Agent engineering toolkit

Agent Engineering Toolkit

Current Skill version: 1.13.0 (Quick compatibility + opt-in Lab)

Activation policy

Default: OFF. Installing this Skill does not authorize its use. Load or run AET only when the user explicitly asks to use AET for the current task (for example, “use AET”, “run aet audit”, or “produce an AET Evidence Pack”). A repository containing AET files, a generic request to test/review a change, or the availability of the aet executable is not opt-in. Do not carry opt-in across tasks.

After explicit activation, choose only the smallest surface needed. Daily work should route to /aet-check, /aet-scope, /aet-proof, or /aet-fresh. Each Quick Skill emits one bounded result and stops. It never enters another Quick Skill or AET Lab automatically.

Use the aet CLI as the source of truth. The host agent may choose its own shell or package runner, but must preserve the commands' exit status and attach the emitted evidence instead of paraphrasing it as unverified fact.

<!-- aet-learn:immutable -->

UNKNOWN is never a pass. Only aet trace executes explicit argv after --. Audit, review, and Evidence Pack compilation stay deterministic and local. AET may propose, replay, gate, and stage a Constitution-bound asset candidate, but it never adopts a candidate, commits it, pushes it, or lowers an evidence contract automatically.

<!-- aet-learn:end -->

Quick routes

This section applies only after explicit activation. Choose one initial surface. If the requested AET surface is ambiguous, ask which claim needs evidence; do not infer permission for a broader workflow.

User need	Preferred Skill	Deterministic CLI
Check Agent instructions and Skills	/aet-check	aet quick check
Investigate whether a diff fits the task	/aet-scope	aet quick scope
Execute and record one real verification	/aet-proof	aet quick proof
Check whether a proof still applies	/aet-fresh	aet quick fresh

When a slash command and its question are both Chinese, the host explanation uses natural Simplified Chinese while code and required technical terms remain English. Every other request uses English.

Advanced and Lab compatibility

The legacy audit, review, trace, Evidence Pack, Context, Decision, Repository Audit Showcase, Evolve, Quality, Learn, Gate, Shadow, Stage, and Adopt surfaces remain available for 1.x compatibility. They are not default Quick routes. Load only the matching reference after the user explicitly asks for that advanced capability.

Load one workflow reference

After routing, load only the matching reference; do not preload the whole AET manual into the Agent context.

Surface	Read only when needed
Audit, Review, Trace, Pack, Run	delivery workflow
Repository Audit Showcase	repository audit showcase
Context, Decision, Evolve	provenance workflow
Quality diagnose/promote	quality workflow
Learn, real-host Gate, Stage/Adopt	evolution workflow
Runner, privacy and trust boundaries	security boundaries

<!-- aet-learn:editable id="routing-guidance" -->

For repeated Evidence Only failures, use harvest → inspect → mine → propose. For observed behavior, freeze aet learn plan before any rollout, then pass the exact plan to aet learn gate --gate-plan <plan.json>. Use --resume only for an exact observed replay binding. Gate may stop on a hard regression or a pre-registered statistical boundary; history is planning-only and never enters PASS. Stage remains human review, and only explicit adopt --yes may write the target. Never auto-adopt, commit, push, release, or treat INCONCLUSIVE as PASS.

<!-- aet-learn:end -->

Portable use

This folder is the canonical, tool-neutral Skill. Install or load the complete folder in any host that supports SKILL.md folders. If the host has no native Skill loader, place this file in that agent's project instructions and ensure the aet executable is available. The optional agents/ metadata is OpenAI-specific UI metadata; other hosts may ignore it.

For compatibility rules and output contracts, read only when the chosen route requires them: cross-agent use, then the applicable audit contract, review contract, or Evidence Pack and Trace contract, then the v1 product contract.

The immutable boundary above always applies. Detailed runner and privacy rules live in security boundaries. AET is not a general benchmark, LLM-Judge platform, semantic RCA engine, auto-repair daemon, release bot, or online ticket/metrics platform.