fix(video): tighten Seedance capability tables to token360's own schema (0.39.0)

CHANGELOG.md

@@ -2,6 +2,81 @@
 
 All notable changes to BlockRun MCP will be documented in this file.
 
+## 0.39.0
+
+The Seedance capability tables, re-derived from token360's OWN published
+parameter schema (GET /v1/models/{id} → parameter_schema) — a source nobody
+had ever read; every table in every repo predated knowing it existed. Sibling
+of blockrun PR #353 and blockrun-sol PR #133, which land the same truth in the
+gateways, so client and server can no longer disagree.
+
+- **`fix(video)` — per-model resolution sets tightened to the schema.** 0.38.1
+  (yesterday) tightened toward the gateway's tables; live probing shows those
+  tables were themselves wrong, in both directions:
+
+  | model | 0.38.1 allowed | schema truth |
+  |---|---|---|
+  | 1.5-pro | 360p-1K | **480p / 720p / 1080p** |
+  | 2.0-fast | 360p-1K | **480p / 720p** |
+  | 2.0 | t2v narrow · i2v 360p-1K+4K | **480p / 720p / 1080p / 4K, both modes** |
+  | 2.5 | 360p-720p | **480p / 720p** |
+
+  The t2v/i2v split is gone — the schema is per-model, everything it lists
+  passes in both modes, and 540p/1K hard-reject upstream even image-conditioned
+  (the pay-then-lose path the 0.38.1 guard half-closed). A few off-schema
+  values do pass upstream validation (2.0-fast 1080p, 1.5-pro 540p) and stay
+  blocked anyway: "passes validation" is not "renders" — 1.5-pro historically
+  echoed 2K/4K, billed the requested tier, and rendered 720p.
+
+- **`fix(video)` — `360p`/`540p`/`1K` leave the resolution enum, `9:21` leaves
+  the aspect-ratio enum.** No Seedance model lists any of them, and 9:21 is
+  rejected upstream for the whole family. Probed at $0: token360 validates
+  duration → resolution → ratio → other params, so a deliberately invalid
+  LATER parameter (ratio "9:99"; output_format "avi" when testing ratios)
+  guarantees rejection after the parameter under test is judged, and a failed
+  submit never creates a task or bills. RESOLUTION_TOKEN_FACTOR drops the same
+  tiers — the estimator throws on anything missing, so a stale entry is a loud
+  failure, not a silent 720p-baseline reserve.
+
+- **`fix(video)` — the resolution guard's advice now names the real ceiling.**
+  2.0-fast is told 720p is its ceiling (0.38.1 wrongly offered it 1080p/1K —
+  the exact bill-high-render-low trap shape).
+
+- **`test(video)` — the t2v/i2v split test became the no-split test**, and the
+  factor-coverage test iterates the trimmed enum. verify:prices drops the
+  2.0@360p tracking probe (inexpressible from this client now) and keeps
+  2.5@1080p: the gateway still quotes $3.55 for a render token360 refuses —
+  the defect blockrun PR #353 fixes — so that probe reports the moment the
+  fleet is consistent.
+
+Breaking (deliberately — hence a minor bump, not a patch): a stale caller
+sending `resolution:"360p"` or `aspect_ratio:"9:21"` now gets a zod enum
+error instead of a paid round trip or an upstream 400. For Sora/Grok
+callers, removed `resolution` values were accepted no-ops; `9:21` was NOT a
+no-op on Sora — it selected portrait orientation — and `9:16` is the exact
+replacement. One removed value did produce billed output before this
+release — seedance-2.0 image-conditioned 360p passed every layer — but it
+is exactly the off-schema shape that can silently downscale while billing
+the requested tier, which is why the schema is the line. The
+`aspect_ratio` description also stops claiming Sora/Grok both ignore it:
+the handler forwards it for every model, SORA uses it to pick portrait vs
+landscape, and Grok is the one that truly ignores it (the gateway never
+forwards it to xAI).
+
+303 tests pass; live gate: 0 under-reserving.
+
+Also shipping in 0.39.0 (landed on main as #97, recorded here because
+releases are cut from this file): the price sweep now probes BOTH gateways —
+Base charges base + $0.001, Solana base + $0.000, and a Solana price DEARER
+than Base fails the gate since an agent can switch chains mid-session;
+`blockrun_defi`/`blockrun_modal` gained the Base-only chain guard six other
+tools already had; eleven skills' flat prices were re-read off the live 402;
+and dFlow left the catalog (all three routes 404 at the gateway). The
+dual-gateway sweep also gives the 2.5@1080p fleet-consistency probe its
+other half for free: Solana already 400s it (blockrun-sol#130's cap) while
+Base still quotes it — one line of output now shows the split until
+blockrun#353 deploys.
+
 ## 0.38.1
 
