friday-sh 0.1.0

LLM-powered shell agent — a ZSH extension for coding, debugging, and shell assistance

friday

LLM-powered shell agent that lives in your ZSH terminal.

friday:auto(glm-5-turbo)> explain why this test is failing

Friday is a conversational AI assistant with coding, debugging, and documentation modes. It routes tasks to specialized sub-agents, remembers context across sessions, and integrates with your shell workflow via keybindings and completions.

Quick Start

Install with uvx (no clone needed)

uvx --from git+https://github.com/oornnery/friday-sh friday chat

Install with uv tool (global)

uv tool install git+https://github.com/oornnery/friday-sh
friday chat

Install from source

git clone https://github.com/oornnery/friday-sh.git
cd friday
uv sync --group dev --group test

Set up API keys

cp .env.example .env
# Edit .env with at least one provider key

# Anthropic
ANTHROPIC_API_KEY=

# OpenAI
OPENAI_API_KEY=

# Mistral
MISTRAL_API_KEY=

# Z.AI (ZhipuAI / GLM)
ZAI_API_KEY=
ZAI_BASE_URL=https://api.z.ai/api/coding/paas/v4

ZSH plugin

Source the plugin in your .zshrc:

source /path/to/friday/src/friday/shell/friday.plugin.zsh

Now use f from anywhere in your shell:

f "why did my last command fail?"
f "how do I use find to delete tmp files?"

Requirements

• Python 3.13+
• uv
• At least one LLM provider key (or local Ollama)

Usage

CLI

# Ask a single question (uses router agent by default)
friday ask "explain this error"

# Pipe input
cat error.log | friday ask "what went wrong?"
git diff | friday ask "review this"

# Force a specific mode
friday ask --mode debug "why is pytest failing?"
friday ask --mode reader "explain the router module"

# Use a specific model
friday ask --model mistral:codestral-latest "refactor this function"

# Interactive chat
friday chat
friday chat --mode code --model zai:glm-5-turbo

REPL

The chat REPL has a styled prompt showing mode and model:

friday:auto(glm-5-turbo)> hello
friday:code(codestral-latest)> fix the failing test
friday:debug(glm-5-turbo)[debug]> /mode show

Commands:

  /model       Interactive model picker (or /model show | /model <name>)
  /mode        Interactive mode picker (or /mode show | /mode <name>)
  /session     Resume session picker (or /session show | resume | new | delete)
  /setting     Show settings (or /setting <key> | /setting <key>=<value>)
  /memory      List memories (or /memory show | search | add | delete)
  /debug       Toggle debug (or /debug on | off | show)
  /clear       Clear conversation
  /quit        Exit Friday

Type @ to trigger file completion, / for command completion.

The /model and /mode pickers support search -- type to filter:

  Select model  (3/87)
  / glm
    zai:glm-4.5
  > zai:glm-5-turbo  (active)
    zai:glm-5.1

ZSH Keybindings

Key	Action
f "question"	Ask Friday from anywhere
Ctrl+F	Ask about current buffer / last command
Ctrl+G	Fuzzy session picker (requires fzf)
friday-select-model	Interactive model picker with fzf

Resource Commands (CLI)

friday model show                 # List all models from providers
friday model set                  # Interactive picker
friday mode show                  # List modes
friday session show               # List saved sessions
friday session resume <id>        # Resume session
friday setting show               # Show all settings
friday memory show                # List shared memories
friday memory add "user prefers concise answers"
friday memory search "preferences"

Architecture

                          +-------------------+
                          |     ZSH Shell     |
                          |  f "question"     |
                          |  Ctrl+F / Ctrl+G  |
                          +---------+---------+
                                    |
                          +---------v---------+
                          |    friday CLI     |
                          |  ask | chat       |
                          +---------+---------+
                                    |
                    +---------------v---------------+
                    |        Router Agent            |
                    |   (auto mode -- conversational)|
                    +--+--------+--------+--------+-+
                       |        |        |        |
              +--------v--+ +---v----+ +-v------+ +--v-------+
              | Code Agent| | Reader | | Writer | |  Debug   |
              | read,write| | read   | | read,  | | read,    |
              | patch,    | | list,  | | write, | | list,    |
              | shell,    | | search | | list,  | | search,  |
              | search    | |        | | search | | shell    |
              +-----------+ +--------+ +--------+ +----------+
                       |        |        |        |
                    +--v--------v--------v--------v--+
                    |         Tool Layer              |
                    |  filesystem | shell | memory    |
                    +---------------+----------------+
                                    |
                    +---------------v---------------+
                    |        Infrastructure          |
                    | config | sessions | memory.db  |
                    | MCP servers | model providers  |
                    +-------------------------------+

Agent Modes

Mode	Purpose	Tools
auto	Routes to the right specialist	delegation only
code	Write, edit, test code	read, write, patch, list, search, shell
reader	Analyze and explain code	read, list, search
write	Generate docs and text	read, write, list, search
debug	Diagnose errors and bugs	read, list, search, shell

Each mode is defined in src/friday/agent/prompts/<mode>.md with YAML frontmatter specifying tools, max steps, model override, and thinking.

