WB-Bench Run Setup

Interactively guide the user from an unconfigured checkout to a running (and then analyzed) WorkBuddy Bench evaluation. This is a seven-phase playbook (0–6). This file is the router: read the matching references/ file completely before executing each phase, complete one phase at a time, and confirm with the user before moving on.

Phases → reference files

Phase	Goal	Read
0. Datasets	subset(s) downloaded under datasets/	references/00-datasets.md
1. Environment	a checkout that can run	references/01-environment.md
2. Model	configs/models/<provider>/<slug>.yaml	references/02-model.md
3. Credentials	.env URL + key for the chosen names	references/03-credentials.md
4. Job	configs/jobs/<slug>.yaml	references/04-job.md
5. Launch	dry-run, confirm, run	references/05-run.md
6. Analysis	hand off to wbbench-report-skills	references/06-analysis.md

Interaction rules

• Reply in the user's language. Detect the language of the user's messages and use it throughout (they write Chinese → you answer in Chinese). Config keys, file contents, and shell commands stay verbatim; only your prose adapts.
• Read the live templates — don't trust field names memorized here or in the references. Schemas evolve; the authoritative sources are configs/models/_template.model.yaml, configs/jobs/_template.job.yaml + configs/jobs/_reference.yaml, and .env.example.
• Discover choices at runtime, never hardcode a list that may be stale: ls configs/harnesses/*/versions/, ls -d datasets/wb-bench-*/. Datasets are not in git — a fresh checkout has none until Phase 0 downloads them, so an empty datasets/ means "download first" (see references/00-datasets.md), not "no datasets exist".
• One phase at a time. Show the user what you're about to write, then write it.

Guardrails (apply across all phases)

• Secrets (API keys, base URLs) live only in .env. YAML references env-var names only — never a real key or URL. Never git add .env; never echo secret values back.
• The model wire-protocol key is protocols (plural). The manifest resolver reads only that and defaults to openai when it's absent, so a singular protocol: silently mis-routes a non-openai backend. Always write protocols. (Details in references/02-model.md.)
• model_connection: local_proxy is the recommended default — a job-private proxy that adds protocol bridging, extra_body injection, and full request logging at the transport layer without altering content, and run.sh starts it for you. Prefer it unless the user asks for direct. It is mandatory (set it automatically, don't make the user discover it) when the model uses extra_body (top_k / min_p / chat_template_kwargs, ...) or the harness protocol isn't in the model's protocols; direct fails fast in both cases. (Details in references/04-job.md.)
• Before dry-run, verify every env-var name the model references exists in .env.
• Ask the user to confirm before the real run (Phase 5) — it is time- and cost-intensive.
• Treat dataset/task files as read-only; only write under configs/ and .env. Datasets come from HuggingFace via scripts/dataset/fetch-dataset.sh (Phase 0); never hand-edit downloaded task files.