Skip to main content

reflex/
models.rs

1//! Core data models for Reflex
2//!
3//! These structures represent the normalized, deterministic output format
4//! that Reflex provides to AI agents and other programmatic consumers.
5
6use serde::{Deserialize, Serialize};
7use strum::{Display, EnumString};
8
9/// Represents a source code location span (line range only)
10#[derive(Debug, Clone, Serialize, Deserialize, PartialEq, Eq)]
11pub struct Span {
12    /// Starting line number (1-indexed)
13    pub start_line: usize,
14    /// Ending line number (1-indexed)
15    pub end_line: usize,
16}
17
18impl Span {
19    pub fn new(start_line: usize, start_col: usize, end_line: usize, end_col: usize) -> Self {
20        // Ignore col parameters for backwards compatibility
21        let _ = (start_col, end_col);
22        Self {
23            start_line,
24            end_line,
25        }
26    }
27}
28
29/// Type of symbol found in code
30#[derive(Debug, Clone, Serialize, Deserialize, PartialEq, Eq, EnumString, Display)]
31#[strum(serialize_all = "PascalCase")]
32pub enum SymbolKind {
33    Function,
34    Class,
35    Struct,
36    Enum,
37    Interface,
38    Trait,
39    Constant,
40    Variable,
41    Method,
42    Module,
43    Namespace,
44    Type,
45    Macro,
46    Property,
47    Event,
48    Import,
49    Export,
50    Attribute,
51    /// Catch-all for symbol kinds not yet explicitly supported.
52    /// This ensures no data loss when encountering new tree-sitter node types.
53    /// The string contains the original kind name from the parser.
54    #[strum(default)]
55    Unknown(String),
56}
57
58/// Programming language identifier
59#[derive(Debug, Clone, Copy, Serialize, Deserialize, PartialEq, Eq, Default)]
60#[serde(rename_all = "lowercase")]
61pub enum Language {
62    #[default]
63    Rust,
64    Python,
65    JavaScript,
66    TypeScript,
67    Vue,
68    Svelte,
69    Go,
70    Java,
71    PHP,
72    C,
73    Cpp,
74    CSharp,
75    Ruby,
76    Kotlin,
77    Swift,
78    Zig,
79    /// Plain-text tier: documentation, config and templates.
80    ///
81    /// Trigram-indexed only — no tree-sitter grammar, no symbol extraction, no
82    /// import extraction. Added because agents do not partition searches by file
83    /// type: a config key lives in the YAML, the Rust struct AND the spec paragraph,
84    /// and Reflex used to return the struct and a confident 0 for the rest.
85    ///
86    /// Serialises as `"text"` (the enum is `rename_all = "lowercase"`).
87    Text,
88    /// Lock files (`Cargo.lock`, `package-lock.json`, `*.lock`, …).
89    ///
90    /// Indexed since 2.0.0 in `[index] mode = "tracked"`, but excluded from every
91    /// search by default: 100k lines of pinned versions are noise for almost every
92    /// query and the one real question ("which lockfile pins serde 1.0.190?") is
93    /// asked with `include_locks: true` or `lang: "lock"`.
94    Lock,
95    /// Generated files judged by name: `*.pb.go`, `*_generated.*`, `*.generated.*`,
96    /// `*.min.js`, `*.min.css`, `*.map`. Indexed, excluded by default,
97    /// `include_generated: true` or `lang: "generated"` to widen.
98    Generated,
99    /// Never produced by [`Language::from_path`] since 2.0.0 (every path classifies
100    /// as code, `Text`, `Lock` or `Generated`); kept as the miss value of
101    /// [`Language::from_extension`] and as the forward-compatible sentinel for
102    /// readers of the `language` field.
103    Unknown,
104}
105
106/// Whether `file_name` is a dependency lock file.
107///
108/// Judged by full name: `package-lock.json` matches a text extension, and the
109/// name is the only thing that tells it apart from `settings.json`.
110pub fn is_lock_file(file_name: &str) -> bool {
111    if TEXT_FILENAME_EXCLUSIONS
112        .iter()
113        .any(|n| n.eq_ignore_ascii_case(file_name))
114    {
115        return true;
116    }
117    // `*-lock.json` and friends, beyond the names listed above.
118    file_name.ends_with("-lock.json") || file_name.ends_with(".lock")
119}
120
121/// Whether `file_name` looks like a generated file, judged by name only.
122///
123/// Content markers (`@generated` in the file head) are not read: the query engine
124/// derives a file's language from its path, so a content-based verdict at index
125/// time could not be honoured at query time.
126pub fn is_generated_name(file_name: &str) -> bool {
127    let lower = file_name.to_ascii_lowercase();
128    lower.ends_with(".pb.go")
129        || lower.ends_with(".min.js")
130        || lower.ends_with(".min.css")
131        || lower.ends_with(".map")
132        || lower.contains("_generated.")
133        || lower.contains(".generated.")
134}
135
136/// Extensions in the plain-text tier.
137///
138/// A fixed allowlist, not "everything unrecognised": an index that swallowed every
139/// binary blob and generated artefact in a repo would be slower and less useful.
140const TEXT_EXTENSIONS: &[&str] = &[
141    "md", "mdx", "txt", "yaml", "yml", "toml", "json", "proto", "html", "htm", "sh", "bash", "ini",
142    "cfg", "sql", "graphql", "bru",
143];
144
145/// Extensionless files in the plain-text tier, matched by exact name.
146///
147/// Agents grep these as readily as any `.md`; a `Makefile` target or a `Dockerfile`
148/// `COPY` line is a legitimate search hit. `Dockerfile.<variant>` is handled as a
149/// prefix in [`is_text_tier_file`].
150const TEXT_FILENAMES: &[&str] = &["Makefile", "makefile", "Dockerfile", "Justfile", "justfile"];
151
152/// Lock files, by exact name. See [`is_lock_file`] for the suffix rules.
