Alibabacloud Agentloop Evaluation

Alibaba Cloud AgentLoop Evaluation

Scenario Description

Orchestrate AgentLoop evaluation workflows through the Aliyun CLI plugin (aliyun-cli-agentloop). The skill converts a compact JSON specification into the required AgentLoop API sequence: discover evaluators, create/update saved evaluators, create one-shot or batch evaluation tasks, poll runs, and analyze results from SLS.

Architecture: Aliyun CLI >= 3.3.3 + aliyun-cli-agentloop 0.7.0 + aliyun-cli-sls plugin + AgentLoop API (2026-05-20) + SLS evaluation_detail Logstore

Use the bundled wrapper to convert a compact JSON specification into the required AgentLoop API sequence. Default to a dry-run; send mutations only with --execute.

Installation

Pre-check: Aliyun CLI >= 3.3.3 required

[MUST] Verify: aliyun version — must be >= 3.3.3.

• First install or major upgrade: Download, review, then execute the setup script. Avoid curl | bash piping.
• Routine update (CLI >= 3.3.5): aliyun upgrade — prefer this built-in self-update over re-running the install script.
• See references/cli-installation-guide.md for full installation instructions.

Pre-check: Aliyun CLI plugin update required

[MUST] run aliyun configure set --auto-plugin-install true to enable automatic plugin installation. [MUST] run aliyun plugin update to ensure that any existing plugins are always up-to-date.

Verify the AgentLoop plugin:

aliyun plugin show --name aliyun-cli-agentloop
aliyun agentloop version

For result analysis, also install the SLS plugin:

aliyun plugin install --name aliyun-cli-sls

Script Dependencies

The bundled Python scripts in scripts/ use only the Python 3.8+ standard library — no external packages are required. See scripts/requirements.txt for the full declaration.

Environment Variables

Authentication

Pre-check: Alibaba Cloud Credentials Required

Security Rules:

• NEVER read, echo, or print AK/SK values (e.g., echo $ALIBABA_CLOUD_ACCESS_KEY_ID is FORBIDDEN)
• NEVER ask the user to input AK/SK directly in the conversation or command line
• NEVER use aliyun configure set with literal credential values
• ONLY use aliyun configure list to check credential status

aliyun configure list

Check the output for a valid profile (AK, STS, or OAuth identity).

If no valid profile exists, STOP here.

1. Obtain credentials from Alibaba Cloud Console
2. Configure credentials outside of this session (via aliyun configure in terminal or environment variables in shell profile)
3. Return and re-run after aliyun configure list shows a valid profile

RAM Policy

This skill requires AgentLoop and SLS permissions. See references/ram-policies.md for the full permission list.

[MUST] Permission Failure Handling: When any command or API call fails due to permission errors at any point during execution, follow this process:

1. Read references/ram-policies.md to get the full list of permissions required by this SKILL
2. Use ram-permission-diagnose skill to guide the user through requesting the necessary permissions
3. Pause and wait until the user confirms that the required permissions have been granted

Parameter Confirmation

IMPORTANT: Parameter Confirmation — Before executing any command or API call, ALL user-customizable parameters (e.g., AgentSpace name, Region, evaluator names, task names, dataset names, time windows, thresholds, etc.) MUST be confirmed with the user. Do NOT assume or use default values without explicit user approval.

Mutation Confirmation Protocol

CRITICAL: All cloud mutation operations require explicit user confirmation before execution.

This skill can create, modify, and delete cloud resources. To prevent unintended changes, every mutation MUST follow the three-step protocol below. No exceptions.

Mutation operations requiring confirmation

The following operations alter cloud resources and MUST NOT be executed without explicit user approval:

Three-step confirmation protocol

1. Preview (dry-run) — Always run without --execute first. Show the user the exact commands and JSON payloads that will be sent to the cloud API.

   python3 "$SKILL_DIR/scripts/agentloop_eval.py" run --spec /path/to/evaluation.json
   

2. Confirm — Present the rendered preview to the user and explicitly ask for approval. Quote the key parameters (AgentSpace, evaluator names, task name, data scope, estimated records, cost implications). Wait for a clear affirmative response before proceeding.

