A property suggestion for the Features group (or Identity, your call):
The property
envIdentity — does the agent identify itself in the environment of the commands it executes (env vars like AI_AGENT, GEMINI_CLI=1, CURSOR_AGENT), so that CLI tools, loggers, and build output can detect it and adapt?
"envIdentity": {
"value": "yes" | "no" | "partial",
"detail": "AI_AGENT=claude-code_<version>_<mode> (documented)"
}Why it fits the matrix
It's a real differentiation axis across your 24 agents, and the ecosystem is currently diverged:
- Documented flags: Claude Code (
AI_AGENT), Gemini CLI (GEMINI_CLI=1), OpenCode (OPENCODE=1), Copilot CLI (COPILOT_CLI=1), Qwen Code
- Undocumented but real: Cursor (
CURSOR_AGENT), Goose (AGENT=goose, some paths only), Aider (accidental OR_APP_NAME)
- No signal at all: Crush, Continue, and others — subprocesses can't tell they're inside an agent
That maps cleanly onto yes / partial / no, and it matters to the same audience comparing agentsMdSupport — it's the other direction of the same handshake (AGENTS.md = human tells agent; envIdentity = agent tells tools).
Data offer
We maintain ai-agent-detect, which documents exactly this per agent with evidence tiers (vendor docs > source code > community reports > live probes) — see docs/harnesses. Happy to send a PR filling envIdentity for all agents in the matrix, each with a sourced detail. Just say the word (and pick the property name/group you prefer).
A property suggestion for the Features group (or Identity, your call):
The property
envIdentity— does the agent identify itself in the environment of the commands it executes (env vars likeAI_AGENT,GEMINI_CLI=1,CURSOR_AGENT), so that CLI tools, loggers, and build output can detect it and adapt?Why it fits the matrix
It's a real differentiation axis across your 24 agents, and the ecosystem is currently diverged:
AI_AGENT), Gemini CLI (GEMINI_CLI=1), OpenCode (OPENCODE=1), Copilot CLI (COPILOT_CLI=1), Qwen CodeCURSOR_AGENT), Goose (AGENT=goose, some paths only), Aider (accidentalOR_APP_NAME)That maps cleanly onto yes / partial / no, and it matters to the same audience comparing
agentsMdSupport— it's the other direction of the same handshake (AGENTS.md = human tells agent; envIdentity = agent tells tools).Data offer
We maintain ai-agent-detect, which documents exactly this per agent with evidence tiers (vendor docs > source code > community reports > live probes) — see docs/harnesses. Happy to send a PR filling
envIdentityfor all agents in the matrix, each with a sourceddetail. Just say the word (and pick the property name/group you prefer).