@lythos/skill-arena 0.9.38 → 0.9.40

package/README.md

@@ -49,26 +49,23 @@ Note: Claude `-p` mode has known issues with web tools in Bun.spawn (deferred to
 ```bash
 bun add -d @lythos/skill-arena
 # or use directly
-bunx @lythos/skill-arena@0.9.38 <command>
+bunx @lythos/skill-arena@0.9.40 <command>
 ```
 
 ## Quick Start
 
 ```bash
-# Mode 1: Compare two skills on the same task
-bunx @lythos/skill-arena@0.9.38 \
-  --task "Generate auth flow diagram" \
-  --skills "design-doc-mermaid,mermaid-tools" \
-  --criteria "syntax,context,token"
-
-# Mode 2: Compare full deck configurations
-bunx @lythos/skill-arena@0.9.38 \
-  --task "Generate auth flow diagram" \
-  --decks "./decks/minimal.toml,./decks/rich.toml" \
-  --criteria "quality,token,maintainability"
+# Single: test a deck with one agent (exec shortcut)
+bunx @lythos/skill-arena@0.9.40 single \
+  --brief "Generate auth flow diagram" \
+  --deck ./examples/decks/documents.toml
+
+# Vs: compare multiple decks side by side (declarative)
+bunx @lythos/skill-arena@0.9.40 vs \
+  --config examples/arena/research-compare/arena.toml
 
 # Visualize results
-bunx @lythos/skill-arena@0.9.38 viz tmp/arena-<id>/
+bunx @lythos/skill-arena@0.9.40 viz tmp/arena-<id>/
 ```
 
 ## Commands
@@ -77,16 +74,16 @@ bunx @lythos/skill-arena@0.9.38 viz tmp/arena-<id>/
 
 ```bash
 # Print execution plan without running
-bunx @lythos/skill-arena@0.9.38 run --config arena.toml --dry-run
+bunx @lythos/skill-arena@0.9.40 vs --config arena.toml --dry-run
 
 # Execute with per-side runs_per_side and statistical aggregation
-bunx @lythos/skill-arena@0.9.38 run --config arena.toml
+bunx @lythos/skill-arena@0.9.40 vs --config arena.toml
 ```
 
 ### CLI-flag mode (backward compat)
 
 ```
-bunx @lythos/skill-arena@0.9.38 run \
+bunx @lythos/skill-arena@0.9.40 run \
   --task ./TASK-arena.md \
   --players ./players/claude.toml \
   --decks ./decks/run-01.toml,./decks/run-02.toml \
@@ -96,13 +93,13 @@ bunx @lythos/skill-arena@0.9.38 run \
 ### Scaffold mode (legacy, manual execution)
 
 ```
-bunx @lythos/skill-arena@0.9.38 scaffold --task "..." --skills a,b
+bunx @lythos/skill-arena@0.9.40 scaffold --task "..." --decks a.toml,b.toml
 ```
 
 ### Viz
 
 ```bash
-bunx @lythos/skill-arena@0.9.38 viz runs/arena-<id>/
+bunx @lythos/skill-arena@0.9.40 viz runs/arena-<id>/
 ```
 
 ## Skill Documentation
@@ -116,7 +113,7 @@ The agent-visible **Skill** layer documentation is here:
 Part of the [lythoskill](https://github.com/lythos-labs/lythoskill) ecosystem — the thin-skill pattern separates heavy logic (this npm package) from lightweight agent instructions (SKILL.md).
 
 ```
-Starter (this package) → npm publish → bunx @lythos/skill-arena@0.9.38 ...
+Starter (this package) → npm publish → bunx @lythos/skill-arena@0.9.40 ...
 Skill   (packages/<name>/skill/)     → build → SKILL.md + thin scripts
 Output  (skills/<name>/)             → git commit → agent-visible skill
 ```
@@ -137,15 +134,6 @@ arena.toml  →  ArenaToml (Zod)  →  ExecutionPlan (pure)  →  per-cell agent
 
 Built on `@lythos/test-utils` shared infrastructure.
