heyspex 0.1.0 → 0.2.0

package/dev-factory/model-catalog.json

@@ -0,0 +1,222 @@
+{
+  "_comment": "SINGLE SOURCE OF TRUTH for model ids, aliases, retirements, context windows, capabilities and list prices — every cloud provider the fleet can call (DeepSeek, Anthropic, OpenAI, Gemini, OpenRouter, GitHub Models) AND every local runtime (Ollama, vLLM). Hand-edit THIS file when a vendor ships or retires a model; never scatter a new literal id through the code. Readers: dev-factory/lib/model-catalog.mjs (typed accessors + self-test), dev-factory/cloud-clients.mjs (the whole PROVIDERS table is derived from `providerKeys` below), dev-factory/lib/fleet-manifest.mjs (context/VRAM/cost tables), spexharness/src/chat/compact.mjs (window sizes), spex-console/src/lib/model-catalog.ts (the console's ONLY import path — keep it so when dev-factory merges into SpexFleet one line moves), dev-factory/judge.mjs, dev-factory/router.mjs --self-test (refuses retired ids in routing-table.json), scripts/check-model-ids.mjs (refuses retired ids in live code). ROUTING (which seat runs which model, which machine hosts which local model) stays in routing-table.json — this file says what a model IS and how it is reached, not who uses it. UPGRADE RECIPE: (1) add the new id under `models` with its window/price; (2) move the old id into `aliases` (old → new) and `retired`; (3) repoint the `roles` / `providerKeys` you want moved; (4) run `node dev-factory/lib/model-catalog.mjs --self-test && node dev-factory/router.mjs --self-test && npm run check:model-ids`. `verifiedOn` per provider records when the prices were last checked against the vendor's sheet.",
+  "verified": "2026-09-18",
+  "cloud": {
+    "deepseek": {
+      "kind": "openai",
+      "company": "deepseek",
+      "baseUrl": "https://api.deepseek.com/v1",
+      "betaBaseUrl": "https://api.deepseek.com/beta",
+      "anthropicBaseUrl": "https://api.deepseek.com/anthropic",
+      "userIdField": "user_id",
+      "_userIdFieldNote": "Wire field for per-seat cache/scheduling isolation (DeepSeek rate-limit doc: user_id isolates KVCache and scheduling). cloud-clients.mjs sends opts.userId (defaulting to opts.seat) under this name only for providers that declare it; betaBaseUrl is used only when a caller opts into strict tool schemas (opts.strictTools) — DeepSeek's strict mode is served from /beta.",
+      "keyEnv": "HEYSPEX_DEEPSEEK_API_KEY",
+      "keyEnvFallback": "DEEPSEEK_API_KEY",
+      "source": "https://api-docs.deepseek.com/quick_start/pricing",
+      "verifiedOn": "2026-09-18",
+      "peakHoursUtc": { "days": [1, 2, 3, 4, 5], "ranges": [["01:00", "04:00"], ["06:00", "10:00"]], "note": "Off-peak rates are half of peak; all other hours are off-peak (DeepSeek pricing note 3)." },
+      "roles": {
+        "cheap": "deepseek-flash",
+        "strong": "deepseek-v4-pro",
+        "chatDefault": "deepseek-flash",
+        "judge": "deepseek-flash",
+        "legacyJudge": "deepseek-chat"
+      },
+      "models": {
+        "deepseek-flash": {
+          "displayName": "DeepSeek-V4.1-Flash",
+          "released": "2026-09-10",
+          "contextWindow": 1048576,
+          "maxOutputTokens": 393216,
+          "tier": "cheap",
+          "capabilities": ["api", "metered", "tools", "thinking", "json", "vision", "strict-tools", "fim", "prefix"],
+          "thinkingDefault": true,
+          "effortLevels": ["low", "high", "max"],
+          "concurrencyLimit": 2500,
+          "pricePer1M": {
+            "peak":    { "cacheHit": 0.006, "cacheMiss": 0.30, "output": 1.20 },
+            "offPeak": { "cacheHit": 0.003, "cacheMiss": 0.15, "output": 0.60 }
+          }
+        },
+        "deepseek-v4-pro": {
+          "displayName": "DeepSeek-V4-Pro-0813",
+          "released": "2026-08-13",
+          "contextWindow": 1048576,
+          "maxOutputTokens": 393216,
+          "tier": "strong",
+          "capabilities": ["api", "metered", "tools", "thinking", "json", "strict-tools", "fim", "prefix", "max_reasoning"],
+          "thinkingDefault": true,
+          "effortLevels": ["low", "high", "max"],
+          "concurrencyLimit": 500,
+          "note": "Retirement announced 2026-09-10 was reversed; service continues past 2026-09-14 at unchanged billing (pricing note 2).",
+          "pricePer1M": {
+            "peak":    { "cacheHit": 0.044, "cacheMiss": 1.32, "output": 3.96 },
+            "offPeak": { "cacheHit": 0.022, "cacheMiss": 0.66, "output": 1.98 }
+          }
+        }
+      },
+      "aliases": {
+        "deepseek-v4-flash": "deepseek-flash",
+        "deepseek-v4-flash-vision-exp": "deepseek-flash",
+        "deepseek-chat": "deepseek-flash"
+      },
+      "retired": [
+        { "id": "deepseek-v4-flash", "on": "2026-09-10", "servedBy": "deepseek-flash", "note": "Legacy name still accepted by the API and routed to V4.1 Flash at the Flash price 'temporarily'." },
+        { "id": "deepseek-v4-flash-vision-exp", "on": "2026-09-10", "servedBy": "deepseek-flash" }
+      ],
+      "unlisted": {
+        "deepseek-chat": "Not on the 2026-09-10 model sheet. Kept only as the held-constant LLM-judge alias (PROVIDERS key `deepseek`); priced as deepseek-flash. Verify with --liveness before relying on it."
