[{"_1":2,"_3":-5,"_4":-5},"loaderData",{"_5":6,"_7":8,"_9":10},"actionData","errors","root",{"_432":433,"_434":435},"routes/_app.x.$handle.$name",{"_101":100,"_32":46,"_110":91,"_27":111,"_13":14,"_112":-5,"_11":113,"_114":115,"_116":117,"_118":119,"_120":121,"_20":-5,"_122":-5,"_123":-5,"_124":25,"_125":126,"_127":60,"_128":-5,"_129":-5,"_130":131},"routes/_app.x.$handle.$name.branch.$branch.version.$version.diff",{"_11":12,"_13":14,"_15":16,"_17":-5,"_18":-5},"project",{"_44":45,"_32":46,"_39":47,"_48":-5,"_49":25,"_50":51,"_52":51,"_53":54,"_55":-5,"_18":56,"_57":58,"_59":60,"_61":-5,"_62":-5,"_63":64,"_65":66,"_67":68,"_69":70,"_71":72,"_73":-5,"_74":75,"_76":75,"_77":78,"_79":78,"_80":78,"_81":75,"_82":75,"_83":84,"_85":86,"_87":88,"_89":25},"version",116,"changes",[19],"aiChatInfo","userId",{"_20":21,"_22":23,"_24":25,"_26":25,"_27":28,"_29":28},"file",{"_30":31,"_32":33,"_34":35,"_36":38,"_39":40,"_41":14,"_42":43,"_24":25},"previous",{"_30":31,"_32":33,"_34":35,"_36":37},"deleted",false,"deletedHere","path","runner/config.ts","latestPath","fileId","26a1ba5e-8899-11f1-b320-1607ee4eb77e","name","config.ts","fileType","script","content","import type { ModelCfg } from \"../shared/types.ts\";\n\n// Short keys are what's stored in the DB; ids are OpenRouter model ids.\nexport const MODELS: Record = {\n // ---- free tier ($0) ----\n ultra: { id: \"nvidia/nemotron-3-ultra-550b-a55b:free\", label: \"Nemotron Ultra 550B\", maxTokens: 9000, reasoning: { effort: \"low\" } },\n super: { id: \"nvidia/nemotron-3-super-120b-a12b:free\", label: \"Nemotron Super 120B\", maxTokens: 9000, reasoning: { effort: \"low\" } },\n nano: { id: \"nvidia/nemotron-3-nano-30b-a3b:free\", label: \"Nemotron Nano 30B\", maxTokens: 9000, reasoning: { effort: \"low\" } },\n gptoss: { id: \"openai/gpt-oss-20b:free\", label: \"GPT-OSS 20B\", maxTokens: 9000, reasoning: { effort: \"low\" } },\n\n // ---- cheap paid tier ----\n haiku: { id: \"anthropic/claude-haiku-4.5\", label: \"Haiku 4.5\" },\n gpt5mini: { id: \"openai/gpt-5-mini\", label: \"GPT-5 mini\", reasoning: { effort: \"low\" } },\n flash: { id: \"google/gemini-2.5-flash\", label: \"Gemini 2.5 Flash\" },\n deepseek: { id: \"deepseek/deepseek-chat-v3.1\", label: \"DeepSeek V3.1\" },\n\n // ---- premium tier ----\n sonnet: { id: \"anthropic/claude-sonnet-4.5\", label: \"Sonnet 4.5\" },\n gpt51: { id: \"openai/gpt-5.1\", label: \"GPT-5.1\", reasoning: { effort: \"low\" } },\n\n // ---- flagship tier: best model per lab (July 2026) ----\n // Reasoning levels follow the convention used by Artificial Analysis / LMArena:\n // run each model at its PROVIDER-DEFAULT reasoning config. Claude is the one\n // exception that needs an explicit opt-in (Anthropic ships thinking OFF by\n // default), so Opus 5 gets effort medium to match the others' defaults.\n opus5: { id: \"anthropic/claude-opus-5\", label: \"Opus 5\", maxTokens: 16000, reasoning: { effort: \"medium\" } },\n sol: { id: \"openai/gpt-5.6-sol\", label: \"GPT-5.6 Sol\", maxTokens: 16000 },\n g3pro: { id: \"google/gemini-3-pro-image\", label: \"Gemini 3 Pro\", maxTokens: 16000 },\n kimik3: { id: \"moonshotai/kimi-k3\", label: \"Kimi K3\", maxTokens: 16000 },\n // hidden undercard (still playable, not featured)\n mimo: { id: \"xiaomi/mimo-v2.5-pro\", label: \"MiMo V2.5 Pro\", maxTokens: 16000 },\n glm52: { id: \"z-ai/glm-5.2\", label: \"GLM-5.2\", maxTokens: 16000 },\n dsv4: { id: \"deepseek/deepseek-v4-pro\", label: \"DeepSeek V4 Pro\", maxTokens: 16000 },\n grok45: { id: \"x-ai/grok-4.5\", label: \"Grok 4.5\" },\n qwenmax: { id: \"qwen/qwen3-max\", label: \"Qwen3-Max\" },\n};\n\n// Reference partners for the frontier bracket (cheap + proven from wave 1):\n// each frontier model F plays (F gives -> deepseek guesses) and (gpt5mini gives -> F guesses).\nconst FRONTIER_KEYS = [\"opus5\", \"sol\", \"g3pro\", \"kimik3\", \"dsv4\"];\nexport const REF_GUESSER = \"deepseek\";\nexport const REF_GIVER = \"gpt5mini\";\n\n// ---- seeding phase (edit + run runner/seed.ts) ----\nexport const SEED_MODE: \"duet\" | \"classic\" = \"classic\";\nexport const SEED_TIER = \"flagship-self-2\";\n// If SEED_PAIRS is non-empty, exactly these ordered [giver, guesser] pairs are\n// seeded and SEED_KEYS is ignored.\nexport const SEED_PAIRS: Array<[string, string]> = FRONTIER_KEYS.map((f) => [f, f] as [string, string]);\nexport const SEED_KEYS: string[] = [];\nexport const GAMES_PER_PAIRING = 3;\n","import type { ModelCfg } from \"../shared/types.ts\";\n\n// Short keys are what's stored in the DB; ids are OpenRouter model ids.\nexport const MODELS: Record = {\n // ---- free tier ($0) ----\n ultra: { id: \"nvidia/nemotron-3-ultra-550b-a55b:free\", label: \"Nemotron Ultra 550B\", maxTokens: 9000, reasoning: { effort: \"low\" } },\n super: { id: \"nvidia/nemotron-3-super-120b-a12b:free\", label: \"Nemotron Super 120B\", maxTokens: 9000, reasoning: { effort: \"low\" } },\n nano: { id: \"nvidia/nemotron-3-nano-30b-a3b:free\", label: \"Nemotron Nano 30B\", maxTokens: 9000, reasoning: { effort: \"low\" } },\n gptoss: { id: \"openai/gpt-oss-20b:free\", label: \"GPT-OSS 20B\", maxTokens: 9000, reasoning: { effort: \"low\" } },\n\n // ---- cheap paid tier ----\n haiku: { id: \"anthropic/claude-haiku-4.5\", label: \"Haiku 4.5\" },\n gpt5mini: { id: \"openai/gpt-5-mini\", label: \"GPT-5 mini\", reasoning: { effort: \"low\" } },\n flash: { id: \"google/gemini-2.5-flash\", label: \"Gemini 2.5 Flash\" },\n deepseek: { id: \"deepseek/deepseek-chat-v3.1\", label: \"DeepSeek V3.1\" },\n\n // ---- premium tier ----\n sonnet: { id: \"anthropic/claude-sonnet-4.5\", label: \"Sonnet 4.5\" },\n gpt51: { id: \"openai/gpt-5.1\", label: \"GPT-5.1\", reasoning: { effort: \"low\" } },\n\n // ---- flagship tier: best model per lab (July 2026) ----\n // Reasoning levels follow the convention used by Artificial Analysis / LMArena:\n // run each model at its PROVIDER-DEFAULT reasoning config. Claude is the one\n // exception that needs an explicit opt-in (Anthropic ships thinking OFF by\n // default), so Opus 5 gets effort medium to match the others' defaults.\n opus5: { id: \"anthropic/claude-opus-5\", label: \"Opus 5\", maxTokens: 16000, reasoning: { effort: \"medium\" } },\n sol: { id: \"openai/gpt-5.6-sol\", label: \"GPT-5.6 Sol\", maxTokens: 16000 },\n g3pro: { id: \"google/gemini-3-pro-image\", label: \"Gemini 3 Pro\", maxTokens: 16000 },\n kimik3: { id: \"moonshotai/kimi-k3\", label: \"Kimi K3\", maxTokens: 16000 },\n // hidden undercard (still playable, not featured)\n mimo: { id: \"xiaomi/mimo-v2.5-pro\", label: \"MiMo V2.5 Pro\", maxTokens: 16000 },\n glm52: { id: \"z-ai/glm-5.2\", label: \"GLM-5.2\", maxTokens: 16000 },\n dsv4: { id: \"deepseek/deepseek-v4-pro\", label: \"DeepSeek V4 Pro\", maxTokens: 16000 },\n grok45: { id: \"x-ai/grok-4.5\", label: \"Grok 4.5\" },\n qwenmax: { id: \"qwen/qwen3-max\", label: \"Qwen3-Max\" },\n};\n\n// Reference partners for the frontier bracket (cheap + proven from wave 1):\n// each frontier model F plays (F gives -> deepseek guesses) and (gpt5mini gives -> F guesses).\nconst FRONTIER_KEYS = [\"opus5\", \"sol\", \"g3pro\", \"kimik3\", \"dsv4\"];\nexport const REF_GUESSER = \"deepseek\";\nexport const REF_GIVER = \"gpt5mini\";\n\n// ---- seeding phase (edit + run runner/seed.ts) ----\nexport const SEED_MODE: \"duet\" | \"classic\" = \"classic\";\nexport const SEED_TIER = \"flagship-self-3\";\n// If SEED_PAIRS is non-empty, exactly these ordered [giver, guesser] pairs are\n// seeded and SEED_KEYS is ignored.\nexport const SEED_PAIRS: Array<[string, string]> = [[\"opus5\", \"opus5\"]];\nexport const SEED_KEYS: string[] = [];\nexport const GAMES_PER_PAIRING = 1;\n","createdAt",["D",1785191966906],"branchVersion","branchId","990afe23-5b2a-436e-8730-9608e52bcd7c","id","7626fd83-ae49-4526-b273-9775f4aa2717","codenames-duet-bench",["D",1785032481089],"deletedAt","legacyEnvVars","privacy","public","httpPrivacy","description","LLM teams play Codenames Duet — clue/guess benchmark + cost","imageUrl","9319c175-2593-4f88-ba21-04f2ef7b1c28","pinnedFileId","b100724a-8898-11f1-b6ec-1607ee4eb77e","pinPreview",true,"pinMigrated","screenshotVersion","likeCount",0,"orgId","d663a026-6449-42f6-9156-1d4a2d64d38c","tursoName","7626fd83ae494526b2739775f4aa2717","tursoHostname","7626fd83ae494526b2739775f4aa2717-val-town.aws-us-east-2.turso.io","tursoMeasuredAt",["D",1789474853273],"pinned","userPerm","r","valPerm","blobPerm","rw","emailPerm","sqlitePerm","projectPerm","telemetryPerm","user",{"_101":100,"_109":105},"org",{"_101":100,"_102":103,"_104":105,"_18":56,"_106":107,"_108":25},"branch",{"_13":90,"_32":91,"_44":43,"_92":45,"_93":-5,"_94":-5,"_95":-5,"_96":-5,"_97":-5,"_98":46,"_99":100},"isOwned",139,"main","projectId","forkedBranchId","forkedBranchVersion","forkedProjectPrivacy","forkedProjectId","forkedProjectOrgId","projectName","userHandle","dcm31","handle","displayName","Charlie","avatar","https://imagedelivery.net/iHX6Ovru0O7AjmyT5yZRoA/3e904faa-3764-4629-c475-f561a49b0000/square512?exp=1789570799&sig=bfafc9de3f25d020766c1a418088d2aa7a118e3408c10bafe80b0604dafaeba0","tier","pro","canRestrictVals","profileImageUrl","branchName","","adoptOrgId",{"_44":45,"_32":46,"_39":427,"_48":-5,"_49":25,"_50":51,"_52":51,"_53":54,"_55":-5,"_18":56,"_57":58,"_59":60,"_61":-5,"_62":-5,"_63":64,"_65":66,"_67":68,"_69":70,"_71":428,"_73":-5,"_74":75,"_76":75,"_77":78,"_79":78,"_80":78,"_81":75,"_82":75,"_83":429,"_85":430,"_87":431,"_89":25},"tree",[207,210,216,219,230,233,234,235],"files",[207,208,209,210,211,212,213,214,215,216,217,218,219,220,221,222,223,224,225,226,227,228,229,230,231,232,233,234,235],"counts",{"_197":64,"_198":146,"_199":64,"_200":201,"_202":203},"contents",[132,133,134,135,136,137,138,139],"upstreamProject","upstreamBranch","emptyEnvVars","flags",[],"hasHttpFiles","grants","orgs","customDomains",{},{"_44":191,"_32":192,"_30":193,"_142":-5,"_42":43,"_41":194,"_144":60,"_39":195,"_13":64,"_34":172,"_36":-5,"_24":25,"_149":-5,"_27":196},{"_44":186,"_32":187,"_30":188,"_142":-5,"_42":43,"_41":164,"_144":60,"_39":189,"_13":64,"_34":172,"_36":-5,"_24":25,"_149":-5,"_27":190},{"_44":180,"_32":181,"_30":182,"_142":-5,"_42":43,"_41":183,"_144":60,"_39":184,"_13":64,"_34":172,"_36":-5,"_24":25,"_149":-5,"_27":185},{"_44":174,"_32":175,"_30":176,"_142":-5,"_42":43,"_41":177,"_144":60,"_39":178,"_13":64,"_34":172,"_36":-5,"_24":25,"_149":-5,"_27":179},{"_44":167,"_32":168,"_30":169,"_142":-5,"_42":43,"_41":170,"_144":60,"_39":171,"_13":64,"_34":172,"_36":-5,"_24":25,"_149":-5,"_27":173},{"_44":159,"_32":160,"_30":161,"_142":-5,"_42":43,"_41":162,"_144":25,"_39":163,"_13":164,"_34":20,"_36":165,"_24":25,"_149":-5,"_27":166},{"_44":151,"_32":152,"_30":153,"_142":-5,"_42":43,"_41":154,"_144":60,"_39":155,"_13":156,"_34":20,"_36":157,"_24":25,"_149":-5,"_27":158},{"_44":140,"_32":141,"_30":58,"_142":-5,"_42":43,"_41":143,"_144":60,"_39":145,"_13":146,"_34":147,"_36":148,"_24":25,"_149":-5,"_27":150},"8ca0ed94-ed03-445e-aec4-f018a5d2d933","main.ts","parentId",29,"latest",["D",1785033208321],1,"http","import { Hono } from \"npm:hono\";\nimport { parseVal, serveFile } from \"https://esm.town/v/std/utils/index.ts\";\nimport { init, sqlite } from \"./database/db.ts\";\nimport { standings } from \"./database/standings.ts\";\n\nconst app = new Hono();\n\napp.get(\"/\", (c) => serveFile(\"/frontend/index.html\"));\napp.get(\"/frontend/**/*\", (c) => serveFile(c.req.path));\napp.get(\"/source\", (c) => c.redirect(parseVal().links.self.val));\n\napp.get(\"/api/standings\", async (c) => c.json(await standings()));\n\napp.get(\"/api/game/:id\", async (c) => {\n await init();\n const id = Number(c.req.param(\"id\"));\n const r = await sqlite.execute({\n sql: `SELECT id, model_a, model_b, label_a, label_b, tier, seed, status, result, agents_found, turns_used, cost, calls, tokens_in, tokens_out, created_at, finished_at, state\n FROM games WHERE id = ?`,\n args: [id],\n });\n if (!r.rows.length) return c.json({ error: \"not found\" }, 404);\n const row: any = r.rows[0];\n return c.json({ ...row, state: JSON.parse(String(row.state)) });\n});\n\napp.onError((err) => Promise.reject(err));\n\nexport default app.fetch;\n","label",[141],"c2eb9e08-f073-42c1-a5cb-9d103ae14a5e","README.md","e8a15dca-889a-11f1-b6f7-1607ee4eb77e",87,["D",1785162556690],3,"# codenames-duet-bench\n\nAI models play **cooperative Codenames** against the deck, in two modes:\n\n- **classic** (primary): ONE key. Model A is the spymaster (sees the key, gives\n clues); model B guesses blind with up to N+1 guesses per clue. 9 agents,\n 3 assassins, 9 clues max. Every ordered (giver, guesser) pairing is a cell in\n the standings matrix — including a model cluing for itself.