Analyzer

Analyzer Skill

Purpose

Analyzer Skill is the Codex/Agent-facing capability for running LoopAI Analyzer independently. It analyzes evaluation outputs, writes Analyzer reports, emits stream events, returns unified success/error payloads, and compares completed versions of the same task and benchmark. Code, Text2SQL and Math Rollout share seven text reports, a training-plan JSON and a new annotated OJ copy.

Python Implementation

The Python skill layer and Analyzer business implementation live in:

loopai/skills/Analyzer

Analyzer no longer depends on loopai/agents/Analyzer. The legacy agents-side Analyzer directory has been removed; Codex and WebUI should call the skill entry directly.

Runtime Entry

from loopai.skills.Analyzer import run, resume_run

run(state=None, resume=False, from_node=None, baseline_result_path=None)
resume_run(state=None, from_node=None, baseline_result_path=None)

run(...) is the Codex/sub-agent process entry. It emits the unified LoopAI payload to stdout and exits, matching the latest Judger skill pattern. For in-process calls use run_analyzer_standalone(...).

Use resume_run(...) for continuation. It always passes resume=True and selects the latest incomplete version checkpoint for the task. A normal run(...) also resumes the latest incomplete version by default. To explicitly start a new run, pass new_version=True (or use the CLI --new-version).

{
  "ok": true,
  "status": "completed",
  "message": "Analyzer completed.",
  "data": {},
  "error": null
}

Direct runner:

from loopai.skills.Analyzer.runner import run_analyzer_standalone

run_analyzer_standalone(...) keeps the legacy behavior and returns the final state directly.

LangGraph-compatible class import:

from loopai.skills.Analyzer.analyzer_agent import AnalyzerAgent

CLI

python examples/scripts/run_analyzer_standalone.py   --config-path /tmp/analyzer_full_demo.json   --baseline-result-path /tmp/analyzer_demo_baseline.jsonl   --print-result