Skip to content

cross-llm-coverage.md

Latest commit

 

History

History
44 lines (32 loc) · 5.75 KB

cross-llm-coverage.md

File metadata and controls

44 lines (32 loc) · 5.75 KB
Raw

Cross-LLM Capability Coverage

Host-by-host capability matrix for the Autonomous Dev Kit skills system.

✅ = fully supported ⚠️ = partial / workaround required ❌ = not supported

Capability Claude Code Codex CLI OpenCode Cursor Hermes Agent Zed Agent
SKILL.md import ✅ native ✅ native ✅ native ✅ plugin manifest defines skills/agents/commands/hooks; install via /plugin-add autodev ✅ native via ~/.hermes/skills/ symlink or hermes skills install ✅ native via flat direct children under ~/.agents/skills/ or <worktree>/.agents/skills/
Sub-agent dispatch ✅ Agent tool ✅ natural language ⚠️ @mention to peer sessions ❌ not documented ✅ delegate_task tool ⚠️ native subagent tool when available; otherwise parallel Zed Agent threads
Agent Teams (persistent multi-agent DM) ✅ experimental flag ❌ ❌ ❌ ❌ (uses sequential delegate_task) ❌ (use Threads Sidebar / worktree isolation)
Background agents ✅ run_in_background ⚠️ thread-based; no explicit background flag ❌ not documented ❌ not documented ✅ terminal(background=true) ⚠️ multiple threads; no ADK lifecycle hook reinvocation
MCP servers ✅ ✅ config.toml ✅ ⚠️ partial ✅ hermes mcp ✅ Zed Agent Settings / context_servers
Slash commands ✅ ✅ 30+ built-ins incl. /plan, /agent, /review ✅ ✅ ✅ 40+ built-ins ✅ skill slash commands and @-mentions
Plan mode ✅ EnterPlanMode + Shift-Tab ✅ /plan slash ⚠️ not documented; use prose planning ⚠️ built-in Composer; not slash-invokable ⚠️ prose planning in chat ⚠️ prose planning / Agent Panel workflow
Task list / TodoWrite ✅ built-in ❌ no documented equivalent ⚠️ update_plan mapping (see .opencode/INSTALL.md) ⚠️ unknown ✅ todo tool (built-in) ⚠️ orchestrator checklist; no ADK shared queue
AGENTS.md / project context CLAUDE.md AGENTS.md (+ .override.md) AGENTS.md n/a AGENTS.md (cwd only), .hermes.md (parent walk) ~/.config/zed/AGENTS.md; project AGENTS.md and compatible instruction files
Host declaration for skill conditionals Host: claude-code in CLAUDE.md Host: codex in ~/.codex/AGENTS.md Host: opencode in ~/.config/opencode/AGENTS.md n/a Auto-detected as hermes-agent Host: zed-agent in ~/.config/zed/AGENTS.md or %APPDATA%\\Zed\\AGENTS.md
Skill discovery path (user scope) ~/.claude/skills/ (personal skills); autodev installed to ~/.claude/plugins/marketplace/autodev/ via marketplace ~/.agents/skills/ ~/.config/opencode/skills/ via plugin (no manual symlink) ~/.hermes/skills/ (symlink or hermes skills install) ~/.agents/skills/<skill-name>/ (flat; no namespace folder)
Model tier vocabulary role names → haiku/sonnet/opus (see agents/model-tiers.md) role names → gpt-5.4-mini/gpt-5.4/gpt-5.5 role names → host-pass-through role names → host-pass-through role names → host-pass-through role names → host-pass-through
Visual companion output ⚠️ markdown/Mermaid best-effort; browser deferred ⚠️ markdown/Mermaid best-effort; browser deferred ⚠️ markdown/Mermaid best-effort; browser deferred ⚠️ markdown/Mermaid best-effort; browser deferred ⚠️ markdown/Mermaid best-effort; browser deferred ⚠️ markdown plus best-effort Mermaid rendering; browser deferred

Notes

Sub-agent dispatch (Codex): Codex uses natural-language spawn ("spawn one agent per X") rather than an explicit Agent tool call. The <host: codex> blocks in skills provide the correct phrasing.

Agent Teams: The TeamCreate / SendMessage persistent-chat pattern is exclusive to Claude Code (experimental flag). Skills fall back to Sequential Mode (one sub-agent at a time) on all other hosts — see skills/subagent-driven-development/SKILL.md.

Task list (Codex): No built-in task-tracking tool is documented in Codex CLI. Skills that reference TodoWrite wrap those references in <host: claude-code> blocks; the host-neutral path uses prose checklists.

Cursor: The .cursor-plugin/plugin.json manifest defines skills, agents, commands, and hooks. Installation is via /plugin-add autodev in the Cursor agent chat (same marketplace mechanism as Claude Code). Skill discovery path (user scope) is managed through the plugin; no manual symlink required.

Zed Agent: Zed Skills apply only to Zed Agent, not ACP External Agents or Terminal Threads. Zed requires a flat skills layout, so ADK installs each skills/<name>/ directory directly under ~/.agents/skills/ or <worktree>/.agents/skills/; project-local skills require a trusted worktree. Zed does not expose the Claude/Codex lifecycle hook contract for skill packages. Current Zed task hooks are narrower (create_worktree is the documented event), so ADK hook-driven reminders use manual skill fallbacks and optional Zed Tool Permissions on Zed.

Visual companion output: Brainstorming may emit markdown, mockups, or Mermaid diagrams where useful, but rendered diagrams are best-effort and every visual needs a text fallback. Browser companion and click/event capture are deferred.

Related files

  • tests/cross-llm-coverage.md — per-skill host-conditional vs host-neutral audit
  • tests/skill-content-grep.sh — CI guard: fails if forbidden tokens appear outside <host: claude-code> blocks
  • .codex/INSTALL.md — Codex setup instructions
  • .opencode/INSTALL.md — OpenCode setup instructions
  • .zed/INSTALL.md — Zed Agent setup instructions
  • agents/model-tiers.md — role-to-model-name resolution table (fast / balanced / frontier / coding-specialist)