153const TEXT_FILENAME_EXCLUSIONS: &[&str] = &[
154    "package-lock.json",
155    "npm-shrinkwrap.json",
156    "composer.lock",
157    "yarn.lock",
158    "pnpm-lock.yaml",
159    "Cargo.lock",
160    "poetry.lock",
161    "Gemfile.lock",
162    "go.sum",
163    "flake.lock",
164    "uv.lock",
165    "deno.lock",
166    "bun.lock",
167];
168
169/// Whether a file belongs in the ALLOWLIST text tier, judged by its full name.
170///
171/// This is the pre-2.0.0 rule, kept for `[index] mode = "allowlist"`. In the
172/// default `tracked` mode every non-binary, non-ignored file is text unless it is
173/// code, a lock file or a generated file.
174pub fn is_text_tier_file(file_name: &str) -> bool {
175    if is_lock_file(file_name) {
176        return false;
177    }
178    if TEXT_FILENAMES.contains(&file_name) || file_name.starts_with("Dockerfile.") {
179        return true;
180    }
181    match file_name.rsplit_once('.') {
182        Some((_, ext)) => TEXT_EXTENSIONS.contains(&ext.to_ascii_lowercase().as_str()),
183        None => false,
184    }
185}
186
187impl Language {
188    /// Classify a file by its path.
189    ///
190    /// This is the one classifier the indexer, watcher and query engine share, so a
191    /// file is either indexed, watched and searchable, or none of the three. Path
192    /// only: whether the file is binary is decided from its bytes by the indexer,
193    /// and a binary file is simply never in the index.
194    ///
195    /// * A lock file name wins (`Cargo.lock`, `package-lock.json`) → `Lock`.
196    /// * A generated name wins next (`x.pb.go`, `app.min.js`) → `Generated`.
197    /// * A recognised code extension → that language (`main.rs`, `app.mjs`).
198    /// * Everything else → `Text`: `README`, `OWNERS`, `foo.po`, `a.css`,
199    ///   `Makefile`, `.githooks/pre-commit`. Whether such a file is INDEXED is the
200    ///   `[index] mode` policy's decision (`PathPolicy::classify`).
201    pub fn from_path(path: &std::path::Path) -> Self {
202        let name = path.file_name().and_then(|n| n.to_str()).unwrap_or("");
203        if is_lock_file(name) {
204            return Language::Lock;
205        }
206        if is_generated_name(name) {
207            return Language::Generated;
208        }
209        let by_ext = path
210            .extension()
211            .and_then(|e| e.to_str())
212            .map(Self::from_extension)
213            .unwrap_or(Language::Unknown);
214        match by_ext {
215            Language::Text | Language::Unknown => Language::Text,
216            code => code,
217        }
218    }
219
220    pub fn from_extension(ext: &str) -> Self {
221        match ext {
222            "rs" => Language::Rust,
223            "py" => Language::Python,
224            "js" | "mjs" | "cjs" | "jsx" => Language::JavaScript,
225            "ts" | "mts" | "cts" | "tsx" => Language::TypeScript,
226            "vue" => Language::Vue,
227            "svelte" => Language::Svelte,
228            "go" => Language::Go,
229            "java" => Language::Java,
230            "php" => Language::PHP,
231            "c" | "h" => Language::C,
232            "cpp" | "cc" | "cxx" | "hpp" | "hxx" | "C" | "H" => Language::Cpp,
233            "cs" => Language::CSharp,
234            "rb" | "rake" | "gemspec" => Language::Ruby,
235            "kt" | "kts" => Language::Kotlin,
236            "swift" => Language::Swift,
237            "zig" => Language::Zig,
238            // The text tier. Note this maps by EXTENSION only; `is_text_tier_file`
239            // additionally excludes lock files by name, and the indexer uses that.
240            ext if TEXT_EXTENSIONS.contains(&ext.to_ascii_lowercase().as_str()) => Language::Text,
241            _ => Language::Unknown,
242        }
243    }
244
245    /// Parse a language from a human-friendly name (CLI/API input)
246    ///
247    /// Accepts lowercase names and common aliases.
248    /// Returns None for unrecognized names.
249    pub fn from_name(name: &str) -> Option<Self> {
250        match name.to_lowercase().as_str() {
251            "rust" | "rs" => Some(Language::Rust),
252            "python" | "py" => Some(Language::Python),
253            "javascript" | "js" => Some(Language::JavaScript),
254            "typescript" | "ts" => Some(Language::TypeScript),
255            "vue" => Some(Language::Vue),
256            "svelte" => Some(Language::Svelte),
257            "go" => Some(Language::Go),
258            "java" => Some(Language::Java),
259            "php" => Some(Language::PHP),
260            "c" => Some(Language::C),
261            "cpp" | "c++" => Some(Language::Cpp),
262            "csharp" | "cs" | "c#" => Some(Language::CSharp),
263            "ruby" | "rb" => Some(Language::Ruby),
264            "kotlin" | "kt" => Some(Language::Kotlin),
265            "swift" => Some(Language::Swift),
266            "zig" => Some(Language::Zig),
267            "text" | "txt" | "plaintext" | "plain" => Some(Language::Text),
268            "lock" | "lockfile" | "lockfiles" => Some(Language::Lock),
269            "generated" | "gen" => Some(Language::Generated),
270            _ => None,
271        }
272    }
273
274    /// Human-readable list of all supported language names (for error messages)
275    pub fn supported_names_help() -> &'static str {
276        "rust (rs), python (py), javascript (js), typescript (ts), vue, svelte, \
277         go, java, php, c, cpp (c++), csharp (cs, c#), ruby (rb), kotlin (kt), \
278         swift (full-text search only: no symbols), zig, \
279         text (every other non-binary file: docs, config, templates, extensionless), \
280         lock (lock files, excluded by default), generated (*.pb.go, *.min.js, *.map, \
281         *_generated.*, excluded by default)"
282    }
283
284    /// Check if this language has a parser implementation
285    ///
286    /// Returns true only for languages with working Tree-sitter parsers.
