delegation-lora

Context-adaptive delegation. Auto-delegates to subagents when thrashing, session is long (>30 calls), or write-heavy. Includes cross-model review trigger.

SKILL.md

GitHub

Delegation — Context-Adaptive Auto-Delegate

长会话不退化、不打转、不盲写。自动委托子 agent 保持上下文干净。

5 Triggers

1. Circular Search

Detect: Same tool + same target ≥3 calls Action: Stop → delegate to Explore agent or general-purpose agent Why: Self-looping = context polluted. Clean agent finds it in one pass.

2. Long Session

Detect: Session >30 tool calls Action: New independent tasks → subagent. Search → Explore. Review → adversarial-reviewer. Deep analysis → general-purpose. Why: Context window is finite. Independent tasks deserve independent windows.

3. Write-Heavy

Detect: 3 consecutive Edit/Write without Read between Action: Pause → Read target + references → resume

4. Context Pressure

Detect: Window feels crowded or >60% Action: Multi-task → parallel agents. Completed subtasks → compress to summaries. Unused file contents → discard.

5. Cross-Model Review

Detect: Security code, architecture change, >20 lines new logic Action: Delegate to adversarial-reviewer / security-auditor / code-review Why: Self-review has 64.5% blind spot. Different model = different blind spot.

Agent Quick Reference

Task Agent Why
Multi-file search Explore Returns conclusions, not file dumps
Code review adversarial-reviewer Different model, truly independent
Security audit security-auditor Specialized capability
System diagnosis system-auditor Comprehensive scan
Deep research general-purpose Multi-step reasoning
Simple mechanical general-purpose (haiku) Cheap and fast

Success Story

Without: 80-tool-call session → context degradation → hallucinated phantom function → 1h debugging. With: 30th call triggers delegation → clean subagent → zero hallucination → 50% faster.

Integration

  • second-brain: Trigger #5 (cross-model review) delegates to Brain 2 adversarial-reviewer
  • memory-lora: 🔜 Sub-agent queries memory.db before starting work
  • precommit-pipeline: Step 4 cross-model review = delegation #5 in action