lninl/ i

Efficient file tools#

The browser and native agent share these built-in tools. The model receives their schemas automatically. They do not need a plugin or shell.

This is a new interface, not a compatibility layer: read.path/offset/limit and edit.old/new are replaced by the batch forms below. Existing conversation history can still be displayed, but old tool arguments must not be replayed.

  1. Discover paths with glob, or locate code directly with search.
  2. Read the relevant ranges from multiple files in one read.
  3. Put all same-file replacements in one edit.
  4. Batch independent tool calls in the same assistant response. Do not reread

unchanged files or repeat their contents in the final answer.

Searching first reduces irrelevant content in model context; batching reduces model round trips. Neither guarantees a particular token saving with a live model: model behavior, schema overhead and task size also matter.

Examples#

Find all source and test files without listing each directory:

json
{"patterns":["src/**/*.rs","tests/**/*.rs"]}

Search for relevant code (search):

json
{"pattern":"configure|validate","regex":true,"glob":"**/*.rs","context_lines":1,"limit":20}

Use mode: "files" when only matching filenames are needed. Search is literal and case-sensitive by default. Regex is line-based Rust regex, not JavaScript regex; lookaround and backreferences are not supported. Snippets are bounded around the first match on each line, with explicit truncation markers.

Batch reads (read):

json
{"files":[{"path":"src/app.rs","start_line":20,"end_line":50},{"path":"tests/app.rs","start_line":10,"end_line":30}],"max_bytes":12000}

Lines are 1-based; end is inclusive. Default is 200 lines per range, maximum 1000 lines per range and 16 ranges per call. Repeated ranges of one file reuse the same in-call read, not a persistent cache. Per-file failures are returned alongside successful ranges. Invalid ranges/paths reject the whole request.

Batch edits (edit):

json
{"path":"src/app.rs","edits":[{"old":"return false;","new":"return true;"},{"old":"old label","new":"new label","line_number":42}]}

Up to 64 replacements run sequentially against the evolving text. Each old must match exactly once; line_number disambiguates a single-line occurrence in the current text. All replacements and size limits are validated in memory before one write. An invalid replacement leaves the file unchanged. The file is reread before committing to detect concurrent changes. Line endings are preserved. This is single-file batching, not a multi-file transaction, and the concurrent-change check is not an OS-level compare-and-swap.

Output and scan bounds#

  • list, glob, search, read return compact JSON inside output, containing

results, truncated, next, and skipped.

  • Output budget defaults to 12000 bytes shared by all results in the call.

max_bytes is 1024..64000, also capped by the host's output allowance. This is a byte budget, not a tokenizer-specific token estimate.

  • list/glob: default 100 results, maximum 500. Pass the returned numeric

next as cursor to continue.

  • search: default 50 hits, maximum 200; 0..3 context lines. Pass the returned

{file_index,line} as cursor. line in this opaque cursor is zero-based; displayed result line numbers are one-based.

  • read: next identifies the original files index and the first unread

start_line. Retry that range and remaining files; preserve its original end. A single line too large for the budget returns an explicit error and a continuation, not an apparently complete but clipped line. Increase the budget or use search for a bounded snippet.

  • Cursors assume unchanged files/directories and unchanged query options;

restart discovery if files change. No workspace-wide index/cache is maintained.

  • Recursive tools visit at most 10000 entries. Larger trees fail explicitly:

narrow path. path limits traversal; glob patterns still match full workspace-relative paths, including the directory prefix.

  • By default recursion skips .git, node_modules, target, .i directories.

Set include_ignored: true to traverse them. This is a fixed directory list, not a .gitignore interpreter. An explicitly selected root is still searched.

  • Search skips binary, non-UTF-8 and oversize files and reports their count.

It reads at most 64 MiB per page based on file sizes, with a continuation for the next file. Read batches retain at most 64 MiB of text. Native/browser per-file limits and path confinement remain enforced.

  • A small result budget does not make disk scans free: a selective search can

still scan every eligible file. Narrow path/glob when possible.

Permissions and native manifests#

Built-in glob, search, list, read, view_image are read-only and do not ask for modification approval by default. Native approval hooks still see every call and can deny or require confirmation. edit/write keep their normal approval and persistence checks. Browser shell remains disabled.

Native full/standard manifests no longer advertise the redundant external search/find_files tools. glob replaces file discovery; built-in search uses pattern, not the old plugin's query. Regenerate old static manifests and remove custom entries colliding with built-in names. The search profile now has no extra tools because search is already built in.

在 GitHub 改进本页