# multi — example config with every backend type and every field.
# Real file: ~/.config/multi/config.toml. `scripts/setup.sh init` writes a
# smaller default there; copy sections from here as you need them.
# Keys are never in this file — see providers.example.env.

default_profile = "normal"
verbose_prompt = false              # true: skills show each full prompt before sending, without waiting for approval
# no_train = false                  # true: keep only models with training verdict no; unknown sits out too
# Account overrides apply to every model of a backend, ahead of built-in rules.
# Put trains = false under that backend only when your account does not train.

# ---- backends: a name, a type, and what that type needs -------------------

[backends.codex]                     # type "codex": the Codex CLI, read-only sandbox
type = "codex"
models = []                          # empty = the CLI's default; at most one model here
timeout = 600                        # seconds for one run; default 300 (claude-headless: 1800, a ceiling)
# trains = false                    # e.g. ChatGPT with training off; true = may train, on any backend

[backends.opencode]                  # type "opencode": OpenCode CLI, plugin's read-only agent
type = "opencode"
models = []                          # empty = a free model from `opencode models`, others as fallback
stall = 180                          # opencode, kilo, claude-headless: seconds of silence before it counts as dead (headless default 600)

[backends.kilo]                      # type "kilo": Kilo Code CLI (an opencode fork), same read-only agent; not in the default profile
type = "kilo"
models = []                          # empty = a free kilo/...:free model from `kilo models` (never the kilo-auto / openrouter routers)

[backends.opencode-paid]             # the same type twice, under another name
type = "opencode"
models = ["opencode-go/deepseek-v4-flash", "opencode-go/qwen3.8-flash"]   # first is tried first

[backends.openrouter]                # type "claude-headless": claude -p against an
type = "claude-headless"             # Anthropic-compatible endpoint. OpenRouter is one.
base_url = "https://openrouter.ai/api"
models = ["qwen/qwen3.8-flash", "deepseek/deepseek-v4-flash-0731", "z-ai/glm-5.2:free"]
# api_key_env not set = OPENROUTER_API_KEY (backend name upper-cased + _API_KEY)

[backends.glm]                       # a second endpoint: z.ai, 9router, Moonshot, self-hosted
type = "claude-headless"
base_url = "https://api.z.ai/api/anthropic"   # https required; plain http only on localhost
models = ["GLM-5.3-Flash"]
api_key_env = "ZAI_API_KEY"          # this backend reads its key from that variable

[backends.local]                     # a router on this machine, like 9router or omnirouter
type = "claude-headless"
base_url = "http://localhost:4000/api"
models = ["local/coder"]
api_key_env = "LOCAL_ROUTER_KEY"
timeout = 1200

[backends.deepseek]                  # DeepSeek's own Anthropic-compatible endpoint, key DEEPSEEK_API_KEY
type = "claude-headless"
base_url = "https://api.deepseek.com/anthropic"
models = ["deepseek-flash"]          # V4.1 Flash; claude-* names are mapped to it too
# Exact model only; explicit --effort overrides every supporting backend.
# No entry: Codex keeps medium, other harnesses decide. Gemini unsupported.
# effort = { "deepseek-flash" = "low" }
# Windows this backend sits out, in UTC only (the zone word is required).
# DeepSeek bills double in these hours (its published peak, as of 2026-09-10 —
# check api-docs.deepseek.com/quick_start/pricing, they moved on 2026-08-16).
# Any backend can carry a list: "Sat-Sun 00:00-24:00 UTC", "22:00-02:00 UTC"
# (every day, wraps midnight). Decided once, at launch: a run that starts in
# an open hour finishes. Inside a window the backend is not launched and its
# answer file says so — it is named in the report, not skipped in silence.
avoid = ["Mon-Fri 01:00-04:00 UTC", "Mon-Fri 06:00-10:00 UTC"]

[backends.gemini]                    # type "gemini": the Gemini CLI, key GEMINI_API_KEY
type = "gemini"
models = []                          # empty = CLI default; at most one model here

# ---- profiles: who runs, in parallel --------------------------------------
# An entry is a backend name (its whole models list, in order) or name:model
# (exactly that model, no fallback). The same entry twice runs twice.
# ask.sh --backend <profile> picks one; no --backend = default_profile;
# --backend a,b:model is a one-off; `all` = every backend above.

# OpenRouter's free and flash pools queue: measured 2026-09-06, 25–40 minutes
# for one review while the same model on a direct key took 5. They belong in a
# profile you pick on purpose, not in the one every review runs with.
[profiles]
normal  = ["codex", "glm"]
free    = ["opencode", "openrouter:z-ai/glm-5.2:free"]   # pinned: the bare backend walks its paid models first
paid    = ["codex", "opencode-paid", "deepseek"]   # in deepseek's peak hours this runs two reviewers, and says so
compare = ["openrouter:qwen/qwen3.8-flash", "openrouter:deepseek/deepseek-v4-flash-0731", "glm"]
solo    = ["codex"]