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            "zig" => Some(Language::Zig),
266            "text" | "txt" | "plaintext" | "plain" => Some(Language::Text),
267            "lock" | "lockfile" | "lockfiles" => Some(Language::Lock),
268            "generated" | "gen" => Some(Language::Generated),
269            _ => None,
270        }
271    }
272
273    /// Human-readable list of all supported language names (for error messages)
274    pub fn supported_names_help() -> &'static str {
275        "rust (rs), python (py), javascript (js), typescript (ts), vue, svelte, \
276         go, java, php, c, cpp (c++), csharp (cs, c#), ruby (rb), kotlin (kt), zig, \
277         text (every other non-binary file: docs, config, templates, extensionless), \
278         lock (lock files, excluded by default), generated (*.pb.go, *.min.js, *.map, \
279         *_generated.*, excluded by default)"
280    }
281
282    /// Check if this language has a parser implementation
283    ///
284    /// Returns true only for languages with working Tree-sitter parsers.
285    /// This determines which files will be indexed by Reflex.
286    pub fn is_supported(&self) -> bool {
287        match self {
288            Language::Rust => true,
289            Language::TypeScript => true,
290            Language::JavaScript => true,
291            Language::Vue => true,
292            Language::Svelte => true,
293            Language::Python => true,
294            Language::Go => true,
295            Language::Java => true,
296            Language::PHP => true,
297            Language::C => true,
298            Language::Cpp => true,
299            Language::CSharp => true,
300            Language::Ruby => true,
301            Language::Kotlin => true,
302            Language::Swift => false, // Temporarily disabled - parser queries out of date with tree-sitter-swift 0.7.x grammar
303            Language::Zig => true,
304            // No tree-sitter grammar, by design.
305            Language::Text => false,
306            Language::Lock => false,
307            Language::Generated => false,
308            Language::Unknown => false,
309        }
310    }
311
312    /// Whether this is the plain-text tier.
313    pub fn is_text(&self) -> bool {
314        matches!(self, Language::Text)
315    }
316
317    /// Whether files of this language are indexed but left out of every search
318    /// unless asked for (`include_locks` / `include_generated`, or `lang`).
319    pub fn is_excluded_by_default(&self) -> bool {
320        matches!(self, Language::Lock | Language::Generated)
321    }
322
323    /// Whether this is a code language, supported or not (Swift is code without a
324    /// working grammar). Code files are assumed to be text; every other file is
325    /// sniffed for a NUL byte before it is indexed.
326    pub fn is_code(&self) -> bool {
327        !matches!(
328            self,
329            Language::Text | Language::Lock | Language::Generated | Language::Unknown
330        )
331    }
332
333    /// Whether files of this language can be in the index at all.
334    ///
335    /// Distinct from [`Self::is_supported`], which means "has a tree-sitter parser".
336    /// The text tier is indexed but never parsed, so symbol search, AST queries and
337    /// dependency analysis skip it while full-text search covers it. Whether a given
338    /// path IS indexed also depends on `[index] mode` (`PathPolicy::classify`).
339    pub fn is_indexable(&self) -> bool {
340        !matches!(self, Language::Unknown)
341    }
342}
343
344/// Which files the indexer takes, beyond the code languages.
345#[derive(Debug, Clone, Copy, Serialize, Deserialize, PartialEq, Eq, Default)]
346#[serde(rename_all = "lowercase")]
347pub enum IndexMode {
348    /// Every file not ignored by `.gitignore` / `.ignore` / `[index] exclude`,
349    /// unless a NUL byte anywhere in it says it is binary. Lock and generated
350    /// files are indexed and excluded from searches by default. This is ripgrep's
351    /// rule, and what an agent that greps expects.
352    #[default]
353    Tracked,
354    /// The pre-2.0.0 rule: code by extension plus the fixed docs/config extension
355    /// list (`is_text_tier_file`). Lock and generated files are not indexed. For
356    /// trees where the long tail of data files is not worth the index size.
