research(skills): MetaAgent tool meta-learning for autonomous knowledge gap detection

[bug-ops]

Summary

Autonomous skill evolution in two phases: Phase 1 distills improvements from successful execution trajectories (SkillRL pattern); Phase 2 detects knowledge gaps and synthesizes new skills from scratch (MetaAgent pattern).

Sources:

• arXiv 2602.08234 — SkillRL: Evolving Agents via Recursive Skill-Augmented Reinforcement Learning (Feb 2026)
• arXiv 2508.00271 — MetaAgent: Toward Self-Evolving Agent via Tool Meta-Learning (Aug 2025)

Phase 1: Trajectory Distillation (SkillRL pattern)

A 7B model outperforms GPT-4o by 41% on ALFWorld/WebShop when augmented with this approach.

When a session ends with high user satisfaction, extract successful tool-use patterns as skill refinement candidates:

1. Add trajectory_capture = true config under [skills.learning] to record full tool-use sequences tagged with outcome.
2. Implement TrajectoryDistiller: background task (runs on session shutdown or timer) that feeds recent positive trajectories to LLM with "what worked?" prompt → structured skill delta.
3. Merge delta into candidate skill version via existing SkillVersionManager.
4. Gate behind [skills.learning] trajectory_distillation = true (off by default, experimental).

Complexity: HIGH — trajectory capture storage + background distillation loop + skill evolution pipeline integration. Implement in phases.

Phase 2: Gap Detection + Skill Synthesis (MetaAgent pattern)

When skill registry has no matching skill for a task, instead of failing silently:

1. Detect gap: GapDetector in zeph-skills fires when top skill score < gap_threshold
2. Generate help-seeking query: structured request "I need a skill that can do X"
3. Route: to web scraper (find docs) or LLM self-generation of new SKILL.md stub
4. Reflect and persist: new skill enters registry at TrustLevel::Community with low initial confidence → after N successful uses → promoted to TrustLevel::Trusted

Integration point: SkillRegistry::match() + FeedbackDetector

Config:

[skills.meta_learning]
enabled = false
gap_threshold = 0.3
trajectory_distillation = false

Integration Points

• crates/zeph-skills: TrajectoryDistiller, GapDetector, SkillSynthesizer
• Extends existing: FeedbackDetector, Wilson score re-ranking, SkillVersionManager
• Phase 1 prerequisite: trajectory capture storage (new SQLite table or append to audit log)
• Phase 2 prerequisite: Phase 1 (gap detection accuracy improves with trajectory data)

Key Results

• SkillRL: 7B model outperforms GPT-4o by 41% on ALFWorld/WebShop with trajectory distillation
• MetaAgent: matches/exceeds end-to-end trained agents on GAIA, WebWalkerQA, BrowseCamp

See Also

• research(testing): TDAD behavioral spec testing for skills and system prompt blocks #1842 (TDAD behavioral spec testing for skills)
• Existing zeph-skills self-learning: FeedbackDetector, SkillVersionManager