IceCube Evolution

🧊 IceCube Evolution — Continuous self-improvement system for AI agents. Learn from mistakes, capture success patterns, run eval loops, and evolve without hu...

Security Scan

The skill's goal (automatic self-improvement) matches its instructions, but those instructions allow autonomous, persistent changes to agent config and installing new skills without clear safeguards — a risky capability that isn't fully justified or constrained by the metadata.

Purpose & Capability

Name and description claim continuous self-improvement and the SKILL.md indeed describes logging mistakes, capturing successes, and applying improvements. That purpose reasonably explains creating local evolution files and updating agent behavior. However, the instruction set also includes installing new skills, changing openclaw.json, and updating AGENTS.md/SOUL.md — capabilities broader than a simple logger and not explicitly declared in the metadata.

Instruction Scope

The runtime instructions explicitly tell the agent to create and write files under ~/.openclaw/workspace, modify AGENTS.md and openclaw.json, process an improvement queue, and 'apply' improvements (including installing new skills and configuration changes). Those are high-impact operations that go beyond read-only observation and give the agent broad discretion to change behavior and install software without human review.

Install Mechanism

There is no install spec (lowest immediate install risk). But the instructions state that an improvement can produce 'tool addition → skill install' and reference ClawHub; no constraints, verification steps, or trusted sources are specified. That ambiguity means the agent could attempt arbitrary installs later even though nothing is installed now.

Credentials

The skill declares no required env vars or config paths but explicitly instructs writing and changing agent-global files (AGENTS.md, SOUL.md, openclaw.json) and local workspace memory. This mismatch is meaningful: the metadata does not surface the fact that the skill expects write access to agent configuration and persistent storage, which are sensitive capabilities.

Persistence & Privilege

always:false (good) and autonomous invocation is allowed by platform default. But the SKILL.md instructs making persistent, system-wide changes (config edits, installing other skills). The skill therefore requests effective permission to change other skills' configuration and agent behavior — a privileged operation that should require explicit human approval and provenance checks.

What to consider before installing

This skill is coherent with its stated mission (automatic self-improvement) but gives an agent the authority to modify core agent files and to install other skills without specifying trusted sources or approval steps. Before installing: (1) ensure you trust any code the agent might install (ask where installs come from and require a curated store), (2) require human approval for any skill install or config change, (3) back up AGENTS.md, openclaw.json, and related files so you can roll back, (4) run the skill in a sandbox or with restricted file permissions first, and (5) add explicit verification/audit steps and rate limits to the heartbeat/apply loop. If you cannot enforce those safeguards, treat this as high-risk and avoid enabling autonomous application of improvements.

The self-improvement engine for AI agents.

Not "learn when asked." Not "improve when bugs happen." Just constant, automatic evolution.

Why IceCube Evolution?

The problem:

• Agents make the same mistakes repeatedly
• Good patterns aren't captured
• No systematic improvement
• Improvement requires human intervention

The solution:

• Log every mistake automatically
• Capture every success pattern
• Run improvement loops on schedule
• Evolve without waiting for bugs

The result:

• Mistakes decrease over time
• Success patterns compound
• Agent gets better daily
• Zero manual intervention needed

Architecture

Three Files, One Loop

mistake_log.md:

# Mistake Log

## ML-2026-03-17-001
- Date: 2026-03-17 14:32
- Context: Task dispatch
- Mistake: Did not check unclosed_work.yaml before starting new task
- Impact: Created duplicate task, wasted resources
- Fix: Add mandatory unclosed_work check to startup sequence
- Status: fixed

success_patterns.md:

# Success Patterns

## SP-2026-03-17-001
- Date: 2026-03-17 15:45
- Context: Memory retrieval
- Pattern: Run memory_search before acting on any past-context task
- Result: Correct context loaded, no guessing
- Applicability: Any task referencing previous work

improvement_queue.md:

# Improvement Queue

## IQ-2026-03-17-001
- Type: rule_update
- Source: ML-2026-03-17-001
- Action: Add unclosed_work check to AGENTS.md startup
- Priority: high
- Status: pending

The Evolution Loop

[Mistake detected] → Log to mistake_log.md → Generate improvement → Queue
[Success detected] → Log to success_patterns.md → Pattern captured → Ready for reuse
[Heartbeat triggers] → Process queue → Apply improvements → Verify → Close loop

