Handoff agent: README with the stack, plugins by ecosystem, and measured performance

[yuanhao]

A README for the handoff agent example, integrations/yoagent-frontend/examples/README.md, linked from the frontend README:

• The idea: questions that wait, a gate for the owner's files, a checked report plus host facts, a page that replays.
• The stack: one Rust process (yoagent loop, yoagent-rutis bridge, yoagent-frontend session and web server) and three rutis Node runtimes (ui, pi, dsh), with a diagram.
• Plugins by ecosystem: pi's @juicesharp/rpiv-todo and DSH's dsh-free-search (with DSH's core packages) are used unchanged from npm. Ours are the pi-format approval gate, the UI plugin, the adapters, and the Rust tool and extension.
• How a tool call travels through the bridge and the pi policy.
• Measured performance and memory, and the cost model.

The example now prints plugins ready in … at start, and in --demo the time of each tool call, so anyone can re-measure.

Measured (2026-10-11, Apple M2 16 GB, release build, DeepSeek)

Startup to plugins ready	1.7–2.3 s with 3 Node runtimes; 1.5–1.7 s with --no-dsh; ~9 s on the first run after boot
A tool call through the loop, including the trip to the pi policy in Node	write_file 0.7 ms · todo 1.3–6 ms · ask_owner 2–2.5 ms · list_files ~10 ms (spawns find)
Live research run	30.5 s wall: model 15.7 s over 6 requests, web searches 2.7–6.7 s each, the agent's own tool calls 1.7–3.6 ms
Memory	Rust process 21 MB; each Node runtime 137–244 MB

The loop's cost is milliseconds per call, against seconds per model request or web search: about 0.1% of that run. Most memory is the Node runtimes, which are the cost of running other ecosystems' plugins unchanged.

🤖 Generated with Claude Code

https://claude.ai/code/session_01T7iq5hpndSiHQcnAsywKuG