github.com/odvcencio/gotreesitter

Added

• Add grammars.ParseFilePooledStrict. It rejects a partial tree and returns
  the parser stop error.

• Add Tagger.TagStrict. It returns tags only after a complete parse and
  preserves parser errors.

Changed

• Bound pending fork-stack retention. Keep a bounded reserve during parsing,
  and drop oversized storage at parse and parser-pool boundaries.

• Separate C-runtime recovery walk, convergence, and fallback state. Reset
  these states across parse, retry, snippet, pool, and iteration boundaries.

• Provision dense reach marks when generalized LR (GLR) slabs grow. Owned nodes
  no longer use the pointer-map fallback.

Fixed

• Clear stale C-runtime recovery state after a clean condense. A clean suffix
  no longer inherits an active recovery-cost gate.

Tests

• Align the Application Binary Interface version 14 (ABI-14) lexer-mode probe
  with generated end-of-file behavior. The probe now checks the emitted layout
  without injecting an end-of-file state.

Continuous integration (CI)

• The project merged pull request (PR) #732.
  Its CI-only change separates query-fleet smoke tests from regular shards and
  is not part of this release.

Open work

The following items remain open:

• Issue #454 tracks
  downstream field reports.

• Issue #576 tracks
  Swift recovery divergence.

• Issue #586 tracks
  shared GLR error-cost bounds.

• Issue #728 tracks
  external-scanner incremental reuse.

Added

• Add grammars.ParseFilePooledStrict. It rejects a partial tree and returns the parser stop error.

• Add Tagger.TagStrict. It returns tags only after a complete parse and preserves parser errors.

Changed

• Bound pending fork-stack retention. Keep a bounded reserve during parsing, and drop oversized storage at parse and parser-pool boundaries.

• Separate C-runtime recovery walk, convergence, and fallback state. Reset these states across parse, retry, snippet, pool, and iteration boundaries.

• Provision dense reach marks when generalized LR (GLR) slabs grow. Owned nodes no longer use the pointer-map fallback.

Fixed

• Clear stale C-runtime recovery state after a clean condense. A clean suffix no longer inherits an active recovery-cost gate.

Tests

• Align the Application Binary Interface version 14 (ABI-14) lexer-mode probe with generated end-of-file behavior. The probe now checks the emitted layout without injecting an end-of-file state.

Continuous integration (CI)

• The project merged pull request (PR) #732. Its CI-only change separates query-fleet smoke tests from regular shards and is not part of this release.

Open work

The following items remain open:

• Issue #454 tracks downstream field reports.

• Issue #576 tracks Swift recovery divergence.

• Issue #586 tracks shared GLR error-cost bounds.

• The Swift #586 compact-parser receipt records a NO-GO recovery-cost and token-producer blocker. Keep the issue open.

• Issue #728 tracks external-scanner incremental reuse.

Added

• Add grammars.ParseFilePooledStrict. It rejects a partial tree and returns the parser stop error.

• Add Tagger.TagStrict. It returns tags only after a complete parse and preserves parser errors.

Changed

• Bound pending fork-stack retention. Keep a bounded reserve during parsing, and drop oversized storage at parse and parser-pool boundaries.

• Separate C-runtime recovery walk, convergence, and fallback state. Reset these states across parse, retry, snippet, pool, and iteration boundaries.

• Provision dense reach marks when generalized LR (GLR) slabs grow. Owned nodes no longer use the pointer-map fallback.

Fixed

• Clear stale C-runtime recovery state after a clean condense. A clean suffix no longer inherits an active recovery-cost gate.

Tests

• Align the Application Binary Interface version 14 (ABI-14) lexer-mode probe with generated end-of-file behavior. The probe now checks the emitted layout without injecting an end-of-file state.

Continuous integration (CI)

• The project merged pull request (PR) #732. Its CI-only change separates query-fleet smoke tests from regular shards and is not part of this release.

Open work

The following items remain open:

• Issue #454 tracks downstream field reports.

• Issue #576 tracks Swift recovery divergence.

• Issue #586 tracks shared GLR error-cost bounds.