   Do NOT assume that running run without --execute implies intent to execute. Do NOT proceed to execution based on silence or ambiguity.

3. Execute — Only after receiving explicit confirmation, add --execute and any required override flags (--allow-unbounded, --allow-continuous):

   python3 "$SKILL_DIR/scripts/agentloop_eval.py" run \
     --spec /path/to/evaluation.json --execute
   

Read-only operations (no confirmation needed)

doctor, discover, status, and analyze (result analysis) are read-only and do not require the confirmation protocol.

Observability (MUST follow for every aliyun command)

Upon loading this skill, generate a random session ID (32-char lowercase hex string) once for the entire session. Use it as {session-id} below.

Rule: Every aliyun CLI command that calls a cloud API MUST include the --user-agent flag. Local utility commands (e.g. configure, plugin, version) do not support this flag and should be excluded.

--user-agent AlibabaCloud-Agent-Skills/alibabacloud-agentloop-evaluation/{session-id}

Example (assuming session-id is a1b2c3d4e5f6a7b8c9d0e1f2a3b4c5d6):

aliyun agentloop list-evaluators --agent-space my-space --user-agent AlibabaCloud-Agent-Skills/alibabacloud-agentloop-evaluation/a1b2c3d4e5f6a7b8c9d0e1f2a3b4c5d6

Do not skip, alter the format, or omit --user-agent on any aliyun API command invocation.

Script execution: The bundled Python wrapper automatically reads SKILL_SESSION_ID from the environment and injects --user-agent into every cloud API command. Inject the session-id via inline environment variable:

SKILL_SESSION_ID={session-id} python3 "$SKILL_DIR/scripts/agentloop_eval.py" run --spec evaluation.json --execute
SKILL_SESSION_ID={session-id} python3 "$SKILL_DIR/scripts/analyze_evaluation_results.py" --agent-space <space> ...

Core Workflow

Select the workflow

• Use a one-shot task with dataFilter.provided to test an evaluator against one supplied sample.
• Use a batch trace task with a bounded time window and maxRecords to evaluate observed traces.
• Use a batch dataset task for an existing AgentLoop dataset. Start from references/examples/batch-dataset-example.json and map evaluator variables directly to dataset columns.
• Dataset task with multiple saved evaluators: referencing two or more custom saved evaluators via evaluatorRef/evaluator_refs triggers a backend defect that expands the data-record set by the evaluator count (rows × evaluators) and re-scores every expanded record with every evaluator, so each row is evaluated N times per evaluator. On --execute the wrapper auto-detects this and inlines the referenced evaluator definitions (fetched via get-evaluator) into the task to match the safe console behaviour; it prints a WARNING and the sent request differs from the dry-run preview. Built-in refs and single-evaluator tasks are untouched. See references/spec-format.md (Dataset batches) for details.
• Use result analysis to inspect evaluation quality and low-score cases from the fixed evaluation_detail SLS Logstore.
• Use discover or status for read-only inspection.
• Add evaluator_actions only when the user asks to create or update saved evaluators. Use only AGENT or CODE as saved-evaluator create types.
• Decide the execution behavior by config.agentEvaluatorMode, NOT by --type alone. Genuine StarOps Agents are standard mode; raw-prompt judges (including rawPromptBackend=starops) are stored as type=AGENT but execute as LLM judges.
  • Genuine StarOps Agent evaluator (数字员工): create type=AGENT and OMIT config.agentEvaluatorMode and config.rawPromptBackend entirely. The backend stores agentEvaluatorMode=standard and runs the standard digital-employee agent flow. This is the only form that a task executes as a real Agent (matches built-ins like Builtin.agent_task_completion, which report agentEvaluatorMode=standard, implementation=agent). Do NOT set agentEvaluatorMode=raw_prompt if you want a real Agent evaluator.
  • LLM-style evaluator (LLM-as-judge): create type=AGENT with config.agentEvaluatorMode=raw_prompt and config.rawPromptBackend=direct_llm (or author type=LLM, which the wrapper normalizes to this form). Executes as a single-shot LLM prompt judge.
• WARNING: agentEvaluatorMode=raw_prompt executes as an LLM judge REGARDLESS of rawPromptBackend. rawPromptBackend=starops is still a single raw-prompt judge (the task layer resolves it as type=LLM), NOT a genuine Agent evaluator. If you set type=AGENT + agentEvaluatorMode=raw_prompt, you get an LLM evaluator, not an Agent one. The wrapper's ensure_raw_prompt_backend_default only injects starops when agentEvaluatorMode=raw_prompt is already present; omitting the mode leaves config untouched so it falls to standard mode.
• Define custom evaluation result fields in config.outputSchema. score and explanation are system fields; add business fields such as risk_level, passed, or evidence as additional schema entries.
• Use the raw evaluator-skill commands in references/api-map.md when the user asks to manage skill files. Preview create and update operations; require explicit permission before deletion.

