Harvey Configuration Reference

# Knowledge base
knowledge_db: "agents/knowledge.db"

# Session recordings
sessions_dir: "agents/sessions"

# Skills/agents directory
agents_dir: "agents"

# Auto-record setting (null = use default)
auto_record: true

# Model capability cache
model_cache_db: "agents/model_cache.db"

# RAG configuration
rag:
  enabled: true
  active: "default"
  stores:
    - name: "default"
      db_path: "agents/rag/default.db"
      embedding_model: "nomic-embed-text"
      model_map:
        "llama3.1:latest": "nomic-embed-text"
        "granite-code:3b": "nomic-embed-text"
        "mistral:latest": "nomic-embed-text"
    - name: "deno_typescript"
      db_path: "agents/rag/deno_typescript.db"
      embedding_model: "nomic-embed-text"
      model_map:
        "llama3.1:latest": "nomic-embed-text"

llamafile:
  models_dir: ~/Models
  active: qwen-coding
  url: http://localhost:8080
  gpu_layers: 99
  startup_timeout: 120s
  models:
    - name: qwen-coding
      path: ~/Models/Qwen2.5-Coder-7B-Q5_K_S.llamafile
      context_length: 32768
    - name: phi-mini
      path: ~/Models/Phi-3.5-mini-instruct-Q4_K_M.llamafile

permissions:
  ".":           [read, write, exec, delete]   # workspace root — full access
  "docs/":       [read]                         # docs tree — read-only
  "scripts/":    [read, exec]                   # scripts — run but not modify

model_aliases:
  coder: qwen2.5-coder:7b
  phi: phi-mini          # registered llamafile name
  claude: claude-sonnet-4-6

tools:
  enabled: true
  max_tool_calls_per_turn: 10
  max_output_bytes: 65536
  tool_result_compaction: true

memory:
  enabled: true
  top_k: 5
  inject_on_start: true
  budget_pct: 0.25
  rolling_summary:
    enabled: true
    warn_at_pct: 0.80
    keep_turns: 6
  knowledge_base:
    path: agents/knowledge.db
  rag:
    enabled: true
    active: golang
    stores:
      - name: golang
        db_path: agents/rag/golang.db
        embedding_model: nomic-embed-text

syntax_highlight: true
auto_format: true

# Paths
sessions_dir: agents/sessions
agents_dir: agents
model_cache_db: agents/model_cache.db
auto_record: true

# Llamafile backend (primary)
llamafile:
  models_dir: ~/Models
  active: qwen-coding
  startup_timeout: 120s
  gpu_layers: 99
  models:
    - name: qwen-coding
      path: ~/Models/Qwen2.5-Coder-7B-Q5_K_S.llamafile
      context_length: 32768

# Model short-name aliases
model_aliases:
  coder: qwen-coding
  phi: phi-mini

# Behaviour
syntax_highlight: true
auto_format: true

# Security
safe_mode: true
allowed_commands: [ls, cat, grep, head, tail, wc, find, stat, git, go, make]
run_timeout: "5m"
ollama_timeout: ""

permissions:
  ".":      [read, write, exec, delete]
  "docs/":  [read]

# Tool calling
tools:
  enabled: true
  max_tool_calls_per_turn: 10

# Memory system
memory:
  enabled: true
  top_k: 5
  budget_pct: 0.25
  rolling_summary:
    enabled: true
    warn_at_pct: 0.80
    keep_turns: 6
  rag:
    enabled: true
    active: golang
    stores:
      - name: golang
        db_path: agents/rag/golang.db
        embedding_model: nomic-embed-text

{
  "enabled": true,
  "endpoints": [
    {
      "name": "pi2",
      "url": "ollama://192.0.2.12:11434",
      "model": "llama3.1:8b",
      "kind": "ollama"
    },
    {
      "name": "claude",
      "url": "anthropic://",
      "model": "claude-3-haiku",
      "kind": "anthropic"
    },
    {
      "name": "pi3",
      "url": "ollama://192.0.2.13:11434",
      "model": "mistral:latest",
      "kind": "ollama"
    }
  ]
}

You are Harvey, a terminal coding agent. You are working in a Go project
located at /home/user/myproject.

