Release v6.7.8: agent investigation-efficiency guidance

[code-crusher]

Summary

Version bump to 6.7.8. This release improves the agent's investigation efficiency when answering comprehension questions.

Changes

• Investigation efficiency section in system prompt (src/prompts/system.ts): directs the agent to classify comprehension vs implementation questions, form a one-line hypothesis before searching, read the call site rather than the implementation, never read prose/content for control-flow questions, and stop exploring once it can answer.
• Zero-result guidance in search_files executor (src/tools/executors/searchFiles.ts): appends actionable guidance when a search returns 0 matches — tighten/simplify the regex, widen path scope, try a different glob, stop after 2+ failed attempts.
• Tool description improvements: read_file schema tells the model not to read file contents when investigating control flow and not to re-read regions already read; search_files schema tells it to scope path to the narrowest plausible directory and to stop after 2+ zero-result searches.
• CHANGELOG.md updated under [Unreleased]; package-lock.json resynced to 6.7.8.

Testing

• npm run typecheck passes

After merge: tag v6.7.8 on main to publish.

[matterai-app]

Summary By MatterAI

🔄 What Changed

Bumped CLI version to 6.7.8 and integrated agent investigation efficiency guidelines across system prompts, tool schemas (read_file, search_files), and search execution handlers to minimize redundant token usage and unproductive exploratory tool calls.

🔍 Impact of the Change

Improves LLM agent responsiveness and token economy by enforcing stricter hypotheses-driven investigation workflows, preventing circular search retries when zero matches occur, and directing focus to call sites rather than deep implementation reads.

📁 Total Files Changed

Click to Expand

File	ChangeLog
Bump Version package.json	Updated project release version to 6.7.8.
System Prompts src/prompts/system.ts	Added investigation efficiency guidelines covering hypothesis testing and read discipline.
Search Execution src/tools/executors/searchFiles.ts	Appended recovery guidance when search returns zero matches to prevent redundant retries.
Read Schema src/tools/schemas/read_file.ts	Updated tool description to enforce batching and discourage prose reads for control flow.
Search Schema src/tools/schemas/search_files.ts	Updated tool description to mandate narrow scoping and strict regex tightening on zero results.

🧪 Test Added/Recommended

Recommended

• Add unit tests validating prompt inclusion of efficiency directives and verifying search executor response formatting on zero matches.

🔒Security Vulnerabilities

No security vulnerabilities introduced.

[matterai-app]

🧪 PR Review is completed: Release v6.7.8 adds agent investigation-efficiency guidance to system prompts and tool schemas, plus no-match guidance in searchFiles executor. One logic issue found: the zero-match detection uses fragile string matching instead of checking the data directly.

Skipped files

• CHANGELOG.md: Skipped file pattern
• package-lock.json: Skipped file pattern

⬇️ Low Priority Suggestions (1)

src/tools/executors/searchFiles.ts (1 suggestion)

Location: src/tools/executors/searchFiles.ts (Lines 90-90)

🟡 Logic Error

Issue: The zero-match detection uses output.includes("Matches: 0") which performs a substring search across the entire formatted output. If a search returns matches but one of the matched lines contains the literal text Matches: 0 (e.g., searching for Matches in this very codebase), the guidance text will be incorrectly appended even though matches were found. The page object is directly available and has matches.length.

Fix: Check page.matches.length === 0 directly instead of string-matching the formatted output.

Impact: Eliminates false positives where match guidance is shown when matches actually exist, preventing confusing output for the AI agent.

-  		if (output.includes("Matches: 0")) {
+  		if (page.matches.length === 0) {

[matterai-app]

✅ Reviewed the changes: The PR correctly replaces the fragile string-based zero-match detection with a direct data check on page.matches.length. Previous review comment is resolved; no new issues found.