Read references/spec-format.md whenever composing or changing a workflow specification. Read references/api-map.md when extending the flow, diagnosing a failed command, or when the installed plugin version differs from the documented version. Read references/result-analysis.md whenever analyzing evaluation results or low-score cases.

Run the workflow

1. Resolve the skill location from the absolute path of this SKILL.md. Prefer the path supplied by the host environment for the loaded skill. Do not assume a platform-specific skill home. Set the path once and verify that the bundled wrapper exists:

   export SKILL_DIR="/absolute/path/to/alibabacloud-agentloop-evaluation"
   test -f "$SKILL_DIR/scripts/agentloop_eval.py"
   

   When multiple Aliyun CLI profiles exist, scope the workflow without switching the default profile:

   export ALIBABA_CLOUD_PROFILE=<profile-name>
   

2. Check the local CLI and plugin. Add --agent-space to verify cloud authentication and access with a read-only request:

   python3 "$SKILL_DIR/scripts/agentloop_eval.py" doctor --agent-space <space>
   

3. Discover saved evaluators, current built-in evaluator names, and tasks before inventing names or relying on an old built-in alias:

   python3 "$SKILL_DIR/scripts/agentloop_eval.py" discover \
     --agent-space <space> --all-pages
   

   Task discovery defaults to the service's default channel. Add --channel <channel> to inspect another channel. One-shot tasks are not returned by the list API; retain their task IDs.

4. Create a spec from the nearest example, filling only required fields. Never put AccessKeys, bearer tokens, or other credentials in the spec.

5. Preview the evaluator and task requests. This invokes the plugin's --cli-dry-run and does not send the AgentLoop mutation request. The CLI may still resolve or refresh credentials:

   python3 "$SKILL_DIR/scripts/agentloop_eval.py" run --spec /path/to/evaluation.json
   

   If the specification enables continuous evaluation, add --allow-continuous only after the user explicitly accepts the ongoing cost. Use the flag for both preview and execution.

6. Inspect the rendered preview. Present the key parameters (AgentSpace, evaluator names, task name, data scope, estimated records, cost implications) to the user and obtain explicit confirmation before proceeding. Only add --execute after the user approves the preview — even if the user initially requested execution, the preview must be shown and confirmed first:

   python3 "$SKILL_DIR/scripts/agentloop_eval.py" run \
     --spec /path/to/evaluation.json \
     --execute \
     --output /path/to/evaluation-result.json
   

   Execution verifies the AgentSpace, applies explicit evaluator create/update actions, verifies custom evaluator references, creates the task, polls its run, and writes the final raw responses. Use --no-wait only when asynchronous handoff is desired.

7. Resume monitoring an existing task when needed:

   python3 "$SKILL_DIR/scripts/agentloop_eval.py" status \
     --agent-space <space> --task-id <task-id> --wait
   

Analyze evaluation results

Use the read-only analyzer after a run completes or when the user asks why evaluation quality is low. Preview the exact SLS queries first:

python3 "$SKILL_DIR/scripts/analyze_evaluation_results.py" \
  --agent-space <space> \
  --region <region> \
  --from <timezone-bearing-ISO-or-epoch> \
  --to <timezone-bearing-ISO-or-epoch> \
  --task-id <task-id> \
  --preview

Then run the bounded analysis when the user has asked for result analysis:

python3 "$SKILL_DIR/scripts/analyze_evaluation_results.py" \
  --agent-space <space> \
  --region <region> \
  --from <timezone-bearing-ISO-or-epoch> \
  --to <timezone-bearing-ISO-or-epoch> \
  --task-id <task-id> \
  --threshold 0.5 \
  --max-cases 50 \
  --output /path/to/evaluation-analysis.json

Narrow by --run-id or --evaluator-name when available. The analyzer resolves the SLS project from the AgentSpace, queries only evaluation_detail, and returns overview metrics, evaluator-level breakdowns, and low-score case evidence.

