Reference /docs/cheat-sheet

Cheat Sheet

Yoke (Yoke) — Cheat Sheet & Tactics Guide

A comprehensive quick-reference guide for Yoke ( yoke ) , covering CLI flags, REPL slash commands, prompt syntax ( @ , ! ), rule engine scoping, branch code reviews, local transcripts, and execution tactics. For a complete walkthrough of onboarding and customization, see the Getting Started Guide .


Table of Contents

  1. CLI Invocation & Command Line Flags
  2. Context Expansion ( @ ) & Shell Execution ( ! )
  3. Slash Commands Reference
  4. Rule Engine & Scoping Tactics ( /rules )
  5. Code Review & Git Workflow ( /cr , /diff , /commit )
  6. Session Management & Persistence ( -c , /resume , Transcripts)
  7. Model Selection Strategy ( /model )
  8. MCP & Ragex Code Analysis ( /mcp , /ragex )
  9. Subagents & Distributed Execution ( /subagent , /mode , /nodes )
  10. Customizable Workflows ( /workflow ) 11. Keyboard Shortcuts & Navigation

CLI Invocation & Command Line Flags

Launch yoke from any directory in interactive REPL mode, one-shot mode, or resume mode:

# Launch interactive REPL mode (generates new UUID session)
yoke

# Resume specific conversation ID across CLI restarts
yoke -c df97eb34-cb33-4f21-bada-2e9c3cf75d46
yoke --conversation=df97eb34-cb33-4f21-bada-2e9c3cf75d46

# One-shot mode with inline prompt
yoke "Analyze project architecture in @mix.exs"

# Specify model alias for one-shot execution
yoke -m deepseek-coder "Write a binary search algorithm in @lib/search.ex"
yoke --model deepseek-reasoner "Diagnose race condition in @lib/worker.ex"

# Run in background self-update mode
yoke --update

# Display CLI help menu
yoke --help

Supported Flags & Aliases

Flag Short Description



--conversation -c Resume a persisted conversation ID
--resume -r Alias for --conversation
--prompt -p Input prompt string for one-shot execution
--model -m Select model ( deepseek-chat , deepseek-coder , deepseek-reasoner )
--update -u Trigger background release update to latest git codebase
--node — Start local Erlang distributed node
--connect — Connect to remote Erlang Hands worker node
--plugin — Load external Elixir plugin script ( .ex or .exs )

Context Expansion ( @ ) & Shell Execution ( ! )

@ File & Context Reference Syntax

Type @ anywhere in user prompts to attach contents:

  • File references : @mix.exs , @lib/yoke/cli/line_editor.ex
  • URI references : @https://raw.githubusercontent.com/...
  • Interactive File Picker : Typing @ in REPL triggers an interactive fuzzy file picker modal. It respects .gitignore rules ( git ls-files ) and strictly excludes _build/ , deps/ , .elixir_ls/ , and .git/ .

! Shell Command Execution Syntax

Prefix any line with ! to execute a shell command directly without sending it to DeepSeek:

📁 my_app 🤖 deepseek-chat ❯ !git status
📁 my_app 🤖 deepseek-chat ❯ !mix test
📁 my_app 🤖 deepseek-chat ❯ !cargo check

!! Pure Console Mode (Flip-Flop)

Type !! on its own to flip yoke into pure console mode : a bare shell passthrough with no Brain/Hands actors, no LLM turns, and no slash-command dispatch in between. Every line you type executes directly via sh -c , with output streamed live and cd applied to yoke 's own working directory (like a real shell builtin). Type !! again (or Ctrl+D ) to flip back into the normal harness REPL, picking up right where you left off -- no need to spawn a second terminal tab just to run a few plain commands:

my_app deepseek-chat > !!
Flipped into pure console mode -- plain shell passthrough, no AI/tooling in between. Type !! again to return to Yoke.
console my_app $ cd ..
console Proyectos $ ls
console Proyectos $ !!
Back to Yoke -- pure console mode OFF.
my_app deepseek-chat >

