model-onboarding Skill
>- 将新一代或同类模型接入 oh-my-hermes:探测路由器识别、研究官方契约、编写特性到计数器的校准、在两条通道中放置路由、仅按文档价目表定价、基于已上线路由设置机器配置、用关卡验证,并以一组基准测试收尾。当模型家族发布新一代时使用(例如“Fable 5.1 出了”、“GLM 5.4 发布了”、“接入新模型”、“把模型加进 chains”)。
安装方式:把技能目录放入 ~/.claude/skills/(Claude Code)或在 claude.ai 设置中启用;也可复制右侧安装命令一键添加。
技能指令原文(SKILL.md)
Model onboarding
The procedure lives in docs/MODEL-ONBOARDING.md.
Read that file and follow it in order.
It is kept there rather than here because Codex, Hermes handoffs, and generic
executor profiles run the same loop, and AGENTS.md requires that no single
executor own a shared surface. This file exists so the loop is reachable as
/model-onboarding; it holds no rules of its own.
Start by reading, in this order:
cat docs/MODEL-ONBOARDING.md
cat MODEL_OPTI.md
Nine things the loop gets wrong most often:
- Recognition before research. Probe every served id and every bare chat
name with omh coding model-route first; a model_family of unknown
means the calibration never attaches. Probe the dated snapshot form too
(-YYYY-MM-DD): some providers serve only that spelling, and it must
resolve the base's contract, chain position, price, and HUD label through
dated_snapshot_base() rather than falling to generic.
- The chain names the id the vendor's API serves, not the model card's
spelling. Read the Hermes provider profile for the family before choosing
the alias — DeepSeek serves deepseek-flash and 400s deepseek-v4.1-flash;
the versioned contract sits behind the pointer as a declared projection
and in EXACT_CONTRACT_POINTER_ALIASES.
- Research runs as four parallel read-only lanes (official docs, Hermes
runtime source, the vendor's own harness, community harnesses), each a
labeled dossier under .omc/research/; the Hermes lane is the one that
shows what the wire carries (id folding, effort overrides, passback).
- Chains move as a set. The Hermes-lane table, its plugin mirror, the
Maestro-lane table, the model-setup skill text, the public docs and site
pages, and the pinned-chain tests all name the old id (the "Files that move
together" table in the procedure lists them). docs/INSTALLATION.md's chain
table is generated (docs chain-table); the two site pages are edited by
hand; the READMEs carry no table. git grep the old id across the whole
tree and move every site in the same commit. Superseded generations leave
the shipped chains and join the recognition-only list; pins hide in
fixtures, so start the full suite in the background at the chain edit.
- Machine config stays provider-neutral. Chains name aliases; the
provider row is a separate concern in model-providers.json. Place the id
with omh model-chains set, never by hand-editing the JSON; an operator
may keep an older generation behind it there, the shipped table does not.
- Served is not released. Prove the route with one
hermes --oneshotcall
and read the usage file (model, provider, cost_status) before any
measurement or placement; gateways want the vendor-prefixed id.
- The override is measured against the block it replaced, on cost. Run the
family arm next to baseline and optimized; expect pass rate to tie
and read the paired token delta, tool calls, and turns. Same pass with
more tokens on the tasks it fails is a sentence that pushes — cut it.
- A routing signal is only as good as the tier's chain head. Measure the
head on the request class it will receive before shipping the signal,
and name the head in every routing claim.
- A separate reviewer lane reads the diff before the commit. The DeepSeek
round's reviewer caught a reverse projection that let -pro / -fast /
-flex aliases label their base id; self-review did not.
Arguments pass through verbatim: the model ids as served (for example
claude-fable-5-1 claude-mythos-5-1).