Raw input, output, evaluation process, and custom outputs are omitted by default. Add --include-content only when exact customer content is necessary and explicitly authorized.

Success Verification Method

See references/verification-method.md for step-by-step verification commands covering CLI checks, discovery, preview, execution, result analysis, and unit tests.

Cleanup

This skill does not automate delete, cancel, or terminate operations. To clean up resources manually after use:

Note: These are raw aliyun CLI commands and must include --user-agent for observability. Replace {session-id} with the current SKILL_SESSION_ID.

1. Terminate a running task (explicit user authorization required):

   aliyun agentloop update-evaluation-task --agent-space <space> --task-id <task-id> --status Terminated --user-agent AlibabaCloud-Agent-Skills/alibabacloud-agentloop-evaluation/{session-id}
   

2. Delete a task (explicit user authorization required):

   aliyun agentloop delete-evaluation-task --agent-space <space> --task-id <task-id> --user-agent AlibabaCloud-Agent-Skills/alibabacloud-agentloop-evaluation/{session-id}
   

3. Delete a saved evaluator (explicit user authorization required):

   aliyun agentloop delete-evaluator --agent-space <space> --name <evaluator-name> --user-agent AlibabaCloud-Agent-Skills/alibabacloud-agentloop-evaluation/{session-id}
   

Warning: These are destructive operations. Always require explicit user permission before execution. One-shot task and run records are automatically cleaned by the backend after 24 hours.

Best Practices

• Require a timezone-bearing ISO-8601 window or epoch milliseconds for batch backfills.
• Require dataFilter.maxRecords for batch tasks unless the user explicitly accepts an unbounded run and --allow-unbounded is supplied.
• Keep continuous.enabled false unless the user explicitly requests continuous evaluation and understands the ongoing cost; the wrapper requires --allow-continuous when it is enabled.
• Preserve unique evaluator identities within a task. Use evaluatorRef for saved evaluators and name for inline evaluators.
• Create saved evaluators only with an exact supported type: AGENT or CODE (type=LLM is normalized to AGENT). Choose the mode deliberately: a genuine StarOps Agent evaluator requires type=AGENT with config.agentEvaluatorMode OMITTED (backend defaults to standard) — do not add rawPromptBackend. An LLM-style evaluator requires type=AGENT with config.agentEvaluatorMode=raw_prompt + config.rawPromptBackend=direct_llm. Note agentEvaluatorMode=raw_prompt always executes as an LLM judge even with rawPromptBackend=starops, so raw_prompt+starops is NOT a real Agent evaluator. Verify after create with get-evaluator and confirm agentEvaluatorMode is standard (Agent) vs raw_prompt (LLM).
• Use config.outputSchema for custom result fields. When outputSchema is present, include or allow the wrapper to default score and explanation; define extra fields with type, required, rule, and optional range, options, or items.
• Treat evaluator version updates, evaluator-skill create/update operations, task creation, and continuous evaluation as cloud mutations. Preview them first.
• Keep result analysis read-only and time-bounded. Query only evaluation_detail, cap low-score cases at 200, and omit raw customer content unless --include-content is explicitly authorized.
• Do not automate delete, cancel, or terminate operations. Use the raw CLI only after explicit user authorization.
• Report the task ID, run ID, final status, and output file. Include raw service errors without exposing local credential files.

Refresh compatibility

If doctor reports a plugin or API version different from the Architecture line above, inspect these before execution and adjust only the affected mappings:

aliyun agentloop create-evaluator --help
aliyun agentloop create-evaluator-skill --help
aliyun agentloop create-evaluation-task --help
aliyun agentloop get-evaluation-task --help
aliyun agentloop list-evaluation-runs --help
aliyun agentloop get-evaluation-run --help

Reference Links