Slash Commands Reference

Slash Command Description Example



!! Flip into/out of pure console mode (plain shell passthrough, no AI/tooling) !!
/review [ [head] \| ] Human-in-the-loop PR review with inline triage ( [i] Include , [e] Edit , [d] Drop ) /review 142
/prboard | /review queue Live Pull Request review queue board (CI, approvals, status) /prboard
/review next Auto-advance and review next queued PR /review next
/review skip | unskip Commit-pinned PR skip (hidden until author pushes new commits) /review skip 142
/review patterns View living codebase review patterns ( .yoke/review_patterns.md ) /review patterns
/cr [base] Shortcut to review active branch against main or custom base /cr main
/diff [target] View colorized git diff of workspace or against target branch /diff main
/resume [id] Resume specific session ID or open interactive conversation picker modal /resume
/rules [add\|delete] List, add, or open checkbox modal to delete prompt rules /rules add cr: format tables 80 cols
/model Switch model ( chat , coder , reasoner , openrouter-r1 , ollama-r1 ) /model reasoner
/mode Set execution target ( local , docker , remote ) /mode local
/compact Compress conversation context window to save tokens /compact
/undo Roll back conversation state to previous temporal checkpoint /undo
/checkpoint [lbl] Create a manual temporal state snapshot /checkpoint "Before refactoring"
/config Manage prompt styles ( starship , compact ) and UI settings /config style starship
/plugins [reload] List tools or hot-reload custom plugins live without resetting state /plugins reload
/mcp [list\|add] Mount or inspect Model Context Protocol (MCP) servers /mcp list
/ragex Mount Ragex MCP server for AST refactoring & SCIP semantic search /ragex
/skills [name] List available skills or execute a skill instruction file /skills
/update Refresh production OTP release in background ( yoke --update ) /update
/commit Auto-stage all modified workspace files and create git commit /commit "feat: add user login"
/cost | /tokens Display token usage breakdown and cumulative session cost /cost
/permissions Set tool execution safety mode ( auto or ask ) /permissions ask
/subagent Spawn a background subagent worker for heavy sub-tasks /subagent "Search all TODOs"
/workflow [cmd] Run customizable multi-step workflows: list\|run\|status\|resume\|abort\|init /workflow run elixir Add dark mode
/cb | /clipboard Copy latest assistant response to system clipboard /cb
/clear Clear terminal output screen /clear
/reset Reset conversation context, history, checkpoints, and clear screen /reset
/help Print REPL help menu /help
/exit | /quit Exit Yoke and print conversation resume banner /exit

Keyboard Shortcuts & Hotkeys

  • Ctrl+P : Toggle tool execution safety mode ( ask_confirm ⇄ auto_approve )
  • Ctrl+G : Toggle workspace sandbox bounds on/off
  • Ctrl+B : Toggle idle status bar mode (gauge ⇄ compact session line)
  • Ctrl+O : Toggle tool call log expansion mode (collapsed/truncated ⇄ expanded/full text)
  • Ctrl+Q : Interrupt active AI turn while responding

Rule Engine & Scoping Tactics ( /rules )

The Rule Engine allows defining persistent prompt preambles and formatting constraints saved in .yoke/rules.json .

Rule Scope Syntax

  • all: — Injected into system preamble for every prompt turn.
  • : — Injected only when that specific command is executed.

Pre-seeded Default Rules

  1. all: typographic quotes “” mean the exact quote
  1. all: backticks mean a code quote
  1. cr: format table cells multiline to fit in 80 symbols width

Rule Commands

  • List Rules : /rules
  • Add Rule : /rules add all: Always prefer pattern matching over if/else
  • Add Command Rule : /rules add cr: Highlight security vulnerabilities first
  • Delete Rules (Interactive TUI Modal) : /rules delete or /rules rm (Opens a checkbox modal with OK/Cancel buttons to delete selected rules)
  • Toggle Rule : /rules toggle

Code Review & Git Workflow ( /review , /prboard , /cr , /diff , /commit )

Human-in-the-Loop PR Code Review ( /review , /cr )