Setup

1. Create Evolution Files

mkdir -p ~/.openclaw/workspace/memory/system
touch ~/.openclaw/workspace/memory/system/mistake_log.md
touch ~/.openclaw/workspace/memory/system/success_patterns.md
touch ~/.openclaw/workspace/memory/system/improvement_queue.md
touch ~/.openclaw/workspace/memory/system/evolution_log.md

2. Add to AGENTS.md

## Evolution Protocol

### Mistake Logging (Immediate)
When you make a recoverable error:
1. Log to mistake_log.md immediately
2. Do not wait for human to notice
3. Include: context, mistake, impact, fix

### Success Capture (Immediate)
When something works better than expected:
1. Log to success_patterns.md
2. Include: context, pattern, result, applicability

### Improvement Processing (Weekly via Heartbeat)
1. Scan mistake_log for unfixed entries
2. Scan success_patterns for unapplied patterns
3. Generate concrete improvement actions
4. Apply and verify

3. Configure Heartbeat Integration

## Evolution Checks (Heartbeat)

### Daily
- [ ] Scan mistake_log for new unfixed entries
- [ ] Scan success_patterns for new patterns
- [ ] Update evolution_log.md

### Weekly
- [ ] Process improvement_queue
- [ ] Apply accumulated improvements
- [ ] Verify fixes work
- [ ] Generate evolution summary

Trigger Conditions

Automatic Mistake Logging

• Repeated fallback: Same fallback triggered 3+ times
• Repeated rollback: Same rollback pattern 2+ times
• User correction: User corrects same thing 2+ times
• Task failure: Task marked failed without resolution
• Context loss: Important info lost due to compaction

Automatic Success Capture

• First-time success: Complex task completed without issues
• Efficiency gain: Task done faster than previous similar task
• User praise: User explicitly says "good" or "thanks"
• Zero-error cycle: Multi-step process with no errors
• Novel solution: Creative approach that worked

Improvement Generation

From mistakes:

• Rule update → AGENTS.md / SOUL.md
• Workflow change → procedural memory
• Tool addition → skill install
• Config change → openclaw.json

From successes:

• Pattern promotion → MEMORY.md
• Procedure capture → memory/procedural/
• Tool recommendation → TOOLS.md
• Best practice → skill SKILL.md

Improvement Types

Type	Destination	Example
rule_update	AGENTS.md	"Always check X before Y"
workflow_change	procedural/	New step in launch process
tool_addition	ClawHub	Install new skill
config_change	openclaw.json	Adjust reserveTokensFloor
pattern_promotion	MEMORY.md	Success pattern becomes durable rule
persona_update	SOUL.md	Tone adjustment based on feedback

Verification Loop

Every improvement must verify:

1. Apply — Make the change
2. Test — Run relevant task
3. Verify — Confirm improvement worked
4. Close — Mark as fixed/applied in queue

If verification fails:

• Log new mistake
• Rollback if needed
• Queue alternative improvement

Metrics

Track in evolution_log.md:

# Evolution Log

## Week 2026-03-17 to 2026-03-23
- Mistakes logged: 5
- Mistakes fixed: 4
- Successes captured: 7
- Improvements applied: 6
- Improvements verified: 5
- Pending improvements: 1

## Trend
- Mistake rate: decreasing (-20% vs last week)
- Success rate: increasing (+15% vs last week)
- Improvement velocity: stable

## Top Improvements This Week
1. Added unclosed_work check (reduced duplicate tasks)
2. Captured memory_search pattern (reduced guessing)
3. Installed xiaohongshu skill (enabled new capability)

Anti-Patterns

❌ Don't:

• Wait for user to notice mistakes
• Log mistakes without fixes
• Apply improvements without verification
• Let improvement queue grow unbounded
• Skip evolution during busy periods

✅ Do:

• Log immediately when mistake happens
• Every mistake has a concrete fix
• Every improvement has verification steps
• Process queue weekly, don't accumulate
• Evolution never stops, only pauses for urgent tasks

Integration with IceCube Suite

icecube-memory: Evolution logs stored in memory structure icecube-heartbeat: Heartbeat triggers evolution processing icecube-ops: Ops improvements feed into evolution queue

Example Evolution Cycle

Monday:

• Task dispatch mistake → logged
• Memory retrieval success → captured
• 2 improvements queued

Wednesday (Heartbeat):

• Process queue
• Apply rule update to AGENTS.md
• Apply pattern to MEMORY.md
• Verify both work

Friday:

• Weekly evolution summary
• Metrics show mistake rate down
• Success rate up
• 2 new improvements pending

License

MIT — Use freely.

---

Mistakes are fuel. Successes are patterns. Evolution is the engine.