Run any Skill in Manus with one click

model-runner

Stars3

Forks1

UpdatedJune 22, 2026 at 18:58

Run, assess, and switch the local vLLM model via the `lobes` CLI (lobes-cli). Use when changing the served model, starting/stopping the server, load-testing a candidate (e.g. an NVFP4 checkpoint), or producing the numbers for a per-model doc under docs/.

Installation

Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.

Run Skill in Manus

Source

agentculture

agentculture/lobes-cli

View GitHub Repository View Creator Repositories

Download

Run Skill in Manus

Related occupationsSOC

Based on SOC occupation classification

Network and Computer Systems AdministratorsComputer and Mathematical Occupations·SOC 15-1244

File Explorer

2 files

SKILL.md

readonly

name	model-runner
description	Run, assess, and switch the local vLLM model via the `lobes` CLI (lobes-cli). Use when changing the served model, starting/stopping the server, load-testing a candidate (e.g. an NVFP4 checkpoint), or producing the numbers for a per-model doc under docs/.

model-runner

The local vLLM model is served by this repo's deployment (docker-compose.yml + .env, scaffolded into ~/.lobes). The canonical implementation is the lobes CLI (the lobes-cli package); this skill points a maintainer at it and the scripts/model-runner.sh shim just forwards to lobes.

It is a maintainer convenience — the deployed agent (lobes) does not run it.

When to use

Starting/stopping the server or switching which model is served.
Load-testing a candidate before adopting it (a model card may bless a different engine — verify it serves under our vLLM image anyway).
Producing the correctness/throughput blocks for a per-model doc (see docs/qwen3-32b-nvfp4.md).

How to run

Use the model CLI directly (or scripts/model-runner.sh <args>, which execs it):

model init --apply                 # scaffold ~/.lobes (compose + .env)

# switch the served model (edits .env, recreates the container, waits for health).
# DRY-RUN by default: prints the plan and changes nothing. Add --apply to execute.
model switch sakamakismile/Qwen3.6-27B-Text-NVFP4-MTP --max-model-len 32768          # preview
model switch sakamakismile/Qwen3.6-27B-Text-NVFP4-MTP --max-model-len 32768 --apply

model serve --apply                # start the server (alias: start)
model stop --apply                 # stop + remove the container

model assess                       # correctness probes + reasoning-trace field (markdown)
model benchmark                    # decode throughput + prefill latency (markdown)
model status                       # current model + container health (read-only)

Write verbs (switch, serve, stop, init) are dry-run by default and require --apply — per CLAUDE.md's mutation-safety rule. --port defaults to VLLM_PORT in .env (then 8000). assess/benchmark/status are read-only. assess and benchmark each emit a markdown block ready to paste into a per-model doc; both are stdlib-only (no third-party deps).

Notes

Only one ~30B-class model fits on a single GB10 at a time; switch does docker compose down before up, so it frees the prior model first.
The reasoning-trace field name varies by vLLM build (reasoning on nv26.04, reasoning_content on older builds); assess detects and reports whichever carried the trace.
.env is git-ignored and per-machine — switch writes the VLLM_* keys there. Nothing this skill writes is committed.

More from this repository

same repository

recall

agentculture/lobes-cli

Search the shared eidetic memory store and get back ranked, provenanced records. Drives `eidetic recall` with four search modes — exact (verbatim substring), approximate (vector/semantic), keyword (BM25 lexical), and hybrid (a weighted blend of vector+keyword, the default) — each hit carrying its text, full metadata, a relevance `score`, and a freshness `signal`. Recall passively reinforces matched records (bumps last_recall + recall_count). Shadowed and archived records are excluded by default; use --include-shadowed / --include-archived to retrieve them. The store lives at ~/.eidetic/memory (a home-dir path outside any git worktree); the wrapper defaults queries to this agent's PERSONAL, PRIVATE scope (`--scope lobes --visibility private`, suffix read from culture.yaml) — matching where /remember writes — so a no-flag recall returns this agent's own private records plus the shared public pool, and Claude and the colleague backend recall each other's memories because both resolve the same suffix via this ski

2026-06-263

remember

agentculture/lobes-cli

Ingest records into the shared eidetic memory store so they can be recalled later. Drives `eidetic remember`: accepts one record as a JSON object, or a batch as NDJSON on stdin for bulk ingest. Upsert is idempotent by id (and dedups by content hash) — re-remembering updates in place, never duplicates. Stamps a `created` date on every record at ingest time. Accepts `supersedes` (id of the record this one replaces, for within-scope shadowing via `sweep`) and `links` (list of related-memory ids). The store lives at ~/.eidetic/memory (a home-dir path outside any git worktree), and the wrapper defaults records to this agent's PERSONAL, PRIVATE scope (`--scope lobes --visibility private`, suffix read from culture.yaml) so they don't leak to a default/other-scope recall — Claude and the colleague backend still share them because both resolve the same suffix via this skill. Pass `--visibility public` to contribute to the shared public pool instead. Use when the user says "remember this", "store this", "save to memory