 
-## Test Coverage
-
-| Layer | Count | CI | Notes |
-|-------|-------|----|-------|
-| Unit tests | 41 | ✅ | TOML parser, player resolution, Pareto, stats |
-| Agent BDD | — | ❌ | Requires `claude` CLI; run locally |
-
-Pareto frontier is a **deterministic algorithm** — never delegated to LLM. 8 unit tests cover dominance, cross-dominance, transitive chains, partial criteria, and empty scores.
-
 ## License
 
 MIT

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@lythos/skill-arena",
-  "version": "0.9.38",
+  "version": "0.9.40",
   "description": "Skill Arena — benchmark skill effectiveness with controlled-variable comparison",
   "keywords": [
     "ai-agent",

package/src/cli.ts

@@ -37,52 +37,49 @@ function printHelp(): void {
   console.log(`🎭 lythoskill-arena — Skill comparison runner
 
 Usage:
-  lythoskill-arena agent-run --task <path> --deck <path> [--player kimi] [--out <dir>] [--timeout <ms>]
-  lythoskill-arena agent-run --brief "<prompt>" --deck <path> [--out <dir>] [--timeout <ms>]
-  lythoskill-arena run --task <path> --players <A.toml,B.toml> --decks <A.toml,B.toml> --criteria <c1,c2,...> [--out <dir>]
-  lythoskill-arena scaffold --task "<description>" --skills <skill1,skill2,...>
+  lythoskill-arena single --task <path> --deck <path> [--player kimi] [--out <dir>] [--timeout <ms>]
+  lythoskill-arena single --brief "<prompt>" --deck <path> [--out <dir>] [--timeout <ms>]
+  lythoskill-arena vs --config arena.toml [--dry-run]
   lythoskill-arena scaffold --task "<description>" --decks <deck1,deck2,...>
   lythoskill-arena viz <arena-dir>
 
 Commands:
-  run       Run arena programmatically (declarative arena.toml or CLI flags)
+  single    Single-player deck test (exec shortcut): test a deck with one player
+  vs        Multi-side comparison: run arena from declarative arena.toml
   scaffold  Create arena directory structure (legacy, manual subagent execution)
   viz       Visualize arena report (ASCII charts)
 
 Options:
-  -t, --task <path|desc> Task description or path to TASK-arena.md
-  -s, --skills <list>    Comma-separated skill names (scaffold only)
-      --decks <list>     Comma-separated deck paths
-  -c, --criteria <list>  Evaluation criteria (default: syntax,context,logic,token)
-      --players <list>   Comma-separated player.toml paths (CLI run only)
-      --config <path>    Path to arena.toml (declarative mode, k8s-style)
-      --dry-run          Print execution plan without running (with --config)
-      --control <skill>  Control skill for comparison (scaffold only)
-      --out <dir>        Output directory (run: defaults to runs/arena-<id>)
-  -d, --dir <dir>        Output directory (scaffold: defaults to tmp)
-  -p, --project <dir>    Project directory (default: .)
+  -t, --task <path|desc> Task description or path to TASK-arena.md / .agent.md
+      --deck <path>      Deck path (single only)
+      --brief "<text>"   Inline task description (single only, alternative to --task)
+      --player <name>    Agent player (single only, default: kimi)
+  -c, --criteria <list>  Evaluation criteria (scaffold only, default: syntax,context,logic,token)
+      --config <path>    Path to arena.toml (vs only)
+      --dry-run          Print execution plan without running (vs --config only)
+      --out <dir>        Output directory
+  -d, --dir <dir>        Parent dir (scaffold: defaults to tmp)
+  -p, --project <dir>    Project root (default: .)
