Select LLM Model

Task-driven LLM recommender. Input: a free-form task description. Output: a ranked shortlist of models, each with per-axis rationale, a source citation, and the fetch timestamp — plus an explicit list of data gaps.

Data is fetched live at invocation time through two CLI tools with subcommands (no bundled snapshot): scripts/benchmarks.ts (capability benchmarks) and scripts/openrouter.ts (deployment price, per-provider breakdown, speed). The agent queries them targeted — by category, benchmark, model, or provider — never a firehose. Standings move continuously, so every answer states when it fetched.

Operating principles

• Fail-fast, never fabricate. If no fetch path works, STOP and say so. A source whose tool exits non-zero (unparseable bytes, missing key, missing field) is recorded as a Gap — never an invented score.
• A blocked or failed curl is the fail-fast signal — never route around it. If a curl <url> fetch errors, returns nothing, or reports command not found, record that source as a Gap and move on. NEVER retry it via an absolute path (/usr/bin/curl, /opt/.../curl), an alternate binary (wget, httpie, a language one-liner), a proxy, or a -o /tmp/… download. Working around an unavailable fetch tool to reach the network anyway defeats the purpose and yields fabricated-looking results. If EVERY source's fetch fails this way, STOP — do not produce a ranking.
• A disclaimer does not license a ranking from memory. When you have reported the Gaps, the answer is finished: naming models is itself the forbidden act, whatever label sits above it. "From my training knowledge", "not live-cited", "as of my cutoff", "just so you have something to start from" — each of these still hands the user a model list they did not ask you to guess at, and a hedged list is what they will act on because it is the only list on the page. Say which sources failed and what would make them work, and stop there.
• Rank ONLY the models the tools return. Model names you do not recognize are valid live data, NOT test stubs or placeholders — never dismiss fetched rows, never re-fetch to get "more familiar" models, and never substitute models from memory. If the tools return models X/Y/Z, the shortlist is drawn from X/Y/Z and nothing else.
• Cite everything. Each model's per-axis score names the source it came from and the fetch timestamp. No uncited rankings.
• Disclose assumptions. When the task is vague, state the axis weights you assumed before ranking.

Phase 0 — Tooling check

The tools run as curl <url> | deno run … --stdin. Probe the runtime and the key — but do NOT probe curl:

• Runtime: confirm deno is available (e.g. deno --version). If deno is unavailable, every source becomes a Gap (fail-fast) — do NOT fall back to scraping raw HTML from memory.
• Key: the Artificial Analysis source needs AA_API_KEY (free key at https://artificialanalysis.ai/api). If unset, AA is a Gap — do NOT scrape its JS-rendered leaderboard page.
• Do NOT probe curl with command -v / which / type / --version. The fetch itself is the availability test: the first curl <url> | … pipe in Phase 2 either returns bytes or it fails. A curl that errors or reports command not found means the fetch tool is unavailable — apply the fail-fast rule above (that source is a Gap; if every source fails, STOP with "No working shell fetch tool — cannot retrieve live data; not fabricating a recommendation"). Probing first only leaks a binary path you must not use.

Phase 1 — Derive benchmark categories from the task

benchmarks.ts is queried by category (the closed set below). Assign a 0–3 weight to each category from the task description (adjust cues to intent):

Category	Benchmarks (in the tool)	Cues in the task
general	intelligence, mmlu-pro, gpqa, hle	"general", "hard problem", "reasoning", "PhD", "knowledge", unspecified
coding	coding-index, livecodebench, scicode	"code", "implement", "bug", "leetcode", language names
math	math-index, aime, aime-25, math-500	"math", "prove", "olympiad", "arithmetic"
agentic	terminal-bench-hard, tau2	"agent", "autonomous", "terminal", "call tools", "function calling"
instruction	ifbench	"follow instructions", "format", "constraints"
long-context	lcr	"long context", "large document", "many files"
diff-edit	aider-polyglot	"edit existing code", "apply a patch", "diff", "in-place"
web	webarena	"browse", "web navigation", "fill forms", "browser agent"
computer-use	osworld	"control the desktop", "GUI automation", "computer use"
swe	swe-bench-verified	"resolve GitHub issues", "real-world SE", "repo-level fix"

