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.
pyproject.toml
View on GitHub[project]
name = "multi-llm"
version = "0.1.9"
description = "Multi-LLM automation for plan review, task generation, implementation, and code review"
requires-python = ">=3.10"
dependencies = ["pyyaml>=6.0"]
[dependency-groups]
dev = ["pytest>=7.0", "pytest-cov>=4.0", "pytest-xdist>=3.0", "ruff>=0.1.0", "jsonschema>=4.0"]
[tool.pytest.ini_options]
markers = [
"robustness: marks tests for LLM error handling robustness scenarios",
"live: marks tests that require Claude Code CLI and API (skipped in CI)",
"timeout: marks tests with a timeout limit (requires pytest-timeout)",
"config_override: opts out of the autouse per-project config isolation fixture (tests that exercise override discovery/env themselves)",
"model_catalog: opts out of the autouse stub that pins OpenCodeProvider.describe_unlisted_model to known=None (tests that drive the real `opencode models` read, or patch it themselves)",
"debug_config: opts out of the autouse stub that answers `opencode debug config` with a failed read (tests that drive the reader themselves, with a fake binary or a patched spawn)",
"sandbox_probe: opts out of the autouse guard that answers a real provider CLI's sandbox probe with `unknown` (tests that deliberately drive the developer's own `codex sandbox`)",
"shell_probe: lets a test's own sandbox probes be shell text or a prefixed command line (`sh -c …`, `python -c`, `env …`, `timeout …`), which the autouse probe guard otherwise refuses because it cannot see which files that text runs (the gate's tests whose fake probes are `sh -c` scripts)",
"slow: marks tests that drive several real orchestrator invocations end to end (the instruction-refactor golden guard); runs in CI, not deselected by default",
"container: marks tests that build and drive the E2E container image (needs a Docker daemon; opt in with -m container). Deselected by default via addopts below",
"package: marks rows that need a packaging artifact plain CI does not build (an `npm pack` tarball, a staged tree). Named for what it *needs*, not for npm specifically, so a second host's packaging plan can reuse it. Deselected by default via addopts below and run by the Linux-only PR package job with `-m \"package and not container\"`; note `pytest -m package` with zero matches exits 5, so a per-host leg must not run it for a host that declares no such rows",
"live_provider: marks the credentialed provider lane (harness-expansion.md §7.4) — a real CLI against https://openrouter.ai on the user's own key, so it SPENDS CREDITS. Carried in addition to `container`, so the addopts default already deselects it and CI's `not container` does too; `make e2e-live-provider` is the only thing that sources the key",
"live_native_phase: marks the host lane's claude-code per-phase rows (chain-lane-write-runner-matrix.md §3.4) — each one a native Task dispatch the host performs itself, anchored on `claude-code-apply-native`'s measured $0.586302 pass rather than on a read row. Carried IN ADDITION TO `container` and `live`, so the addopts default and CI's `not container` keep excluding it with no further edit, and `make e2e-container-host` excludes it explicitly (`-m \"container and live and not live_native_phase\"`) so the cheap tier's cost does not move. `make e2e-container-host-phases` is the only target that selects it, with its own MLLM_E2E_BUDGET_USD: the $2.00 default is a BETWEEN-ROWS pytest.exit, and a ~$3-5 tier crosses it around the fourth row, stranding the rest",
"live_next_command: marks the host lane's T10 next-command rows (next-phase-command.md T10) — one `generate-tasks` row per host and route plus a short typed follow-up session, graded by `TestTheNextCommandLiveCheck`. Carried IN ADDITION TO `container` and `live`, so the addopts default and CI's `not container` keep excluding it, and `make e2e-container-host` excludes it explicitly so the ordinary tier's cost does not move. `make e2e-container-host-next-command` is the only target that selects it",
"landlock: marks tier 2/3 rows of the plugin-landlock sandbox rung (codex-sandbox-unavailable-findings-and-fixes §8.1): the real `utils/landlock_exec.py` against the real kernel. They need a kernel whose Landlock ABI query returns >= 3 and SKIP otherwise, naming the errno or ABI found; `MULTI_LLM_REQUIRE_LANDLOCK=1` turns that skip into a failure, so a host that must run them cannot pass by skipping. Not deselected by default: zero-spend and fast where the kernel supports it",
"codex_jail: marks tier 3 of the plugin-landlock rung — the installed `codex` inside the real jail, against the localhost stub (throwaway CODEX_HOME, zero spend). Opt in with -m codex_jail; deselected by default via addopts below, like `container`, because it drives the developer's own codex binary. Carried IN ADDITION TO `landlock`, so it also skips without a usable kernel",
]
# A marker deselects nothing on its own — it only gives a -m expression something
# to exclude. This default is what keeps a bare `pytest` green on a machine with no
# container runtime. It is a *default*: an explicit -m on the command line replaces
# it, which is why .github/workflows/test.yml has to say `not container` (and
# `not codex_jail`) itself.
addopts = "-m 'not container and not package and not codex_jail'"
# Filter out warnings about dataclass fixtures being collected by pytest
filterwarnings = [
"ignore:cannot collect test class.*FixturePlan.*:pytest.PytestCollectionWarning",
]
# Lint gate (stage 1). The command is `make lint`: it passes this file with
# --config because packaging/ sits outside the project directory, so ruff would
# not find it there. With --config the globs below resolve from the repo root.
# See docs/development/ for the stages that follow.
[tool.ruff]
extend-exclude = [
"packaging/claude/multi-llm",
"packaging/codex/multi-llm",
"packaging/opencode/dist",
"packaging/opencode/skills",
]
[tool.ruff.lint]
select = ["E9", "F", "B006", "B008", "B023", "B904", "PLE"]
ignore = ["PLE2502"]
[tool.ruff.lint.per-file-ignores]
# The tests keep only the rules that find real defects: E9, F82x (undefined
# names), F63x and F70x. Everything else the selection brings in is dropped for
# them: unused imports and variables (F40x, F84x), format-string checks (F5),
# dict and assert checks (F60x, F62x), annotations (F72x), `__all__` (F90x),
# fixture-name shadowing (F811), the B and PLE families, late imports (E402),
# one-line statements (E702) and short names (E741) are how the suite is written.
"skills/multi-llm/tests/**" = [
"B", "PLE", "E402", "E702", "E741",
"F40", "F5", "F60", "F62", "F72", "F811", "F84", "F90",
]
# These two modules export their moved names lazily through a module-level
# `__getattr__`, so `__all__` lists names the module does not bind.
"skills/multi-llm/utils/subagent_executor.py" = ["F822"]
"skills/multi-llm/utils/write_batch_scheduler.py" = ["F822"]
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: 321k | Updated | Claude CodeCodex | Verified | Install |
| requesting-code-reviewobra/superpowersUse when completing tasks, implementing major features, or before merging to verify work meets requirements | Stars: 321k | 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: 321k | 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: 321k | Updated | Claude CodeCodex | Verified | Install |