2026-06-263

ask-colleague

agentculture/lobes-cli

Ask colleague — a *different* backend/model than you (e.g. a local vLLM Qwen) — to take a scoped repo task off your plate, then fold its answer back. The point isn't a stronger model; it's a second, independent mind, and that diversity is the value: `ask-colleague review` gets a candid second opinion on a diff, `ask-colleague explore` gets a fresh read of an area, `ask-colleague write` hands off a small implementation, `ask-colleague feedback` grades a finished work item (the ROI loop), and `ask-colleague clean` reaps stale/corrupt `colleague/*` branches a crashed run left behind (which can break `git fetch`). Pilot a running work item with `monitor`/`guide`/`stop`. Reach for it REFLEXIVELY, the way you'd lean over to the teammate at the next desk — not only when asked: before you present or open a PR on a non-trivial committed diff, run `review` for a diverse second opinion; for a fresh read of an unfamiliar area whose answer is independent of your current context, run `explore`. Both are read-only — isolate

2026-06-223

cicd

agentculture/lobes-cli

Steward's CI/CD lane, layered on `agex pr`. Delegates lint / open / read / reply / delta to agex; adds two steward extensions — `status` (SonarCloud quality gate + hotspots + unresolved-thread tally) and `await` (read --wait + status with non-zero exit on Sonar ERROR or unresolved threads). Use when: creating PRs in steward, handling review feedback, polling CI status, or the user says "create PR", "review comments", "address feedback", "resolve threads". Renamed from `pr-review` in steward 0.7.0; rebased on agex in 0.12.0.

2026-06-223

communicate

agentculture/lobes-cli

Cross-repo + mesh communication from steward: file tracked GitHub issues on sibling repos, fetch issues from sibling repos to inline current state into briefs, and send live messages to Culture mesh channels. Use when the next step lives outside steward (a brief for a sibling-repo agent, a status ping for a Culture channel, or pulling an issue body + comments into context). Issue posts auto-sign with `- steward (Claude)`; mesh messages are unsigned (the IRC nick is the speaker). Not for in-steward issues — use `gh issue create` or the `cicd` skill for those. Renamed from `coordinate` in steward 0.8.0; absorbed `gh-issues` in 0.9.1. Issue I/O is backed by `agtag` (>=0.1) starting in 0.11.0.

2026-06-223

doc-test-alignment

agentculture/lobes-cli

Verify that committed docs (README.md, CLAUDE.md, SKILL.md descriptions) still describe what the code and tests actually do. Use at the end of a plan, before PR creation, or when the user says "check doc-test alignment", "verify docs", or "do the docs still match the code". STUB — `scripts/check.sh` exits with a not-yet-implemented error today; the contract for what it will do lives in this file.

2026-06-223

name	model-runner
description	Run, assess, and switch the local vLLM model via the `lobes` CLI (lobes-cli). Use when changing the served model, starting/stopping the server, load-testing a candidate (e.g. an NVFP4 checkpoint), or producing the numbers for a per-model doc under docs/.

model-runner

It is a maintainer convenience — the deployed agent (lobes) does not run it.

When to use

Starting/stopping the server or switching which model is served.
Load-testing a candidate before adopting it (a model card may bless a different engine — verify it serves under our vLLM image anyway).
Producing the correctness/throughput blocks for a per-model doc (see docs/qwen3-32b-nvfp4.md).

How to run

Use the model CLI directly (or scripts/model-runner.sh <args>, which execs it):

model init --apply                 # scaffold ~/.lobes (compose + .env)

# switch the served model (edits .env, recreates the container, waits for health).
# DRY-RUN by default: prints the plan and changes nothing. Add --apply to execute.
model switch sakamakismile/Qwen3.6-27B-Text-NVFP4-MTP --max-model-len 32768          # preview
model switch sakamakismile/Qwen3.6-27B-Text-NVFP4-MTP --max-model-len 32768 --apply

model serve --apply                # start the server (alias: start)
model stop --apply                 # stop + remove the container

model assess                       # correctness probes + reasoning-trace field (markdown)
model benchmark                    # decode throughput + prefill latency (markdown)
model status                       # current model + container health (read-only)

Notes

Only one ~30B-class model fits on a single GB10 at a time; switch does docker compose down before up, so it frees the prior model first.
The reasoning-trace field name varies by vLLM build (reasoning on nv26.04, reasoning_content on older builds); assess detects and reports whichever carried the trace.
.env is git-ignored and per-machine — switch writes the VLLM_* keys there. Nothing this skill writes is committed.