The conversational mediary for your tmux panes ("yes sir"). One conversation covers everything running in your terminal: read, scroll back through, and operate any TUI in a pane; keep a scratchpad of tasks, todos, and schedules; run and watch commands; produce diagrams and detailed reports on request; and drive multi-step work toward completion — without ever leaving the conversation or looking at a screen.
yscr sits between you and the programs in your panes (Claude Code CLIs, shells, anything with scrollback) and translates both directions:
- Coworker voice, medium-aware. Reports high-level by default; detail only when asked, or when it matters — a KPI moved, was added, or is at risk. Verbosity is tuned to the medium: text can be dense; audio must be sparse (the audio layer lands later, but the contract assumes it now).
- Reads and operates. Pulls live output and scrollback on demand; answers prompts, submits inputs, steers sessions — through conversation.
- Manages horizon work. Keeps a durable scratchpad — tasks, todos, schedules/cron — and runs terminal commands in the foreground or background, watching them and reporting what's material.
- Plans toward goals. Decomposes a goal into a plan with dependencies, works everything that isn't blocked, and parks ambiguity in an explicit queue of open questions for you — never stalling on what it could be doing anyway.
- Shows its work. On request, produces diagrams (dependency graphs, session maps, flows) and detailed written reports from the state it already sees — the same data that drives the high-level voice, rendered deeply when you want to look at it.
- Remembers your decisions. Records the choices you make (with a log); later questions whose answer is already obvious from those preferences are resolved automatically, citing the decision they came from.
One line: yscr makes your tmux panes conversational — a coworker who sees every pane, talks to the medium it's in, keeps the work-list and schedule for you, plans around ambiguity, draws what it's working on when asked, and remembers what you've already decided.
The concierge is an agentkit session on corrallm (the LLM endpoint; corrallm's own downstream model adapters are its business), with audio via oidio (STT/TTS).
It drives every pane through one plugin contract — source.Source — with a
program Adapter per supported application:
| adapter | observe | spawn | act |
|---|---|---|---|
| claude | live pane + JSONL transcript tail | new tmux session | answer questions (verified keystroke protocol) |
| terminal | scrollback + pipe-pane stream | shell window (run commands) | — |
A new supported program = a new Adapter, no new tmux/source code.
The crux — source.Questionnaire: structured input requests from a
program (Claude's AskUserQuestion, any schema-shaped prompt) are rendered
conversationally by the handler model and parsed back into a
schema-validated structured answer — the user faces a conversation, never a
form.
Shipped: the source.Source plugin contract + program adapters (claude,
terminal) with live-pane adoption, question detect/answer (hook-primary,
pane-parse fallback), Watch (SSE live-tail) and Narrate (LLM delta summary →
TTS); the concierge on agentkit with a serialized, coalescing per-session
dispatch queue; scratchpad (tasks/todos/schedules + run & watch commands);
goal plans (dependency graphs over the cue release gate) + open-questions
queue; decision memory (log + auto-resolve); diagrams (SVG) and persisted
reports on request; medium-aware verbosity; ambient narration with a
materiality gate; installable PWA (chat, session cards, detail sheet, tap-to-
answer "Needs you", work list, reports, streaming STT via PCM worklet, web
push); Postgres durable store; systemd supervision.
On the way (see plan/plan.md): streaming STT live browser drive; LLM
near-match decision proposals; more diagram kinds.
⚠️ Early project in places. Seeplan/plan.mdfor current state, active work, and decisions.