hive

name	hive
description	Coordinate a governed hive of specialized agents for broad research, planning, review, or implementation work when roles, blackboard state, dissent, synthesis, and verification materially improve the outcome.
argument-hint	["task"]
disable-model-invocation	true

/hive

Coordinate a governed multi-agent hive for:

$ARGUMENTS

/hive is a coordination mode, not a new execution engine.

It sits above /parallel and adds role assignment, blackboard discipline, dissent, synthesis, and arbitration. When execution packets need workers, /hive reuses /parallel packet DAGs, ownership matrices, sidecars, and aggregation instead of inventing a second worker format.

Command Boundary

• /metacognition: decides whether hive coordination is useful.
• /hive: coordinates multiple specialized agents around shared evidence.
• /parallel: executes independent packets with ownership and aggregation.
• /hiveworkflow: runs the full file-changing lifecycle with hive coordination.
• /agent-view: optional operator-managed visibility for independent Claude Code background sessions. It is not hive coordination and does not satisfy blackboard, dissent, aggregation, or verification requirements.
• /goal-mode: static readiness and safe-template guidance for native /goal. Native /goal can continue an already-owned packet, but it is not hive coordination and does not satisfy blackboard, dissent, synthesis, aggregation, or verification requirements.
• /workflow: remains the default single-supervisor lifecycle.
• /introspect: remains the hard-gate audit. Hive consensus never replaces it.

Use /hive when a task benefits from multiple independent perspectives, research branches, adversarial reviews, or role-specialized work packets. Downgrade to /workflow or local work when the task is tightly coupled.

Hive Vs Parallel

Pick /parallel when the problem is mostly execution throughput:

• independent packets
• owned files or surfaces
• sync barriers
• worker sidecars
• aggregate verification

Pick /hive when the problem is mostly judgment breadth:

• independent roles
• blackboard claims
• dissent or skeptic lanes
• synthesis and arbitration
• evidence-backed consensus checks

Pick /hiveworkflow when the task needs both coordinated judgment and the full file-changing lifecycle. If the work only needs owned implementation packets, use /parallel. If it only needs one supervisor loop, use /workflow.

For file-changing hive work, prefer /opusworkflow with inner_contract=hiveworkflow. Use /hive directly for read-only coordination, research, dissent, or synthesis where no files are changed.

Non-Negotiable Contract

• The queen/supervisor owns route, scope, safety, final synthesis, and closeout.
• Agents receive bounded roles, evidence requirements, stop conditions, and owned files/surfaces when they may edit.
• The hive uses a visible blackboard artifact for claims, evidence, conflicts, decisions, locks, and open questions.
• Consensus is not evidence. Agreement only becomes usable when supported by source, repo, command, runtime, or verification evidence.
• Dissent is required for non-trivial decisions. If no dissent appears, run an explicit skeptic lane before synthesis.
• Shared-state edits require ownership, locks, or merge barriers.
• Worker and peer summaries are claims until the supervisor verifies them.
• MAX_PARALLEL_AGENTS, Codex max_threads, and hardware ceiling are ceilings, not quotas.
• Agent teams are opt-in experimental only when capacity says they are available and the operator explicitly enables them.

Hive Roles

Use the smallest useful role set:

Role	Purpose	Typical Output
queen	supervisor, route owner, final arbiter	synthesis and decision log
scout	independent research or repo discovery	source/evidence ledger
builder	bounded implementation packet	changed files and command evidence
reviewer	adversarial critique	risks, blockers, missing tests
verifier	independent proof lane	commands, runtime proof, pass/fail
scribe	blackboard and artifact keeper	durable hive artifact

Do not create a role without a concrete evidence return.

Capacity And Budget

Every hive run must read or cite:

bash scripts/parallel-capacity.sh --json 2>/dev/null || true

Compute:

effective_hive_budget =
  min(MAX_PARALLEL_AGENTS, codex_max_threads, hardware_recommended_ceiling,
      independent_roles_or_packets, supervisor_review_capacity,
      verifier_capacity, blackboard_merge_capacity)

Use:

• 1: local work; hive not useful.
• 2-3: scout + reviewer, or scout + verifier.
• 4-6: scouts/builders/reviewer/verifier with clear ownership.
• 7+: broad audits or dense DAGs only when blackboard and verification can absorb the results.

Never claim linear speedup or that a larger hive is inherently smarter.

Blackboard Artifact

Create a durable artifact for non-trivial hive runs:

mkdir -p .taste/workflow-runs
STAMP="$(date +%Y%m%d-%H%M%S)"
HIVE_ARTIFACT=".taste/workflow-runs/${STAMP}-hive.md"

Required sections:

# Hive Run: [task]

## Task
## Metacognitive Route
## Capacity Profile
## Hive Eligibility
## Role Map
## Blackboard
## Dissent And Conflict Log
## Synthesis And Arbitration
## Packet DAG
## Verification Evidence
## Introspection
## Outcome

## Blackboard must include:

Claim ID	Owner	Claim	Evidence	Status	Conflicts	Lock/Merge Barrier

Status values: candidate, verified, rejected, blocked.

Machine-Readable Sidecar

When a hive run is used for planning, implementation, verification, or durable learning, also write:

.taste/hive/{run_id}/hive-run.json

Validate it before closeout:

bash scripts/artifact-lint.sh .taste/hive/{run_id}/hive-run.json
bash scripts/hive-aggregate.sh .taste/hive/{run_id}

The sidecar records role_map, blackboard_claims, capacity, dissent_log, synthesis, consensus_policy, and verification.

Hive Eligibility

Use hive only when all pass:

• independent roles or packets exist
• supervisor can review returned evidence
• blackboard can preserve provenance and conflict state
• verification can aggregate the result
• role prompts can be thin and current
• shared-state ownership or locks can be written

Downgrade when:

• one tight reasoning loop dominates
• all agents need the same context and files
• disagreement cannot be resolved with evidence
• tool or credential state is shared and risky
• verification depends on trusting peer summaries

Coordination Protocol

1. Queen writes the route, budget, roles, blackboard schema, and stop conditions.
2. Scouts/reviewers/builders work in bounded lanes.
3. Scribe updates the blackboard from returned evidence.
4. Queen runs a sync barrier and checks stale context, conflicts, and locks.
5. If consensus appears, queen still checks evidence and dissent.
6. If conflict remains, queen arbitrates using evidence or downgrades/block.
7. Verifier proves the aggregate result.
8. /introspect attacks the synthesis before closeout.

Relationship To /parallel

When hive work changes files or executes worker packets, reuse /parallel:

• packet DAG
• ownership matrix
• sync barriers
• .taste/parallel/{run_id}/packet-dag.json
• .taste/parallel/{run_id}/ownership.json
• .taste/parallel/{run_id}/worker-results/*.json
• scripts/parallel-plan-lint.sh
• scripts/parallel-aggregate.sh
• scripts/artifact-lint.sh
• scripts/hive-aggregate.sh

Hive adds cognition and coordination; /parallel remains the execution and aggregation substrate.

Claude Code Agent View can help an operator monitor independent background sessions used outside the hive or as manually managed parallel-instances, but the hive still needs its own queen/supervisor, blackboard, dissent log, sidecars, aggregation, and verification evidence. Agent View rows are operator status, not verified hive claims.

Native /goal can make one Claude Code session persistent inside an already owned packet, but it does not create hive memory, blackboard state, dissent, or verified synthesis. Treat every /goal completion as a claim until the queen/supervisor verifies command evidence and repo state.

Output

## Hive Decision
- Decision: local / hive-research / hive-review / hive-implementation / blocked
- Effective hive budget: N of ceiling M
- Roles: ...
- Blackboard: [path or not needed]
- Execution substrate: local / subagents / parallel-instances / agent-teams-experimental
- Dissent required: yes / no, reason
- Verification path: ...
- Route: /workflow / parallel / hiveworkflow / direct answer / blocked
- Confidence: high / medium / low, downgrade reason

Anti-Patterns

• "Hive mind" as branding for unowned parallel work.
• Consensus without evidence.
• Majority vote as verification.
• Shared file edits without locks or merge barriers.
• Agents reading one another's unverified summaries as truth.
• Using every lane because the ceiling exists.
• Letting the hive replace /introspect, /verify, /workflow, or /parallel.