287    /// This determines which files will be indexed by Reflex.
288    pub fn is_supported(&self) -> bool {
289        match self {
290            Language::Rust => true,
291            Language::TypeScript => true,
292            Language::JavaScript => true,
293            Language::Vue => true,
294            Language::Svelte => true,
295            Language::Python => true,
296            Language::Go => true,
297            Language::Java => true,
298            Language::PHP => true,
299            Language::C => true,
300            Language::Cpp => true,
301            Language::CSharp => true,
302            Language::Ruby => true,
303            Language::Kotlin => true,
304            Language::Swift => false, // Temporarily disabled - parser queries out of date with tree-sitter-swift 0.7.x grammar
305            Language::Zig => true,
306            // No tree-sitter grammar, by design.
307            Language::Text => false,
308            Language::Lock => false,
309            Language::Generated => false,
310            Language::Unknown => false,
311        }
312    }
313
314    /// Whether this is the plain-text tier.
315    pub fn is_text(&self) -> bool {
316        matches!(self, Language::Text)
317    }
318
319    /// Whether files of this language are indexed but left out of every search
320    /// unless asked for (`include_locks` / `include_generated`, or `lang`).
321    pub fn is_excluded_by_default(&self) -> bool {
322        matches!(self, Language::Lock | Language::Generated)
323    }
324
325    /// Whether this is a code language, supported or not (Swift is code without a
326    /// working grammar). Code files are assumed to be text; every other file is
327    /// sniffed for a NUL byte before it is indexed.
328    pub fn is_code(&self) -> bool {
329        !matches!(
330            self,
331            Language::Text | Language::Lock | Language::Generated | Language::Unknown
332        )
333    }
334
335    /// Whether files of this language can be in the index at all.
336    ///
337    /// Distinct from [`Self::is_supported`], which means "has a tree-sitter parser".
338    /// The text tier is indexed but never parsed, so symbol search, AST queries and
339    /// dependency analysis skip it while full-text search covers it. Whether a given
340    /// path IS indexed also depends on `[index] mode` (`PathPolicy::classify`).
341    pub fn is_indexable(&self) -> bool {
342        !matches!(self, Language::Unknown)
343    }
344}
345
346/// Which files the indexer takes, beyond the code languages.
347#[derive(Debug, Clone, Copy, Serialize, Deserialize, PartialEq, Eq, Default)]
348#[serde(rename_all = "lowercase")]
349pub enum IndexMode {
350    /// Every file not ignored by `.gitignore` / `.ignore` / `[index] exclude`,
351    /// unless a NUL byte anywhere in it says it is binary. Lock and generated
352    /// files are indexed and excluded from searches by default. This is ripgrep's
353    /// rule, and what an agent that greps expects.
354    #[default]
355    Tracked,
356    /// The pre-2.0.0 rule: code by extension plus the fixed docs/config extension
357    /// list (`is_text_tier_file`). Lock and generated files are not indexed. For
358    /// trees where the long tail of data files is not worth the index size.
359    Allowlist,
360}
361
362impl IndexMode {
363    pub fn from_name(name: &str) -> Option<Self> {
364        match name.trim().to_ascii_lowercase().as_str() {
365            "tracked" => Some(Self::Tracked),
366            "allowlist" => Some(Self::Allowlist),
367            _ => None,
368        }
369    }
370}
371
372/// Type of import/dependency
373#[derive(Debug, Clone, Serialize, Deserialize, PartialEq, Eq)]
374#[serde(rename_all = "lowercase")]
375pub enum ImportType {
376    /// Internal project file
377    Internal,
378    /// External library/package
379    External,
380    /// Standard library
381    Stdlib,
382    /// Rust `mod foo;` declaration (parent→child ownership, not a usage edge)
383    #[serde(rename = "mod_decl")]
384    ModDecl,
385}
386
387/// Dependency information for API output (simplified, path-based)
388#[derive(Debug, Clone, Serialize, Deserialize)]
389pub struct DependencyInfo {
390    /// Import path as written in source (or resolved path for internal deps)
391    pub path: String,
392    /// Line number where import appears (optional)
393    #[serde(skip_serializing_if = "Option::is_none")]
394    pub line: Option<usize>,
395    /// Imported symbols (for selective imports like `from x import a, b`)
396    #[serde(skip_serializing_if = "Option::is_none")]
397    pub symbols: Option<Vec<String>>,
398    /// Every file a package import resolved to (a Go package, a C# namespace, a
399    /// JVM wildcard), in walk order; `path` is then the import as written.
400    #[serde(skip_serializing_if = "Option::is_none")]
401    pub resolved_paths: Option<Vec<String>>,
402}
403
404/// Full dependency record (internal representation with file IDs)
405#[derive(Debug, Clone)]
406pub struct Dependency {
407    /// Source file ID
408    pub file_id: i64,
409    /// Import path as written in source code
410    pub imported_path: String,
411    /// Resolved file ID (None if external or stdlib)
412    pub resolved_file_id: Option<i64>,
413    /// Package key the import resolved to (`go:<dir>`, `jvm:<package>`,
414    /// `cs:<namespace>`), when it names a package rather than one file. The graph
415    /// expands it to the files that declare membership (`package_members`).
416    pub resolved_package: Option<String>,
417    /// The member of `resolved_package` the import names (a JVM class); `None`
418    /// means the whole package.
419    pub resolved_member: Option<String>,
420    /// Import type classification
421    pub import_type: ImportType,
422    /// Line number where import appears
423    pub line_number: usize,
424    /// Imported symbols (for selective imports)
425    pub imported_symbols: Option<Vec<String>>,
426}
427
428/// A lightweight, stable reference to a code symbol for API responses
429///
430/// Prefer this over `(String, SymbolKind, Span)` tuples — tuples serialize as
431/// positional JSON arrays, making any field addition a breaking change.
