release: 3.40.0 — surface the models 3.39.0 priced but never exposed

CHANGELOG.md

@@ -1,5 +1,116 @@
 # Changelog
 
+## Franklin Agent 3.40.0 — priced but unreachable: the catalog gets a front door
+
+**3.39.0 taught Franklin what the new models cost. It never taught anyone how
+to ask for them.** That release added eleven chat models to `src/pricing.ts`
+and stopped there — no shortcut, no picker row, no router wiring. The models
+were live on the gateway, correctly priced, and reachable only by typing the
+full `provider/model` id from memory. This release finishes the job.
+
+**38 new shortcuts (107 → 145), covering every model 3.39.0 priced:**
+
+| family | now reachable as |
+|---|---|
+| GPT-5.6 pro reasoning tier | `sol-pro`, `terra-pro`, `luna-pro` |
+| GPT-5.5 Pro / ChatGPT default | `gpt-5.5-pro`, `chatgpt` / `instant` |
+| Gemini 3.6 Flash + the Flash Lite line | `flash` (now 3.6), `flash-lite`, `gemini-3.1-flash-lite` |
+| Qwen3.7 Plus / Flash | `qwen-plus`, `qwen-flash` |
+| Tencent, Xiaomi (new providers) | `hy3` / `tencent`, `mimo` / `xiaomi` |
+| GPT-4o + 4.1 mini/nano, o3-mini | `4o`, `gpt-4o-mini`, `gpt-4.1-mini`, `o3-mini` |
+
+Worth knowing about the pro tier: **Terra Pro ($1/$6) and Luna Pro
+($0.1/$0.6) undercut their own base tiers** while adding pro reasoning mode.
+Sol Pro matches Sol exactly. `gpt` stays pinned to Sol — bare aliases follow
+the gateway's flagship, not the cheapest sibling.
+
+**New arrivals since 3.39.0.** `zai/glm-5.3` is Z.AI's flagship (1M context,
+always-on reasoning, priced at 5.2's $1.4/$4.4) and `xai/grok-imagine-video-1.5`
+joins VideoGen at $0.08/s.
+
+One catalog fact worth writing down: **GLM-5.3 is Base-only.** The Solana
+gateway lists 92 of Base's 93 models and 5.3 is the one it's missing (prices
+are otherwise identical across chains). So `glm` stays pinned to 5.2 rather
+than following the flagship — a bare alias is what people type from muscle
+memory and has to resolve on *both* chains; pointing it at a Base-only id
+would hand every Solana user an HTTP 400. 5.3 is one keystroke away as
+`glm-5.3` and has its own picker row, which reconciliation drops automatically
+for Solana users.
+
+**The router's LLM classifier has been dead since 2026-07-27, silently.**
+Its default model was `nvidia/qwen3-next-80b-a3b-instruct` — the same id
+NVIDIA EOL'd, which the gateway now rides on `nemotron-3-super-120b`. That
+substitute opens with *"Okay, let's see. The user wants to…"* and never
+reaches a verdict inside the classifier's 16-token budget, so every
+classification failed the strict parse and fell through to keyword-only
+routing. The fallback is invisible by design, which is exactly why this went
+three weeks without anyone noticing. Re-probed the free pool on the real
+classifier prompt:
+
+| free model | reply to a classification prompt |
+|---|---|
+| `qwen3-next` (old default) | prose, served by `nemotron-3-super-120b` |
+| `nemotron-nano-9b-v2` | prose — *"Okay, let's see…"* |
+| `nemotron-3-nano-omni` | **`MEDIUM`** — one bare word, served by itself |
+
+`nemotron-3-nano-omni` is the new classifier. The lesson worth keeping: "good
+free chat model" and "answers in one bare word under a tight cap" are
+different requirements, and the free default is not automatically both.
+
+**The 30B omni model started serving itself again.** In August it came back
+as `gpt-oss-120b` — the pooled-substitute trap — which is why it had no alias
+and led no chain. Re-probed 2026-08-19 on both the streaming and non-streaming
+paths, it now answers as itself. It gets a shortcut (`omni`), a free picker
+row, and second place in every free chain; the still-degraded
+`mistral-nemotron` slides to third but stays as a genuinely different family.
+`nvidia/step-3.7-flash` remains deliberately unaliased — it is billed at $0
+and listed in the catalog, but live probes come back served by
+`nemotron-3-super-120b`, and an alias that promises a model you don't get is
+worse than no alias.
+
+**Picker stays at its 24-row cap — new models displaced old ones.** GLM-5.3
+takes GLM-5.2's slot. GPT-5.6 Terra Pro takes the row
+`grok-4-1-fast-reasoning` held, which the gateway hides and reconciliation
+dropped on every live render anyway. Qwen3.7 Flash ($0.03/$0.13, 1M context,
+the cheapest paid model on the gateway) takes GLM-5's budget slot. Gemini 2.5
+Pro's row retires from directly under 3.1 Pro. Every retired row keeps its
+shortcut, per the long-standing hide-the-row-keep-the-alias rule, and `gemini`
+now tracks the 3.1 Pro flagship like `gpt`, `grok` and `glm` already did.
+
+**Also:** the offline picker fallback had stale prices (deepseek at
+$0.20/$0.40, corrected to $0.14/$0.28 — live renders were already right since
+the gateway is the source of truth for price). Brand numbers re-synced: 95
+models visible, 8 video, 5 free. Model-family guidance now recognises
+`chat-latest` as a strong model and routes Tencent/Xiaomi to balanced
+guidance.
+
+**Second pass — the tables the aliases exposed.** Surfacing the models turned
+up three hand-curated tables that had drifted behind the catalog:
+
+- **`xai/grok-4.5` was missing from the vision allowlist.** It is what `grok`
+  resolves to, and it is vision-capable — so every image turn on the xAI
+  flagship was quietly rerouted to a "vision sibling" the user never asked
+  for. Eleven other vision models were missing too (the GPT-5.6 pro tier,
+  `chat-latest`, `gpt-5.3`, `gpt-4o`, Gemini 3.6 Flash, the free omni model).
+  A new test pins the invariant: every bare flagship alias must be in the
+  allowlist, so the next flagship fails CI instead of someone's session.
+- **Context windows.** `grok-4.5` (500K), `grok-4.3` (1M) and `glm-5.3` (1M)
+  had no entry and match no inference pattern, so a cold catalog cache fell
+  through to the blind 128k default — compacting a 1M window eight times too
+  early. Twelve entries added.
+- **Model-family guidance.** The weak-model branch matched on bare `glm`,
+  which dates from GLM-4.x. `glm` now resolves to GLM-5.3 — 1M context,
+  always-on reasoning, priced above Gemini 3.1 Pro on input — and it was being
+  told to make ONE tool call and stay under 300 words. GLM-5.x moves to
+  balanced guidance.
+- The Brain's third extraction fallback was `nvidia/nemotron-super-49b`, gone
+  from the catalog; now the current free default.
+
+All 647 local tests pass. The five hidden-but-resolvable ids Franklin still
+pins (`opus-4.6`, `gpt-5-nano`, `grok-3`, `grok-4-0709`,
+`grok-4-1-fast-reasoning`) were each re-probed and still return 402, not 400 —
+they stay.
+
 ## Franklin Agent 3.39.0 — the gateway moved for three weeks and Franklin didn't
 
 **Franklin's free default was a dead model, and the whole catalog had drifted.**

brand-numbers.json

@@ -3,22 +3,22 @@
   "version": 1,
   "models": {
     "chatVisible": 71,
-    "totalVisible": 93,
-    "free": 6,
-    "freeWithheld": 19,
+    "totalVisible": 95,
+    "free": 5,
+    "freeWithheld": 20,
     "image": 9,
-    "video": 6,
+    "video": 8,
     "music": 1,
     "speech": 5,
     "soundfx": 1,
-    "withFallback": 46,
-    "withFallbackAllEntries": 79
+    "withFallback": 44,
+    "withFallbackAllEntries": 78
   },
   "clawrouter": {
     "dimensions": 15,
     "tiers": 4,
     "profiles": 4,
-    "aliases": 202
+    "aliases": 229
   },
   "mcp": {
     "tools": 20

package.json

@@ -1,6 +1,6 @@
 {
   "name": "@blockrun/franklin",
-  "version": "3.39.0",
+  "version": "3.40.0",
   "description": "Franklin Agent — The AI agent with a wallet. Spends USDC autonomously to get real work done. Pay per action, no subscriptions.",
   "type": "module",
   "exports": {

src/agent/context.ts

@@ -254,7 +254,7 @@ You run on the BlockRun AI Gateway. When the user asks you to "test the BlockRun
 - \`GET /.well-known/x402\` — x402 resource list with prices
 
 **LLM (POST, x402-paid)**
-- \`POST /v1/chat/completions\` — OpenAI-compatible. Body: \`{ model, messages, stream?, tools?, max_tokens?, temperature? }\`. \`model\` MUST come from \`GET /v1/models\` (real frontier examples on the gateway, verified live 2026-08-12: \`anthropic/claude-sonnet-5\`, \`anthropic/claude-opus-5\`, \`deepseek/deepseek-v4-pro\`, \`zai/glm-5.2\`, \`xai/grok-4.5\`, \`nvidia/nemotron-nano-9b-v2\` (free)). Do NOT invent versions like \`openai/gpt-5.1\` or \`xai/grok-5\` — those don't exist; the gateway 400s with the valid list in the error body, so when in doubt fetch \`GET /v1/models\` first.
+- \`POST /v1/chat/completions\` — OpenAI-compatible. Body: \`{ model, messages, stream?, tools?, max_tokens?, temperature? }\`. \`model\` MUST come from \`GET /v1/models\` (real frontier examples on the gateway, verified live 2026-08-19: \`anthropic/claude-sonnet-5\`, \`anthropic/claude-opus-5\`, \`openai/gpt-5.6-sol\`, \`deepseek/deepseek-v4-pro\`, \`zai/glm-5.3\`, \`xai/grok-4.5\`, \`qwen/qwen3.7-flash\`, \`nvidia/nemotron-nano-9b-v2\` (free)). Do NOT invent versions like \`openai/gpt-5.1\` or \`xai/grok-5\` — those don't exist; the gateway 400s with the valid list in the error body, so when in doubt fetch \`GET /v1/models\` first.
 - \`POST /v1/messages\` — Anthropic-compatible. Body: \`{ model, messages, max_tokens, system?, tools? }\`.
 
 **Media (POST, x402-paid; GET to poll async jobs)**
@@ -548,8 +548,14 @@ export function assembleInstructions(workingDir: string, model?: string): string
 export function getModelGuidance(model: string): string {
   const m = model.toLowerCase();
 
-  // Weak/cheap models: strict discipline to prevent looping and hallucination
-  if (m.includes('glm') || m.includes('gpt-oss') || m.includes('nemotron') ||
+  // Weak/cheap models: strict discipline to prevent looping and hallucination.
+  // The bare `glm` match dates from the GLM-4.x era. The paid Z.AI GLM-5 line
+  // is a different animal — 1M context, always-on reasoning, priced above
+  // Gemini 3.1 Pro on input — and `glm` now resolves to 5.3, so matching it
+  // here was telling a flagship to make ONE tool call and stay under 300
+  // words. GLM-5.x moves to the balanced branch; glm-4.7 and friends stay.
+  if ((m.includes('glm') && !m.includes('glm-5')) ||
+      m.includes('gpt-oss') || m.includes('nemotron') ||
       m.includes('minimax') || m.includes('devstral') || m.includes('llama-4')) {
     return `# Execution Discipline (strict — this model requires guardrails)
 - Make ONE tool call per task. Do NOT retry the same tool with query variations.
@@ -560,10 +566,13 @@ export function getModelGuidance(model: string): string {
   }
 
   // Medium models: balanced guidance. The bare `qwen` match dates from when
-  // every qwen id on the gateway was a free NVIDIA SKU — qwen3.7-max is a paid
-  // 1M-context flagship and belongs in the strong branch below.
+  // every qwen id on the gateway was a free NVIDIA SKU. The paid Qwen line is
+  // now Max + Plus + Flash — all 1M-context with reasoning — so only the Max
+  // flagship graduates to the strong branch; Plus and Flash stay here, and the
+  // legacy free `nvidia/qwen*` ids keep matching as before.
   if (m.includes('kimi') || m.includes('grok') || m.includes('flash') ||
       m.includes('haiku') || m.includes('deepseek') ||
+      m.includes('hy3') || m.includes('mimo') || m.includes('glm-5') ||
       (m.includes('qwen') && !m.includes('qwen3.7-max'))) {
     return `# Execution Guidance
 - Use tools to verify facts before stating them. Do not answer from memory when a tool can confirm.
@@ -576,7 +585,7 @@ export function getModelGuidance(model: string): string {
   if (m.includes('claude') || m.includes('gpt-5') || m.includes('opus') ||
       m.includes('sonnet') || m.includes('gemini-2.5-pro') || m.includes('gemini-3') ||
       m.includes('o3') || m.includes('o1') || m.includes('codex') ||
-      m.includes('qwen3.7-max')) {
+      m.includes('chat-latest') || m.includes('qwen3.7-max')) {
     return `# Quality Standards (strong model)
 - Keep calling tools until the task is complete AND the result is verified. Don't stop at "this should work" — prove it works.
 - Before finalizing: check correctness, grounding in tool output, and formatting.

src/agent/tokens.ts

@@ -249,8 +249,15 @@ const MODEL_CONTEXT_WINDOWS: Record<string, number> = {
   'openai/gpt-5.2': 128_000,
   'openai/gpt-5-mini': 128_000,
   'openai/gpt-5-nano': 128_000,
+  'openai/gpt-5.2-pro': 400_000,
   'openai/gpt-4.1': 1_000_000,
+  'openai/gpt-4.1-mini': 128_000,
+  'openai/gpt-4.1-nano': 128_000,
+  'openai/gpt-4o': 128_000,
+  'openai/gpt-4o-mini': 128_000,
+  'openai/o1': 200_000,
   'openai/o3': 200_000,
+  'openai/o3-mini': 128_000,
   'openai/o4-mini': 200_000,
   // Google
   'google/gemini-2.5-pro': 1_000_000,
@@ -259,18 +266,27 @@ const MODEL_CONTEXT_WINDOWS: Record<string, number> = {
   'google/gemini-3.1-pro': 1_000_000,
   'google/gemini-3.5-flash': 1_000_000,
   'google/gemini-3.1-flash-lite': 1_000_000,
+  'google/gemini-3-flash-preview': 1_048_576,
   // DeepSeek (V4 family — gateway aliased deepseek-chat / -reasoner to V4
   // Flash on 2026-05-03; context bumped 128K → 1M for both, 65K out)
   'deepseek/deepseek-chat': 1_000_000,
   'deepseek/deepseek-reasoner': 1_000_000,
   'deepseek/deepseek-v4-pro': 1_000_000,
-  // xAI
+  // xAI. grok-4.5 / 4.3 / build were missing until 2026-08-20: neither matches
+  // any inference pattern below, so a cold catalog cache fell through to the
+  // blind 128k default and compacted a 500K–1M window ~4-8x too early.
+  'xai/grok-4.5': 500_000,
+  'xai/grok-4.3': 1_000_000,
+  'xai/grok-build-0.1': 256_000,
   'xai/grok-3': 131_072,
   'xai/grok-4-0709': 131_072,
   'xai/grok-4-1-fast-reasoning': 131_072,
   // Others
+  'zai/glm-5.3': 1_000_000, // flagship 2026-08 — 1M context, always-on reasoning
   'zai/glm-5.2': 1_000_000, // flagship bump 2026-06 — context jumped 200K → 1M
   'zai/glm-5.1': 200_000,
+  'zai/glm-5': 200_000,
+  'zai/glm-5-turbo': 200_000,
   'moonshot/kimi-k3': 1_048_576,
   'moonshot/kimi-k2.7': 256_000,
   'moonshot/kimi-k2.6': 256_000,

src/brain/extract.ts

@@ -11,10 +11,15 @@ import {
   addObservation, upsertRelation, isJunkEntityName,
 } from './store.js';
 
+// Last entry is the free safety net. nvidia/nemotron-super-49b held it until
+// 2026-08-20, by which point it had left the catalog entirely — the free pool
+// answers for it with a substitute, so the "free fallback" was neither the
+// model named nor reliably shaped. nemotron-nano-9b-v2 is the current free
+// default and verifiably serves itself.
 const EXTRACTION_MODELS = [
   'google/gemini-2.5-flash-lite',
   'google/gemini-2.5-flash',
-  'nvidia/nemotron-super-49b',
+  'nvidia/nemotron-nano-9b-v2',
 ];
 
 const VALID_TYPES = new Set<EntityType>(['person', 'project', 'company', 'product', 'concept']);

src/pricing.ts

@@ -145,6 +145,7 @@ export const MODEL_PRICING: Record<string, { input: number; output: number; perC
   'zai/glm-5': { input: 1.0, output: 3.2 }, // flat promo ended 2026-06-06; raised upstream 2026-08-07
   'zai/glm-5.1': { input: 1.40, output: 4.40 }, // launch promo ended 2026-06-05 — per-token now
   'zai/glm-5.2': { input: 1.40, output: 4.40 }, // new flagship 2026-06 — 1M context, same per-token price as 5.1
+  'zai/glm-5.3': { input: 1.40, output: 4.40 }, // flagship 2026-08 — 1M context, always-on reasoning, priced at 5.2
   'zai/glm-5-turbo': { input: 1.2, output: 4.0 }, // flat promo ended 2026-06-06 — per-token now
   'zai/glm-5.1-turbo': { input: 1.2, output: 4.0 },  // client alias for zai/glm-5-turbo
 };

src/router/index.ts

@@ -414,10 +414,23 @@ function classicRouteRequest(
 //     that can't be async (proxy, LLM-client bootstrap) keep using the sync
 //     `routeRequest`, which silently does keyword-only routing.
 
-// llama-4-maverick: clean one-word classification output. glm-4.7 + qwen-
-// thinking emit reasoning into thinking blocks and leave text empty under
-// tight max_tokens — fine for chat, wrong shape for single-word dispatch.
-const CLASSIFIER_MODEL = process.env.FRANKLIN_ROUTER_MODEL || 'nvidia/qwen3-next-80b-a3b-instruct';
+// The classifier needs a free model that answers with ONE BARE WORD under a
+// tight max_tokens. That is a narrower requirement than "is a good free chat
+// model", and most of the free pool fails it by streaming chain-of-thought
+// into `content` (glm-4.7 and the qwen-thinking builds emit reasoning and
+// leave text empty; nemotron-nano-9b-v2 opens with "Okay, let's see. The user
+// wants to…" and never reaches a verdict inside the budget).
+//
+// 2026-08-19: the previous default, nvidia/qwen3-next-80b-a3b-instruct, was
+// EOL'd by NVIDIA (410 on 2026-07-27) — the gateway now rides its calls on
+// nemotron-3-super-120b, which leaks prose. Every classification has been
+// failing the strict parse and falling through to keyword-only routing ever
+// since, silently, because the fallback is by design invisible.
+//
+// nemotron-3-nano-omni is the replacement: live-probed on this exact prompt
+// shape it returns "MEDIUM" and nothing else, and it verifiably serves itself
+// rather than riding a pooled substitute.
+const CLASSIFIER_MODEL = process.env.FRANKLIN_ROUTER_MODEL || 'nvidia/nemotron-3-nano-omni-30b-a3b-reasoning';
 const CLASSIFIER_TIMEOUT_MS = 2_500;
 
 const CLASSIFIER_SYSTEM = `You classify a user's message into ONE routing tier for a CLI agent. Reply with EXACTLY ONE WORD from the allowed set. No explanation, no punctuation, no quotes.
@@ -771,22 +784,31 @@ export function getFallbackChain(
 // verifiably serves ITSELF on the streaming path Franklin uses (verified live
 // through the binary today). mistral-nemotron is DEGRADED at NVIDIA — stream
 // calls 400 ("DEGRADED function cannot be invoked") and non-stream calls ride
-// the gateway's disclosed fallback — so it sits second: it still answers
-// non-streaming turns, and it upgrades the chain automatically if NVIDIA
-// restores it. step-3.7-flash leaks thinking prose into content, and the 30B
-// omni model came back served as nvidia/gpt-oss-120b (the pooled-substitute
-// trap this comment already warns about).
+// the gateway's disclosed fallback.
+// 2026-08-19: nvidia/nemotron-3-nano-omni-30b-a3b-reasoning promoted to
+// second. When it was last probed it came back served as gpt-oss-120b — the
+// pooled-substitute trap this comment warns about — but NVIDIA has since
+// fixed it: it now answers as ITSELF on both the streaming and non-streaming
+// paths, and it returns clean single-word output under tight max_tokens (it
+// is also the router's classifier for that reason). mistral-nemotron drops to
+// third: still DEGRADED at NVIDIA, kept as a genuinely different family for
+// the case where both Nemotron nano builds are down. The catalog's fifth free
+// id, step-3.7-flash, stays out of every chain — it leaks thinking prose into
+// content AND comes back served by nemotron-3-super-120b.
+const OMNI = 'nvidia/nemotron-3-nano-omni-30b-a3b-reasoning';
+
 const FREE_MODELS_BY_CATEGORY: Record<Category, string[]> = {
-  coding:    ['nvidia/nemotron-nano-9b-v2', 'nvidia/mistral-nemotron'],
-  trading:   ['nvidia/nemotron-nano-9b-v2', 'nvidia/mistral-nemotron'],
-  research:  ['nvidia/nemotron-nano-9b-v2', 'nvidia/mistral-nemotron'],
-  reasoning: ['nvidia/nemotron-nano-9b-v2', 'nvidia/mistral-nemotron'],
-  chat:      ['nvidia/nemotron-nano-9b-v2', 'nvidia/mistral-nemotron'],
-  creative:  ['nvidia/nemotron-nano-9b-v2', 'nvidia/mistral-nemotron'],
+  coding:    ['nvidia/nemotron-nano-9b-v2', OMNI, 'nvidia/mistral-nemotron'],
+  trading:   ['nvidia/nemotron-nano-9b-v2', OMNI, 'nvidia/mistral-nemotron'],
+  research:  ['nvidia/nemotron-nano-9b-v2', OMNI, 'nvidia/mistral-nemotron'],
+  reasoning: ['nvidia/nemotron-nano-9b-v2', OMNI, 'nvidia/mistral-nemotron'],
+  chat:      ['nvidia/nemotron-nano-9b-v2', OMNI, 'nvidia/mistral-nemotron'],
+  creative:  ['nvidia/nemotron-nano-9b-v2', OMNI, 'nvidia/mistral-nemotron'],
 };
 
 const DEFAULT_FREE_CHAIN: string[] = [
   'nvidia/nemotron-nano-9b-v2',
+  OMNI,
   'nvidia/mistral-nemotron',
 ];