Skip to content

tui.md

Latest commit

 

History

History
142 lines (111 loc) · 4.23 KB

tui.md

File metadata and controls

142 lines (111 loc) · 4.23 KB
Raw

🖥️ TUI & chat

Two terminal faces. One harness. Zero mystery meat. 🥩➡️🥗

⌨️

Pick your fighter: Premium TUI for day-to-day board + live events; Chat REPL for quick prompts and slash-command speedruns.

Face Best for
Premium TUI (slmcode / slmcode tui) Day-to-day — board, live events, agents, stats 🕹️
Chat REPL (slmcode chat) Quick prompts + slash commands 💬

Premium TUI ✨

slmcode
# or
slmcode tui

What you see 👀

  • Connection / provider health
  • Live kanban (columns move as waves complete — ASMR optional)
  • Agent stream: @worker, scope files, truncated outputs
  • Errors without the doom-scroll
  • Diffs and query history

Slash commands 🪄

Command Effect
/compact [heuristic|llm|auto] Shrink noisy context 🧹
/models Catalog / search enabled models 🔎
/mcp MCP server status 🔌
/auth Show / set provider keys (.slmcode/auth.json) 🔑
/schema Dump config JSON schema 📐
/sessions Browse saved runs 📼
/stats Phase latency + token estimates 📊
/permission Write / shell policy 🛡️
/plan [auto|ask] Plan-approve gate before execute 📋
/escalate … Answer escalate HITL: re_scope | retry | mark_done | abort ⚖️
/agents List / create / edit / delete agents 🧩
/stop Checkpoint mid-run 🛑
/resume Continue from checkpoint ▶️
/help Remind your future self 🛟

!!! tip "⚖ Escalate banner" When a task hits max review retries, the TUI shows an ESCALATE banner and the pipeline pauses that task. Answer with /escalate retry (etc.), or wait for the timeout — then @escalate (SLM) decides. Same modal exists in Studio, and slmcode run draws the same card inline and takes a single keystroke.

Before you answer, `slmcode task show <id>` in another terminal prints the scope, the
reviewer's verdict and issues, the gate that refused the task, and the diff of its focus
files. `mark_done` (`[d]one`) **overrides** the gate that refused the work: the task is
recorded as a human override, `slmcode board` marks it `⚑ forced done`, and the run summary
counts it separately from a verified pass.

!!! tip "💪 Keyboard muscle memory" Ctrl+C mid-run checkpoints board + ReAct history under .slmcode/queries/<id>/react/. Prefer /resume over “start over and pray”.

Agent CRUD from the TUI 🧩

/agent new
/agent new id=night-auditor title=Night provider=ollama model=qwen2.5-coder:14b
/agent edit worker model=qwen2.5-coder:14b
/agent show night-auditor
/agent delete night-auditor

Chat REPL 💬

slmcode chat
slm › /help
slm › /permission review
slm › Add validation to the login handler
slm › /board
slm › /diff
slm › /quit
Command Effect
/run <q> Explicit full pipeline
/board /status /diff /skills /doctor Inspect
/permission auto|dry-run|review Write policy
/model <id> Switch model (rebuild as needed)
/quit Exit (politely)

Plain lines (no /) also run the full pipeline. Yes, even typos. Choose carefully. 😅

Classic one-shot CLI 🛠️

Still the power tool:

slmcode run -v "…"
slmcode board
slmcode watch
slmcode session list

See ⌨️ CLI reference for every subcommand.

When to use which 🎯

Situation Reach for
Exploring a new repo TUI + Studio side-by-side 🥊
Scripted / CI-ish slmcode run + flags
“Just fix this” chat or run -v
Teaching / demos permission dry-run + TUI 🎭
🍿

Pro move: run slmcode studio in one pane and slmcode watch in another. Feel like mission control. Resist the urge to narrate aloud. Or don’t — we can’t hear you.

☀️ Made with ♥ by UnicoLab