Skip to content

guide.md

Latest commit

 

History

History
179 lines (134 loc) · 4.49 KB

guide.md

File metadata and controls

179 lines (134 loc) · 4.49 KB
Raw

🧭 User guide

Daily-driving SLMCode without turning your repo into modern art. 🎨 (Unless modern art was the goal. In that case… carry on.)

🛤️

Recommended path: Install → Quick start → Concepts → you are here. Welcome to adulthood.

Day-to-day loop 🔁

cd your-project
slmcode init
# optional: AGENTS.md at repo root (be stern, be kind)
# edit .slmcode/PROJECT.md

slmcode                      # premium TUI
slmcode run -v "Add validation to the login handler"
slmcode board
slmcode studio
Habit Command Vibe
🔎 Explore only run --agent explorer "…" Maps, not edits
🎯 Pin craft run --skill atomic-coding "…" Tiny diffs or bust
👀 Watch board watch ASMR for kanban
🛡️ Safe writes config set permission review → apply Trust, but verify

More copy-paste → 🧪 Recipes.

Pipeline (what happens) 🏭

flowchart TD
  Q[Query 🎯] --> I[Instructions + skills]
  I --> C[Context]
  C --> E{Explore?}
  E -->|reuse ♻️| P[Plan / split]
  E -->|deep 🕵️| X[Explorer] --> P
  P --> Coord[Coordinator 🧭]
  Coord --> W[Workers 🛠️]
  W --> R[Review ↔ correct 🔍]
  R --> L[Learn / test / skills 🦋]
Loading

Force a fresh dig: SLMCODE_FORCE_EXPLORE=1 slmcode run -v "…".

Deep dive → 🧠 Concepts.

Surfaces 🎛️

Surface Doc Mood
Premium TUI + chat 🖥️ TUI & chat Keyboard wizard
Studio GUI + API 🎨 Studio Click enjoyer
Full CLI ⌨️ CLI reference Script goblin

Skills & specialists 🧩

Skills teach house style; specialists execute roles.

slmcode skills
slmcode run --skill multipass-quality "…"
slmcode run "Fix login @skill:atomic-coding"

→ 🦋 Skills · 🧩 Agents

Project memory (the real DB) 💾

Markdown. On disk. Boring on purpose. Future-you can grep it. Past-you cannot gaslight it.

File Role
.slmcode/PROJECT.md Durable facts
.slmcode/CONTEXT.md Working focus + discoveries
.slmcode/MEMORY.md Lessons / pitfalls
.slmcode/SKILLS.md Index + recent lessons
.slmcode/skills/learned/ Auto-grown conventions
.slmcode/sessions/ Resumable runs
.slmcode/pending/ Staged review writes
AGENTS.md / CLAUDE.md / .cursorrules Auto-injected instructions 
slmcode context append "Prefer table-driven tests"
slmcode docs show MEMORY.md

Permissions 🛡️

Mode Behavior
auto Write now ✍️
dry-run Simulate 🎭
review Stage → slmcode apply (interactive) / slmcode reject 👀 
slmcode config set permission review
slmcode run "refactor foo"
slmcode apply && slmcode diff   # `apply` is interactive; use --all in scripts

Shell: shell_permission: allow | ask | deny — independent of file writes. → ⚙️ Config

Quality knobs 🎛️

slmcode config set think_passes 2
slmcode config set retries 2
slmcode config set parallel 2
# The pack budget comes from the model's real window, not from bytes.
# model_profiles is a structured field — edit .slmcode/config.yaml:
#   model_profiles:
#     qwen2.5-coder:
#       context_limit: 32768
slmcode config get model_profiles
Symptom Try
🥴 Wanders Lower context, pin atomic-coding, stern AGENTS.md
🥴 Weak JSON More think passes / retries; better tool-calling model
🐢 Too slow Lower parallel; explore reuse; faster provider

→ ❓ FAQ

Sessions & resume 🛟

slmcode session list
slmcode session resume run-…
# TUI: /stop then /resume

Ctrl+C mid-run is a feature, not a crime — it checkpoints.

Develop from a checkout 🛠️

make lint && make test && make docs-build
RUN_E2E=1 make e2e
🧃

Hydration tip: if a run feels cursed, check doctor, then permissions, then whether you asked for a rewrite of the universe. Usually it’s #3.

☀️ Made with ♥ by UnicoLab