 Review pass over 0.38.0. The model addition held up; the guards and the tests

package.json

@@ -1,6 +1,6 @@
 {
   "name": "@blockrun/mcp",
-  "version": "0.38.1",
+  "version": "0.39.0",
   "mcpName": "io.github.BlockRunAI/blockrun-mcp",
   "description": "BlockRun MCP Server - Give your AI agent web search, deep research, prediction markets, and crypto data. Paid via x402 micropayments.",
   "type": "module",

scripts/verify-prices.ts

@@ -162,14 +162,24 @@ const PROBES: Probe[] = [
     ["bytedance/seedance-2.0", undefined, "4K"],
     ["bytedance/seedance-2.5", undefined, undefined],
     ["bytedance/seedance-2.5", 30, undefined],
-    // Combinations the client-side guards REFUSE, probed anyway because a guard
-    // is a claim about the gateway and this is the only thing that checks it.
-    // 2.5@1080p is the interesting one: the gateway still QUOTES it ($3.55) even
-    // though token360 rejects it at submit (probed 2026-08-07), so this line
-    // tracks a known gateway defect. The day it stops quoting, the gateway has
-    // been fixed and this probe will report `no 402` like 2.0@360p does.
+    // A combination the client-side guard REFUSES, probed anyway because a
+    // guard is a claim about the gateway and this is the only thing checking
+    // it: the gateway still QUOTES 2.5@1080p ($3.55) even though token360
+    // rejects it at submit (probed 2026-08-07) — a known gateway defect, fixed
+    // by blockrun PR #353. Once that deploys, this probe reports `no 402` and
+    // the guard is vindicated. (2.0@360p left the matrix: 360p is out of the
+    // schema for every model and out of this client's enum, so the estimator
+    // now throws on it — it can no longer even be expressed from here.)
     ["bytedance/seedance-2.5", undefined, "1080p"],
-    ["bytedance/seedance-2.0", undefined, "360p"],
+    // Each model probed at (or near) its ceiling and floor tier, so a gateway
+    // whose capability surface diverges from SEEDANCE_RESOLUTIONS — in either
+    // direction — shows up here as `no 402` or a price mismatch. 1.5-pro@1080p
+    // was the only above-720p schema tier with no live pin anywhere until the
+    // 0.39.0 review probed it: charge $0.797311 vs reserve $0.798311.
+    ["bytedance/seedance-1.5-pro", undefined, "1080p"],
+    ["bytedance/seedance-2.0-fast", undefined, "480p"],
+    ["bytedance/seedance-2.5", undefined, "480p"],
+    ["bytedance/seedance-2.0", undefined, "480p"],
   ] as Array<[string, number | undefined, string | undefined]>).map(([model, seconds, resolution]) => ({
     label: `video ${model.split("/")[1]}${seconds ? ` ${seconds}s` : ""}${resolution ? ` ${resolution}` : ""}`,
     path: "videos/generations",