357    Allowlist,
358}
359
360impl IndexMode {
361    pub fn from_name(name: &str) -> Option<Self> {
362        match name.trim().to_ascii_lowercase().as_str() {
363            "tracked" => Some(Self::Tracked),
364            "allowlist" => Some(Self::Allowlist),
365            _ => None,
366        }
367    }
368}
369
370/// Type of import/dependency
371#[derive(Debug, Clone, Serialize, Deserialize, PartialEq, Eq)]
372#[serde(rename_all = "lowercase")]
373pub enum ImportType {
374    /// Internal project file
375    Internal,
376    /// External library/package
377    External,
378    /// Standard library
379    Stdlib,
380    /// Rust `mod foo;` declaration (parent→child ownership, not a usage edge)
381    #[serde(rename = "mod_decl")]
382    ModDecl,
383}
384
385/// Dependency information for API output (simplified, path-based)
386#[derive(Debug, Clone, Serialize, Deserialize)]
387pub struct DependencyInfo {
388    /// Import path as written in source (or resolved path for internal deps)
389    pub path: String,
390    /// Line number where import appears (optional)
391    #[serde(skip_serializing_if = "Option::is_none")]
392    pub line: Option<usize>,
393    /// Imported symbols (for selective imports like `from x import a, b`)
394    #[serde(skip_serializing_if = "Option::is_none")]
395    pub symbols: Option<Vec<String>>,
396}
397
398/// Full dependency record (internal representation with file IDs)
399#[derive(Debug, Clone)]
400pub struct Dependency {
401    /// Source file ID
402    pub file_id: i64,
403    /// Import path as written in source code
404    pub imported_path: String,
405    /// Resolved file ID (None if external or stdlib)
406    pub resolved_file_id: Option<i64>,
407    /// Import type classification
408    pub import_type: ImportType,
409    /// Line number where import appears
410    pub line_number: usize,
411    /// Imported symbols (for selective imports)
412    pub imported_symbols: Option<Vec<String>>,
413}
414
415/// A lightweight, stable reference to a code symbol for API responses
416///
417/// Prefer this over `(String, SymbolKind, Span)` tuples — tuples serialize as
418/// positional JSON arrays, making any field addition a breaking change.
419/// Named fields here are additive-safe: new optional fields can be added without
420/// shifting positions or bumping the version.
421#[derive(Debug, Clone, Serialize, Deserialize, PartialEq, Eq)]
422pub struct SymbolRef {
423    /// Symbol name (e.g., function name, class name)
424    pub name: String,
425    /// Symbol kind (function, class, struct, etc.)
426    pub kind: SymbolKind,
427    /// Location span in source file
428    pub span: Span,
429}
430
431/// Helper function to skip serializing "Unknown" symbol kinds
432fn is_unknown_kind(kind: &SymbolKind) -> bool {
433    matches!(kind, SymbolKind::Unknown(_))
434}
435
436/// A search result representing a symbol or code location
437#[derive(Debug, Clone, Serialize, Deserialize)]
438pub struct SearchResult {
439    /// Absolute or relative path to the file
440    pub path: String,
441    /// Detected programming language (internal use only, not serialized to save tokens)
442    #[serde(skip)]
443    pub lang: Language,
444    /// Type of symbol found (only included for symbol searches, not text matches)
445    #[serde(skip_serializing_if = "is_unknown_kind")]
446    pub kind: SymbolKind,
447    /// Symbol name (e.g., function name, class name)
448    /// None for text/regex matches where symbol name cannot be accurately determined
449    #[serde(skip_serializing_if = "Option::is_none")]
450    pub symbol: Option<String>,
451    /// Location span in the source file
452    pub span: Span,
453    /// Code preview (few lines around the match)
454    pub preview: String,
455    /// File dependencies (only populated when --dependencies flag is used)
456    /// DEPRECATED: Use FileGroupedResult.dependencies instead for file-level grouping
457    #[serde(skip_serializing_if = "Option::is_none")]
458    pub dependencies: Option<Vec<DependencyInfo>>,
459}
460
461/// An individual match within a file (no path or dependencies)
462#[derive(Debug, Clone, Serialize, Deserialize)]
463pub struct MatchResult {
464    /// Type of symbol found (only included for symbol searches, not text matches)
