mise.toml

[tasks.check]
description = "Run cargo check"
run = "cargo check"

[tasks.build]
description = "Build the project"
run = "cargo build"

[tasks.build-release]
description = "Build release binary"
run = "cargo build --release"

[tasks.fmt]
description = "Format code with rustfmt"
run = "cargo fmt"

[tasks.fmt-check]
description = "Check code formatting"
run = "cargo fmt --check"

[tasks.clippy]
description = "Run clippy lints"
run = "cargo clippy --all-features -- -D warnings"

[tasks.test]
description = "Run tests"
run = "cargo test --all-features"

[tasks.run]
description = "Run the TUI application"
run = "cargo run"

[tasks.refresh-benchmarks]
description = "Fetch latest benchmark data from Artificial Analysis API"
run = """
#!/usr/bin/env bash
set -euo pipefail

if [ ! -f .env ]; then
  echo "Error: .env file not found. Create one with AA_API_KEY=your_key"
  exit 1
fi

source .env

if [ -z "${AA_API_KEY:-}" ]; then
  echo "Error: AA_API_KEY not set in .env"
  exit 1
fi

echo "Fetching from Artificial Analysis API..."
xh GET https://artificialanalysis.ai/api/v2/data/llms/models X-API-Key:"$AA_API_KEY" \
  | jq '[.data[] | {
    id: .id,
    name: .name,
    slug: .slug,
    creator: (.model_creator?.slug // null),
    creator_id: (.model_creator?.id // null),
    creator_name: (.model_creator?.name // null),
    release_date: .release_date,
    intelligence_index: (.evaluations?.artificial_analysis_intelligence_index // null),
    coding_index: (.evaluations?.artificial_analysis_coding_index // null),
    math_index: (.evaluations?.artificial_analysis_math_index // null),
    mmlu_pro: (.evaluations?.mmlu_pro // null),
    gpqa: (.evaluations?.gpqa // null),
    hle: (.evaluations?.hle // null),
    livecodebench: (.evaluations?.livecodebench // null),
    scicode: (.evaluations?.scicode // null),
    ifbench: (.evaluations?.ifbench // null),
    lcr: (.evaluations?.lcr // null),
    terminalbench_hard: (.evaluations?.terminalbench_hard // null),
    tau2: (.evaluations?.tau2 // null),
    math_500: (.evaluations?.math_500 // null),
    aime: (.evaluations?.aime // null),
    aime_25: (.evaluations?.aime_25 // null),
    output_tps: (.median_output_tokens_per_second | if . == 0 then null else . end),
    ttft: (.median_time_to_first_token_seconds | if . == 0 then null else . end),
    ttfat: (.median_time_to_first_answer_token | if . == 0 then null else . end),
    price_input: (.pricing?.price_1m_input_tokens // null),
    price_output: (.pricing?.price_1m_output_tokens // null),
    price_blended: (.pricing?.price_1m_blended_3_to_1 // null)
  }]' > data/benchmarks.json

echo "Updated data/benchmarks.json ($(jq length data/benchmarks.json) entries)"
"""

[tasks.refresh-sources]
description = "Fetch AA/Arena/Epoch/LLM Stats data and regenerate all v2 source files"
run = """
#!/usr/bin/env bash
set -euo pipefail

if [ ! -f .env ]; then
  echo "Error: .env file not found. Create one with AA_API_KEY=... and LLM_STATS_API_KEY=..."
  exit 1
fi

source .env

if [ -z "${AA_API_KEY:-}" ]; then
  echo "Error: AA_API_KEY not set in .env"
  exit 1
fi
if [ -z "${LLM_STATS_API_KEY:-}" ]; then
  echo "Error: LLM_STATS_API_KEY not set in .env"
  exit 1
fi

raw=$(mktemp)
work=$(mktemp -d)
trap 'rm -f "$raw"; rm -rf "$work"' EXIT

echo "Fetching from Artificial Analysis API..."
xh GET https://artificialanalysis.ai/api/v2/data/llms/models X-API-Key:"$AA_API_KEY" > "$raw"

# Legacy lane (frozen): byte-identical jq contract with refresh-benchmarks.
jq '[.data[] | {
  id: .id,
  name: .name,
  slug: .slug,
  creator: (.model_creator?.slug // null),
  creator_id: (.model_creator?.id // null),
  creator_name: (.model_creator?.name // null),
  release_date: .release_date,
  intelligence_index: (.evaluations?.artificial_analysis_intelligence_index // null),
  coding_index: (.evaluations?.artificial_analysis_coding_index // null),
  math_index: (.evaluations?.artificial_analysis_math_index // null),
  mmlu_pro: (.evaluations?.mmlu_pro // null),
  gpqa: (.evaluations?.gpqa // null),
  hle: (.evaluations?.hle // null),
  livecodebench: (.evaluations?.livecodebench // null),
  scicode: (.evaluations?.scicode // null),
  ifbench: (.evaluations?.ifbench // null),
  lcr: (.evaluations?.lcr // null),
  terminalbench_hard: (.evaluations?.terminalbench_hard // null),
  tau2: (.evaluations?.tau2 // null),
  math_500: (.evaluations?.math_500 // null),
  aime: (.evaluations?.aime // null),
  aime_25: (.evaluations?.aime_25 // null),
  output_tps: (.median_output_tokens_per_second | if . == 0 then null else . end),
  ttft: (.median_time_to_first_token_seconds | if . == 0 then null else . end),
  ttfat: (.median_time_to_first_answer_token | if . == 0 then null else . end),
  price_input: (.pricing?.price_1m_input_tokens // null),
  price_output: (.pricing?.price_1m_output_tokens // null),
  price_blended: (.pricing?.price_1m_blended_3_to_1 // null)
}]' "$raw" > data/benchmarks.json

