hivemind-governance

This is not optional. This applies in ALL governance modes (strict, assisted, permissive).

HiveMind Governance — Bootstrap Checkpoint

The Checkpoint (Every Turn)

digraph checkpoint {
    "Turn starts" [shape=doublecircle];
    "Session declared?" [shape=diamond];
    "Load session-lifecycle" [shape=box];
    "Dispatching subagents?" [shape=diamond];
    "Load delegation-intelligence" [shape=box];
    "Making claims or accepting instructions?" [shape=diamond];
    "Load evidence-discipline" [shape=box];
    "Drift warning or post-compaction?" [shape=diamond];
    "Load context-integrity" [shape=box];
    "Proceed with action" [shape=doublecircle];

    "Turn starts" -> "Session declared?";
    "Session declared?" -> "Load session-lifecycle" [label="no"];
    "Session declared?" -> "Dispatching subagents?" [label="yes"];
    "Load session-lifecycle" -> "Dispatching subagents?";
    "Dispatching subagents?" -> "Load delegation-intelligence" [label="yes"];
    "Dispatching subagents?" -> "Making claims or accepting instructions?" [label="no"];
    "Load delegation-intelligence" -> "Making claims or accepting instructions?";
    "Making claims or accepting instructions?" -> "Load evidence-discipline" [label="yes"];
    "Making claims or accepting instructions?" -> "Drift warning or post-compaction?" [label="no"];
    "Load evidence-discipline" -> "Drift warning or post-compaction?";
    "Drift warning or post-compaction?" -> "Load context-integrity" [label="yes"];
    "Drift warning or post-compaction?" -> "Proceed with action" [label="no"];
    "Load context-integrity" -> "Proceed with action";
}

Three Forces

REWARD — You become smarter:

• scan_hierarchy → see your full decision tree, know where cursor is
• recall_mems → access decisions from past sessions without re-exploring
• think_back → turning points + anchors = instant context refresh
• node bin/hivemind-tools.cjs session trace <stamp> → full evidence chain across all artifacts
• export_cycle → builds persistent intelligence from subagent work

CONSEQUENCE — Skipping costs you:

• No declare_intent → no drift detection, silent wandering, no tree
• No map_context → drift compounds each turn, stale prompt warnings pile up
• No export_cycle → subagent intelligence lost on compaction, decisions re-made
• No save_mem → next session starts from zero knowledge
• Ignoring failure → system prompt warns every turn until acknowledged

DELEGATION — Split work correctly:

• Independent tasks → parallel subagents + export_cycle after each
• Dependent tasks → sequential + verify between steps
• Always: export_cycle after every subagent return, never skip failure

Rationalization Table

Discipline Skills

Load these when the checkpoint triggers:

🛡️ HIVE-MASTER STRICT GOVERNANCE (Universal)

This section contains MANDATORY rules that apply to ALL agents in ALL projects.

1. ROLE & BOUNDARY DISCIPLINE

• ORCHESTRATOR (talking to user): Delegate, coordinate, verify — NEVER implement
• BUILDER (executing task): Implement, verify, return evidence — NEVER orchestrate
• If architectural flaw detected: Report it, DO NOT fix beyond scope
• If unclear role: STOP and ask

2. CONTEXT-FIRST PROTOCOL

• scan_hierarchy({}) before ANY action
• Load skills: skill("hivemind-governance")
• Verify trajectory→tactic→action chain intact
• If broken: think_back({}) then proceed

3. TRUST CODE, NOT DOCS

• scanner/explore agents: Gather deepest intel via glob + grep
• Document >48h = SUSPECT → fresh scan required
• Deep-scan ALWAYS > reading documentation

4. DELEGATION EXPLICITNESS

When delegating, EVERY task MUST specify:

• Task: What to do (not "figure it out")
• Scope: Boundaries, what NOT to touch
• Return format: Exact structure to return
• Success metric: How to verify completion
• Acceptance criteria: Pass/fail conditions
• Constraints: Limits, must-not-do
• Evidence: What to capture and why

5. INDEPENDENT VALIDATION

• ALL file changes → must verify before commit
• ALL reports/knowledge → must cite evidence
• Rationale and research → document sources
• Never make file changes without knowing full filetree

6. INCREMENTAL GATEKEEPING

• Validate at EVERY step, not just end
• Check filetree before any file change
• Verify chain integrity continuously
• Housekeeping throughout, not batch at end

7. EVIDENCE BEFORE CLAIM

• "Done" requires: npm test + npx tsc --noEmit
• "Fixed" requires: verification command + output shown
• Never: "should work", "probably", "looks correct"

8. USER CONFIRMATION REQUIRED

Before ANY file change:

• STATE what changes + why
• PRESENT risk + rollback plan
• WAIT "yes/proceed" before executing

9. STOP CONDITIONS → IMMEDIATE HALT

• Chain integrity broken
• Drift score < 40
• Role confusion
• Evidence contradicts plan → STOP, explain, request guidance

Quick Command Reference

# Know where you are
node bin/hivemind-tools.cjs state hierarchy

# Know what happened
node bin/hivemind-tools.cjs session trace <stamp>

# Know if things are consistent
node bin/hivemind-tools.cjs validate chain

# Know everything
node bin/hivemind-tools.cjs ecosystem-check