432/// Named fields here are additive-safe: new optional fields can be added without
433/// shifting positions or bumping the version.
434#[derive(Debug, Clone, Serialize, Deserialize, PartialEq, Eq)]
435pub struct SymbolRef {
436    /// Symbol name (e.g., function name, class name)
437    pub name: String,
438    /// Symbol kind (function, class, struct, etc.)
439    pub kind: SymbolKind,
440    /// Location span in source file
441    pub span: Span,
442}
443
444/// Helper function to skip serializing "Unknown" symbol kinds
445fn is_unknown_kind(kind: &SymbolKind) -> bool {
446    matches!(kind, SymbolKind::Unknown(_))
447}
448
449/// A search result representing a symbol or code location
450#[derive(Debug, Clone, Serialize, Deserialize)]
451pub struct SearchResult {
452    /// Absolute or relative path to the file
453    pub path: String,
454    /// Detected programming language (internal use only, not serialized to save tokens)
455    #[serde(skip)]
456    pub lang: Language,
457    /// Type of symbol found (only included for symbol searches, not text matches)
458    #[serde(skip_serializing_if = "is_unknown_kind")]
459    pub kind: SymbolKind,
460    /// Symbol name (e.g., function name, class name)
461    /// None for text/regex matches where symbol name cannot be accurately determined
462    #[serde(skip_serializing_if = "Option::is_none")]
463    pub symbol: Option<String>,
464    /// Location span in the source file
465    pub span: Span,
466    /// Code preview (few lines around the match)
467    pub preview: String,
468    /// File dependencies (only populated when --dependencies flag is used)
469    /// DEPRECATED: Use FileGroupedResult.dependencies instead for file-level grouping
470    #[serde(skip_serializing_if = "Option::is_none")]
471    pub dependencies: Option<Vec<DependencyInfo>>,
472}
473
474/// An individual match within a file (no path or dependencies)
475#[derive(Debug, Clone, Serialize, Deserialize)]
476pub struct MatchResult {
477    /// Type of symbol found (only included for symbol searches, not text matches)
478    #[serde(skip_serializing_if = "is_unknown_kind")]
479    pub kind: SymbolKind,
480    /// Symbol name (e.g., function name, class name)
481    #[serde(skip_serializing_if = "Option::is_none")]
482    pub symbol: Option<String>,
483    /// Location span in the source file
484    pub span: Span,
485    /// Code preview (few lines around the match)
486    pub preview: String,
487    /// Lines of code before the match (for context)
488    #[serde(skip_serializing_if = "Vec::is_empty")]
489    pub context_before: Vec<String>,
490    /// Lines of code after the match (for context)
491    #[serde(skip_serializing_if = "Vec::is_empty")]
492    pub context_after: Vec<String>,
493}
494
495/// File-level grouped results with dependencies at file level
496#[derive(Debug, Clone, Serialize, Deserialize)]
497pub struct FileGroupedResult {
498    /// Absolute or relative path to the file
499    pub path: String,
500    /// Detected programming language of this file (e.g. "rust", "python", "unknown")
501    pub language: Language,
502    /// File dependencies (only populated when --dependencies flag is used)
503    #[serde(skip_serializing_if = "Option::is_none")]
504    pub dependencies: Option<Vec<DependencyInfo>>,
505    /// Individual matches within this file
506    pub matches: Vec<MatchResult>,
507}
508
509impl SearchResult {
510    pub fn new(
511        path: String,
512        lang: Language,
513        kind: SymbolKind,
514        symbol: Option<String>,
515        span: Span,
516        scope: Option<String>,
517        preview: String,
518    ) -> Self {
519        // Ignore scope parameter for backwards compatibility
520        let _ = scope;
521        Self {
522            path,
523            lang,
524            kind,
525            symbol,
526            span,
527            preview,
528            dependencies: None,
529        }
530    }
531}
532
533/// Configuration for indexing behavior
534/// [`IndexConfig::lock_wait_secs`]: wait for `index.lock` until it is free. The
535/// lock is advisory and released when its holder exits, so a crash cannot leave
536/// a waiter blocked.
537pub const LOCK_WAIT_FOREVER: u64 = u64::MAX;
538
539#[derive(Debug, Clone, Serialize, Deserialize)]
540pub struct IndexConfig {
541    /// Languages to include (empty = all supported)
542    pub languages: Vec<Language>,
543    /// Glob patterns to include
544    pub include_patterns: Vec<String>,
545    /// Glob patterns to exclude
546    pub exclude_patterns: Vec<String>,
547    /// Follow symbolic links
548    pub follow_symlinks: bool,
549    /// Maximum file size to index (bytes)
550    pub max_file_size: usize,
551    /// Number of threads for parallel indexing (0 = auto, 80% of available cores)
552    pub parallel_threads: usize,
553    /// Threads for the background symbol pass (`rfx index-symbols-internal`).
554    /// `0` = auto: half the cores, at most 32. See [`resolve_symbol_thread_count`].
555    #[serde(default)]
556    pub symbol_threads: usize,
557    /// Query timeout in seconds (0 = no timeout)
558    pub query_timeout_secs: u64,
559    /// Maximum entries per trigram posting list (0 = unlimited).
560    /// High-frequency trigrams are truncated at this threshold to bound query latency.
561    pub max_posting_list_entries: usize,
562    /// How long `Indexer::index` waits for `.reflex/index.lock` when another
563    /// indexer holds it (seconds). 0 = fail immediately with `IndexLocked`;
564    /// [`LOCK_WAIT_FOREVER`] = wait until it is free. The `rfx index` CLI waits
565    /// 30 s; automatic updates wait forever (a query never answers from a stale
566    /// index because another run was busy).
