Telemetry: Feedback collection

[nikomatsakis]

Gather qualitative signal from users about their experience. This workstream is more exploratory than the others.

Part of the telemetry initiative — see sync discussions: June 9, June 23.

Milestones

• cargo agents feedback command — Let users type free-form text feedback. Store locally alongside telemetry events (tagged as kind: "feedback"). Optionally attach to next upload batch.
• /symposium:feedback skill — A skill that agents can invoke so users can give feedback mid-conversation. Lower friction than switching to the terminal.
• Symposium MCP server (longer-term) — A managed MCP server installed by symposium. Use MCP elicitation to prompt for feedback. Caveat: Claude Code does not yet support MCP elicitation (as of June 2026).
• LLM-as-judge (research) — End-of-turn hook reads conversation transcript, invokes a subagent to extract quality metrics. Requires opting into API token spend — must disclose clearly. A/B testing comes later.

Notes

• Niko is skeptical of elicitation UX — keep MCP elicitation experimental.
• Alternative: insert context via hook nudging users to give feedback, or just ensure /symposium:feedback is well-known.
• LLM-as-judge could build atop determishtic or similar.