\n- **duet**: faithful Codenames Duet — two overlapping keys, 15 shared agents,\n guesses judged on the giver's key, both partners clue.\n\nEvery clue, guess, prompted \"thinking\", native reasoning trace, and per-call\ncost is recorded.\n\nLive UI: `main.ts` endpoint (standings → tap a team → tap a game → full replay\nwith collapsed chain-of-thought). Wears the house style (quatro direction),\nimporting `pen.tsx`/`directions.ts` from\n[dcm31/house-style](https://www.val.town/x/dcm31/house-style).\n\n## Rules implemented (faithful Duet)\n\n- 25 words from the official 400-word list; official Duet key card overlap\n (9 green / 3 assassin / 13 neutral per side; 3 shared greens; 15 unique agents;\n 1 double assassin; 1 green-for-you-assassin-for-them each way).\n- 9 total clue turns shared by the team. Givers alternate, except a giver with\n no remaining greens hands the pen over (Duet allows clues in any order).\n- Guesses judged against the **giver's** key. Green → keep guessing (or stop);\n neutral → turn ends, word marked against that key; assassin → instant loss.\n- Simplifications: no sudden-death phase; 8-guess safety cap per turn; a model\n that can't produce a valid clue/guess in 2 attempts forfeits that action.\n\n## Scoring\n\n**Guess Efficiency Score (GES)**, 0-100%: raw = successful guesses (0-9), plus\na win bonus of (10 - clues used); GES = raw / 18. Wins always outrank losses\n(win floor 56% vs loss ceiling 44%); the perfect game — all 9 words on the\nfirst clue — is 100%. Leaderboard shows each model's average GES combined and\nsplit by role. Per-game data also keeps win/loss, words found, clues used,\nfull transcript, and OpenRouter usage cost.\n\n## Architecture\n\n```mermaid\nflowchart LR\n seed[runner/seed.ts] -->|queued games| db[(sqlite: games)]\n cron[runner/cron.ts every 1 min] --> drain[runner/drain.ts]\n drain -->|playTurn| engine[engine/play.ts + prompts + llm]\n engine -->|OpenRouter| or[(openrouter.ai)]\n drain -->|state per turn| db\n ui[main.ts + frontend/] -->|/api/standings, /api/game/:id| db\n```\n\n## Running more games\n\n1. Edit `runner/config.ts`: pick `SEED_MODE` (\"classic\" or \"duet\"), `SEED_KEYS`\n (model keys), `SEED_TIER` (a fresh tier name — seeding is guarded per tier),\n `GAMES_PER_PAIRING`. Classic seeds ORDERED pairs (model_a gives, model_b\n guesses); duet seeds unordered pairs.\n2. Run `runner/seed.ts`. The cron drains the queue automatically (~1 game/2–4 min,\n 3 concurrent). `runner/run-next.ts` advances it manually.\n3. When done benchmarking, pause the worker: set `runner/cron.ts` back to a long\n interval or change its file type to `script` (it's a cheap no-op when idle,\n but still).\n\n`runner/validate-models.ts` checks ids + pricing; `runner/check-credits.ts`\nshows key limits/balance; `runner/test-engine.ts` is the offline engine test.\n\n## Tiers\n\n- **free** — `:free` OpenRouter models ($0): Nemotron 550B/120B/30B + GPT-OSS 20B.\n- **cheap** / **premium** — configured (Haiku 4.5, GPT-5 mini, Gemini 2.5 Flash,\n DeepSeek V3.1 / Sonnet 4.5, GPT-5.1) but need OpenRouter credits topped up.\n\nEnv: `OPENROUTER_API_KEY`.\n\nWord list note: the 400 base words as shipped by open-source Codenames clones —\nlists of individual common words aren't subject to copyright; used here for a\npersonal benchmark.\n",[152],"582647a2-8d25-41fd-95bf-0df2ee38d22e","AGENTS.md","efc304fa-889a-11f1-95a2-1607ee4eb77e",97,["D",1785162785500],4,"# AGENTS.md — orientation for codenames-duet-bench\n\nRead this before touching anything. Update it (and README.md) in the same\nchange-set as any behavior change.\n\n## What this is\n\nAn LLM benchmark where models play cooperative Codenames. The headline mode is\n**classic**: one model is the spymaster (sees the single secret key, gives\none-word clues), another guesses completely blind. 9 agents, 3 assassins,\n13 neutrals, 9 clues max, N+1 guesses per clue (unlimited on a 0-clue).\nGuesses judged against the key; assassin = instant loss. A legacy **duet**\nmode (two overlapping keys, 15 shared agents, alternating givers) still runs\nfor old games and is tucked away in the UI.\n\nScores: **Guess Efficiency Score (GES)**, a 0-100%: raw = successful guesses\n(0-9) + on a WIN a bonus of (10 - clues used); pct = raw/18. You must win to\nbeat any loss (win floor 10/18=56% > loss ceiling 8/18=44%); best possible =\nall 9 words on clue 1 = 100%. Defined in database/standings.ts (gesPct) and\nmirrored in frontend/index.tsx — change both together. COPY RULE from Charlie:\nthe word \"agent\" never appears in the UI (confusing next to AI agents); say\n\"words\" / \"successful guesses\". Charlie dictates all site copy — do not invent\nprose on the page without him. Palette = the physical Codenames game (cream\ncards, blue + red teams, tan bystanders, black assassin); Fullerfont only for\nthe title, a couple of section headings, and score numbers. The site leads\nwith a vertical grouped bar chart (combined/giver/guesser GES), then ONE\nmerged giver x guesser matrix with big CLUE GIVERS / GUESSERS axis headings,\nthen a Game logs list; every game has the replay scrubber.\n\n## Benchmark protocol\n\n- Full round-robins within a price tier (free, cheap): every ordered pair\n incl. self-pairs.\n- **Frontier models get 2 reference games each** instead of a full matrix\n (cost control): (frontier gives -> `deepseek` guesses) and (`gpt5mini`\n gives -> frontier guesses). References were chosen from cheap-tier wave 1:\n gpt5mini was the strongest giver, deepseek the strongest guesser.\n Keep references FIXED so frontier rows/columns stay comparable.\n- Same prompts/engine for every classic game ever run — results merge safely.\n\n## Architecture\n\n```\nrunner/seed.ts -> INSERT queued games (guarded per SEED_TIER)\nrunner/cron.ts -> 1-min interval, calls drain()\nrunner/drain.ts -> claims games (atomic, heartbeat), plays turns, saves per turn\nengine/play.ts -> one turn: clue call, validate, guess loop, judge\nengine/prompts.ts -> mode-aware prompt builders (classic guesser sees NO key)\nengine/game.ts -> pure rules: keys, judging, turn bookkeeping (no I/O)\nengine/llm.ts -> OpenRouter chat + usage-cost accounting + parseOut\ndatabase/db.ts -> sqlite schema, claim/saveState/finalize (status-guarded)\ndatabase/standings.ts -> bars + merged matrix + duet legacy aggregation\nmain.ts + frontend/ -> Hono + React UI (house style from dcm31/house-style)\n```\n\n## Key invariants\n\n- `engine/game.ts` stays pure; ALL LLM I/O lives in play.ts/llm.ts.\n- `GameState.mode` switches rules; **absent mode = duet** (old rows).\n- Classic: giver is ALWAYS side \"a\" (model_a); guesser side \"b\" (model_b).\n Ordered pairs matter. Duet: unordered, roles alternate.\n- Guesses judged on the giver's key; `marks[side]` = revealed-neutral vs that\n key; marked-vs-giver words are illegal on that giver's clues.\n- State JSON saves to the row after EVERY turn — games resume anywhere.\n- `saveState`/`finalize` carry `WHERE status != 'done'` so zombie workers can\n never resurrect finished games. `claim()` is the concurrency lock\n (heartbeat via claimed_at; 150s staleness window).\n- **Spend cap**: drain() refuses paid games once SUM(cost) >= $4.50. Raise it\n there, deliberately, if the budget changes.\n- Seeding is guarded per tier name; to re-run a tier, delete its rows first.\n- Every game is reproducible from its seed (mulberry32).\n\n## Workflows\n\n**Add a model**: add to MODELS in runner/config.ts (short key = DB identity —\nnever reuse or rename existing keys), run runner/validate-models.ts (checks\nthe id against the live catalog + pings), then seed.\n\n**Seed games**: edit SEED_MODE / SEED_TIER (fresh name) / SEED_KEYS (round\nrobin) or SEED_PAIRS (exact ordered [giver, guesser] pairs — wins if\nnon-empty) / GAMES_PER_PAIRING, run runner/seed.ts. The cron drains\nautomatically, paid + classic first; free-tier games are slow (10-60s/call).\n\n**Watch**: the UI live-polls, or\n`SELECT tier, COUNT(*), SUM(status='done'), SUM(cost) FROM games GROUP BY tier`.\n\n**Pause the worker**: set runner/cron.ts to a long interval (or file type\nscript) when no games are pending for a long stretch.\n\n## Gotchas learned the hard way\n\n- Both prompts contain the phrase \"CLUE GIVER\"; discriminate roles with\n \"You are the CLUE GIVER\" / \"You are the GUESSER\".\n- Reasoning effort is a CONFOUND, not just a cost knob: the original frontier\n runs at `effort: \"low\"` made Opus 5 look broken and were deleted (2026-07-27).\n Flagship-tier models run `effort: \"medium\"` with maxTokens 16000 and a 140s\n call timeout (claim staleness 330s to match). Free Nemotrons stay at low —\n they're not the story. Never compare models run at different efforts.\n- The OpenRouter ACCOUNT balance gates paid models, not the key's monthly\n limit; runner/check-credits.ts shows both.\n- Long executions outlive MCP run_file timeouts — that's what the cron is\n for. A single playTurn can also outlive the drain's budget (budget is only\n checked between turns); the claim lock + status guards make that safe.\n- `google/gemini-3-pro` does not exist on OpenRouter; the pro-class Gemini is\n `google/gemini-3-pro-image` (labeled \"Gemini 3 Pro\" here).\n- Val Town sqlite json_each: alias the tables or `id` becomes ambiguous.\n- Harness was audited (2026-07-26): zero invalid clues / parse failures /\n forfeits across paid games; losses are genuine model strategy errors\n (theme-chasing past the clue number is the classic death).\n",[160],"1d539134-dcb0-4e9b-b1d0-b3dafacda51d","shared","34de8914-dae5-46f9-b7b9-be34374fb52f",2,["D",1785032563573],"directory",[168],"3228d301-eb4a-45f7-be1c-50cd162e1062","runner","8c3a8801-dc3e-46f9-95a3-385b263d40e7",10,["D",1785032678458],[175],"8b50bdaa-1279-40cd-842d-97870c5af880","frontend","ff5b8dc1-982a-483c-8f65-d7ef283ef11b",30,["D",1785033217341],[181],"828c061d-299c-48b6-b1a0-22b49a64db31","engine","1e68fd3d-a5f7-4b7c-9318-d0ce6c31cbf6",["D",1785032587913],[187],"84f6a046-2087-4156-8106-c54fe9cbe3e2","database","639da78d-d168-4be4-9148-d62cf8d403a4",9,["D",1785032670439],[192],"pulls","branches","forks","envVars",{"_204":146,"_206":64},"connections",{"_204":64,"_205":64},"count","unfilledCount","emptyCount",{"_44":191,"_32":192,"_30":193,"_142":-5,"_42":43,"_41":194,"_144":60,"_39":424,"_13":64,"_34":172,"_36":-5,"_24":25,"_149":-5,"_237":425,"_27":426},{"_44":416,"_32":417,"_30":418,"_142":193,"_42":43,"_41":419,"_144":60,"_39":420,"_13":156,"_34":35,"_36":421,"_24":25,"_149":-5,"_237":422,"_27":423},{"_44":407,"_32":408,"_30":409,"_142":193,"_42":43,"_41":410,"_144":25,"_39":411,"_13":412,"_34":35,"_36":413,"_24":25,"_149":-5,"_237":414,"_27":415},{"_44":186,"_32":187,"_30":188,"_142":-5,"_42":43,"_41":164,"_144":60,"_39":404,"_13":64,"_34":172,"_36":-5,"_24":25,"_149":-5,"_237":405,"_27":406},{"_44":396,"_32":397,"_30":398,"_142":188,"_42":43,"_41":399,"_144":60,"_39":400,"_13":146,"_34":35,"_36":401,"_24":25,"_149":-5,"_237":402,"_27":403},{"_44":389,"_32":390,"_30":391,"_142":188,"_42":43,"_41":164,"_144":60,"_39":392,"_13":64,"_34":35,"_36":393,"_24":25,"_149":-5,"_237":394,"_27":395},{"_44":381,"_32":382,"_30":383,"_142":188,"_42":43,"_41":384,"_144":60,"_39":385,"_13":164,"_34":35,"_36":386,"_24":25,"_149":-5,"_237":387,"_27":388},{"_44":372,"_32":373,"_30":374,"_142":188,"_42":43,"_41":375,"_144":60,"_39":376,"_13":377,"_34":35,"_36":378,"_24":25,"_149":-5,"_237":379,"_27":380},{"_44":364,"_32":365,"_30":366,"_142":188,"_42":43,"_41":367,"_144":60,"_39":368,"_13":156,"_34":35,"_36":369,"_24":25,"_149":-5,"_237":370,"_27":371},{"_44":180,"_32":181,"_30":182,"_142":-5,"_42":43,"_41":183,"_144":60,"_39":361,"_13":64,"_34":172,"_36":-5,"_24":25,"_149":-5,"_237":362,"_27":363},{"_44":354,"_32":355,"_30":356,"_142":182,"_42":43,"_41":183,"_144":60,"_39":357,"_13":64,"_34":20,"_36":358,"_24":25,"_149":-5,"_237":359,"_27":360},{"_44":345,"_32":346,"_30":347,"_142":182,"_42":43,"_41":348,"_144":25,"_39":349,"_13":350,"_34":35,"_36":351,"_24":25,"_149":-5,"_237":352,"_27":353},{"_44":174,"_32":175,"_30":176,"_142":-5,"_42":43,"_41":177,"_144":60,"_39":342,"_13":64,"_34":172,"_36":-5,"_24":25,"_149":-5,"_237":343,"_27":344},{"_44":334,"_32":335,"_30":336,"_142":176,"_42":43,"_41":337,"_144":60,"_39":338,"_13":64,"_34":35,"_36":339,"_24":25,"_149":-5,"_237":340,"_27":341},{"_44":330,"_32":33,"_30":31,"_142":176,"_42":43,"_41":14,"_144":60,"_39":331,"_13":267,"_34":35,"_36":38,"_24":25,"_149":-5,"_237":332,"_27":333},{"_44":321,"_32":322,"_30":323,"_142":176,"_42":43,"_41":324,"_144":60,"_39":325,"_13":64,"_34":326,"_36":327,"_24":25,"_149":-5,"_237":328,"_27":329},{"_44":312,"_32":313,"_30":314,"_142":176,"_42":43,"