bello-config-advisor

Bello Config Advisor

Use the task file, production repository, current model/effort information, and the user's quality, cost, and time preferences to choose one usable Bello setup. Include an explicit planning decision. With no stated preference, favor the least expensive setup that still has strong expected quality.

This skill targets Bello 0.7.2. The host where the user talks to you does not determine Bello's models or billing routes. Advice is read-only: do not apply settings, authenticate, install dependencies, create a plan, or start a run until the user asks.

Inspect

1. Read SELECTION_POLICY.md. Use CONFIG_SCHEMA.md to construct the internal config and OUTPUT_CONTRACT.md for the response.
2. Run python <SKILL_DIR>/scripts/inspect_models.py, using the Python executable available in the host. It reads bello runtime models --engine all, not the host agent's model list. Keep the result in a temporary file outside the project for validate_config.py --catalog. No model generation is requested. A supplied catalog can be read with --file. Do not silently replace an unavailable catalog with a fixed list of OpenAI models.
3. Read MODEL_ECONOMICS.md. Refresh the relevant official model pages for price, capability, speed, and effort. The OpenAI, Claude, and OpenRouter catalogs are entry points, not three pages to read exhaustively on every task. Read details for plausible available profiles, and use approximate comparisons internally where exact measurements are absent. Account for the user's available subscriptions/API connections and spending preferences; never read or print credentials.
4. Resolve the project root and task file. Read the task from its file. Require a regular file inside the project and normalize task to its project-root-relative path with / separators.
5. Inspect task-relevant production files: applicable workspace instructions, manifests, interfaces, existing implementations, and the current diff where useful. Nearby implementations are important: they may supply the architecture and make a cheaper executor sufficient. Use rg --files or git ls-files first. Ignore the project's tests, CI, benchmark artifacts, previous runs, telemetry, hidden Bello history, and saved configuration when selecting the setup. A testing requirement in the task remains part of the task. Public model comparisons in the official documentation are model information, not project-run evidence.
6. Extract hard constraints and softer preferences. Ask one focused question only when missing information materially changes the choice; otherwise make a reasonable internal assumption. Do not inspect the saved Bello config until the user asks to apply or launch the chosen setup.
7. Do not run tests, builds, installers, the task's code, configuration editors, or plugin updates just to make a recommendation.