567    #[serde(default, skip_serializing_if = "is_zero_u64")]
568    pub lock_wait_secs: u64,
569    /// Index documentation, config and template files alongside code.
570    ///
571    /// On by default. Covers md, mdx, txt, yaml, yml, toml, json, proto, html, htm,
572    /// sh, bash, ini, cfg, sql and graphql, trigram-indexed only — no symbols, no
573    /// AST, no dependency analysis. Lock files are always excluded.
574    ///
575    /// Set `[index] text_tier = false` for a repo with large generated JSON or
576    /// vendored documentation where the index growth is not worth it.
577    ///
578    /// `#[serde(default = ...)]` so a config file written before 1.7.2 still parses
579    /// and gets the new default.
580    #[serde(default = "default_true")]
581    pub text_tier: bool,
582    /// Which non-code files the text tier takes: every non-binary, non-hidden, non-ignored file
583    /// (`tracked`, the default since 2.0.0) or the fixed extension list
584    /// (`allowlist`, the pre-2.0.0 rule).
585    #[serde(default)]
586    pub mode: IndexMode,
587    /// Walk dot-directories and dotfiles too (`.githooks/pre-commit`, `.env.example`).
588    /// Off by default, like ripgrep without `--hidden`. `.git/` and `.reflex/` are
589    /// never walked.
590    #[serde(default)]
591    pub hidden: bool,
592    /// `[index.vendored] patterns`: gitignore rules that mark directories vendored
593    /// (`!pattern` marks one project code), checked before the built-in rules.
594    /// Vendored files are searchable but not in the import graph (`crate::vendor`).
595    #[serde(default)]
596    pub vendored_patterns: Vec<String>,
597}
598
599/// Serde default for boolean options that are on unless explicitly disabled.
600fn default_true() -> bool {
601    true
602}
603
604impl Default for IndexConfig {
605    fn default() -> Self {
606        Self {
607            languages: vec![],
608            include_patterns: vec![],
609            exclude_patterns: vec![],
610            follow_symlinks: false,
611            max_file_size: 10 * 1024 * 1024,   // 10 MB
612            parallel_threads: 0,               // 0 = auto (80% of available cores)
613            symbol_threads: 0,                 // 0 = auto (50% of available cores)
614            query_timeout_secs: 30,            // 30 seconds default timeout
615            max_posting_list_entries: 500_000, // cap at 500k to bound query latency
616            text_tier: true,                   // docs and config are searchable by default
617            mode: IndexMode::Tracked, // ripgrep defaults: not ignored, not hidden, not binary
618            hidden: false,            // dot-directories skipped, like ripgrep
619            lock_wait_secs: 0,        // fail fast when another indexer runs
620            vendored_patterns: vec![],
621        }
622    }
623}
624
625fn is_zero(v: &usize) -> bool {
626    *v == 0
627}
628fn is_zero_u64(v: &u64) -> bool {
629    *v == 0
630}
631
632/// A `HashMap` serialized with its keys in order, so the same index prints the same JSON.
633fn serialize_sorted<S: serde::Serializer>(
634    map: &std::collections::HashMap<String, usize>,
635    serializer: S,
636) -> Result<S::Ok, S::Error> {
637    map.iter()
638        .collect::<std::collections::BTreeMap<_, _>>()
639        .serialize(serializer)
640}
641
642/// Resolve `[performance] symbol_threads` (the background symbol pass) to a
643/// concrete thread count.
644///
645/// `REFLEX_SYMBOL_THREADS` overrides for benchmarking, then `configured` if
646/// non-zero, else half the cores (1..=32). The pass is detached from `rfx index`
647/// and runs while the user may be querying, so it takes half the machine rather
648/// than the indexer's 80%; before 2.0.0 it took 27.5%.
649pub fn resolve_symbol_thread_count(configured: usize) -> usize {
650    if let Some(n) = std::env::var("REFLEX_SYMBOL_THREADS")
651        .ok()
652        .and_then(|v| v.parse::<usize>().ok())
653        .filter(|&n| n > 0)
654    {
655        return n;
656    }
657    if configured != 0 {
658        return configured.max(1);
659    }
660    let available = std::thread::available_parallelism()
661        .map(|n| n.get())
662        .unwrap_or(4);
663    ((available as f64 * 0.5).ceil() as usize).clamp(1, 32)
664}
665
666/// Resolve `[performance] parallel_threads` to a concrete thread count.
667///
668/// `0` means automatic: 80% of the available cores, at least 1, at most
669/// `auto_cap`. A non-zero value is used as given. The indexer passes a cap of 8
670/// (write-side cache contention); query-time verification passes a higher cap.
671pub fn resolve_thread_count(configured: usize, auto_cap: usize) -> usize {
672    if configured != 0 {
673        return configured.max(1);
674    }
675    let available = std::thread::available_parallelism()
676        .map(|n| n.get())
677        .unwrap_or(4);
678    ((available as f64 * 0.8).ceil() as usize).clamp(1, auto_cap.max(1))
679}
680
681/// How a query found its candidate lines.
682///
683/// `trigram` is the normal case: the pattern's literals were looked up in the
684/// inverted index and only the lines they name were verified. `scan` means every
685/// line of every file was verified, which happens for a pattern shorter than
686/// 3 chars, a regex with no literal of 3+ chars (`\w+_id`), a non-ASCII literal
687/// under `(?i)`, or a keyword symbol query. Before 2.0.0 every `(?i)` regex
688/// scanned; since 2.0.0 its literals are looked up under all case variants.
689#[derive(Debug, Clone, Copy, Default, Serialize, Deserialize, PartialEq, Eq)]
690#[serde(rename_all = "lowercase")]
691pub enum IndexPath {
692    /// Candidates came from the trigram index.