465    #[serde(skip_serializing_if = "is_unknown_kind")]
466    pub kind: SymbolKind,
467    /// Symbol name (e.g., function name, class name)
468    #[serde(skip_serializing_if = "Option::is_none")]
469    pub symbol: Option<String>,
470    /// Location span in the source file
471    pub span: Span,
472    /// Code preview (few lines around the match)
473    pub preview: String,
474    /// Lines of code before the match (for context)
475    #[serde(skip_serializing_if = "Vec::is_empty")]
476    pub context_before: Vec<String>,
477    /// Lines of code after the match (for context)
478    #[serde(skip_serializing_if = "Vec::is_empty")]
479    pub context_after: Vec<String>,
480}
481
482/// File-level grouped results with dependencies at file level
483#[derive(Debug, Clone, Serialize, Deserialize)]
484pub struct FileGroupedResult {
485    /// Absolute or relative path to the file
486    pub path: String,
487    /// Detected programming language of this file (e.g. "rust", "python", "unknown")
488    pub language: Language,
489    /// File dependencies (only populated when --dependencies flag is used)
490    #[serde(skip_serializing_if = "Option::is_none")]
491    pub dependencies: Option<Vec<DependencyInfo>>,
492    /// Individual matches within this file
493    pub matches: Vec<MatchResult>,
494}
495
496impl SearchResult {
497    pub fn new(
498        path: String,
499        lang: Language,
500        kind: SymbolKind,
501        symbol: Option<String>,
502        span: Span,
503        scope: Option<String>,
504        preview: String,
505    ) -> Self {
506        // Ignore scope parameter for backwards compatibility
507        let _ = scope;
508        Self {
509            path,
510            lang,
511            kind,
512            symbol,
513            span,
514            preview,
515            dependencies: None,
516        }
517    }
518}
519
520/// Configuration for indexing behavior
521/// [`IndexConfig::lock_wait_secs`]: wait for `index.lock` until it is free. The
522/// lock is advisory and released when its holder exits, so a crash cannot leave
523/// a waiter blocked.
524pub const LOCK_WAIT_FOREVER: u64 = u64::MAX;
525
526#[derive(Debug, Clone, Serialize, Deserialize)]
527pub struct IndexConfig {
528    /// Languages to include (empty = all supported)
529    pub languages: Vec<Language>,
530    /// Glob patterns to include
531    pub include_patterns: Vec<String>,
532    /// Glob patterns to exclude
533    pub exclude_patterns: Vec<String>,
534    /// Follow symbolic links
535    pub follow_symlinks: bool,
536    /// Maximum file size to index (bytes)
537    pub max_file_size: usize,
538    /// Number of threads for parallel indexing (0 = auto, 80% of available cores)
539    pub parallel_threads: usize,
540    /// Threads for the background symbol pass (`rfx index-symbols-internal`).
541    /// `0` = auto: half the cores, at most 32. See [`resolve_symbol_thread_count`].
542    #[serde(default)]
543    pub symbol_threads: usize,
544    /// Query timeout in seconds (0 = no timeout)
545    pub query_timeout_secs: u64,
546    /// Maximum entries per trigram posting list (0 = unlimited).
547    /// High-frequency trigrams are truncated at this threshold to bound query latency.
548    pub max_posting_list_entries: usize,
549    /// How long `Indexer::index` waits for `.reflex/index.lock` when another
550    /// indexer holds it (seconds). 0 = fail immediately with `IndexLocked`;
551    /// [`LOCK_WAIT_FOREVER`] = wait until it is free. The `rfx index` CLI waits
552    /// 30 s; automatic updates wait forever (a query never answers from a stale
553    /// index because another run was busy).
554    #[serde(default, skip_serializing_if = "is_zero_u64")]
555    pub lock_wait_secs: u64,
556    /// Index documentation, config and template files alongside code.
557    ///
558    /// On by default. Covers md, mdx, txt, yaml, yml, toml, json, proto, html, htm,
559    /// sh, bash, ini, cfg, sql and graphql, trigram-indexed only — no symbols, no
560    /// AST, no dependency analysis. Lock files are always excluded.
561    ///
562    /// Set `[index] text_tier = false` for a repo with large generated JSON or
563    /// vendored documentation where the index growth is not worth it.
