Krewinkel).
  * Remove CPP instructions for GHC versions < 7.10 (Albert Krewinkel).
  *  update list of GHC versions used for testing (Albert Krewinkel).
  * Fix compiler and hlint warnings (Pete Ryland).

[1.17.5.4]

  * Put NFData in scope for ghc < 7.10.
  * Reduce deepseq lower bound for ghc < 7.10.

[1.17.5.3]

  * For ghc < 7.10, constrain deepseq-generics to >= 0.2, which no
    longer exprots NFData from deepseq. Add deepseq dependency.

[1.17.5.2]

  * Bump upper bound for deepseq-generics, QuickCheck, criterion.
  * Implement QuickCheck shrinking for Inlines and Blocks (Alexander Krotov).

[1.17.5.1]

  * Declare the ToMetaValue instance for String as OVERLAPPING (#46).

[1.17.5]


  * Bump upper bounds for aeson, base.
  * Allow building on older ghc versions (George Wilson).
  * Text.Pandoc.Arbitrary: generate SoftBreaks and LineBreaks
    (Alexander Krotov).
  * Pad table rows up to maximum row length, to guarantee that
    all rows have the same number of columns
    (see jgm/pandoc#4059, Francesco Occhipinti).
  * Make String an instance of ToMetaValue (Alexander Krotov).

[1.17.4.2]

  * Fix compiler warnings.

[1.17.4.1]

  * Import Semigroups when needed rather than using CPP.
  * Bump criterion upper bound.

[1.17.4]

  * Add Semigroup instances for Pandoc, Meta, Inlines, Blocks
    (if base >= 4.9).  This is needed for the library to compile
    with ghc 8.4.
  * Bumped criterion upper bound.

[1.17.3.1]

  * Bumped upper bounds for criterion and QuickCheck.

[1.17.3]

  * Added Walkable instances for `[Inline] Inline` and `[Block] Block`.

[1.17.2]

  * Provide default implementation for walk (Albert Krewinkel).
    The implementations for `walk` and `walkM` are very similar, so a
    default method is provided which implements the former in terms of the
    latter. This change should not affect performance, as the `Identity`
    functor, which is used in the default definition, is a newtype that
    should be eliminated at compile time.  (This requires a dependency
    on transformers for ghc 7.8.x.)
  * Force optimizations when compiling Walk module (Albert Krewinkel).
  * Add `Applicative m` to the context of walkM (Albert Krewinkel).
    The acceptance of AMP makes this a natural change.
  * Add `Walkable [Block]` and `Walkable [Inline]` instances (Albert
    Krewinkel).
  * Factored out duplicate code in Walk.
  * Added benchmark.
  * Text.Pandoc.JSON: Use `walk` instead of `bottomUp` in the
    `ToJSONFilter` instance for `a -> [a]`.  Note that behavior will be
    slightly different, since bottomUp's treatment of a function `[a] -> [a]`
    is to apply it to each sublist of a list, while walk applies it only to
    maximal sublists.  Usually the latter behavior is what is wanted, and the
    former can be simulated when needed.  But there may be existing filters
    that need to be rewritten in light of the new behavior.

[1.17.1]

  * Better consistency in simpleTable and table (jgm/pandoc#3648).
    If `headers` is empty, we populate it with empty cells, using the rows
    to determine number of columns.  We also ensure that there are numcols
    alignments and column widths.
  * Make sure Div and Span occur in Arbitrary instances.
  * Bump dependency upper bounds.
  * Removed unused mapConst.

[1.17.0.5]

  * Allow aeson 1.1.
  * Added tests for Walk (originally from pandoc).
  * Renamed README -> README.md, fix link (Kolen Cheung).

[1.17.0.4]

  * Re-add Functor constraint to walkM, needed for ghc 7.8.
  * Turn off redundant-constraints warning for Walk (for ghc 8.0.1+).
  * Added necessary Data.Traversable import.
  * Allow HUnit 1.5.

[1.17.0.3]

  * More ghc 7.8 compatibility imports.

[1.17.0.2]

  * Added a necessary import from Data.Monoid (for ghc 7.8).

[1.17.0.1]

  * Fix compiler warnings around Data.Monoid import on some platforms.

[1.17]

  * Remimplement json encoding of inlines manually (Jesse Rosenthal).
    This is the first step to doing manual encoding and decoding of pandoc
    JSON. This will replace the current generic deriving, which can be a
    moving target.
  * Move Arbitrary instances for types from pandoc (Jesse Rosenthal).
  * Remove empty arrays for leaf elements (Jesse Rosenthal).
    Elements with no children (Space, SoftBreak, LineBreak, HorizontalRule,
    Null) previously had an empty array for their "c" value. We remove that
    here.  This is a breaking change for the JSON format.
  * New toplevel JSON format with api-version (Jesse Rosenthal). The version
    number is the pandoc-types version.  The toplevel format was previously:

        [{"unMeta": META}, [BLOCKS]]

    It is now:

        {
         "pandoc-api-version" : [MAJ, MIN, REV],
         "meta" : META,
         "blocks": BLOCKS
        }

    Decoding fails if the major and minor version numbers don't match.
  * Add simple quickcheck tests to verify that roundtrip works properly
    (Jesse Rosenthal).
  * Add a LineBlock block element (Albert Krewinkel).
  * Add explicit unit tests for encoding and decoding (Jesse Rosenthal).
  * Fixed warnings in Walk, Builder.
  * Test with travis and appveyor.

[1.16.1.1]

  * Allow aeson 1.0.*.

[1.16.1]

  * Allow aeson 0.11.*.
  * Export pandocTypesVersion from Text.Pandoc.Definition.

[1.16.0.1]

  * Use deepseq instead of deepseq-generics.  deepseq now allows deriving
    generic NFData instances, so we don't need deepseq-generics.
  * Removed unneeded instance, use OVERLAPPING pragma for ghc 7.10.
  * Added CPP so that deepseq-generics is used for ghc < 7.10.
  * Added tested-with, generate .travis.yml using make_travis_yml.hs
  * Added stack.yaml

[1.16]

* Added Attr field to Image and Link.
* Added SoftBreak constructor to Inline