• Issue #728 tracks external-scanner incremental reuse.

Fixed

• Restore single-language grammar_subset builds. Shared lexer helpers and
  Python-derived scanner state now compile with each grammar that uses them.
  Derivative-only builds do not register Python scanner metadata.

• Add a blocking subset build sweep. Continuous integration now builds all
  206 registered grammars with their individual grammar_subset tags.

Fixed

• Restore single-language grammar_subset builds. Shared lexer helpers and Python-derived scanner state now compile with each grammar that uses them. Derivative-only builds do not register Python scanner metadata.

• Add a blocking subset build sweep. Continuous integration now builds all 206 registered grammars with their individual grammar_subset tags.

Added

• TestOutlineOracleDifferential (cgo_harness/outline_differential_test.go)
  runs each language's resolved tags query through both the pure-Go query
  engine and the official C tree-sitter runtime, then diffs the two capture
  streams. It hard-asserts capture parity for the core nine outline
  languages (go, python, javascript, typescript, tsx, rust,
  java, c, cpp) and logs a census for every other language with a
  resolvable tags query.

• TestOutlineCoverageWitnesses (grammars/outline_coverage_witness_test.go)
  pins 30 languages against the real Outliner pipeline. Each case names an
  exact symbol its resolved tags query must produce, so a pattern that
  compiles but never fires now fails the test instead of hiding behind a
  non-empty query string.

• File-outline tags-query coverage rose from a 30-language floor to 84 of
  206 registered languages: 83 with a real-corpus fixture and 1 with a
  smoke-sample fixture (grammars/testdata/outline_census/baseline.json).
  TestInferredTagsQueryCoverage (grammars/registry_test.go) now enforces
  84 as the floor. See docs/outline.md for the full coverage tiers and the
  file outline API.

• OutlineSymbol.Owner now resolves on every OutlineTree call. A rule
  attached through WithOutlineOwnerRules matches a symbol by NodeType,
  reads its OwnerField, and descends through the rule's Unwrap node
  types until it reaches exactly one NameTypes terminal. Any other
  outcome — an absent field, or a walk that reaches zero or more than one
  terminal — leaves Owner empty and counts one
  OutlineReport.OwnerRuleMisses; a NodeType no attached rule names
  touches neither field.

• grammars.OutlineOwnerRules(entry) gates the shipped owner-rule table by
  symbol and field presence in each language's own compiled grammar, the
  same way ResolveTagsQuery's inference table is gated, and composes
  directly with gotreesitter.WithOutlineOwnerRules. The shipped Go rule
  resolves all four receiver shapes — value, pointer, generic value, and
  generic pointer — to the receiver's base type name. See docs/outline.md
  for the full resolution contract and a worked example.

Removed

• Native root-extra folding and reduction now own Elixir comment placement
  and map entry grouping. Parsing drops the hidden
  _newline_before_comment scanner token without a post-parse pass.
  Reduction groups map keyword pairs and wraps update and arrow entries in
  the map grammar's binary_operator node. This removes the Elixir
  result-compatibility dispatcher arm.

Removed

• Native scheduling and reduction now own the Enforce const int formal
  parameter shape. Parsing classifies const as a
  formal_parameter_modifier and int as type_int directly. It keeps
  the parameter's own name and default value instead of losing them to a
  misread type/name pair. This removes the Enforce result-compatibility
  dispatcher arm.

Fixed

• TypeScript and TSX no longer split a signed right-shift operator into two
  generic closers. splitCompactCloseAngleToken narrows a >> token to a
  single > so nested closers such as Array<Array<string>> parse. Java
  gated that split behind an "unclosed < precedes this run" check, because
  >> is also a shift operator there; TypeScript and TSX carried no such
  gate. Any >> whose next byte was one of ( ) [ ] { } , . ; : ? was torn
  apart, so x = a >> (b) failed to parse while x = a >> b succeeded. The
  angle-depth gate now covers TypeScript and TSX, and runs only when a >>
  symbol is actually active in the state: with no shift alternative to
  protect, narrowing to > remains the only way to make progress. Found
  while migrating the GoSX browser runtime to TypeScript, on
  indices[i] = (src[i >> 3] >> (i & 7)) & 1;.
  on, off, yes, no) at lex time. The grammar's word token pattern can
  absorb leading whitespace before a keyword, and the keyword re-lex
  previously required its match to start at byte zero, so it missed that
  case and left the value as a generic string token. DFA keyword promotion
  now skips the leading run first, the same way tree-sitter's own generated
  keyword lexer does. This retires the Hyprlang dispatcher arm and fixes a
  related bug: a trailing space after the keyword no longer produces an
  incorrect boolean node.

