AgentSkillsCN

Agent Tooling Discipline

代理工具规范。

SKILL.md

Skill: Agent Tooling Discipline (Structured Outputs Only)

Intent

  • Enable an agent to produce machine-checkable outputs (structured, versioned, rejectable) alongside concise human-readable summaries.
  • Ensure agents can be safely automated: outputs are deterministic, validated, and compatible with conformance gates.

Scope boundaries

  • In-scope: structured output discipline, error/diagnostic contracts, validation and rejection conditions, and replayability expectations.
  • Out-of-scope: implementation of agent tools/CLI; output formatting and UI details.

Primary concepts

  • Structured output: well-defined fields and identifiers that validators can accept/reject deterministically.
  • Rejectability: ability to fail fast on unknown fields, incompatible versions, missing provenance, or nondeterminism.
  • Replayability: same declared inputs produce same structured outputs.

Required inputs/context

  • Tool contract index: what structured outputs are expected per operation (validate/plan/explain/audit) and their versions.
  • Error taxonomy: categories/codes with remediation guidance expectations (conceptual).
  • Determinism & provenance requirements: what must be included for traceability and audit.
  • Golden structured outputs: representative examples used as oracles for validators.

Expected outputs

  • Structured outputs (minimum):
    • Operation result envelope: operation identifier, input contract versions referenced, output contract version produced.
    • Diagnostics: errors/warnings with category/code, stable location references, and remediation hints.
    • Provenance hooks: stable references that allow explain/provenance tracing.
  • Human-readable outputs:
    • Concise summary of what was validated/emitted/denied and why, consistent with structured diagnostics.

Acceptance criteria

  • Machine-checkable: outputs can be validated without ambiguity.
  • Deterministic: stable ordering and stable identifiers for repeated runs.
  • Strict: unknown fields or incompatible versions are rejected (especially for regulated workflows).
  • Consistent: human-readable summaries do not contradict structured outputs.

Validation signals

  • Schema/contract validation rejects unknown fields, missing required fields, and incompatible versions (deny).
  • Determinism checks fail if ordering/serialization changes without semantic change (deny).
  • Explain/provenance checks fail if required provenance anchors are missing (deny).

Guardrails & forbidden behaviors

  • Forbidden: free-form outputs that cannot be validated or deterministically parsed.
  • Forbidden: hiding important decisions only in prose (must be in structured outputs).
  • Forbidden: “best-effort” acceptance of unknown fields (must reject).
  • Escalation (HITL required):
    • Any weakening of validation strictness for high-assurance workflows.

Used by roles

  • All roles (cross-cutting requirement)