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.
hosts/capabilities.yaml
View on GitHub# Host capability data (D5b) — one entry per host id, validated against
# schemas/host_capabilities.schema.json.
#
# WHAT THIS FILE IS. The verb -> mechanism map the host-neutral instruction core
# defers to. The core says *spawn subagents* / *ask the user* / *run detached*; the
# three fields `subagents`, `ask` and `detach` say what each one resolves to for a
# given host. That resolution is paid **once**, when a shim is rendered, and never
# per run by an agent.
#
# WHAT THIS FILE IS NOT. It is not on any agent read path. It is consumed by
# `multi_llm.py`, by `install-host`, by `hosts/render_capabilities.py` and by the
# tests. No instruction file and no shim ever tells an agent to open it or to open
# `references/host-capabilities.md` — the shim's baked sentence *is* the answer this
# matrix would have given. An agent reasoning over a capability table mid-workflow
# would be a worse instruction-following failure mode than the one this design
# removes.
#
# PROVENANCE IS LOAD-BEARING. Host docs churn faster than this design does, so
# every value carries a provenance: `verified` (a recorded live smoke observed it),
# `documented` (host docs only) or `unknown` (the shipping default). **A
# `documented`-only value never selects the non-conservative path.** Concretely,
# today: Codex and OpenCode are `subagents: executor` because spawn-from-exec is
# unconfirmed, and `detach: python`/`required` rather than trusting a raised command
# timeout. Promoting either is a one-line edit here, not a shim rewrite — but only
# once a smoke run has recorded it.
#
# A cell is `documented` until a recorded run has observed it, and most of them
# still are — which is why the conservative resolution is what executes. The
# `verified` cells each name the run that promoted them in a comment beside the
# entry, and `tests/test_host_capabilities.py`'s smoke ledger fails any promotion
# that does not.
version: 1
hosts:
claude-code:
# The reference host. There is deliberately no hosts/claude-code/shim.md.tmpl:
# Claude Code loads a skill from <skill-root>/SKILL.md and nowhere else, so a
# copy under hosts/ could only be dead text or a duplicate to keep in sync.
# `installable: false` -> `install-host claude-code` is --check-only, and the
# shim-correctness test asserts SKILL.md's three hand-written verb sentences are
# byte-identical to what the renderer emits for this entry.
shim: {template: null, installed: SKILL.md, format: markdown+frontmatter}
# Rationale: this host's mandatory manifest is 23 lines on its own (a long
# `description`, the rendered `argument-hint`, and 18 `allowed-tools` grants
# that cannot be shortened without losing a permission). The shim *body* — what
# the ~100-line budget is actually about — is 97 lines and still inside the
# 100-line contract `DEFAULT_SHIM_MAX_LINES` holds every shim to; this key only
# ever absorbed the frontmatter, and 120 is 23 + 97 exactly, so the next
# addition is the same deliberate decision again rather than slack to spend.
#
# 110 -> 120 (2026-09-22): step 4 gained the truncation rule. Measured, not
# defensive: Claude Code persists a Bash result over ~30 000 characters to a
# file and returns a ~2.2 KB stub holding the output's *first 2 KB*, and seven
# of the eleven cores are over that cap — so the read step this shim exists to
# name silently delivered a fragment. Run 20260922T191142Z has one row that took
# the fragment for the whole, never spawned a subagent, and reported four
# batches applied with zero edits made. Step 4 now reads the core through
# `util instructions <mode> --to-file` and a `Read` of the file it names
# (docs/development/hosts.md, "The read step comes in three forms"); the printed
# form keeps its head banner for the hosts and fallbacks that still use it.
shim_max_lines: 120
# The plugin-host baseline: the lowest build this skill was RUN on with all of
# its host mechanisms working. A record, not a gate: nothing reads it to decide
# anything, and the reviewer adapter's floor in utils/providers/claude_code.py
# protects a different thing (the CLI's JSON result envelope).
#
# Measured 2026-11-07 on a throwaway HOME, `claude -p --permission-mode default
# --plugin-dir <frozen tree>`, `openai/gpt-6-luna` through OpenRouter. A build
# passes if the skill loads, `Agent` creates a subagent and `SendMessage` resumes
# it. 2.1.319, 2.1.300, 2.1.280, 2.1.250 and 2.1.200 pass; 2.1.319 defers
# `SendMessage` behind `ToolSearch`. 2.1.150 and 2.1.100 load the skill and create
# a subagent but expose no `SendMessage`, so a batch cannot be resumed there.
# Versions between 2.1.150 and 2.1.200 were not tried. Every run was headless, so
# no interactive prompt was observed, and none reported a permission denial on
# `Agent` or `SendMessage`.
min_host_version: "2.1.200"
installable: false
install_paths:
project: [".claude/skills/multi-llm"]
global: ["$HOME/.claude/skills/multi-llm"]
legacy: []
reload: none
arguments: placeholder # $ARGUMENTS / ${CLAUDE_SKILL_DIR}
# The plugin's own invocation, namespace included. `documented`, not
# `verified`: no row has ever typed it into a Claude Code session — the
# container lane drives this host headless with a prompt, exactly as the two
# installable hosts' `documented` cells were before 2026-08-24.
#
# The trailing clause names the second install route: a copy under one of
# `install_paths` is a plain skill, invoked without the plugin namespace. It
# is appended, so the leading ``as `/`` `hosts/render_capabilities.py` keys
# on is unchanged, and it is the prose `next_command.skill` is held to.
invocation: "as `/multi-llm:multi-llm <mode-flag> <plan-path> [options]`, or as `/multi-llm <mode-flag> <plan-path> [options]` when the skill is installed as a plain skill under one of `install_paths`."