564    ///
565    /// `#[serde(default = ...)]` so a config file written before 1.7.2 still parses
566    /// and gets the new default.
567    #[serde(default = "default_true")]
568    pub text_tier: bool,
569    /// Which non-code files the text tier takes: every non-binary, non-hidden, non-ignored file
570    /// (`tracked`, the default since 2.0.0) or the fixed extension list
571    /// (`allowlist`, the pre-2.0.0 rule).
572    #[serde(default)]
573    pub mode: IndexMode,
574    /// Walk dot-directories and dotfiles too (`.githooks/pre-commit`, `.env.example`).
575    /// Off by default, like ripgrep without `--hidden`. `.git/` and `.reflex/` are
576    /// never walked.
577    #[serde(default)]
578    pub hidden: bool,
579}
580
581/// Serde default for boolean options that are on unless explicitly disabled.
582fn default_true() -> bool {
583    true
584}
585
586impl Default for IndexConfig {
587    fn default() -> Self {
588        Self {
589            languages: vec![],
590            include_patterns: vec![],
591            exclude_patterns: vec![],
592            follow_symlinks: false,
593            max_file_size: 10 * 1024 * 1024,   // 10 MB
594            parallel_threads: 0,               // 0 = auto (80% of available cores)
595            symbol_threads: 0,                 // 0 = auto (50% of available cores)
596            query_timeout_secs: 30,            // 30 seconds default timeout
597            max_posting_list_entries: 500_000, // cap at 500k to bound query latency
598            text_tier: true,                   // docs and config are searchable by default
599            mode: IndexMode::Tracked, // ripgrep defaults: not ignored, not hidden, not binary
600            hidden: false,            // dot-directories skipped, like ripgrep
601            lock_wait_secs: 0,        // fail fast when another indexer runs
602        }
603    }
604}
605
606fn is_zero(v: &usize) -> bool {
607    *v == 0
608}
609fn is_zero_u64(v: &u64) -> bool {
610    *v == 0
611}
612
613/// Resolve `[performance] symbol_threads` (the background symbol pass) to a
614/// concrete thread count.
615///
616/// `REFLEX_SYMBOL_THREADS` overrides for benchmarking, then `configured` if
617/// non-zero, else half the cores (1..=32). The pass is detached from `rfx index`
618/// and runs while the user may be querying, so it takes half the machine rather
619/// than the indexer's 80%; before 2.0.0 it took 27.5%.
620pub fn resolve_symbol_thread_count(configured: usize) -> usize {
621    if let Some(n) = std::env::var("REFLEX_SYMBOL_THREADS")
622        .ok()
623        .and_then(|v| v.parse::<usize>().ok())
624        .filter(|&n| n > 0)
625    {
626        return n;
627    }
628    if configured != 0 {
629        return configured.max(1);
630    }
631    let available = std::thread::available_parallelism()
632        .map(|n| n.get())
633        .unwrap_or(4);
634    ((available as f64 * 0.5).ceil() as usize).clamp(1, 32)
635}
636
637/// Resolve `[performance] parallel_threads` to a concrete thread count.
638///
639/// `0` means automatic: 80% of the available cores, at least 1, at most
640/// `auto_cap`. A non-zero value is used as given. The indexer passes a cap of 8
641/// (write-side cache contention); query-time verification passes a higher cap.
642pub fn resolve_thread_count(configured: usize, auto_cap: usize) -> usize {
643    if configured != 0 {
644        return configured.max(1);
645    }
646    let available = std::thread::available_parallelism()
647        .map(|n| n.get())
648        .unwrap_or(4);
649    ((available as f64 * 0.8).ceil() as usize).clamp(1, auto_cap.max(1))
650}
651
652/// How a query found its candidate lines.
653///
654/// `trigram` is the normal case: the pattern's literals were looked up in the
655/// inverted index and only the lines they name were verified. `scan` means every
656/// line of every file was verified, which happens for a pattern shorter than
657/// 3 chars, a regex with no literal of 3+ chars (`\w+_id`), a non-ASCII literal
658/// under `(?i)`, or a keyword symbol query. Before 2.0.0 every `(?i)` regex
659/// scanned; since 2.0.0 its literals are looked up under all case variants.