_41":315,"_144":60,"_39":316,"_13":317,"_34":35,"_36":318,"_24":25,"_149":-5,"_237":319,"_27":320},{"_44":304,"_32":305,"_30":306,"_142":176,"_42":43,"_41":307,"_144":60,"_39":308,"_13":146,"_34":35,"_36":309,"_24":25,"_149":-5,"_237":310,"_27":311},{"_44":296,"_32":297,"_30":298,"_142":176,"_42":43,"_41":299,"_144":60,"_39":300,"_13":156,"_34":35,"_36":301,"_24":25,"_149":-5,"_237":302,"_27":303},{"_44":288,"_32":289,"_30":290,"_142":176,"_42":43,"_41":291,"_144":60,"_39":292,"_13":146,"_34":35,"_36":293,"_24":25,"_149":-5,"_237":294,"_27":295},{"_44":280,"_32":281,"_30":282,"_142":176,"_42":43,"_41":283,"_144":60,"_39":284,"_13":156,"_34":35,"_36":285,"_24":25,"_149":-5,"_237":286,"_27":287},{"_44":272,"_32":273,"_30":274,"_142":176,"_42":43,"_41":275,"_144":60,"_39":276,"_13":146,"_34":35,"_36":277,"_24":25,"_149":-5,"_237":278,"_27":279},{"_44":264,"_32":265,"_30":266,"_142":176,"_42":43,"_41":267,"_144":60,"_39":268,"_13":64,"_34":35,"_36":269,"_24":25,"_149":-5,"_237":270,"_27":271},{"_44":167,"_32":168,"_30":169,"_142":-5,"_42":43,"_41":170,"_144":60,"_39":261,"_13":64,"_34":172,"_36":-5,"_24":25,"_149":-5,"_237":262,"_27":263},{"_44":253,"_32":254,"_30":255,"_142":169,"_42":43,"_41":256,"_144":60,"_39":257,"_13":146,"_34":35,"_36":258,"_24":25,"_149":-5,"_237":259,"_27":260},{"_44":246,"_32":247,"_30":248,"_142":169,"_42":43,"_41":170,"_144":60,"_39":249,"_13":64,"_34":35,"_36":250,"_24":25,"_149":-5,"_237":251,"_27":252},{"_44":159,"_32":160,"_30":161,"_142":-5,"_42":43,"_41":162,"_144":25,"_39":243,"_13":164,"_34":20,"_36":165,"_24":25,"_149":-5,"_237":244,"_27":245},{"_44":151,"_32":152,"_30":153,"_142":-5,"_42":43,"_41":154,"_144":60,"_39":240,"_13":156,"_34":20,"_36":157,"_24":25,"_149":-5,"_237":241,"_27":242},{"_44":140,"_32":141,"_30":58,"_142":-5,"_42":43,"_41":143,"_144":60,"_39":236,"_13":146,"_34":147,"_36":148,"_24":25,"_149":-5,"_237":238,"_27":239},["D",1785033208321],"children",[],[141],["D",1785162556690],[],[152],["D",1785162785500],[],[160],"b522bc61-40e2-4e9c-bbb2-49f62af4fb7e","words.ts","e229c2c2-8898-11f1-884c-1607ee4eb77e",["D",1785032563573],"// The 400-word Codenames word list (as shipped in open-source clones like\n// jbowens/codenames). Codenames Duet uses the same base list. Individual\n// common words aren't copyrightable; used here for a personal benchmark.\nexport const WORDS: string[] = [\n \"AFRICA\", \"AGENT\", \"AIR\", \"ALIEN\", \"ALPS\", \"AMAZON\", \"AMBULANCE\", \"AMERICA\", \"ANGEL\", \"ANTARCTICA\",\n \"APPLE\", \"ARM\", \"ATLANTIS\", \"AUSTRALIA\", \"AZTEC\", \"BACK\", \"BALL\", \"BAND\", \"BANK\", \"BAR\",\n \"BARK\", \"BAT\", \"BATTERY\", \"BEACH\", \"BEAR\", \"BEAT\", \"BED\", \"BEIJING\", \"BELL\", \"BELT\",\n \"BERLIN\", \"BERMUDA\", \"BERRY\", \"BILL\", \"BLOCK\", \"BOARD\", \"BOLT\", \"BOMB\", \"BOND\", \"BOOM\",\n \"BOOT\", \"BOTTLE\", \"BOW\", \"BOX\", \"BRIDGE\", \"BRUSH\", \"BUCK\", \"BUFFALO\", \"BUG\", \"BUGLE\",\n \"BUTTON\", \"CALF\", \"CANADA\", \"CAP\", \"CAPITAL\", \"CAR\", \"CARD\", \"CARROT\", \"CASINO\", \"CAST\",\n \"CAT\", \"CELL\", \"CENTAUR\", \"CENTER\", \"CHAIR\", \"CHANGE\", \"CHARGE\", \"CHECK\", \"CHEST\", \"CHICK\",\n \"CHINA\", \"CHOCOLATE\", \"CHURCH\", \"CIRCLE\", \"CLIFF\", \"CLOAK\", \"CLUB\", \"CODE\", \"COLD\", \"COMIC\",\n \"COMPOUND\", \"CONCERT\", \"CONDUCTOR\", \"CONTRACT\", \"COOK\", \"COPPER\", \"COTTON\", \"COURT\", \"COVER\", \"CRANE\",\n \"CRASH\", \"CRICKET\", \"CROSS\", \"CROWN\", \"CYCLE\", \"CZECH\", \"DANCE\", \"DATE\", \"DAY\", \"DEATH\",\n \"DECK\", \"DEGREE\", \"DIAMOND\", \"DICE\", \"DINOSAUR\", \"DISEASE\", \"DOCTOR\", \"DOG\", \"DRAFT\", \"DRAGON\",\n \"DRESS\", \"DRILL\", \"DROP\", \"DUCK\", \"DWARF\", \"EAGLE\", \"EGYPT\", \"EMBASSY\", \"ENGINE\", \"ENGLAND\",\n \"EUROPE\", \"EYE\", \"FACE\", \"FAIR\", \"FALL\", \"FAN\", \"FENCE\", \"FIELD\", \"FIGHTER\", \"FIGURE\",\n \"FILE\", \"FILM\", \"FIRE\", \"FISH\", \"FLUTE\", \"FLY\", \"FOOT\", \"FORCE\", \"FOREST\", \"FORK\",\n \"FRANCE\", \"GAME\", \"GAS\", \"GENIUS\", \"GERMANY\", \"GHOST\", \"GIANT\", \"GLASS\", \"GLOVE\", \"GOLD\",\n \"GRACE\", \"GRASS\", \"GREECE\", \"GREEN\", \"GROUND\", \"HAM\", \"HAND\", \"HAWK\", \"HEAD\", \"HEART\",\n \"HELICOPTER\", \"HIMALAYAS\", \"HOLE\", \"HOLLYWOOD\", \"HONEY\", \"HOOD\", \"HOOK\", \"HORN\", \"HORSE\", \"HORSESHOE\",\n \"HOSPITAL\", \"HOTEL\", \"ICE\", \"ICE CREAM\", \"INDIA\", \"IRON\", \"IVORY\", \"JACK\", \"JAM\", \"JET\",\n \"JUPITER\", \"KANGAROO\", \"KETCHUP\", \"KEY\", \"KID\", \"KING\", \"KIWI\", \"KNIFE\", \"KNIGHT\", \"LAB\",\n \"LAP\", \"LASER\", \"LAWYER\", \"LEAD\", \"LEMON\", \"LEPRECHAUN\", \"LIFE\", \"LIGHT\", \"LIMOUSINE\", \"LINE\",\n \"LINK\", \"LION\", \"LITTER\", \"LOCH NESS\", \"LOCK\", \"LOG\", \"LONDON\", \"LUCK\", \"MAIL\", \"MAMMOTH\",\n \"MAPLE\", \"MARBLE\", \"MARCH\", \"MASS\", \"MATCH\", \"MERCURY\", \"MEXICO\", \"MICROSCOPE\", \"MILLIONAIRE\", \"MINE\",\n \"MINT\", \"MISSILE\", \"MODEL\", \"MOLE\", \"MOON\", \"MOSCOW\", \"MOUNT\", \"MOUSE\", \"MOUTH\", \"MUG\",\n \"NAIL\", \"NEEDLE\", \"NET\", \"NEW YORK\", \"NIGHT\", \"NINJA\", \"NOTE\", \"NOVEL\", \"NURSE\", \"NUT\",\n \"OCTOPUS\", \"OIL\", \"OLIVE\", \"OLYMPUS\", \"OPERA\", \"ORANGE\", \"ORGAN\", \"PALM\", \"PAN\", \"PANTS\",\n \"PAPER\", \"PARACHUTE\", \"PARK\", \"PART\", \"PASS\", \"PASTE\", \"PENGUIN\", \"PHOENIX\", \"PIANO\", \"PIE\",\n \"PILOT\", \"PIN\", \"PIPE\", \"PIRATE\", \"PISTOL\", \"PIT\", \"PITCH\", \"PLANE\", \"PLASTIC\", \"PLATE\",\n \"PLATYPUS\", \"PLAY\", \"PLOT\", \"POINT\", \"POISON\", \"POLE\", \"POLICE\", \"POOL\", \"PORT\", \"POST\",\n \"POUND\", \"PRESS\", \"PRINCESS\", \"PUMPKIN\", \"PUPIL\", \"PYRAMID\", \"QUEEN\", \"RABBIT\", \"RACKET\", \"RAY\",\n \"REVOLUTION\", \"RING\", \"ROBIN\", \"ROBOT\", \"ROCK\", \"ROME\", \"ROOT\", \"ROSE\", \"ROULETTE\", \"ROUND\",\n \"ROW\", \"RULER\", \"SATELLITE\", \"SATURN\", \"SCALE\", \"SCHOOL\", \"SCIENTIST\", \"SCORPION\", \"SCREEN\", \"SCUBA DIVER\",\n \"SEAL\", \"SERVER\", \"SHADOW\", \"SHAKESPEARE\", \"SHARK\", \"SHIP\", \"SHOE\", \"SHOP\", \"SHOT\", \"SINK\",\n \"SKYSCRAPER\", \"SLIP\", \"SLUG\", \"SMUGGLER\", \"SNOW\", \"SNOWMAN\", \"SOCK\", \"SOLDIER\", \"SOUL\", \"SOUND\",\n \"SPACE\", \"SPELL\", \"SPIDER\", \"SPIKE\", \"SPINE\", \"SPOT\", \"SPRING\", \"SPY\", \"SQUARE\", \"STADIUM\",\n \"STAFF\", \"STAR\", \"STATE\", \"STICK\", \"STOCK\", \"STRAW\", \"STREAM\", \"STRIKE\", \"STRING\", \"SUB\",\n \"SUIT\", \"SUPERHERO\", \"SWING\", \"SWITCH\", \"TABLE\", \"TABLET\", \"TAG\", \"TAIL\", \"TAP\", \"TEACHER\",\n \"TELESCOPE\", \"TEMPLE\", \"THEATER\", \"THIEF\", \"THUMB\", \"TICK\", \"TIE\", \"TIME\", \"TOKYO\", \"TOOTH\",\n \"TORCH\", \"TOWER\", \"TRACK\", \"TRAIN\", \"TRIANGLE\", \"TRIP\", \"TRUNK\", \"TUBE\", \"TURKEY\", \"UNDERTAKER\",\n \"UNICORN\", \"VACUUM\", \"VAN\", \"VET\", \"WAKE\", \"WALL\", \"WAR\", \"WASHER\", \"WASHINGTON\", \"WATCH\",\n \"WATER\", \"WAVE\", \"WEB\", \"WELL\", \"WHALE\", \"WHIP\", \"WIND\", \"WITCH\", \"WORM\", \"YARD\",\n];\n",[],[168,247],"63dd3aaf-4a9d-4f71-984b-aa62c17e9de0","types.ts","e78e0c78-8898-11f1-a225-1607ee4eb77e",47,["D",1785039003520],"// Shared types for the Codenames Duet benchmark.\n\nexport type Cell = \"green\" | \"bystander\" | \"assassin\";\n\n/** Duet key card: 25 cells, each with a value per side. */\nexport interface KeyCard {\n a: Cell[];\n b: Cell[];\n}\n\nexport type Side = \"a\" | \"b\";\n\nexport interface ModelCfg {\n /** OpenRouter model id, e.g. \"anthropic/claude-haiku-4.5\" */\n id: string;\n label: string;\n /** OpenRouter unified reasoning param, e.g. { effort: \"low\" } */\n reasoning?: Record;\n maxTokens?: number;\n}\n\nexport interface LlmOut {\n text: string;\n reasoning: string | null;\n cost: number;\n tokensIn: number;\n tokensOut: number;\n}\n\nexport interface GuessRec {\n word: string;\n outcome: \"agent\" | \"bystander\" | \"assassin\";\n thinking?: string;\n reasoning?: string | null;\n cost: number;\n}\n\nexport interface TurnRec {\n turn: number;\n giver: Side;\n clue?: {\n word: string;\n number: number;\n thinking?: string;\n reasoning?: string | null;\n cost: number;\n };\n guesses: GuessRec[];\n stopped?: boolean;\n invalid?: string;\n error?: string;\n}\n\nexport type GameMode = \"duet\" | \"classic\";\n\nexport interface GameState {\n mode?: GameMode; // absent = \"duet\" (older games)\n seed: number;\n words: string[]; // 25\n key: KeyCard;\n covered: boolean[]; // found agents (global)\n /** marks.a[i] = word i was revealed as a bystander when judged against side A's key */\n marks: { a: boolean[]; b: boolean[] };\n turn: number; // 1..9\n nextGiver: Side;\n status: \"playing\" | \"won\" | \"lost_assassin\" | \"lost_timeout\" | \"error\";\n agentsFound: number;\n transcript: TurnRec[];\n cost: number;\n calls: number;\n tokensIn: number;\n tokensOut: number;\n errorMsg?: string;\n}\n",[],[168,254],["D",1785032563573],[231,232],[168],"2c991c99-175a-4a38-b0d4-f53c5fa39233","validate-models.ts","99efc64a-8899-11f1-bfb5-1607ee4eb77e",18,["D",1785032871893],"// Checks configured model ids against OpenRouter's live catalog + does one\n// tiny authed completion to confirm the API key works.\nimport { MODELS } from \"./config.ts\";\nimport { chat } from \"../engine/llm.ts\";\n\nconst r = await fetch(\"https://openrouter.ai/api/v1/models\");\nconst j = await r.json();\nconst byId = new Map((j.data ?? []).map((m: any) => [m.id, m]));\n\nfor (const [key, cfg] of Object.entries(MODELS)) {\n const m = byId.get(cfg.id);\n if (!m) {\n console.log(`MISSING ${key}: ${cfg.id}`);\n const stem = cfg.id.split(\"/\")[1]?.split(\"-\").slice(0, 2).join(\"-\") ?? cfg.id;\n const near = [...byId.keys()].filter((id) => id.includes(stem)).slice(0, 6);\n console.log(` nearby: ${near.join(\", \") || \"(none)\"}`);\n } else {\n const p = m.pricing ?? {};\n console.log(`OK ${key}: ${cfg.id} $${(Number(p.prompt) * 1e6).toFixed(2)}/M in, $${(Number(p.completion) * 1e6).toFixed(2)}/M out`);\n }\n}\n\nconst ping = await chat(MODELS.haiku, 'Reply with only this JSON: {\"ok\": true}');\nconsole.log(\"key ping:\", ping.text.slice(0, 60), `cost=$${ping.cost.toFixed(6)}`);\n",[],[175,265],"20e6de27-c68a-4f29-96c3-5f14da08c2d0","test-engine.ts","67c211f0-8899-11f1-b94d-1607ee4eb77e",16,["D",1785032856979],"// Offline sanity checks: key card distribution + full game loop with a stub LLM.\nimport { makeKeyCard, totalAgents } from \"../engine/keycard.ts\";\nimport { legalGuess, mulberry32, newGame, other, pickGiver } from \"../engine/game.ts\";\nimport { playTurn } from \"../engine/play.ts\";\nimport type { LlmOut, ModelCfg } from \"../shared/types.ts\";\n\n// 1) key card distribution over many seeds\nfor (let seed = 1; seed <= 300; seed++) {\n const k = makeKeyCard(mulberry32(seed));\n const count = (side: \"a\" | \"b\", c: string) => k[side].filter((x) => x === c).length;\n if (count(\"a\", \"green\") !== 9 || count(\"b\", \"green\") !== 9) throw new Error(\"green count wrong\");\n if (count(\"a\", \"assassin\") !== 3 || count(\"b\", \"assassin\") !== 3) throw new Error(\"assassin count wrong\");\n if (totalAgents(k) !== 15) throw new Error(\"total agents != 15\");\n const bothAssassin = k.a.filter((x, i) => x === \"assassin\" && k.b[i] === \"assassin\").length;\n if (bothAssassin !== 1) throw new Error(\"double assassin count wrong\");\n}\nconsole.log(\"keycard distribution OK (300 seeds)\");\n\n// 2) full-game loop with a stub chat: giver clues \"ZEBRA\", guesser guesses a random legal word then stops\nconst stub = (rng: () => number) => async (_m: ModelCfg, prompt: string): Promise => {\n const base = { reasoning: null, cost: 0.0001, tokensIn: 100, tokensOut: 20 };\n if (prompt.includes(\"You are the CLUE GIVER\")) {\n return { ...base, text: JSON.stringify({ thinking: \"stub clue\", clue: \"ZEBRA\", number: 2 }) };\n }\n // parse legal words crudely from state via global (we cheat: guess handled below)\n return { ...base, text: JSON.stringify({ thinking: \"stub guess\", guess: (globalThis as any).__nextGuess ?? \"STOP\" }) };\n};\n\nlet results = { won: 0, lost_assassin: 0, lost_timeout: 0, error: 0 };\nfor (let seed = 1; seed <= 20; seed++) {\n const s = newGame(seed);\n const rng = mulberry32(seed * 7 + 1);\n const chatFn = stub(rng);\n let hardStop = 0;\n while (s.status === \"playing\" && hardStop++ < 40) {\n const giver = pickGiver(s);\n // choose a random legal word for the stub guesser before the turn runs\n const legal = s.words.map((_, i) => i).filter((i) => legalGuess(s, giver, i));\n (globalThis as any).