-
Notifications
You must be signed in to change notification settings - Fork 2.8k
refactor(read_file): Codex-inspired read_file refactor EXT-617 #10981
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Conversation
Re-review for 1a9bbc8 complete. One remaining item still applies.
Mention @roomote in a comment to request specific changes to this pull request or fix all unresolved issues. |
5c7a773 to
5785b36
Compare
|
This requires us to enable parallel tool calling by default, moving it out of experimental |
68ef238 to
9c670b4
Compare
123aacc to
1c3ef7d
Compare
3ae0e4d to
ff54851
Compare
mrubens
left a comment
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Sorry to keep nagging on this, but are we confident that this won't break existing conversations for people? If there's a chance it will I think we need to discuss the rollout plan more.
My tests worked without a problem |
Added fallback for legacy format so if the model gets confused and sends the old format Roo will not break! |
3cea212 to
1704ec7
Compare
|
@roomote Your review comment about the rooignore validation passing an absolute path in Looking at the validateAccess(filePath: string): boolean {
// ...
const absolutePath = path.resolve(this.cwd, filePath)
// Follow symlinks...
const relativePath = path.relative(this.cwd, realPath).toPosix()
return !this.ignoreInstance.ignores(relativePath)
}The method correctly handles both absolute and relative paths by:
This is also confirmed by the existing test suite in All other issues from your review (max_lines coercion, unused imports, anchor_line fallback) have been fixed. Please approve the PR. |
1704ec7 to
6029495
Compare
- Replace multi-file read_file with single-file-per-call design - Add two reading modes: slice (default) and indentation - Implement bidirectional expansion algorithm for indentation mode - Add line truncation (500 chars) and limit (2000 lines default) - Remove legacy token-budget-based reading approach - Remove maxReadFileLine setting (replaced by limit parameter) - Add new IndentationParams and ReadFileParams types - Clean up stale FileEntry/LineRange types and helpers Known limitations: - Lines >500 chars are truncated (content lost) - No server-side max limit enforcement
This setting was never used after removing the batch file reading system. Removes dead code from UI, state management, and types.
- Fix formatPathTooltip to add space before additionalContent (fixes display bug where 'index.ts' + 'up to 50 lines' showed as 'index.tsup to 50 lines') - Add startLine field to ClineSayTool type for read_file operations - Update ReadFileTool to include startLine in the message - Update ChatRow to pass startLine to openFile for navigation to the correct line
…ines - Add explicit 'when to use' guidance for slice and indentation modes - Add indentation mode example alongside existing slice example - Fix confusing 'indentation.anchor_line' text in mode description - Add 'when NOT to use' guidance in anchor_line description - Enhance descriptions to help non-OpenAI models choose appropriate mode This follows Anthropic's documented best practices for tool descriptions: - Explain what the tool does - Specify when to use and when NOT to use - Describe parameter impact - Aim for 3-4+ sentences
- Remove terminalCompressProgressBar from SettingsView.tsx destructuring - Remove terminalCompressProgressBar from test mock objects - Property was never defined in GlobalSettings or ExtensionState types
When mode='indentation' but anchor_line is not provided, now defaults to the offset parameter (or 1 if neither is provided), rather than silently falling back to slice mode. This aligns with the documented behavior in packages/types/src/tool-params.ts.
- Add truncation support to extractTextFromFile via readWithSlice (2000 line limit) - Update parseMentions to return file content as separate MentionContentBlock objects - Update processUserContentMentions to handle new contentBlocks structure - Add Gemini-style truncation warnings with IMPORTANT header - Sync truncation message format between read_file tool and @ mentions - Put truncation warning at TOP (before content) in both implementations - Update test mocks and expectations for new behavior
- Changed getLineSnippet() to always return 'up to X lines' even for default limit - Previously only showed line count when limit < DEFAULT_LINE_LIMIT - Now users always see how many lines will be read (e.g., 'up to 2000 lines')
When the model makes multiple parallel read_file tool calls, the UI now consolidates consecutive read_file ask messages into a single batch view showing 'Roo wants to read these files' instead of showing repeated 'Roo wants to read this file' messages for each file. The groupedMessages useMemo in ChatView now detects consecutive read_file asks and creates a synthetic batch message with batchFiles, which triggers the existing BatchFilePermission component to render the files as a group. This improves the UX by reducing visual noise when reading multiple files.
Remove IndentationReadResult and MAX_LINE_LENGTH imports that were not being used, as flagged by the review bot.