660#[derive(Debug, Clone, Copy, Default, Serialize, Deserialize, PartialEq, Eq)]
661#[serde(rename_all = "lowercase")]
662pub enum IndexPath {
663    /// Candidates came from the trigram index.
664    #[default]
665    Trigram,
666    /// Every line was verified.
667    Scan,
668}
669
670/// Per-phase wall-clock timings for one query, in microseconds.
671///
672/// Present in a [`QueryResponse`] only when the caller asked for it
673/// (`rfx query --timing`, or `REFLEX_MCP_TIMING=1` for the MCP server).
674#[derive(Debug, Clone, Default, Serialize, Deserialize, PartialEq, Eq)]
675pub struct QueryTimings {
676    /// Whether the trigram index or a full scan produced the candidates.
677    #[serde(default)]
678    pub index_path: IndexPath,
679    /// Opening (or reusing) the index handle.
680    pub open_us: u64,
681    /// Trigram lookup and posting-list intersection.
682    pub candidates_us: u64,
683    /// Verifying candidate lines against the pattern (and any enrichment).
684    pub verify_us: u64,
685    /// Time the query waited for the freshness check after the search finished.
686    /// The check runs on its own thread alongside the search, so this is usually
687    /// near zero; `status_compute_us` is what the check itself cost.
688    pub status_us: u64,
689    /// The freshness check's own duration (git spawns, or a tree walk outside git).
690    #[serde(default)]
691    pub status_compute_us: u64,
692    /// Grouping by file, context lines, dependencies.
693    pub group_us: u64,
694    /// Whole query as seen by the engine.
695    pub total_us: u64,
696    /// Automatic updates of a stale index before the answer (0: none ran).
697    /// `total_us` includes them.
698    #[serde(default, skip_serializing_if = "is_zero_u64")]
699    pub update_us: u64,
700}
701
702/// Statistics about the index
703#[derive(Debug, Clone, Serialize, Deserialize, Default)]
704pub struct IndexStats {
705    /// Total files indexed
706    pub total_files: usize,
707    /// Index size on disk (bytes)
708    pub index_size_bytes: u64,
709    /// Last update timestamp
710    pub last_updated: String,
711    /// File count breakdown by language
712    pub files_by_language: std::collections::HashMap<String, usize>,
713    /// Line count breakdown by language
714    pub lines_by_language: std::collections::HashMap<String, usize>,
715    /// New files added since last index run (0 if not an incremental run)
716    #[serde(default, skip_serializing_if = "is_zero")]
717    pub new_files: usize,
718    /// Modified files re-indexed since last run (0 if not an incremental run)
719    #[serde(default, skip_serializing_if = "is_zero")]
720    pub modified_files: usize,
721    /// Unchanged files (same hash as last run, still re-indexed due to other changes)
722    #[serde(default, skip_serializing_if = "is_zero")]
723    pub unchanged_files: usize,
724    /// Files dropped from the index because they no longer exist on disk
725    #[serde(default, skip_serializing_if = "is_zero")]
726    pub deleted_files: usize,
727    /// Files skipped because they exceeded max_file_size
728    #[serde(default, skip_serializing_if = "is_zero")]
729    pub skipped_too_large: usize,
730    /// Total bytes of files skipped due to max_file_size
731    #[serde(default, skip_serializing_if = "is_zero_u64")]
732    pub skipped_bytes_too_large: u64,
733    /// Files skipped because a NUL byte says they are binary (ripgrep's rule)
734    #[serde(default, skip_serializing_if = "is_zero")]
735    pub skipped_binary: usize,
736    /// Raw bytes of indexed source held in content.bin (0 if unknown)
737    #[serde(default, skip_serializing_if = "is_zero_u64")]
738    pub corpus_bytes: u64,
739    /// Size of trigrams.bin on disk (0 if absent)
740    #[serde(default, skip_serializing_if = "is_zero_u64")]
741    pub trigram_index_bytes: u64,
742}
743
744/// Information about an indexed file
745#[derive(Debug, Clone, Serialize, Deserialize)]
746pub struct IndexedFile {
747    /// File path
748    pub path: String,
749    /// Detected language
750    pub language: String,
751    /// Last indexed timestamp
752    pub last_indexed: String,
753}
754
755/// Index status for query responses
756#[derive(Debug, Clone, Serialize, Deserialize, PartialEq, Eq)]
757#[serde(rename_all = "snake_case")]
758pub enum IndexStatus {
759    /// Index is fresh and up-to-date
760    Fresh,
761    /// Index is stale (any issue: branch not indexed, commit changed, files modified)
762    Stale,
763}
764
765/// Warning details when index is stale
766#[derive(Debug, Clone, Serialize, Deserialize)]
767pub struct IndexWarning {
768    /// Human-readable reason why index is stale
769    pub reason: String,
770    /// Command to run to fix the issue
771    pub action_required: String,
772    /// Tracked files edited since the index was built.
