model-based-agentic-software-engineering

Turn the agent-governance-mechanisms catalogue on the current repository — the governance skill of the MAGE method. Two modes: AUDIT (survey the repo for missing guardrails and propose a prioritized adopt/adapt/skip plan) and INTERPRET-FAILURE (a failure just recurred — classify it and convert the class into a durable governance mechanism: a constraint that prevents it or a sensor that detects it). Also sets an ambient stance from the Davis AI-First Engineering Method. Use when the user wants to harden an agent-collaborative codebase, reduce recurring agent-caused regressions, set up guardrails / lints / gates / typed seams, review their governance posture, or when a failure class just recurred and should be prevented structurally, not only patched. ALSO use at DESIGN time — reviewing a new subsystem, component, state machine, queue, cross-service seam, or data model; or introducing concurrency, a trust boundary, an irreversible operation, or duplicated state/logic — to catch predictive smells whose failure c

Operate an agent-fleet repository as a DevOps engineer — the positive lifecycle map of how the substrate works (managing agents, context windows, the git repo, deploys, the dev machine, cron, and the orchestrator's own hooks), plus a symptom → resolving-doc catalog for when something breaks. Use when running an orchestrator/operator session: dispatching or recovering agents, keeping the mainline deployable, reclaiming disk, colima/host-tool trouble, cron or merge-train health, hook misbehavior, or RCA of a substrate failure. LEADS with how things should work; troubleshooting is the fallback. Partner to the self-governance skill — operate routes a break to its fix, and when a failure RECURS, hand it to self-governance to convert into a durable control. NOT for the product's domain logic (its checkers, generators, business rules) — that is product code, not repo operation.

Write and diagram engineer-facing prose well — the third partner skill in the govern / operate / communicate trio. It governs the prose the fleet AND its operator produce: the fleet's docs (a control's description, a mechanism entry, a design doc, a runbook, a handoff, a README, a tutorial, an ADR, an API reference) AND the operator's own reports to the human (a status report, explaining a tradeoff, surfacing a decision — the orchestrator's prose to the user is prose too). Use when authoring or auditing any of it. Two legs. PROSE — the rhetoric toolkit (classical figures, used with variety so the prose doesn't read as machine-uniform), the Diátaxis engineering register (tutorial / how-to / reference / explanation), a house lexicon that names concepts consistently, the target voice, and an audit procedure that grades a passage and emits concrete fixes. VISUALIZATION — technical-diagram types, Mermaid-first, with a narrow hand-authored-SVG escape hatch. Also bootstraps a house vocabulary by walking a codebase a