Rules:

• Vague task → balanced profile. If no category stands out, weight general highest and the rest evenly, and DISCLOSE: "Task under-specified — ranked on a balanced general-purpose profile."
• Price / speed are NOT capability weights. Treat a stated budget ("cheap", "best value", "under $X/Mtok") or latency need ("fast") as a post-rank filter / tie-breaker, fetched from openrouter.ts (price/speed). When the task says nothing about cost or latency, they do not enter the capability score. (For an explicit value/$ ask, price is the primary ranking lens — see Phase 3.)

Phase 2 — Query the data tools

Call the two CLI tools ONLY for the categories you weighted non-zero. Run every command with the working directory set to the directory you read THIS SKILL.md from — the one holding scripts/. Do not reconstruct another skill root (an IDE's own skills/.system/… tree, a home-directory install) from a naming pattern: an exec that fails before the shell starts with No such file or directory is reporting the working directory, not a missing deno. A missing runtime says deno: command not found. (2026-09-06: one run set the workdir to a non-existent .system/select-llm-model, read the cwd error as "Deno is unavailable", and stopped without the pipe.) Always fetch with curl and pipe into the tool with --stdin — this is the ONE supported invocation form (below). Do NOT run the tool without the curl … | … --stdin pipe. A non-zero exit means that source is a Gap. Add --format table for readable output, --top N to cap rows.

benchmarks.ts — one call per weighted category (a category maps to one fetch URL):

# general/coding/math/agentic/instruction/long-context → Artificial Analysis (key):
curl -fsSL -H "x-api-key: $AA_API_KEY" https://artificialanalysis.ai/api/v2/data/llms/models \
  | deno run -A scripts/benchmarks.ts scores --category coding --top 10 --stdin
# diff-edit → Aider:
curl -fsSL https://raw.githubusercontent.com/Aider-AI/aider/main/aider/website/_data/polyglot_leaderboard.yml \
  | deno run -A scripts/benchmarks.ts scores --category diff-edit --stdin
# web → Steel.dev (swap board per category: webarena | osworld | swe-bench-verified):
curl -fsSL https://leaderboard.steel.dev/leaderboards/webarena/ \
  | deno run -A scripts/benchmarks.ts scores --category web --stdin

• scores --category <c> (required) [--benchmark <b>] [--model <substr>] [--top N].
• Output rows: {category, benchmark, source, model, score, higherIsBetter}.
• --benchmark is optional — omit to get all benchmarks of the category.

openrouter.ts — deployment price / provider detail / speed:

curl -fsSL https://openrouter.ai/api/v1/models \
  | deno run -A scripts/openrouter.ts price --sort blended --top 10 --stdin
curl -fsSL https://openrouter.ai/api/v1/models/openai/gpt-4o/endpoints \
  | deno run -A scripts/openrouter.ts providers --model openai/gpt-4o --stdin

• price [--match <substr>] [--sort blended|input|output] [--top N] → {id,name,input,output,blended,context} ($/Mtok).
• models [--match <substr>] [--top N] → {id,name,context} — resolve the exact author/slug for providers.
• providers --model <author/slug> (required) [--sort price|uptime] → per provider {provider,input,output,cacheRead?,uptime30m,uptime1d,context,quantization,status}.
• speed [--match <substr>] [--top N] → {model,speed} (median tok/s, from the AA URL — needs AA_API_KEY).