# The next-phase command a finished phase ends its report with, per install
# route (plans/next-phase-command.md §2.1). Read by `utils/next_phase.py` at
# run time, so it is required. `default` is the plugin (the route
# `invocation` documents first); `skill` is a copy under `install_paths`,
# which `resolve_route` recognises from the dispatcher's own skill root.
# `skill` is `verified` (2026-10-01, T10 of plans/next-phase-command.md,
# claude-code 2.1.285): the container lane stages this host as a plain skill
# under `$HOME/.claude/skills/multi-llm`, and `claude-code-next-command` ran
# twice on that route: runs/20261001T114907Z on `openai/gpt-6-luna` through
# OpenRouter, and runs/20261001T121548Z on `sonnet` through the Claude
# subscription (`--creds fallback`). Each printed `/multi-llm --implement
# "/work/sample-project/plans/next phase.md"`, and that command typed as the
# whole prompt of a fresh `claude -p` session dispatched `implement` with the
# fixture unquoted. NOT covered: `default` (the plugin route) has never been
# typed by a recorded run (the lane cannot stage it) and stays `documented`;
# the plugin's Stop hook (plain-skill staging loads no plugin hooks); and
# typing into the TUI (both runs were headless `claude -p`).
next_command:
default: "/multi-llm:multi-llm {args}"
skill: "/multi-llm {args}"
multi_turn: true
subagents: native
ask: structured
detach: native
native_tool:
# `Agent` is the tool's current name, and `Task` is its compatibility alias in
# settings and agent definitions, so SKILL.md's `allowed-tools` keeps `Task`.
# The spelling is not a permission fix: subagent launches need no permission.
subagents: Agent
ask: AskUserQuestion
detach: run_in_background
# No session rung: this host dispatches subagents natively through `Agent`
# and picks their model from `subagent_tiers`, so there is nothing for a
# session model to be substituted into — and nothing on the skill's side of
# the boundary reads the session's selected model in the first place.
session_runner_provider: null
default_subagent_runner: "claude-code:sonnet"
# subagents: native, so the tier -> model map lives here rather than in
# providers.yaml driver.subagent_runner{,_cheap} (D2e).
subagent_tiers: {full: sonnet, cheap: haiku}
runners:
claude-code: {bundled: true, needs_auth: false, login_cmd: "",
profiles: [generate, agent-readonly, agent-write], usage: full}
cursor-agent: {bundled: false, needs_auth: true,
login_cmd: "cursor-agent login",
profiles: [generate], usage: partial}
# In-session native dispatch: no process of ours sits between parent and child,
# so there is nothing to strip and nothing to neutralize.
nested_exec: {spawn: allowed, network: allowed, credentials: allowed,
notes: "Native subagents run in-session; the executor is not on
the default path for this host."}
# Audited and non-empty: these are the variables Claude Code injects, which a
# dispatched third-party runner must not inherit. Literal names only — a
# wildcard would strip the runner's own credentials along with the guard.
#
# `passthrough` is EMPTY, and the entry that used to sit here is the reason
# this comment is worth reading. It carried `ANTHROPIC_API_KEY`, sanctioned
# as "an API key the host owns". `subagent_executor` re-admits every
# passthrough name last, so that one cell silently cancelled M3/§2.1's
# credential strip on the agent-profile lane while leaving it in force on
# the fan-out lane: the same claude-code slot stripped the key on Lane A and
# inherited it on Lane B, and the money bug the strip exists to close stayed
# open for every subagent dispatch. Re-adding it would restore exactly that.
#
# The strip is not unconditional and does not need a passthrough as its
# escape hatch: it is gated on the sanitized `claude auth status` preflight
# and fails OPEN — an API-key-only user's key is preserved byte-for-byte —
# and a user who wants it back unconditionally sets
# `providers.claude-code.strip_vendor_keys: never` in providers.yaml, which
# is a per-provider decision rather than a host-wide one.
#
# NOTE: this is a HOST-side allowlist, keyed by the host multi-llm is
# *running inside*, not by the runner it dispatches. `codex` and `opencode`
# both already carry `passthrough: []`, so emptying this one changes exactly
# one situation: claude-code dispatched from inside claude-code.
env_guards:
guards: ["CLAUDECODE", "CLAUDE_CODE_ENTRYPOINT"]
session_scoped: ["CLAUDE_SESSION_ID", "CLAUDE_SKILL_DIR"]
passthrough: []
max_concurrent_subagents: 10
timeout_ceiling_s: 600 # Bash tool hard cap, silently clamped above
timeout_override: {frontmatter: none, env: none, config: none}
provenance:
shim: documented
shim_max_lines: verified # counted on the file in the repo
min_host_version: verified # the runs in the comment beside the key
installable: verified
install_paths: documented
reload: documented
arguments: verified
invocation: documented
next_command: {default: documented, skill: verified}
multi_turn: verified
subagents: verified
ask: verified
detach: verified
native_tool: verified
session_runner_provider: documented
default_subagent_runner: verified
subagent_tiers: verified
runners: documented
nested_exec: documented
env_guards: documented
max_concurrent_subagents: documented
timeout_ceiling_s: documented
timeout_override: documented
codex:
# Same skills format, and no longer on paper. The 2026-08-24 container
# documented-invocation row installed this shim into a fresh `$HOME` and codex
# enumerated it, advertising `multi-llm` to the model inside
# `<skills_instructions>` with its `SKILL.md` locator — so the format claim
# itself is measured. What the same row measured is the half that claim never
# covered, and it is a negative: see `invocation` below. A codex skill is
# reached by the MODEL choosing to read it; the composer rejects `/multi-llm`
# outright. `invocation` is therefore the one cell promoted to `verified` here.
# shim/arguments/install_paths stay `documented`: that row graded the host's
# command surface in one scope on one platform, not placeholder substitution
# and not the rest of the path matrix, and a promotion may not claim more than
# the run behind it.
shim: {template: hosts/codex/shim.md.tmpl, installed: SKILL.md,
format: markdown+frontmatter}
installable: true
install_paths:
# A preferred entry names the directory to write into (the file name comes
# from `shim.installed`); a legacy entry names the FILE, because a
# deprecated layout's file name is the host's own old convention and not
# ours. Deprecated Codex prompts were `~/.codex/prompts/<command>.md`, so
# the copy to detect is `multi-llm.md` — probing `prompts/SKILL.md` would
# be a path no install of any vintage ever created.
project: [".codex/skills/multi-llm"]
global: ["$HOME/.codex/skills/multi-llm"]
legacy: ["$HOME/.codex/prompts/multi-llm.md"]
reload: new-session
arguments: placeholder
# MEASURED 2026-08-24, codex-cli 0.146.0, container documented-invocation row,
# and it is a NEGATIVE result: typing `/multi-llm --review-plan <plan>` into a
# fresh composer prints `Unrecognized command '/multi-llm'` and sends nothing
# at all — no request reached the stub. The `/` popup lists no skill, `/multi`
# offers no completion, and a file at the deprecated `install_paths.legacy`
# prompts location is rejected the same way. Headless, `codex exec "/multi-llm
# …"` passes the line through verbatim. The shim IS discovered — codex
# advertises it to the model inside `<skills_instructions>` with its `SKILL.md`
# locator — so it is model-invoked rather than typed, which is what this
# sentence has to say. Read into the generated pages after "Invoke it ".
invocation: >-
by name rather than as a slash command — `Use the multi-llm skill with
exactly these arguments: <the same arguments the Claude Code skill takes>`.
Codex installs this shim as a **skill**, which is model-invoked rather than
typed: its composer answers `/multi-llm` with `Unrecognized command`.
# Per route (plans/next-phase-command.md §2.1). `default` is the file shim
# `invocation` documents; `package` is the Codex plugin, the spelling
# `package.invocation` documents. Both `verified` since T10 (2026-10-01) —
# by a recorded run that typed the PRINTED command, never by derivation
# from `invocation` (see the provenance row below for the runs).
next_command:
default: "Use the multi-llm skill with exactly these arguments: {args}"
package: "$multi-llm:multi-llm {args}"
multi_turn: true # documented, not observed: `codex exec`
# runs a tool loop and `codex exec resume`
# continues the session — either would give
# the agent the follow-up turn.
subagents: executor # agents are GA in the TUI; spawn-from-exec is
# unconfirmed, and the failure is
# silent-then-fatal, so the exec profile is
# `executor` until a recorded run says
# otherwise. native_tool names the tool for
# documentation, not for dispatch.
#
# EVIDENCE, NOT A PROMOTION (2026-08-25,
# codex-cli 0.149.1, read out of the captured
# tools[] array of a real headless request):
# `multi_agent_v1` IS offered in `codex exec`,
# carrying spawn_agent, wait_agent, send_input,
# resume_agent and close_agent with a
# documented model-override list. That closes
# "is it offered" and leaves "does it work"
# open, which is the half this cell grades — a
# tool in the array is not a recorded run, and
# the failure mode here is silent-then-fatal.
# One detail a `native` promotion would have to
# clear: spawn_agent's own description reads
# "Do not spawn sub-agents unless the user or
# applicable AGENTS.md/skill instructions
# explicitly ask" — a skill CAN grant that
# permission, so the lever exists.
ask: free-text # request_user_input is Plan-Mode-only
#
# EVIDENCE, NOT A PROMOTION (same capture):
# request_user_input is in the tool list and
# its own description says "This tool is only
# available in Plan mode." — which CONFIRMS the
# line above rather than contradicting it, so
# the cell does not move. What is new is a flag
# to watch: `codex features list` reports
# `default_mode_request_user_input under
# development false`. That is the switch that
# would make the tool generally available, so
# re-probe it on each codex bump.
detach: python
native_tool:
subagents: agent
# The plugin's server exports the turn metadata's `model` (codex 0.156.1 sent `--model
# stub-model` as `"model": "stub-model"`, 2026-09-25 package rows); no effort key rides along.
session_runner_provider: codex
# UNVALIDATED MODEL ID — THE ROSTER OWNS THE ID; THIS PIN FOLLOWS IT. This is
# a credential-free floor (the last rung of `resolve_subagent_runner`'s
# ladder, chosen so a zero-config install dispatches at all), not an
# independent opinion about which model is current. `providers.yaml` is where
# the codex model is actually chosen and re-measured (3f95a6e moved the whole
# fan-out to gpt-6-astra against a live catalog), so the edit only ever runs
# one way: the roster moves, then this pin. The fan-out's `:xhigh` effort
# suffix does NOT come across — the floor takes the model's default effort
# until somebody measures that it is not enough for a subagent dispatch.
# No test resolves this id against a catalog (only kilocode has a captured
# one), so this rule buys consistency, not correctness: after it, the pin is
# wrong only when the roster is. Whoever finds a dead id edits
# `providers.yaml` and lets this line follow.
default_subagent_runner: "codex:gpt-6-astra"
subagent_tiers: null # forbidden to carry a map: subagents != native
runners:
codex: {bundled: true, needs_auth: false, login_cmd: "",
profiles: [generate, agent-readonly, agent-write], usage: full}
claude-code: {bundled: false, needs_auth: true, login_cmd: "claude login",
profiles: [generate, agent-readonly, agent-write], usage: full}
nested_exec: {spawn: unknown, network: unknown, credentials: unknown,
notes: "Filled by the Phase 2 nested-exec smoke: --exec-subagents
launched from inside `codex exec`'s own sandboxed shell, for
both the same-host runner and a third-party one."}
# Empty because the env diff has not run yet, not because the host is clean —
# which is what provenance: unknown records.
env_guards: {guards: [], session_scoped: [], passthrough: []}
max_concurrent_subagents: 4
timeout_ceiling_s: 0 # 0 = no documented cap on `codex exec`
timeout_override: {frontmatter: none, env: unknown, config: unknown}
# Shell output budget, read off codex-cli 0.167.0 (2026-11-08); a comment,
# not a field, because no dispatch decision reads it. Every model in its
# catalog carries `truncation_policy: {mode: tokens, limit: 10000}`, and both
# shell tools (`exec_command`, code mode's `exec`) document `max_output_tokens`
# as "Defaults to 10000 tokens", a budget the model may lower per call (the
# `exec` tool's own example passes 1000). Over budget, codex keeps the head
# and the tail, drops the middle (`…N tokens truncated…`) and counts about 4
# bytes a token, so the file shim reads the core with `util instructions
# --page` (multi_llm.CORE_PAGE_SIZE_DEFAULT; docs/development/hosts.md,
# "Paged delivery").
# A second install route — the Codex plugin marketplace — NOT a second host,
# for the same reason as opencode's block below: every harness property above
# is the same however the skill got installed. What the plugin adds is a
# model-callable `multi_llm` tool (an MCP server bundled with the plugin) and
# a thin skill advertised as `multi-llm:multi-llm` that points at it. A
# plugin exposes no slash command (R21), so `provides` has no `command`.
#
# `update: reinstall` is two different commands from the install, not the
# install re-run: `marketplace add` on a marketplace already added is a
# no-op ("Marketplace … is already added", R37), so the upgrade is
# `codex plugin marketplace upgrade` then `codex plugin add`, then a new
# session (R30). The renderer's registry row owns that sentence.
#
# No `version`: the number is owned by `skills/multi-llm/pyproject.toml`,
# and the codex end of that chain is the GENERATED
# `.codex-plugin/plugin.json`, never a hand-typed copy here.
package:
registry: codex-marketplace
name: multi-llm
marketplace: beastlabai
repo: beastlabai/multi-llm-plugin
ref: release
sparse: [.agents, packaging/codex]
# The slug, the ref and the two sparse paths the first
# install command needs. Cells rather than literals in
# the shared form string, which every later
# codex-marketplace package would inherit.
config_key: plugins # [plugins."multi-llm@beastlabai"] in $CODEX_HOME/config.toml (R15)
provides: [skill, tool] # NOT command (R21)
update: reinstall # `marketplace upgrade` then `plugin add` (R30/R37)
write_posture: refuses-under-readonly
# The server runs outside the session's exec sandbox
# (R46), so an install hands a `--sandbox read-only`
# session a write path it had refused (R45).
# The server reads the session's posture per call
# (`sandbox_mode` in codex's turn metadata, 0.156.1):
# under `read-only`, or a posture it cannot read,
# write-capable modes are refused before the spawn
# unless MULTI_LLM_CODEX_UNSANDBOXED_WRITES=1 is in the
# environment (D3b); `workspace-write` and
# `danger-full-access` allow them. The renderer reads this cell into
# the generated install paragraph, so a user meets it
# before installing, not after.
min_host_version: "0.159.2" # the lowest codex the plugin route was RUN on (T10's container
# rows, 2026-10-01, below), the schema's own wording — never the
# container pin or packaging/codex/ingestion_rules.py's stamp, which
# follow each pin bump; it moves on a measured break, and test_codex_route
# holds it at or below both. Not bisected (Open question 5)
invocation: >-
with a `$` mention of the namespaced skill — `$multi-llm:multi-llm <the same
arguments the Claude Code skill takes>`; plugins add no slash command, so
`/multi-llm` stays unrecognized. If the model ever reworks the arguments,
spell it out — `Use the multi-llm:multi-llm skill with exactly these
arguments: <args>` — or let the model call the `multi_llm` tool directly.