…pproval flow Adds comprehensive tests for ReadFileTool covering: - Input validation (missing path parameter) - RooIgnore blocking - Directory read error handling - Image handling (memory limits, format detection, model support) - Binary file handling (PDF, DOCX, unsupported formats) - Text file processing (slice and indentation modes) - Approval flow (approve, deny, feedback) - Output structure formatting - Error handling (file read errors, stat errors) 29 new tests covering the orchestration layer that was previously untested after the Codex-inspired refactor.
- Restore convertFileEntries() function in NativeToolCallParser
- Detect and handle both old { files: [...] } and new { path, mode, ... } formats
- Handle double-stringified JSON from models (files as string vs array)
- Add executeLegacy() method in ReadFileTool for multi-file batch processing
- Add READ_FILE_LEGACY_FORMAT_USED telemetry event
- Add console.warn indicator when legacy format is detected (for testing)
- Add 7 new tests for backward compatibility scenarios
- Add legacy param names (files, line_ranges) to toolParamNames
…itted - getLineSnippet() now always displays 'indentation mode at line X' when mode is indentation - getStartLine() now always returns the effective anchor line for indentation mode - Both methods use fallback chain: anchor_line ?? offset ?? 1 Addresses review feedback about approval UI not reflecting defaulted anchor_line
- Clarify when to use each mode with balanced positive framing - Slice: ideal for exploration, understanding structure, config files - Indentation: PREFERRED when you have a line number for complete blocks - Add warning about anchor_line requirement for indentation mode - Add concrete example from condensed file summaries format - Strengthen guidance in mode, anchor_line, and indentation field descriptions - Increase MAX_LINE_LENGTH from 500 to 2000 characters
061496b to
3e25f9c
Compare
…p line info in brackets
| if (params.indentation?.anchor_line !== undefined && params.indentation.anchor_line < 1) { | ||
| const errorMsg = `anchor_line must be a 1-indexed line number (got ${params.indentation.anchor_line}). Line numbers start at 1.` | ||
| pushToolResult(`Error: ${errorMsg}`) | ||
| return |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
The current behavior rejects offset/anchor_line <= 0 with an error, but earlier reviews asked to clamp to the documented defaults (>= 1) so invalid model output degrades gracefully (and avoids “line 0” approval UI). If clamping is the intended UX, these early returns should normalize to 1 instead of failing.
Fix it with Roo Code or mention @roomote and request a fix.
Summary
closes #10239
This PR introduces a smarter way for Roo to read your code files. Instead of just grabbing arbitrary line ranges, Roo can now intelligently extract complete functions, classes, and code blocks based on their structure.
What's New
Indentation Mode for Smart Code Extraction
When Roo needs to read code around a specific line (like when investigating an error or following a search result), it can now use indentation mode to automatically extract the complete containing function or class—not just a fixed number of lines that might cut off mid-block.
Example: If Roo finds an error on line 42 inside a function, indentation mode will return the entire function definition with proper context, instead of an arbitrary range like "lines 40-60" that might miss important context or include unrelated code.
Simplified File Reading
The
read_filetool now has two clear modes:Breaking Changes
Settings Removed
The following settings have been removed from the UI:
If you previously relied on
maxReadFileLineto limit file reads, this behavior is now handled automatically by the agent's per-request parameters.API Changes (for Custom Modes/Prompts)
The
line_rangesparameter has been replaced with a simpleroffset/limitapproach:Before:
{ "files": [{ "path": "app.ts", "line_ranges": [[1, 50]] }] }After:
{ "path": "app.ts", "mode": "slice", "offset": 1, "limit": 50 }For semantic code extraction, use the new indentation mode:
{ "path": "app.ts", "mode": "indentation", "indentation": { "anchor_line": 42 } }Backward Compatibility
Legacy Format Support
To ensure smooth mid-chat updates, the parser maintains full backward compatibility with the old
read_fileformat. If a model uses the legacy format (based on chat history from before the update), it will still work correctly.Supported legacy formats:
{ files: [{ path: "...", line_ranges: [[start, end]] }] }- Array format{ files: "[{...}]" }- Double-stringified JSON (model quirk)line_rangesas tuples, objects, or stringsHow it works:
NativeToolCallParserdetects the presence of afilesparameterReadFileTool.executeLegacy()handles multi-file batch processingREAD_FILE_LEGACY_FORMAT_USED) is captured for monitoring adoptionTesting indicator: A
console.warn("[read_file] Legacy format detected")message appears when the fallback path is used (temporary, for validation).Benefits