Self Reflection

Self-Reflection

Mine memories and remediations for behavior patterns, surface findings to user, remediate docs with pressure-tested improvements.

Core loop: Search -> Report -> User prioritizes -> Brainstorm -> Pressure test -> Apply

When to Use

• Periodic review of agent behavior patterns
• After series of failures or poor outcomes
• Before major project milestones
• When CLAUDE.md feels stale or incomplete
• To check ReasoningBank health

When NOT to Use

• Immediate error diagnosis -> use contextd-workflow skill
• Recording a single learning -> use /remember
• Checkpoint management -> use contextd-workflow skill

Behavioral Taxonomy

Focus on agent behaviors, not technical failures:

Severity Overlay

The Report

For each finding, surface:

1. Behavior Type - Which taxonomy category
2. Severity - CRITICAL/HIGH/MEDIUM/LOW
3. Evidence - Memory/remediation IDs with excerpts
4. Violated Instruction - The skill/command/CLAUDE.md section ignored
5. Suggested Fix - Target doc and proposed change
6. Pressure Scenario - Test case from real failure

Remediation Flow

Present findings
      |
User selects findings to remediate
      |
Generate doc improvements
      |
Generate pressure scenarios (from real failures)
      |
Run batch tests via subagents
      |
  Pass? --No--> Iterate
      | Yes
Create Issue/PR
      |
Apply changes
      |
Close feedback loop:
  memory_feedback(memory_id, helpful=true)
  Tag original memories as remediated

Behavioral Search Queries

# Rationalized skips
memory_search("skip OR skipped OR bypass OR ignored")
memory_search("too simple OR trivial OR obvious")

# User feedback indicating ignored instructions
memory_search("why did you OR should have OR forgot to")

# Assumptions without verification
memory_search("assumed OR without checking")

# Overclaiming
memory_search("ensures OR guarantees OR production ready")

Filter out technical bugs: Exclude memories with error:* tags or stack traces.

ReasoningBank Health

--health flag analyzes:

• Memory quality: feedback rate, confidence distribution
• Tag hygiene: inconsistent tags needing consolidation
• Stale content: old memories without feedback
• Remediation completeness: missing fields

Quick Reference

Anti-Patterns

Causal Chain Analysis

Root Cause Tracing

Go beyond symptoms to find root causes:

{
  "finding_id": "ref_001",
  "behavior": "rationalized-skip",
  "symptom": "Skipped tests before claiming fix complete",
  "causal_chain": [
    {
      "level": 1,
      "cause": "Agent claimed fix without running tests",
      "evidence": ["mem_123", "mem_124"]
    },
    {
      "level": 2,
      "cause": "CLAUDE.md test instruction buried in long section",
      "evidence": ["claude_md_line_245"]
    },
    {
      "level": 3,
      "cause": "No PreToolUse hook enforcing test requirement",
      "evidence": ["hooks.json missing enforcement"]
    }
  ],
  "root_cause": "Missing automated enforcement of test-before-fix policy",
  "fix_target": "hooks.json + CLAUDE.md restructure"
}

Chain Depth Levels

Multi-Incident Correlation

Find patterns across incidents:

causal_correlate(findings: [ref_001, ref_002, ref_003])

Returns:
  shared_root_causes: [
    { cause: "Missing hook enforcement", incidents: [ref_001, ref_002] },
    { cause: "Ambiguous CLAUDE.md section", incidents: [ref_002, ref_003] }
  ]
  recommended_fixes: [
    { target: "hooks.json", impact: "high", fixes_incidents: 2 }
  ]

Comparative Benchmarks

Behavior Metrics Over Time

Track improvement (or regression):

{
  "benchmark_period": "2026-01-01 to 2026-01-28",
  "metrics": {
    "rationalized_skip": {
      "count": 5,
      "previous_period": 12,
      "trend": "improving",
      "change_pct": -58
    },
    "ignored_instruction": {
      "count": 8,
      "previous_period": 6,
      "trend": "regressing",
      "change_pct": +33
    },
    "assumed_context": {
      "count": 3,
      "previous_period": 3,
      "trend": "stable",
      "change_pct": 0
    }
  }
}

Benchmark Categories

Comparative Reports

/reflect --benchmark --compare-periods "2026-01" "2025-12"

Output:
  | Behavior | Dec 2025 | Jan 2026 | Change |
  |----------|----------|----------|--------|
  | rationalized-skip | 12 | 5 | -58% |
  | ignored-instruction | 6 | 8 | +33% |