provenance:
shim: documented
installable: documented
install_paths: documented
reload: documented
arguments: documented
invocation: verified
# PROMOTED BY T10 (plans/next-phase-command.md), codex-cli 0.159.2,
# container rows `codex-next-command` / `codex-package-next-command`. Each
# route ran `generate-tasks` on a fixture plan whose path holds a space
# (`plans/next phase.md`), then handed the command its dispatcher printed
# (`user_next_step.command`), verbatim and as the whole prompt, to a FRESH
# headless `codex exec` session on the same route
# (`entrypoint.run_typed_followup`). In both sessions on both routes the
# dispatcher's own receipt then recorded `implement` with the plan
# argument naming the fixture and holding no quote character:
# default: `Use the multi-llm skill with exactly these arguments:
# --implement "plans/next phase.md"` — runs/20261001T113914Z
# (~deepseek/deepseek-v4-flash-latest) and runs/20261001T115800Z
# (deepseek/deepseek-v4.1-flash), argv `plans/next phase.md`;
# package: `$multi-llm:multi-llm --implement "/work/sample-project/plans/
# next phase.md"` — the `$` mention, typed into `codex exec` — runs/
# 20261001T113244Z (~deepseek/deepseek-v4-flash-latest) and runs/
# 20261001T115346Z (deepseek/deepseek-v4.1-flash), a `multi_llm` tool
# call whose argv held the absolute fixture path unquoted. The plugin
# was installed from the image's local marketplace mirror, which is
# why `package` (the published Git route) below stays `documented`.
# What it does NOT cover, and the cell claims no more: the interactive
# composer (no row has typed either spelling into the TUI against a real
# model — the `$` popup's own resolution is unmeasured), other platforms,
# and the four other workflow modes' commands (one template, one row).
next_command: {default: verified, package: verified}
multi_turn: documented
subagents: documented
ask: documented
detach: documented
native_tool: documented
session_runner_provider: verified # the model key in codex-package-sandbox-posture's mcp-tap
default_subagent_runner: documented
subagent_tiers: documented
runners: documented
nested_exec: unknown
env_guards: unknown
max_concurrent_subagents: documented
timeout_ceiling_s: unknown
timeout_override: unknown
# PROMOTED 2026-10-23 with the first release, codex-v0.1.0 (ae089856),
# on the two Git-route records rule (j) asks for, never the container's
# local-path installs: the offline Git-route row `codex-package-git-route`
# (an ssh://localhost bare repository with `release` one commit behind,
# `--ref release` and both `--sparse` paths; green on codex 0.159.2,
# 0.163.0 and 0.164.0 in the stamped gate session 2bc2b6ca), and the
# manual network check from GitHub itself,
# tests/container/records/20261023T132656Z-codex-release-network-check.json
# (a fresh stock-image container, codex 0.164.0, all six criteria pass
# for the tag). After promotion the published `--ref release` line was
# run the same way in a second fresh container: the snapshot resolved to
# the tagged commit, all six criteria passed, and `marketplace upgrade` +
# `plugin add` left the install byte-identical. Linux only.
package: verified
opencode:
# commands/ is PLURAL — the singular command/ path is simply not scanned.
shim: {template: hosts/opencode/shim.md.tmpl,
installed: commands/multi-llm.md, format: markdown+frontmatter}
installable: true
install_paths:
project: [".opencode/commands"]
global: ["$XDG_CONFIG_HOME/opencode/commands", "$HOME/.config/opencode/commands"]
# Empty on purpose, and `.opencode/command` is the entry that is NOT here:
# OpenCode never scans the singular directory, so a file sitting in it is
# not a copy that loads. Listing it would make `--check` report — and, as
# the highest-precedence project location, *rank first* — a file the host
# will never discover, marking the real install shadowed. A compatibility
# location is a place an install of an older vintage really works, never
# a place a file might merely exist.
legacy: []
reload: new-session
arguments: placeholder # $ARGUMENTS / $1 — verified: the message
# text reaches the command body.
# Measured, and a trap worth shipping: `opencode run` parses the message
# with yargs, so an argument list that STARTS with a flag (`--ask …`,
# `--review-plan …` — i.e. every invocation this skill takes) is swallowed
# as an option and the run silently prints yargs' own help instead. The
# `--` separator is mandatory headless. Harmless in the TUI, where
# `/multi-llm --ask …` needs no separator.
headless_invocation: 'opencode run --command multi-llm -- "<args>"'
# MEASURED 2026-08-24, opencode 1.18.18, container documented-invocation row:
# the published spelling WORKS in a fresh TUI session. Typing
# `/multi-llm --review-plan <plan>` expanded the installed command into the
# turn the host sent — the shim's own text, its dispatcher block with
# `{SKILL_DIR}` already substituted, the baked `--host opencode --yes
# --exec-subagents`, and the mode flag still adjacent to its argument. Read
# into the generated pages after "Invoke it "; the headless correction above
# is rendered separately, because `--command` is what resolves the shim
# there, not this spelling.
invocation: "as `/multi-llm <the same arguments the Claude Code skill takes>`."
# Per route (plans/next-phase-command.md §2.1): the file shim and the npm
# package both answer to `/multi-llm` (the package's `config` hook injects
# the same command). Both `verified` since T10 (2026-10-01); the runs are
# named on the provenance row below.
next_command:
default: "/multi-llm {args}"
package: "/multi-llm {args}"