+      },
+      "emitAliasesAsProviderKeys": true,
+      "providerKeys": {
+        "deepseek-flash":  { "model": "deepseek-flash" },
+        "deepseek-v4-pro": { "model": "deepseek-v4-pro" },
+        "deepseek":        { "model": "deepseek-chat", "legacyAliasOf": "deepseek-flash", "note": "Legacy alias, held-constant LLM judge. Sends `deepseek-chat` on the wire, priced as deepseek-flash." }
+      }
+    },
+    "anthropic": {
+      "kind": "anthropic",
+      "company": "anthropic",
+      "baseUrl": "https://api.anthropic.com/v1",
+      "keyEnv": "HEYSPEX_ANTHROPIC_API_KEY",
+      "source": "claude-api skill model table (cached 2026-06-24) — first-party API rates",
+      "verifiedOn": "2026-06-24",
+      "_note": "Anthropic models are excluded from the dev-factory council by construction and are the ceiling/reference arm in benches; the PROVIDERS keys `claude` / `claude-opus` still point at the 4.x generation on purpose (flip `providerKeys` to the 5-family when the owner rules on it — the ids and prices are already here). Claude Code CLI launches use the bare alias words in `cliAliases` (resolved by spex-console/src/lib/model-registry/launch-resolve.ts), not these ids.",
+      "roles": {
+        "default": "claude-opus-5",
+        "strong": "claude-opus-5",
+        "mid": "claude-sonnet-5",
+        "cheap": "claude-haiku-4-5",
+        "max": "claude-fable-5-1",
+        "previousStrong": "claude-opus-4-8",
+        "previousMid": "claude-sonnet-4-6"
+      },
+      "cliAliases": { "opus": "claude-opus-5", "sonnet": "claude-sonnet-5", "haiku": "claude-haiku-4-5", "fable": "claude-fable-5-1" },
+      "models": {
+        "claude-fable-5-1": { "displayName": "Claude Fable 5.1", "contextWindow": 1048576, "maxOutputTokens": 131072, "tier": "max",    "capabilities": ["api", "metered", "tools", "thinking", "vision", "json"], "noTemp": true, "thinkingAlwaysOn": true, "pricePer1M": { "input": 10.00, "output": 50.00 } },
+        "claude-opus-5":    { "displayName": "Claude Opus 5",    "contextWindow": 1048576, "maxOutputTokens": 131072, "tier": "strong", "capabilities": ["api", "metered", "tools", "thinking", "vision", "json"], "noTemp": true, "pricePer1M": { "input": 5.00,  "output": 25.00 } },
+        "claude-sonnet-5":  { "displayName": "Claude Sonnet 5",  "contextWindow": 1048576, "maxOutputTokens": 131072, "tier": "strong", "capabilities": ["api", "metered", "tools", "thinking", "vision", "json"], "noTemp": true, "pricePer1M": { "input": 2.00,  "output": 10.00 } },
+        "claude-haiku-4-5": { "displayName": "Claude Haiku 4.5", "contextWindow": 200000,  "maxOutputTokens": 65536,  "tier": "cheap",  "capabilities": ["api", "metered", "tools", "thinking", "vision"], "pricePer1M": { "input": 1.00,  "output": 5.00 } },
+        "claude-opus-4-8":  { "displayName": "Claude Opus 4.8",  "contextWindow": 1048576, "maxOutputTokens": 131072, "tier": "strong", "capabilities": ["api", "metered", "tools", "thinking", "vision", "json"], "noTemp": true, "generation": "previous", "pricePer1M": { "input": 5.00,  "output": 25.00 } },
+        "claude-sonnet-4-6":{ "displayName": "Claude Sonnet 4.6","contextWindow": 1048576, "maxOutputTokens": 131072, "tier": "strong", "capabilities": ["api", "metered", "tools", "thinking", "vision", "json"], "generation": "previous", "pricePer1M": { "input": 3.00,  "output": 15.00 } }
+      },
+      "aliases": {},
+      "retired": [],
+      "providerKeys": {
+        "claude":      { "model": "claude-sonnet-4-6", "note": "Generic Claude key — previous generation kept on purpose; see _note." },
+        "claude-opus": { "model": "claude-opus-4-8", "noTemp": true, "note": "Opus key — previous generation kept on purpose. Old table priced this at $15/$75; the sheet says $5/$25." }