• Seven of the file outline's core nine languages carried tags-query rows
  that reference a child node type the grammar never produces at that
  position: javascript, typescript, tsx, java, c, cpp, and
  rust. The C query compiler rejects such a row as an "Impossible
  pattern" and drops the entire multi-pattern query with it. The pure-Go
  query engine has no matching check, so it compiled the same row and
  silently matched nothing. The dead rows are now removed or corrected.
  TestOutlineOracleDifferential's core-nine tier reports capture-stream
  parity for all nine languages, and the fix changed no Go outline output —
  confirmed by unchanged golden fixtures.

Removed

• The Bash assignment-wrapper and if-condition-field repairs. Native
  reduction already builds the C-shaped variable_assignments wrapper for
  two or more consecutive assignments. It already sets the condition
  field on an if-statement's condition tokens too. The Bash compatibility
  pass no longer splices that wrapper's children into the enclosing node or
  rebuilds the field afterward.
  tree-sitter-bash's own corpus (test/corpus/literals.txt) pins the same
  wrapper for two top-level assignments.
  Production, compact, forest, and incremental routes match the raw parse
  exactly, and the isolated C-oracle comparison matches for every case.
  One unrelated Bash subpass remains live.
• The FIDL result-compatibility dispatcher arm is retired. Native recovery
  already builds the C-equivalent error shape for a versioned-layout-modifier
  declaration whose modifier keyword carries a stray (name=value) argument
  list. Production, compact, forest, and incremental routes produce the same
  tree, and an isolated C-oracle parity check confirms the shape.
• The HLSL subscript-assignment declarator member of the result-compatibility
  dispatcher arm is retired. structured_binding_declarator carries a
  negative dynamic precedence in the grammar. Native parsing already elects
  the C-equivalent subscript-assignment expression for Name[index] = value;
  without a post-parse pass. Production, compact, forest, and incremental
  routes stay exact. The negative-number cast and unorm-buffer members of the
  HLSL arm remain live.

Added

• TestOutlineOracleDifferential (cgo_harness/outline_differential_test.go) runs each language's resolved tags query through both the pure-Go query engine and the official C tree-sitter runtime, then diffs the two capture streams. It hard-asserts capture parity for the core nine outline languages (go, python, javascript, typescript, tsx, rust, java, c, cpp) and logs a census for every other language with a resolvable tags query.

• TestOutlineCoverageWitnesses (grammars/outline_coverage_witness_test.go) pins 30 languages against the real Outliner pipeline. Each case names an exact symbol its resolved tags query must produce, so a pattern that compiles but never fires now fails the test instead of hiding behind a non-empty query string.

• File-outline tags-query coverage rose from a 30-language floor to 84 of 206 registered languages: 83 with a real-corpus fixture and 1 with a smoke-sample fixture (grammars/testdata/outline_census/baseline.json). TestInferredTagsQueryCoverage (grammars/registry_test.go) now enforces 84 as the floor. See docs/outline.md for the full coverage tiers and the file outline API.

• OutlineSymbol.Owner now resolves on every OutlineTree call. A rule attached through WithOutlineOwnerRules matches a symbol by NodeType, reads its OwnerField, and descends through the rule's Unwrap node types until it reaches exactly one NameTypes terminal. Any other outcome — an absent field, or a walk that reaches zero or more than one terminal — leaves Owner empty and counts one OutlineReport.OwnerRuleMisses; a NodeType no attached rule names touches neither field.

