Converse

Sticky conversation listen-back: once on, every finished assistant reply in this session is spoken aloud (rewrite → TTS → play), until the user turns it off. No per-message /speak.

Uses the same engine as /speak (scripts/speak.ts, Grok Voice by default). Does not replace a dictation tool for voice input, it only automates voice output.

Usage

/converse              # turn ON (default)
/converse on           # turn ON
/converse off          # turn OFF
/converse status       # report whether mode is on

Natural-language equivalents (treat the same as the slash forms):

• ON: "start converse", "converse on", "auto speak from now on", "speak every reply"
• OFF: "stop converse", "converse off", "stop auto speak", "stop speaking every reply"

State file

Persist mode in the repo root's scratch/ working folder (keep scratch/ gitignored):

• Path: scratch/converse-mode
• ON contents: a single line on (optionally a second line with ISO timestamp)
• OFF / absent: mode is off

Write or clear this file whenever you change mode so status and later turns can check it.

Instructions

1. Parse intent

If the user runs /converse with no args → on.

2. Turn ON

1. Write scratch/converse-mode with contents on (and optional timestamp line).
2. Reply with a short confirmation, e.g. 🔊 Converse on, every reply will be spoken until you say /converse off.
3. Run the speak pipeline (below) on that confirmation so they hear that mode started.
4. Stay sticky for the rest of this conversation (see § Sticky duty).

3. Turn OFF

1. Delete scratch/converse-mode, or write a single line off.
2. Reply with a short confirmation, e.g. Converse off, replies are silent again. Use /speak when you want one-off audio.
3. Do not speak that confirmation (mode is already off).

4. Status

Read scratch/converse-mode. If missing or not starting with on → off. Report in one line. No TTS.

5. Sticky duty (the whole point)

While converse is on (state file is on, or you turned it on earlier this conversation and have not turned it off):

For every subsequent assistant turn that produces a real reply to the user:

1. Produce the normal visible reply first (full substance, do not collapse to only 🔊 like bare /speak).
2. Then run the speak pipeline on that reply (rewrite + play).
3. You may append a trailing 🔊 after the visible reply; optional.

Do not wait for the user to re-run /converse or /speak. Re-check the state file if unsure; if it says on, keep speaking every reply.

Skip the speak pipeline only when:

• Mode is off
• The turn is tool-only with no user-facing message
• The user is mid-turn only approving a dangerous action and you are asking a one-line confirm (still prefer speaking short confirms when mode is on)
• TTS already failed twice in a row this session, note once, keep replying in text, do not spam retries every turn

6. Speak pipeline (same rules as /speak)

Rewrite, do not read the reply verbatim:

• Plain, conversational tone, sharp friend over coffee. Warm, casual, not verbose.
• Gist and why in simple terms. Drop jargon, or define it in a few words if it matters.
• No code, no file paths, no markdown, no headings, no bullet lists read aloud, natural sentences only.
• Tight: ~20–60 seconds (~60–180 words). Key points only. Hard cap ~4000 characters.
• If the reply was already short, smooth it for speech; do not pad.

Then:

1. Write the rewritten explainer to scratch/speak-input.txt (overwrite).
2. Run: npx tsx scripts/speak.ts scratch/speak-input.txt
3. If the script errors, add a one-line note under the visible reply. Do not dump the rewritten explainer text unless the user asks.

Unlike bare /speak, the visible chat reply stays the full answer. Audio is an add-on.

7. Interaction with /speak

• /speak while converse is on: still works as a one-shot of whatever source /speak uses; then continue auto-speaking future replies.
• /speak while converse is off: unchanged (one-shot only).
• Prefer /converse off over telling the user to "ignore speak" when they want silence.

Voice setup

Same as /speak, see .claude/skills/speak/SKILL.md (or its .agents/skills/ mirror):

• SPEAK_PROVIDER, grok (default) or hume
• SPEAK_GROK_VOICE, eve | ara | rex | sal | leo (default)
• Hume: SPEAK_VOICE_ID via scripts/speak-setup.ts

What this is not

• Not hands-free mic input (dictation / push-to-talk / always-listen tooling stays separate).
• Not multi-session routing.
• Not a permanent global preference across new sessions, each session starts off unless the user turns converse on again (stale scratch/converse-mode from a prior session: if the file says on at the start of a new conversation, treat it as off until the user explicitly re-enables, and clear or rewrite the file when they choose).