Swarm Protocol v2 Skill

Swarm Protocol v2

Turn complex projects into parallel agent swarms. One command to go from idea to merged PR.

Orchestrates specialized subagents across planning, implementation, and review phases with worktree isolation for conflict-free parallelization, progressive status tracking for real-time monitoring, and a multi-lens review council (via council-review) that catches architecture issues, bugs, security holes, compliance gaps, silent failures, type design issues, stale comments, and unnecessary complexity — all before you merge.

Why Swarm Protocol?

One command — /swarm-protocol <name> creates docs, branches, worktree, and starts planning
Specialized agents — Each role maps to the right subagent type (architect, explorer, reviewer, etc.)
Maximum parallelism — Agents run concurrently with worktree isolation to prevent file conflicts
Review council — Delegates review to council-review: architect, correctness, security, compliance, silent-failure, type/comment, and simplification lenses — all in parallel, synthesized to consensus
Real-time monitoring — swarm-status.json with progressive auto-updates for ScopeTUI integration
Smart resumption — Pick up where you left off with --resume, retry single agents with --agent=N.M
Plan conversion — Already have a plan? Convert it directly with --from-plan=<name>

Commands

| Command | Purpose | | --------------------------------------------- | ------------------------------------------------------- | | /swarm-protocol <project-name> | Initialize new project with full planning phase | | /swarm-protocol | Continue existing project, convert a plan, or start new | | /swarm-protocol --from-plan=<name> | Convert a specific plan file to swarm project | | /swarm-protocol --review-only | Run only the Council Review phase on current changes |

`/swarm-protocol` (No Arguments)

Smart detection: Check branch (feature/<name>) → docs/projects/ → ~/.claude/plans/ → prompt user.

Detection priority:

Current branch feature/<name> → match project in docs/projects/<name>/
List recent projects by modified date with status from CHANGELOG.md
List plans from ~/.claude/plans/ available for conversion
Offer new project with random name suggestions

Project names: Random from Barcelona region pool (rubi, sitges, mura, tiana, etc.)

Plan conversion: Parse plan → auto-populate PROJECT_PLAN.md, IMPLEMENTATION_PLAN.md, AGENT_SWARM_SPEC.md → review before execution.

`/swarm-protocol <project-name>`

Initialize and execute a multi-agent development project.

Phase 0: Project Setup

Create docs/projects/<name>/ with: PROJECT_PLAN.md, IMPLEMENTATION_PLAN.md, AGENT_SWARM_SPEC.md, CODE_REVIEW.md, CHANGELOG.md
Create worktree: git worktree add ../<repo>-<name> -b feature/<name>
Supabase preview (if DB changes): supabase branches create <name>

swarm-status.json — Progressive Auto-Update System (MANDATORY)

CRITICAL: swarm-status.json MUST be created at Phase 0 and updated automatically at every action. The ScopeTUI Swarm Monitor (scopetui --swarm) depends on this file being accurate. Stale status = broken monitoring.

Location: docs/projects/<name>/swarm-status.json inside the worktree (not master).

Full schema + progressive update protocol: See references/swarm-status-schema.md.

The Anti-Lag Rule

Update BEFORE and AFTER every Agent dispatch — never batch updates.

BEFORE dispatching agent → set agent "running" + write file
AFTER agent returns      → set agent "completed"/"failed" + write file
BEFORE phase transition  → advance currentPhaseId + write file

Quick Update Helper

# Atomic status update via jq (define once, use throughout)
swarm_update() {
  local file="$1" agent_id="$2" new_status="$3" now
  now=$(date -u +"%Y-%m-%dT%H:%M:%SZ")
  jq --arg aid "$agent_id" --arg st "$new_status" --arg now "$now" '
    (.phases[].agents[] | select(.id == $aid)) |= (
      .status = $st |
      if $st == "running" then .startedAt = $now
      elif $st == "completed" then .completedAt = $now
      else . end
    ) | .updatedAt = $now
  ' "$file" > "${file}.tmp" && mv "${file}.tmp" "$file"
}

# Usage: swarm_update docs/projects/<name>/swarm-status.json "2.1" "running"

Orchestrator Owns Status

The main session (orchestrator) writes all status updates. Subagents NEVER touch swarm-status.json. This prevents race conditions and ensures consistent state.

