跳到主要内容

viben agent chat

Non-interactive conversation with a specified Agent.

Overview​

Adds a chat subcommand to viben agent for non-interactive Agent conversations. The Agent calls the corresponding underlying executor based on its configured type (e.g., claude-code, gemini), while automatically loading the Agent's memory, MCP configuration, and Skills.

Command Interface​

viben agent chat [OPTIONS] -n <AGENT_ID>

OPTIONS:
    -n, --name <AGENT_ID>       Agent ID (required)
    -p, --prompt <PROMPT>       Prompt (optional, reads from stdin if not provided)
    -C, --cwd <DIR>             Working directory (default: current directory)

    --input-format <FORMAT>     Input format: text (default), stream-json
    --output-format <FORMAT>    Output format: text (default), stream-json
    --verbose                   Verbose output

    -s, --session <SESSION_ID>  Specify session ID
    --resume <SESSION_ID>       Resume existing session
    --new-session               Force create new session

    --model <MODEL>             Override Agent's configured model
    --no-memory                 Don't load Agent memory
    --dangerously-skip-permissions  Skip permission checks

    --json                      JSON format output

Data Flow Architecture​

Usage Examples​

# ============================================================
# Basic Usage
# ============================================================

# Conversation with default agent
viben agent chat -p "Analyze this code"

# Specify agent
viben agent chat -n my-agent -p "Write a sorting function"

# Read prompt from stdin
echo "Explain this error" | viben agent chat -n my-agent

# Specify working directory
viben agent chat -n my-agent -p "Analyze project structure" -C /path/to/project

# ============================================================
# Session Management
# ============================================================

# Use specific session
viben agent chat -n my-agent -p "Continue previous work" -s main

# Resume existing session
viben agent chat -n my-agent -p "Continue" --resume abc123

# Force create new session
viben agent chat -n my-agent -p "Start new task" --new-session

# ============================================================
# Advanced Options
# ============================================================

# Override model
viben agent chat -n my-agent -p "Complex reasoning task" --model claude-3-opus

# Don't load memory (clean state)
viben agent chat -n my-agent -p "Independent task" --no-memory

# Skip permission checks
viben agent chat -n my-agent -p "Automation script" --dangerously-skip-permissions

# ============================================================
# Programmatic Calls (JSON Stream)
# ============================================================

# JSON stream input/output
echo '{"type":"user","message":{"role":"user","content":"Analyze code"}}' | \
  viben agent chat -n my-agent --input-format stream-json --output-format stream-json

# JSON format output
viben agent chat -n my-agent -p "Quick task" --json

Agent Memory Injection​

Agent chat automatically loads and injects the Agent's memory:

Use --no-memory to skip memory loading.

Configuration Merge Order​

When agent chat executes, configuration is merged by priority:

Priority from low to high:

1. Agent Base Configuration
   ~/.viben/agents/<id>/config.yaml
   - type, model, mcp, skills, etc.
────────────────────────────────────────────────────
2. Workspace Configuration (based on agent type)
   <cwd>/.claude/settings.json  (if type=claude-code)
   <cwd>/.cursor/settings.json  (if type=cursor)
────────────────────────────────────────────────────
3. Command Line Arguments
   --model, --session, --no-memory, etc.

Session Storage​

Conversation sessions are stored in the Agent directory:

~/.viben/agents/<agent-id>/.agent_sessions/
└── <session_id>/
    ├── config.yaml              # Session configuration
    │   session_id: abc123
    │   created_at: 2024-01-16T10:30:00Z
    │   last_active: 2024-01-16T14:15:00Z
    │   cwd: /path/to/project
    │   model: claude-3-sonnet
    │
    └── messages.rollout.jsonl   # Message history (JSONL format)

Output Examples​

viben agent chat -n my-agent -p "Hello" (Human):

Hello! I'm your coding assistant. How can I help you today?

I have access to your project context and remember our previous conversations.
What would you like to work on?

viben agent chat -n my-agent -p "Hello" --json:

{
  "success": true,
  "agent_id": "my-agent",
  "session_id": "abc123",
  "response": "Hello! I'm your coding assistant...",
  "tokens": {
    "input": 150,
    "output": 42
  },
  "memory_loaded": true,
  "duration_ms": 1234
}

Error: Agent not found:

Error: Agent not found: unknown-agent

Available agents:
  main         claude-code
  my-agent     claude-code
  researcher   gemini

Use `viben agent list` to see all agents.

Error: Agent type doesn't support chat:

Error: Chat not supported for agent type: custom

Supported types: claude-code, gemini, codex

Difference from executor chat​

Featureviben executor chatviben agent chat
TargetDirectly call underlying executorCall configured Agent
MemoryNoneAutomatically loads Agent memory
ConfigurationCommand line args onlyAgent config + workspace + command line
SessionTemporaryPersisted to Agent directory
MCPManual configuration neededUses Agent's MCP configuration
SkillsNoneUses Agent's Skills
Use CaseOne-off calls, script integrationOngoing conversations, project development

Acceptance Criteria​

Basic Functionality​

  • viben agent chat -n <id> -p <prompt> executes conversation
  • Read prompt from stdin (when -p not provided)
  • Select correct executor based on Agent type
  • -C specifies working directory

Memory System​

  • Automatically load MEMORY.md
  • Automatically load today + yesterday logs
  • --no-memory skips memory loading
  • Update daily log after conversation

Session Management​

  • -s specifies session ID
  • --resume resumes existing session
  • --new-session forces new session creation
  • Session persisted to .agent_sessions/

Configuration Merge​

  • Load Agent base configuration
  • Merge workspace configuration
  • Command line args override configuration
  • --model overrides model setting

Input/Output Formats​

  • --input-format text (default)
  • --input-format stream-json
  • --output-format text (default)
  • --output-format stream-json
  • --json outputs JSON result

Error Handling​

  • Show available agents when Agent not found
  • Error when Agent type doesn't support chat
  • Error when no prompt and stdin is empty

Permissions & Security​

  • Permission checks required by default
  • --dangerously-skip-permissions skips checks
  • --verbose verbose output