Vendor capability matrix
LumenFlow stays vendor-agnostic by anchoring on shared concepts first:
- Persistent repo instructions
- Reusable skill or workflow surfaces
- Optional subagents or background agents
- Memory or personalization layers
- Model and reasoning controls
- Plugin or marketplace distribution where the vendor publishes a stable surface
The details vary by vendor. Use this matrix to decide which LumenFlow surfaces can be projected directly and which need a vendor-specific overlay.
Matrix
Section titled “Matrix”| Vendor | Instructions / rules surface | Skills / reusable workflows | Sub-agents | Hooks | A2A signal delivery | Slash commands | Memory / personalization | Model / effort controls |
| ----------- | ------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------- | --------------------------------------------------------- | ----------------------------------------- | --------------------------------------------------------- | ---------------------------------------------------------------------------------------------------- |
| Claude Code | CLAUDE.md, .claude/CLAUDE.md, .claude/rules/ | Native skills in .claude/skills/; custom slash commands | Yes. Native subagents in .claude/agents/ | Yes. Project/user/local hooks via settings, plus skill/agent-scoped hooks | Supported through LumenFlow PostToolUse hook delivery. | Yes | CLAUDE.md plus auto memory | /model and explicit effort controls are documented |
| Codex CLI | AGENTS.md (root + nested), configurable fallback filenames | Native skills from .agents/skills/ | Yes, but explicit ask only | Yes, via Codex hooks | Supported through foreground mem:watch daemon fallback. | Yes | Native memories are documented | /model; reasoning effort available when the selected model supports it |
| Cursor | .cursor/rules/, root AGENTS.md, root CLAUDE.md in CLI | No documented SKILL.md surface; use rules instead | Yes, background agents | No documented project hook surface in the public docs reviewed | Supported through foreground mem:watch daemon fallback. | CLI slash commands documented | Native Memories plus user rules | Model picker and Auto/Max modes documented |
| Windsurf | .windsurf/rules/ and location-scoped AGENTS.md | Native skills in .windsurf/skills/; also discovers .agents/skills/; workflows in .windsurf/workflows/ | Caution: public docs describe Cascade, skills, and workflows, but do not document user-authored subagents the way Claude/Codex/Gemini do | No documented project hook surface in the public docs reviewed | Supported through foreground mem:watch daemon fallback. | Yes, workflows run via /[workflow-name] | Native Memories and Rules | Model picker documented; no separate reasoning-effort control documented in the public docs reviewed |
| Gemini CLI | GEMINI.md by default; configurable to also read AGENTS.md | Custom commands and extensions are documented; no official .agents/skills/ skill surface found in the docs reviewed | Yes. Local and remote subagents under .gemini/agents | No documented project hook surface in the public docs reviewed | Supported through foreground mem:watch daemon fallback. | Yes | Hierarchical GEMINI.md memory plus /memory management | /model and per-agent model configuration are documented |
Plugin distribution overlay
Section titled “Plugin distribution overlay”| Vendor | Plugin status for LumenFlow |
| ----------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| Claude Code | Native plugins and marketplaces are documented. Plugin skills are namespaced, for example /lumenflow:design-first. Anthropic official marketplace submission exists through in-app submission forms. |
| Codex CLI | Native plugins can bundle skills, apps, MCP servers, and hooks. Repo and personal marketplaces are available; OpenAI documents official self-serve publishing as coming soon. Codex subagents remain outside the plugin manifest. |
| Cursor | Keep using rules and repo instruction surfaces. LumenFlow does not add a Cursor plugin format until a stable public packaging contract exists. |
| Windsurf | Keep using rules, workflows, and shared skills. LumenFlow does not add a Windsurf plugin format until a stable public packaging contract exists. |
| Gemini CLI | Keep using GEMINI.md, commands/extensions, and .gemini/agents. LumenFlow does not add a Gemini plugin format. |
Agent plugins are distribution packages, not LumenFlow runtime packs. A plugin can help a client discover skills and first-run guidance; Work Units, gates, evidence, shared memory, and pack policies still require the LumenFlow CLI/runtime inside the repository.
A2A signal delivery
Section titled “A2A signal delivery”The A2A signal contract is vendor-neutral, but local push delivery is adapter-specific. Claude Code
is currently supported through generated PostToolUse hook delivery for signal:received events.
Clients without an installed hook adapter use the foreground mem:watch daemon
fallback, which drains the same signal:received event queue and records
delivered receipts for the active reader identity.
Shared instruction surfaces
Section titled “Shared instruction surfaces”AGENTS.mdis a strong cross-vendor baseline today for Codex, Cursor, and Windsurf.- Claude Code still centers
CLAUDE.md, notAGENTS.md. - Gemini CLI centers
GEMINI.md, but its documented config allows additional context filenames, includingAGENTS.md.
Shared skills surface
Section titled “Shared skills surface”.agents/skills/is a real cross-vendor projection target today for Codex and Windsurf.- Claude Code still expects
.claude/skills/. - Cursor’s current public docs emphasize rules rather than a native
SKILL.mdskills surface. - Gemini CLI currently documents commands/extensions/subagents rather than a
SKILL.mdskill registry.
Sub-agent caution
Section titled “Sub-agent caution”Windsurf is intentionally marked cautiously in the sub-agent column. Its public docs clearly cover
Cascade, skills, rules, workflows, and memories, but the docs reviewed for this page do not expose
an explicit user-authored subagent surface comparable to Claude Code’s .claude/agents/, Codex’s
subagents, or Gemini CLI’s .gemini/agents.
Official vendor docs
Section titled “Official vendor docs”- Claude Code: Plugins, Discover plugins, Plugin marketplaces, Subagents, Hooks, Memory, Model configuration, Slash commands
- Codex: Plugins, Build plugins, AGENTS.md, Skills, Subagents, CLI features, CLI slash commands, Memories, GPT-5.3-Codex model
- Cursor: Rules, Memories, Background agents, CLI usage, Models
- Windsurf: AGENTS.md, Skills, Workflows, Memories & Rules, AI models
- Gemini CLI: GEMINI.md and context files, Subagents, Commands, Configuration