AdvancingTitans/agent-engineering-toolkit

Agent Engineering Toolkit

English · 简体中文

AET turns AI coding work into portable, inspectable evidence, then compiles that evidence into traceable maps, bounded improvement prompts, and a reviewable read-only Plan before a code Agent edits.

An AI coding Agent says it fixed the task and ran the tests. AET helps answer seven concrete questions:

• Did the change stay relevant to the user's task?
• Did the claimed verification actually run on this workspace?
• Does that proof still apply to the current code?
• Can another Agent review the same evidence without installing AET?
• Can a human trace a conclusion, counter-evidence, and UNKNOWN through a

readable investigation map?

• Can the same evidence become an actionable prompt without silently turning

advice into evidence or granting an Agent permission to edit?

• Before an Agent edits, can its Planner name the required source locations,

linked change sites, verification requirements, and remaining unknowns?

Tools produce reproducible facts. The host LLM may recover intent, form competing hypotheses, call authorized tools, test counter-explanations, and make a conditional engineering judgment. AET's deterministic Grounding API validates the recorded references, permissions, evidence strength, and declared budget usage before rendering. AET can then export the result as a portable, self-describing evidence handoff. A human decides what happens next.

deterministic facts → bounded investigation → grounding validation
                    → Portable Evidence Bundle
                       ├─→ Evidence Atlas
                       ├─→ Human report + bounded Agent prompt
                       └─→ read-only Evidence-Guided Plan
                    → independent review → human decision

Four Quick Skills

Each Skill answers one question, emits one bounded result, and stops.

| Skill | Use it when | Main result | | --- | --- | --- | | /aet-check | Agent instructions, Skills, or completion rules may be unsafe or unverifiable | Up to five evidence-backed engineering findings | | /aet-scope | You need to know whether a diff fits the task, including necessary cross-module work | A disposition for each change group, with counter-explanation | | /aet-proof | A command must be executed now and bound to the current workspace | One minimal JSON proof receipt | | /aet-fresh | You need to know whether an older proof still applies | Exact, relevant-file, artifact, environment, or unknown freshness |

The portable host Skills live in skills/aet-check, skills/aet-scope, skills/aet-proof, and skills/aet-fresh. The deterministic runtime is:

aet quick check .
aet quick scope . --base main --intent aet.intent.json
aet quick proof --output .aet/proofs/auth.json \
  --relevant-path src/auth/session.py -- pytest tests/auth
aet quick fresh --proof .aet/proofs/auth.json

The legacy aet audit, review, trace, and evidence receipt commands remain compatible throughout 1.x. They are advanced native vocabulary, not the default Quick product surface.