# A second, recommended install route — NOT a second host. `reload`,
# `subagents`, `ask`, `detach`, `max_concurrent_subagents` and
# `timeout_ceiling_s` are properties of this harness, not of how the skill
# got installed, so an `opencode-plugin` host id would duplicate every one
# of them and split the matrix.
#
# What it buys over the file shim: the plugin computes the skill dir from
# its own on-disk location at every startup, so there is no baked absolute
# prefix to go stale and no `install-host --all` owed after an update. The
# baked flags stop being prose a model may improvise past and become code
# the plugin appends.
#
# `update: pinned-spec` is not a preference. The npm cache key is the
# literal spec string and the installer early-returns whenever the package
# directory already exists, so an unpinned entry resolves ONCE and never
# again — bumping the pin is the only clean upgrade, and there is no
# `opencode plugin update` command.
#
# `version` is a COPY. The number is owned by
# `skills/multi-llm/pyproject.toml` — which ships *inside* the tarball, so a
# package declaring a different one would ship a payload that disagrees with
# it — and `test_version_source_is_singular` asserts the chain.
#
# `public_version` is the last version whose registry install check passed,
# the one the generated paragraph and the hand-written README and
# docs/hosts.md pins tell a user to pin. `version` is the candidate and moves
# at every bump, so a recommendation built from it would name a version npm
# does not serve yet. The schema requires `public_version` while
# `provenance.package` below is `verified`, never newer than `version`. It
# was set in the edit that promoted that row (0.1.1, backed by
# records/20261026T175524Z-opencode-release-network-check.json), and it
# moves only once the next version's registry record validates.
package:
registry: npm
name: opencode-multi-llm
version: "0.1.9"
public_version: "0.1.8"
config_key: plugin
provides: [command, tool]
min_host_version: "1.19.5" # the one supported opencode release (the
# newest; `make opencode-bump` moves it with
# every other pin), never a lower one that
# might work: `engines` is a hard gate, and
# a floor asserted from an assumption either
# admits a version that then fails somewhere
# less legible or gets defended by a claim no
# evidence backs.
update: pinned-spec
invocation: "as `/multi-llm <the same arguments the Claude Code skill takes>`, or by calling the `multi_llm` tool directly."
multi_turn: true # measured: its agent takes further tool
# turns inside one `opencode run` and acts
# on what they printed — see the promotion
# note below.
subagents: executor # `mode: subagent` dispatch is documented only;
# dispatch is instructed in the command body,
# never via frontmatter `subtask:` (which the
# parser accepts and then ignores).
ask: free-text # the `question` tool is unreliable headless
detach: required # the ~120 s bash cap is documented, not
# measured, and every fan-out mode runs far
# longer than it, as does every apply mode
# and `implement`, whose write batches the
# executor runs inside the command — so no
# foreground path is offered for either.
native_tool:
