[Maxaubert]

Summary

Closes #323. Sonara side of the approved plan docs/plans/2026-10-16-question-follow.md; the PrismTerminal side follows in its repo.

Claude Code tells plugins nothing while the user moves between the questions of one AskUserQuestion. A terminal that draws the widget can: the new agent message question_shown (protocol 1.12) carries the question on screen, and Sonara reads it.

• {channel | host_tab, text} or {..., review: true}. host_tab is the terminal tab the hook already reports (PRISM_TAB_ID), so PrismTerminal only needs its own tab id; the newest session of that tab with a match is used.
• Matching: case and anything but letters and digits ignored; a screen cut short matches the question it starts (20 characters on).
• With question_step on, the set lands on that question (as next_question); off, the reading goes on from that question's entry. The current question, an unknown text and a set already reviewed do nothing. The review screen says "Review your answers." once per set.
• Reply: moved, plus channel, question, of (or review).
• Both SDKs: questionShown (TS AgentApi) and question_shown (Python).

Tests

• Rules: stepping set, read-in-order set, wrapped and cut-off texts, review once.
• Agent: read when stepping (forward, back, review), jump in a set read in order.
• Conformance: test_question_shown_follows_the_question_on_screen, test_question_shown_finds_the_session_by_its_terminal_tab. SDK namespace tests (TS 39/39, Python).
• python packaging/gate.py --keep-going: all pass except the tests/embed curl example (local only).

Docs: protocol-v1 (message, 1.12), README (Prism follows the questions). Version 0.33.0. Releases are unsigned.

🤖 Generated with Claude Code

https://claude.ai/code/session_014cnwwWNiEmM5ghBD933h59

[Maxaubert]

Review fixes (2026-10-16 full code review), commit b4b82c0:

• Daemon lifecycle: real shutdown/start commands, install stop-then-swap, uninstall ordering #23 question_shown spoke while muted: with mute_level >= 1 nothing is read (the set still follows the screen, the entry is still revised); "Review your answers." goes through Rules::speak, so it is stored while muted. Test: a_question_shown_on_screen_is_not_read_while_muted (stepping, review, in order).
• Control cues wait out cold Chatterbox reloads; speak them via the instant Windows voice #60 host_tab fell through to older sessions of the tab: it now resolves to the newest session of the tab with a question set (stepping or read in order) and stops there. Test: host_tab_stops_at_the_newest_session_of_the_tab_with_a_question_set (sonarad).
• Instant Windows-voice control cues (#60) #61 fold kept word boundaries: matching now folds to letters and digits with no separators, so a long token cut mid-word on screen matches; the 20-character prefix minimum stays. Test: a_long_token_wrapped_mid_word_on_screen_matches_its_question.
• Unverified point 1, confirmed real: a set read in order restarted the question already being read when the screen showed it. Now it goes on. Test: a_question_shown_on_screen_does_not_restart_the_question_being_read.
• Unverified point 2, confirmed real: host_tab only came with channel_open (SessionStart, UserPromptSubmit), so a runtime started mid-turn had the channel without its tab. In a host tab the hook now sends channel_open (with host_tab, keep_label) before an AskUserQuestion's questions. New golden case ask_user_question_in_a_host_tab_reopens_the_channel.json (a deliberate mapping change; existing goldens unchanged).
• docs/protocol-v1.md question_shown row updated.

Tests: each new test failed first, then passed. cargo test -p sonara-agent / sonara-hook / sonarad all green, clippy clean, gate --quick all ok except the known local-only tests/embed curl example. No version bump.

https://claude.ai/code/session_014cnwwWNiEmM5ghBD933h59