Runs a comprehensive AI code review comparing branches or evaluating a GitHub Pull Request:

# Review active branch against default base 'main'
/cr
/review main HEAD

# Review specific GitHub PR by number
/review 142

# Inline review help & standards
/review help

Key Review Capabilities :

  • Granular Triage : Review findings one by one with [i] Include , [e] Edit (modify comment inline), or [d] Drop .
  • Atomic GitHub Review Posting : Submits review events ( APPROVE , COMMENT , REQUEST_CHANGES ) alongside native right-side inline diff comments and committable suggestion ` blocks via `gh api`. - **10-Point Self-Critique Checklist**: Discards suggestions if gated by feature flags/supervisors, speculative, bikeshedding, or already flagged by linters/CI. - **Typographic Standards**: Enforces real em dashes (`—`) for parenthetical clauses/breaks and proper typographic quotes (`“ ”` and `‘ ’`) in prose explanations. ### Live PR Review Queue Board (`/prboard`, `/review queue`) Manage and triage your repository's open pull requests without switching contexts: bash# Open live review queue board /prboard /review queue # Step through the queue automatically /review next # Commit-pinned skip (hides PR until author pushes new commits) /review skip 142 /review unskip 142 Categorizes open PRs into **Visible** (ranked by readiness: `re-review` > `new`) and **Hidden** (`draft`, `yours`, `approved`, `waiting on author`, `skipped`). ### Living Codebase Review Memory (`/review patterns`) Maintains an evolving codebase memory in `.yoke/review_patterns.md`: bash# View active review patterns and tripwires /review patterns - Real accepted comments are recorded to `.yoke/review_patterns.md` after each review. - Patterns with `[Seen: 3+]` are automatically enforced as **mandatory tripwire checks** on future reviews. ### Workspace & Branch Diffing (`/diff`) bash# View colorized diff of unstaged/staged workspace changes /diff # View colorized diff against target branch or commit /diff main /diff origin/main ### Auto-Commit (`/commit`) Stages all workspace modifications and creates a git commit: bash/commit "fix(cli): resolve terminal cursor offset on emoji rendering" --- ## Session Management & Persistence (`-c`, `/resume`, Transcripts) ### Session State Persistence All REPL and one-shot sessions automatically save their GenServer state, snapshots, and conversation messages to `.yoke/sessions/.lmml` -- an `lmml` narrative (a Markdown-superset markup for LLM conversations). The file is plain, self-contained Markdown that any Markdown viewer renders sensibly, yet round-trips every structured message losslessly through its inline-embed model. Legacy `.json` session files from earlier versions are still read transparently on resume. ### Exit Resume Banner Upon exit via `/exit`, `/quit`, or `Ctrl+D`, `yoke` prints your exact conversation UUID: Resume with -c (or command below): yoke --conversation=df97eb34-cb33-4f21-bada-2e9c3cf75d46 ### Local Transcripts Log Files Every session maintains dual JSONL transcript logs in `.yoke/sessions//`: - `transcript_full.jsonl`: Complete, untruncated step-by-step logs of all user prompts, tool calls, and model responses. - `transcript_compact.jsonl`: Token-efficient version with large binary/text outputs truncated for rapid grepping. ### Ambiguous Context Expansion & Issue Tracking `yoke` automatically detects ambiguous context references in prompts (such as *"error above"*, *"the build failure"*, or `@error`): - Resolves previous tool execution tracebacks from session history. - Maintains an internal Issue Tracker categorizing issues as `:open` or `:resolved`. - Automatically marks issues as `:resolved` when subsequent tool calls succeed. --- ## Model & Endpoint Selection Strategy (`/model`, `/endpoint`) Switch model architectures or LLM providers on the fly during a session: bash/model chat # Switch to DeepSeek-V3 (Default agentic model) /model coder # Switch to DeepSeek-Coder-V2.5 (Idiomatic code writing) /model reasoner # Switch to DeepSeek-R1 (Deep chain-of-thought debugging) # OpenRouter & Local Model shortcuts /endpoint openrouter # Switch endpoint to OpenRouter API /model openrouter-free # Switch to meta-llama/llama-3.3-70b-instruct:free /endpoint ollama # Switch endpoint to local Ollama (http://localhost:11434) /model ollama-qwen # Switch to local qwen2.5-coder:14b | Scenario | Recommended Model / Provider | Rationale | | --- | --- | --- | | **Multi-turn file editing & bash commands** | `deepseek-chat` | Highest tool call precision and execution reliability | | **Writing complex algorithms & tests** | `deepseek-coder` | Trained on 338+ languages; idiomatic syntax generation | | **Diagnosing stack traces & deadlocks** | `deepseek-reasoner` | Streams live `[DeepSeek-R1 Reasoning]` thoughts before execution | | **Open-source & Free AI Models** | OpenRouter (`/endpoint openrouter`) | Free tier access to Llama 3.3, Qwen 2.5 Coder, Gemini 2.0 | | **Offline & Air-gapped Execution** | Local Ollama / LM Studio (`/endpoint ollama`) | Complete privacy, offline tool execution without API key requirements | --- ## MCP & Ragex Code Analysis (`/mcp`, `/ragex`) ### Model Context Protocol (MCP) Management bash# List all connected MCP servers and tools /mcp list # Connect a stdio MCP server /mcp add sqlite npx -y @modelcontextprotocol/server-sqlite --db-path ./app.db ### Native Ragex Integration (`/ragex`) Mounts the **Ragex** code analysis MCP server targeting the workspace: - SCIP Code Indexing & Graph Queries - AST Pattern Searching (`metaast_search`) - Refactoring Effort Estimation & Security Auditing - Graph Stats (`/ragex stats`) --- ## Subagents & Distributed Execution (`/subagent`, `/mode`, `/nodes`) ### Background Subagents (`/subagent`) Delegate non-blocking sub-tasks to concurrent subagent processes running in separate BEAM GenServers: bash/subagent "Analyze all TODO comments across lib/ and generate a report" ### Execution Target Modes (`/mode`) Control where Hands tools execute: bash/mode local # Local host filesystem (Default) /mode docker container_id # Execute bash/tools inside running Docker container /mode remote hands@host # Execute tools on remote Erlang node --- ## Customizable Workflows (`/workflow`) Runs a named, customizable, multi-step process on top of the ordinary agent loop -- branch, describe the task, propose a non-clashing parallel split, require tests + docs, lint, and commit -- with the entire run persisted under `.yoke/workflows/`. Full reference: [`docs/WORKFLOW_ENGINE.md`](WORKFLOW_ENGINE.md). bash/workflow list # discover built-in + custom workflows /workflow run elixir Add dark mode # run the bundled "elixir" workflow /workflow status [run-id] # inspect a run (works even after a crash) /workflow resume # continue a failed/interrupted run /workflow abort # mark a run as aborted /workflow init [--from ] # scaffold a custom workflow definition Parallel subtasks each get their own git worktree and branch, so concurrent agent processes can never clash on disk even if the model's proposed split wasn't perfectly non-overlapping. --- ## Keyboard Shortcuts & Navigation | Key Combination | Action | | --- | --- | | `Tab` | Trigger slash-command or path autocomplete | | `Ctrl+R` | Reverse history search (type query to search past inputs) | | `Up Arrow` / `Down Arrow` | Navigate input history (or navigate options in TUI modals) | | `Left Arrow` / `Right Arrow` | Move cursor left/right (or accept ghost auto-suggestion) | | `Home` / `Ctrl+A` | Jump cursor to line start | | `End` / `Ctrl+E` | Jump cursor to line end | | `Ctrl+U` | Clear text from cursor to start of line | | `Ctrl+K` | Clear text from cursor to end of line | | `Ctrl+W` | Delete word backward | | `Spacebar` | Toggle checkbox option in TUI Question Modals | | `Enter` | Submit input / confirm modal selection | | `Ctrl+C` | Cancel line input | | `Ctrl+D` | Signal EOF and exit `yoke` |