Published on Infostart: Code Index — структурный поиск по выгрузке кода 1С через MCP

code-index-mcp

Русская версия

Rust-native code index for AI agents. Static binary. Production-grade BSL/1C support.

One static binary for Windows/Linux/macOS — no runtime, no dependencies. Indexes large repositories in seconds, returns results to AI agents over MCP in milliseconds. 31 tools: 20 universal + 11 BSL-specific for 1C:Enterprise configurations.

What's inside

• Performance. 62,000 files indexed in 43 seconds, sub-ms search per query. Production-grade for 100K+ file monorepos.
• 31 MCP tools. 20 universal (functions, classes, callers/callees, file content, grep) + 11 BSL-tools (object structure & profile, form handlers, event subscriptions, call graph, data links, register writers, impact map, read-only SQL).
• Native BSL/1C. Parses both Configurator (XML) and 1C:EDT (.mdo) exports of 1C:Enterprise 8.3 configurations. Data-link graph (object→object edges via reference types in attributes) — ~60,000 edges in seconds for a typical accounting configuration.
• Federation. One MCP server can serve multiple repositories across machines — pass repo: "alias" in each tool call.
• Compressed content storage. File contents stored in SQLite via zstd, cheap random-access reads for AI agents.
• Tree-sitter AST. 10 languages with full parsing (Rust, Python, JavaScript, TypeScript, Java, Kotlin, C#, Go, Objective-C, Zig) + fallback for 50+ formats.

Connects to Claude Code, Cursor, any MCP client over HTTP.

Problem

AI models waste enormous time on repeated grep/find calls just to locate a single symbol. A real example: finding RuntimeErrorProcessing in a Java project required 14 sequential grep/find calls, each scanning thousands of files. With Code Index, that is one query returning results in under a millisecond.

Solution

A compiled Rust binary with one-writer / many-readers architecture:

1. Parses source code into AST via tree-sitter
2. Indexes everything into SQLite with FTS5 full-text search
3. A separate background daemon is the sole writer: one process per machine watches a list of folders from its config and keeps .code-index/index.db up to date.
4. The MCP server is a thin read-only client: any number of Claude Code / VS Code / subagent sessions can connect to the same project in parallel — no pidlock conflicts, no per-session re-indexing.

Supported Languages

Text files (.md, .json, .yaml, .toml, .xml, .sql, .env, etc.) are also indexed for full-text search.

HTML — entity mapping (v0.7.2)

HTML has no native concept of "function" or "class", so the mapping is conventional. Dual-indexing: html files go through both AST parser AND text_files (so search_text / grep_text / read_file keep working alongside the new structural queries).

All MCP tools that work for HTML files after re-indexing:

# === Discovery & metadata ===
list_files(repo="X", pattern="**/*.html")                # all html (returns language="html")
list_files(repo="X", path_prefix="src/templates/")
stat_file(repo="X", path="src/templates/base.html")      # returns language="html", category="text"
get_stats(repo="X")                                       # totals

# === Structural (AST) — new in 0.7.x ===
# Elements with id, forms, css-classes, links, inline blocks → AST tables
get_class(repo="X", name="cart")                          # outerHTML of <... id="cart">
get_class(repo="X", name="form_login")                    # full <form id="login">
search_class(repo="X", query="container", language="html")
get_function(repo="X", name="inline_script_42")           # body of <script> at line 42
search_function(repo="X", query="inline_script", language="html")
find_symbol(repo="X", name="form_login")                  # exact-name lookup across all 4 tables
find_symbol(repo="X", name="class:htmx-indicator")        # CSS class usage
get_imports(repo="X", module="https://unpkg.com/htmx.org@1.9.12")  # who depends on this CDN
get_file_summary(repo="X", path="src/templates/base.html")         # full map (functions/classes/imports/variables)

