使い方
- インストールコマンドをコピーし、自分の環境へ導入する前にソースを確認します。
- 必要なツール、認証情報、対応環境を原文で確認します。
- 入力と成果物を指定し、出力を資料と照合します。
npx skills add https://github.com/firegoat-ai/skills --skill harnesses-and-models開始前の確認
- 導入にはモデル契約、API枠、ファイルへのアクセス許可は含まれません。
- スター数や導入数は活動の指標であり、独立した品質・安全性試験ではありません。
見出し、例、コード、表、リンク、参照画像を含む原文を表示しています。
Harnesses and models
Every run executes on a harness · model combo: the harness is the coding agent (the product being measured is used through it), the model is the LLM inside it. The same prompt can behave very differently across combos, so knowing which combo produced a run is load-bearing for any comparison.
The registry
The only combos the queue accepts:
| Harness | Models |
|---|---|
claude-code | claude-opus-5-5 (Opus 5.5), claude-opus-5 (Opus 5), claude-opus-4-8 (Opus 4.8), claude-opus-4-6 (Opus 4.6), claude-sonnet-4-6 (Sonnet 4.6), claude-sonnet-5-5 (Sonnet 5.5), claude-sonnet-5 (Sonnet 5), claude-haiku-5-5 (Haiku 5.5), claude-haiku-4-5-20251001 (Haiku 4.5), claude-fable-5-1 (Fable 5.1), claude-fable-5 (Fable 5) |
codex | openai/gpt-5.5 (GPT-5.5), openai/gpt-5.6 (GPT-5.6), openai/gpt-5.6-sol (GPT-5.6 Sol), openai/gpt-5.6-terra (GPT-5.6 Terra), openai/gpt-5.6-luna (GPT-5.6 Luna) |
opencode | openai/gpt-5.5 (GPT-5.5), openai/gpt-5.6 (GPT-5.6), anthropic/claude-opus-4-8 (Opus 4.8), anthropic/claude-opus-4-6 (Opus 4.6), xai/grok-4.3 (Grok 4.3), xai/grok-4.5 (Grok 4.5), xai/grok-4.6 (Grok 4.6), xai/grok-4.7 (Grok 4.7) |
openclaw | openai/gpt-5.5 (GPT-5.5), openai/gpt-5.6 (GPT-5.6) |
cursor-cli | cursor/auto (Auto) |
fx | vercelaigateway/anthropic/claude-sonnet-4.5 (Sonnet 4.5), vercelaigateway/anthropic/claude-opus-4.6 (Opus 4.6) |
v0 | v0/v0-mini (v0 Mini), v0/v0-pro (v0 Pro), v0/v0-max (v0 Max) |
grok-build | grok-build-0.1 (Grok Build 0.1) |
Default when nothing is pinned: claude-code + claude-opus-4-8.
Historical quirk: null harness
Older runs recorded a model but no harness. Resolution rule: a null harness on a model that claude-code can run means it WAS the claude-code default; only models claude-code can't run (gpt-5.5, gpt-5.6, cursor/auto, Kimi, Grok) stay genuinely unknown.
Where harness · model is exposed
The run-centric tools carry the submitted combo directly: list_runs and get_run return top-level harness and model fields (the config the run was triggered with) — that's the reliable source. The report-centric tools do NOT: get_report may carry a model inside providerMeta.model but no harness, and list_reports / get_prompt_benchmarks history items return neither. So if a user asks "which harness/model ran this?": for a run, read list_runs/get_run harness/model; for a report slug only, try get_report().providerMeta.model, else say the report tools don't expose the harness and point them at the run's dashboard report page (see the linking-to-the-dashboard skill) rather than guessing.
Pinning at trigger time
request_run({ promptId, harnessModel: { harness, model } }) pins the combo — it's validated against the registry above and unknown combos are rejected. Omit harnessModel to let the queue use the default.
Comparability
A prompt accumulates runs across multiple harness·model combos over time. Metrics like time and cost shift with the harness — there is no normalization. Compare metrics within the same combo when you can; when combos differ, say so explicitly instead of presenting a raw delta as a clean win. A lane's history[] from get_prompt_benchmarks can silently mix combos, so cross-check the combo (via list_runs/get_run) before reading a trajectory as a clean trend.
Practical loop
- When a question hinges on harness/model, read
list_runs/get_runharness/model; for a report-only slug, tryget_report().providerMeta.model, else say the report tools don't expose it and link the dashboard report page. - Before comparing runs or reading a trend, ask whether the combos match; flag mixed-combo deltas explicitly.
- When triggering a run that must repeat a specific setup, pin it with
request_run'sharnessModel; otherwise omit it for the default.