773    ///
774    /// BREAKING in 1.7.2: this was a `u32` count. It is now the paths themselves,
775    /// because a count told an agent something was wrong without telling it what, so
776    /// the only safe reaction was to distrust the whole result. The list is capped;
777    /// `truncated` says when.
778    #[serde(skip_serializing_if = "Option::is_none")]
779    pub files_modified: Option<Vec<String>>,
780    /// Files present on disk but absent from the index (new or untracked).
781    #[serde(skip_serializing_if = "Option::is_none")]
782    pub files_added: Option<Vec<String>>,
783    /// Files in the index but no longer on disk. These produce ghost hits.
784    #[serde(skip_serializing_if = "Option::is_none")]
785    pub files_deleted: Option<Vec<String>>,
786    /// Total changed paths, which may exceed the lengths of the lists above.
787    #[serde(skip_serializing_if = "Option::is_none")]
788    pub changed_count: Option<usize>,
789    /// Whether the lists were cut short. Set on a fresh checkout or a huge rebase.
790    #[serde(default, skip_serializing_if = "std::ops::Not::not")]
791    pub truncated: bool,
792    /// Additional context (git branch info, etc.)
793    #[serde(skip_serializing_if = "Option::is_none")]
794    pub details: Option<IndexWarningDetails>,
795}
796
797impl IndexWarning {
798    /// A warning with only a reason and an action, no file lists.
799    pub fn new(reason: impl Into<String>, action_required: impl Into<String>) -> Self {
800        Self {
801            reason: reason.into(),
802            action_required: action_required.into(),
803            files_modified: None,
804            files_added: None,
805            files_deleted: None,
806            changed_count: None,
807            truncated: false,
808            details: None,
809        }
810    }
811
812    /// Attach git branch/commit context.
813    pub fn with_details(mut self, details: IndexWarningDetails) -> Self {
814        self.details = Some(details);
815        self
816    }
817}
818
819/// Detailed information about index staleness
820#[derive(Debug, Clone, Serialize, Deserialize)]
821pub struct IndexWarningDetails {
822    /// Current branch (if in git repo)
823    #[serde(skip_serializing_if = "Option::is_none")]
824    pub current_branch: Option<String>,
825    /// Indexed branch (if in git repo)
826    #[serde(skip_serializing_if = "Option::is_none")]
827    pub indexed_branch: Option<String>,
828    /// Current commit SHA (if in git repo)
829    #[serde(skip_serializing_if = "Option::is_none")]
830    pub current_commit: Option<String>,
831    /// Indexed commit SHA (if in git repo)
832    #[serde(skip_serializing_if = "Option::is_none")]
833    pub indexed_commit: Option<String>,
834    /// When the index was last written (Unix seconds).
835    #[serde(default, skip_serializing_if = "Option::is_none")]
836    pub indexed_at: Option<i64>,
837    /// How the working tree was compared to the index: `"git"` (candidates from
838    /// `git status`, confirmed by fingerprint) or `"walk"` (every file stat'ed).
839    #[serde(default, skip_serializing_if = "Option::is_none")]
840    pub checked_by: Option<String>,
841}
842
843/// The full answer to `check_index_status`.