+      --timeout <ms>     Subagent timeout (single only)
 
 Examples:
-  # Single agent run (simplest path)
-  lythoskill-arena agent-run --task ./TASK.md --deck ./deck.toml
-  lythoskill-arena agent-run --task ./TASK.md --deck ./deck.toml --player kimi --out ./output
+  # Single-player deck test (exec shortcut)
+  lythoskill-arena single --task ./TASK.agent.md --deck ./deck.toml
+  lythoskill-arena single --brief "Generate auth flow diagram" --deck ./deck.toml --player kimi
 
-  # Declarative mode (k8s-style)
-  lythoskill-arena run --config ./arena.toml
-  lythoskill-arena run --config ./arena.toml --dry-run
-
-  # CLI-flag mode (backward compat)
-  lythoskill-arena run --task ./TASK-arena.md --players ./players/claude.toml --decks ./decks/run-01.toml,./decks/run-02.toml --criteria coverage,relevance
+  # Multi-side comparison (declarative)
+  lythoskill-arena vs --config ./arena.toml
+  lythoskill-arena vs --config ./arena.toml --dry-run
 
   # Legacy scaffolding
-  lythoskill-arena scaffold --task "Refactor auth module" --skills skill-a,skill-b
+  lythoskill-arena scaffold --task "Refactor auth module" --decks ./decks/a.toml,./decks/b.toml
   lythoskill-arena viz runs/arena-20260504
 `)
 }
 
-// ── agent-run: single agent execution (simplest path) ────────────────────
+// ── single: single-player deck test (exec shortcut) ──────────────────────
 
-async function agentRun(args: string[]) {
+async function singleRun(args: string[]) {
   const opts: Record<string, string | undefined> = {}
   for (let i = 0; i < args.length; i++) {
     if (args[i] === '--task' || args[i] === '-t') opts.task = args[++i]
@@ -94,11 +91,16 @@ async function agentRun(args: string[]) {
   }
 
   if (!opts.deck) {
-    console.error('❌ --deck <path> is required')
+    console.error(`❌ --deck <path> is required.
+   Usage: lythoskill-arena single --deck ./deck.toml --task ./scenario.agent.md
+          lythoskill-arena single --deck ./deck.toml --brief "your task description"
+   Example decks: examples/decks/scout.toml, examples/decks/documents.toml`)
     process.exit(1)
   }
   if (!opts.task && (!opts.brief || !opts.brief.trim())) {
-    console.error('❌ --task <path> or --brief "<prompt>" is required and cannot be empty')
+    console.error(`❌ --task <path> or --brief "<prompt>" is required.
+   Usage: lythoskill-arena single --deck ./deck.toml --task ./scenario.agent.md
+          lythoskill-arena single --deck ./deck.toml --brief "your task description"`)
     process.exit(1)
   }
 
@@ -122,7 +124,10 @@ async function agentRun(args: string[]) {
     deckPath = dest
   } else {
     deckPath = resolve(opts.deck)
-    if (!deckExists(deckPath)) { console.error(`❌ Deck file not found: ${deckPath}`); process.exit(1) }
+    if (!deckExists(deckPath)) { console.error(`❌ Deck file not found: ${deckPath}
+   Create one: examples/decks/scout.toml (minimal), examples/decks/documents.toml (documents)
+   Or fetch:   curl -fsSL https://raw.githubusercontent.com/lythos-labs/lythoskill/main/examples/decks/scout.toml > deck.toml
+   Or create:  see https://github.com/lythos-labs/lythoskill/tree/main/examples/decks/`); process.exit(1) }
   }
 
   const { useAgent } = await import('@lythos/test-utils/agents')
@@ -142,7 +147,10 @@ async function agentRun(args: string[]) {
   const scenarioOpt: Record<string, unknown> = {}
   if (opts.task) {
     const taskPath = resolve(opts.task)
-    if (!existsSync(taskPath)) { console.error(`❌ Task file not found: ${taskPath}`); process.exit(1) }
+    if (!existsSync(taskPath)) { console.error(`❌ Task file not found: ${taskPath}
+   Create a .agent.md scenario or use --brief for inline tasks.
+   Format: frontmatter (name, description, timeout) + Given/When/Then/Judge sections.
+   Example: see playground/arena-one-shot/TASK-arena.agent.md`); process.exit(1) }
     scenarioOpt.scenarioPath = taskPath
   } else {
     scenarioOpt.scenario = {
@@ -268,10 +276,8 @@ function parseArgs(argv: string[]) {
 
   const options: Record<string, string | undefined> = {
     task: undefined,
-    skills: undefined,
     decks: undefined,
     criteria: 'syntax,context,logic,token',
-    control: 'lythoskill-project-scribe',
     dir: 'tmp',
     project: '.',
     config: undefined,
@@ -284,13 +290,10 @@ function parseArgs(argv: string[]) {
     const arg = argv[i]
     if (arg === '--task' || arg === '-t') {
       options.task = argv[++i]
-    } else if (arg === '--skills' || arg === '-s') {
-      options.skills = argv[++i]
     } else if (arg === '--decks') {
       options.decks = argv[++i]
     } else if (arg === '--criteria' || arg === '-c') {
       options.criteria = argv[++i]
-    } else if (arg === '--control') {
       options.control = argv[++i]
     } else if (arg === '--dir' || arg === '-d') {
       options.dir = argv[++i]
@@ -319,39 +322,13 @@ export function runArena(argv: string[]) {
     process.exit(1)
   }
 
-  const HAS_DECKS = !!options.decks
-  const HAS_SKILLS = !!options.skills
+  const DECK_PATHS = (options.decks || '').split(',').map(s => s.trim()).filter(Boolean)
 
-  if (!HAS_DECKS && !HAS_SKILLS) {
-    console.error('❌ 请提供 --skills 或 --decks')
-    process.exit(1)
-  }
-  if (HAS_DECKS && HAS_SKILLS) {
-    console.error('❌ --skills 和 --decks 不能同时使用')
-    process.exit(1)
-  }
-
-  const DECK_PATHS = HAS_DECKS
-    ? (options.decks || '').split(',').map(s => s.trim()).filter(Boolean)
-    : []
-
-  const SKILLS = HAS_SKILLS
-    ? (options.skills || '').split(',').map(s => s.trim()).filter(Boolean)
-    : []
-
-  if (HAS_SKILLS && SKILLS.length < 2) {
-    console.error('❌ 至少需要 2 个 skill 才能进行 arena')
-    process.exit(1)
-  }
-  if (HAS_SKILLS && SKILLS.length > 5) {
-    console.error('❌ 一次 arena 最多 5 个 skill')
-    process.exit(1)
-  }
-  if (HAS_DECKS && DECK_PATHS.length < 2) {
+  if (DECK_PATHS.length < 2) {
     console.error('❌ 至少需要 2 个 deck 才能进行 arena')
     process.exit(1)
   }
-  if (HAS_DECKS && DECK_PATHS.length > 5) {
+  if (DECK_PATHS.length > 5) {
     console.error('❌ 一次 arena 最多 5 个 deck')
     process.exit(1)
   }
@@ -359,9 +336,6 @@ export function runArena(argv: string[]) {
   const CRITERIA = (options.criteria || 'syntax,context,logic,token')
     .split(',').map(s => s.trim()).filter(Boolean)
 
-  const CONTROL_SKILLS = (options.control || 'lythoskill-project-scribe')
-    .split(',').map(s => s.trim()).filter(Boolean)
-
   const PROJECT_DIR = resolve(options.project!)
   const ARENA_SLUG = slugify(TASK)
   const ARENA_ID = `arena-${timestamp()}-${ARENA_SLUG.slice(0, 30)}`
@@ -373,37 +347,20 @@ export function runArena(argv: string[]) {
   mkdirSync(join(ARENA_DIR, 'sides'), { recursive: true })
 
   // ── 生成参与者与 deck ───────────────────────────────────────
-  let participants: { id: string; name: string; skill_name: string; deck_path: string }[]
-  let mode: 'single-skill' | 'full-deck'
-
-  if (HAS_DECKS) {
-    mode = 'full-deck'
-    participants = DECK_PATHS.map((deckPath, i) => {
-      const id = `run-${String(i + 1).padStart(2, '0')}`
-      const name = basename(deckPath).replace(/\.toml$/, '')
-      const destPath = join(ARENA_DIR, 'decks', `arena-${id}.toml`)
-      // Copy the provided deck to arena directory
-      if (existsSync(deckPath)) {
-        const content = readFileSync(deckPath, 'utf-8')
-        writeFileSync(destPath, content)
-      } else {
-        console.error(`❌ Deck 文件不存在: ${deckPath}`)
-        process.exit(1)
-      }
-      return { id, name, skill_name: name, deck_path: destPath }
-    })
-  } else {
-    mode = 'single-skill'
-    participants = SKILLS.map((skill, i) => {
-      const id = `run-${String(i + 1).padStart(2, '0')}`
-      return {
-        id,
-        name: skill,
-        skill_name: skill,
-        deck_path: join(ARENA_DIR, 'decks', `arena-${id}.toml`),
-      }
-    })
-  }
+  const participants = DECK_PATHS.map((deckPath, i) => {
+    const id = `run-${String(i + 1).padStart(2, '0')}`
+    const name = basename(deckPath).replace(/\.toml$/, '')
+    const destPath = join(ARENA_DIR, 'decks', `arena-${id}.toml`)
+    // Copy the provided deck to arena directory
+    if (existsSync(deckPath)) {
+      const content = readFileSync(deckPath, 'utf-8')
+      writeFileSync(destPath, content)
+    } else {
+      console.error(`❌ Deck 文件不存在: ${deckPath}`)
+      process.exit(1)
+    }
+    return { id, name, skill_name: name, deck_path: destPath }
+  })
 
   const criteria = CRITERIA.map((c) => ({
     name: c,
@@ -411,29 +368,6 @@ export function runArena(argv: string[]) {
     weight: 1,
   }))
 
-  if (mode === 'single-skill') {
-    for (const p of participants) {
-      const deckContent = `# ============================================================
-# Arena Deck: ${p.id} — ${p.name}
-# ============================================================
-# 变量：${p.name}
-# 控制变量：${CONTROL_SKILLS.join(', ')}
-# ============================================================
-
-[deck]
-working_set = ".claude/skills"
-cold_pool   = "~/.agents/skill-repos"
-max_cards   = 10
-
-[tool]
-skills = [
-${[...new Set([p.skill_name, ...CONTROL_SKILLS])].map(s => `  "${s}",`).join('\n')}
-]
-`
-      writeFileSync(p.deck_path, deckContent)
-    }
-  }
-
   // ── 为每个 side 创建隔离工作空间 ────────────────────────────
   for (const p of participants) {
     const sideDir = join(ARENA_DIR, 'sides', p.id)
@@ -481,14 +415,11 @@ ${criteria.map(c => `  - ${c.label}`).join('\n')}
 arena_decks:
 ${participants.map(p => `  - ${p.deck_path.replace(PROJECT_DIR, '.')}`).join('\n')}
 judge_persona: |
-  ${mode === 'full-deck'
-    ? `你是一个多目标优化分析师。不要选 Winner。