mkdir -p data/v2
cargo run --features pipeline --bin transform -- aa "$raw" -o data/v2/aa.json

# --- Arena: latest.json -> snapshot dir -> 6 board JSONs --------------------
echo "Fetching Arena leaderboard snapshot..."
arena_repo="https://raw.githubusercontent.com/oolong-tea-2026/arena-ai-leaderboards/main"
snapshot=$(curl -sf "$arena_repo/data/latest.json" | jq -r '.path // .date // empty')
if [ -z "$snapshot" ]; then
  echo "Error: could not resolve Arena snapshot pointer from latest.json"
  exit 1
fi
mkdir -p "$work/arena_boards"
for board in text vision code agent search document; do
  curl -sf "$arena_repo/data/$snapshot/$board.json" -o "$work/arena_boards/$board.json" \
    || echo "warning: Arena board $board missing in snapshot $snapshot"
done
cargo run --features pipeline --bin transform -- arena "$work/arena_boards" -o data/v2/arena.json

# --- Epoch: download ZIP -> unzip -> transform -----------------------------
echo "Fetching Epoch AI benchmark ZIP..."
curl -sf "https://epoch.ai/data/benchmark_data.zip" -o "$work/epoch.zip"
unzip -q -o "$work/epoch.zip" -d "$work/epoch_csv"
epoch_dir="$work/epoch_csv"
if [ -z "$(find "$epoch_dir" -maxdepth 1 -name '*.csv' -print -quit)" ]; then
  epoch_dir=$(dirname "$(find "$work/epoch_csv" -name '*.csv' -print -quit)")
fi
cargo run --features pipeline --bin transform -- epoch "$epoch_dir" -o data/v2/epoch.json

# --- LLM Stats: bounded paged fetch (key sourced, never echoed) ------------
echo "Fetching LLM Stats rankings + models..."
ls_base="https://api.llm-stats.com/stats/v1"
ls_auth="Authorization: Bearer $LLM_STATS_API_KEY"
: > "$work/ls_rankings_parts.json"
for cat in agents code finance frontend_development general healthcare legal math multimodal reasoning vision; do
  curl -sf -H "$ls_auth" "$ls_base/rankings?category=$cat&limit=50" >> "$work/ls_rankings_parts.json"
  echo >> "$work/ls_rankings_parts.json"
done
jq -s '{ rankings: . }' "$work/ls_rankings_parts.json" > "$work/ls_rankings.json"
: > "$work/ls_models_parts.json"
cursor=""
page=0
while [ "$page" -lt 10 ]; do
  if [ -z "$cursor" ]; then
    resp=$(curl -sf -H "$ls_auth" "$ls_base/models?limit=50")
  else
    resp=$(curl -sf -H "$ls_auth" "$ls_base/models?limit=50&cursor=$cursor")
  fi
  echo "$resp" | jq -c '.models[]?' >> "$work/ls_models_parts.json"
  cursor=$(echo "$resp" | jq -r '.next_cursor // empty')
  page=$((page + 1))
  [ -z "$cursor" ] && break
done
jq -s '{ models: . }' "$work/ls_models_parts.json" > "$work/ls_models.json"
cargo run --features pipeline --bin transform -- \
  llmstats "$work/ls_rankings.json" --models "$work/ls_models.json" -o data/v2/llmstats.json

echo "Updated data/benchmarks.json ($(jq length data/benchmarks.json) entries)"
echo "Updated data/v2/aa.json ($(jq '.models | length' data/v2/aa.json) models)"
echo "Updated data/v2/arena.json ($(jq '.models | length' data/v2/arena.json) models)"
echo "Updated data/v2/epoch.json ($(jq '.models | length' data/v2/epoch.json) models)"
echo "Updated data/v2/llmstats.json ($(jq '.models | length' data/v2/llmstats.json) models)"
"""

[tasks.snapshot-stats]
description = "Snapshot daily jsDelivr CDN hits per data file into data/stats/jsdelivr-history.json (app-launch proxy)"
run = "bash scripts/snapshot-stats.sh"

[tools]
xh = "latest"