844///
845/// `details` is present even when the index is fresh, so a human can see that the
846/// indexed commit differs from HEAD without that difference being called staleness:
847/// since 2.0.0 freshness is judged by file content, not by commit.
848#[derive(Debug, Clone)]
849pub struct IndexStatusReport {
850    pub status: IndexStatus,
851    pub can_trust_results: bool,
852    pub warning: Option<IndexWarning>,
853    pub details: Option<IndexWarningDetails>,
854}
855
856/// Pagination information for query results
857#[derive(Debug, Clone, Serialize, Deserialize)]
858pub struct PaginationInfo {
859    /// Total number of results before offset/limit — **`null` whenever
860    /// `total_is_exact` is false**. A list-mode search with a `limit` stops
861    /// verifying once the page is full, and the number verified by then is not a
862    /// total; reporting it as one made agents stop paginating early (2.0.0 field
863    /// test: `total: 851` for a term with 18,752 matches). Use `approx_total` for an
864    /// estimate, or a count-mode / no-limit search for the exact number.
865    #[serde(default)]
866    pub total: Option<usize>,
867    /// Number of results in this response (after offset/limit)
868    pub count: usize,
869    /// Offset used (starting position)
870    pub offset: usize,
871    /// Limit used (max results per page)
872    #[serde(skip_serializing_if = "Option::is_none")]
873    pub limit: Option<usize>,
874    /// Whether there are more results after this page
875    pub has_more: bool,
876    /// `true` when `total` counts every match. `false` when a list-mode search
877    /// stopped verifying once the page was full: `total` is then `null` and
878    /// `approx_total` carries an estimate. Count mode, no-limit searches,
879    /// symbol/AST searches and `require_exact_total` callers are always exact.
880    #[serde(default = "default_true")]
881    pub total_is_exact: bool,
882    /// Estimated total when `total_is_exact` is false. After the page filled, a
883    /// spread sample of the remaining candidate lines was verified and the
884    /// measured hit rate scaled over the rest (see
885    /// `query::ESTIMATE_SAMPLE_FILES`). Typically within ±30% on the synthetic
886    /// corpus; wider when hits cluster in a few files. Absent when no estimate was
887    /// possible (a regex with no literal, where every line is a candidate).
888    #[serde(default, skip_serializing_if = "Option::is_none")]
889    pub approx_total: Option<usize>,
890}
891
892impl PaginationInfo {
893    /// The total, only when it is a real one.
894    pub fn exact_total(&self) -> Option<usize> {
895        if self.total_is_exact {
896            self.total
897        } else {
898            None
899        }
900    }
901
902    /// The best number available for a threshold or a log line: the exact total,
903    /// else the estimate, else the end of this page. Never show it as a total.
904    pub fn best_total(&self) -> usize {
905        self.exact_total()
906            .or(self.approx_total)
907            .unwrap_or(self.offset + self.count)
908    }
909}
910
911/// Query response with results and index status
912#[derive(Debug, Clone, Serialize, Deserialize)]
913pub struct QueryResponse {
914    /// AI-optimized instruction for how to handle these results
915    /// Only present when --ai flag is used or in MCP mode
916    /// Provides guidance to AI agents on response format and next actions
917    #[serde(skip_serializing_if = "Option::is_none")]
918    pub ai_instruction: Option<String>,
919    /// Status of the index (fresh or stale)
920    pub status: IndexStatus,
921    /// Whether the results can be trusted
922    pub can_trust_results: bool,
923    /// Warning information (only present if stale)
924    #[serde(skip_serializing_if = "Option::is_none")]
925    pub warning: Option<IndexWarning>,
926    /// Pagination information
927    pub pagination: PaginationInfo,
928    /// File-grouped search results
929    /// Results are always grouped by file path, with dependencies populated when --dependencies flag is used
930    pub results: Vec<FileGroupedResult>,
931    /// For a whole-identifier search that found nothing: how many candidate lines
932    /// contain the pattern as a substring. Lets the caller explain a zero without
933    /// running a second search. Absent whenever there were results, or when the
934    /// search was not a whole-identifier one.
935    #[serde(default, skip_serializing_if = "Option::is_none")]
936    pub substring_hint_count: Option<usize>,
937    /// Things the engine did to the query that the caller should know about.