• grammars.OutlineOwnerRules(entry) gates the shipped owner-rule table by symbol and field presence in each language's own compiled grammar, the same way ResolveTagsQuery's inference table is gated, and composes directly with gotreesitter.WithOutlineOwnerRules. The shipped Go rule resolves all four receiver shapes — value, pointer, generic value, and generic pointer — to the receiver's base type name. See docs/outline.md for the full resolution contract and a worked example.

Removed

• Native root-extra folding and reduction now own Elixir comment placement and map entry grouping. Parsing drops the hidden _newline_before_comment scanner token without a post-parse pass. Reduction groups map keyword pairs and wraps update and arrow entries in the map grammar's binary_operator node. This removes the Elixir result-compatibility dispatcher arm.

Removed

• Native scheduling and reduction now own the Enforce const int formal parameter shape. Parsing classifies const as a formal_parameter_modifier and int as type_int directly. It keeps the parameter's own name and default value instead of losing them to a misread type/name pair. This removes the Enforce result-compatibility dispatcher arm.

Fixed

• TypeScript and TSX no longer split a signed right-shift operator into two generic closers. splitCompactCloseAngleToken narrows a >> token to a single > so nested closers such as Array<Array<string>> parse. Java gated that split behind an "unclosed < precedes this run" check, because >> is also a shift operator there; TypeScript and TSX carried no such gate. Any >> whose next byte was one of ( ) [ ] { } , . ; : ? was torn apart, so x = a >> (b) failed to parse while x = a >> b succeeded. The angle-depth gate now covers TypeScript and TSX, and runs only when a >> symbol is actually active in the state: with no shift alternative to protect, narrowing to > remains the only way to make progress. Found while migrating the GoSX browser runtime to TypeScript, on indices[i] = (src[i >> 3] >> (i & 7)) & 1;. on, off, yes, no) at lex time. The grammar's word token pattern can absorb leading whitespace before a keyword, and the keyword re-lex previously required its match to start at byte zero, so it missed that case and left the value as a generic string token. DFA keyword promotion now skips the leading run first, the same way tree-sitter's own generated keyword lexer does. This retires the Hyprlang dispatcher arm and fixes a related bug: a trailing space after the keyword no longer produces an incorrect boolean node.

• Seven of the file outline's core nine languages carried tags-query rows that reference a child node type the grammar never produces at that position: javascript, typescript, tsx, java, c, cpp, and rust. The C query compiler rejects such a row as an "Impossible pattern" and drops the entire multi-pattern query with it. The pure-Go query engine has no matching check, so it compiled the same row and silently matched nothing. The dead rows are now removed or corrected. TestOutlineOracleDifferential's core-nine tier reports capture-stream parity for all nine languages, and the fix changed no Go outline output — confirmed by unchanged golden fixtures.

Removed

• The Bash assignment-wrapper and if-condition-field repairs. Native reduction already builds the C-shaped variable_assignments wrapper for two or more consecutive assignments. It already sets the condition field on an if-statement's condition tokens too. The Bash compatibility pass no longer splices that wrapper's children into the enclosing node or rebuilds the field afterward. tree-sitter-bash's own corpus (test/corpus/literals.txt) pins the same wrapper for two top-level assignments. Production, compact, forest, and incremental routes match the raw parse exactly, and the isolated C-oracle comparison matches for every case. One unrelated Bash subpass remains live.
• The FIDL result-compatibility dispatcher arm is retired. Native recovery already builds the C-equivalent error shape for a versioned-layout-modifier declaration whose modifier keyword carries a stray (name=value) argument list. Production, compact, forest, and incremental routes produce the same tree, and an isolated C-oracle parity check confirms the shape.
• The HLSL subscript-assignment declarator member of the result-compatibility dispatcher arm is retired. structured_binding_declarator carries a negative dynamic precedence in the grammar. Native parsing already elects the C-equivalent subscript-assignment expression for Name[index] = value; without a post-parse pass. Production, compact, forest, and incremental routes stay exact. The negative-number cast and unorm-buffer members of the HLSL arm remain live.