feat!: multi-tool adapters — Codex CLI, Cursor, ORGN CDE, OpenCode (v0.2.0)

[Pycomet]

What: One unified capability map across every supported AI coding tool on the machine — adapter registry with detection, four new adapters, cross-tool diagnostics, and per-tool tuning bars (SPEC_V2 M5–M8).

Why: v1 only answered "what can my Claude Code setup do?" The real question is "what can my whole AI setup do, where does it overlap, and what am I paying for but never using?" — duplicate MCP servers across tools and capability imbalance are invisible in any single-tool view.

How:

• ToolAdapter v2: detect() per adapter, tool-namespaced ids (claude-code/skill:x), --list-tools / --tool CLI surface, tools + crossTool in --json
• Codex CLI (full usage): config.toml MCP servers + AGENTS.md + session JSONL mining; the desktop-app variant has no stable public surface and honestly reports "not detected"
• Cursor (inventory-only): mcp.json (project wins collisions), skills-cursor/*/SKILL.md, rules files; nodes render fixed-size with a "usage unavailable" badge — no faked numbers
• ORGN CDE + OpenCode (partial usage): shared opencode-family core — string-aware JSONC parser, read-only SQLite session mining (copy-on-read under WAL locks, degrades to inventory-only on schema surprises); schema survey in docs/opencode-db-survey.md
• Cross-tool diagnostics: duplicate MCP servers by normalized command/URL identity with usage contrast, per-axis capability imbalance, rules-file overlap flags (human review, no content diffing)
• Dead weight is now tool-aware: usage-less tools excluded (no data → no "never used" claim), token math uses the owning tool's session count
• Prereq fix from the M1–M3 review: plugin-skill usage credited across prefix mismatches when unambiguous

Breaking: item ids are tool-namespaced (skill:x → claude-code/skill:x); --json gains tools/crossTool; version bumped to 0.2.0. See CHANGELOG.md.

Testing: 96 vitest tests green (fixture trees per adapter incl. a generated SQLite fixture); verified live on a real machine — 4/5 tools detected, 113 items, real cross-tool findings ("playwright MCP installed in Claude Code + Cursor, never fired in any tool with usage data").

Size note: >400 lines by necessity — four adapters built against one interface landed as one reviewable unit; each adapter is an isolated module + fixtures + tests, reviewable per-commit (one commit per plan task).

🤖 Generated with Claude Code

https://claude.ai/code/session_01WjehirYUmeq8GdcYe5U6RJ