938    ///
939    /// Today this is one message at most: a whole-identifier pattern containing
940    /// brackets was escaped and run as a regex (see
941    /// `query::prepare_literal_pattern`). Every surface — CLI, MCP, HTTP — carries it,
942    /// so a rewrite is never silent on any of them.
943    #[serde(default, skip_serializing_if = "Vec::is_empty")]
944    pub warnings: Vec<String>,
945    /// For a whole-identifier search that found nothing while substring matches
946    /// exist: a ready-to-show sentence naming the count and the switch that shows
947    /// them. Absent whenever there were results, or when the search was not a
948    /// whole-identifier one.
949    #[serde(default, skip_serializing_if = "Option::is_none")]
950    pub hint: Option<String>,
951    /// Machine-readable cause behind `hint`: `hidden` (the filter names a
952    /// dot-directory), `not_indexed` (the `file` filter names a path the index does
953    /// not hold), `lock_or_generated`, `whole_identifier`. Absent when no rule
954    /// applies, so a harness can branch without parsing prose.
955    #[serde(default, skip_serializing_if = "Option::is_none")]
956    pub excluded_reason: Option<crate::query::ExcludedReason>,
957    /// For a search that found nothing: how many candidate files were lock or
958    /// generated files, which every search leaves out unless `include_locks` /
959    /// `include_generated` (or `lang`) asks for them. The `hint` says so too.
960    #[serde(default, skip_serializing_if = "Option::is_none")]
961    pub excluded_by_default: Option<usize>,
962    /// Count-only searches: the number of files with at least one match. `results`
963    /// is empty in that mode, so this is the only place the file count lives.
964    #[serde(default, skip_serializing_if = "Option::is_none")]
965    pub file_count: Option<usize>,
966    /// Per-phase timings, only when requested.
967    #[serde(default, skip_serializing_if = "Option::is_none")]
968    pub timings: Option<QueryTimings>,
969}
970
971/// Report from cache compaction operation
972#[derive(Debug, Clone, Serialize, Deserialize)]
973pub struct CompactionReport {
974    /// Always 0 since auto-update: index runs remove deleted files (kept for JSON compatibility)
975    pub files_removed: usize,
976    /// Space saved in bytes
977    pub space_saved_bytes: u64,
978    /// Duration in milliseconds
979    pub duration_ms: u64,
980}
981
982#[cfg(test)]
983mod tests {
984    use super::*;
985
986    #[test]
987    fn test_symbol_ref_json_shape() {
988        let sym = SymbolRef {
989            name: "my_function".to_string(),
990            kind: SymbolKind::Function,
991            span: Span {
992                start_line: 10,
993                end_line: 20,
994            },
995        };
996        let json = serde_json::to_value(&sym).unwrap();
997        assert_eq!(json["name"], "my_function");
998        assert_eq!(json["kind"], "Function");
999        assert_eq!(json["span"]["start_line"], 10);
1000        assert_eq!(json["span"]["end_line"], 20);
1001        assert!(json.as_array().is_none());
1002    }
1003
1004    #[test]
1005    fn test_symbol_ref_roundtrip() {
1006        let original = SymbolRef {
1007            name: "MyStruct".to_string(),
1008            kind: SymbolKind::Struct,
1009            span: Span {
1010                start_line: 1,
1011                end_line: 5,
1012            },
1013        };
1014        let json = serde_json::to_string(&original).unwrap();
1015        let decoded: SymbolRef = serde_json::from_str(&json).unwrap();
1016        assert_eq!(original, decoded);
1017    }
1018
1019    #[test]
1020    fn test_symbol_ref_exact_json() {
1021        let sym = SymbolRef {
1022            name: "Foo".to_string(),
1023            kind: SymbolKind::Class,
1024            span: Span {
1025                start_line: 3,
1026                end_line: 7,
1027            },
1028        };
1029        let json = serde_json::to_string(&sym).unwrap();
1030        assert_eq!(
1031            json,
1032            r#"{"name":"Foo","kind":"Class","span":{"start_line":3,"end_line":7}}"#
1033        );
1034    }
1035}