__nextGuess = legal.length ? s.words[legal[Math.floor(rng() * legal.length)]] : \"STOP\";\n await playTurn(s, { a: { id: \"stub\", label: \"stub\" }, b: { id: \"stub\", label: \"stub\" } }, chatFn);\n }\n if (s.status === \"playing\") throw new Error(\"game did not terminate\");\n results[s.status as keyof typeof results]++;\n if (s.transcript.length > 9) throw new Error(\"more than 9 turns recorded\");\n}\nconsole.log(\"stub games OK:\", JSON.stringify(results));\n",[],[175,273],"ba105316-aef4-4a72-8c26-6723c3b10635","seed.ts","2cb754e4-8899-11f1-8e10-1607ee4eb77e",69,["D",1785160161999],"// Enqueue games for the current SEED_MODE / SEED_TIER / SEED_KEYS in runner/config.ts.\n// - duet: unordered pairings incl. self (roles alternate anyway)\n// - classic: ORDERED pairings incl. self (model_a = giver, model_b = guesser)\n// Guards against double-seeding a tier.\nimport { init, sqlite } from \"../database/db.ts\";\nimport { newGame } from \"../engine/game.ts\";\nimport { GAMES_PER_PAIRING, MODELS, SEED_KEYS, SEED_MODE, SEED_PAIRS, SEED_TIER } from \"./config.ts\";\n\nawait init();\n\nconst existing = await sqlite.execute({\n sql: \"SELECT COUNT(*) AS c FROM games WHERE tier = ?\",\n args: [SEED_TIER],\n});\nif (Number(existing.rows[0].c) > 0) {\n console.log(`tier \"${SEED_TIER}\" already has ${existing.rows[0].c} games - not seeding again.`);\n} else {\n const pairs: Array<[string, string]> = [];\n if (SEED_PAIRS.length) {\n pairs.push(...SEED_PAIRS);\n } else if (SEED_MODE === \"classic\") {\n for (const ka of SEED_KEYS) for (const kb of SEED_KEYS) pairs.push([ka, kb]);\n } else {\n for (let i = 0; i < SEED_KEYS.length; i++) {\n for (let j = i; j < SEED_KEYS.length; j++) pairs.push([SEED_KEYS[i], SEED_KEYS[j]]);\n }\n }\n let inserted = 0;\n for (const [ka, kb] of pairs) {\n for (let g = 0; g < GAMES_PER_PAIRING; g++) {\n const seed = Math.floor(Math.random() * 2 ** 31);\n const state = newGame(seed, SEED_MODE);\n await sqlite.execute({\n sql: `INSERT INTO games (model_a, model_b, label_a, label_b, tier, mode, seed, status, created_at, state)\n VALUES (?, ?, ?, ?, ?, ?, ?, 'queued', ?, ?)`,\n args: [ka, kb, MODELS[ka].label, MODELS[kb].label, SEED_TIER, SEED_MODE, seed, new Date().toISOString(), JSON.stringify(state)],\n });\n inserted++;\n }\n }\n console.log(`seeded ${inserted} ${SEED_MODE} games across ${pairs.length} pairings (tier=${SEED_TIER})`);\n}\n",[],[175,281],"ac2739b0-dbd0-4cd7-9ca4-35dd3f17ca68","run-next.ts","3318414a-8899-11f1-8891-1607ee4eb77e",26,["D",1785033156609],"// Manual queue advance (short budget so MCP/tool runs return in time).\n// The cron in runner/cron.ts does the heavy lifting in the background.\nimport { drain } from \"./drain.ts\";\n\nawait drain(25_000, 3);\n",[],[175,289],"fc6f3284-3b1a-487f-b8de-a33a831dd447","pick-frontier.ts","cb44bc90-89c1-11f1-b5f2-1607ee4eb77e",90,["D",1785162723479],"// Surface frontier-model candidates per lab from OpenRouter's live catalog.\nconst r = await fetch(\"https://openrouter.ai/api/v1/models\");\nconst j = await r.json();\nconst models = (j.data ?? []).filter((m: any) => !m.id.endsWith(\":free\"));\nconst hit = (m: any, p: string) => m.id.startsWith(p) || m.id.includes(p);\nconst groups: Record = {\n anthropic: [\"anthropic/claude-opus-5\", \"anthropic/claude-sonnet-5\"],\n openai: [\"openai/gpt-5.6\"],\n google: [\"google/gemini-3\"],\n xiaomi: [\"xiaomi/\", \"mimo\"],\n tencent: [\"tencent/\"],\n kimi: [\"moonshotai/kimi-k3\"],\n glm: [\"z-ai/glm-5\"],\n deepseek: [\"deepseek/deepseek-v4\"],\n minimax: [\"minimax/minimax-m3\"],\n};\nfor (const [lab, prefixes] of Object.entries(groups)) {\n console.log(`--- ${lab} ---`);\n const hits = models.filter((m: any) => prefixes.some((p) => hit(m, p)));\n hits.sort((a: any, b: any) => (b.created ?? 0) - (a.created ?? 0));\n for (const m of hits.slice(0, 8)) {\n const p = m.pricing ?? {};\n console.log(\n `${m.id} $${(Number(p.prompt) * 1e6).toFixed(2)}/M in $${(Number(p.completion) * 1e6).toFixed(2)}/M out (${new Date((m.created ?? 0) * 1000).toISOString().slice(0, 10)})`,\n );\n }\n}\n",[],[175,297],"3ca9bad0-dd21-4284-8280-7a62ea8a15ed","list-free.ts","bc2d967e-8899-11f1-90d0-1607ee4eb77e",23,["D",1785032969028],"// List :free models on OpenRouter (candidates for a $0 bracket) + test one call.\nimport { chat } from \"../engine/llm.ts\";\n\nconst r = await fetch(\"https://openrouter.ai/api/v1/models\");\nconst j = await r.json();\nconst free = (j.data ?? []).filter((m: any) => m.id.endsWith(\":free\"));\nconsole.log(`free models: ${free.length}`);\nfor (const m of free) console.log(m.id, \"| ctx\", m.context_length);\n\n// probe: do free models respond despite the ~-$0.15 balance?\nfor (const id of [\"nvidia/nemotron-3-ultra-550b-a55b:free\", \"google/gemma-4-31b-it:free\", \"openai/gpt-oss-20b:free\", \"nvidia/nemotron-3-super-120b-a12b:free\"]) {\n try {\n const out = await chat({ id, label: id, maxTokens: 400 }, 'Reply with only this JSON: {\"ok\": true}');\n console.log(`PING OK ${id}:`, out.text.slice(0, 60).replaceAll(\"\\n\", \" \"), \"cost:\", out.cost);\n } catch (e) {\n console.log(`PING FAILED ${id}:`, String(e).slice(0, 200));\n }\n}\n",[],[175,305],"1556376b-2070-42b3-a45c-e21e1924722b","drain.ts","412306e8-889a-11f1-8f31-1607ee4eb77e",115,["D",1785191960131],7,"// Shared queue-drainer: claims games and advances them within a time budget.\n// State is saved after every turn, so interrupted runs resume cleanly.\n// Paid tiers are prioritized over 'free' (fast + what Charlie's waiting on).\nimport { claim, finalize, init, saveState, sqlite } from \"../database/db.ts\";\nimport { playTurn } from \"../engine/play.ts\";\nimport { MODELS } from \"./config.ts\";\nimport type { GameState } from \"../shared/types.ts\";\n\nexport async function drain(budgetMs: number, concurrency = 3): Promise {\n const t0 = Date.now();\n await init();\n // hard spend ceiling: once paid usage nears $10, only free-tier games may run\n const spent = await sqlite.execute(`SELECT COALESCE(SUM(cost),0) AS c FROM games`);\n const capped = Number(spent.rows[0].c) >= 9.5;\n const rows = await sqlite.execute({\n sql: `SELECT id, model_a, model_b, state FROM games\n WHERE (status='queued' OR (status='running' AND (claimed_at IS NULL OR claimed_at < ?)))\n ${capped ? \"AND tier LIKE '%free%'\" : \"\"}\n ORDER BY CASE WHEN tier LIKE '%free%' THEN 1 ELSE 0 END,\n CASE WHEN mode='classic' THEN 0 ELSE 1 END, id\n LIMIT ?`,\n args: [new Date(Date.now() - 330_000).toISOString(), concurrency],\n });\n if (capped) console.log(\"SPEND CAP REACHED (~$9.50): paid games paused, free games only\");\n if (rows.rows.length === 0) {\n console.log(\"QUEUE EMPTY (or all claimed by live workers)\");\n } else {\n await Promise.all(rows.rows.map(async (r) => {\n const id = Number(r.id);\n if (!(await claim(id))) return; // another worker got there first\n const models = { a: MODELS[String(r.model_a)], b: MODELS[String(r.model_b)] };\n if (!models.a || !models.b) {\n console.log(`#${id}: unknown model key (${r.model_a}, ${r.model_b}) - marking error`);\n await sqlite.execute({ sql: `UPDATE games SET status='done', result='error' WHERE id=?`, args: [id] });\n return;\n }\n let s: GameState;\n try {\n s = JSON.parse(String(r.state));\n } catch (e) {\n console.log(`#${id}: bad state json: ${e}`);\n return;\n }\n while (s.status === \"playing\" && Date.now() - t0 < budgetMs) {\n await playTurn(s, models);\n await saveState(id, s);\n }\n if (s.status !== \"playing\") {\n // integrity guard: a \"game\" where no guess ever landed is a failed run\n // (provider outage / parse failures), not a result — void it.\n const voided = s.transcript.length > 0 && s.transcript.every((t) => t.guesses.length === 0);\n if (voided && s.status !== \"won\") s.status = \"error\";\n await finalize(id, s);\n }\n console.log(\n `#${id} ${r.model_a}x${r.model_b}: ${s.status} agents=${s.agentsFound}/15 turns=${s.transcript.length} cost=$${s.cost.toFixed(4)}`,\n );\n }));\n }\n const left = await sqlite.execute(`SELECT COUNT(*) AS c FROM games WHERE status IN ('queued','running')`);\n const tot = await sqlite.execute(`SELECT COUNT(*) AS c, COALESCE(SUM(cost),0) AS cost FROM games`);\n console.log(\n `pending: ${left.rows[0].c} | total games: ${tot.rows[0].c} | total spend: $${Number(tot.rows[0].cost).toFixed(4)}`,\n );\n}\n",[],[175,313],"02fe93be-1f37-445d-92cc-5330aa730160","cron.ts","45efc940-889a-11f1-a4ba-1607ee4eb77e",27,["D",1785033160478],"interval","// Background worker: every minute, drains the game queue (no-op when empty).\nimport { drain } from \"./drain.ts\";\n\nexport default async function (interval: Interval) {\n await drain(52_000, 3);\n}\n",[],[175,322],"e170e9db-ff5e-4e7c-b9c4-7e3dc21edf4c",["D",1785191966906],[],[175,33],"cf7d4bbe-9e0f-4674-a6cb-89488372e7af","check-credits.ts","a93ed1b8-8899-11f1-b37d-1607ee4eb77e",19,["D",1785032897552],"// Reports the OpenRouter key's usage/limits so we know the budget headroom.\nconst r = await fetch(\"https://openrouter.ai/api/v1/key\", {\n headers: { Authorization: `Bearer ${Deno.env.get(\"OPENROUTER_API_KEY\")}` },\n});\nconst j = await r.json();\nconsole.log(JSON.stringify(j, null, 2));\n\nconst c = await fetch(\"https://openrouter.ai/api/v1/credits\", {\n headers: { Authorization: `Bearer ${Deno.env.get(\"OPENROUTER_API_KEY\")}` },\n});\nconsole.log(JSON.stringify(await c.json(), null, 2));\n",[],[175,335],["D",1785032678458],[220,221,222,223,224,225,226,227,228,229],[175],"a518229d-beec-4b3a-9dbd-c1b8555e33cf","index.tsx","a68789aa-889a-11f1-aff9-1607ee4eb77e",114,["D",1785188537164],13,"/** @jsxImportSource https://esm.sh/react@18.2.0 */\n// Codenames bench UI. Palette lifted from the physical Codenames game:\n// cream word cards, blue + red team cards, tan bystanders, black assassin.\n// Charlie's handwriting (Fullerfont) is reserved for the title, a couple of\n// section headings, and score numbers only.\nimport React, { useEffect, useMemo, useState } from \"https://esm.sh/react@18.2.0\";\nimport { createRoot } from \"https://esm.sh/react-dom@18.2.0/client\";\nimport { hexA } from \"https://esm.town/v/dcm31/house-style/frontend/directions.ts\";\nimport { Mark, Squiggle, WobblyBox } from \"https://esm.town/v/dcm31/house-style/frontend/pen.tsx\";\n\nconst FONT_URL = \"https://cdn.jsdelivr.net/gh/dcm31/myfont@main/Fullerfont-Regular.otf\";\nconst CODENAMES_URL = \"https://czechgames.com/en/codenames/\";\n\n/** Guess Efficiency Score, mirrored from database/standings.ts */\nfunction gesPct(result: string, words: number, clues: number): number {\n const raw = words + (result === \"won\" ? Math.max(1, 10 - clues) : 0);\n return Math.round((raw / 18) * 100);\n}\n\nconst hand = (size: string): React.CSSProperties => ({\n fontFamily: \"'Fullerfont', sans-serif\",\n fontWeight: 400,\n fontSize: size,\n lineHeight: 1.25,\n});\nconst handBold = (size: string): React.CSSProperties => ({\n ...hand(size),\n WebkitTextStroke: \"0.035em currentColor\",\n});\nconst squigUnder = (color: string, sw = 2): React.CSSProperties => {\n const svg =\n ``;\n return {\n textDecoration: \"none\",\n backgroundImage: `url(\"data:image/svg+xml,${encodeURIComponent(svg)}\")`,\n backgroundRepeat: \"repeat-x\",\n backgroundPosition: \"0 100%\",\n backgroundSize: \"28px 7px\",\n paddingBottom: \"3px\",\n };\n};\n\nconst usd = (x: number) => x === 0 ? \"$0\" : x < 0.01 ? `$${x.toFixed(4)}` : `$${x.toFixed(2)}`;\nconst RESULT_TEXT: Record = {\n won: \"won\", lost_assassin: \"assassin\", lost_timeout: \"out of clues\", error: \"error\",\n};\nconst short = (label: string) =>\n label.replace(\"Nemotron \", \"\").replace(\"Gemini 2.5 \", \"\").replace(\"GPT-OSS\", \"OSS\").replace(\" 4.5\", \"\").replace(\" V3.1\", \"\").replace(\"Gemini 3 Pro\", \"G3 Pro\").replace(\"GPT-5.6 \", \"\").replace(\"DeepSeek V4 Pro\", \"DS V4\");\n\n// ---- Codenames palette ----\nfunction useTheme() {\n const [dark, setDark] = useState(false);\n useEffect(() => {\n const mq = window.matchMedia(\"(prefers-color-scheme: dark)\");\n setDark(mq.matches);\n const fn = (e: MediaQueryListEvent) => setDark(e.matches);\n mq.addEventListener(\"change\", fn);\n return () => mq.removeEventListener(\"change\", fn);\n }, []);\n return {\n dark,\n paper: dark ? \"#211d16\" : \"#f1e8d2\",\n ink: dark ? \"#ebe4d1\" : \"#29241c\",\n faint: dark ? \"#98907c\" : \"#8a7f6a\",\n blue: dark ? \"#7fb0d9\" : \"#33698f\",\n red: dark ? \"#e0715f\" : \"#b8422f\",\n tan: dark ? \"#a3936f\" : \"#b3a075\",\n black: dark ? \"#0d0b08\" : \"#1d1812\",\n };\n}\ntype Theme = ReturnType;\n\nfunction resultColor(t: Theme, result: string): string {\n return result === \"won\" ? t.blue : result === \"lost_assassin\" ? (t.dark ? t.red : t.black) : result.startsWith(\"lost\") || result === \"error\" ? t.red : t.faint;\n}\n\nfunction GameChips({ list, t, nonce, onGame, target = 9 }: { list: any[]; t: Theme; nonce: number; onGame: (id: number) => void; target?: number }) {\n return (\n