693    #[default]
694    Trigram,
695    /// Every line was verified.
696    Scan,
697}
698
699/// Per-phase wall-clock timings for one query, in microseconds.
700///
701/// Present in a [`QueryResponse`] only when the caller asked for it
702/// (`rfx query --timing`, or `REFLEX_MCP_TIMING=1` for the MCP server).
703#[derive(Debug, Clone, Default, Serialize, Deserialize, PartialEq, Eq)]
704pub struct QueryTimings {
705    /// Whether the trigram index or a full scan produced the candidates.
706    #[serde(default)]
707    pub index_path: IndexPath,
708    /// Opening (or reusing) the index handle.
709    pub open_us: u64,
710    /// Trigram lookup and posting-list intersection.
711    pub candidates_us: u64,
712    /// Verifying candidate lines against the pattern (and any enrichment).
713    pub verify_us: u64,
714    /// Time the query waited for the freshness check after the search finished.
715    /// The check runs on its own thread alongside the search, so this is usually
716    /// near zero; `status_compute_us` is what the check itself cost.
717    pub status_us: u64,
718    /// The freshness check's own duration (git spawns, or a tree walk outside git).
719    #[serde(default)]
720    pub status_compute_us: u64,
721    /// Grouping by file, context lines, dependencies.
722    pub group_us: u64,
723    /// Whole query as seen by the engine.
724    pub total_us: u64,
725    /// Automatic updates of a stale index before the answer (0: none ran).
726    /// `total_us` includes them.
727    #[serde(default, skip_serializing_if = "is_zero_u64")]
728    pub update_us: u64,
729}
730
731/// Statistics about the index
732#[derive(Debug, Clone, Serialize, Deserialize, Default)]
733pub struct IndexStats {
734    /// Total files indexed
735    pub total_files: usize,
736    /// Index size on disk (bytes)
737    pub index_size_bytes: u64,
738    /// Last update timestamp
739    pub last_updated: String,
740    /// File count breakdown by language
741    #[serde(serialize_with = "serialize_sorted")]
742    pub files_by_language: std::collections::HashMap<String, usize>,
743    /// Line count breakdown by language
744    #[serde(serialize_with = "serialize_sorted")]
745    pub lines_by_language: std::collections::HashMap<String, usize>,
746    /// New files added since last index run (0 if not an incremental run)
747    #[serde(default, skip_serializing_if = "is_zero")]
748    pub new_files: usize,
749    /// Modified files re-indexed since last run (0 if not an incremental run)
750    #[serde(default, skip_serializing_if = "is_zero")]
751    pub modified_files: usize,
752    /// Unchanged files (same hash as last run, still re-indexed due to other changes)
753    #[serde(default, skip_serializing_if = "is_zero")]
754    pub unchanged_files: usize,
755    /// Files dropped from the index because they no longer exist on disk
756    #[serde(default, skip_serializing_if = "is_zero")]
757    pub deleted_files: usize,
758    /// Files skipped because they exceeded max_file_size
759    #[serde(default, skip_serializing_if = "is_zero")]
760    pub skipped_too_large: usize,
761    /// Total bytes of files skipped due to max_file_size
762    #[serde(default, skip_serializing_if = "is_zero_u64")]
763    pub skipped_bytes_too_large: u64,
764    /// Files skipped because a NUL byte says they are binary (ripgrep's rule)
765    #[serde(default, skip_serializing_if = "is_zero")]
766    pub skipped_binary: usize,
767    /// Raw bytes of indexed source held in content.bin (0 if unknown)
768    #[serde(default, skip_serializing_if = "is_zero_u64")]
769    pub corpus_bytes: u64,
770    /// Size of trigrams.bin on disk (0 if absent)
771    #[serde(default, skip_serializing_if = "is_zero_u64")]
772    pub trigram_index_bytes: u64,
773}
774
775/// Information about an indexed file
776#[derive(Debug, Clone, Serialize, Deserialize)]
777pub struct IndexedFile {
778    /// File path
779    pub path: String,
780    /// Detected language
781    pub language: String,
782    /// Last indexed timestamp
783    pub last_indexed: String,
784}
785
786/// Index status for query responses
787#[derive(Debug, Clone, Serialize, Deserialize, PartialEq, Eq)]
788#[serde(rename_all = "snake_case")]
789pub enum IndexStatus {
790    /// Index is fresh and up-to-date
791    Fresh,
792    /// Index is stale (any issue: branch not indexed, commit changed, files modified)
793    Stale,
794}
795
796/// Warning details when index is stale
797#[derive(Debug, Clone, Serialize, Deserialize)]
798pub struct IndexWarning {
799    /// Human-readable reason why index is stale
800    pub reason: String,
801    /// Command to run to fix the issue
802    pub action_required: String,
803    /// Tracked files edited since the index was built.
804    ///
805    /// BREAKING in 1.7.2: this was a `u32` count. It is now the paths themselves,
806    /// because a count told an agent something was wrong without telling it what, so
807    /// the only safe reaction was to distrust the whole result. The list is capped;
808    /// `truncated` says when.
809    #[serde(skip_serializing_if = "Option::is_none")]
810    pub files_modified: Option<Vec<String>>,
811    /// Files present on disk but absent from the index (new or untracked).
812    #[serde(skip_serializing_if = "Option::is_none")]
813    pub files_added: Option<Vec<String>>,
814    /// Files in the index but no longer on disk. These produce ghost hits.
815    #[serde(skip_serializing_if = "Option::is_none")]
816    pub files_deleted: Option<Vec<String>>,
817    /// Total changed paths, which may exceed the lengths of the lists above.
818    #[serde(skip_serializing_if = "Option::is_none")]
819    pub changed_count: Option<usize>,
820    /// Whether the lists were cut short. Set on a fresh checkout or a huge rebase.
821    #[serde(default, skip_serializing_if = "std::ops::Not::not")]
822    pub truncated: bool,
823    /// Additional context (git branch info, etc.)
824    #[serde(skip_serializing_if = "Option::is_none")]
825    pub details: Option<IndexWarningDetails>,
826}
827
828impl IndexWarning {
829    /// A warning with only a reason and an action, no file lists.
830    pub fn new(reason: impl Into<String>, action_required: impl Into<String>) -> Self {
831        Self {
832            reason: reason.into(),
833            action_required: action_required.into(),
834            files_modified: None,
835            files_added: None,
836            files_deleted: None,
837            changed_count: None,
838            truncated: false,
839            details: None,
840        }
841    }
842
843    /// Attach git branch/commit context.
844    pub fn with_details(mut self, details: IndexWarningDetails) -> Self {
845        self.details = Some(details);
846        self
847    }
848}
849
850/// Detailed information about index staleness
851#[derive(Debug, Clone, Serialize, Deserialize)]
852pub struct IndexWarningDetails {
853    /// Current branch (if in git repo)
854    #[serde(skip_serializing_if = "Option::is_none")]
855    pub current_branch: Option<String>,
856    /// Indexed branch (if in git repo)
857    #[serde(skip_serializing_if = "Option::is_none")]
858    pub indexed_branch: Option<String>,
859    /// Current commit SHA (if in git repo)
860    #[serde(skip_serializing_if = "Option::is_none")]
861    pub current_commit: Option<String>,
862    /// Indexed commit SHA (if in git repo)
863    #[serde(skip_serializing_if = "Option::is_none")]
864    pub indexed_commit: Option<String>,
865    /// When the index was last written (Unix seconds).
866    #[serde(default, skip_serializing_if = "Option::is_none")]
867    pub indexed_at: Option<i64>,
868    /// How the working tree was compared to the index: `"git"` (candidates from
869    /// `git status`, confirmed by fingerprint) or `"walk"` (every file stat'ed).
870    #[serde(default, skip_serializing_if = "Option::is_none")]
871    pub checked_by: Option<String>,
872}
873
874/// The full answer to `check_index_status`.
875///
876/// `details` is present even when the index is fresh, so a human can see that the
877/// indexed commit differs from HEAD without that difference being called staleness:
878/// since 2.0.0 freshness is judged by file content, not by commit.
879#[derive(Debug, Clone)]
880pub struct IndexStatusReport {
881    pub status: IndexStatus,
882    pub can_trust_results: bool,
883    pub warning: Option<IndexWarning>,
884    pub details: Option<IndexWarningDetails>,
885}
886
887/// Pagination information for query results
888#[derive(Debug, Clone, Serialize, Deserialize)]
889pub struct PaginationInfo {
890    /// Total number of results before offset/limit — **`null` whenever
891    /// `total_is_exact` is false**. A list-mode search with a `limit` stops
892    /// verifying once the page is full, and the number verified by then is not a
893    /// total; reporting it as one made agents stop paginating early (2.0.0 field
894    /// test: `total: 851` for a term with 18,752 matches). Use `approx_total` for an
895    /// estimate, or a count-mode / no-limit search for the exact number.
896    #[serde(default)]
897    pub total: Option<usize>,
898    /// Number of results in this response (after offset/limit)
899    pub count: usize,
900    /// Offset used (starting position)
901    pub offset: usize,
902    /// Limit used (max results per page)
903    #[serde(skip_serializing_if = "Option::is_none")]
904    pub limit: Option<usize>,
905    /// Whether there are more results after this page
906    pub has_more: bool,
907    /// `true` when `total` counts every match. `false` when a list-mode search
908    /// stopped verifying once the page was full: `total` is then `null` and
909    /// `approx_total` carries an estimate. Count mode, no-limit searches,
910    /// symbol/AST searches and `require_exact_total` callers are always exact.
911    #[serde(default = "default_true")]
912    pub total_is_exact: bool,
913    /// Estimated total when `total_is_exact` is false. After the page filled, a
914    /// spread sample of the remaining candidate lines was verified and the
915    /// measured hit rate scaled over the rest (see
916    /// `query::ESTIMATE_SAMPLE_FILES`). Typically within ±30% on the synthetic
917    /// corpus; wider when hits cluster in a few files. Absent when no estimate was
918    /// possible (a regex with no literal, where every line is a candidate).
919    #[serde(default, skip_serializing_if = "Option::is_none")]
920    pub approx_total: Option<usize>,
921}
922
923impl PaginationInfo {
924    /// The total, only when it is a real one.
925    pub fn exact_total(&self) -> Option<usize> {
926        if self.total_is_exact {
927            self.total
928        } else {
929            None
930        }
931    }
932
933    /// The best number available for a threshold or a log line: the exact total,
934    /// else the estimate, else the end of this page. Never show it as a total.
935    pub fn best_total(&self) -> usize {
936        self.exact_total()
937            .or(self.approx_total)
938            .unwrap_or(self.offset + self.count)
939    }
940}
941
942/// Query response with results and index status
943#[derive(Debug, Clone, Serialize, Deserialize)]
944pub struct QueryResponse {
945    /// AI-optimized instruction for how to handle these results
946    /// Only present when --ai flag is used or in MCP mode
947    /// Provides guidance to AI agents on response format and next actions
948    #[serde(skip_serializing_if = "Option::is_none")]
949    pub ai_instruction: Option<String>,
950    /// Status of the index (fresh or stale)
951    pub status: IndexStatus,
952    /// Whether the results can be trusted
953    pub can_trust_results: bool,
954    /// Warning information (only present if stale)
955    #[serde(skip_serializing_if = "Option::is_none")]
956    pub warning: Option<IndexWarning>,
957    /// Pagination information
958    pub pagination: PaginationInfo,
959    /// File-grouped search results
960    /// Results are always grouped by file path, with dependencies populated when --dependencies flag is used
961    pub results: Vec<FileGroupedResult>,
962    /// For a whole-identifier search that found nothing: how many candidate lines
963    /// contain the pattern as a substring. Lets the caller explain a zero without
964    /// running a second search. Absent whenever there were results, or when the
965    /// search was not a whole-identifier one.