+      }
+    },
+    "openai": {
+      "kind": "openai",
+      "company": "openai",
+      "baseUrl": "https://api.openai.com/v1",
+      "keyEnv": "HEYSPEX_OPENAI_API_KEY",
+      "source": "carried over from cloud-clients.mjs PROVIDERS (no vendor re-verification in this pass)",
+      "verifiedOn": "2026-06-28",
+      "roles": { "default": "gpt-4o" },
+      "models": {
+        "gpt-4o": { "contextWindow": 128000, "tier": "strong", "capabilities": ["api", "metered", "tools", "vision", "json"], "pricePer1M": { "input": 2.50, "output": 10.00 } }
+      },
+      "aliases": {},
+      "retired": [],
+      "providerKeys": { "openai": { "model": "gpt-4o" } }
+    },
+    "gemini": {
+      "kind": "openai",
+      "company": "google",
+      "baseUrl": "https://generativelanguage.googleapis.com/v1beta/openai",
+      "keyEnv": "HEYSPEX_GEMINI_API_KEY",
+      "source": "carried over from cloud-clients.mjs PROVIDERS (no vendor re-verification in this pass)",
+      "verifiedOn": "2026-06-28",
+      "roles": { "default": "gemini-2.5-pro", "strong": "gemini-2.5-pro" },
+      "models": {
+        "gemini-2.5-pro": { "contextWindow": 1048576, "tier": "strong", "capabilities": ["api", "metered", "tools", "vision", "json", "thinking"], "pricePer1M": { "input": 1.25, "output": 10.00 } }
+      },
+      "aliases": {},
+      "retired": [],
+      "providerKeys": { "gemini": { "model": "gemini-2.5-pro" } }
+    },
+    "openrouter": {
+      "kind": "openai",
+      "company": "openrouter",
+      "baseUrl": "https://openrouter.ai/api/v1",
+      "keyEnv": "OPENROUTER_API_KEY",
+      "source": "carried over from cloud-clients.mjs PROVIDERS (OpenRouter list prices as dated per model)",
+      "verifiedOn": "2026-07-10",
+      "_note": "Same-family fallbacks hosted via OpenRouter (council key resilience, 2026-07-08 dogfood audit): a customer fleet with ONE OpenRouter key can still seat all three decorrelated council families — the MODEL family is what decorrelates blind spots, not the API host. grok-4.3 is the decorrelated director-seat primary (owner directive 2026-07-10, HS-E1O0); glm-5.2 is the 'beat Opus cheaper' candidate (HS-XNJI), not yet wired into any tier.",
+      "roles": { "cheap": "openai/gpt-4o-mini", "director": "x-ai/grok-4.3", "directorFallback": "z-ai/glm-5.2", "geminiFallback": "google/gemini-2.5-pro", "gpt4oFallback": "openai/gpt-4o" },
+      "models": {
+        "openai/gpt-4o-mini":    { "contextWindow": 128000,  "tier": "cheap",  "capabilities": ["api", "metered", "tools", "json"], "pricePer1M": { "input": 0.15, "output": 0.60 } },
+        "openai/gpt-4o":         { "contextWindow": 128000,  "tier": "strong", "capabilities": ["api", "metered", "tools", "vision", "json"], "pricePer1M": { "input": 2.50, "output": 10.00 } },
+        "google/gemini-2.5-pro": { "contextWindow": 1048576, "tier": "strong", "capabilities": ["api", "metered", "tools", "vision", "json"], "pricePer1M": { "input": 1.25, "output": 10.00 } },
+        "z-ai/glm-5.2":          { "displayName": "GLM-5.2 (Z.ai, open-weight 744B)", "contextWindow": 200000, "tier": "strong", "capabilities": ["api", "metered", "tools", "json"], "verifiedOn": "2026-06-30", "pricePer1M": { "input": 0.95, "output": 3.00 } },
+        "x-ai/grok-4.3":         { "displayName": "Grok 4.3 (xAI)", "contextWindow": 256000, "tier": "strong", "capabilities": ["api", "metered", "tools", "json", "thinking", "structured-outputs"], "verifiedOn": "2026-07-10", "pricePer1M": { "input": 1.25, "output": 2.50 } }
+      },
+      "aliases": {},
+      "retired": [],
+      "providerKeys": {
+        "openrouter":        { "model": "openai/gpt-4o-mini" },
+        "gemini-openrouter": { "model": "google/gemini-2.5-pro" },
+        "gpt4o-openrouter":  { "model": "openai/gpt-4o" },
+        "glm-5.2":           { "model": "z-ai/glm-5.2" },
+        "grok-4.3":          { "model": "x-ai/grok-4.3" }
+      }
+    },
+    "github-models": {
+      "kind": "openai",
+      "company": "github",
+      "baseUrl": "https://models.github.ai/inference",
+      "keyEnv": "HEYSPEX_GITHUB_TOKEN",
+      "keyEnvFallback": "GITHUB_TOKEN",
+      "source": "GitHub Models bills via Copilot credits; gpt-4o list price mirrored for the ledger (estimate)",
+      "verifiedOn": "2026-06-28",
+      "_note": "Copilot-ecosystem credits as a cloud tier (HS-32OM fleet onboarding). Model ids are publisher-namespaced.",
+      "roles": { "default": "openai/gpt-4o" },
+      "models": {
+        "openai/gpt-4o": { "contextWindow": 128000, "tier": "strong", "capabilities": ["api", "metered", "tools", "vision", "json"], "pricePer1M": { "input": 2.50, "output": 10.00 } }
+      },
+      "aliases": {},
+      "retired": [],
+      "providerKeys": { "github-models": { "model": "openai/gpt-4o" } }
+    }
+  },
+  "local": {
+    "ollama": {
+      "kind": "ollama",
+      "_comment": "Local fleet models. WHERE each fits (canHost) and WHO uses it (purposes) live in routing-table.json; per-machine speeds live in fleet-manifest.mjs STATIC_SPEEDS. Ids are Ollama tags exactly as `ollama list` prints them. Local models have no PROVIDERS entry — callModel treats an unknown id as an Ollama id.",
+      "roles": {
+        "builder": "gpt-oss:20b",
+        "coder": "qwen3-coder:30b",
+        "judgeFallback": "qwen3-coder:30b",
+        "vision": "qwen3.6:27b",
+        "small": "qwen3:8b"
+      },
+      "models": {
+        "gpt-oss:20b":       { "contextWindow": 131072, "vramRequiredMb": 13800, "architecture": "gptoss",   "capabilities": ["tools", "thinking"] },
+        "qwen3-coder:30b":   { "contextWindow": 262144, "vramRequiredMb": 18600, "architecture": "qwen3moe", "capabilities": ["tools"] },
+        "qwen3.6:27b":       { "contextWindow": 262144, "vramRequiredMb": 17400, "architecture": "qwen35",   "capabilities": ["tools", "thinking", "vision"] },
+        "qwen2.5-coder:32b": { "contextWindow": 131072, "vramRequiredMb": 19900, "architecture": "qwen2",    "capabilities": ["tools", "insert"] },
+        "qwen3:8b":          { "contextWindow": 32768,  "vramRequiredMb": 5200,  "architecture": "qwen3",    "capabilities": ["tools", "thinking"] },
+        "deepseek-r1:70b":   { "contextWindow": 131072, "vramRequiredMb": 42500, "architecture": "deepseek", "capabilities": ["tools", "thinking"], "note": "Q2_K quant — lower quality than Q4_K_M." },
+        "llama3.3:70b":      { "contextWindow": 131072, "vramRequiredMb": 42500, "architecture": "llama3",   "capabilities": ["tools"] },
+        "VibeThinker-3B":    { "contextWindow": 8192,   "vramRequiredMb": 3300,  "architecture": "qwen2",    "capabilities": [], "note": "Fits anywhere ≤12GB." }
+      },
+      "aliases": {},
+      "retired": []
+    },
+    "vllm": {
+      "kind": "openai",
+      "_comment": "Night Action vLLM + DSpark pilot (RTX 3060, matched draft `deepseek-ai/dspark_qwen3_8b_block7`). AWQ quant target (bf16 ~16GB does not fit 12GB VRAM alongside the draft; AWQ ~5.6GB does). Reference: HS-XNJI Experiment 1/2. Local, free, no key — vLLM OpenAI-compat endpoint on port 8000.",
+      "baseUrl": "http://192.168.1.252:8000/v1",
+      "keyEnv": null,
+      "roles": { "default": "Qwen/Qwen3-8B-AWQ" },
+      "models": {
+        "Qwen/Qwen3-8B-AWQ": { "contextWindow": 32768, "vramRequiredMb": 5600, "architecture": "qwen3", "tier": "local", "capabilities": ["api", "tools", "thinking"] }
+      },
+      "aliases": {},
+      "retired": [],
+      "providerKeys": { "qwen3-8b-dspark": { "model": "Qwen/Qwen3-8B-AWQ", "local": true } }
+    }
+  }
+}