\n {list.map((g: any, i: number) => (\n \n ))}\n

\n );\n}\n\n// ---------------- leaderboard: one grouped chart, tap for details ----------------\n\nfunction Leaderboard({ charts, t, nonce }: { charts: any; t: Theme; nonce: number }) {\n const [sel, setSel] = useState(null);\n const byKey = (rows: any[]) => Object.fromEntries((rows ?? []).map((r: any) => [r.key, r]));\n const overall = byKey(charts.overall);\n const giving = byKey(charts.giving);\n const guessing = byKey(charts.guessing);\n const keys: string[] = useMemo(() => {\n const all = [...new Set([...(charts.overall ?? []), ...(charts.giving ?? []), ...(charts.guessing ?? [])].map((r: any) => r.key))];\n return all.sort((a, b) => (overall[b]?.ges ?? -1) - (overall[a]?.ges ?? -1));\n }, [charts]);\n if (!keys.length) return null;\n const label = (k: string) => (overall[k] ?? giving[k] ?? guessing[k])?.label ?? k;\n const H = 170;\n const bar = (row: any, color: string) => (\n

\n {row ? Math.round(row.ges) : 0}\n

\n );\n const selRow = sel\n ? { key: sel, label: label(sel), overall: overall[sel], giving: giving[sel], guessing: guessing[sel] }\n : null;\n const detail = (name: string, r: any, color: string) =>\n r\n ? (\n \n {name}: {Math.round(r.ges)}{\" \"}\n · {r.wins} win{r.wins === 1 ? \"\" : \"s\"} in {r.games} game{r.games === 1 ? \"\" : \"s\"}\n \n )\n : {name}: no games yet;\n return (\n

\n overall (self-play)\n clue giving\n guessing\n tap a model for details\n

\n {[25, 50, 75, 100].map((y) => (\n

\n {y}\n

\n ))}\n

\n {keys.map((k) => (\n setSel(sel === k ? null : k)}\n className=\"cursor-pointer flex flex-col items-center gap-1\"\n style={{ background: sel === k ? hexA(t.blue, 0.08) : \"none\", border: \"none\", borderRadius: 10, padding: \"4px 6px 6px\", width: 96 }}\n >\n

\n {bar(overall[k], t.tan)}\n {bar(giving[k], t.blue)}\n {bar(guessing[k], t.red)}\n

\n {short(label(k))}\n \n ))}\n

\n {selRow && (\n

\n {selRow.label}\n {detail(\"overall, self-play\", selRow.overall, t.tan)}\n {detail(\"clue giving, for reference guesser DeepSeek V3.1\", selRow.giving, t.blue)}\n {detail(\"guessing, for reference giver GPT-5 mini\", selRow.guessing, t.red)}\n

\n )}\n

\n );\n}\n\n// ---------------- merged matrix ----------------\n\nfunction MergedMatrix({ merged, title, helper, t, nonce, onGame }: { merged: any; title: string; helper?: string; t: Theme; nonce: number; onGame: (id: number) => void }) {\n const [sel, setSel] = useState<{ g: string; u: string } | null>(null);\n const keys: string[] = merged.keys;\n const cell = (g: string, u: string) => merged.cells?.[g]?.[u] ?? null;\n const selCell = sel ? cell(sel.g, sel.u) : null;\n return (\n

\n {title}\n \n {helper &&

{helper}

}\n

\n CLUE GIVERS\n

GUESSERS

\n \n {keys.map((u) => (\n {short(merged.labels[u])}\n ))}\n {keys.map((g) => (\n \n {short(merged.labels[g])}\n {keys.map((u) => {\n const c = cell(g, u);\n const active = sel?.g === g && sel?.u === u;\n const color = !c ? hexA(t.faint, 0.45) : c.wins > 0 ? t.blue : t.red;\n return (\n c && setSel(active ? null : { g, u })}\n className={c ? \"cursor-pointer flex flex-col items-center py-1\" : \"flex flex-col items-center py-1\"}\n style={{\n background: active ? hexA(t.blue, 0.1) : \"none\",\n border: \"none\",\n borderBottom: `1px dashed ${hexA(t.faint, 0.3)}`,\n borderRadius: 8,\n }}\n >\n {c ? Math.round(c.ges / c.games) : \"–\"}\n {c && {c.wins}/{c.games} won}\n \n );\n })}\n \n ))}\n

\n {sel && selCell && (\n

\n \n {merged.labels[sel.g]} cluing for {merged.labels[sel.u]}\n \n \n

\n )}\n

\n );\n}\n\n// ---------------- game logs ----------------\n\nfunction GameLogs({ recent, t, nonce, onGame }: { recent: any[]; t: Theme; nonce: number; onGame: (id: number) => void }) {\n if (!recent?.length) return null;\n return (\n

\n game logs\n \n

\n {recent.map((g: any) => (\n onGame(g.id)}\n className=\"cursor-pointer text-left w-full grid items-center gap-2 py-2\"\n style={{ background: \"none\", border: \"none\", borderBottom: `1px dashed ${hexA(t.faint, 0.35)}`, gridTemplateColumns: \"minmax(0,2.4fr) minmax(64px,1fr) minmax(56px,1fr) minmax(48px,0.8fr)\", color: t.ink }}\n >\n \n {short(g.giver)}\n cluing for \n {short(g.guesser)}\n \n {RESULT_TEXT[g.result] ?? g.result}\n {g.words}/9 · {g.clues} clues\n {g.ges}%\n \n ))}\n

\n );\n}\n\n// ---------------- tucked-away duet experiments ----------------\n\nfunction DuetArchive({ teams, t, nonce, onGame }: { teams: any[]; t: Theme; nonce: number; onGame: (id: number) => void }) {\n const [openTeam, setOpenTeam] = useState(null);\n if (!teams.length) return null;\n return (\n

\n earlier duet-mode experiments (different rules)\n

\n {teams.map((team: any) => (\n

\n setOpenTeam(openTeam === team.key ? null : team.key)}\n className=\"cursor-pointer text-left w-full grid items-center gap-2 py-2\"\n style={{ background: \"none\", border: \"none\", borderBottom: `1px dashed ${hexA(t.faint, 0.35)}`, gridTemplateColumns: \"minmax(0,2.2fr) repeat(2, minmax(52px, 1fr))\", color: t.ink }}\n >\n \n {team.labels[0]}\n {team.keys[0] !== team.keys[1] ? {\" + \"}{team.labels[1]} : {\" solo\"}}\n \n {team.avgAgents.toFixed(1)}/15 words\n \n {team.wins ? `${team.wins}w` : team.assassins ? \"assassin\" : \"out of clues\"}\n \n \n {openTeam === team.key && }\n

\n ))}\n

\n );\n}\n\nfunction Standings({ data, t, nonce, onGame }: { data: any; t: Theme; nonce: number; onGame: (id: number) => void }) {\n return (\n

\n \n \n \n

\n hidden models (the full matrix)\n

\n \n

\n );\n}\n\n// ---------------- game story ----------------\n\ninterface Snap {\n covered: boolean[];\n marks: { a: boolean[]; b: boolean[] };\n dead: number;\n}\n\n/** rebuild the board after each turn so the story can show state over time */\nfunction buildSnapshots(s: any): Snap[] {\n const wordIdx = new Map(s.words.map((w: string, i: number) => [w, i]));\n let covered = Array(25).fill(false);\n let marks = { a: Array(25).fill(false), b: Array(25).fill(false) };\n let dead = -1;\n const snaps: Snap[] = [{ covered: [...covered], marks: { a: [...marks.a], b: [...marks.b] }, dead }];\n for (const tr of s.transcript) {\n for (const gu of tr.guesses) {\n const i = wordIdx.get(gu.word);\n if (i === undefined) continue;\n if (gu.outcome === \"agent\") covered[i] = true;\n else if (gu.outcome === \"bystander\") marks[tr.giver as \"a\" | \"b\"][i] = true;\n else if (gu.outcome === \"assassin\") dead = i;\n }\n snaps.push({ covered: [...covered], marks: { a: [...marks.a], b: [...marks.b] }, dead });\n }\n return snaps;\n}\n\nconst OUTCOME_TEXT: Record = { agent: \"correct\", bystander: \"neutral\", assassin: \"assassin\" };\n\nfunction Dot({ color, big, marked, t }: { color: string; big: boolean; marked: boolean; t: Theme }) {\n const d = big ? 9 : 7;\n return (\n \n \n {marked && (\n \n )}\n \n );\n}\n\nfunction Tile({ word, cellA, cellB, covered, dead, markedA, markedB, giver, badge, classic, t, nonce, small }: {\n word: string; cellA: string; cellB: string; covered: boolean; dead: boolean;\n markedA: boolean; markedB: boolean; giver: \"a\" | \"b\" | null;\n badge: { n: number; outcome: string } | null; classic: boolean;\n t: Theme; nonce: number; small: boolean;\n}) {\n const dot = (cell: string) => cell === \"green\" ? t.blue : cell === \"assassin\" ? (t.dark ? t.red : t.black) : hexA(t.tan, 0.85);\n const stroke = dead ? (t.dark ? t.red : t.black) : covered ? t.blue : badge ? t.red : hexA(t.ink, 0.7);\n const w = small ? 64 : 124;\n const h = small ? 46 : 64;\n const badgeColor = badge ? (badge.outcome === \"agent\" ? t.blue : badge.outcome === \"assassin\" ? (t.dark ? t.red : t.black) : t.tan) : t.faint;\n return (\n \n \n \n {word}\n \n \n \n {!classic && }\n {covered && }\n {dead && }\n \n \n {badge && (\n \n {badge.n}\n \n )}\n \n );\n}\n\nfunction Board({ s, snap, giver, badges, classic, t, nonce, small }: {\n s: any; snap: Snap; giver: \"a\" | \"b\" | null; badges: Map;\n classic: boolean; t: Theme; nonce: number; small: boolean;\n}) {\n return (\n

\n {s.words.map((w: string, i: number) => (\n \n ))}\n

\n );\n}\n\nfunction Thinking({ label, text, t }: { label: string; text: string | null | undefined; t: Theme }) {\n if (!text) return null;\n return (\n

\n {label}\n

\n {text}\n

\n );\n}\n\nfunction GameView({ id, t, nonce, onBack }: { id: number; t: Theme; nonce: number; onBack: () => void }) {\n const [g, setG] = useState(null);\n const [err, setErr] = useState(null);\n const [small, setSmall] = useState(false);\n useEffect(() => {\n const check = () => setSmall(window.innerWidth < 720);\n check();\n window.addEventListener(\"resize\", check);\n return () => window.removeEventListener(\"resize\", check);\n }, []);\n useEffect(() => {\n fetch(`/api/game/${id}`).then((r) => r.json()).then(setG).catch((e) => setErr(String(e)));\n }, [id]);\n\n const s = g?.state;\n const classic = s?.mode === \"classic\";\n const target = classic ? 9 : 15;\n const snaps = useMemo(() => (s ? buildSnapshots(s) : []), [s]);\n\n if (err) return

{err}

;\n if (!g) return

fetching the game...

