노이즈 가득한 실행 로그에서 근본 원인 찾기: Agent 최적화를 위한 구조적 Trajectory 분석과 인과 추출

# STRACE 4단계 파이프라인 개념 구현 스케치

# Phase 1: Execution Dependency Graph 구성
system_prompt_phase1 = """
You are analyzing an agent system repository.
Identify all functional modules (V) and infer edges (E) between them:
- Data dependency: Module B consumes artifacts produced by Module A
- Control dependency: Module A governs whether Module B is invoked

Output as a text edge list:
<edges>
module_A -> module_B : data | artifact: intermediate_plan
module_A -> module_C : control | condition: error_type == 'syntax'
</edges>
"""

# Phase 2: 실패 패턴 요약 및 trace 필터링
failure_summary_prompt = """
Given execution traces, compute:
1. Statistical Severity: P(Task Fail | Node_i Fail) for each module
2. Structural Path Patterns: recurring pathological sequences (loops, dead-ends)

Select {s} representative traces covering distinct failure clusters.
"""

# Phase 3: Causal Localization - backward slicing
causal_localization_prompt = """
Given:
- Manifestation Node (vm): where error surfaces
- Dependency Graph (G)
- Execution Trace

Step 1: Starting from vm, traverse dependency edges BACKWARD.
        Retain only steps reachable via Data/Control dependency closure.
        This is the Causal Slice (C_slice).

Step 2: Within C_slice, identify Root Cause Node (vr):
        The earliest module where execution first deviated from intended logic.
        Example: Code Interpreter (vm) crashed because Planner (vr) hallucinated
        a parameter 45 steps earlier.

Output:
- Causal Slice: [step_1, step_5, step_12, step_48]
- Root Cause Node: planner_module
- Root Cause Explanation: ...
"""

# Phase 4: 귀납적 정책 최적화
policy_optimization_prompt = """
Given causal slices from multiple failures all traced to root cause module: {vr}

Synthesize GENERALIZED heuristic rules (not instance-specific patches):
- Identify recurring failure patterns across episodes
- Abstract into reusable preventive guidelines
- Format as instruction additions for the module's system prompt

Example output:
"When attempting action X more than 2 times with consistent rejection,
 switch to alternative action Y. Track attempt counts per action type."

Inject these rules ONLY into module {vr}'s prompt.
"""

# GitHub: https://github.com/moomight/STRACE