Minimum viable format (TUI validates only `phases` array):

{
  "project": "<name>",
  "branch": "feature/<name>",
  "currentPhaseId": "<active-phase-id>",
  "phases": [
    {
      "id": "1", "name": "Setup", "status": "completed",
      "agents": [{ "id": "1.1", "name": "Types", "status": "completed", "subagentType": "general-purpose" }]
    }
  ],
  "updatedAt": "<ISO 8601>"
}

Phase 1: Planning (Sequential — use `feature-dev:code-architect`)

Generate all planning documents before implementation. See references/templates.md for document templates.

Dispatch: Use Agent tool with subagent_type: "feature-dev:code-architect" for architecture design, and subagent_type: "feature-dev:code-explorer" for codebase analysis.

Planning sequence:

Explore — Dispatch feature-dev:code-explorer agent to analyze existing codebase patterns, trace execution paths, map dependencies
Architect — Dispatch feature-dev:code-architect agent to design implementation blueprint based on exploration results
PROJECT_PLAN.md → Goals, success criteria, dependencies, risks
IMPLEMENTATION_PLAN.md → Architecture, file changes, migrations, API endpoints
AGENT_SWARM_SPEC.md → Phase breakdown, agent definitions with subagent_type assignments, parallelization strategy

Phase 1.5: Doc Review (Sequential — gate before execution)

Mandatory review of all planning docs before any execution agent runs. Garbage plan → garbage swarm. Catch it here.

In swarm-status.json, this is a phase with id "1.5" and name "Doc Review".

Dispatch (parallel — launch ALL in a single message):

| Agent | Subagent Type | Focus | |-------|--------------|-------| | 1.5.1 Plan Reviewer | feature-dev:code-reviewer | PROJECT_PLAN goals, success criteria, dependencies, risk coverage | | 1.5.2 Architecture Reviewer | feature-dev:code-architect | IMPLEMENTATION_PLAN soundness, file change scope, migration safety | | 1.5.3 Swarm Spec Reviewer | feature-dev:code-reviewer | AGENT_SWARM_SPEC parallelization correctness, file conflict analysis, subagent_type assignments |

Each reviewer answers:

Is the doc internally consistent?
Are success criteria measurable?
Are there missing edge cases, unstated assumptions, or unverified dependencies?
Does the plan match the actual codebase (run feature-dev:code-explorer checks against claims)?
For AGENT_SWARM_SPEC specifically: do parallel agents have real file isolation, or will worktree isolation be required?

Output: Append findings to PROJECT_PLAN.md under ## Doc Review Findings section, categorized by severity (Blocker / Major / Minor).

Gate criteria:

Blockers → STOP. Loop back to Phase 1 (Planning) to revise docs. Do NOT proceed to execution.
Major → Decision required: revise docs or document accepted trade-off + rationale.
Minor → Note in PROJECT_PLAN.md, proceed.

Commit:

git commit -m "docs(<project>): doc review complete - <N> blockers, <N> major, <N> minor"

Why this phase exists: planning agents hallucinate. A reviewer pass on the plan costs minutes; executing a flawed plan costs hours. Catch architecture mismatches, missing migrations, and false parallelization claims before they cascade into broken execution.

Phase 2-N: Execution

FULL EXECUTION, MAXIMUM PARALLELIZATION

Agent dispatch: Use the Agent tool for all agent invocations. See references/agent-dispatch.md for the centralized dispatch protocol mapping roles → subagent_types.

Worktree isolation: For parallel agents that touch overlapping directories, use isolation: "worktree" on the Agent tool call. This gives each agent an isolated repo copy, eliminating file conflicts and enabling more aggressive parallelization.

Parallelization rules:

Agents within a phase run parallel when no file conflicts exist (or use worktree isolation)
Phases with dependencies wait for blockers to complete
Launch all independent agents in a single message with multiple Agent tool calls
Use run_in_background: true for agents whose results aren't needed immediately

Progress display: See references/templates.md for colored box format, ANSI codes, and ASCII art variants.

Status symbols: ● (green/complete), ◐ (yellow/in-progress), ○ (dim gray/pending)

Milestone commits (after each phase):

git add -A
git commit -m "feat(<project>): complete phase N - <description>

- Agent N.1: <summary>
- Agent N.2: <summary>

Co-Authored-By: Claude Code <noreply@anthropic.com>"

Quality gates (invoke after relevant phases):

After UI phases → invoke react-doctor skill if React project
After implementation → invoke smart-test skill for intelligent test selection
After test writing → invoke full-test-coverage skill to verify pyramid coverage

Phase Review & Repair: Council Review + Repair

Review: Invoke the council-review skill on the phase diff. Council is the shared review primitive — it runs a parallel team of specialized read-only reviewers, synthesizes their findings, resolves conflicts via a debate-to-consensus pass, and writes the verdict to CODE_REVIEW.md. Swarm does NOT inline its own reviewer dispatch anymore; council owns review, swarm owns repair. One source of truth.

