[Bug] v2.6.0/v2.7.0 Inline Diff triggers false-positive Linter/Compiler errors (duplicate symbols) and breaks multi-prompt diff state

*(Note to the Product & UX Team: I am a native Spanish-speaking software developer actively building a production marketplace on Antigravity IDE. I used an AI assistant solely to translate and structure my real daily experience into clear English—this is genuine human feedback from a daily power user.)*

Over the last 7–10 days, my Antigravity IDE silently updated from the previous review architecture (v2.5.5) to the new v2.6.0 / v2.7.0 inline diff workflow. After trying hard to adapt to it for over a week on a large full-stack codebase, I have to report that this update has severely degraded code safety and cut my productivity in half. Here is why the previous diff viewer (v2.5.5) was vastly superior and how the new version breaks real-world development workflows:

  1. **Loss of the Floating Navigation Bar (Hunk-to-Hunk & File-to-File):** Previously, after an agent modified one or multiple files, a floating bottom navigation bar allowed me to jump instantly from change to change within a file, switch cleanly between modified files, or accept/reject all edits in a single file with one click. Now that bar is completely gone. I am forced to manually scroll through hundreds of lines of code trying to guess where edits occurred.

2. **Microscopic Inline Action Buttons:** In the new inline view, the Accept/Reject controls per hunk are tiny and visually buried in the code, making fast, confident code auditing painful.

3. **False-Positive Linter & Compiler Errors (Virtual Phantom Compilation):** The new inline overlay renders deleted (red) and inserted (green) blocks inside the active buffer in a way that triggers live inline compiler/linter errors (such as duplicate variable or constant declarations). Instead of helping me spot real bugs, the editor floods the file with false syntax errors until the diff is resolved.

4. **Destruction of Multi-Prompt Accumulated Reviews (Auto-Accept on Next Prompt):** In the previous version, if I wasn’t 100% sure about an edit, I could leave changes pending across 5 to 10 consecutive prompts while testing and iterating—and then bulk-reject all accumulated edits in a file or workspace if the AI went off track. By default in the new version, sending a follow-up prompt silently auto-accepts all pending changes (`antigravity.autoAcceptOnChat`), stripping away our safety net when an agent hallucinates or over-edits.

**What we need:** Please do not force this stripped-down review UX on developers who rely on strict code auditing. Give us an explicit toggle in Settings (e.g., `“antigravity.diffReviewStyle”: “classic-v2.5.5”`) to restore the floating hunk/file navigator bar, clear action buttons, and multi-prompt pending diff buffer.


Technical Details & Reproduction

  • Version: Antigravity IDE v2.6.0 / v2.7.0 (auto-updated from v2.5.5)
  • Steps to Reproduce:
    1. Ask any agent to modify existing variables, imports, or functions across one or more files.
    2. Open the modified file in the editor while the inline diff is active.
    3. Observe that the language server / linter evaluates both the deleted (red) and inserted (green) lines simultaneously, throwing false-positive syntax/duplicate declaration errors.
    4. Send a follow-up prompt in chat without accepting the previous diff: pending edits are silently auto-accepted (antigravity.autoAcceptOnChat defaults to true), and the bottom hunk/file navigation bar from v2.5.5 is missing.