Release v6.6.7

[code-crusher]

Release v6.6.7

FFF-backed search_files with ripgrep fallback, pagination, and compact results.

Added

• FFF-backed search_files with ripgrep fallback. search_files now runs through the native @ff-labs/fff-node file finder first, falling back to a streaming ripgrep implementation when FFF is unavailable or fails. Results are compact (at most three matches per file), paginated, and capped at a configurable page size so the model reads only the relevant regions instead of a context-flooding dump.
  • New src/services/search-files module (index.ts, format.ts, types.ts) is the single entry point: it tries FFF, falls back to ripgrep, and formats a compact Engine / Matches / Next cursor page.
  • New src/services/fff/index.ts wraps the ESM-only FFF SDK, keeps it out of the CommonJS bundle via a dynamic import(), caches up to three FileFinder instances per base path, encodes the user regex for FFF's query parser, and continues through FFF pages until .orbitalignore-filtered matches fill the requested page.
  • New searchFilesWithRipgrep in src/services/ripgrep/index.ts streams JSON ripgrep output, applies .orbitalignore before consuming the page budget, enforces the per-file match cap, and returns a ripgrep-engine cursor for continuation.
  • New cursor, max_results (1-100, default 50), and context_lines (0-2, default 0) parameters on both the XML and native JSON search_files tool schemas; cursors are engine-scoped (fff:<offset> / ripgrep:<offset>) and validated by parseSearchCursor / normalizeNullableSearchString.
  • searchFilesTool now parses the new params and delegates to searchFiles; toolParamNames and SearchFilesToolUse extended accordingly.
  • esbuild.mjs copies the FFF runtime packages (and their platform-specific native libraries) under dist/fff/node_modules so the packaged extension can resolve the native SDK; root package.json declares pnpm.supportedArchitectures so all platform FFF binaries install.
  • extension.ts disposes FFF finders and closes the native library on deactivate.
  • Tests for the FFF wrapper, ripgrep fallback, search-files orchestration, and cursor/format helpers.

Changed

• search_files system prompt rewritten. The tool description and capabilities section now describe compact, paginated matches and direct the model to read the relevant file region rather than relying on large context windows. The verbose file_pattern quoting examples were removed in favor of a concise parameter list.
• ChatRow shows the file_pattern for search_files rows. Parallel searches are now distinguishable in the chat history by their glob, rendered as a <code> chip next to the regex.
• Pinned todo list horizontal padding. ChatView pinned-todo container padding aligned with the shared mx-3.5 chat column padding.

Version bump

src/package.json bumped from 6.6.5 to 6.6.7.

---

• pnpm lint
• pnpm check-types

[matterai-app]

Context

This release refines the FFF (Fast File Finder) integration by addressing a reliability issue where transient native library load failures would permanently disable the high-performance search engine for the duration of a session.

Implementation

Added a .catch block to the dynamic import() call for the FFF native module. If the promise rejects, the cached fffModulePromise is reset to undefined before re-throwing the error. This ensures that subsequent calls attempt to reload the library rather than returning a permanently rejected promise.

Screenshots

before	after
N/A	N/A

How to Test

1. Simulate a transient failure by temporarily blocking access to the native binary.
2. Trigger a search and observe the failure/fallback.
3. Restore access to the binary.
4. Trigger another search and verify that the FFF engine successfully initializes without a session restart.

Summary By MatterAI

🔄 What Changed

Implemented error recovery for the FFF native module loader. The system now resets the internal module promise on rejection, preventing a single failed import from permanently degrading the search experience to ripgrep for the entire session.

🔍 Impact of the Change

Improves system resilience and search performance reliability. Users will no longer be locked into the slower ripgrep fallback if a transient environment issue occurs during the initial FFF load attempt.

📁 Total Files Changed

Click to Expand

File	ChangeLog
src/services/fff/index.ts	Resilient Loading: Added catch-and-reset logic to the native module import promise.

🧪 Test Added/Recommended

Recommended

• Unit Test: Mock the dynamic import() to reject once and then resolve, verifying that the service successfully recovers on the second call.

🔒 Security Vulnerabilities

N/A. This is a logic-level reliability fix for internal module loading.

Caution

Package Vulnerabilities

Package	Version	Severity	CVE	Fix Version	Vulnerability
esbuild	^0.25.0		GHSA-gv7w-rqvm-qjhr	0.28.1	Withdrawn Advisory: esbuild: Missing binary integrity verification in Deno module enables remote code execution via NPM_CONFIG_REGISTRY
glob	^11.0.3		CVE-2025-64756	11.1.0	glob CLI: Command injection via -c/--cmd executes matches with shell:true
turbo	^2.5.6		CVE-2026-45773 CVE-2026-45772	2.9.14	Turbo: Login callback CSRF/session fixation Turbo: Unexpected local code execution during Yarn Berry detection

[matterai-app]

🧪 PR Review is completed: Release v6.6.7 introduces a new FFF-based search engine with ripgrep fallback, paginated cursors, and compact output formatting. Two issues found: FFF module load failures are never retried, and non-integer search options can reach native APIs.

Skipped files

• CHANGELOG.md: Skipped file pattern
• pnpm-lock.yaml: Skipped file pattern

⬇️ Low Priority Suggestions (1)

src/services/search-files/types.ts (1 suggestion)

Location: src/services/search-files/types.ts (Lines 47-48)

🟡 Type Safety

Issue: clampSearchOptions does not floor maxResults or contextLines. If the AI passes a fractional value like "50.5", Number("50.5") produces 50.5 which passes through clampSearchOptions and reaches the FFF native API as pageSize: 50.5 and the ripgrep comparator. Native FFI APIs typically expect integer page sizes, which could cause undefined behavior or errors.

Fix: Apply Math.floor in clampSearchOptions to guarantee integer values.

Impact: Prevents fractional values from reaching native APIs and ensures consistent integer page sizes across both search engines.

-  		maxResults: Math.min(Math.max(options.maxResults ?? DEFAULT_SEARCH_RESULTS, 1), MAX_SEARCH_RESULTS),
-  		contextLines: Math.min(Math.max(options.contextLines ?? 0, 0), MAX_SEARCH_CONTEXT_LINES),
+  		maxResults: Math.min(Math.max(Math.floor(options.maxResults ?? DEFAULT_SEARCH_RESULTS), 1), MAX_SEARCH_RESULTS),
+  		contextLines: Math.min(Math.max(Math.floor(options.contextLines ?? 0), 0), MAX_SEARCH_CONTEXT_LINES),

[matterai-app]

✅ Reviewed the changes: The PR correctly adds a .catch() handler to reset fffModulePromise on rejection, addressing the transient failure retry issue. The implementation is sound with no new issues introduced.