In swarm-status.json, this is a single phase with id "review-repair" and name "Council Review & Repair", containing the council invocation + repair agent.

Review (invoke council-review)

# Review the current phase diff through the council.
council-review --paste   # or council-review <branch> against the phase base

The council runs these lenses in parallel (see council-review/SKILL.md for full dispatch):

| # | Lens | Subagent Type | |---|------|--------------| | C.1 | Architect | feature-dev:code-architect | | C.2 | Correctness | feature-dev:code-reviewer | | C.3 | Security | feature-dev:code-reviewer (security prompt) | | C.4 | Silent Failures | pr-review-toolkit:silent-failure-hunter | | C.5 | Types + Comments | pr-review-toolkit:type-design-analyzer | | C.6 | Simplification | pr-review-toolkit:code-simplifier | | C.7 | Compliance (HIPAA/EU) | feature-dev:code-reviewer (compliance prompt) — conditional, PHI surfaces only |

Council adds the architect and compliance lenses swarm previously lacked — swarm inherits them for free. Council writes CODE_REVIEW.md per the schema in council-review/references/code-review-schema.md; the Resolution column is left empty for the repair conductor to fill.

After council returns: CODE_REVIEW.md holds a ranked, deduplicated, conflict-resolved findings table. Repair works against it directly — no re-consolidation needed.

Repair (Sequential — after Council Review completes)

Repair steps:

High-severity issues: Fix all critical bugs, security vulnerabilities, and breaking issues
Medium-severity issues: Resolve performance problems, missing validation, incomplete error handling
Low-severity issues: Fix code style violations, naming inconsistencies, minor improvements
Recommendations (Immediate): Implement all "Before Merge" recommendations from CODE_REVIEW.md
Simplifications: Apply code-simplifier suggestions — reduce complexity while preserving functionality
Suggestions & nits: Apply code suggestions, improve readability, clean up dead code
Enhancements: Add small enhancements identified during review (better types, missing edge cases, improved UX)
Update CODE_REVIEW.md — mark all resolved issues with resolution details
Commit all repair changes

Repair rules:

Work through issues by severity: High → Medium → Low → Recommendations → Simplifications → Nits → Enhancements
Each fix must not introduce new issues
Update the Resolution column in CODE_REVIEW.md for every addressed item
If a fix requires significant refactoring, document the trade-off and proceed

Phase Verify: Final Verification

Validate that all repairs are correct and the project is ready for merge.

Verification steps:

npm run test && npm run typecheck && npm run lint
Dispatch feature-dev:code-reviewer on repair diff to confirm no regressions introduced
Verify all CODE_REVIEW.md issues marked as resolved
Update project CLAUDE.md
If Supabase preview used: Execute database deployment workflow (see below)
Final commit with verification summary

Phase Debrief: One-Page TLDR

Mandatory final phase before merge. Write a single-page debrief that any operator can read in under 2 minutes to understand the project's state.

Deliverables:

docs/projects/<name>/DEBRIEF.md — one page (≤ 80 lines, excluding diagram)
docs/projects/<name>/diagrams/<NN>-debrief-stack.svg — single B/W SVG showing phase pipeline + key metrics

Required sections (use template):

Built — 2-4 sentence prose + surface table (what was built, where the code lives)
Open issues — solution options + recommendation — every remaining issue gets ≥ 2 options + chosen option + one-sentence rationale
Next steps — numbered action list for post-merge / rollout
Cross-references — links to PROJECT_PLAN, CODE_REVIEW, runbooks

Hard rules:

One page (≤ 80 lines excluding diagram)
Every open issue: ≥ 2 options + recommendation + why
No marketing speak, no superlatives
No new architectural decisions (debrief summarizes; doesn't introduce)
Diagrams B/W only (rendered in dark-mode markdown viewers)
Cross-link every claim to canonical artifact (PROJECT_PLAN, CODE_REVIEW, etc.)

Template: references/debrief-template.md (full structure + visual style guide + anti-patterns)

Why this phase exists: at merge gate, the project owner needs ONE artifact that captures state without reading 10 docs. Debrief is the executive summary; everything else is the supporting evidence. If the debrief can't fit on one page, the project state is too vague to merge.

