Environment Variables

Required

Variable	Description
ISSUE_BODY	The task description. Required unless prompt is passed via CLI args.

Core Config

Variable	Default	Description
VERBOSE	false	Enable verbose streaming output (true/false)
AGENTS_DIR	.github/agents	Directory containing agent instruction files
SWARM_DIR	.swarm	Root directory for all swarm output
SESSION_TIMEOUT_MS	1800000	Agent session timeout in ms (default: 30 minutes)
MAX_AUTO_RESUME	3	Auto-resume attempts on failure (0 to disable)
MAX_RETRIES	2	Retry attempts for failed agent responses
AUTO_MODEL	false	Auto-select model per task based on complexity

Model Overrides

These override values in swarm.config.yaml:

Variable	Default (YAML)	Description
PRIMARY_MODEL	claude-opus-4-6	Primary model for all main agent sessions
REVIEW_MODEL	gpt-5.3-codex	Model for cross-model review sessions
FAST_MODEL	claude-haiku-4.5	Model for coordination tasks (cheaper, faster)

Analysis Config

Variable	Default	Description
ANALYZE_CHUNK_THRESHOLD	500	File count to trigger chunked analysis
ANALYZE_CHUNK_MAX_FILES	300	Max files per chunk
MODEL_CONTEXT_LIMIT	128000	Model context window in tokens
TOKEN_BUDGET_RATIO	0.7	Fraction of context used for prompt content

Prepare Config

Variable	Default	Description
PREPARE_DEEP_THRESHOLD	10	Source file count for deep directory analysis

Examples

# Use different models
PRIMARY_MODEL=gpt-5.2 REVIEW_MODEL=claude-opus-4-6 swarm "Add feature"

# Skip cross-model review (same model for both)
PRIMARY_MODEL=claude-opus-4-6 REVIEW_MODEL=claude-opus-4-6 swarm "Fix bug"

# Use a cheaper fast model
FAST_MODEL=gpt-5-mini swarm task "Update docs"

# Increase timeout for complex tasks
SESSION_TIMEOUT_MS=3600000 swarm "Major refactoring"

# Disable auto-resume
MAX_AUTO_RESUME=0 swarm "Risky change"