hashgraph-online/awesome-codex-plugins / plugins/beastlabai/multi-llm-plugin/skills/multi-llm
multi-llm
FlaggedOrchestrate plan reviews, task generation, implementation and code reviews across several LLM CLIs in parallel, or ask every model one free-text question about a plan. Use when the user says: review my plan, multi-llm, run a code review, generate tasks from a plan, implement this plan, ask the models about my plan.
providers.yaml
View on GitHub# providers.yaml - Multi-LLM provider configuration (BASE layer)
#
# Layers, lowest first, each deep-merged over the one below (a LIST replaces
# the one below wholesale):
# 1. this file (base; always present)
# 2. <git-root>/.multi-llm/providers.yaml (per project; scaffold with `--init`)
# 3. MULTI_LLM_PROVIDERS_CONFIG=/path.yaml (env escape hatch)
#
# What each key means and why: docs/configuration.md#base-config-reference in
# the repository (https://github.com/beastlabai/multi-llm-plugin); a bare
# `#anchor` below is a section of that page. Comments here keep only what you
# need to set a value correctly. Flags, arguments and
# the binary itself live in the adapters under utils/providers/, never here.
# default_timeout: slot seconds. max_concurrent: parallel slots of that CLI.
# models: the --interactive menu, the first being the provider's default.
# Model ids: keep each CLI's own spelling; an id OpenRouter serves may not exist
# in a native CLI catalog. Most providers take an optional ":effort" suffix,
# split on the LAST colon; an unknown suffix stays part of the model name.
# Listing commands, effort rules: docs/configuration.md#provider-blocks.
# Catalog refreshed 2026-11-19.
providers:
claude-code:
default_timeout: 1800
max_concurrent: 2
# ":effort" -> --effort: low, medium, high, xhigh, max. Aliases move with
# releases; explicit ids pin a generation.
models:
- sonnet
- opus
- fable
- haiku
- claude-sonnet-5-5
- claude-opus-5-5
- claude-fable-5-2
- claude-haiku-5-5
# --fallback-model stand-ins; empty = no flag. Each entry must be a model
# listed above, not repeated, not its own primary (#claude-code-keys):
# fallback_models:
# sonnet: [haiku]
# fallback_models_default: [haiku] # primaries with no entry above
fallback_models: {}
# auto | always | never: strip ANTHROPIC_API_KEY/_AUTH_TOKEN/_CUSTOM_HEADERS
# so a subscription login is not billed per token (#claude-code-keys).
strip_vendor_keys: auto
# auto | keep | strip: an inherited CLAUDE_CODE_USE_* backend; `auto` keeps
# only one named in this block's `env:` (#claude-code-keys).
vendor_backend: auto
cursor-agent:
default_timeout: 1200
max_concurrent: 2
# auto | on | off: run cursor on a throwaway profile so --model and --trust
# never touch yours; auto = Linux only (#cursor-agent-keys).
isolate_profile: "auto"
# Effort is part of cursor's own model names; no ":effort" suffix.
models:
- auto
- composer-2.5
- gpt-5.6-sol-xhigh
- gpt-5.6-terra-xhigh
- gpt-5.6-luna-high
- gemini-4-carbon-high
- gemini-4-neon-high
- gemini-4-hydrogen-high
- gemini-3.8-flash-high
- kimi-k3-high
- kimi-k3-max
- glm-5.2-high
- glm-5.2-max
- grok-4.7-high
- claude-sonnet-5-5-xhigh
- claude-opus-5-5-xhigh
- claude-fable-5-2-high
- claude-haiku-5-5-thinking-high
- muse-wonder-1.0-xhigh
- muse-spark-1.3-xhigh
gemini:
default_timeout: 1200 # Gemini can be slower
# No ":effort": set thinking with a custom alias in ~/.gemini/settings.json
# and list the alias here (#provider-blocks).
models:
- gemini-4-carbon
- gemini-4-neon
- gemini-4-hydrogen
- gemini-3.8-flash
- gemini-3.5-flash-lite
- gemini-flash-latest
grok:
default_timeout: 1200
# ":effort" -> --reasoning-effort: none, minimal, low, medium, high, xhigh,
# max. `grok models` lists the catalog.
models:
- grok-4.6:high
fx:
default_timeout: 1200
max_concurrent: 2
# Vercel fx (not the npm/Homebrew JSON viewer of the same name). OpenRouter
# slugs; no ":effort".
models:
- openai/gpt-6-luna
opencode:
default_timeout: 1200
# First opencode version to refuse, e.g. "2.0.0"; "" = no ceiling. Write the
# breakage it guards in a comment beside it (#version-ceilings).
max_version_exclusive: ""
# on | auto | off: run with an empty config home so your opencode.json cannot
# change the model, agent or permissions (#opencode-keys). Quoted on purpose.
hermetic: "on"
# A provider defined only in your own opencode config must be declared here
# to stay usable under `hermetic` (no credentials; #opencode-keys):
# config_providers:
# my-gateway:
# options:
# baseURL: https://gateway.example.com/v1
# models:
# my-model: {}
#
# ":effort" -> run --variant <name>; names are model-specific. Catalog:
# `opencode models`; Zen (`opencode/*`) ids retire server-side.
models:
- opencode/big-pickle
- opencode/nemotron-3.5-super-vl-free
- google/gemini-4-carbon
- openai/gpt-6.2-sol
- openai/gpt-6.2-astra
# Needs stored OpenRouter auth (`opencode auth login`). Leave it without
# ":effort": a --variant hung the earlier V4 Flash, and this one is unprobed.
- openrouter/deepseek/deepseek-v4.1-flash-1017
codex:
default_timeout: 1800
# strip_vendor_keys is deliberately unset (= never): codex has no
# zero-network login check to strip against (#codex-keys).
# ":effort" -> -c model_reasoning_effort=: none, minimal, low, medium, high,
# xhigh, max (ultra = max). Without one, ~/.codex/config.toml decides.
models:
- gpt-6.2-sol
- gpt-6.2-sol:xhigh
- gpt-6.2-astra
- gpt-6.2-astra:xhigh
- gpt-6.1-terra
- gpt-6-luna
# auto | contained | off: what codex's read-only phases do where its own
# sandbox cannot start. A project layer can only narrow it. Never enable
# use_legacy_landlock globally. See docs/configuration.md#safety.
sandbox_fallback: auto
kilocode:
default_timeout: 1200
# backend/provider/model. ":variant" -> run --variant: none, minimal, low,
# medium, high, xhigh, max, thinking, instant; but a variant the MODEL does
# not publish is ignored silently: check tests/fixtures/kilocode/
# model_variants.json (`kilocode models <backend> --refresh --verbose`).
max_concurrent: 2
models:
- openrouter/moonshotai/kimi-k3.1-pro:max
- openrouter/nvidia/nemotron-3-ultra-550b-a55b
- openrouter/z-ai/glm-5.3:max
- kilo/z-ai/glm-5.3
- openrouter/minimax/minimax-m3.1-flash
- openrouter/x-ai/grok-4.7
- openrouter/xiaomi/mimo-v2.6-pro
- openrouter/qwen/qwen4-max:xhigh
- openrouter/tencent/hy4:high
- openrouter/stepfun/step-5-preview:high
- openrouter/openai/gpt-6.2-astra-pro
- openai/gpt-6.2-astra
- openrouter/nex-agi/nex-n2.6
- openrouter/google/gemini-3.8-flash
- openrouter/google/gemini-4-carbon
- openrouter/google/gemini-4-neon
- google/gemini-4-carbon
- openrouter/deepseek/deepseek-v4-pro-0813
- openrouter/anthropic/claude-sonnet-5.5
- openrouter/anthropic/claude-fable-5.2
- openrouter/meta/muse-wonder
cline:
default_timeout: 1200
# <cline-backend>/<model-id>, e.g. openrouter/z-ai/glm-5.3. -m rewrites
# the saved default in ~/.cline, hence the cap. ":effort" -> --thinking:
# none, low, medium, high, xhigh (anything else: cline exits 1).
max_concurrent: 2
models:
- openrouter/z-ai/glm-5.3
- openrouter/moonshotai/kimi-k3.1-pro
- openrouter/openai/gpt-6.2-astra
goose:
default_timeout: 1200
# <goose-backend>/<model-id>; openrouter needs OPENROUTER_API_KEY exported.
# ":effort" -> GOOSE_THINKING_EFFORT: off, none, low, medium, high, max,
# xhigh (ignored silently where goose sees no reasoning model).
models:
- openrouter/z-ai/glm-5.3
- openrouter/moonshotai/kimi-k3.1-pro
- openrouter/openai/gpt-6.2-astra
aider:
default_timeout: 1200
# litellm specs passed verbatim to --model; openrouter/ needs
# OPENROUTER_API_KEY exported. ":effort" -> --reasoning-effort: none,
# minimal, low, medium, high, xhigh.
models:
- openrouter/z-ai/glm-5.3
- openrouter/moonshotai/kimi-k3.1-pro
- openrouter/openai/gpt-6.2-astra
agy:
# Also sets agy's own --print-timeout (30 s under it).
default_timeout: 1200
# Exact display names from `agy models` (an unknown one is rejected).
# ":effort" -> --effort: low, medium, high, e.g. "Gemini 4 Neon (High):high".
models:
- "Gemini 4 Neon (High)"
- "Gemini 4 Neon (Medium)"
- "Gemini 4 Neon (Low)"
- "Gemini 3.8 Flash (High)"
- "Gemini 3.8 Flash (Medium)"
- "Gemini 3.8 Flash (Low)"
- "Claude Opus 5.5 (High)"
- "Claude Opus 5.5 (Medium)"
- "Claude Opus 5.5 (Low)"
- "Claude Sonnet 5.5 (High)"
- "Claude Sonnet 5.5 (Medium)"
- "Claude Sonnet 5.5 (Low)"
- "GPT-OSS 120B (Medium)"
muse:
# Generous on purpose: muse retries a failing call up to 10 times.
default_timeout: 1800
# ":effort" -> --reasoning-effort: none, minimal, low, medium, high, xhigh,
# ultra (anything else fails). Without one, ~/.config/muse/settings.json
# decides. Wonder Full needs a Power or Maximum Usage subscription. muse
# has no OpenRouter provider (`--provider openrouter` is refused); Wonder
# through OpenRouter is kilocode:openrouter/meta/muse-wonder.
models:
- muse-wonder-1.0
- muse-wonder-1.0:xhigh
- muse-wonder-full-1.0
- muse-spark-1.3
- muse-spark-1.3:xhigh
pi:
default_timeout: 1200
# pi model patterns, verbatim to --model (`pi --list-models`). ":effort" ->
# --thinking: off, minimal, low, medium, high, xhigh, max.
models:
- openrouter/z-ai/glm-5.3
- openrouter/moonshotai/kimi-k3.1-pro
- openrouter/openai/gpt-6.2-astra
kimi:
default_timeout: 1200
max_concurrent: 2
# OpenRouter slugs (a Kimi Code subscription is a different catalog). No
# ":effort": a third ":" segment is part of the id.
models:
- moonshotai/kimi-k3.1-pro
qwen-code:
default_timeout: 1200
max_concurrent: 2
# OpenRouter slugs (a Qwen first-party plan is a different catalog). No
# ":effort": a stray suffix reaches the backend as part of the id.
models:
- qwen/qwen3.8-27b
- qwen/qwen4-max
copilot:
default_timeout: 1200
max_concurrent: 2
# OpenRouter slugs, BYOK (a Copilot seat's catalog does not apply).
# ":effort": none, minimal, low, medium, high, xhigh, max.
models:
- openai/gpt-6-luna
vibe:
default_timeout: 1200
max_concurrent: 2
# Aliases, not slugs: each needs a matching [[models]] entry in
# $VIBE_HOME/config.toml. ":effort" -> VIBE_THINKING: off, low, medium,
# high, max.
models:
- openai/gpt-6-luna
omp:
default_timeout: 1200
# omp SELECTORS, verbatim to --model: copy `selector` (not `id`) from
# `omp models openrouter --json`. ":effort" -> --thinking: off, minimal,
# low, medium, high, xhigh, max (omp itself validates nothing).
models:
- openrouter/openai/gpt-6-luna
- openrouter/z-ai/glm-5.3
- openrouter/moonshotai/kimi-k3.1-pro
- openrouter/openai/gpt-6.2-astra
# Default provider when model name doesn't include provider prefix
default_provider: cursor-agent
# ---------------------------------------------------------------------------
# driver: how --exec-subagents dispatches materialized subagent prompts (D3)
# ---------------------------------------------------------------------------
# Only consulted when the Python executor is the invoker (`--exec-subagents`,
# or a host whose capability entry has no native subagent tool). A host with
# native subagents (Claude Code, Cursor) never reads these on a default run —
# the key configures the *executor's* runner, not the session's model.
#
# Merge semantics are the ordinary three-layer chain: this base block is always
# present and fully populated, an override layer merges key-by-key (so setting
# `subagent_runner` alone inherits the rest), a blank value (`subagent_runner:`
# → YAML None) is SKIPPED and keeps the base value, and unknown keys under
# `driver:` are rejected at load time so a typo is never silently a no-op.
# Leaving `subagent_runner` unset in a repo shared by several hosts is the
# healthy arrangement: each host then resolves its own CLI.
driver:
# Any provider:model spec with an implemented adapter. Validated at load time
# (parseable spec, known adapter, declares the profile the phase needs).
subagent_runner: claude-code:sonnet
# Used for tasks whose materialized `capability_tier` is `cheap`; falls back
# to `subagent_runner` when unset.
subagent_runner_cheap: claude-code:haiku
# User preference, not the effective limit: the executor takes the minimum of
# this, the runner provider's own `max_concurrent` above, and the host's
# documented ceiling.
max_concurrent_subagents: 4
# WRITE phases (apply-*, implement) get their own ceiling and it defaults to 1:
# the batchers partition by artifact (one plan file per apply batch, one source
# file per code-fix batch), so the read path's fan-out would be two agents
# editing one file. Raising this only PERMITS parallelism — the batch must also
# declare `can_parallelize`, the declared file sets must be pairwise disjoint at
# dispatch time, and shared-artifact phases stay pinned to 1 regardless.
max_concurrent_write_subagents: 1
# null = the runner provider's `default_timeout`. A positive integer overrides it.
subagent_timeout_s: null
# Seconds a WRITE subagent may print nothing, with no tool call running, before
# it is stopped and what it printed is read: an unreadable closing report after
# edits gets the one report re-ask, as for any runner that died. Not a cap on
# how long it works: a long test suite in its shell keeps the clock off. Only
# for runners that print events as they work (codex); claude-code prints its
# answer at exit, so it never has this timer. 0 turns it off.
subagent_idle_timeout_s: 900
# Re-dispatches for TRANSPORT-shaped failures only (timeout, no parsed text,
# parse_output failure). A result that parsed but failed schema validation is
# never retried — that is a deterministic disagreement a retry would launder.
# A write batch that already changed files is never re-dispatched: a retry
# would apply it twice. It stops as a partial application, or, in an apply
# phase whose closing report was unreadable, is asked once to restate it.
subagent_retries: 1
# The repository's OWN post-write check, run ONCE by the executor after a
# code-editing phase (implement, apply-code-fixes) applied at least one batch.
# A shell command string — `pnpm typecheck && pnpm lint:fix`, `cargo check`,
# `make lint test` — run with cwd at the pinned repo root. The markdown apply
# phases (apply-suggestions, apply-task-suggestions) skip it: a typecheck has
# nothing to say about a plan file.
# null (the default) means THIS REPO DECLARES NO CHECK, and the phase says so
# in one line rather than completing silently — the two are different facts.
# A non-zero exit rolls NOTHING back (every batch was already verified against
# its own snapshot, and a revert would destroy what a human needs to read) but
# ends the phase at exit 3, naming the command and the tail of its output: a
# repo whose own check fails needs a human, not the next phase.
verify_command: null
# Wall-clock ceiling for the command above. Not optional and not inherited from
# a provider — nothing else knows how long this repo's build takes.
verify_timeout_s: 600
# How apply-code-fixes batches run `verify_command` under --exec-subagents:
# `full` (the default) has every batch run it as written. `scoped` has each
# batch run it narrowed to the tests of the files it changed, and the phase
# then owes one full run before it can complete. Worth it only for a slow
# suite; it applies to a pytest command only, and a later batch can break a
# test the narrowed runs skipped (the full run then stops the phase at exit 3).
batch_checks: full
# The declared write policy for edit-capable dispatches, resolved against the
# pinned repo root and enforced TWICE: pre-dispatch against the batch's declared
# file set (refused before the runner starts) and post-run against the set the
# snapshot diff actually observed (batch failed, files restored, run exits 3).
# `.multi-llm/**` is denied to the RUNNER because the EXECUTOR owns those bytes.
# The runner's own sandbox stays the preventive control; this is the second,
# detective layer that makes a violation loud.
write_paths:
allow:
- '**'
deny:
- .git/**
- .multi-llm/**
- '**/.env'
- '**/.env.*'
- '**/*.pem'
- '**/id_rsa*'
- '**/.ssh/**'
- '**/.aws/**'
- '**/.npmrc'
- '**/node_modules/**'
# ---------------------------------------------------------------------------
# progress: how often a running fan-out says it is still alive
# ---------------------------------------------------------------------------
# A closed set: an unknown key is rejected at load. Read once per run.
# docs/configuration.md#good-to-know explains the two cadences.
progress:
# Seconds between `[PROGRESS]` lines; 0 silences them, otherwise >= 5.
interval_s: 30
# Seconds between host polls, and the `--wait` default; >= 5, no 0, clamped
# under the host's shell cap.
poll_s: 90
# ---------------------------------------------------------------------------
# git_env: the non-interactive git environment every spawned CLI inherits
# ---------------------------------------------------------------------------
# auto | wipe | keep. auto wipes credential.helper on the read-only lanes only,
# keep never wipes, wipe wipes on agent-write too. Settable per provider under
# providers.<name>.git_env; MULTI_LLM_GIT_CREDENTIAL_HELPER overrides both.
# An unknown value is rejected at load (docs/configuration.md#git_env).
git_env:
credential_helper: auto
# ---------------------------------------------------------------------------
# code_review: how `review-code` groups the paths in its scope
# ---------------------------------------------------------------------------
# collapse_paths: glob lists (repo-relative; `**/` any depth, a pattern with no
# `/` matches the file name anywhere); `[]` switches a rule off. A path matching
# `generated` (or marked linguist-generated) or `fixtures` is listed, not diffed.
# commit_reviews_dir: relative to the repository root; no absolute path, `..`,
# `.` or empty string. docs/configuration.md#good-to-know has both keys.
code_review:
collapse_paths:
generated: []
fixtures:
- "**/tests/fixtures/**"
- "**/tests/golden/**"
commit_reviews_dir: plans/commit-reviews
# off | on: on passes provider children only an allow-listed environment. Ships
# off until its per-CLI matrix is measured (docs/configuration.md#env_allowlist).
# Quoted on purpose: YAML 1.1 reads a bare on/off as a boolean.
env_allowlist: "off"
# Default models to use when no models are specified on CLI
# If set, skips interactive selection and uses these models automatically
# Use --interactive flag to override and force interactive selection
defaults:
models:
#- cursor-agent:composer-2.5
#- cursor-agent:gpt-5.6-sol-xhigh
- cursor-agent:gemini-4-carbon-high
- cursor-agent:grok-4.7-high
#- cursor-agent:kimi-k3-high
#- opencode:opencode/big-pickle
- opencode:openrouter/deepseek/deepseek-v4.1-flash-1017
# kilocode ignores an unpublished variant silently; these match
# tests/fixtures/kilocode/model_variants.json.
- kilocode:openrouter/moonshotai/kimi-k3.1-pro:max
- kilocode:openrouter/nvidia/nemotron-3-ultra-550b-a55b:high
- kilocode:openrouter/z-ai/glm-5.3:max
- kilocode:openrouter/minimax/minimax-m3.1-flash
- kilocode:openrouter/xiaomi/mimo-v2.6-pro:high
- kilocode:openrouter/qwen/qwen4-max:xhigh
- codex:gpt-6.2-astra:xhigh
- claude-code:claude-fable-5-2:high
#- grok:grok-4.6
#- cline:openrouter/z-ai/glm-5.3
#- goose:openrouter/z-ai/glm-5.3
#- aider:openrouter/z-ai/glm-5.3
#- "agy:Gemini 4 Neon (High)"
#- pi:openrouter/z-ai/glm-5.3
#- omp:openrouter/openai/gpt-6-luna
#- muse:muse-wonder-1.0:xhigh
#- kimi:moonshotai/kimi-k3.1-pro
#- qwen-code:qwen/qwen4-max
#- copilot:openai/gpt-6-luna
#- vibe:openai/gpt-6-luna
#- fx:openai/gpt-6-luna
# Lightweight model subset for the --quick / -q flag
# Used instead of defaults.models for faster reviews with fewer models
# Omit --quick to use the full defaults.models list, or pass --interactive to pick manually
quick_models:
- cursor-agent:grok-4.7-high
- codex:gpt-6.2-astra:xhigh
- claude-code:claude-fable-5-2:high
# Optional: only use defaults for specific modes
# modes:
# review-plan:
# models:
# - cursor-agent:auto
# - gemini:gemini-3.8-flash
# quick:
# - cursor-agent:auto
# code-review:
# models:
# - cursor-agent:gpt-5.6-sol-xhigh
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| receiving-code-reviewobra/superpowersUse when receiving code review feedback, before implementing suggestions, especially if feedback seems unclear or technically questionable - requires technical rigor and verification, not performative agreement or blind implementation | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| requesting-code-reviewobra/superpowersUse when completing tasks, implementing major features, or before merging to verify work meets requirements | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| using-git-worktreesobra/superpowersUse when starting feature work that needs isolation from current workspace or before executing implementation plans - ensures an isolated workspace exists via native tools or git worktree fallback | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| verification-before-completionobra/superpowersUse when about to claim work is complete, fixed, or passing, before committing or creating PRs - requires running verification commands and confirming output before making any success claims; evidence before assertions always | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |