pix-data

Pi coding agent extension — shared model data layer. Warms two cached data sources on session start so other extensions (model picker, footer, subagent resolver) can read context window, pricing, and a coding-focused score/rank synchronously without redundant network calls:

• modelgrep — the model catalog (context window, pricing, modalities, capabilities, raw benchmark fields) used as the authoritative source when present.
• benchlm.ai — a leaderboard of 0–100 coding scores used as a fallback when modelgrep's artificial_analysis block is null (currently the common case for the long tail of models).

Both caches live under ~/.cache/pi/ and are shared across every Pi extension using the same DataSource class — whichever extension loads first populates the cache; subsequent extensions read from disk.

Data sources

• modelgrep — GET /api/v1/models?sort=coding&order=desc&limit=200, paginated up to 10 pages (meta.has_more / next_offset). Free, no API key. modelgrep aggregates benchmark numbers from Artificial Analysis. Context window, pricing, and modalities are taken verbatim from the catalog.
• benchlm — GET https://benchlm.ai/api/data/leaderboard. Free, no API key. Each entry has an overallScore (0–100) used as the fallback score when modelgrep's artificial_analysis block is null.

Cache files:

• ~/.cache/pi/modelgrep.json (TTL 24h)
• ~/.cache/pi/benchlm.json (TTL 24h)

On outage the stale cache keeps the picker working until it can refresh.

Scoring methodology

The score a model receives is the first of the following that succeeds, in order:

1. Primary = Artificial Analysis Intelligence Index when present on the modelgrep entry — AA's authoritative composite of 9 independent evals (agents, coding, scientific reasoning, general), already weighted toward agentic work. Rescaled to 0–100 (intelligence / 65 × 100; the current leader scores ~65).
2. Heuristic from modelgrep's raw benchmark fields when the AA index is absent. Weighted blend of the same family of evals AA uses, then mapped onto the index scale by a least-squares line. Both the heuristic weights and the line were jointly tuned against the index on the models that carry both it and the raw benches (index100 ≈ 120.6·heuristic − 10.6, deduped n=29, R²=0.901, leave-one-out RMSE 6.55pt) — a data calibration, not a guessed penalty. The picker exists to choose a model for coding work in an agent, so the heuristic is weighted toward exactly that:

bench	range	measures
coding	0–100	code generation index
scicode	0–1	scientific coding
tau2	0–1	agentic tool-use
agentic	0–100	agentic index
gpqa	0–1	graduate-level reasoning
hle	0–1	hard-exam reasoning

When the index is absent, three sub-scores combine, each a weighted blend of its benches (all normalized to 0–1):

coding_score    = 0.60·(coding/100) + 0.40·scicode
agentic_score   = 0.70·tau2         + 0.30·(agentic/100)
reasoning_score = 0.60·gpqa         + 0.40·hle

heuristic = 0.30·coding_score + 0.60·agentic_score + 0.10·reasoning_score
score     = round(clamp₀₁₀₀(120.6·heuristic − 10.6))   // fitted to the index

1. benchlm.ai fallback — if the model exists in benchlm but modelgrep has no AA index and no raw benches, look up the benchlm overallScore (0–100) and use it verbatim. Match strategy (in lookupBenchlmScore): exact normalized slug, then prefix overlap either way, then take the highest-scoring match on a tie.

Why a heuristic at all, and why these raw evals only: the AA Intelligence Index is the ideal number — but only ~16% of the catalog has it. For the rest we rebuild a comparable score from the same family of raw evals. Crucially we use each raw eval once and never feed intelligence and its components together, nor any _pct field (which is just a percentile-rank of a raw field) — doing so would double-count the same measurement and silently inflate weights you can't see. Independent inputs only → honest weighted average.

Why these weights: an agentic coding model lives or dies on tool-calling and code generation, so agentic_score (0.60) and coding_score (0.30) carry the score; pure reasoning (0.10) is a tiebreaker, not the headline. The split is not arbitrary — a grid search over weight combinations, scored by how well the heuristic predicts the AA index (leave-one-out cross-validation), landed on this agentic-heavy mix. Within each group the dominant bench (tau2 for agentic, raw coding, gpqa) carries most of the weight and a secondary bench refines it.

Missing benchmarks: every blend renormalizes over the fields actually present, so a model missing one bench is diluted only within its own group — it is never zero-penalized or dropped. A model with no benchmarks at all gets a null score (shown as a bare row) and sorts to the bottom.

The exact implementation is codingScore() in src/data.ts; the weights are intentionally easy to tune in one place if your priorities differ.

What's included

Export	Description
modelgrep	DataSource<ModelGrepModel[]> — the modelgrep catalog. TTL 24h → ~/.cache/pi/modelgrep.json
benchlm	DataSource<BenchLMRawEntry[]> — the benchlm.ai leaderboard (fallback scores). TTL 24h → ~/.cache/pi/benchlm.json
DataSource	Generic cached data source class
CACHE_DIR	Resolved cache directory (~/.cache/pi)
buildModelsDevIndex	Build a lookup Map from the catalog (context/cost/modalities)
fetchModelsDevIndex	Fetch the catalog and build the lookup index in one call
lookupInIndex	Fuzzy-match a router model id against an index
lookupModelsDev	Sync lookup by id from in-memory cache (joined on slug)
lookupBenchmark	Sync lookup a model by id — returns score + rank + pricing
benchScoreColor	Map a 0–100 score to a success/warning/error/muted token

Unified config + collapse — now in pix-runtime

The shared ~/.pi/agent/pix.json config loader and the auto-collapse state machine used to live here. They now live in @xynogen/pix-runtime:

• Config: @xynogen/pix-runtime/config (sections in @xynogen/pix-runtime/sections)
• Collapse: @xynogen/pix-runtime/collapse
• The /pix settings command and config APIs are documented in pix-runtime's README.

pix-data no longer exports pixConfig, reloadPixConfig, shouldCollapse, collapseDelayMs, or tickCollapse — update any imports to pix-runtime.

How it works

On session start the extension fires two non-blocking fetches in parallel (modelgrep.get() and benchlm.get()) — Pi session start is not gated on either. If the cache is fresh both fetches are skipped. The cache files live in ~/.cache/pi/ — any Pi extension using the same DataSource shares them automatically.

Install

pi install npm:@xynogen/pix-data

Foundation library. Feature packages install it as a dependency. Install it directly only when you build your own extension on it.

Full distro

This package is part of Pix. The installer sets up Pi and the full distro. See Install for the notes for each OS.

# Linux / macOS
curl -fsSL https://raw.githubusercontent.com/xynogen/pix-mono/main/scripts/install.sh | sh

# Windows
irm https://raw.githubusercontent.com/xynogen/pix-mono/main/scripts/install.ps1 | iex

License

MIT. See LICENSE.