swe-guidelines

Run a benchmark scenario (a skill, a command, or a question) in a declared runtime and have frontier models score it against the rubric. Use to measure a change to a skill or the text.

Run the benchmark prompt in the chatgpt.com, claude.ai, and gemini.google.com products, signed in, and save each answer with its conversation URL as proof.

Record a deliberate deviation from the guideline as an ADR under docs/adr/: the rule quoted, the decision, the consequences. Use when a finding is accepted as intentional.

Explain how the guideline applies to a question, a file, or a change, citing the sections and lenses that govern it. Use when onboarding or before a cross-layer change.

Incorporate a new aspect into the guideline and cascade it through the lenses, skills, docs, and README, naming the release level. Runs in a checkout of the guideline repository.

Architecture review through the Async lenses: infra, queues, workers, park vs fail. For a change in this area, or as one leg of arch-review-full.

Architecture review through the Context lenses: stages, scopes, OperatorContext, authorization, tenancy, provenance. For a change in this area, or as one leg of arch-review-full.

Architecture review through the Contracts lenses: interfaces, injection, wiring. For a change in this area, or as one leg of arch-review-full.

Architecture review through the Delivery lenses: apps, deployment, repo layout, client architecture, logs and telemetry, conventions, substitutions. For a change in this area, or as one leg of arch-review-full.

Full architecture review: the eight lens groups of the guideline run in parallel and merge into one report. Use before a pull request, or when a change crosses layers.

Architecture review through the Network lenses: topology, gateway, public types, clients, realtime, push-first. For a change in this area, or as one leg of arch-review-full.

Architecture review through the Object Model lenses: source of truth, mixins, immutability, identifiers, namespaces, pure rules. For a change in this area, or as one leg of arch-review-full.

Architecture review through the Operations lenses: roles, credentials, ops skills, alarms, scale-out, cost, environments, traffic, READMEs. For a change in this area, or as one leg of arch-review-full.

Architecture review through the Storage lenses: storage principles, tables, translation, roles, migrations. For a change in this area, or as one leg of arch-review-full.

Create a client app: a browser app (portal or operator console) with generated types, one transport, and one realtime channel; or a Python CLI over REST. TypeScript or Python.

Add one entity to an object-model namespace: the frozen type, the table and migration, storage in Postgres and memory, manager operations, wire types, router, and tests. Python.

Create an object-model namespace (a swimlane): the manager interface, types, impl, storage in Postgres and memory, tables, and the wiring into both roots. Python.

Bootstrap a whole new system in the guideline's shape into an empty folder: the monorepo skeleton, the first API, a worker, a portal, deployment, CI, then the first namespace and entity.

Create a web service: the app factory, container, gateway, routers and wire types, service interfaces, health endpoints, the ops CLI, the image, and tests. Python (FastAPI).

Create a worker role: the claim, handle, complete loop over the work queue, lease renewal and self-fencing, liveness, drain-first shutdown, the sweep, the image, and tests. Python.

Upgrade every dependency of the repository to its latest stable or LTS release (runtimes, tools, images, CI, Terraform, libraries), then run the gates and hold back what breaks.