Autocomplete

tallow has two independent autocomplete systems that work together in the editor input:

System	Trigger	What it does
Structural	/, @, file paths, Tab	Completes commands, file references, and paths deterministically
LLM ghost text	Typing 4+ characters	Calls a fast model to predict how you’ll finish your sentence

Structural autocomplete

Structural autocomplete is built into the TUI editor. It uses no LLM — completions come from the filesystem, registered commands, and fuzzy matching. Results appear in a dropdown menu.

Slash commands

Type / at the start of a line to see all available commands. Keep typing to fuzzy-filter the list.

/rev       → matches /rewind, /reviewer
/task      → matches /tasks

After selecting a command and pressing Space, some commands offer argument completions (for example, /context-fork suggests branch names).

@file references

Type @ followed by a filename to fuzzy-search your project. The search uses fd under the hood — it’s fast, respects .gitignore, and searches the full directory tree.

@login     → matches src/auth/login.ts, tests/auth/login.test.ts
@pack      → matches package.json, packages/

Each suggestion shows:

Field	Content
Label	Filename (with trailing / for directories)
Description	Relative path from project root
Value	Full @path insertion text

Directories are sorted first. Filenames with spaces are automatically quoted (@"path with spaces/file.ts").

When you select a file, the file-reference extension reads it and inlines the contents into your prompt.

File path completions

Paths are completed via directory listing whenever the input contains path-like patterns:

• ./src/ — relative paths
• ~/ — home directory paths
• ../ — parent directory paths
• Any text containing /

Press Tab to force file completion even when the input doesn’t look like a path yet.

Keys

Key	Behavior
Tab	Accept selected completion / force file completions
↑ / ↓	Navigate completion list
Escape	Dismiss completions
Typing	Filters the completion list

LLM ghost text

When you type 4 or more characters (and aren’t typing a / command), tallow calls a fast, cheap model to suggest how you might finish your sentence. The suggestion appears as dim ghost text after your cursor.

fix the login bug on the  ← you typed this
                          authentication page  ← ghost text

Keys

Key	Behavior
Tab	Accept ghost text into the editor
Enter (empty input)	Accept idle suggestion and submit
Escape	Dismiss ghost text
Any character	Dismiss ghost text, type normally

Model and cost

The default model is Groq Llama 3.1 8B at $0.05 / $0.08 per million tokens. If it’s unavailable, tallow falls back through a chain of cheap models. See the prompt-suggestions extension for the full fallback chain, conversation context details, and configuration.

Ghost text is capped at 200 API calls per session. At ~50 tokens per call, a full session of autocomplete costs roughly $0.004.

Idle suggestions

When the editor is empty and the agent is idle, a random prompt suggestion appears as ghost text (no model call — these are picked from a curated template list). Press Enter to accept and submit, or start typing to dismiss.

Configuration

All LLM autocomplete settings live in ~/.tallow/settings.json:

{
  "prompt-suggestions.enabled": true,
  "prompt-suggestions.autocomplete": true,
  "prompt-suggestions.model": "groq/llama-3.1-8b-instant",
  "prompt-suggestions.debounceMs": 600
}

Setting	Default	Description
prompt-suggestions.enabled	true	Enable/disable the entire extension (idle + LLM)
prompt-suggestions.autocomplete	true	Enable/disable LLM autocomplete only
prompt-suggestions.model	groq/llama-3.1-8b-instant	Model for autocomplete (provider/model-id)
prompt-suggestions.debounceMs	600	Delay in ms before calling the model

Structural autocomplete has no configuration — it’s always available.