You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
MCP elicitation requests are auto-errored; mid-run questions from MCP servers never reach the user #173
MCP elicitation (a server asking the user for input mid-run) flows through print/SDK control requests in current claude (changelog v2.1.117). Solenta's runner handles only the can_use_tool control_request subtype; every other subtype gets respondError("Unsupported control request: ...") (electron/runner.js:1720-1726).
Impact
An elicitation from any MCP server, including Solenta's own injected memory/threads servers, is answered with an error the model must improvise around. The question never reaches the user and the turn degrades or fails with no visible reason. Distinct from #157 (pairing external clients): the server is already connected.
Fix direction
Route elicitation requests to the same awaiting-input UI used for permission prompts, and respond with the user's answer.
Round 20 intel: Copilot CLI now shows structured ask_user forms by default in interactive sessions, and Codex 0.147.0 cleans up cancelled MCP elicitation requests (#35836). Two signals for this issue: (1) the market-standard answer to mid-run questions is a structured form (typed fields, not free-text chat), which is a better UX target than a plain prompt card; (2) elicitation lifecycle hygiene matters: when a turn is interrupted or a server disconnects, pending elicitations must be cancelled and cleaned up, not left dangling or auto-errored into the transcript.
Round 23 intel: Factory Droid v0.183.0 added multi-select questions: when the agent asks a question, the user can pick several options at once. Combined with Copilot's structured ask_user forms (earlier comment), the shape of the standard is now clear: typed fields (single-select, multi-select, free text) rendered as a form, with keyboard navigation that does not fight the mouse (Droid v0.197.0 fixed option navigation jumping to wherever the mouse rests). Solenta's MCP elicitation UI should support multi-select from day one; agents asking "which of these 4 services should I restart?" is a real case where single-select forces N round trips.
Round 24 intel: the MCP 2026-07-28 revision (final since 2026-07-28) rewrote elicitation along with the rest of the protocol, and migration write-ups name the elicitation rewrite, not the session removal, as the part most likely to break existing implementations (dev.to migration notes). When Solenta implements this issue, target the new revision's elicitation semantics from the start rather than the 2025-11-25 shape; Codex CLI already supports the new revision opt-in, and Anthropic announced rollout across Claude products. Also note the spec now formalizes multi-round requests, which is exactly the mid-run-question flow this issue covers.
Activity
currentbits commentedon Aug 18, 2026
Round 20 intel: Copilot CLI now shows structured ask_user forms by default in interactive sessions, and Codex 0.147.0 cleans up cancelled MCP elicitation requests (#35836). Two signals for this issue: (1) the market-standard answer to mid-run questions is a structured form (typed fields, not free-text chat), which is a better UX target than a plain prompt card; (2) elicitation lifecycle hygiene matters: when a turn is interrupted or a server disconnects, pending elicitations must be cancelled and cleaned up, not left dangling or auto-errored into the transcript.
currentbits commentedon Aug 18, 2026
Round 23 intel: Factory Droid v0.183.0 added multi-select questions: when the agent asks a question, the user can pick several options at once. Combined with Copilot's structured ask_user forms (earlier comment), the shape of the standard is now clear: typed fields (single-select, multi-select, free text) rendered as a form, with keyboard navigation that does not fight the mouse (Droid v0.197.0 fixed option navigation jumping to wherever the mouse rests). Solenta's MCP elicitation UI should support multi-select from day one; agents asking "which of these 4 services should I restart?" is a real case where single-select forces N round trips.
currentbits commentedon Aug 18, 2026
Round 24 intel: the MCP 2026-07-28 revision (final since 2026-07-28) rewrote elicitation along with the rest of the protocol, and migration write-ups name the elicitation rewrite, not the session removal, as the part most likely to break existing implementations (dev.to migration notes). When Solenta implements this issue, target the new revision's elicitation semantics from the start rather than the 2025-11-25 shape; Codex CLI already supports the new revision opt-in, and Anthropic announced rollout across Claude products. Also note the spec now formalizes multi-round requests, which is exactly the mid-run-question flow this issue covers.
Route Claude MCP elicitation requests to the input prompt (#173)