Options

| Flag | Purpose | | -------------------- | --------------------------------------------------------------------------------------- | | --skip-worktree | Use current branch instead of creating worktree | | --skip-preview | Skip Supabase preview branch creation | | --phases=1,2 | Run only specific phases | | --dry-run | Generate plans only, no execution | | --from-phase=N | Resume from specific phase | | --agent=N.M | Run single agent only | | --from-plan=<name> | Convert a specific plan file to project (e.g., --from-plan=scalable-forging-lovelace) | | --resume | Auto-detect last incomplete agent and resume from there | | --max-agents=N | Limit concurrent parallel agents (default: 4) | | --graph | Output agent dependency graph in Mermaid format | | --review-only | Run only the Council Review phase on current changes (no planning/execution) | | --isolate | Force worktree isolation for all parallel agents |

Supabase Database Deployment

If project uses Supabase preview branches, see references/supabase-deployment.md for the full deployment workflow including migration validation loops, advisor checks, and production deployment steps.

Quick ref: supabase db push --linked → fix errors → run advisors → consolidate → deploy to production.

Agent Dispatch Protocol

CRITICAL: Always use the correct subagent_type when dispatching agents. See references/agent-dispatch.md for the complete mapping.

Quick reference:

| Agent Role | Subagent Type | When to Use | |-----------|--------------|-------------| | Codebase exploration | Explore | Understanding existing code, finding patterns | | Architecture design | feature-dev:code-architect | Planning implementation, designing components | | Deep codebase analysis | feature-dev:code-explorer | Tracing execution paths, mapping dependencies | | Code review | feature-dev:code-reviewer | Reviewing for bugs, security, conventions | | Silent failure hunting | pr-review-toolkit:silent-failure-hunter | Finding swallowed errors, bad fallbacks | | Type design review | pr-review-toolkit:type-design-analyzer | Reviewing type encapsulation, invariants | | Comment review | pr-review-toolkit:comment-analyzer | Checking comment accuracy, staleness | | Code simplification | pr-review-toolkit:code-simplifier | Reducing complexity, improving clarity | | Test coverage analysis | pr-review-toolkit:pr-test-analyzer | Reviewing test completeness | | General implementation | general-purpose | Default for coding agents | | Planning | Plan | Designing implementation strategies |

Dispatch rules:

Single message, multiple agents: Launch all independent agents in ONE message with multiple Agent tool calls
Background when possible: Use run_in_background: true for agents whose output isn't needed to proceed
Worktree isolation: Use isolation: "worktree" when parallel agents may touch overlapping files
Resume capability: Store agent IDs to resume failed/interrupted agents via the resume parameter

Agent Design Principles

Parallelization: Safe when different files/dirs, independent features, tests for completed work. Sequential for dependency chains (Schema→Types→API), base→composite components, core→integration. Use isolation: "worktree" to unlock parallelism when agents share directories.

Agent scope: Single responsibility, own specific files (no overlap), define completion criteria, include validation.

Communication: IMPLEMENTATION_PLAN.md + type definitions (shared context), file system (deliverables), CHANGELOG.md (progress).

Agent Failure Protocol

| Failure | Action | | -------------------- | ------------------------------------------------------------------- | | Timeout (>15min) | Log to CHANGELOG, mark blocked, continue non-dependent, prompt user | | Error | Log, retry once, if fails pause and prompt, block dependents | | File Conflict | Halt agents, alert user, request resolution |

Recovery: /swarm-protocol --resume (auto-resume) | /swarm-protocol-protocol --agent=2.3 --retry | /swarm-protocol-protocol --skip-agent=2.3

Project Archival

After merge: move docs to docs/archive/, remove worktree, delete feature branch, clean Supabase preview if used.

mv docs/projects/<name> docs/archive/<name>
git worktree remove ../<repo>-<name>
git branch -d feature/<name>
supabase branches delete <name>  # if applicable

Best Practices

Plan thoroughly → Use specialized subagent types → Keep agents focused → Define interfaces first → Launch parallel agents in single messages → Commit per phase → Review via council → Repair all issues → Verify before merge → Document decisions in CODE_REVIEW.md

Reference Files

| File | Load When | | ----------------------------------- | ---------------------------------- | | references/agent-dispatch.md | Always (subagent type mapping) | | references/templates.md | Always (document templates) | | references/agent-patterns.md | Always (agent configurations) | | references/supabase-deployment.md | DB migrations present | | references/progress-art.md | Display customization |

Agent Skills: Swarm Protocol v2

Install this agent skill to your local

Skill Files