What Gets Scanned

The scanner reads your local AI tool configuration directories — never modifying any files — to build a complete picture of your environment.

Architecture

Source Reference

Source	What's Extracted
~/.gemini/settings.json	MCP servers, model config, experimental flags, auth type, autoMemory
~/.gemini/skills/	All installed skills with descriptions
~/.gemini/agents/	Custom agent definitions
~/.gemini/extensions/	Extension catalog with enable/disable status
~/.gemini/policies/	Policy TOML configurations (v0.40+ fields: modes, mcpName, denyMessage, interactive, commandRegex, subagent, toolAnnotations)
~/.gemini/GEMINI.md	Global context file (Tier 1 memory)
~/.gemini/tmp/*/chats/*.jsonl	Conversation intelligence: tool frequency, models, topics, prompts, tokens, tool call chains
~/.gemini/tmp/*/memory/	Private project memory (Tier 2): MEMORY.md, sibling files, word counts
~/.gemini/tmp/*/memory/.extraction-state.json	Skill extraction state: run history, skills created, sessions processed
~/.gemini/tmp/*/memory/skills/	Inbox skills: awaiting approval, pending patches
/Library/Application Support/GeminiCli/policies/	Admin (system-level) policies — silently skipped if inaccessible
/etc/gemini-cli/policies/	Admin policies (Linux) — silently skipped if inaccessible
~/.claude/skills/	Claude Code skill catalog
~/.claude/CLAUDE.md	Claude context files
~/.gemini/antigravity/	Antigravity Desktop: brain conversations, skills, MCP servers, knowledge items
~/.gemini/antigravity-cli/	Antigravity CLI: brain sessions, skills, plugins, settings, history, import manifest, MCP servers
~/.gemini/antigravity-ide/	Antigravity IDE: brain conversations, skills, plugins, MCP servers
~/.continue/skills/	Continue skills (symlink-aware)
~/.codeium/windsurf/skills/	Windsurf skills (symlink-aware)
~/Library/.../JetBrains/Air/	JetBrains AI rules and presence
.opencode/	OpenCode config (JSONC), agents, skills, commands, custom tools, plugins, themes, glossary
AGENTS.md	OpenCode project context file (analogous to GEMINI.md)
--repos paths	Project-level .gemini/settings.json, skills, agents, GEMINI.md, .claude/ configs

Memory Architecture (4 Tiers)

The scanner maps and audits all four tiers of the Gemini CLI memory hierarchy:

Tier	Scope	Location	What's Checked
1. Global Personal	All sessions	~/.gemini/GEMINI.md	Word count, sections, bullet facts
2. Extension Context	Extension-scoped	~/.gemini/extensions/*/context/	Presence detection
3. Project Shared	Project team	<repo>/.gemini/GEMINI.md	Via --repos scanning
4. Private Project	User per-project	~/.gemini/tmp/<hash>/memory/	MEMORY.md content, sibling files, bloat detection

The advisory engine checks for:

• Missing tiers (e.g., no global memory configured)
• Memory bloat (>5,000 words in a single MEMORY.md)
• Cross-tier fact duplication (same bullet in global and private)
• autoMemory configuration status
• Under-utilized tier hierarchy

Skill Extraction State

The scanner detects the CLI's background skill extraction agent activity:

• Run history: How many extraction runs, when was the last one, how many sessions processed
• Skills created: Which skills were auto-generated vs. user-created (origin tagging)
• Inbox: Skills awaiting user approval, pending patches for existing skills
• Health: Stale extraction locks (crashed runs >30 minutes old)

Policy Engine Depth (v0.40+)

Policy scanning now covers all v0.40+ fields:

Field	What It Does	Advisory Check
modes	Scopes rules to default/plan/yolo	Warns if no mode-specific rules exist
mcpName	Scopes rules to specific MCP servers	Warns about ungoverned MCP servers
denyMessage	Custom message shown when rule blocks	Warns if deny rules lack explanations
interactive	Distinguishes headless vs. interactive	Info if no headless-mode rules exist
commandRegex	Regex-based command matching	Info about regex pattern brittleness
subagent	Scopes rules to specific sub-agents	Passed through for audit
toolAnnotations	Inline table metadata for tools	Parsed via upgraded TOML parser

Behavioral Pattern Detection

The scanner extracts tool call chains from each session (ordered sequences of tool calls) and aggregates them into behavioral fingerprints. These are used for:

• Skill suggestion gating: Repeating tool chain patterns (≥3 occurrences, ≥2 projects) provide evidence for reusable skill candidates
• Workflow visualization: The top 15 chain patterns are reported with their frequency counts
• Cross-project analysis: Chains spanning multiple projects indicate generalizable workflows

Recursive Repo Discovery

When you pass a parent directory like ~/Code to --repos, the scanner walks up to 3 levels deep (configurable via --repo-depth) looking for directories containing .git/. It automatically skips noise directories:

node_modules, .git, vendor, __pycache__, dist, build, .next, .venv, venv, .cache, .npm, .yarn, coverage, .terraform

Each discovered repo is logged during scanning so you can see exactly what's being processed.

Partial Environments (v3.5.8+)

When ~/.gemini is not found, the scanner automatically skips Gemini-specific sub-scans (settings, skills, agents, conversations, memory, policies) and continues with all other ecosystem and repository scans. The JSON manifest includes a gemini_cli_installed: false flag, and the markdown report displays a prominent installation banner.

Repository-level scans (--repos) still detect project-specific .gemini/ configurations, GEMINI.md, and CLAUDE.md files regardless of whether the global CLI directory exists.

AI Tool Ecosystem Detection

The scanner detects and reports on multiple AI coding tools:

• Gemini CLI — full config, skills, agents, extensions, policies, conversations, memory, extraction state (skipped if ~/.gemini missing)
• Claude Code — skills, CLAUDE.md, project configs
• Antigravity Desktop — brain conversations, skills, knowledge items, MCP servers
• Antigravity CLI — brain sessions, skills, plugins, settings (model, permissions), history, import manifest (Gemini CLI migration), cached projects
• Antigravity IDE — brain conversations, skills, plugins, MCP servers
• Continue — skills (symlink-aware)
• Windsurf — skills (symlink-aware)
• JetBrains AI — rules and presence detection
• OpenCode — config (JSONC), agents, skills, commands, custom tools, plugins, themes, glossary, AGENTS.md

Cross-tool skill overlap analysis highlights where the same workflow is covered by multiple tools, helping you consolidate and standardize.