Context Integrity

Core principle: Detect loss. Repair state. Never work on stale context.

When to Use

• Drift score warning appears in <hivemind> prompt
• Post-compaction (context was just compressed)
• Resuming work after a gap (hours/days between sessions)
• Hierarchy feels wrong ("what was I doing?")
• Multiple compactions in same conversation
• Switching from one subagent's context back to main flow

Detect — Know When Context Is Broken

Automatic Detection (hooks do this for you)

The <hivemind> block in your system prompt already contains:

• Drift score — drops below 50 = you're wandering
• Turn count — 5+ turns without map_context = stale
• Timestamp gaps — hours between hierarchy nodes = context likely lost
• Post-compaction alert — context was just compressed
• Failure flag — subagent reported failure, not yet acknowledged
• Tool-hierarchy mismatch — writing files but no action declared

You don't need to detect these yourself. Read the <hivemind> block.

Manual Detection (when you suspect problems)

# Full hierarchy tree with cursor position
node bin/hivemind-tools.cjs state hierarchy

# Validate parent-child chain integrity
node bin/hivemind-tools.cjs validate chain

# Check all timestamps parse correctly
node bin/hivemind-tools.cjs validate stamps

# See gap between last activity and now
node bin/hivemind-tools.cjs state get metrics.turn_count

# Check if compaction happened
node bin/hivemind-tools.cjs state get compaction_count

Repair — Fix Broken Context

After Drift Warning

// 1. Check where you are
scan_hierarchy({})
// → Shows full tree with cursor marker

// 2. Update to re-engage
map_context({
  level: "action",    // or tactic if action is wrong too
  content: "What you're actually doing right now",
  status: "active"
})
// → Resets turn count, boosts drift score

After Compaction

// 1. Refresh context from persistent stores
think_back({})
// → Shows: tree position + anchors + chain analysis + session plan

// 2. Check what mems survived
recall_mems({ shelf: "decisions" })
recall_mems({ shelf: "cycle-intel" })
// → Key decisions and subagent intelligence from before compaction

// 3. Re-declare if trajectory feels wrong
// Only if think_back shows misalignment
map_context({
  level: "tactic",
  content: "Continuing: [what think_back showed as active]"
})

After Session Gap (Hours/Days)

# 1. What session was active?
node bin/hivemind-tools.cjs session active

# 2. What's the full tree?
node bin/hivemind-tools.cjs state hierarchy

# 3. Trace the last session's work
node bin/hivemind-tools.cjs session trace <stamp>

# 4. What mems exist?
recall_mems({})
recall_mems({ query: "decisions" })

Then start a new session:

declare_intent({
  mode: "plan_driven",
  focus: "Continuing: [what trace showed]",
  reason: "Resuming after gap — checked trace + mems"
})

After Chaos (User Changed Mind Multiple Times)

// 1. See the FULL tree — shows every pivot as a node
scan_hierarchy({})

// 2. Save an anchor at the stable point
save_anchor({
  type: "decision",
  content: "After 3 pivots, settling on: [X]. Reasons: [Y].",
  tags: ["pivot-resolution"]
})

// 3. Prune completed/abandoned branches
hierarchy_manage({ action: "prune" })

// 4. Re-set trajectory clearly
map_context({
  level: "trajectory",
  content: "Final direction: [X]",
  status: "active"
})

Survive — Patterns for Long-Haul Sessions

The Anchor Pattern (Compaction-Proof Decisions)

Before making any significant decision:

save_anchor({
  type: "decision",
  content: "Chose JWT over session tokens because: stateless, scales horizontally",
  tags: ["auth", "architecture"]
})

This survives compaction. think_back will show it.

The Mem Pattern (Session-Proof Knowledge)

When you learn something that future sessions need:

save_mem({
  shelf: "architecture",
  content: "Auth uses jose library for JWT, passport was abandoned at turn 15",
  tags: ["auth", "jwt", "jose"]
})

This survives session boundaries. recall_mems from any future session.

The Checkpoint Pattern (Drift-Proof Progress)

Every 3-5 meaningful actions:

map_context({
  level: "action",
  content: "Completed: X, Y, Z. Next: W.",
  status: "active"
})

This resets drift counter. Keeps the tree current.

Red Flags