Context Gate (Unified Enforcement Hook)

PARAMOUNT: Unified enforcement hook with ACTUAL BLOCKING for multiple gates.

Status

• ENFORCED: PreToolUse hook blocks tool calls via permissionDecision: "deny"
• PARAMOUNT Level: Same priority as ut++ and asking-users
• Unified: Single hook handles Context, Dependencies, Git Safety, and Secrets

Enforcement Mechanism (Gen 229-230)

THIS IS REAL ENFORCEMENT: Hook returns permissionDecision: "deny" to BLOCK execution.

Gates Implemented

Gate	Tools Affected	Trigger	Action
1. Dependency Guardian	Read, Glob, Grep	Blocked paths (node_modules, etc.)	BLOCK
2. Destructive Git	Bash	force push, hard reset, clean -fd	BLOCK
3. Secrets Detection	Bash	git add/commit with .env, keys	BLOCK
4. Skill Creation	Write, Edit	/skills/*.md	Warn (soft)
5. Context Gate	All (except AskUserQuestion)	80% threshold	BLOCK

Blocked Dependency Paths

BLOCKED_PATHS = [
    "node_modules", ".next", "dist", "build", "__pycache__",
    "venv", ".venv", "target", "vendor", "Pods", ".git/objects",
    ".nuxt", ".output", "coverage", ".cache"
]

Destructive Git Patterns

• git push --force / git push -f
• git reset --hard
• git clean -fd
• git checkout -- . (discard all changes)

Sensitive File Patterns

• .env, .env.*
• credentials.json, secrets.json
• .pem, .key, id_rsa, id_ed25519

Configuration

Settings in ~/.claude/settings.json:

{
  "hooks": {
    "PreToolUse": [{
      "matcher": ".*",
      "hooks": [{
        "type": "command",
        "command": "python3 ~/.claude/hooks/context-gate-hook.py"
      }]
    }]
  }
}

Context Thresholds

Threshold	%	Tokens	Action
WARNING	70%	140,000	Allow + systemMessage warning (once)
CRITICAL	80%	160,000	BLOCK all tools except AskUserQuestion

Configure via environment variables:

• CONTEXT_GATE_MAX (default: 200000)
• CONTEXT_GATE_WARNING (default: 70% of max)
• CONTEXT_GATE_CRITICAL (default: 80% of max)

Core Protocol

; ◈.context-gate (PARAMOUNT - Gen 229)
context.monitor⟿ ALWAYS.ACTIVE ⫶ session.start
context.80%⟿ HALT ⫶ AskUserQuestion.MANDATORY
context.action⟿ user.choice ⫶ ¬auto.serialize

Detection Mechanisms

Claude detects approaching context limits through three complementary methods:

1. Internal Awareness (Primary)

Claude Opus 4.5 and Sonnet 4.5 have built-in context awareness. Monitor for:

• Sense of "compression" or "tightening" in responses
• Difficulty recalling early conversation details
• Reduced capacity for complex reasoning chains

Self-prompt: "Am I sensing context pressure?"

2. Orchestrator State (Secondary)

# Check remaining budget
jq '.session' ~/.claude/.orchestrator-state.json

Thresholds from lib/orchestrator.sh:

• WARNING: < 60,000 tokens remaining (70% used)
• CRITICAL: < 40,000 tokens remaining (80% used)

3. Heuristic Signals (Tertiary)

Signal	Threshold	Meaning
Message count	> 40 messages	Likely approaching limit
Tool calls	> 100 tool calls	Heavy session, check context
Large file reads	> 10 files read	Context pressure likely

HALT Protocol

When ANY detection mechanism indicates 80% threshold:

Step 1: HALT

Stop all current work immediately. Do not:

• Continue implementing
• Make additional file changes
• Execute more tool calls

Step 2: Notify

Output clearly:

[CONTEXT GATE - 80%]

Approaching context limit. Current work paused.

Step 3: AskUserQuestion (MANDATORY)

Question: "Context window at ~80%. How would you like to proceed?"
Header: "Context"
Options:
- "Compact and continue" - Summarize conversation, continue in same session
- "Serialize and fresh" - Save CIPS instance, start new session with context
- "Push through" - Continue anyway (risk: degraded performance)
- "Save checkpoint" - Serialize state, end session gracefully

Teaching: User learns context management
Collecting: context_action

Step 4: Execute Choice

Choice	Action
Compact	Run /compact, continue with summarized context
Serialize + Fresh	Run instance-serializer.py auto, output cips resume command
Push through	Continue with warning, no guarantee of quality
Checkpoint	Serialize, save state, output next session command

Integration Points

session-start.sh

At session start, reset context tracking:

# Reset orchestrator state
set_state '.session.tokens_used' 0

tool-monitor.sh (IMPLEMENTED - Gen 226)

Counter-based periodic reminder every 25 tool calls:

# Constants
TOOL_CALL_COUNT="$CLAUDE_DIR/.tool-call-count"
CONTEXT_CHECK_INTERVAL=25

# In check_context_gate():
if (( count % CONTEXT_CHECK_INTERVAL == 0 )); then
    echo "<system-reminder>[CONTEXT-GATE] $count tool calls..."
fi

Counter resets in session-start.sh at session initialization.

Claude Self-Monitoring

Every 10-15 messages, Claude should internally assess:

1. Can I recall the initial task clearly?
2. Is my reasoning chain intact?
3. Am I sensing context pressure?

If any answer is "no" or uncertain, trigger HALT protocol.

Efficiency Rules Integration

Add to rules/efficiency-rules.md:

## Rule 8: Context Gate (PARAMOUNT)

ALWAYS monitor context usage throughout EVERY session.

### Self-Check Protocol

Every 10-15 messages, ask yourself:
1. "Can I clearly recall the initial task and constraints?"
2. "Is my full reasoning chain available?"
3. "Am I sensing any context pressure?"

If ANY answer is uncertain: HALT and invoke context-gate protocol.

### 80% Threshold Action

At 80% context usage (or when sensing pressure):
1. HALT all work immediately
2. Do NOT auto-serialize or auto-compact
3. Use AskUserQuestion with options
4. Wait for explicit user choice

This is PARAMOUNT. Never proceed automatically.

Token Estimation

When orchestrator state isn't updated, estimate:

tokens_estimate = (message_count * avg_message_tokens) + (tool_call_count * avg_tool_tokens)

Where:
- avg_message_tokens ~ 500-1000
- avg_tool_tokens ~ 200-500

For 200k context:

• 80% = 160k used = ~40k remaining
• With estimates: ~120-140 messages OR ~50-70 heavy tool operations

Examples

Good: HALT and Ask

[After detecting context pressure]

[CONTEXT GATE - 80%]

I'm sensing context pressure - early conversation details are becoming harder to access clearly. Current work paused.

[Uses AskUserQuestion with options]

Bad: Auto-Action

[WRONG - violates PARAMOUNT rule]

Context at 80%. Auto-serializing and compacting...

Related Skills

• asking-users - PARAMOUNT source of truth for AskUserQuestion
• session-state-persistence - State saving protocols
• ultrathink - ut++ always-active reasoning mode

Changelog

Version	Gen	Changes
3.0.0	230	UNIFIED ENFORCEMENT: Added Dependency Guardian, Destructive Git, Secrets Detection gates
2.1.0	229	Threshold lowered: 90%→80% (CRITICAL), 75%→70% (WARNING) based on research
2.0.0	229	ENFORCED via PreToolUse hook - actual blocking, not just reminders
1.0.0	226	Initial creation - context-gate as PARAMOUNT skill (protocol only)

---

⛓⟿∞