AntiAgent v0.3.0: Unified Antigravity Conversation Browser & Native Context Management

Summary

Introduced the Unified Antigravity Conversation Browser + Native agy Context Management Integration. AntiAgent now discovers, indexes, inspects, and manages conversations across all Antigravity runtimes—Desktop, IDE, and CLI/agy. Rather than duplicating LLM context compression, AntiAgent deeply integrates with Antigravity's native compaction engine: detecting checkpoints, compactions, and resets while exposing rich context telemetry. Includes robust security isolation (blocking path traversal, symlink escapes, and credential leakage), strict chain-of-thought protection (never exposing internal thinking), high-performance bounded indexing with SQLite summary caching, full CLI commands, and a sleek dashboard interface.

Architectural & Functional Highlights

Area / Component	Improvement
Unified Multi-Source Discovery (engine/conversations.py)	Centralized session discovery across Desktop (~/.gemini/antigravity), IDE (~/.gemini/antigravity-ide), and CLI (~/.gemini/antigravity-cli, ~/.gemini/agy) using canonical (source, conversation_id) identity with environment variable isolation (ANTIAGENT_DESKTOP_DIR, ANTIAGENT_IDE_DIR, ANTIAGENT_CLI_DIR).
Native Context & Compaction State Inspector	Exposes native Antigravity auto-compaction and transcript lifecycle events (CHECKPOINT, COMPACTION, CONTEXT_RESET, SESSION_RESET). Computes turn counts, tool executions, transcript byte size, and extracts the latest compaction checkpoint summary directly from transcripts.
Strict Security & Privacy Enforcement	Rejects directory traversal (../, absolute paths) and verifies path containment preventing symlink escapes. Automatically redacts sensitive credentials in tool arguments (password, token, secret, api_key, authorization, Bearer tokens, GitHub PATs, OpenAI keys, PEM private keys).
Chain-of-Thought (CoT) Protection	Strips internal agent reasoning (thinking) across normalized message streams, dashboard UI, and Markdown/JSON transcript exports. Only user requests, assistant responses, and sanitized tool calls are rendered.
High-Performance Bounded Indexing	Optimized for massive histories using bounded head/tail byte inspections (<8KB head for first prompt and model, <16KB tail for latest prompt and compactions) combined with read-only SQLite conversation_summaries.db caching and in-memory mtime/size checks.
Transcript Path Forwarding (hook.py & context_extractor.py)	Hook payloads now capture and forward transcriptPath, enabling instant context resolution without redundant filesystem scans.
CLI Conversation Management (antiagent/cli.py)	Added full CLI suite: antiagent conversations list (with --source, --search, --limit, --json), antiagent conversations show <id>, `antiagent conversations export [--format markdown
REST API Suite (antiagent/dashboard/server.py)	Added 6 secure endpoints: GET /api/conversations, GET /api/conversations/detail, GET /api/conversations/context, POST /api/conversations/resume, POST /api/conversations/open, and POST /api/conversations/export.
Unified Dashboard UI (index.html)	Added glassmorphic Conversations card with real-time source tabs (All, Desktop, IDE, CLI), 300ms debounced search, native context telemetry banner, conversation detail & context modal, transcript viewer, and resume/export actions.
Doctor Diagnostics (engine/doctor.py)	Enhanced run_doctor_check to report structured session breakdowns across Desktop, IDE, and CLI sources with direct jump actions in the dashboard.

📥 Which download is right for you?

Package	Operating System	Description
AntiAgent.dmg	macOS	Recommended for Mac. Drag-and-drop installer disk image with automatic application bundle placement.
AntiAgent-Windows.zip	Windows (Public Beta)	Recommended for Windows (Beta). Standalone portable package with native application launcher.
AntiAgent.pkg	macOS	Standard auto-launching installer package for enterprise / MDM deployment.
AntiAgent.zip	macOS	Standalone portable .app bundle without disk image mounting.

ℹ️ Windows Support (Public Beta): The Windows standalone app mode is currently in Public Beta. If you encounter any platform-specific quirks or edge cases, please submit an issue on GitHub!