.env.example

# Backend selection: ollama (default), lm-studio, mlx, llamacpp, openrouter, openai
BACKEND=ollama

# Override the model for the chosen backend (otherwise uses the backend default)
# AGENT_MODEL=qwen2.5-coder:14b

# Per-backend endpoint overrides (defaults are the standard local ports)
# OLLAMA_BASE_URL=http://localhost:11434/v1
# LM_STUDIO_BASE_URL=http://localhost:1234/v1
# MLX_BASE_URL=http://localhost:8080/v1
# LLAMACPP_BASE_URL=http://localhost:8080/v1

# Optional only if llama-server was started with API-key enforcement.
# LLAMACPP_API_KEY=sk-no-key-required

# Required only when BACKEND=openrouter or you want the cloud-compare slash command
# OPENROUTER_API_KEY=sk-or-...

# Required only when BACKEND=openai. Point OPENAI_BASE_URL at a compatible
# proxy if you don't want to hit api.openai.com directly.
# OPENAI_API_KEY=sk-...
# OPENAI_BASE_URL=https://api.openai.com/v1

# BACKEND=openai-codex signs in with a ChatGPT/Codex subscription via OAuth
# (run `/login openai-codex`) instead of an API key — no key needed here.
# Override the Codex endpoint only if you proxy it.
# OPENAI_CODEX_BASE_URL=https://chatgpt.com/backend-api

# Approval policy for mutating tools: always (default) | never | dangerous-only
APPROVAL_POLICY=always

# Tools to enable, comma-separated. Each tool you remove saves ~100 tokens
# of prompt eval per turn — meaningful on small local models.
# Available: apply_patch, file_read, file_write, file_edit, glob, grep, list_dir,
#   shell, repo_search, run_tests, batch_edit, ship_status, web_fetch, update_plan, task
# Default (core): file_read, grep, list_dir, file_edit, file_write, shell, update_plan, task
# AGENT_TOOLS=file_read,grep,list_dir,file_edit,file_write,shell,update_plan,task

# Tool schema selection. auto sends only prompt-relevant enabled tools;
# fixed sends every enabled tool schema every turn.
# AGENT_TOOL_SELECTION=auto

# Tuning
AGENT_MAX_STEPS=20

# Workspace safety. Relative file paths are normalized before tool execution.
# prompt (default) asks for approval outside the root; deny blocks; allow permits.
# WORKSPACE_ROOT=/Users/me/my-project
# OUTSIDE_WORKSPACE=prompt

# Context and input history
# AGENT_CONTEXT_MAX_MESSAGES=40
# AGENT_CONTEXT_MAX_BYTES=262144
# AGENT_HISTORY=true
# AGENT_HISTORY_MAX_ENTRIES=200

# Pre-load the model + cache the system prompt at startup so the first prompt
# is fast (llama.cpp-derived engines cache prompt-eval results for matching prefixes).
# WARMUP=false to skip.
# WARMUP=true

# Skip the first-run setup wizard and rely on env vars / built-in defaults.
# SMALL_HARNESS_NO_WIZARD=true

# Skip the background GitHub release check. Only outbound call the harness
# makes that isn't user-initiated; opt out if you want zero network at idle.
# SMALL_HARNESS_NO_UPDATE_CHECK=true

# Ephemeral managed hooks for process-owning launchers such as terminal
# orchestrators or agent status dashboards. These must be exported in the
# launcher process; repo .env files are intentionally ignored for managed hooks.
# Prefer the file form for larger hook documents to avoid shell quoting JSON.
# SMALL_HARNESS_MANAGED_HOOKS_FILE=/tmp/agent-status-hooks.json
# SMALL_HARNESS_MANAGED_HOOKS_JSON={"source":"terminal-orchestrator","hooks":{}}