which-agent

which(1) for AI agents — resolve which agent CLI would run a task, and how to invoke it headlessly.

Status: name claimed, v0 in development. This package does not do anything useful yet.

What it will be

Programs that shell out to "the user's AI agent" today hardcode one (claude -p). which-agent resolves the runner the way which resolves a binary:

• Resolution ladder: explicit config → environment override → running-agent hint → detected installed agents → not found.
• Headless invocation adapters: the per-agent spelling of prompt, model, permissions, and output flags (claude -p / codex exec / gemini -p / amp -x / …), including a normalized permission knob (readonly | edit | full) expanded to each agent's dialect.
• which semantics: prints the resolution and exits 0; silent exit 1 when no agent is found — script-friendly from day one.

# planned
which-agent            # claude-code 2.1.205 → claude -p
if which-agent >/dev/null; then …; fi

Relationship to ai-agent-detect

ai-agent-detect answers the inbound question — who is running me. which-agent answers the outbound one — who would run this. It may read the former as one rung of its ladder; neither contains the other.