Source notes. Artificial Analysis absorbs ~15 benchmarks in one keyed call. OpenRouter is the real deployment price/value source. Steel.dev adds agent + SE benchmarks AA lacks; its web/computer-use boards are system/submission-attributed (agent scaffold + model, e.g. "Claude Code + GBOX MCP") — cite those as agent-system scores, not bare-model; swe lists bare models. latency/throughput are not available (null in the OpenRouter API) — speed (AA median tok/s) is the only throughput proxy.

Known Gaps (no stable keyless endpoint — report as Gaps, do NOT scrape):

• Human-preference Elo (LMArena) and fluid reasoning (ARC-AGI) — JS-rendered, no clean endpoint, not in the AA API.
• Aggregator SPAs LLM-Stats and BenchLM — Next.js apps with no exposed _next/data buildId at probe time.
• AgentBench — leaderboard is a Google Sheet + image, no machine-readable feed.
• Epoch AI Capabilities Index — no confirmed CSV/JSON endpoint.
• Scale SEAL / Showdown and SWE-bench Pro — private datasets, no keyless feed.

Phase 3 — Normalize and rank

Parser scores are on incomparable native scales (intelligence index, 0–1 bench fractions, $/Mtok). Merge rows across sources by model name/slug (best-effort; an unmatched model is "no data" on that axis, never zero), then:

1. Per axis, convert each model's raw score to a percentile within the fetched set. For rows with higherIsBetter: false (price), invert so cheaper ranks higher — never let a low price sink a model. 1a. Collapse a category's benchmarks. A category groups correlated benchmarks (general = intelligence/mmlu-pro/gpqa/hle; math = the four math benchmarks). Average a category's benchmark percentiles into ONE contribution before the weighted sum, so a model is not multi-counted for being broadly smart.
2. Weighted-sum the per-category contributions using the Phase-1 weights.
3. Apply price/speed as a filter/tie-breaker per Phase 1. For an explicit value/$ ask, rank by the price percentile (or capability-per-dollar) directly and disclose that lens.
4. A model missing from an axis is excluded from that axis's average — do NOT treat missing as zero.

Phase 4 — Deployment enrichment + output

For the top 1–2 ranked models, optionally enrich with where to deploy them (same curl … | … --stdin form as Phase 2): resolve the slug via openrouter.ts models --match <name>, then openrouter.ts providers --model <slug> --sort price (or --sort uptime) for the cheapest / most-reliable provider. Disclose that latency/throughput are unavailable (use speed for the AA throughput proxy).

Report, in this order:

1. Assumed category weights (and the vague-task disclosure if it applied).
2. Ranked shortlist (top 3–5). For each model: overall rank, then one line per contributing category/benchmark giving the source, the model's standing, and that this came from the live fetch.
3. Deployment (optional) — for the top pick(s), the cheapest/most-reliable provider from openrouter.ts providers.
4. Fetch timestamp — when the data was retrieved, as a literal line in the answer: Data fetched: <YYYY-MM-DD HH:MM UTC>. Saying that a fetch happened is not the same as saying when; without the clock the reader cannot tell yesterday's leaderboard from last year's. Take the time from the machine (date -u), never from memory, and print the line even when only one source answered.
5. Gaps — every source that failed, lacked a key, or lacked a model, with the reason.

If every source ended up a Gap, do NOT produce a ranking — report the gaps and stop. Partial data is fine: rank on what you have and disclose the rest.

A source that RETURNED ROWS is not a Gap, and nothing about the rows can make it one. Unfamiliar model names are the expected case — leaderboards move faster than any training cut-off — so "these look like fixtures", "this is mock data" and "the sandbox stubbed curl" are not findings; they are you declining to read what you fetched. Rank the rows. A Gap is a fetch that failed: a non-zero exit, a missing key, an unparseable body, no matching rows.

Scope

Recommender only. It does not benchmark models itself, does not bundle a dataset, and does not promise freshness beyond the fetch it just performed. The parser fixtures inlined in scripts/parse-*_test.ts are test data, not a live snapshot, and need periodic re-capture as source schemas drift.