{
  "markdown": "# Hive v4.0: Bio-Inspired Swarm Orchestrator for Claude Code\n\nA multi-agent orchestration skill that coordinates parallel Claude Code agents using mechanisms from ant colonies, honeybee swarms, and AI research.\n\n## What It Does\n\n`/hive <task>` breaks your task into subtasks, spawns agents in parallel waves, coordinates them through shared findings, resolves conflicts, and learns from each run.\n\n## Skills in This Plugin\n\n| Skill | What it does |\n|-------|--------------|\n| `/hive` | Parallel swarm orchestration for execution tasks: research, fixes, refactors, audits. |\n| `/hiverev` | Persona-driven review of any artifact from any angle: a UI, a landing page, a cold email, a README, a pitch deck, a flow. Pick who is looking and what lens they bring, get a prioritized friction punch list. Runs in-context, no sub-agents. |\n| `/hivesim` | Strategic wargame engine. 7 phases: scout, decompose, quantitative buyer sim (10-15 personas), deep buying journeys, 3-adversary red team with FATAL verification, pre-mortem, and a one-page Go/No-Go Decision Brief with kill criteria. |\n\n## What's New in v4.0 (April 8, 2026)\n\n- **Concurrency unlocked**: Old limit was 2 heavy agents (from a March 2026 incident). Stress-tested to 25 read / 20 write / 15 mixed agents with zero failures across 161 launches. Default start concurrency raised from 5 to 10.\n- **Mechanism audit**: Honest assessment of all 27 mechanisms. 10 fire regularly, 17 are theoretical/broken. Dead mechanisms documented, not removed (platform may support them in future).\n- **Cross-skill sync**: Updated concurrency limits across /hivesim, /sim, /gtm, /hivefront. All skills now use proven ceilings.\n- **Hivesim personas**: Now spawn 10-15 personas per wave instead of 2. ~4x faster wall clock time for strategic analysis.\n\n## Performance (April 2026 benchmarks)\n\n| Scenario | Agents | Wall Time | Speedup vs Sequential |\n|----------|--------|-----------|----------------------|\n| 10-file audit | 10 haiku | ~60s | 8x |\n| 15-fix code wave | 2 opus + 4 sonnet + 9 haiku | ~200s | 4.4x |\n| 109-file catalog | 5 waves x 25 haiku | ~3 min | 18x |\n| 20-file write batch | 20 haiku | ~73s | 10x |\n\n## Proven Concurrency (April 8, 2026 test suite: 78 agents, 10 experiments, 0 errors)\n\n| Config | Agents | Result | Avg Latency |\n|--------|--------|--------|-------------|\n| 5 haiku | 5 | PASS | 5.4s |\n| 10 haiku | 10 | PASS | 9.2s |\n| 4 sonnet | 4 | PASS | 9.7s |\n| 2 opus | 2 | PASS | 11.9s |\n| 10h + 3s | 13 | PASS | h:7.6s / s:12.9s |\n| 8h + 1o | 9 | PASS | h:5.5s / o:10.5s |\n| 5h + 2s + 1o | 8 | PASS | h:4.9s / s:7.1s / o:10.7s |\n\n**Key findings:**\n- No hard 529 wall found up to 13 mixed agents\n- Model tiers don't conflict (haiku/sonnet/opus share the same pool)\n- 10+ agents get queued ~2 min before executing (Claude Code internal queue)\n- 529 errors are transient API load spikes. Always retry with 30s delay.\n- Safe default: 13 mixed agents. Flat count, not weighted.\n\n## Installation\n\nIn Claude Code:\n\n```\n/plugin marketplace add CipherandRow/claude-hive-plugin\n/plugin install hive@cipherandrow\n```\n\nOr manually copy any skill folder from `skills/` into your `~/.claude/skills/` directory.\n\n## Usage\n\n```\n/hive research the top 5 competitors and compare pricing\n/hive fix all failing tests in src/\n/hive --dry-run refactor the auth module\n/hive --resume\n/hive --isolate update all API endpoints to v2\n```\n\n### Flags\n\n| Flag | Description |\n|------|-------------|\n| `--resume` | Resume from last checkpoint |\n| `--isolate` | Force git worktree isolation for all agents |\n| `--dry-run` | Show execution plan without launching agents |\n| `--verbose` | Show mechanism trace after each wave (default: on) |\n| `--quiet` | Suppress execution trace |\n\n## How It Works\n\n1. **Mode Detection**: Counts subtasks, selects Lite/Standard/Full mode\n2. **Strategy Selection**: Reads past run history, picks best strategy\n3. **Planning**: Breaks task into waves with dependency ordering and reserve pool\n4. **Pre-flight**: Context ceiling check, confidence calibration, agent TTL calculation\n5. **Execution**: Spawns agents in waves with stigmergy (shared findings board), velocity scaling, and error handling\n6. **Synthesis**: Merges results using decision protocols (vote, consensus, or AAD)\n7. **Learning**: Appends scored results to history, updates playbook with time-decay\n\n## Mechanisms\n\n### Active (fire regularly, proven in production)\n\n| # | Mechanism | Origin | Purpose |\n|---|-----------|--------|---------|\n| 2 | Self-Validation Gates | Leaf-cutter ants | Agents validate own output before returning |\n| 3 | Reasoning Tree Conflicts | AgentAuditor | Find exact divergence point in reasoning |\n| 4 | Stigmergy | Ants | Shared findings for indirect coordination |\n| 6 | Semantic Quorum | Honeybees + LLM | Early commitment via semantic similarity |\n| 7 | Scout Retirement (TTL) | Honeybees | Kill stuck agents proactively |\n| 11 | Cross-Inhibition | Honeybee stop signals | Competing proposals dampen each other |\n| 13 | Assembly Line QC | Leaf-cutter ants | Pipeline handoff quality gates |\n| 15 | Adaptive Mode | Ant response thresholds | Auto-scale mechanisms to task size |\n| 17 | Auto-Verify Gate | Manufacturing QC | Run tsc + tests after code-writing waves |\n| 27 | Debug Trace | Distributed tracing | JSONL event log for every decision |\n\n### Theoretical (defined but rarely/never fire in current Claude Code)\n\n| # | Mechanism | Why Inactive |\n|---|-----------|-------------|\n| 1 | Pheromone Evaporation | History is read but scoring never changes agent routing |\n| 5 | Completion Velocity | Concurrency is now fixed at proven tiers, not dynamically scaled |\n| 8 | Decision Protocols | Vote/consensus/AAD never explicitly activates as distinct step |\n| 9 | Swarm Playbook | No evidence of playbook influencing future runs |\n| 10 | Ready-Up Signal | Pre-flight exists but not as a distinct mechanism |\n| 12 | Inspector Agents | Never fires in production |\n| 14 | Checkpoint/Resume | Defined but never used across 50 runs |\n| 16 | Worktree Isolation | Broken on Windows (file delivery fails on cleanup) |\n| 18-26 | Advanced mechanisms | Platform doesn't yet support (JSON schema enforcement, dynamic tool loading, hot reassignment, etc.) |\n\nThese remain in the skill definition for forward compatibility. As Claude Code and the Agent SDK evolve, more mechanisms may activate.\n\n## Model Selection (data-driven)\n\n| Model | Avg Duration | Avg Tool Calls | Best For |\n|-------|-------------|----------------|----------|\n| Haiku | 7-29s | 1-5 | File reads, grep, single-file fixes, summaries |\n| Sonnet | 85-95s | 20-25 | Multi-file fixes, code review, medium complexity |\n| Opus | 94-115s | 10-12 | Complex rewrites, architecture decisions, synthesis |\n\n## Prompt Discipline (proven by A/B test)\n\nStrict prompts beat loose prompts in every dimension:\n- 44% faster per-agent execution\n- 10x less code bloat\n- 0 scope violations vs 1\n\nAlways include in write agent prompts:\n```\nONLY EDIT THIS FILE: [path]. DO NOT edit any other file.\nDO NOT add inline comments, refactor, or change code outside the task.\n```\n\n## Requirements\n\n- Claude Code CLI\n- Git (for worktree isolation, optional)\n\n## Version History\n\n| Version | Date | Changes |\n|---------|------|---------|\n| v4.2 | July 23, 2026 | Added /hiverev (persona-driven review of any artifact) and /hivesim (strategic wargame engine) skills |\n| v4.1 | April 9, 2026 | Full concurrency test suite (78 agents, 0 errors), flat ceiling model replaces weighted tokens |\n| v4.0 | April 8, 2026 | Concurrency unlocked (2->25), mechanism audit, cross-skill sync |\n| v3.0 | March 25, 2026 | 14 mechanisms, semantic quorum, reasoning tree, A2A protocol |\n| v2.0 | March 2026 | Adaptive mode, worktree isolation, stigmergy |\n| v1.0 | March 2026 | Basic parallel agent orchestration |\n\n## License\n\nMIT\n",
  "bytes": 7808,
  "sha": "e90915be5f1ca1330be8b8faa297ae3b5354df9c155f5d9a9cd1901bbe06fad8",
  "repo_slug": "cipherandrow/claude-hive-plugin",
  "fonte": "repo",
  "truncated": false,
  "api": "https://agentalog.com/api/listings/plg_cipherandrow_claude_hive_plugin_hive_eafb612b/readme"
}