ADFA-2621 | Add file-edit tool for local LLM models

[jatezzz]

Description

What:
This PR introduces the edit_file tool to the AI Assistant, allowing models to perform surgical find-and-replace changes on project files instead of attempting to rewrite them entirely.

How:

• Targeted Edits: The tool uses old_string and new_string arguments to locate and replace exact snippets, falling back to line-ending adaptations if needed.
• Diff-based Review UI: Replaces the generic JSON approval dialog with a dedicated before/after diff preview for file edits.
• Interactive Corrections: Adds a "Correct" option in the approval dialog, allowing users to reject an edit but provide written guidance back to the agent so it can retry in the same loop.
• Buffer & Disk Safety: Applies changes directly to the active editor buffer when a file is open (preserving unsaved work and allowing Ctrl+Z undo). For closed files, it uses atomic disk writes.
• Prompt Engineering: Updates system prompts to explicitly guide the LLM on using edit_file, emphasizing the need to copy old_string verbatim after reading the file.

Why:
Local LLMs generally lack the context window and output budget to reliably overwrite entire files using update_file. A targeted edit approach makes local, on-device models substantially more viable for iterative coding tasks.

Details

• Added EditFileHandler, EditorBufferApplier, and AtomicFileWriter to safely manage the lifecycle of an edit.
• Updated ToolApprovalManager to support CORRECTED states and strictly prevent edit_file from ever being granted blanket "Always Allow" session approval.
• Introduced AgentTrace for detailed lifecycle logging of prompt, tool execution, parsing, and execution.
• Bumped plugin version to 1.1.0.

document_5152462601887680330.mp4

Ticket

ADFA-2621

Observation

• Security/Safety: edit_file validations explicitly block writes to .git/, keystores, and build configuration files.

[claude]

Claude Code Review

This repository is configured for manual code reviews. Comment @claude review for a one-time review, or @claude review always to subscribe this PR to a review on every future push.

_{Tip: disable this comment in your organization's Code Review settings.}

[jatezzz]

@claude review

[claude]

⚠️ Code review skipped — your organization's overage spend limit has been reached.

Code review is billed via overage credits. To resume reviews, an organization admin can raise the monthly limit at claude.ai/admin-settings/claude-code.

Once credits are available, comment @claude review on this pull request to trigger a review.