src/tools/video.ts

@@ -59,13 +59,12 @@ const VIDEO_BASE_PRICE_PER_SECOND: Record<string, number> = {
 // is 0.3, not 0.25), and the live 402 agrees with the table, not the geometry.
 // So a new tier must be read off a real quote (`npm run verify:prices`), never
 // derived. Sora/Grok bill per second and ignore this entirely.
+// 360p/540p/1K trimmed 2026-08-07: token360's schema lists them for no model,
+// so they left the enum too; estimateVideoCost throws on anything missing here.
 const RESOLUTION_TOKEN_FACTOR: Record<string, number> = {
-  "360p": 0.3,
   "480p": 0.5,
-  "540p": 0.7,
   "720p": 1,
   "1080p": 2.25,
-  "1K": 2.25,
   "4K": 9,
 };
 
@@ -113,37 +112,41 @@ const VIDEO_DURATION_RANGE: Record<string, { min: number; max: number; allowed?:
   "azure/sora-2": { min: 4, max: 12, allowed: [4, 8, 12] },
 };
 
-// Which resolutions each Seedance SKU actually accepts. Mirrors the gateway's
-// checkUnsupportedVideoInput (blockrun/src/lib/video-input.ts), which is the
-// authority — it rejects PRE-402, so a mismatch here costs a wasted round trip
-// rather than money. The asymmetry that makes it worth gating anyway: where the
-// gateway DOES quote, token360 bills the tier it was asked for even if it
-// renders lower, so an over-tall request pays 2.25x (1080p) or 9x (4K) for 720p.
+// Which resolutions each Seedance SKU actually accepts — token360's OWN
+// published parameter schema (GET /v1/models/{id} -> parameter_schema), verified
+// live 2026-08-07. The gateway's checkUnsupportedVideoInput will derive from
+// the same source once blockrun PR #353 lands (machine-written snapshot);
+// until it deploys, production still runs the OLD wide tables and this client
+// is the only gate. THIS table is a hand-copy of that probe run — re-check it
+// whenever the gateway resyncs its snapshot.
+// The verify:prices matrix probes each model at its ceiling tier, so a gateway
+// that starts quoting (or stops quoting) a tier this table disagrees with
+// shows up in the free 402 sweep.
 //
-// seedance-2.0 splits on generation mode — token360 400s 360p/540p/1K in
-// text-to-video but accepts them when a first frame is supplied (live-probed
-// upstream 2026-06-26). 4K is real 3840x2160 on 2.0 and on nothing else.
-const RES_LOW = ["360p", "480p", "540p", "720p", "1080p", "1K"];
-const SEEDANCE_RESOLUTIONS: Record<string, { t2v: Set<string>; i2v: Set<string>; note: string }> = {
-  "bytedance/seedance-1.5-pro": { t2v: new Set(RES_LOW), i2v: new Set(RES_LOW), note: "1080p is the ceiling — only bytedance/seedance-2.0 renders true 4K" },
-  "bytedance/seedance-2.0-fast": { t2v: new Set(RES_LOW), i2v: new Set(RES_LOW), note: "1080p is the ceiling — only bytedance/seedance-2.0 renders true 4K" },
+// No t2v/i2v split: the schema is per-model, and probing shows everything it
+// lists passes in both modes, while 540p/1K hard-reject upstream even
+// image-conditioned. We allow exactly the schema, nothing more — a few
+// off-schema values PASS submit validation (2.0-fast 1080p, 1.5-pro 540p) but
+// "passes validation" is not "renders": 1.5-pro historically echoed 2K/4K,
+// billed the requested tier, and rendered 720p. Off-schema acceptance can
+// silently downscale while billing the higher tier, so it stays out.
+export const SEEDANCE_RESOLUTIONS: Record<string, { resolutions: Set<string>; note: string }> = {
+  "bytedance/seedance-1.5-pro": {
+    resolutions: new Set(["480p", "720p", "1080p"]),
+    note: "1080p is the ceiling — only bytedance/seedance-2.0 renders true 4K",
+  },
+  "bytedance/seedance-2.0-fast": {
+    resolutions: new Set(["480p", "720p"]),
+    note: "720p is the ceiling on 2.0-fast — for 1080p use bytedance/seedance-1.5-pro (cheapest) or bytedance/seedance-2.0 (also 4K)",
+  },
   "bytedance/seedance-2.0": {
-    t2v: new Set(["480p", "720p", "1080p", "4K"]),
-    i2v: new Set([...RES_LOW, "4K"]),
-    note: "text-to-video on 2.0 accepts 480p / 720p / 1080p / 4K only; the lower tiers need an image_url first frame",
+    resolutions: new Set(["480p", "720p", "1080p", "4K"]),
+    note: "the only model that renders true 4K (3840x2160)",
+  },
+  "bytedance/seedance-2.5": {
+    resolutions: new Set(["480p", "720p"]),
+    note: "2.5 caps at 720p — it trades resolution for length (up to 30s). Use bytedance/seedance-2.0 for 1080p or 4K",
   },
-  // 2.5 trades resolution for length, and this is PROVEN, not inferred: a paid
-  // probe on 2026-08-07 sent 1080p and 1K at both t2v and i2v, and token360
-  // rejected all four — "the parameter resolution specified in the request is
-  // not valid for model dreamina-seedance-2-5". The probe cost $0.00 because
-  // nothing settles on a failed submit.
-  //
-  // Worth knowing why this guard earns its keep: the GATEWAY does not enforce
-  // this. checkUnsupportedVideoInput's `is20` regex does not match "2.5", so 2.5
-  // falls to a permissive fallback and the gateway happily quotes $3.551218 for
-  // 1080p — a 402 signed for a request upstream will refuse. Without this check
-  // the caller signs a payment and gets a 500 the gateway labels "temporary".
-  "bytedance/seedance-2.5": { t2v: new Set(["360p", "480p", "540p", "720p"]), i2v: new Set(["360p", "480p", "540p", "720p"]), note: "2.5 caps at 720p — it trades resolution for length (up to 30s). Use bytedance/seedance-2.0 for 1080p or 4K" },
 };
 