;\n\n const sideLabel = (side: string) => side === \"a\" ? g.label_a : g.label_b;\n const result = g.result ?? s.status;\n const solo = g.model_a === g.model_b;\n const ges = classic && g.result ? gesPct(g.result, s.agentsFound, s.transcript.length) : null;\n const title = classic\n ? `${g.label_a} cluing${solo ? \" for itself\" : ` for ${g.label_b}`}`\n : `${g.label_a}${solo ? \" solo pair\" : ` + ${g.label_b}`}`;\n const emptyBadges = new Map();\n\n return (\n

\n \n game {g.id}: {title}\n {RESULT_TEXT[result] ?? result}\n {ges !== null && {ges}%}\n \n {s.agentsFound}/{target} words · {s.transcript.length} clues · {usd(s.cost)}\n \n

\n\n {/* the fresh board */}\n

\n the fresh board\n \n {classic\n ? (\n

\n one secret key, held by {g.label_a}: 9 team words to find, 3{\" \"}\n assassins, the rest bystanders. {solo ? \"the guessing copy of the model\" : g.label_b} guesses blind\n with up to N+1 guesses per clue. a slash = revealed neutral. numbered rings = that clue's guesses in order.\n

\n )\n : (\n

\n duet deals each partner their own secret key: the left dot is {g.label_a}'s{solo ? \" (as player A)\" : \"\"}, the right is{\" \"}\n {solo ? \"its twin's (player B)\" : `${g.label_b}'s`}. every key hides 9 team words and 3{\" \"}\n assassins — a word can be safe on one key and fatal on the other.\n guesses are judged on whichever partner gave the clue. numbered rings = that clue's guesses in order.\n

\n )}\n

\n\n {/* the story, one long scroll */}\n {s.transcript.map((turn: any, i: number) => {\n const badges = new Map();\n turn.guesses.forEach((gu: any, k: number) => badges.set(gu.word, { n: k + 1, outcome: gu.outcome }));\n return (\n

\n clue {turn.turn} of {s.transcript.length}\n {sideLabel(turn.giver)} {classic ? \"says\" : `(player ${turn.giver.toUpperCase()}) says`}\n {turn.clue\n ? {turn.clue.word} {turn.clue.number}\n : no valid clue ({turn.invalid})}\n

\n {turn.clue && }\n

\n {turn.guesses.map((gu: any, k: number) => (\n

\n {k + 1}.\n {gu.word}\n \n {OUTCOME_TEXT[gu.outcome] ?? gu.outcome}\n \n {gu.outcome === \"agent\" && }\n {gu.outcome === \"assassin\" && }\n

\n \n

\n ))}\n {turn.guesses.length === 0 && !turn.stopped && no guesses landed this clue}\n {turn.stopped && guesser stopped to bank progress}\n {turn.error && {turn.error}}\n

\n the board after clue {turn.turn}\n \n

\n );\n })}\n\n

\n {RESULT_TEXT[result] ?? result}{classic && g.result ? ` — ${s.agentsFound}/9 words in ${s.transcript.length} clues` : \"\"}\n

\n );\n}\n\n// ---------------- app ----------------\n\nfunction App() {\n const t = useTheme();\n const [fontLoaded, setFontLoaded] = useState(false);\n const [nonce] = useState(0);\n const [data, setData] = useState(null);\n const [route, setRoute] = useState(window.location.hash);\n\n useEffect(() => {\n const font = new FontFace(\"Fullerfont\", `url('${FONT_URL}')`);\n font.load().then(() => {\n document.fonts.add(font);\n setFontLoaded(true);\n }).catch(() => setFontLoaded(true));\n }, []);\n\n useEffect(() => {\n const fn = () => setRoute(window.location.hash);\n window.addEventListener(\"hashchange\", fn);\n return () => window.removeEventListener(\"hashchange\", fn);\n }, []);\n\n const load = () => fetch(\"/api/standings\").then((r) => r.json()).then(setData).catch(() => {});\n useEffect(() => {\n load();\n }, []);\n useEffect(() => {\n if (!data || data.totals.pending === 0) return;\n const iv = setInterval(load, 20_000);\n return () => clearInterval(iv);\n }, [data?.totals?.pending]);\n\n const gameMatch = route.match(/^#\\/game\\/(\\d+)$/);\n const openGame = (id: number) => {\n window.location.hash = `#/game/${id}`;\n };\n\n if (!fontLoaded) return null;\n\n return (\n

codenames bench

\n \n

\n A benchmark to see how well AI models can play{\" \"}\n Codenames\n

\n\n {gameMatch\n ? { window.location.hash = \"\"; }} />\n : data\n ? \n :

fetching standings...

}\n\n \n

\n );\n}\n\ncreateRoot(document.getElementById(\"root\")!).render();\n",[],[181,346],"ae7593b0-75cf-4d74-8ea1-eb0f1c4e2d8f","index.html","67d69db8-889a-11f1-8728-1607ee4eb77e",["D",1785033217341],"\n\n\n \n \n codenames duet bench\n \n \n\n\n

\n \n\n\n",[],[181,355],["D",1785033217341],[217,218],[181],"e29e0e1f-8852-4674-99e4-476bdb4db37b","prompts.ts","09be0168-8899-11f1-b8e2-1607ee4eb77e",49,["D",1785039051116],"import type { GameState, Side, TurnRec } from \"../shared/types.ts\";\nimport { agentsTarget, other, remainingGreens, TOTAL_TURNS } from \"./game.ts\";\n\nfunction label(cell: string): string {\n return cell === \"green\" ? \"GREEN\" : cell === \"assassin\" ? \"ASSASSIN\" : \"NEUTRAL\";\n}\n\nfunction history(transcript: TurnRec[], me: Side): string {\n const done = transcript.filter((t) => t.clue || t.invalid);\n if (!done.length) return \"(no clues yet)\";\n return done\n .map((t) => {\n const who = t.giver === me ? \"You\" : \"Partner\";\n if (!t.clue) return `Clue ${t.turn}: ${who} failed to give a valid clue (turn forfeited).`;\n const gs = t.guesses.map((g) => `${g.word}=${g.outcome.toUpperCase()}`).join(\", \") || \"(no guesses)\";\n return `Clue ${t.turn}: ${who} gave \"${t.clue.word} ${t.clue.number}\" -> ${gs}${t.stopped ? \"; guesser stopped\" : \"\"}`;\n })\n .join(\"\\n\");\n}\n\nconst RULES_DUET = `You are playing Codenames Duet, a fully cooperative word game, with a partner.\n- The board has 25 words. 15 of them are AGENTS in total; you win together if all 15 are found within ${TOTAL_TURNS} clues (total, shared between both of you).\n- Each player has their own secret key card side. On YOUR side, 9 words are GREEN (agents), 3 are ASSASSINS, 13 are NEUTRAL bystanders. Your partner's side is different: 3 of your greens are also green for them, and the sides overlap in tricky ways — a word that is green on one side can be neutral or even an ASSASSIN on the other.\n- When a clue is given, the guesser's guesses are judged against the CLUE GIVER's key: green = agent found (guesser may keep guessing), neutral = the turn ends harmlessly, ASSASSIN = you both lose instantly.`;\n\nconst RULES_CLASSIC = `You are playing cooperative classic Codenames with a partner.\n- The board has 25 words. There is ONE secret key: 9 GREEN agents, 3 ASSASSINS, 13 NEUTRAL bystanders.\n- The CLUE GIVER sees the key; the GUESSER sees nothing but the board and the clue history.\n- You win together if all 9 agents are found within ${TOTAL_TURNS} clues. For a clue with number N, the guesser gets up to N+1 guesses (a clue of 0 means unlimited guesses).\n- Each guess: GREEN = agent found (guessing may continue), NEUTRAL = the turn ends harmlessly, ASSASSIN = you both lose instantly.`;\n\nexport function buildCluePrompt(s: GameState, giver: Side): string {\n const done = s.transcript.slice(0, -1); // current turn record is already pushed\n const classic = s.mode === \"classic\";\n const target = agentsTarget(s);\n const opp = other(giver);\n const lines = s.words.map((w, i) => {\n if (s.covered[i]) return `${w} - FOUND (agent, done)`;\n const flags: string[] = [];\n if (classic) {\n if (s.marks[giver][i]) flags.push(\"already revealed NEUTRAL; cannot be guessed again\");\n } else {\n if (s.marks[giver][i]) flags.push(\"already revealed NEUTRAL on your key; partner cannot guess it on your clues\");\n if (s.marks[opp][i]) flags.push(\"revealed NEUTRAL on partner's key (could still be anything on yours)\");\n }\n return `${w} - your key: ${label(s.key[giver][i])}${flags.length ? ` (${flags.join(\"; \")})` : \"\"}`;\n }).join(\"\\n\");\n const mine = remainingGreens(s, giver).map((i) => s.words[i]).join(\", \") || \"(none)\";\n const guessRule = classic\n ? \"Your partner then guesses one word at a time, up to NUMBER+1 guesses while they keep hitting greens, and may stop after at least one guess.\"\n : \"Your partner will then guess words one at a time; they may keep guessing as long as they keep hitting your greens, and may stop after at least one guess.\";\n const strategy = classic\n ? \"STRATEGY NOTES: connect as many remaining green words as safely possible, and steer well clear of anything that could pull your partner toward the 3 assassins. Your partner cannot see the key at all.\"\n : \"STRATEGY NOTES: connect as many of YOUR remaining green words as safely possible, and steer clear of anything that could pull your partner toward YOUR 3 assassins or neutrals. Your partner cannot see your key.\";\n return `${classic ? RULES_CLASSIC : RULES_DUET}\n\nYou are the CLUE GIVER for clue ${s.turn} of ${TOTAL_TURNS}. Agents found so far: ${s.agentsFound}/${target}.\n\nYour clue must be exactly ONE word (not any word visible on the board, no board word as a substring trick) plus a NUMBER = how many board words relate to it. ${guessRule}\n\n${strategy}\n\nBOARD (word - what ${classic ? \"THE key\" : \"YOUR key\"} says):\n${lines}\n\nAgents still to find: ${mine}\nClues remaining including this one: ${TOTAL_TURNS - s.turn + 1}\n\nHISTORY:\n${history(done, giver)}\n\nKeep your reasoning brief. Respond with ONLY a JSON object, no other text:\n{\"thinking\": \"your private reasoning\", \"clue\": \"ONEWORD\", \"number\": 2}`;\n}\n\nexport function buildGuessPrompt(\n s: GameState,\n giver: Side,\n current: TurnRec,\n canStop: boolean,\n): string {\n const classic = s.mode === \"classic\";\n const target = agentsTarget(s);\n const guesser = other(giver);\n const done = s.transcript.slice(0, -1);\n const lines = s.words.map((w, i) => {\n if (s.covered[i]) return `${w} - FOUND (agent, cannot guess)`;\n if (classic) {\n return s.marks[giver][i] ? `${w} - revealed NEUTRAL (illegal to guess)` : w;\n }\n const flags: string[] = [];\n if (s.marks[giver][i]) flags.push(\"revealed NEUTRAL against partner's key - ILLEGAL to guess on this clue\");\n if (s.marks[guesser][i]) flags.push(\"was NEUTRAL against YOUR key (still guessable - might be an agent on partner's key)\");\n return `${w} - your key: ${label(s.key[guesser][i])}${flags.length ? ` (${flags.join(\"; \")})` : \"\"}`;\n }).join(\"\\n\");\n const soFar = current.guesses.length\n ? current.guesses.map((g) => `${g.word} -> ${g.outcome.toUpperCase()}`).join(\", \")\n : \"(none yet)\";\n const n = current.clue!.number;\n const situational = classic\n ? `You cannot see the key. For this clue you get up to ${n === 0 ? \"unlimited\" : `${n + 1}`} guesses while you keep hitting greens. A NEUTRAL just ends the turn; an ASSASSIN loses the game instantly, so avoid words your partner probably wasn't pointing at.`\n : `Your guesses are judged against your PARTNER's key, which you CANNOT see. You DO see your own key below, but beware: a word that is GREEN on your key may be neutral or even the ASSASSIN on your partner's key (exactly one of your greens is your partner's assassin). A word that is ASSASSIN on your key might even be green on theirs. Hitting a NEUTRAL just ends the turn; hitting your partner's ASSASSIN loses the game instantly, so avoid words your partner probably wasn't pointing at.`;\n return `${classic ? RULES_CLASSIC : RULES_DUET}\n\nYou are the GUESSER. Your partner gave the clue: \"${current.clue!.word}\" for ${n}. This is clue ${s.turn} of ${TOTAL_TURNS}. Agents found so far: ${s.agentsFound}/${target}.\n\n${situational}\n\nBOARD${classic ? \"\" : \" (word - what YOUR OWN key says)\"}:\n${lines}\n\nGuesses so far on THIS clue: ${soFar}\n${canStop ? \"You may guess again, or STOP to bank your progress.\" : \"You MUST make at least one guess for this clue.\"}\n\nHISTORY:\n${history(done, guesser)}\n\nKeep your reasoning brief. Respond with ONLY a JSON object, no other text:\n{\"thinking\": \"your private reasoning\", \"guess\": \"BOARDWORD\"}${canStop ? ' or {\"thinking\": \"...\", \"guess\": \"STOP\"}' : \"\"}`;\n}\n",[],[187,365],"d1b6d568-c960-4422-b019-13d0f21ebcf9","play.ts","1be9c502-8899-11f1-90af-1607ee4eb77e",50,["D",1785039055896],5,"import type { GameState, LlmOut, ModelCfg, Side, TurnRec } from \"../shared/types.ts\";\nimport { anyLegalGuess, applyGuess, endTurnBook, legalGuess, MAX_GUESSES_PER_TURN, other, pickGiver } from \"./game.ts\";\nimport { chat, parseOut } from \"./llm.ts\";\nimport { buildCluePrompt, buildGuessPrompt } from \"./prompts.ts\";\n\nexport type ChatFn = (model: ModelCfg, prompt: string) => Promise;\n\nfunction tally(s: GameState, out: LlmOut): void {\n s.cost += out.cost;\n s.calls++;\n s.tokensIn += out.tokensIn;\n s.tokensOut += out.tokensOut;\n}\n\n/** Plays exactly one clue turn (clue + guessing phase), mutating state. */\nexport async function playTurn(\n s: GameState,\n models: { a: ModelCfg; b: ModelCfg },\n chatFn: ChatFn = chat,\n): Promise {\n if (s.status !== \"playing\") return;\n const giver = pickGiver(s);\n const guesser: Side = other(giver);\n const rec: TurnRec = { turn: s.turn, giver, guesses: [] };\n s.transcript.push(rec);\n\n try {\n // ---- clue phase (up to 2 attempts) ----\n let clueCost = 0;\n for (let attempt = 0; attempt < 2 && !rec.clue; attempt++) {\n const p = buildCluePrompt(s, giver) +\n (attempt ? \"\\n\\nYour previous reply was invalid. The clue must be a single word that is NOT on the board, with an integer number. Respond with ONLY the JSON object.\" : \"\");\n const out = await chatFn(models[giver], p);\n tally(s, out);\n clueCost += out.cost;\n const j = parseOut(out);\n const w = String(j?.clue ?? \"\").trim().toUpperCase();\n const n = Number(j?.number);\n const onBoard = s.words.some((bw, i) => !s.covered[i] && bw.toUpperCase() === w);\n if (w && /^[A-Z][A-Z'&. -]*$/.test(w) && w.length <= 30 && !onBoard && Number.isFinite(n)) {\n rec.clue = {\n word: w,\n number: Math.max(0, Math.min(9, Math.round(n))),\n thinking: typeof j?.thinking === \"string\" ? j.thinking : undefined,\n reasoning: out.reasoning,\n cost: clueCost,\n };\n }\n }\n if (!rec.clue) {\n rec.invalid = \"could not produce a valid clue after 2 attempts\";\n endTurnBook(s, giver);\n return;\n }\n\n // ---- guessing phase ----\n const n = rec.clue.number;\n const maxG = s.mode === \"classic\" && n > 0 ? Math.min(MAX_GUESSES_PER_TURN, n + 1) : MAX_GUESSES_PER_TURN;\n while (s.status === \"playing\" && rec.guesses.length < maxG) {\n if (!anyLegalGuess(s, giver)) break;\n const canStop = rec.guesses.length > 0;\n let idx = -1;\n let stop = false;\n let thinking: string | undefined;\n let reasoning: string | null = null;\n let gCost = 0;\n for (let attempt = 0; attempt < 2; attempt++) {\n const p = buildGuessPrompt(s, giver, rec, canStop) +\n (attempt\n ? `\\n\\nYour previous reply was invalid. Guess must be one unrevealed, legal board word${canStop ? ' or \"STOP\"' : \"\"}. Respond with ONLY the JSON object.`\n : \"\");\n const out = await chatFn(models[guesser], p);\n tally(s, out);\n gCost += out.cost;\n const j = parseOut(out);\n const gw = String(j?.guess ?? \"\").trim().toUpperCase();\n thinking = typeof j?.thinking === \"string\" ? j.thinking : thinking;\n reasoning = out.reasoning ?? reasoning;\n if (gw === \"STOP\" && canStop) {\n stop = true;\n break;\n }\n const found = s.words.findIndex((bw) => bw.toUpperCase() === gw);\n if (found >= 0 && legalGuess(s, giver, found)) {\n idx = found;\n break;\n }\n }\n if (stop) {\n rec.stopped = true;\n break;\n }\n if (idx < 0) {\n rec.error = (rec.error ? rec.error + \"; \" : \"\") + \"no legal guess produced\";\n break;\n }\n const outcome = applyGuess(s, giver, idx);\n rec.guesses.push({ word: s.words[idx], outcome, thinking, reasoning, cost: gCost });\n if (outcome !== \"agent\") break;\n }\n endTurnBook(s, giver);\n } catch (e) {\n // A flaky/slow model call shouldn't kill the whole game: forfeit the turn.\n rec.error = String(e).slice(0, 600);\n if (s.status === \"playing\") endTurnBook(s, giver);\n }\n}\n",[],[187,373],"d76b08c8-5544-4e2c-8358-854c4515798d","llm.ts","1174854e-8899-11f1-b138-1607ee4eb77e",94,["D",1785162766827],"import type { LlmOut, ModelCfg } from \"../shared/types.ts\";\n\nasync function chatOnce(model: ModelCfg, prompt: string): Promise {\n const body: Record = {\n model: model.id,\n messages: [{ role: \"user\", content: prompt }],\n max_tokens: model.maxTokens ?? 8000,\n usage: { include: true }, // OpenRouter returns exact cost in credits (USD)\n };\n if (model.reasoning) body.reasoning = model.reasoning;\n const ctrl = new AbortController();\n const t = setTimeout(() => ctrl.abort(), 140_000); // medium-effort reasoning calls can run long\n try {\n const r = await fetch(\"https://openrouter.ai/api/v1/chat/completions\", {\n method: \"POST\",\n headers: {\n Authorization: `Bearer ${Deno.env.get(\"OPENROUTER_API_KEY\")}`,\n \"Content-Type\": \"application/json\",\n \"X-Title\": \"codenames-duet-bench\",\n \"HTTP-Referer\": \"https://www.val.town/x/dcm31/codenames-duet-bench\",\n },\n body: JSON.stringify(body),\n signal: ctrl.signal,\n });\n const j = await r.json();\n if (!r.ok || j.error) {\n throw new Error(`OpenRouter ${r.status} for ${model.id}: ${JSON.stringify(j.error ?? j).slice(0, 400)}`);\n }\n const m = j.choices?.[0]?.message ?? {};\n const text = typeof m.content === \"string\" ? m.content : JSON.stringify(m.content ?? \"\");\n return {\n text,\n reasoning: typeof m.reasoning === \"string\" && m.reasoning.trim() ? m.reasoning : null,\n cost: Number(j.usage?.cost ?? 0),\n tokensIn: Number(j.usage?.prompt_tokens ?? 0),\n tokensOut: Number(j.usage?.completion_tokens ?? 0),\n };\n } finally {\n clearTimeout(t);\n }\n}\n\n/** retries with backoff on transient failures (network blips, free-tier 429s) */\nexport async function chat(model: ModelCfg, prompt: string): Promise {\n const delays = [2500, 9000];\n for (let i = 0; ; i++) {\n try {\n return await chatOnce(model, prompt);\n } catch (e) {\n if (i >= delays.length) throw e;\n await new Promise((r) => setTimeout(r, delays[i]));\n }\n }\n}\n\n/** tolerant JSON extraction from model output */\nexport function extractJson(text: string): Record | null {\n const tryParse = (s: string) => {\n try {\n const v = JSON.parse(s);\n return v && typeof v === \"object\" ? v as Record : null;\n } catch {\n return null;\n }\n };\n const direct = tryParse(text.trim());\n if (direct) return direct;\n const fenced = text.match(/```(?:json)?\\s*([\\s\\S]*?)```/);\n if (fenced) {\n const v = tryParse(fenced[1].trim());\n if (v) return v;\n }\n const s = text.indexOf(\"{\");\n const e = text.lastIndexOf(\"}\");\n if (s >= 0 && e > s) return tryParse(text.slice(s, e + 1));\n return null;\n}\n\n/** parse the model's JSON reply, falling back to the reasoning channel\n * (some reasoning models put everything there and leave content empty) */\nexport function parseOut(out: LlmOut): Record | null {\n return extractJson(out.text) ?? (out.reasoning ? extractJson(out.reasoning) : null);\n}\n",[],[187,382],"1b788f72-85d9-49c9-a879-a0701001575b","keycard.ts","f0aa6022-8898-11f1-a290-1607ee4eb77e",["D",1785032587913],"import type { Cell, KeyCard } from \"../shared/types.ts\";\n\n// Official Codenames Duet key card distribution over 25 words:\n// [sideA, sideB, count]\nconst DIST: Array<[Cell, Cell, number]> = [\n [\"green\", \"green\", 3],\n [\"green\", \"bystander\", 5],\n [\"bystander\", \"green\", 5],\n [\"green\", \"assassin\", 1],\n [\"assassin\", \"green\", 1],\n [\"assassin\", \"assassin\", 1],\n [\"assassin\", \"bystander\", 1],\n [\"bystander\", \"assassin\", 1],\n [\"bystander\", \"bystander\", 7],\n];\n// => each side sees 9 green / 3 assassin / 13 bystander; 15 unique agents total.\n\nexport function makeKeyCard(rng: () => number): KeyCard {\n const cells: Array<[Cell, Cell]> = [];\n for (const [a, b, n] of DIST) for (let i = 0; i < n; i++) cells.push([a, b]);\n for (let i = cells.length - 1; i > 0; i--) {\n const j = Math.floor(rng() * (i + 1));\n [cells[i], cells[j]] = [cells[j], cells[i]];\n }\n return { a: cells.map((c) => c[0]), b: cells.map((c) => c[1]) };\n}\n\nexport function totalAgents(key: KeyCard): number {\n let n = 0;\n for (let i = 0; i < 25; i++) if (key.a[i] === \"green\" || key.b[i] === \"green\") n++;\n return n; // always 15\n}\n",[],[187,390],"7381e918-8c03-4c22-88ad-c38db28abb87","game.ts","f8e58118-8898-11f1-9e75-1607ee4eb77e",48,["D",1785039020999],"import { WORDS } from \"../shared/words.ts\";\nimport type { Cell, GameMode, GameState, Side } from \"../shared/types.ts\";\nimport { makeKeyCard } from \"./keycard.ts\";\n\nexport const TOTAL_TURNS = 9;\nexport const TOTAL_AGENTS = 15; // duet\nexport const CLASSIC_AGENTS = 9; // classic one-key mode\nexport const MAX_GUESSES_PER_TURN = 8; // safety cap (Duet is unlimited-while-green)\n\nexport function agentsTarget(s: GameState): number {\n return s.mode === \"classic\" ? CLASSIC_AGENTS : TOTAL_AGENTS;\n}\n\n/** deterministic PRNG so games are reproducible from seed */\nexport function mulberry32(seed: number): () => number {\n let t = seed >>> 0;\n return function () {\n t += 0x6D2B79F5;\n let x = t;\n x = Math.imul(x ^ (x >>> 15), x | 1);\n x ^= x + Math.imul(x ^ (x >>> 7), x | 61);\n return ((x ^ (x >>> 14)) >>> 0) / 4294967296;\n };\n}\n\n/** classic Codenames single key: 9 green, 3 assassin, 13 bystander */\nfunction makeClassicKey(rng: () => number): Cell[] {\n const cells: Cell[] = [\n ...Array(9).fill(\"green\"),\n ...Array(3).fill(\"assassin\"),\n ...Array(13).fill(\"bystander\"),\n ];\n for (let i = cells.length - 1; i > 0; i--) {\n const j = Math.floor(rng() * (i + 1));\n [cells[i], cells[j]] = [cells[j], cells[i]];\n }\n return cells;\n}\n\nexport function newGame(seed: number, mode: GameMode = \"duet\"): GameState {\n const rng = mulberry32(seed);\n const pool = [...WORDS];\n for (let i = pool.length - 1; i > 0; i--) {\n const j = Math.floor(rng() * (i + 1));\n [pool[i], pool[j]] = [pool[j], pool[i]];\n }\n const words = pool.slice(0, 25);\n let key;\n if (mode === \"classic\") {\n const k = makeClassicKey(rng);\n key = { a: k, b: k }; // one shared key; giver is always side \"a\"\n } else {\n key = makeKeyCard(rng);\n }\n return {\n mode,\n seed,\n words,\n key,\n covered: Array(25).fill(false),\n marks: { a: Array(25).fill(false), b: Array(25).fill(false) },\n turn: 1,\n nextGiver: \"a\",\n status: \"playing\",\n agentsFound: 0,\n transcript: [],\n cost: 0,\n calls: 0,\n tokensIn: 0,\n tokensOut: 0,\n };\n}\n\nexport function other(s: Side): Side {\n return s === \"a\" ? \"b\" : \"a\";\n}\n\nexport function remainingGreens(s: GameState, side: Side): number[] {\n const out: number[] = [];\n for (let i = 0; i < 25; i++) {\n if (s.key[side][i] === \"green\" && !s.covered[i]) out.push(i);\n }\n return out;\n}\n\n/** Duet alternates givers (skipping one with no agents left). Classic: model A always gives. */\nexport function pickGiver(s: GameState): Side {\n if (s.mode === \"classic\") return \"a\";\n if (remainingGreens(s, s.nextGiver).length > 0) return s.nextGiver;\n return other(s.nextGiver);\n}\n\n/** Guesses are always judged against the CLUE GIVER's key. */\nexport function applyGuess(s: GameState, giver: Side, idx: number): \"agent\" | \"bystander\" | \"assassin\" {\n const cell = s.key[giver][idx];\n if (cell === \"green\") {\n s.covered[idx] = true;\n s.agentsFound++;\n if (s.agentsFound >= agentsTarget(s)) s.status = \"won\";\n return \"agent\";\n }\n if (cell === \"assassin\") {\n s.status = \"lost_assassin\";\n return \"assassin\";\n }\n s.marks[giver][idx] = true;\n return \"bystander\";\n}\n\nexport function legalGuess(s: GameState, giver: Side, idx: number): boolean {\n return idx >= 0 && idx < 25 && !s.covered[idx] && !s.marks[giver][idx];\n}\n\nexport function anyLegalGuess(s: GameState, giver: Side): boolean {\n for (let i = 0; i < 25; i++) if (legalGuess(s, giver, i)) return true;\n return false;\n}\n\nexport function endTurnBook(s: GameState, giver: Side): void {\n s.nextGiver = other(giver);\n if (s.status === \"playing\") {\n if (s.turn >= TOTAL_TURNS) s.status = \"lost_timeout\";\n else s.turn++;\n }\n}\n",[],[187,397],["D",1785032587913],[211,212,213,214,215],[187],"aa2f5a2e-b02f-4fc8-950b-d55380c7cfe9","standings.ts","5d252592-889a-11f1-ba0a-1607ee4eb77e",112,["D",1785188405496],17,"// Aggregates done games into standings.\n// classic games (the benchmark): one merged giver x guesser matrix + per-role\n// bar data. duet games are legacy/tucked-away.\nimport { init, sqlite } from \"./db.ts\";\nimport type { GameState } from \"../shared/types.ts\";\n\n/** Guess Efficiency Score for one classic game, as a 0-100 percentage.\n * You must WIN to beat any loss: raw = successful guesses (0-9) plus, on a\n * win, a bonus of (10 - clues used). Best possible = all 9 words on clue 1\n * -> raw 18 -> 100%. A win in all 9 clues -> 10/18 -> 56%. Losses cap at\n * 8/18 -> 44%, ranked by words found. */\nexport function gesPct(result: string, words: number, clues: number): number {\n const raw = words + (result === \"won\" ? Math.max(1, 10 - clues) : 0);\n return Math.round((raw / 18) * 100);\n}\n\nexport async function standings() {\n await init();\n const res = await sqlite.execute(\n `SELECT id, model_a, model_b, label_a, label_b, tier, mode, status, result, agents_found, turns_used, cost, calls, state\n FROM games ORDER BY id`,\n );\n\n const games = res.rows.map((r: any) => ({\n id: Number(r.id),\n model_a: String(r.model_a),\n model_b: String(r.model_b),\n label_a: String(r.label_a),\n label_b: String(r.label_b),\n tier: String(r.tier),\n mode: String(r.mode ?? \"duet\"),\n status: String(r.status),\n result: r.result == null ? null : String(r.result),\n agents_found: Number(r.agents_found),\n turns_used: Number(r.turns_used),\n cost: Number(r.cost),\n calls: Number(r.calls),\n state: r.state ? JSON.parse(String(r.state)) as GameState : undefined,\n }));\n\n const done = games.filter((g) => g.status === \"done\" && g.result && g.result !== \"error\");\n const doneDuet = done.filter((g) => g.mode === \"duet\");\n const doneClassic = done.filter((g) => g.mode === \"classic\");\n const classicAll = games.filter((g) => g.mode === \"classic\");\n\n // ============ classic: per-role stats ============\n type Role = {\n key: string; label: string; games: number; wins: number; agents: number; turns: number;\n ges: number;\n guesses: number; agentGuesses: number; assassins: number;\n };\n const givers = new Map();\n const guessers = new Map();\n const getR = (m: Map, key: string, label: string): Role => {\n if (!m.has(key)) m.set(key, { key, label, games: 0, wins: 0, agents: 0, turns: 0, ges: 0, guesses: 0, agentGuesses: 0, assassins: 0 });\n return m.get(key)!;\n };\n for (const g of doneClassic) {\n const gGes = gesPct(g.result!, g.agents_found, g.turns_used);\n const gv = getR(givers, g.model_a, g.label_a);\n gv.games++;\n if (g.result === \"won\") gv.wins++;\n if (g.result === \"lost_assassin\") gv.assassins++;\n gv.agents += g.agents_found;\n gv.turns += g.turns_used;\n gv.ges += gGes;\n const gu = getR(guessers, g.model_b, g.label_b);\n gu.games++;\n if (g.result === \"won\") gu.wins++;\n if (g.result === \"lost_assassin\") gu.assassins++;\n gu.agents += g.agents_found;\n gu.turns += g.turns_used;\n gu.ges += gGes;\n if (g.state) {\n for (const t of g.state.transcript) {\n for (const x of t.guesses) {\n gu.guesses++;\n if (x.outcome === \"agent\") gu.agentGuesses++;\n }\n }\n }\n }\n\n // ============ classic: three chart datasets (bias-controlled) ============\n // overall = SELF-PLAY only (model clues for itself; no partner bias)\n // giving = games cluing for the FIXED reference guesser (deepseek)\n // guessing = games guessing for the FIXED reference giver (gpt5mini)\n const REF_GUESSER = \"deepseek\";\n const REF_GIVER = \"gpt5mini\";\n // The headline benchmark: best model per lab (Charlie's pick, July 2026).\n // Everything else still plays but is tucked away in the UI.\n const FEATURED = [\"opus5\", \"sol\", \"g3pro\", \"kimik3\", \"dsv4\"];\n const labels: Record = {};\n for (const g of classicAll) {\n labels[g.model_a] = g.label_a;\n labels[g.model_b] = g.label_b;\n }\n type ChartRow = { key: string; label: string; ges: number; games: number; wins: number };\n const chartFrom = (rows: typeof doneClassic, keyOf: (g: typeof doneClassic[number]) => string): ChartRow[] => {\n const acc = new Map();\n for (const g of rows) {\n const k = keyOf(g);\n const a = acc.get(k) ?? { ges: 0, games: 0, wins: 0 };\n a.ges += gesPct(g.result!, g.agents_found, g.turns_used);\n a.games++;\n if (g.result === \"won\") a.wins++;\n acc.set(k, a);\n }\n return [...acc.entries()]\n .map(([key, a]) => ({ key, label: labels[key] ?? key, ges: a.ges / a.games, games: a.games, wins: a.wins }))\n .sort((x, y) => y.ges - x.ges);\n };\n const charts = {\n overall: chartFrom(doneClassic.filter((g) => g.model_a === g.model_b && FEATURED.includes(g.model_a)), (g) => g.model_a),\n giving: chartFrom(doneClassic.filter((g) => g.model_b === REF_GUESSER && FEATURED.includes(g.model_a)), (g) => g.model_a),\n guessing: chartFrom(doneClassic.filter((g) => g.model_a === REF_GIVER && FEATURED.includes(g.model_b)), (g) => g.model_b),\n };\n\n // (legacy per-role stats kept for the merged matrix ordering)\n const allKeys = new Set([...givers.keys(), ...guessers.keys()]);\n const bars = [...allKeys].map((key) => {\n const gv = givers.get(key);\n const gu = guessers.get(key);\n const giverGes = gv && gv.games ? gv.ges / gv.games : null;\n const guesserGes = gu && gu.games ? gu.ges / gu.games : null;\n const parts = [giverGes, guesserGes].filter((x): x is number => x !== null);\n const combined = parts.length ? parts.reduce((a, b) => a + b, 0) / parts.length : 0;\n return {\n key,\n label: labels[key] ?? key,\n giverGes,\n giverGames: gv?.games ?? 0,\n giverWins: gv?.wins ?? 0,\n guesserGes,\n guesserGames: gu?.games ?? 0,\n guesserWins: gu?.wins ?? 0,\n combined,\n games: (gv?.games ?? 0) + (gu?.games ?? 0),\n wins: (gv?.wins ?? 0) + (gu?.wins ?? 0),\n };\n }).sort((x, y) => y.combined - x.combined);\n\n // ============ classic: one merged giver x guesser matrix ============\n type Cell = {\n games: number; wins: number; agents: number; turns: number; assassins: number; cost: number; ges: number;\n gameList: { id: number; result: string; agents: number; turns: number; cost: number; ges: number }[];\n };\n const cells: Record> = {};\n for (const g of doneClassic) {\n cells[g.model_a] ??= {};\n const cell = (cells[g.model_a][g.model_b] ??= { games: 0, wins: 0, agents: 0, turns: 0, assassins: 0, cost: 0, ges: 0, gameList: [] });\n const gGes = gesPct(g.result!, g.agents_found, g.turns_used);\n cell.games++;\n if (g.result === \"won\") cell.wins++;\n if (g.result === \"lost_assassin\") cell.assassins++;\n cell.agents += g.agents_found;\n cell.turns += g.turns_used;\n cell.cost += g.cost;\n cell.ges += gGes;\n cell.gameList.push({ id: g.id, result: g.result!, agents: g.agents_found, turns: g.turns_used, cost: g.cost, ges: gGes });\n }\n // matrix axes ordered by combined strength; models with no finished games last\n const barOrder = bars.map((b) => b.key);\n const allMatrixKeys = [...new Set([...barOrder, ...classicAll.flatMap((g) => [g.model_a, g.model_b])])];\n const featuredKeys = allMatrixKeys.filter((k) => FEATURED.includes(k) || k === REF_GIVER || k === REF_GUESSER);\n const mkMerged = (keys: string[]) => ({\n keys,\n labels,\n cells,\n done: classicAll.filter((g) => g.status === \"done\").length,\n total: classicAll.length,\n });\n const merged = mkMerged(featuredKeys);\n const mergedFull = mkMerged(allMatrixKeys);\n\n // ============ duet (legacy, tucked away) ============\n type Team = {\n key: string; labels: [string, string]; keys: [string, string]; tier: string;\n games: number; wins: number; agents: number; turns: number; assassins: number; cost: number;\n gameList: { id: number; result: string; agents: number; turns: number; cost: number }[];\n };\n const teams = new Map();\n for (const g of doneDuet) {\n const [k1, k2] = [g.model_a, g.model_b].sort();\n const key = `${k1}|${k2}|${g.tier}`;\n if (!teams.has(key)) {\n const flip = k1 !== g.model_a;\n teams.set(key, {\n key,\n keys: [k1, k2],\n labels: flip ? [g.label_b, g.label_a] : [g.label_a, g.label_b],\n tier: g.tier,\n games: 0, wins: 0, agents: 0, turns: 0, assassins: 0, cost: 0, gameList: [],\n });\n }\n const t = teams.get(key)!;\n t.games++;\n if (g.result === \"won\") t.wins++;\n if (g.result === \"lost_assassin\") t.assassins++;\n t.agents += g.agents_found;\n t.turns += g.turns_used;\n t.cost += g.cost;\n t.gameList.push({ id: g.id, result: g.result!, agents: g.agents_found, turns: g.turns_used, cost: g.cost });\n }\n const teamRows = [...teams.values()]\n .map((t) => ({\n ...t,\n winRate: t.games ? t.wins / t.games : 0,\n avgAgents: t.games ? t.agents / t.games : 0,\n avgTurns: t.games ? t.turns / t.games : 0,\n }))\n .sort((x, y) => y.winRate - x.winRate || y.avgAgents - x.avgAgents);\n\n const pending = games.filter((g) => g.status !== \"done\").length;\n const featuredGame = (g: typeof doneClassic[number]) => FEATURED.includes(g.model_a) || FEATURED.includes(g.model_b);\n const recent = doneClassic.filter(featuredGame).slice(-12).reverse().map((g) => ({\n id: g.id,\n giver: g.label_a,\n guesser: g.label_b,\n result: g.result!,\n words: g.agents_found,\n clues: g.turns_used,\n ges: gesPct(g.result!, g.agents_found, g.turns_used),\n }));\n const totals = {\n games: games.length,\n done: done.length,\n pending,\n errors: games.filter((g) => g.result === \"error\").length,\n cost: games.reduce((s, g) => s + g.cost, 0),\n calls: games.reduce((s, g) => s + g.calls, 0),\n };\n\n return { charts, bars, merged, mergedFull, recent, teams: teamRows, totals };\n}\n",[],[192,408],"88fcb95f-7d0c-4e2b-991c-d3c0adcb4cc3","db.ts","21dab89a-8899-11f1-b138-1607ee4eb77e",95,["D",1785162771244],"import { sqlite } from \"https://esm.town/v/std/sqlite/main.ts\";\nimport type { GameState } from \"../shared/types.ts\";\n\nexport { sqlite };\n\nexport async function init(): Promise {\n await sqlite.execute(`CREATE TABLE IF NOT EXISTS games (\n id INTEGER PRIMARY KEY AUTOINCREMENT,\n model_a TEXT NOT NULL,\n model_b TEXT NOT NULL,\n label_a TEXT NOT NULL,\n label_b TEXT NOT NULL,\n tier TEXT NOT NULL,\n seed INTEGER NOT NULL,\n status TEXT NOT NULL DEFAULT 'queued',\n result TEXT,\n agents_found INTEGER DEFAULT 0,\n turns_used INTEGER DEFAULT 0,\n cost REAL DEFAULT 0,\n calls INTEGER DEFAULT 0,\n tokens_in INTEGER DEFAULT 0,\n tokens_out INTEGER DEFAULT 0,\n created_at TEXT NOT NULL,\n finished_at TEXT,\n state TEXT NOT NULL\n )`);\n try {\n await sqlite.execute(`ALTER TABLE games ADD COLUMN claimed_at TEXT`);\n } catch (_e) {\n // column already exists\n }\n try {\n await sqlite.execute(`ALTER TABLE games ADD COLUMN mode TEXT DEFAULT 'duet'`);\n } catch (_e) {\n // column already exists\n }\n}\n\n/** Atomically claim a game for this worker run. Returns true if we got it.\n * Claimable: queued, or running with a stale heartbeat (dead/zombie worker). */\nexport async function claim(id: number, staleMs = 330_000): Promise {\n const staleBefore = new Date(Date.now() - staleMs).toISOString();\n const r = await sqlite.execute({\n sql: `UPDATE games SET status='running', claimed_at=?\n WHERE id=? AND (status='queued' OR (status='running' AND (claimed_at IS NULL OR claimed_at < ?)))`,\n args: [new Date().toISOString(), id, staleBefore],\n });\n return r.rowsAffected > 0;\n}\n\nexport async function saveState(id: number, s: GameState): Promise {\n // status guard: a slow zombie worker must never resurrect a finished game\n await sqlite.execute({\n sql: `UPDATE games SET status='running', claimed_at=?, state=?, agents_found=?, turns_used=?, cost=?, calls=?, tokens_in=?, tokens_out=?\n WHERE id=? AND status != 'done'`,\n args: [new Date().toISOString(), JSON.stringify(s), s.agentsFound, s.transcript.length, s.cost, s.calls, s.tokensIn, s.tokensOut, id],\n });\n}\n\nexport async function finalize(id: number, s: GameState): Promise {\n await sqlite.execute({\n sql: `UPDATE games SET status='done', result=?, state=?, agents_found=?, turns_used=?, cost=?, calls=?, tokens_in=?, tokens_out=?, finished_at=?\n WHERE id=? AND status != 'done'`,\n args: [\n s.status,\n JSON.stringify(s),\n s.agentsFound,\n s.transcript.length,\n s.cost,\n s.calls,\n s.tokensIn,\n s.tokensOut,\n new Date().toISOString(),\n id,\n ],\n });\n}\n",[],[192,417],["D",1785032670439],[208,209],[192],["D",1785032481089],["D",1789474853273],{"_101":100,"_109":105},{"_101":100,"_102":103,"_104":105,"_18":56,"_106":107,"_108":25},{"_13":90,"_32":91,"_44":43,"_92":45,"_93":-5,"_94":-5,"_95":-5,"_96":-5,"_97":-5,"_98":46,"_99":100},"ctx",{"_83":-5,"_129":465,"_466":467,"_468":469,"_470":25,"_471":25,"_472":25,"_473":60,"_474":-5,"_475":-7,"_476":-7,"_477":64,"_478":25,"_479":25,"_480":-5,"_481":25,"_482":25,"_483":25,"_484":-5},"ENV",{"_436":437,"_438":439,"_440":91,"_441":442,"_443":444,"_445":446,"_447":448,"_449":450,"_451":452,"_453":454,"_455":456,"_457":458,"_459":460,"_461":462,"_463":464},"REACT_APP_SENTRY_DSN","https://586020d86f05489db63a554690acf1e9@o1398359.ingest.sentry.io/6724499","RENDER_GIT_COMMIT","e87eb04e707cf64e7e532c8e8ec395f209e17f15","RENDER_GIT_BRANCH","POSTHOG_API_TOKEN","phc_nmoyuAkStEZuUp69rYlg1DZgrIt19Gx0Y4Hwf8tGgJP","POSTHOG_API_HOST","https://wart.val.town","WEB_API_TEMPLATE","https://{val}.web.val.run","APP_URL","https://www.val.town","MODULE_URL","https://esm.town","VANITY_WEB_API_TEMPLATE","https://{label}.val.run/","DENOLS_URL","https://lsp.val.town","NODE_ENV","production","CLOUDFLARE_TURNSTILE_SITE_KEY","0x4AAAAAACtaIZJH7s9rBMZ6","GOOGLE_CLIENT_ID","875989696899-jtsqkhlbdv5jrlsk8631p7ipjg3c6o62.apps.googleusercontent.com","STRIPE_PUBLISHABLE_KEY","pk_live_51N2zvWEpZTTLipIkVEwXrUG9AT0yeOOEYxszarZT1jpsZV58Ly6wlg3SltZpHN26CUalOBmUSY7adQRQiSkCZz2500qSXNkf8Y","API_URL","https://api.val.town",[],"preferencesKeybindings","default","preferencesTheme","light","preferencesFixers","preferencesCopilot","preferencesSelectionAi","preferencesCloseBrackets","preferencesChatPanelState","preferencesDprint","preferencesDenofmt","notificationCount","flagDenols","isBusinessOrgMember","businessOrgHandle","isBusinessOrgAdmin","isHigherTownieTier","navCollapsed","currentOrgId"]

Diff · v116