Session 38 surfaced real friction: every narration beat, feed post, and DM lookup during live play required a synchronous round-trip in chat, and the table waited on Claude mid-scene. Explore using background agents/tasks more during live sessions (e.g. a standing agent that drafts 2-3 likely next-beat narrations ahead of the table reaching them, so approval is a quick pick-and-post instead of a cold draft), and/or dedicated slash commands for the highest-frequency live actions (draft+post narration, pull a quick rules lookup, sync HP) so they’re faster and more consistent than free-form prompting. Needs a session-mode-specific pass, not just general Claude Code speed – the constraint is turnaround at the table, not model quality. Update (/txt-session-retro, 2026-07-05): part of the mechanical overhead is already fixed – txt-narrate.md/txt-session-open.md now call the pre-existing bin/scribe-event script instead of hand-rolling scribe_id/JSON, which removes plumbing time within a turn. The structural issue (every beat still needs a full agent turn) remains open. Update (/txt-session-retro, 2026-08-21, Session 39): bin/scribe-event/push-loop fixes held cleanly all night – 42/42 narration events synced, zero delivery failures, push log stayed quiet between real pushes. The structural per-beat-turn issue is still the open part; nothing new to add there this session. Update (2026-09-03): two more fast-path commands landed. ADR-0087 made narration/dialogue approval drafts a 4-option pick instead of a single-draft negotiation (cuts the back-and-forth on ambiguous NPC lines, not just the plumbing). ADR-0088 added /txt-turn, wrapping the Stage encounter lifecycle (start/next/end) with a session-cached stage id, so turn advancement joins HP and rules lookups as a single fast call instead of a free-form MCP round-trip. Update (2026-09-03, ADR-0090): /txt-scribe now also auto-applies condition events to the active Stage token (reusing /txt-turn’s stage cache), same two-actions-in-one benefit HP auto-apply already had – closes a gap where a condition was logged to the Scroll/Feed but never shown on the Encounter Tracker/TV view players see. Research note (2026-09-03, docs/session-running-improvements.md): evaluated the standing predictive-drafting-agent half of this item and do NOT recommend building it as originally scoped – the motivating example (Alora’s giants-item barter) turned on a detail generated live in an improvised conversation that no speculative pre-drafting agent could have anticipated; most speculative drafts for improvised moments would be wasted generation. Spun the smaller, more tractable “NPC voice sheet” idea out to its own item (BL-116) rather than pursuing the standing agent. The structural per-beat-turn issue (every narration beat still needs a full agent turn) remains the open core of this item, and per this research note, is likely to stay open – fast commands and cheaper approval drafts are the realistic levers, not prediction.
BL-099
Next Up