## Project Context

Today: <!-- @date -->

Current git status:
<!-- @git-status -->

Workspace files:
<!-- @files -->

## Conventions

- Use Go 1.26+ features
- Prefer stdlib over third-party packages
- All exported symbols need doc comments
- Use `t.TempDir()` in tests

## Available Tools

You have access to the following slash commands: /read, /write, /run, /search,
/git, /apply, /clear, /summarize, /context, /kb, /rag, /ollama, /route, /skill.

## Rules

1. Tag code blocks with file paths for auto-apply
2. Use /run to suggest shell commands
3. Never narrate fake command output

# 1. Start Harvey for the first time (creates defaults)
harvey

# 2. Register remote endpoints
harvey
> /route add pi2 ollama://192.0.2.12:11434 llama3.1:8b
harvey
> /route on

# 3. Configure RAG (optional)
harvey
> /rag new golang
harvey
> /rag use golang
harvey
> /rag on

# 4. Create workspace-specific config
mkdir -p myproject/agents
echo 'knowledge_db: myproject/agents/knowledge.db' > myproject/agents/harvey.yaml

# 5. Run Harvey in the project
harvey -w myproject

# Check workspace routes
harvey --help routing  # or: cat agents/routes.json

# Check workspace config
cat agents/harvey.yaml

# Check knowledge base
sqlite3 agents/knowledge.db ".tables"

# Check model cache
sqlite3 agents/model_cache.db "SELECT count(*) FROM models;"

# Check RAG stores
ls -la agents/rag/*.db

{
  "enabled": true/false,
  "endpoints": [
    {
      "name": "string (required)",
      "url": "string (required)",
      "model": "string (optional)",
      "kind": "string (optional, auto-detected)"
    }
  ]
}

# ── Paths ────────────────────────────────────────────────────────────────────
sessions_dir: agents/sessions          # default
agents_dir: agents                     # default
model_cache_db: agents/model_cache.db  # default

# ── Behaviour ────────────────────────────────────────────────────────────────
auto_record: true          # false to disable session auto-recording
syntax_highlight: true     # false to disable ANSI colour in code blocks
auto_format: true          # false to disable auto-formatting after write_file

# ── Llamafile backend ─────────────────────────────────────────────────────────
llamafile:
  models_dir: ~/Models            # discovery directory; $HOME/Models is the default
  active: qwen-coding             # name of the active registered model
  url: http://localhost:8080      # API base URL; this is the default
  gpu_layers: 99                  # -ngl value; 99 = maximise GPU, -1 = CPU only
  startup_timeout: 120s           # time to wait for the server to become ready
  models:
    - name: qwen-coding
      path: ~/Models/Qwen2.5-Coder-7B-Q5_K_S.llamafile
      context_length: 32768       # optional; probed from server when omitted

# ── Model aliases ─────────────────────────────────────────────────────────────
model_aliases:
  coder: qwen2.5-coder:7b         # short name → full Ollama model ID or llamafile name
  phi: phi-mini

# ── Security ──────────────────────────────────────────────────────────────────
safe_mode: true
allowed_commands:
  - ls
  - cat
  - grep
  - git
run_timeout: "5m"
ollama_timeout: ""                # empty = no timeout (recommended for local hardware)
permissions:
  ".":      [read, write, exec, delete]
  "docs/":  [read]

# ── Tool calling ──────────────────────────────────────────────────────────────
tools:
  enabled: true
  max_tool_calls_per_turn: 10     # hard limit per user turn
  max_output_bytes: 65536         # cap on tool output injected into context (64 KiB)
  tool_result_compaction: true    # compact prior tool-call rounds before new calls

# ── Memory system ─────────────────────────────────────────────────────────────
memory:
  enabled: true
  top_k: 5                        # experience memories retrieved at session start
  inject_on_start: true
  budget_pct: 0.25                # fraction of context window reserved for memory
  rolling_summary:
    enabled: true
    warn_at_pct: 0.80             # compression triggers at 80% context usage
    keep_turns: 6                 # recent turns kept verbatim after compression
  knowledge_base:
    path: agents/knowledge.db
  rag:
    enabled: true
    active: golang
    stores:
      - name: golang
        db_path: agents/rag/golang.db
        embedding_model: nomic-embed-text
        model_map:
          qwen2.5-coder:7b: nomic-embed-text
        embedder_kind: ollama     # or "encoderfile" for a local encoderfile binary
        embedder_url: ""          # base URL when embedder_kind = "encoderfile"