# === Body-level grep (works on inline_script bodies) ===
grep_body(repo="X", regex="fetch\\(", language="html")    # in <script> blocks
grep_body(repo="X", pattern="color:", language="html")    # in <style> blocks
grep_body(repo="X", regex="hx-target", language="html", path_glob="src/templates/**", context_lines=2)

# === Text-level (still works via dual-indexing) ===
read_file(repo="X", path="src/templates/base.html", line_start=1, line_end=20)
search_text(repo="X", query="DOCTYPE", language="html")
grep_text(repo="X", regex="\\{%\\s*include", path_glob="**/*.html", context_lines=1)  # Jinja includes

get_callers / get_callees are not populated for HTML (the parser does not extract call edges between scripts).

Template engines (Jinja/Django/EJS): {{ … }} and {% … %} are tolerated as text content; surrounding HTML elements are still parsed normally.

Quick Start

Install via npm (easiest)

npm install -g @regsorm/code-index-mcp

The postinstall step downloads the prebuilt native binary for your platform (Windows x64, Linux x64, macOS arm64) from GitHub Releases — nothing is compiled. Then run it as an MCP server:

npx @regsorm/code-index-mcp serve --path /path/to/your/repo

Also published to the official MCP Registry as io.github.Regsorm/code-index. This wrapper ships only the public code-index binary (no 1C support); for bsl-indexer build from source.

Build from source

git clone https://github.com/Regsorm/code-index-mcp.git
cd code-index-mcp
cargo build --release -p code-index               # public binary for Python/Rust/Go/Java/JS/TS
cargo build --release -p bsl-indexer --features enrichment   # extra build with 1C support + LLM enrichment

Binaries:

• target/release/code-index[.exe] — main binary (no 1C support).
• target/release/bsl-indexer[.exe] — full 1C support (XML metadata parsers, BSL call graph, data-links graph, MCP tools get_object_structure / get_form_handlers / find_path_bsl / search_terms / get_data_links / find_data_path / get_register_writers, optional LLM enrichment under cargo feature enrichment).

GitHub Releases publish 6 ready artifacts per tag: code-index × {Win, Linux, macOS} + bsl-indexer × {Win, Linux, macOS}.

Set up the background daemon (v0.5+)

Portable layout: one folder for everything (binary + config + runtime files). Pointed to by CODE_INDEX_HOME env var.

1. Create the daemon folder and drop code-index.exe into it (e.g. C:\tools\code-index\).

2. Set the CODE_INDEX_HOME environment variable to point at that folder:

   Windows (persistent, user scope):

   powershell

   setx CODE_INDEX_HOME "C:\tools\code-index"
   # Reopen your shell so the variable is visible.
   

   Linux — add to ~/.bashrc or ~/.zshrc:

   bash

   export CODE_INDEX_HOME="$HOME/.local/code-index"
   

   macOS — same as Linux for shells; for launchd agents use launchctl setenv.

   Any OS — per-project fallback via .mcp.json (no system env var needed):

   json

   {
     "mcpServers": {
       "code-index": {
         "command": "C:\\tools\\code-index\\code-index.exe",
         "args": ["serve", "--path", "."],
         "env": { "CODE_INDEX_HOME": "C:\\tools\\code-index" }
       }
     }
   }
   

3. Create daemon.toml inside that folder and list the paths to watch:

   toml

   [daemon]
   http_port = 0                  # 0 = pick free port automatically
   max_concurrent_initial = 1     # folders processed sequentially during initial indexing
   
   [[paths]]
   path = "C:\\RepoUT"
   
   [[paths]]
   path = "C:\\RepoBP_1"
   debounce_ms = 500              # per-folder override: react faster than the default 1500 ms
   batch_ms    = 1000
   

   Per-folder debounce_ms / batch_ms are optional. If omitted, the daemon falls back to .code-index/config.json inside that project, and then to built-in defaults (1500 ms / 2000 ms).

4. Start the daemon (foreground):

   bash

   code-index daemon run
   

   Or install it as a Windows Scheduled Task (auto-start at user logon; the script also sets CODE_INDEX_HOME via setx):

   powershell

   powershell -ExecutionPolicy Bypass -File scripts\install-daemon-autostart.ps1 `
     -BinaryPath "C:\tools\code-index\code-index.exe" `
     -CodeIndexHome "C:\tools\code-index" `
     -StartNow
   

5. Check status:

   bash

   code-index daemon status        # human-readable
   code-index daemon status --json # JSON
   code-index daemon reload        # re-read daemon.toml after edits
   code-index daemon stop
   

CODE_INDEX_HOME is required — there is no fallback. If it is unset, both daemon and serve exit with an error explaining how to set it.

Troubleshooting — "daemon not running / runtime-info missing" even though the daemon IS running.

The serve process and the daemon find each other only through $CODE_INDEX_HOME/daemon.json. If serve sees a different (or empty) CODE_INDEX_HOME than the daemon, it looks for daemon.json in the wrong place and reports the daemon as offline — while it is actually alive.

The most common cause on Linux/macOS: GUI MCP clients (VS Code, Continue, Cline) do not read ~/.bashrc / ~/.zshrc, so a serve they launch with an empty env never sees the CODE_INDEX_HOME you exported in your shell. Meanwhile the daemon, started from a terminal, does — so they end up pointing at different folders.

Fix: set CODE_INDEX_HOME explicitly in the env section of the client's MCP config, using the same absolute path the daemon uses ($HOME is not expanded there — use a real path). Restart the client and verify with code-index daemon status.

One-shot indexing (no daemon)

code-index index /path/to/project
code-index stats --path /path/to/project --json

Run as MCP server (read-only)

code-index serve --path /path/to/project

This is a thin read-only client of the daemon. It does not index anything itself — the daemon does. If the folder is still being indexed or not in daemon.toml, tools return a structured {status, message, progress} response instead of failing.

Transports (stdio vs HTTP)

serve supports two transports:

# stdio — per-session, alias set at CLI
code-index serve --path ut=/repos/ut --path bp=/repos/bp

# HTTP — shared process, aliases come from daemon.toml
code-index serve --transport http --port 8011 --config /etc/code-index/daemon.toml

--path can be repeated in alias=dir form (multi-repo mode). Each tool call takes a repo parameter to select which repository to query. Without =, the single path uses alias=default (backward-compatible).

In HTTP mode, if --config is provided, aliases are taken from [[paths]] entries of daemon.toml: explicit alias = "...", or derived from the path's last segment (lowercased, spaces → _) when not set. CLI --path takes precedence over the config file.

Connecting to Claude Code

Add to .mcp.json in your project root. For stdio:

{
  "mcpServers": {
    "code-index": {
      "command": "npx",
      "args": ["-y", "@regsorm/code-index-mcp", "serve", "--path", "."]
    }
  }
}

For a shared HTTP process:

{
  "mcpServers": {
    "code-index": {
      "type": "http",
      "url": "http://127.0.0.1:8011/mcp"
    }
  }
}

MCP Tools

All search tools (search_function, search_class, get_function, get_class, find_symbol, search_text, grep_body) accept an optional path_glob parameter (v0.7.0) to scope results to a subtree (e.g., src/auth/**, Documents/**/*.bsl). Implementation: post-filter via the globset crate after the SQL query. Since v0.32.0, path_glob/pattern support {a,b} brace alternates (**/*.{bsl,xml}) — including in grep_code/grep_text/grep_body/list_files, where the filter runs at the SQLite GLOB level (the pattern is expanded into an OR group of conditions).

Code-file content storage (v0.8.0)

Starting with v0.8.0, code-file content is stored in the file_contents table (zstd-compressed) and returned by read_file and searched by grep_code. Large files can be excluded from storage via max_code_file_size_bytes (default 5 MB):

[indexer]
max_code_file_size_bytes = 5242880   # 5 MB global override

[[paths]]
path = "C:/RepoUT"
max_code_file_size_bytes = 10485760  # 10 MB for this repo only

Priority: per-path → [indexer] section → 5 MB default. Files exceeding the limit are stored with oversize=1 and content_blob=NULL; AST parsing, FTS, and call-graph edges still work for them in full. read_file and grep_code return a hint explaining how to query such files via get_function/get_class/grep_body.

Connection pool (v0.24.0)

In serve, each repo is read through a pool of read-only SQLite connections instead of a single connection behind a mutex. Concurrent requests to the same repo now run in parallel — a heavy query (bsl_sql, a full grep_code, recursive find_path/get_call_tree) no longer blocks a fast get_function on that repo. Tune via an optional [pool] section in serve.toml:

[pool]
pool_size = 4               # connections per repo (default 4)
per_conn_cache_kib = 16384  # page-cache per connection, KiB (default 16384 = 16 MB)
busy_timeout_ms = 5000      # SQLite busy_timeout per connection, ms (default 5000)

All fields are optional; omitting the section uses the defaults. The default is memory-neutral: 4 × 16 MB = 64 MB per active repo, the same as the previous single connection. Connections open lazily up to pool_size and return to the pool when a request finishes; 0 is clamped to a safe value. WAL mode (already used by the index) makes the multiple readers safe alongside the indexing daemon's writes.

Additional tools for 1C repos (only in bsl-indexer, v0.6+)

When BSL repos are present in daemon.toml (language = "bsl"), 11 BSL-specific tools are auto-registered:

These tools appear in tools/list only when at least one BSL repo is configured (conditional registration). When the repo set changes in daemon.toml, the server emits notifications/tools/list_changed. On Claude Code 2.1.120 this notification is currently ignored; workaround — manual /mcp Reconnect.

Since v0.8.1, these BSL-tools work in all scenarios (was broken in v0.8.0, see CHANGELOG):

• via bsl-indexer.exe daemon run — daemon now applies schema_extensions and index_extras for each BSL repo on startup (creates metadata_objects / metadata_forms / event_subscriptions / proc_call_graph and fills them from Configuration.xml).
• via federation — extension-tools are forwarded over the universal POST /federate/extension route. Both peers must be on ≥ 0.8.1; older peers will return 404 on the new route.
• on repos without Configuration.xml (e.g. partial dumps containing only forms/processors) — the tables are created empty and the tools return [] instead of throwing no such table: metadata_objects.

Full instructions: docs/bsl-indexer.md.

All tools support a language filter: search_function(query="X", language="python")

grep_body

Unlike FTS search, grep_body supports literal substrings (including dots and special characters) and regular expressions. This is essential for finding references like Catalog.Contractors or Справочники.Контрагенты that break FTS5 syntax.

grep_body(pattern="Справочники.Контрагенты", language="bsl")
grep_body(regex="Catalog\\.(Contractors|Organizations)", language="bsl")

Returns [{file_path, name, kind, line_start, line_end, match_lines, match_count}] — concrete functions/classes containing the match.

Each result includes match_lines — up to 3 absolute line numbers in the file where the pattern was found. If there are more than 3 matches, match_count shows the total.

[
  {
    "file_path": "src/Catalogs/Products/ObjectModule.bsl",
    "name": "OnWrite",
    "kind": "function",
    "line_start": 45,
    "line_end": 82,
    "match_lines": [51, 63, 78]
  }
]

CLI Reference

# Background daemon (writer — one per machine)
code-index daemon run                          # foreground, for Scheduled Task / systemd
code-index daemon status [--json]              # query GET /health via loopback
code-index daemon reload                       # re-read daemon.toml
code-index daemon stop                         # POST /stop

# MCP server (read-only client; used by Claude Code, VS Code, subagents)
code-index serve --path /project

# One-shot indexing (no daemon)
code-index index /project [--force]

# Project management
code-index init --path /project          # Create config
code-index clean --path /project         # Remove stale entries
code-index stats --path /project [--json]

# Symbol search
code-index query "name" --path /project [--language rust] [--json]

# Full-text search (JSON output)
code-index search-function "query" --path /project [--language python] [--limit 20]
code-index search-class "query" --path /project [--language python] [--limit 20]
code-index search-text "query" --path /project [--limit 20]

# Exact lookup (JSON output)
code-index get-function "exact_name" --path /project
code-index get-class "exact_name" --path /project

# Call graph (JSON output)
code-index get-callers "function_name" --path /project [--language python]
code-index get-callees "function_name" --path /project [--language python]

# Navigation (JSON output)
code-index get-imports --path /project [--module "name"] [--file-id 42]
code-index get-file-summary "src/main.rs" --path /project

# Substring / regex search in function and class bodies (supports dots and special chars)
code-index grep-body --pattern "Catalog.Contractors" --path /project [--language bsl] [--limit 100]
code-index grep-body --regex "Catalog\.(Contractors|Organizations)" --path /project

Using CLI from Subagents

Subagents launched via the Agent tool in Claude Code do not have access to MCP servers — they run in isolated subprocesses with no connection to the parent MCP session. All 12 MCP tools are mirrored as CLI subcommands that output JSON, making code-index fully usable from any subprocess, script, or subagent.

# Instead of an MCP tool call, a subagent runs:
code-index search-function "authenticate" --path /my/project --language python

# Call graph from CLI:
code-index get-callers "process_order" --path /my/project

# File map:
code-index get-file-summary "src/auth/login.py" --path /my/project

Every command outputs valid JSON that the subagent can parse and reason over, identical in structure to what the MCP tools return.

Note: CLI read commands use SQLITE_OPEN_READ_ONLY mode, so they work in parallel with the MCP daemon without database locking conflicts.

CLAUDE.md Setup

Add this block to your project's CLAUDE.md to instruct Claude Code subagents to use the CLI indexer instead of grep, find, or reading files manually:

## Code Index — fast code search

For code search, use the CLI indexer instead of grep/find/Read:
- Search: code-index query "name" --path /path/to/project --json
- FTS search: code-index search-function "query" --path /path/to/project
- Call graph: code-index get-callers "function" --path /path/to/project
- File map: code-index get-file-summary "file" --path /path/to/project
- Stats: code-index stats --path /path/to/project --json
All commands output JSON. This is instant search over an indexed database.

Use an absolute path to the binary and adjust /path/to/project to your setup. On Windows, specify the full path to code-index.exe, for example C:\MCP-Servers\code-index\target\release\code-index.exe.

Daemon Mode (v0.5+)

Starting with v0.5, code-index uses a one-writer / many-readers architecture:

Background daemon (single writer)

code-index daemon run starts a long-running process that:

1. Loads the list of watched folders from daemon.toml.
2. For each folder: opens .code-index/index.db, runs full reindex with mtime fast-path (v0.4.0), then switches to a notify watcher that re-indexes on change (1.5s debounce, 2s batch).
3. Exposes a local health / management HTTP endpoint on loopback (port written to daemon.json in the state directory).
4. Holds a global PID-lock (daemon.pid) to prevent two daemons per machine.

Per-folder lifecycle: not_started → initial_indexing → ready ⇄ reindexing_batch / error. Each status transition is visible via daemon status.

MCP servers (many read-only readers)

code-index serve --path <project> opens .code-index/index.db in SQLITE_OPEN_READ_ONLY and exposes MCP tools over stdio. Multiple MCP instances on the same project run in parallel without blocking each other.

Before every tool call the MCP asks the daemon for the per-folder status. If it is not ready, the tool returns a structured JSON:

{ "status": "indexing", "progress": {"files_done": 4200, "files_total": 10000, "percent": 42.0}, "message": "Первичная индексация в процессе" }

If the daemon is offline:

{ "status": "daemon_offline", "message": "Демон code-index не доступен. Запустите 'code-index daemon run' или Scheduled Task." }

Configuration

.code-index/config.json is created automatically on first run. Full reference:

{
  "exclude_dirs": ["node_modules", ".venv", "__pycache__", ".git", "target", "output"],
  "extra_text_extensions": [],
  "max_file_size": 1048576,
  "max_files": 0,
  "bulk_threshold": 10,
  "languages": ["python", "javascript", "typescript", "java", "rust", "go", "bsl"],
  "batch_size": 500,
  "storage_mode": "auto",
  "memory_max_percent": 25,
  "debounce_ms": 1500,
  "batch_ms": 2000
}

Key fields:

• storage_mode — auto selects in-memory or disk SQLite based on available RAM; memory forces in-memory; disk forces on-disk
• memory_max_percent — maximum percentage of system RAM the in-memory database may use before falling back to disk (used in auto mode)
• debounce_ms — milliseconds to wait after a file change before triggering re-indexing (collects burst edits into one pass)
• batch_ms — upper bound on how long the watcher keeps accumulating events after the first one in a batch
• batch_size — number of records per SQLite transaction during indexing (higher = faster bulk inserts, higher peak memory)
• bulk_threshold — minimum number of files that triggers bulk mode (drop indexes, insert, rebuild indexes); faster for large batches

Tuning watcher latency (debounce_ms, batch_ms)

Defaults are 1500 ms / 2000 ms — good for typical IDE save + formatter + linter bursts and for git operations that touch many files at once. For a lively single-user IDE session you can lower the debounce and trade throughput for responsiveness.

The daemon resolves these values in this order (first match wins):

1. Per-folder override in daemon.toml:
   toml

   [[paths]]
   path = "C:/RepoBP_1"
   debounce_ms = 500      # react in ~0.6 s instead of ~1.6 s
   batch_ms    = 1000
   

2. Per-project .code-index/config.json — applies to that project only.
3. Built-in defaults (1500 / 2000).

Re-read after editing daemon.toml:

code-index daemon reload

Recommended values:

Guarding output against disk offload ([cap], v0.38.0)

The client (claude CLI / Claude Code) caps a single tool_result streamed inline into context (MAX_MCP_OUTPUT_TOKENS ≈ 25,000 tokens). A response over the cap is dumped to a file on disk by the harness, handing the model only a path + preview — structured inline access is lost. To keep large outputs (a big module's map, long arrays of values/sources/attributes) from being offloaded, serve trims them at the source. Configured via an optional [cap] section in daemon.toml:

[cap]
max_response_bytes      = 48000   # response budget in JSON bytes; 0 disables cap_response
cap_enabled             = true    # global on/off for cap_response (takes precedence over cap_tools)
cap_tools               = ["get_event_subscriptions", "bsl_sql", "find_references", "get_register_writers"]
max_function_body_chars = 15000   # get_function/get_class body threshold; 0 keeps the full body

Mechanisms (all optional, act on the serve output layer, require no reindex):

• cap_response — while the response JSON exceeds max_response_bytes, halves the heaviest array, leaving <key>_total (original count) and <key>_truncated: true. Applied to tools in cap_tools (when cap_enabled = true). Only arrays are trimmed — large strings (read_file/grep) are untouched. get_file_summary (core) is wired in here too: a giant module's map (hundreds of functions) no longer goes to offload.
• omit_oversize_sections (for get_object_structure) — where an array/map is the full authoritative answer (a 1C object's structure), the heaviest section is dropped WHOLESALE with <section>_omitted: true + <section>_count: N (a partial sample would lie, “here are all the enum values”).
• Navigational body cap (get_function/get_class) — a body longer than max_function_body_chars is returned as a head+tail+marker stub with a hint to read_file(line_start,line_end) / grep_body.

Benchmarks

Tested on 1C:Enterprise configurations (HDD, Windows):

Re-check uses mtime + file_size fast-path: only stat() per file, zero reads, zero SHA-256 hashes.

Comparison with grep:

Architecture

Source Files -> Tree-sitter Parser -> SQLite (in-memory) -> MCP Server -> AI Model
                                           ^
                      File Watcher --------+ (auto re-index)

Key optimizations:

• In-memory SQLite with event-driven flush — all reads and writes go to RAM; disk is written only when data actually changes (see below)
• Rayon parallel parsing — files are parsed across all CPU cores simultaneously
• Bulk mode — for large batches: drop indexes, bulk insert, rebuild indexes; significantly faster than incremental inserts
• mtime/size fast-path — on restart, each file is checked via stat() (mtime + file_size); if both match the stored values, the file is not read at all — zero I/O, zero SHA-256. Only changed files are read and re-hashed
• PID-lock — prevents multiple daemon instances from competing for the same index.db

Flush to disk policy

The daemon works in in-memory mode for maximum performance. The database is flushed to disk only when data actually changes — no periodic timers, no unnecessary I/O:

This means: if you're just chatting with AI and not editing code, the daemon produces zero disk I/O.

• Batch transactions — 500 records per transaction reduces SQLite overhead by orders of magnitude

For 1C Developers

Code Index has first-class support for 1C:Enterprise source files.

From BSL files, it extracts:

• Procedures and functions with full body text
• Compilation directives (&AtServer, &AtClient, &AtServerNoContext)
• Extension annotations (&Instead, &After, &Before)
• Bilingual keywords (Russian and English forms are both indexed)

These are stored in two dedicated fields:

• override_type: "Перед" (Before), "После" (After), or "Вместо" (Instead)
• override_target: name of the original procedure being overridden

From XML configuration exports, it extracts:

• Metadata objects: catalogs, documents, registers, and more
• Attributes and tabular sections
• Forms and their composition

This makes Code Index suitable as an offline search layer over full 1C configuration exports without requiring a running platform instance.

System Requirements

• OS: Windows, Linux, macOS
• RAM: 512 MB for small projects; up to 4 GB for large 1C configurations (60K+ files)
• Disk: index size is approximately 1-2 GB for projects with 60K+ files
• Build: Rust 1.77 or later — install from rustup.rs

MCP tool response format (v0.9.0+)

All data tools return a unified JSON envelope:

{
  "result": <previous plain payload>,
  "_meta": {
    "dependent_files": ["src/X.bsl", "src/Y.bsl"],
    "file_mtimes": { "src/X.bsl": 1717689600, "src/Y.bsl": 1717689600 }
  }
}

_meta.dependent_files lists files the response depends on; _meta.file_mtimes (0.20.0+) maps each of them to its indexed mtime (unix seconds). As of 0.42.0 serve strips _meta from the client response itself — the model only ever receives result. _meta is now an internal signal: serve has a built-in result cache (cross-session, local repos only) with per-file freshness. On a file change the daemon sends POST /mark-dirty {repo, files:[{path, mtime}]} (files changed on disk), then after commit POST /invalidate {repo, file_paths}. A response is not cached/served from cache only if one of its source files is dirty (disk mtime newer than the indexed mtime in _meta.file_mtimes); per-file invalidation drops only the cache keys depending on the changed file (a file→keys reverse index). Requests about untouched files are unaffected — no whole-repo coarsening. The channel is wired by [[cache_targets]] in daemon.toml. The same _meta was previously consumed by the companion caching proxy mcp-cache-ci; with serve's own cache the standalone proxy is no longer required for the ci chain. Observability: GET /cache-stats on serve.

Diagnostic tools (health, get_stats, stat_file) are not wrapped — their format is unchanged.

License

MIT. See LICENSE.

Acknowledgements

• tree-sitter — incremental parsing library
• tree-sitter-onescript — BSL/OneScript grammar by the 1c-syntax community
• rusqlite — SQLite bindings for Rust
• rayon — data parallelism for Rust
• rmcp — Rust MCP SDK