Approval System

Risky tools (write_file, patch_file, run_shell) require user confirmation before executing:

  +-- Confirm ----------------------+
  | run_shell: Execute shell command |
  |                                  |
  | rm -rf node_modules && npm i     |
  +----------------------------------+
  Allow? [y/N]

Configurable via approval_policy:

• ask (default) -- prompt for each risky action
• auto -- execute without asking
• never -- always deny

Memory System

Friday has three layers of memory:

  +---------------------------+
  | Shared Memory (SQLite)    |  <-- cross-session, searchable
  | profile, preferences,     |      agent saves via save_memory
  | notes, decisions          |
  +---------------------------+
  | Working Memory            |  <-- per-session, short-term
  | task, files, notes,       |      auto-managed by harness
  | entities, decisions       |
  +---------------------------+
  | Message History           |  <-- per-session transcript
  | pydantic-ai messages      |      trimmed by history processor
  +---------------------------+

The agent decides what to save to shared memory via the save_memory tool. The harness indexes chat turns for cross-session search.

Supported Providers

Prefix	Provider	Auth
anthropic:	Anthropic (Claude)	ANTHROPIC_API_KEY
openai:	OpenAI (GPT)	OPENAI_API_KEY
mistral:	Mistral AI	MISTRAL_API_KEY
zai:	Z.AI (GLM)	ZAI_API_KEY
ollama:	Ollama (local)	no key needed

Model fallback: if the default model fails (missing key), Friday automatically tries fallback_model from config.

MCP Integration

Friday supports Model Context Protocol servers. Configure in ~/.config/friday/config.toml:

[[mcp_servers]]
name = "github"
transport = "http"
url = "http://localhost:3000/mcp"

[[mcp_servers]]
name = "filesystem"
transport = "stdio"
command = "npx"
args = ["-y", "@modelcontextprotocol/server-filesystem", "/home/user"]

Configuration

Friday reads config from (in priority order):

1. Environment variables (FRIDAY_* prefix)
2. .env file
3. ~/.config/friday/config.toml or friday.toml

Settings

Key	Default	Description
default_model	zai:glm-5-turbo	Default LLM
fallback_model	mistral:devstral-latest	Fallback if default fails
default_mode	auto	Default agent mode
approval_policy	ask	Risky tool confirmation
max_steps	25	Max tool calls per turn
session_dir	~/.local/share/friday/sessions	Session storage
config_dir	~/.config/friday	Config directory
memory_db_path	memory.db	SQLite memory database
memory_top_k	6	Memory results per turn

Runtime override in chat: /setting default_model=mistral:codestral-latest

Project Structure

src/friday/
|-- __init__.py              # package version
|-- __main__.py              # python -m friday
|
|-- agent/                   # core AI runtime
|   |-- core.py              # agent factory, execution loop, approval flow
|   |-- router.py            # delegation tools (delegate_code, etc.)
|   |-- memory.py            # shared memory retrieval + chat indexing
|   |-- context.py           # workspace snapshot (git, env, shell state)
|   |-- contracts.py         # AgentReply, RouterDecision, TurnOutput
|   |-- deps.py              # AgentDeps injected into tools
|   |-- modes.py             # mode config loader (YAML frontmatter)
|   |-- history.py           # history trimming (pair-safe, dedup)
|   |-- stats.py             # token/cost tracking per turn
|   `-- prompts/             # system prompts per mode
|       |-- router.md        # auto mode -- conversational router
|       |-- code.md          # code mode -- full access
|       |-- reader.md        # reader mode -- read-only
|       |-- writer.md        # writer mode -- docs generation
|       `-- debug.md         # debug mode -- error diagnosis
|
|-- cli/                     # terminal interface
|   |-- app.py               # typer entrypoint + all CLI commands
|   |-- chat.py              # interactive REPL loop
|   |-- ask.py               # single-shot ask
|   |-- catalog.py           # command registry (REPL + CLI)
|   |-- completer.py         # / commands + @ file completion
|   |-- picker.py            # interactive list picker with search
|   |-- confirm.py           # approval dialogs for risky tools
|   |-- models.py            # dynamic model listing from providers
|   |-- resources.py         # interactive pickers for model/mode/session
|   |-- output.py            # rich console, markdown rendering
|   |-- theme.py             # unified color palette
|   `-- debug.py             # debug mode logging
|
|-- domain/                  # pure business logic (no IO)
|   |-- models.py            # AgentMode, WorkingMemory, MemoryKind
|   |-- validation.py        # input limits (path, command, content)
|   `-- permissions.py       # safe_path, contains_secret, clip
|
|-- infra/                   # IO boundary
|   |-- config.py            # FridaySettings (pydantic-settings)
|   |-- sessions.py          # JSON session store
|   |-- memory.py            # SQLite + FTS5 shared memory
|   `-- mcp.py               # MCP server factory
|
|-- tools/                   # pydantic-ai tool implementations
|   |-- filesystem.py        # read, write, patch, list, search
|   |-- shell.py             # run_shell with validation
|   |-- memory.py            # search_memory, save_memory, list_memories
|   `-- registry.py          # tool metadata
|
`-- shell/
    `-- friday.plugin.zsh    # ZSH integration (f, Ctrl+F, Ctrl+G)

Development

# Setup
git clone https://github.com/oornnery/friday-sh.git
cd friday
uv sync --group dev --group test

# Run all checks (format + lint + typecheck + test)
uv run task check

# Individual tasks
uv run task format      # ruff format
uv run task lint        # ruff check --fix
uv run task typecheck   # ty check
uv run task test        # pytest -v

# Run a single test
uv run pytest tests/test_security.py::TestContainsSecret -v

# Run friday locally
uv run friday chat
uv run friday ask "hello"

License

MIT