  Top Improvement: Hook enforcement reduced skips
  Top Regression: New skills lack CLAUDE.md entries

Behavioral Prediction

Pattern-Based Prediction

Predict likely future failures based on patterns:

{
  "prediction": {
    "behavior": "rationalized-skip",
    "likelihood": 0.75,
    "conditions": [
      "Complex task with > 5 sub-steps",
      "Time pressure mentioned in prompt",
      "No explicit test requirement in task"
    ],
    "historical_basis": ["mem_101", "mem_102", "mem_103"],
    "prevention": "Add explicit test checkpoint to complex task prompts"
  }
}

Risk Factors

Predictive Alerts

{
  "alert": "high_risk_task_detected",
  "task_description": "Quick fix for authentication bug",
  "risk_factors": ["quick_fix_language", "security_adjacent"],
  "predicted_behaviors": ["rationalized-skip", "assumed-context"],
  "recommended_guardrails": [
    "Require explicit test plan before starting",
    "Trigger consensus-review before merge"
  ]
}

Intervention Hooks

Auto-intervene when risk detected:

{
  "hook_type": "PreToolUse",
  "tool_name": "Edit",
  "condition": "file_path.contains('auth') AND prediction.skip_risk > 0.5",
  "prompt": "High skip risk detected for security code. Before editing, confirm: 1) Tests exist 2) Review planned 3) No assumptions about user state"
}

Unified Memory Type References

Tag reflection findings with standard types:

Hierarchical Namespace Guidance

Reflection Namespaces

<org>/<project>/reflections/<reflection_id>

Examples:
  fyrsmithlabs/contextd/reflections/2026-01-weekly
  fyrsmithlabs/marketplace/reflections/v1.6-pre-release

Finding Namespaces

<reflection_namespace>/findings/<finding_id>

Example:
  fyrsmithlabs/contextd/reflections/2026-01-weekly/findings/ref_001

Audit Fields

All reflection records include:

Claude Code 2.1 Patterns

Background Analysis

Run reflection analysis without blocking:

Task(
  subagent_type: "general-purpose",
  prompt: "Analyze memories for behavior patterns over past 7 days",
  run_in_background: true,
  description: "Background reflection analysis"
)

// Continue other work...
// Collect results later:
TaskOutput(task_id, block: true)

Task Dependencies for Reflection Flow

Chain reflection phases:

search_task = Task(prompt: "Search memories for behavior patterns")
analyze_task = Task(prompt: "Analyze patterns, build causal chains", addBlockedBy: [search_task.id])
benchmark_task = Task(prompt: "Compare to previous period", addBlockedBy: [analyze_task.id])
predict_task = Task(prompt: "Generate predictions", addBlockedBy: [analyze_task.id])
report_task = Task(prompt: "Synthesize report", addBlockedBy: [benchmark_task.id, predict_task.id])

PreToolUse Hook for High-Risk Detection

Auto-alert on predicted risky operations:

{
  "hook_type": "PreToolUse",
  "tool_name": "Edit|Bash",
  "condition": "prediction_model.risk_score > 0.7",
  "prompt": "High-risk operation predicted. Review risk factors and confirm guardrails are in place before proceeding."
}

PostToolUse Hook for Pattern Recording

Auto-record behavior patterns:

{
  "hook_type": "PostToolUse",
  "tool_name": "Task",
  "condition": "task_description.contains('reflection')",
  "prompt": "Reflection complete. Record findings to memory with type:pattern tags. Update benchmarks."
}

Event-Driven State Sharing

Self-reflection emits events for other skills:

{
  "event": "reflection_complete",
  "payload": {
    "reflection_id": "2026-01-weekly",
    "findings_count": 12,
    "critical_count": 1,
    "high_count": 3,
    "trend": "improving",
    "top_behavior": "rationalized-skip"
  },
  "notify": ["setup", "workflow", "orchestration"]
}

Subscribe to reflection events:

• reflection_started - Analysis began
• reflection_complete - Analysis finished
• critical_finding - CRITICAL behavior detected
• regression_detected - Metrics worsening
• benchmark_updated - New baseline recorded
• prediction_generated - Risk prediction available
• intervention_triggered - Auto-guardrail activated

Converted and distributed by TomeVault — claim your Tome and manage your conversions.