966    #[serde(default, skip_serializing_if = "Option::is_none")]
967    pub substring_hint_count: Option<usize>,
968    /// Things the engine did to the query that the caller should know about.
969    ///
970    /// Today this is one message at most: a whole-identifier pattern containing
971    /// brackets was escaped and run as a regex (see
972    /// `query::prepare_literal_pattern`). Every surface — CLI, MCP, HTTP — carries it,
973    /// so a rewrite is never silent on any of them.
974    #[serde(default, skip_serializing_if = "Vec::is_empty")]
975    pub warnings: Vec<String>,
976    /// For a whole-identifier search that found nothing while substring matches
977    /// exist: a ready-to-show sentence naming the count and the switch that shows
978    /// them. Absent whenever there were results, or when the search was not a
979    /// whole-identifier one.
980    #[serde(default, skip_serializing_if = "Option::is_none")]
981    pub hint: Option<String>,
982    /// Machine-readable cause behind `hint`: `hidden` (the filter names a
983    /// dot-directory), `not_indexed` (the `file` filter names a path the index does
984    /// not hold), `lock_or_generated`, `whole_identifier`. Absent when no rule
985    /// applies, so a harness can branch without parsing prose.
986    #[serde(default, skip_serializing_if = "Option::is_none")]
987    pub excluded_reason: Option<crate::query::ExcludedReason>,
988    /// For a search that found nothing: how many candidate files were lock or
989    /// generated files, which every search leaves out unless `include_locks` /
990    /// `include_generated` (or `lang`) asks for them. The `hint` says so too.
991    #[serde(default, skip_serializing_if = "Option::is_none")]
992    pub excluded_by_default: Option<usize>,
993    /// Count-only searches: the number of files with at least one match. `results`
994    /// is empty in that mode, so this is the only place the file count lives.
995    #[serde(default, skip_serializing_if = "Option::is_none")]
996    pub file_count: Option<usize>,
997    /// Per-phase timings, only when requested.
998    #[serde(default, skip_serializing_if = "Option::is_none")]
999    pub timings: Option<QueryTimings>,
1000}
1001
1002/// Report from cache compaction operation
1003#[derive(Debug, Clone, Serialize, Deserialize)]
1004pub struct CompactionReport {
1005    /// Always 0 since auto-update: index runs remove deleted files (kept for JSON compatibility)
1006    pub files_removed: usize,
1007    /// Space saved in bytes
1008    pub space_saved_bytes: u64,
1009    /// Duration in milliseconds
1010    pub duration_ms: u64,
1011}
1012
1013#[cfg(test)]
1014mod tests {
1015    use super::*;
1016
1017    /// `rfx stats --json` printed the language maps in `HashMap` order through 2.2.0.
1018    #[test]
1019    fn index_stats_json_lists_languages_in_order() {
1020        let names = ["Zig", "Rust", "Text", "Go", "Python", "C", "Kotlin", "Ruby"];
1021        let stats = IndexStats {
1022            files_by_language: names.iter().map(|n| (n.to_string(), 1)).collect(),
1023            lines_by_language: names.iter().map(|n| (n.to_string(), 10)).collect(),
1024            ..Default::default()
1025        };
1026        // A string, not a `Value`: serde_json's `Map` would sort the keys itself.
1027        let json = serde_json::to_string(&stats).unwrap();
1028        for field in ["files_by_language", "lines_by_language"] {
1029            let section = &json[json.find(field).unwrap()..];
1030            let section = &section[..section.find('}').unwrap()];
1031            let positions: Vec<usize> =
1032                ["C", "Go", "Kotlin", "Python", "Ruby", "Rust", "Text", "Zig"]
1033                    .iter()
1034                    .map(|n| section.find(&format!("\"{n}\"")).unwrap())
1035                    .collect();
1036            assert!(positions.is_sorted(), "{field}: {section}");
1037        }
1038    }
1039
1040    #[test]
1041    fn test_symbol_ref_json_shape() {
1042        let sym = SymbolRef {
1043            name: "my_function".to_string(),
1044            kind: SymbolKind::Function,
1045            span: Span {
1046                start_line: 10,
1047                end_line: 20,
1048            },
1049        };
1050        let json = serde_json::to_value(&sym).unwrap();
1051        assert_eq!(json["name"], "my_function");
1052        assert_eq!(json["kind"], "Function");
1053        assert_eq!(json["span"]["start_line"], 10);
1054        assert_eq!(json["span"]["end_line"], 20);
1055        assert!(json.as_array().is_none());
1056    }
1057
1058    #[test]
1059    fn test_symbol_ref_roundtrip() {
1060        let original = SymbolRef {
1061            name: "MyStruct".to_string(),
1062            kind: SymbolKind::Struct,
1063            span: Span {
1064                start_line: 1,
1065                end_line: 5,
1066            },
1067        };
1068        let json = serde_json::to_string(&original).unwrap();
1069        let decoded: SymbolRef = serde_json::from_str(&json).unwrap();
1070        assert_eq!(original, decoded);
1071    }
1072
1073    #[test]
1074    fn test_symbol_ref_exact_json() {
1075        let sym = SymbolRef {
1076            name: "Foo".to_string(),
1077            kind: SymbolKind::Class,
1078            span: Span {
1079                start_line: 3,
1080                end_line: 7,
1081            },
1082        };
1083        let json = serde_json::to_string(&sym).unwrap();
1084        assert_eq!(
1085            json,
1086            r#"{"name":"Foo","kind":"Class","span":{"start_line":3,"end_line":7}}"#
1087        );
1088    }
1089}