subagents: task
# A CREDENTIAL-FREE FLOOR WITH A KNOWN EXPIRY, NOT A RECOMMENDATION (D3).
# This is the last rung of `resolve_subagent_runner`'s ladder — what a user
# who configured nothing falls to — so its one requirement is that it
# dispatches on a fresh install with nothing logged in. It is NOT the
# recommended validation runner: validation-grade work configures
# `driver.subagent_runner` with a full-capability model (`providers.yaml`
# ships `claude-code:sonnet` in that shape). `opencode/big-pickle` is a Zen
# stealth id served from the same server-side catalog that retired
# `opencode/sonnet`, so it is measured, dated, and re-probed on each
# opencode bump — that re-probe is the gate on any commit that moves an
# opencode version (container pin, `engines` floor, or a fixture
# recapture), and `tests/test_container_host_lane.py`'s catalog diff is
# what makes a skipped ritual visible.
#
# RE-PROBED 2026-11-06 on opencode 1.19.5 (the pinned build), zero
# credentials: HOME and all four XDG dirs (XDG_CACHE_HOME too) at empty
# scratch dirs (`opencode auth list` → `└ 0 credentials`), every
# `*_API_KEY` unset, never wrapped in `timeout` (which wedges `opencode
# run`). The listing was taken twice with the same answer: from the host's
# installed 1.19.5 and inside mllm-e2e:local (its real binary, not the
# rig's PATH wrapper).
# REFRESH FIRST. With an empty cache, and always under the image's
# OPENCODE_DISABLE_MODELS_FETCH=1, `opencode models` answers from the
# catalog built into the binary and writes no models.json. That copy still
# listed `opencode/nemotron-3-ultra-free` on this day, although models.dev
# already marked it `status: deprecated` and a fetched listing hides it.
# $ opencode models --refresh && opencode models | grep '^opencode/'
# opencode/beam-free opencode/glm-5.3-flash-free
# opencode/big-pickle opencode/mimo-v2.6-flash-free
# opencode/crouton-alpha-free opencode/nemotron-3.5-lightning-free
# opencode/deepseek-v4.1-flash-free opencode/nemotron-3.5-super-vl-free
# opencode/flint-40b-free opencode/pulse-alpha-free
# The user's own HOME (its models.json fetched 2026-11-05) listed the same
# 10 ids: the listing is a catalog, not an entitlement check. Gone since
# 2026-09-11: `opencode/nemotron-3-ultra-free`, the one shipped id, which
# providers.yaml replaced with `opencode/nemotron-3.5-super-vl-free`, the
# listed Nemotron that is not the lightning id measured hanging below. Also
# gone: ling-3.0-flash-fin-free, mimo-v2.5-free and both
# muse-spark-*-contributor-free ids, none of them shipped.
# `opencode run --model <id> --print-logs --log-level ERROR
# "Reply with the single word OK."` (host build, same scratch pair)
# opencode/big-pickle exit 0, stdout `OK`, 2 s
# opencode/nemotron-3.5-super-vl-free exit 0, stdout `OK`, 2 s
# The floor stands. The validation-batch gate below was not re-run because
# the floor did not move.
# FIRST PROBED 2026-09-11 on opencode 1.18.30, the same recipe without the
# refresh. It listed 7 `opencode/` ids: big-pickle, ling-3.0-flash-fin-free,
# mimo-v2.5-free, muse-spark-1.2-contributor-free,
# muse-spark-1.3-contributor-free, nemotron-3-ultra-free and
# nemotron-3.5-lightning-free. It listed no `opencode/sonnet`,
# `opencode/deepseek-v4-flash-free` or `opencode/hy3-free`, and all three
# were dropped from providers.yaml the same day. The same `OK` prompt gave:
# opencode/big-pickle exit 0, stdout `OK`, a few seconds
# opencode/nemotron-3.5-lightning-free zero bytes on either stream >7 min,
# killed (exit 143)
# opencode/sonnet exit 1, `ProviderModelNotFoundError:
# Model not found: opencode/sonnet`
# The `*-free` siblings are NOT automatically the safer alternative. The
# lightning id reproduced the zero-byte hang, which costs `default_timeout`
# per task. One answering `OK` (super-vl, 2026-11-06) has passed only that
# check, not the validation-batch gate below.
# VALIDATION-BATCH GATE (existence is not fitness), MEASURED 2026-09-11
# from the same zero-credential pair, opencode 1.18.30: a scratch git repo
# whose `.multi-llm/providers.yaml` set `driver.subagent_runner:
# opencode:opencode/big-pickle` (and `quick_models` to the same id, so the
# whole run was credential-free), `multi_llm.py review-plan plans/tiny.md
# --host opencode --yes --quick --exec-subagents` over the 285-byte
# tests/fixtures/sample_plan.md: 12 suggestions → 10 groups → 7
# `agent-readonly` validation batches; `subagent_batch_report.json`
# `counts: {ok: 7}`, 0 unparseable_result, 0 schema_invalid, 7 attempts,
# 34,679 tokens at `cost_usd: 0.0`, batches 4.5–29 s each; `validation.json`
# carried 10 per-group verdicts each with its own reason and confidence
# (0.85–0.98). Exit 0, 77 s wall clock.
# The preflight itself, same pair, completes rather than exiting 3:
# `auth: not-required`, `model_catalog: ok`, `nested_exec: allowed`, while
# `opencode:openrouter/deepseek/deepseek-v4-flash-0731` on the same CLI in
# the same process is refused `not_authenticated`.
# If this id vanishes: the replacement is whichever `opencode/*` id answers
# `OK` from a zero-credential HOME AND clears the validation-batch gate; if
# none does, opencode has no credential-free default — drop the entry from
# `OpenCodeProvider.credential_free_models` in the same commit and let the
# out-of-the-box run refuse `not_authenticated` with its login hint.
# THE FLOOR IS THE NEWEST RELEASE (policy since 2026-10-02): opencode is
# supported on its newest release only, so `min_host_version` above,
# `engines.opencode`, the adapter's `min_version`/`verified_version` and the
# container's pin are one version, the one the parser fixtures were
# captured from. `make opencode-bump VERSION=<x.y.z>` moves them together
# (never below the current pin without ALLOW_DOWNGRADE=1), and
# `tests/test_opencode_version_pins.py` fails the build when one did not.
# THE SESSION RUNG (`resolve_subagent_runner`, layer `session`). Names the
# CLI of this host's own that can be pointed at whatever model the *session*
# is currently running on, which the plugin exports as
# `MULTI_LLM_HOST_SESSION_MODEL` (owned by `utils/provider_registry.py`; the
# plugin reads the name out of the generated cli-surface.json and never
# spells it in JS). The rung sits below every explicit `driver.*` layer and
# above `default_subagent_runner`.
#
# WHY. A `--review-plan --quick` run launched from the TUI with
# "Nemotron 3 Ultra Free · OpenCode Zen" selected validated all sixteen of
# its batches on `opencode/big-pickle`: no layer above the floor had an
# opinion, and the floor is a credential-free *default*, not a choice anyone
# made. The session's own model is the nearest thing to an operator opinion
# that exists when no config file does.
#
# MEASURED 2026-09-11, opencode 1.18.30 (the installed build), never wrapped
# in `timeout`:
# $ opencode run --model opencode/nemotron-3-ultra-free \
# "Reply with the single word OK."
# OK exit 0, 62 s
# (That id has since been deprecated and is unlisted on 2026-11-06; see the
# re-probe above. The claim here is about the mechanism, not that id.)
# So `opencode run --model <the session's own id>` really serves a TUI model
# id, which is the whole claim this field makes. `opencode/big-pickle` stays
# the credential-free floor below it, and the fallback is what keeps the
# rung safe: a session model that fails preflight for any reason (a private
# id, a logged-out provider, an unresolvable slug) drops to the next rung
# with one printed line rather than refusing the batch.
session_runner_provider: opencode
default_subagent_runner: "opencode:opencode/big-pickle"
subagent_tiers: null
runners:
opencode: {bundled: true, needs_auth: false, login_cmd: "",
profiles: [generate, agent-readonly, agent-write], usage: full}
claude-code: {bundled: false, needs_auth: true, login_cmd: "claude login",
profiles: [generate, agent-readonly, agent-write], usage: full}
nested_exec: {spawn: unknown, network: unknown, credentials: unknown,
notes: "Filled by the Phase 3 nested-exec smoke from inside
`opencode run`."}
env_guards: {guards: [], session_scoped: [], passthrough: []}
max_concurrent_subagents: 4
timeout_ceiling_s: 120 # documented default bash timeout
# No `timeout` field exists in the V2 command contract, so the ceiling moves
# only through the environment/config surface — and the installer only ever
# writes through a non-`none` channel.
timeout_override: {frontmatter: none, env: unknown, config: unknown}
# Promoted from a recorded smoke on opencode 1.18.12 (2026-08-08): the
# installed command was discovered by name in a new session, its uv
# preflight ran ahead of the first `uv run`, the baked prefix resolved, and
# a `--detach`ed `ask` fan-out was polled through `status --wait` to a
# terminal `done` and its answers file read back. That same run settles
# `multi_turn`, which had shipped `false` on nothing: its agent took four
# tool turns inside a single `opencode run` — `command -v uv`, then
# `util instructions status`, then `status`, then a read of the answers —
# and followed `MULTI_LLM_RESULT.next` on its own initiative, which is
# precisely the follow-up turn the field claims. Everything that run did
# not exercise stays where it was — in particular `subagents` (no subagent
# was dispatched) and `ask` (nothing was asked), which is what keeps the
# conservative resolution in force.
provenance:
shim: verified
installable: verified
install_paths: verified
reload: verified
arguments: verified
headless_invocation: verified
# PROMOTED BY THE FIRST RELEASE'S REGISTRY CHECK (opencode-v0.1.1,
# f4f024c7, 2026-10-26). The criterion was always a *registry* install,
# never a local build, which is why the container lane's `route: package`
# rows (a LOCAL DIRECTORY at `/opt/mllm/opencode-plugin`, since
# 2026-09-19) never promoted it. `make opencode-registry-check
# ENVIRONMENT=container` installed `opencode-multi-llm@0.1.1` from
# registry.npmjs.org into a stock node:22-bookworm-slim container with
# opencode 1.19.4 and empty caches, and passed all eight first-release
# criteria: the spec resolves with the gated integrity, the installed
# tree equals the candidate rebuilt from the tag, `/multi-llm` resolves
# the installed skill, the `multi_llm` tool runs `init` from that payload,
# first use installs the runtime dependencies, a warm `status` runs with
# the network cut, and a restart on the same pin serves the same payload
# (records/20261026T175524Z-opencode-release-network-check.json, validated
# by `registry_record.py validate --profile release`). What it does NOT
# cover: macOS and Windows (Linux only, as packaging/opencode/README.md
# says), or an upgrade from an earlier npm version (none existed at
# that first release). The later-release proof below covers upgrades.
# RECHECKED 2026-10-30 for opencode-v0.1.3 on OpenCode 1.19.5.
# A fresh npm install passed all nine later-release criteria, including
# offline use, same-pin restart, and an upgrade from 0.1.1 executing
# the new payload (records/20261030T214755Z-opencode-release-network-check.json).
# RECHECKED 2026-11-01 for opencode-v0.1.4 on OpenCode 1.19.5.
# All nine registry criteria passed, including offline operation and
# a 0.1.3 -> 0.1.4 upgrade executing the newly installed payload
# (records/20261101T061412Z-opencode-release-network-check.json).
# RECHECKED 2026-11-12 for opencode-v0.1.8 on OpenCode 1.19.5.
# All nine registry criteria passed, including offline operation and
# a 0.1.4 -> 0.1.8 upgrade executing the newly installed payload
# (records/20261112T160945Z-opencode-release-network-check.json).
package: verified
invocation: verified
# PROMOTED BY T10 (plans/next-phase-command.md), opencode 1.18.33,
# container rows `opencode-next-command` / `opencode-package-next-command`.
# Each route ran `generate-tasks` on `plans/next phase.md`, then handed the
# PRINTED command to a FRESH `opencode run` on the same route through the
# host's verified headless spelling, `--command multi-llm -- <args>` — the
# command's own `/multi-llm ` prefix is what `--command multi-llm` names,
# and the argument string is the printed one, byte for byte (`opencode run
# -- "/multi-llm …"` never expands, measured on the documented-invocation
# row, which separately verified the TUI's `/multi-llm` resolution). The
# dispatcher's own receipt then recorded `implement` with the fixture as
# its plan argument and no quote character in it:
# default: `--implement "plans/next phase.md"` — runs/20261001T113344Z
# (~deepseek/deepseek-v4-flash-latest) and runs/20261001T115548Z
# (deepseek/deepseek-v4.1-flash), a shell dispatch;
# package: `--implement "/work/sample-project/plans/next phase.md"` —
# runs/20261001T112645Z (~deepseek/deepseek-v4-flash-latest) and
# runs/20261001T115138Z (deepseek/deepseek-v4.1-flash), a `multi_llm`
# tool call. The plugin was the image's local-directory install, which
# is why `package` (a registry install) above stays `documented`.
# Not covered: a real model typing into the TUI, other platforms, and the
# other modes' commands. (Both default-route runs' FINAL MESSAGES left the
# optional review-tasks command unfenced — criterion 4, a separate
# question this cell does not grade.)
next_command: {default: verified, package: verified}
multi_turn: verified
subagents: documented
ask: documented
detach: verified
native_tool: documented
session_runner_provider: documented
default_subagent_runner: documented
subagent_tiers: documented
runners: documented
nested_exec: unknown
env_guards: unknown
max_concurrent_subagents: documented
timeout_ceiling_s: documented
timeout_override: unknown
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 |