Skip to content

kyr0/defuss-vae

v0.9.0MIT

Lean verified agentic engineering: plan, implement, verify, doc, doc-edit, wrap, status, optimize and research skills plus deterministic verify/review/docs lifecycle gates.

doc

Writes and syncs doc pages: grounded claims, prose catalog, anchored review, Mermaid, static check. The agent may start it only after implementation.

doc-edit

Edits only the named doc pages as instructed, with an anchored catalog review, a review page and the static check. The agent may start it after doc, for targeted fixes.

implement

Implements a task or approved plan: minimum code, probes before assumptions, root-cause fixes, real tests, gated. The agent may start it for a human's task or approved plan.

optimize

Researches the target project's verifiers, gates and checks for a cost-vs-assurance Pareto optimum, then writes an optimization plan into plans/ and stops for review. The agent may start it when verification cost is the task or measured verification dominates the session budget.

plan

Plans risky or complex work into plans/ for human review. The agent may start it only without a human plan, for a task too risky or complex to do directly.

research

Deep research on the project's configured LLM, MCP servers, typed decisions and experiments, with a cited report in tmp/vae/research/. The agent may start it when the human asks for deep research, and may call single capabilities when a claim needs an external source.

status

What runs and whether the record matches: sub-agents, services, resources; reconciles the swarm registry. The agent may start it any time.

verify

Deep review against requirements, callers and tests; fixes confirmed defects. The agent should start it at a goal or plan milestone: on the whole change if large, else on its paths and tests.

wrap

Wraps everything up: fresh gate proof, coherent Conventional Commits, CHANGELOG; reflects lessons learned into new memories and consolidates agent memory. Only the human starts it.