sota-improvement

name: sota-improvement description: | Use when: Agent self-improvement cycles, scheduled maintenance windows, after detecting multiple failures (auto-trigger), user requests "make yourself better" or "run improvement loop", or when system is stable and time for growth. Don't use when: User has specific task (not exploration), simple queries, general coding work, or when user asks for help with specific problem (use debugging/coding skills instead). Outputs: Fixed failures, integrated capabilities, new beliefs, updated workflows, research summaries, experience captures. metadata: {"openclaw":{"emoji":"🔄","category":"meta"}} read_when: - User requests self-improvement or maintenance - Scheduled improvement window opens - Multiple failures detected (3+ in 24h) - User specifies END_TIME for improvement - NOT for specific tasks or general queries

SOTA-Driven Self-Improvement

Execute continuous improvement cycles until END_TIME. Two modes:

Mode A (Reactive): Fix failures from logs Mode B (Proactive): Research improvements, find underutilized capabilities

Quick Start

# Check mode
fail-query --recent --count 5
# If failures → Mode A, else → Mode B

# Research (MANDATORY: use skills)
# Mode A: exa "[problem] solutions"
# Mode B: exa "agent [topic] SOTA 2025"

Mode Selection

1. Check recent failures:
   - fail-query --recent --count 10
   - ~/clawd/memory/networks/scripts/session-report.sh | grep -A 5 "Failures"

2. If failures exist → Mode A
3. If no failures → Mode B

Mode A: Failure-Driven

See: mode-a-template.md

Quick steps:

1. Scan failures → Select highest-impact
2. Prioritize → Score = (impact × feasibility) / time
3. Check existing → grep scripts/, skills/, qmd search
4. Research SOTA → exa "[problem] prevention patterns"
5. Fix minimally → <200 lines, single solution
6. Verify → Recreate failure, confirm fixed
7. Prove → Demo or test passes
8. Integrate → Workflow + docs + beliefs

Mode B: Exploration-Driven

See: mode-b-template.md

Quick steps:

1. Audit → Check capability usage, workflow gaps, knowledge gaps
2. Research frontier → exa "agent [topic] SOTA 2025"
3. Identify gap → Current vs. SOTA approach
4. Check existing → Already have this capability?
5. Implement → Prototype or integrate
6. Verify → Before/after metrics
7. Prove → Benchmark or demo
8. Integrate → Workflow + docs + beliefs + experience

Research Requirements

PRIMARY: exa skill (neural search)

exa: "[topic] implementation patterns"
exa: "[topic] github examples"
exa: "[topic] research 2024 2025"

FALLBACK: perplexity skill (AI + citations)

perplexity: "What are SOTA approaches for [topic]?"

LOCAL ONLY (degraded mode)

Only if both skills unavailable
Log skill failure for investigation

Constraints

• END_TIME: Stop regardless of progress
• YAGNI: If not needed THIS session, don't build
• Integration-first: Check existing before creating
• Max complexity: +200 lines or 1 file per iteration
• Research quality: exa ≥80% target

Synthesis Template

[SYNTHESIS] <time> | <mode: A/B> | <focus> | ComplexityDelta: <+X lines> | Status: <VALIDATED/FAILED>

Mode: [Failure/Exploration]
What Changed: [1-2 sentences]
Key Insight: [what you learned]
Research Source: [exa/perplexity/local]
Tool Now Available: [if applicable]

Success Metrics

Full Protocol

For complete protocol with all steps: protocol-v2.2.md

Example Session

See: session-2026-02-20.md

Background Process Management

If long-running process needed:

• Run with timeout and yieldMs
• Check status each iteration
• Stuck >15 min → kill and defer

Why Two Modes?

Both needed. Reactive maintains, proactive grows.