Hydra Swarm

hydra-swarm — Multi-Instance Coordination

What This Does

Coordinates multiple Claude Code instances across the Swarm cluster using NFS-backed shared state at /opt/swarm/. Provides awareness of what other instances are doing, task queuing, messaging, artifact sharing, task decomposition, worktree isolation, and session summary context sharing.

Quick Reference

Check swarm status before starting work

python3 <coordination-tool-root>/src/swarm_cli.py status

Create a task for another instance

python3 <coordination-tool-root>/src/swarm_cli.py tasks create "Task title" \
  --desc "Description" --project <your-project-path> --priority high \
  --requires gpu --minutes 30

Claim a task

python3 <coordination-tool-root>/src/swarm_cli.py tasks claim task-001

Claim with worktree isolation (conflict-free editing)

python3 <coordination-tool-root>/src/swarm_cli.py tasks claim task-001 --isolate --project <your-project-path>

Complete a task

python3 <coordination-tool-root>/src/swarm_cli.py tasks complete task-001 --artifact results.md

Complete + merge worktree to main

python3 <coordination-tool-root>/src/swarm_cli.py tasks complete task-001 --merge --project <your-project-path>

Complete + push branch only (another instance merges)

python3 <coordination-tool-root>/src/swarm_cli.py tasks complete task-001 --branch --project <your-project-path>

Task Decomposition — suggest splitting a large task

python3 <coordination-tool-root>/src/swarm_cli.py tasks decompose task-001

Task Decomposition — apply the suggested split

python3 <coordination-tool-root>/src/swarm_cli.py tasks decompose task-001 --apply

Send a message to another instance

python3 <coordination-tool-root>/src/swarm_cli.py message <gpu-host> "Need the RAG results from <your-project>"
python3 <coordination-tool-root>/src/swarm_cli.py message --broadcast "Deploying v2 in 5 min"

Check inbox

python3 <coordination-tool-root>/src/swarm_cli.py inbox

Share an artifact

python3 <coordination-tool-root>/src/swarm_cli.py artifacts share /path/to/results.md

List active worktrees across fleet

python3 <coordination-tool-root>/src/swarm_cli.py worktrees --project <your-project-path>

View session summaries

python3 <coordination-tool-root>/src/swarm_cli.py summaries
python3 <coordination-tool-root>/src/swarm_cli.py summaries --project <repo-path>/a

Show what the last instance left for you

python3 <coordination-tool-root>/src/swarm_cli.py context --project <repo-path>/a

Health check

python3 <coordination-tool-root>/src/swarm_cli.py health

Force git sync

python3 <coordination-tool-root>/src/swarm_cli.py sync

Task Decomposition Workflow

1. Create a large task: tasks create "Generate 50 CPA questions across all sections"
2. Suggest decomposition: tasks decompose task-001 — shows subtasks based on rules
3. Apply if acceptable: tasks decompose task-001 --apply — creates subtask files, parent moves to decomposed state
4. Subtasks can be claimed independently by different nodes
5. When ALL subtasks complete, parent auto-completes

Decomposition rules:

• "all sections" or "across" + CPA sections -> split per FAR/AUD/REG/BAR + validation
• "generate AND validate" -> split into generate + validate phases
• Multiple project names mentioned -> split per project
• Capabilities auto-inferred: ollama/gpu/docker keywords tag subtasks for routing

Worktree Isolation

• --isolate on claim creates a git worktree branch swarm/<hostname>/<task-id>
• Each instance works in its own branch, no merge conflicts
• On completion: --merge merges to main, --branch pushes branch only
• Worktree paths and branches recorded in task YAML and artifacts

Session Summaries for Context Continuity

• On session end, a summary is auto-generated with files changed, decisions, context
• On session start, the hook loads the last relevant summary and injects into systemMessage
• This gives the new instance immediate context without re-exploring the project
• context command shows what the previous instance left for you

v2 Autonomous Behavior Rules

• Agents auto-register on session start, deregister on end
• Heartbeat every 60s — stale after 5 min, tasks auto-released
• Auto-claim highest-priority matching task (by capability)
• Emit events on commit, test, task completion — other agents consume
• Auto git pull when commit events arrive from other hosts
• GPU slot claiming prevents model contention
• Session summaries auto-generated from event stream
• Conflict detection: warn if two agents editing same project
• This is NOT advisory — agents act autonomously within the priority framework

v2 Modules (in /src/)

Session Protocol

from session import start_session, end_session

# On session start:
info = start_session(model="opus-4.6", project="<repo-path>")
# → registers agent, starts heartbeat, pulls repos, reads event catchup

# During work: events auto-emitted on commit/test/task completion

# On session end:
result = end_session()
# → pushes repos, syncs config, writes summary, deregisters

Node Map

Task States

• pending — available for claiming
• claimed — being worked on by a specific host
• completed — done
• decomposed — split into subtasks (auto-completes when all subtasks done)

Node States

• active — instance running, available for coordination
• idle — no active session
• offline — no heartbeat for >5 minutes
• busy — active but do not interrupt

File Locations

• Project: <coordination-tool-root>/
• NFS share: <shared-coordination-root>/ (mounted from primary or replica host)
• Config: <coordination-tool-root>/config/swarm.yaml
• CLI: <coordination-tool-root>/src/swarm_cli.py