From Claude Kit
Defines orchestration protocols for multi-phase /workflow pipelines: handoffs, checkpoints, re-routing, parallel dispatch, and metrics tracking.
How this skill is triggered — by the user, by Claude, or both
Slash command
/claude-kit:workflow-protocolsThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
| Protocol | When to Load | Purpose |
| Protocol | When to Load | Purpose |
|---|---|---|
| Handoff | BEFORE forming handoff between phases | 4 pipeline contracts + 1 tool contract (code-researcher) + narrative casting |
| Checkpoint | AFTER completing each phase | Format (12 YAML fields), recovery (5 steps), state persistence |
| State Layer | On-demand (debugging state issues, Phase 5 cleanup) | File contracts, lifecycle rules, cleanup protocol for .claude/workflow-state/ |
| Re-routing | On complexity mismatch signal | Downgrade/upgrade route + tracking + learning |
| Pipeline Metrics | At completion phase only | Format (12 fields), storage (JSONL file), analysis, anomaly detection |
| Agent Memory | Agent startup + completion | Shared memory behavior for all memory: project agents |
| Parallel Dispatch | 2+ independent tasks identified (L/XL) | Multi-agent dispatch patterns (research + future debugging) |
Read this SKILL.md for the protocol overview. Then load Core Deps files: Autonomy, Orchestration Core.
Do NOT load all protocols upfront. Load on-demand per the Event Triggers section below.
CRITICAL: Every phase MUST produce a structured handoff payload (context + artifact + metadata) for the next phase.
Handoff load rule (the single source of truth — Event Triggers and Protocol References point here):
Good — full handoff payload (planner → plan-review):
artifact: ".claude/prompts/add-user-endpoint.md"
metadata:
task_type: "new_feature"
complexity: "L"
sequential_thinking_used: true
key_decisions:
- "Repository pattern over direct SQL — testability"
known_risks:
- "N+1 queries in list endpoint"
Bad — skipping handoff, jumping to code:
# No plan created, no review requested
# → jump directly to /coder
Why: Skipping phases leads to unvalidated code without architectural review. Every phase MUST produce a handoff payload for the next phase (see Handoff Protocol).
For more examples, see Examples & Troubleshooting.
Cause: Requirements unclear or task too broad. Fix: Ask user to clarify scope, break task into smaller pieces.
Cause: Connection lost, timeout, or manual stop.
Fix: Check .claude/prompts/{feature}.md for saved plan, use --from-phase to resume.
Cause: Plan missing critical sections (Scope, Architecture Decision, Tests).
Fix: Check plan against templates/plan-template.md, ensure all sections filled.
For all troubleshooting cases, see Examples & Troubleshooting.
These files define fundamental workflow behavior and are loaded at pipeline startup (step 0.1):
For detailed protocol specifications, read the supporting files in this skill directory:
memory: project agentsnpx claudepluginhub hex0xdeadbeef/claude-kit --plugin claude-kitProvides multi-agent workflow patterns: parallel dispatch, sequential pipelines, QC gates, and retry loops. Use when architecting systems with multiple agents or processing stages.
Orchestrates multi-agent workflows with parallel pipelines, sync barriers, state tracking, checkpointing, and progress metrics. Use for coordinating 3+ agents across sessions.
Executes orchestration workflows using -> || ~> syntax with real-time ASCII visualization, interactive steering at checkpoints, error recovery, and resume from .orchestration/state.json.