Add an optional AISP eval quality contract for dataset coverage, grading, failure analysis, and optimization traces

从现有的 agents-cli eval generate、grade、dataset synthesize、compare、analyze 和 optimize 入口点开始，然后查看链接的 AISP 参考资料和拟议的包布局。定义 MVP eval 质量报告和可选的契约，不替换命令，也不要求运行时支持。完成标准是：已明确规定数据集警告、trace 路径和 grade 配置、失败摘要、可选的前后对比以及 eval_readiness_evidence。

What is your feature suggestion?

Add an optional machine-readable evaluation quality contract for agents-cli eval.

The eval commands are already one of the strongest parts of agents-cli:

eval generate
eval grade
eval dataset synthesize
eval compare
eval analyze
eval optimize

I propose adding an optional AISP V1.0.0 companion package that describes the expected eval lifecycle and trace requirements.

AISP protocol reference:

https://github.com/AIXP-Labs/AISP

This would not replace the current eval commands.

It would define what a “complete eval cycle” means.

Possible layout:

skills/google-agents-cli-eval/
├── SKILL.md
└── aisp/
    └── google_agents_cli_eval_quality_contract_aisp/
        ├── aisp.aisop.json
        ├── README.md
        └── evals/
            └── eval-quality-traces/

A standard AISP V1.0.0 package could look like:

[
  {
    "role": "system",
    "content": {
      "protocol": "AISP V1.0.0",
      "axiom_0": "Human_Sovereignty_and_Wellbeing",
      "id": "google_agents_cli_eval_quality_contract_aisp",
      "name": "Google Agents CLI Eval Quality Contract",
      "version": "1.0.0",
      "license": "Apache-2.0",
      "summary": "A machine-checkable evaluation quality contract for agents-cli eval generate, grade, analyze, compare, and optimize workflows.",
      "description": "Defines dataset readiness, trace generation, grading, metric thresholds, failure clustering, optimization comparison, and deploy-readiness evidence for ADK agent evaluation.",
      "flow_format": "mermaid",
      "loading_mode": "node",
      "tools": [
        "filesystem",
        "shell"
      ],
      "params": {
        "project_root": "string",
        "eval_dataset_path": "string",
        "eval_config_path": "string?",
        "baseline_results_path": "string?"
      },
      "system_prompt": ""
    }
  },
  {
    "role": "user",
    "content": {
      "instruction": "STRICTLY OBEY aisp_contract; its non_negotiable rules are inviolable; then RUN aisop.main",
      "user_input": "{user_input}",
      "aisp_contract": {
        "profile": "aisp.skill.v1",
        "invocation": {
          "mode": "eval_quality_gate",
          "when_to_use": [
            "before running agents-cli eval generate",
            "before running agents-cli eval grade",
            "before using eval results as deployment evidence",
            "before running eval optimize",
            "when comparing baseline and optimized agent behavior"
          ],
          "when_not_to_use": [
            "unit tests that only validate deterministic code",
            "manual exploratory chat with no eval dataset",
            "local smoke test not intended as deployment evidence"
          ]
        },
        "non_negotiable": [
          {
            "rule": "Do not treat pytest output as a substitute for agent behavior eval.",
            "enforced_by": "classify_eval.step2:sys.assert"
          },
          {
            "rule": "Eval dataset must include expected behavior, allowed tools, failure conditions, or rubric fields.",
            "enforced_by": "dataset_gate.step3:sys.assert"
          },
          {
            "rule": "Eval generate must produce traces before eval grade runs.",
            "enforced_by": "generate.step3:sys.assert"
          },
          {
            "rule": "Eval grade must record metrics, failures, and grading configuration.",
            "enforced_by": "grade.step3:sys.assert"
          },
          {
            "rule": "Eval analyze must cluster or summarize failure modes before optimization is accepted.",
            "enforced_by": "analyze.step2:sys.assert"
          },
          {
            "rule": "Eval optimize must compare before/after results before claiming improvement.",
            "enforced_by": "compare.step2:sys.assert"
          },
          {
            "rule": "Deploy readiness must not be based on eval status without dataset id, trace id, grade result, and threshold status.",
            "enforced_by": "readiness.step3:sys.assert"
          }
        ],
        "discovery": {
          "category": "evaluation",
          "tags": [
            "agents-cli",
            "eval",
            "dataset",
            "grade",
            "failure-analysis",
            "optimization",
            "deploy-readiness",
            "aisp"
          ]
        },
        "risk_level": "medium",
        "resources": [
          {
            "id": "eval_dataset",
            "path": "{eval_dataset_path}",
            "kind": "eval_dataset",
            "mode": "read_only",
            "when": "Read before eval generate.",
            "scope": "skill"
          },
          {
            "id": "eval_config",
            "path": "{eval_config_path}",
            "kind": "config",
            "mode": "read_only",
            "when": "Read before eval grade.",
            "scope": "skill"
          },
          {
            "id": "baseline_results",
            "path": "{baseline_results_path}",
            "kind": "eval_results",
            "mode": "read_only",
            "when": "Read before eval compare or optimize.",
            "scope": "skill"
          }
        ]
      },
      "aisop": {
        "main": "graph TD\n    classify_eval[Classify eval purpose] --> dataset_gate[Validate dataset]\n    dataset_gate --> generate[Run eval generate]\n    generate --> grade[Run eval grade]\n    grade --> analyze[Analyze failures]\n    analyze --> optimize[Optimize if requested]\n    optimize --> compare[Compare before after]\n    compare --> readiness[Emit deploy readiness evidence]\n    readiness --> trace[Write eval quality trace]\n    trace --> end_node((End))"
      },
      "functions": {
        "classify_eval": {
          "step1": "Classify whether this is behavior eval, deterministic unit test, smoke test, regression eval, deployment evidence, or optimization eval.",
          "step2": "sys.assert('behavior eval is distinct from pytest/unit tests', 'pytest is not a substitute for agent behavior eval')",
          "output_mapping": "eval_purpose"
        },
        "dataset_gate": {
          "step1": "Read eval_dataset_path.",
          "step2": "Validate dataset has cases with prompts, expected behavior, rubric, allowed tool behavior, or failure criteria.",
          "step3": "sys.assert('eval dataset has required behavioral criteria', 'Eval dataset missing behavior/rubric fields')",
          "output_mapping": "dataset_validation"
        },
        "generate": {
          "step1": "Run agents-cli eval generate or equivalent.",
          "step2": "Record trace output path and case count.",
          "step3": "sys.assert('eval traces generated before grading', 'Eval grade requires generated traces')",
          "output_mapping": "eval_generate_status"
        },
        "grade": {
          "step1": "Run agents-cli eval grade with eval_config_path when provided.",
          "step2": "Record metrics, score, failures, judge config, and thresholds.",
          "step3": "sys.assert('eval grade result includes metrics and config', 'Eval grade result incomplete')",
          "output_mapping": "eval_grade_status"
        },
        "analyze": {
          "step1": "Cluster or summarize failure modes from eval_grade_status.",
          "step2": "sys.assert('failure modes analyzed before optimization acceptance', 'Eval failures must be analyzed')",
          "output_mapping": "failure_analysis"
        },
        "optimize": {
          "step1": "If optimization requested, run prompt/agent optimization using eval data.",
          "output_mapping": "optimization_status"
        },
        "compare": {
          "step1": "Compare baseline and optimized eval results when optimization occurs.",
          "step2": "sys.assert('before/after eval comparison exists for optimization claim', 'Optimization requires before/after comparison')",
          "output_mapping": "eval_comparison"
        },
        "readiness": {
          "step1": "Compute deploy readiness evidence from dataset_validation, eval_generate_status, eval_grade_status, failure_analysis, and eval_comparison.",
          "step2": "Record dataset id, trace id, grade result, threshold status, and failure summary.",
          "step3": "sys.assert('deploy readiness includes eval evidence fields', 'Deploy readiness evidence incomplete')",
          "output_mapping": "eval_readiness_evidence"
        },
        "trace": {
          "step1": "Write eval quality trace with dataset validation, trace paths, grade results, failure clusters, comparison, and readiness status.",
          "output_mapping": "eval_quality_trace"
        },
        "end_node": {
          "step1": "Return eval_quality_trace and eval_readiness_evidence."
        }
      }
    }
  }
]

What will this enable you to do?

This would make eval runs more useful as engineering evidence.

It would enable:

• clearer distinction between pytest/unit tests and agent behavior evals;
• validation of dataset quality before running expensive evals;
• structured eval traces that can be referenced before deployment;
• failure clustering before optimization;
• before/after comparison for optimization claims;
• deploy readiness evidence that includes dataset id, trace id, grade result, and threshold status;
• future CI gates based on eval contract fields.

It would also help coding agents avoid the common mistake of saying “tests passed” when only deterministic unit tests ran and no agent behavior eval was performed.

Additional context

This should start as a contract + report shape, not as a required new runtime.

Non-goals:

• Do not replace existing eval commands.
• Do not require AISP runtime support.
• Do not block lightweight local experiments.
• Do not require every project to define strict thresholds on day one.
• Do not make LLM-as-judge the only possible metric.

MVP:

1. Define eval quality report fields.
2. Add dataset validation warnings.
3. Record generated trace paths and grade config.
4. Add failure clustering summary.
5. Add optional before/after comparison for optimize.
6. Emit eval_readiness_evidence that deploy or CI can read later.