-  对每个 deck 配置，按 evaluation_criteria 输出评分向量（1-5 分）。
-  识别 Pareto 非支配解集——没有"最强"，只有"在不同维度上的最优权衡"。
-  对被支配的解，说明它被谁支配、在哪个维度上劣势。
-  如果发现任何涌现 combo（多个 skill 组合产生 1+1>2 的效果），单独标注。`
-    : `你是一个中立的技能评测员。对比所有 subagent 的输出，
-  按 evaluation_criteria 给出 1-5 分评分，最终给出 Winner 和选型建议。`}
+    你是一个多目标优化分析师。不要选 Winner。
+    对每个 deck 配置，按 evaluation_criteria 输出评分向量（1-5 分）。
+    识别 Pareto 非支配解集——没有"最强"，只有"在不同维度上的最优权衡"。
+    对被支配的解，说明它被谁支配、在哪个维度上劣势。
+    如果发现任何涌现 combo（多个 skill 组合产生 1+1>2 的效果），单独标注。
 acceptance:
 ${participants.map(p => `  - Subagent ${p.id} 在 sides/${p.id}/ 隔离环境完成任务并写入 runs/${p.id}.md`).join('\n')}
   - Judge 读取所有 run 文件并生成 report.md
@@ -527,9 +458,9 @@ cd "${ARENA_DIR}"
 ID:        ${ARENA_ID}
 任务:      ${TASK}
 目录:      ${ARENA_DIR}
-模式:      ${mode === 'full-deck' ? '完整 deck 配置对比' : '单 skill 对比'}
+模式:      deck 配置对比
 参与者:    ${participants.map(p => p.name).join(', ')}
-${mode === 'single-skill' ? `控制变量:  ${CONTROL_SKILLS.join(', ')}\n` : ''}评测维度:  ${CRITERIA.join(', ')}
+评测维度:  ${CRITERIA.join(', ')}
 
 生成文件:
   📋 ${join(ARENA_DIR, 'arena.json')}
@@ -773,7 +704,7 @@ function runViz(argv: string[]) {
 
 // ── Run: programmatic arena execution ───────────────────────
 
-async function runProgrammaticArena(argv: string[]) {
+async function vsRun(argv: string[]) {
   const { options } = parseArgs(argv)
   const { readFileSync } = await import('node:fs')
 
@@ -811,13 +742,15 @@ async function runProgrammaticArena(argv: string[]) {
     return
   }
 
-  // CLI-flag mode (backward compat)
-  if (!options.task || !options.decks) {
-    console.error('❌ --task <path> and --decks <list> are required for "run" (or use --config <arena.toml>)')
-    process.exit(1)
-  }
-
-  const { runArena: runArenaProgrammatic } = await import('./runner')
+  // --config was not provided
+  console.error(`❌ --config <arena.toml> is required.
+   Usage: lythoskill-arena vs --config ./arena.toml
+          lythoskill-arena vs --config ./arena.toml --dry-run
+   Example configs:
+     examples/arena/research-compare/arena.toml   — two-side A/B
+     examples/arena/add-remove/arena.toml          — three-side Pareto
+   Create one: cp examples/arena/research-compare/arena.toml ./arena.toml`)
+  process.exit(1)
 
   const result = await runArenaProgrammatic({
     taskPath: options.task,
@@ -838,18 +771,20 @@ if (import.meta.main) {
   const args = process.argv.slice(2)
   const cmd = args[0]
 
-  if (cmd === 'agent-run') {
-    agentRun(args.slice(1))
+  if (cmd === 'single') {
+    singleRun(args.slice(1))
   } else if (cmd === 'viz') {
     runViz(args.slice(1))
-  } else if (cmd === 'run') {
-    runProgrammaticArena(args.slice(1))
+  } else if (cmd === 'vs') {
+    vsRun(args.slice(1))
   } else if (cmd === 'scaffold' || !cmd || args[0]?.startsWith('-')) {
     // Legacy behavior: if no subcommand or starts with flags, treat as scaffold
     runArena(cmd === 'scaffold' ? args.slice(1) : args)
   } else {
-    console.error(`❌ Unknown command: ${cmd}`)
-    printHelp()
+    console.error(`❌ Unknown command: "${cmd}"
+   Available: single, vs, scaffold, viz
+   Usage: lythoskill-arena <command> [options]
+   Help:  lythoskill-arena --help`)
     process.exit(1)
   }
 }