Alibabacloud Agentloop Management

AgentLoop Skill Router

Positioning: This skill is the single entry point for Alibaba Cloud AgentLoop requests. It only classifies the user's intent and dispatches to one of the four domain playbooks below. All executable rules - prerequisites, credentials, RAM policies, parameter confirmation, safety protocols, command usage, and verification - live inside the domain files. Do not run any cloud operation before reading the matched domain file.

Compatibility: cloud-operation domains require Aliyun CLI 3.3.15 or later; Pipeline requires aliyun-cli-agentloop 0.7.4 or later; bundled evaluation and Pipeline scripts require Python 3.8 or later.

Routing Table

Dispatch Rules

1. Classify the request into one or more domains using the routing table, then read only the matched domain entry file(s). Never preload all domains.
2. Follow the matched domain file completely. Each domain defines its own prerequisites, credentials check, RAM policies, parameter confirmation, execution-safety protocol, and verification method.
3. If the request matches none of the domains, state that it is out of scope for this skill and do not dispatch.
4. If the intent is ambiguous between two domains, ask one clarifying question before dispatching.

Disambiguating Dataset vs Pipeline vs Evaluation

• Writing or reading rows the user already has: Dataset.
• Deriving new rows from LogStore or trace data through processing nodes: Pipeline. Create or confirm the sink Dataset first.
• Judging the quality of existing traces or Dataset rows with an evaluator: Evaluation.

Delegated Domain: Experience

Experience work - recalling prior experience, similar cases, past incidents and fixes, old runbooks, lessons learned, and the lifecycle of experience stores (ContextStore) and their API Keys - is not implemented in this skill. It lives in the separate alibabacloud-agentloop-experience skill.

When a request needs experience, on its own or as one step of a multi-domain request:

1. Check whether the alibabacloud-agentloop-experience skill is available in the current environment.
2. If it is available, hand the experience part off to it and follow that skill's own rules. Do not reimplement recall or ContextStore commands here.
3. If it is not available, tell the user that this part requires the separate skill and point them at https://skills.aliyun.com/skills/alibabacloud-agentloop-experience to install it. Offer to help with the installation, and wait for the user's answer.
4. Never guess at experience behavior in place of the missing skill. Continue with the remaining in-scope domains and report the experience step as blocked on that skill.

Multi-Intent Handling

• Execute multiple domains sequentially in dependency order; finish and verify one mutation stage before starting the next.
• For Logstore-to-Dataset materialization: confirm or create the Dataset schema, preview the Pipeline, create and observe the Pipeline run, then read back and reconcile Dataset contents. Start Evaluation only after the Dataset field contract passes.
• When the user also asks to reuse prior work, resolve that experience step through the delegated skill above before the in-scope domains start, and say so if the skill is missing.

Shared Conventions

• Session ID: generate one 32-character lowercase hex session ID once at the start of the workflow (openssl rand -hex 16) and reuse that same value for the rest of the session. Keep the generated value and write it out literally in every command that needs it. Do not re-derive it per command, and do not reach for it through a shell variable or a cat of a saved file - either one forces an assignment in front of the call and breaks the command shape rule below.
• User-Agent: every aliyun CLI cloud API command must carry --user-agent "AlibabaCloud-Agent-Skills/alibabacloud-agentloop-management/{session-id}". Bundled Python wrappers read SKILL_SESSION_ID; local configure, plugin, and version commands are excluded.
• Command shape: every cloud API call must run as a single-line bare command whose first token is aliyun, or python3 for a bundled wrapper, and whose last token is the final flag of that same call. Nothing may come before it - no VAR=value assignment, no set -o pipefail, no source, no cd, no bash some_script.sh wrapper - and a newline between an assignment and the call still counts as coming before it. Nothing may come after it either - no | tee, no | head, no 2>&1, no > file redirect, and no && or ; chaining onto a second command. Diagnostic probes such as --help follow the same rule. When the environment asks for a log of executed actions, run the bare call first and then write the command text and its output into the log as a separate file-write step; a single action log listing each command and its result satisfies that requirement in full, so piping a call into tee adds nothing and only corrupts the record of what ran. The command that executes must be the API call itself and nothing else, so that run records, audit trails, and CLI tooling all see it verbatim.
• Credential red lines: never read, echo, or print AK/SK/STS-token values or the APM LicenseKey (entryPointInfo.authToken) - in chat answers, summaries, credential tables, generated snippets, or report files. Keep every retrieved credential inside an environment variable, report only whether it was obtained, and reference the variable name instead of the value. Never ask the user to paste literal credentials; never run aliyun configure set with literal credential values; use only aliyun configure list to check identity status. Onboarding redaction recipe: references/onboarding.md.
• RAM permissions: use references/ram-policies.md as the skill-wide index. Never put * in an Action list. Grant destructive actions separately and deliberately.
• Resource names: confirm each resource's exact naming contract before create; Dataset and Pipeline use different character sets.