primexiao/agent-skills

Model Radar

OpenRouter model selection CLI. Parses natural-language constraints into token:value arguments, queries a local cache of model metadata + rankings, and emits JSON for the agent to render as a markdown table.

Setup

• Runtime: Node.js 18+ (no install step and no npm dependencies)
• No API key — uses OpenRouter public endpoints
• Run all commands from this skill directory (the one containing this file)
• First run populates the cache from the network; subsequent calls hit the local cache (models TTL 6h, rankings TTL 24h)
• Self-updates are enabled by default. Each CLI run checks the fixed GitHub source at most once every 24h and applies updates automatically for installed copies. Set MODEL_RADAR_AUTO_UPDATE=0 to disable.

Workflow

0. Run node scripts/self-update.js auto before using the skill. If stderr prints MODEL_RADAR_UPDATED, re-read this SKILL.md before continuing.
1. Parse the user's request into token:value arguments using the Intent → Token mapping below
2. Pick the command: list (filter+sort), recommend (weighted score against a preset), compare (side-by-side)
3. Run from the skill directory. scripts/main.js also runs the same self-update preflight automatically, so direct CLI usage stays current.
4. Render the JSON output as a markdown table (see Output)
5. End with Data as of {fetched_at}; suggest refresh if older than 12h

Commands

node scripts/main.js list      [tokens...]    # filter + sort
node scripts/main.js recommend [tokens...]    # weighted score against a preset
node scripts/main.js compare   <id|name>...   # side-by-side comparison
node scripts/main.js refresh                  # force-refresh both caches
node scripts/self-update.js status            # inspect update state
node scripts/self-update.js check             # check GitHub for updates
node scripts/self-update.js apply             # apply the latest skill update

Self-Update

• Source is fixed by package.json.repository: primexiao/agent-skills, path skills/model-radar.
• Installed copies update automatically. Source checkouts skip auto-update unless MODEL_RADAR_AUTO_UPDATE_SOURCE=1 is set.
• Local uncommitted changes under the skill directory cause auto-update to skip.
• Direct scripts/main.js usage restarts the current command after a successful self-update so the result is produced by the updated code.
• Updates use a temp checkout, validate required files, back up the current skill, replace non-runtime files, and roll back on failure.
• Successful updates clear cache/raw.json, cache/models.json, and cache/rankings.json so new code does not mix with old data cache.
• Self-update status and warnings go to stderr. Command JSON stays on stdout.

Token Syntax

Tailwind-inspired positional token:value arguments. Order doesn't matter.

Token	Meaning	Value format	Examples
sort	sort dimension	preset word (see below)	sort:cheap sort:popular
price	$/MTok blended	..max / min.. / min..max / bare=max	price:..10 price:5..20
ctx	context length	same, k/m suffix, bare=min	ctx:128k ctx:128k..1m
out	max output tokens	same, bare=min	out:32k
cap	capabilities (AND)	comma-separated	cap:tool_use,reasoning
in	input modalities (AND)	comma-separated	in:image in:audio,image
gen	output modalities (AND)	comma-separated	gen:image gen:audio
open	open-source filter	true / false	open:true
top	result limit	number (default: list=20, recommend=3)	top:5
preset	scoring preset (recommend only)	preset name	preset:coding

sort values: cheap expensive new (default) ctx popular throughput latency

Capabilities: tool_use reasoning structured_output vision long_output prompt_caching web_search audio_input video_input file_input image_generation audio_output

Input modalities: text image audio video file Output modalities: text image audio

Recommend presets: coding batch chat vision (see config/presets.json for weights and required capabilities)

Intent → Token

User says	Token(s)
"budget $X" / "under $X"	price:..X
"at least $X" / "premium tier"	price:X..
"at least Nk context"	ctx:Nk
"needs tool use / vision / reasoning"	cap:tool_use / cap:vision / cap:reasoning
"structured output" / "JSON mode"	cap:structured_output
"can take audio / image input"	in:audio / in:image
"can generate images / audio"	gen:image / gen:audio
"open source only"	open:true
"cheapest" / "most affordable"	sort:cheap
"most popular" / "trending"	sort:popular
"newest" / "latest"	sort:new
"fastest" / "lowest latency"	sort:latency
"highest throughput"	sort:throughput
"coding agent"	recommend preset:coding
"batch / bulk job"	recommend preset:batch
"vision agent"	recommend preset:vision

Model Matching (compare)

Match priority for each argument to compare:

1. Exact id — anthropic/claude-sonnet-4-5
2. Suffix-exact — claude-sonnet-4 → anthropic/claude-sonnet-4 (never the -fast variant)
3. Partial substring on id/name, newest release wins — claude-sonnet → latest Sonnet

Output

Always render the JSON output as a markdown table — tables are the primary display format.

list

| # | Model | $/MTok (P/C/Blend) | Ctx | Output | Caps | Pop# |

• $/MTok: prompt / completion / blended prices
• Caps: short tags (tool reason vision …)
• Pop#: popularity rank if rankings cache is fresh, else —

recommend

| Rank | Model | Score | Cost↑ | Ctx↑ | Out↑ | Cap↑ | $/MTok | Caps |

Add a one-line rationale per model. Score and per-dimension scores are in [0, 1].

compare

Vertical table — one column per model, one row per dimension:

| Dimension          | Model A | Model B | Model C |
| ------------------ | ------- | ------- | ------- |
| Price (P/C/Blend)  |   …     |   …     |   …     |
| Context            |   …     |   …     |   …     |
| Max output         |   …     |   …     |   …     |
| Input modalities   |   …     |   …     |   …     |
| Output modalities  |   …     |   …     |   …     |
| Capabilities       |   …     |   …     |   …     |
| Popularity rank    |   …     |   …     |   …     |
| Weekly requests    |   …     |   …     |   …     |

Always end with: Data as of {fetched_at}. If older than 12h, suggest the user run node scripts/main.js refresh.

Caching

cache/
  raw.json       — OpenRouter frontend card data adapted to local raw schema (TTL 6h)
  models.json    — enriched: normalized pricing + capability labels (TTL 6h)
  rankings.json  — popularity / throughput / latency + usage analytics (TTL 24h)
config/
  presets.json
  tier1-categories.json

• Cache files live under cache/ and are not committed — first run hits the network
• On TTL expiry, the CLI refetches with a 10s timeout
• On any fetch failure, it falls back to the stale cache and emits [warn] on stderr — so read-only / offline sandboxes still answer (possibly stale)
• A successful skill self-update deletes the model-data cache files. The next model command repopulates them.

Sandbox Notes

• Codex CLI: needs -s workspace-write (CLI writes to cache/) and network egress to openrouter.ai. Within TTL, no network calls.
• First call in a fresh checkout is the only one that strictly requires network.
• Self-update needs filesystem write access to this skill directory and network access to GitHub. If unavailable, it warns or skips and the model command still runs with the current local copy.

Known Issues

• The OpenRouter frontend API (/api/frontend/v1/models/find?active=true&fmt=cards) is unofficial; if it fails, the CLI falls back to stale cache where available and emits [warn].

Data Sources

• Models: https://openrouter.ai/api/frontend/v1/models/find?active=true&fmt=cards
• Rankings: same endpoint with order=top-weekly, order=throughput-high-to-low, and order=latency-low-to-high