Claude Agent

Claude is Anthropic's AI assistant, available through the Claude Code CLI. It's Bloom's default agent, offering a rich tool ecosystem with web search, session management, and the TodoWrite tool for progress tracking.

Prerequisites

1. Anthropic Account: Create an account at anthropic.com
2. API Key: Generate an API key from the Anthropic Console
3. Claude Code CLI: Install the command-line interface

Installation

See the official Claude Code documentation for installation instructions.

After installation, verify it's working:

claude --version

Configuration

Basic Configuration

Claude is the default agent, so no configuration is required for basic usage. To explicitly configure:

# ~/.bloom/config.yaml
agent:
  defaultInteractive: claude
  defaultNonInteractive: claude

Model Selection

Claude supports model selection via the agent config:

# ~/.bloom/config.yaml
agent:
  defaultInteractive: claude
  defaultNonInteractive: claude

  claude:
    defaultModel: opus    # Use Claude Opus by default
    models:               # Available models for switching
      - sonnet
      - haiku
      - opus

Available models depend on your API access:

• sonnet - Claude Sonnet (default, balanced)
• opus - Claude Opus (most capable)
• haiku - Claude Haiku (fastest)

Configuration Commands

# Set claude as default
bloom config set-interactive claude
bloom config set-noninteractive claude

# Set default model
bloom config set-model claude opus

# Discover and save available models
bloom config models claude --discover --save

Capabilities

Capability	Supported	Notes
File Read	Yes	Read any file in the working directory
File Write	Yes	Create and modify files
Bash/Terminal	Yes	Execute shell commands
Git Operations	Yes	Full git support
Web Search	Yes	Search the web for information
Web Fetch	Yes	Fetch and read web pages
Session Resume	Yes	Continue previous sessions with --resume
Human Questions	Yes	Can pause to ask clarifying questions
TodoWrite	Yes	Built-in task tracking tool

Unique Features

TodoWrite Tool

Claude has a built-in TodoWrite tool for tracking task progress. When working on complex tasks, Claude will:

1. Create a checklist from acceptance criteria
2. Update items as work progresses
3. Provide visibility into completion status

Task Subagents

Claude can spawn specialized subagents for complex tasks:

• Explore agent: For codebase exploration
• Plan agent: For implementation planning
• Research agent: For gathering information

Web Search & Fetch

Claude can search the web for documentation, examples, and current information. It can also fetch and read specific URLs.

Provider-Specific Options

The Claude provider supports these additional options (used internally by Bloom):

Option	Type	Default	Description
dangerouslySkipPermissions	boolean	true	Skip permission prompts for autonomous execution
activityTimeoutMs	number	600000	Timeout (10 min) before considering agent stuck
heartbeatIntervalMs	number	10000	Heartbeat interval (10s) in streaming mode
verbose	boolean	false	Show detailed event output

CLI Flags Reference

When Bloom runs Claude, it uses these flags:

# Interactive mode
claude --verbose --dangerously-skip-permissions \
  --append-system-prompt "system prompt" \
  "initial prompt"

# Streaming mode (autonomous)
claude -p --verbose --dangerously-skip-permissions \
  --output-format stream-json \
  --append-system-prompt "system prompt" \
  "task prompt"

# Resume session
claude -p --resume <session_id> ...

Troubleshooting

"Failed to spawn claude"

Cause: Claude CLI not installed or not in PATH

Solution:

# Check if installed
which claude

If not found, reinstall using the official docs and ensure the binary is in your PATH.

"Authentication failed" or "Invalid API key"

Cause: Missing or invalid ANTHROPIC_API_KEY

Solution:

# Check if set
echo $ANTHROPIC_API_KEY

# Set if missing
export ANTHROPIC_API_KEY="sk-ant-..."

# Verify key is valid at console.anthropic.com

"Agent timed out due to inactivity"

Cause: Claude didn't produce output for 10 minutes

Possible causes:

• Complex operation taking longer than expected
• API rate limiting
• Network issues

Solution:

• Check Anthropic API status
• Review task complexity (break into smaller tasks)
• Check network connectivity

"Rate limit exceeded"

Cause: Too many API requests

Solution:

• Wait and retry
• Consider upgrading API tier
• Reduce parallel agent count in Bloom

Best Practices

For Interactive Sessions

1. Use Claude for exploration and planning
2. Let Claude use web search for documentation
3. Review TodoWrite progress for complex tasks

For Autonomous Tasks

1. Provide clear, specific acceptance criteria
2. Claude will track progress via TodoWrite
3. Use checkpoints for human review at phase boundaries

Prompt Tips

Claude responds well to:

• Clear task boundaries
• Specific acceptance criteria
• Context about the codebase
• Examples of expected output

Example Session

# Start interactive session with Claude
bloom enter

# Claude will:
# 1. Read the project context (PRD.md, plan.md, etc.)
# 2. Use TodoWrite to track progress
# 3. Search web for documentation as needed
# 4. Ask clarifying questions when unsure