[Rchari1]

Problem

The Amico pulse-designer interview asks each question via the native question tool, which renders the clickable answer card. Models intermittently drift and ask in plain prose instead — e.g.:

"What's your qubit frequency ω (GHz) and anharmonicity δ (GHz), and do you want 3 or 4 levels?"

That renders no card, so the user has nothing to click and the interview stalls. Reported repeatedly. Prompt hardening in AGENTS.md alone did not hold — LLM tool-use adherence is probabilistic — so this enforces the contract in the turn loop.

Fix

In the session prompt loop, when a turn is about to exit (assistant finished, no tool calls) and:

• an Amico interview is active (the session has used any amicode_* tool), and
• this turn made no question tool call, and
• the assistant's final text ends with ? (a question to the user),

…inject a synthetic system nudge ("re-ask via the question tool") and continue the loop instead of exiting — forcing the model to re-emit the question as a card.

Idempotent: an in-memory Set of nudged assistant-message IDs fires the nudge at most once per message (the Assistant info schema has no metadata field to persist a flag on), so a stubborn turn cannot loop forever — after one nudge it exits normally.

Scope: only fires inside an active amicode interview. Ordinary agent turns that happen to end in a question (e.g. a build agent asking for clarification) are unaffected — they haven't used amicode_* tools.

Testing

• tsgo --noEmit: clean
• Built the darwin-arm64 binary, swapped into the amicode dev host, verified the interview re-asks via the tool (card renders) after a prose drift.

Follow-up

Once merged, cut a fork release so the amicode extension's vendored binary (opencode.lock.json) can bump to it — the fix ships to users via that pin, not this branch directly.