 /**
@@ -212,8 +215,8 @@ Returns a permanent blockrun-hosted MP4 URL (the gateway mirrors the asset to GC
         real_face_asset_id: z.string().regex(/^ta_[A-Za-z0-9]+$/, "token360 asset id like 'ta_xxxx'").optional().describe("BytePlus RealFace asset id (from blockrun_realface enroll/list) to generate video of a specific real person. Seedance 2.0 / 2.0-fast only (NOT 2.5). Mutually exclusive with image_url."),
         duration_seconds: z.number().int().min(1).max(60).optional().describe("Duration to bill for. Defaults to the model's own default (8s xAI, 5s Seedance, 4s Sora). Per-model range: seedance-1.5-pro 4-12s · seedance-2.0 / 2.0-fast 4-15s · seedance-2.5 4-30s · sora-2 exactly 4, 8 or 12 · grok-imagine-video 1-15s."),
         generate_audio: z.boolean().optional().describe("Seedance only: whether to generate a synced audio track. Defaults ON for text-to-video and OFF for image/RealFace-conditioned. The auto-generated audio is occasionally rejected by upstream moderation ('output audio may contain sensitive information') even for benign prompts — pass false to skip audio and avoid that failure. Ignored by xAI/Sora."),
-        resolution: z.enum(["360p", "480p", "540p", "720p", "1080p", "1K", "4K"]).optional().describe("Seedance only: output resolution. Defaults to 720p. Higher resolutions cost more (token-priced upstream, ~2.25x at 1080p and ~9x at 4K) — the final price comes from the 402 challenge, so the up-front estimate understates 1080p/4K. 4K is bytedance/seedance-2.0 only; seedance-2.5 caps at 720p. Ignored by xAI/Sora."),
-        aspect_ratio: z.enum(["adaptive", "16:9", "9:16", "1:1", "4:3", "3:4", "21:9", "9:21"]).optional().describe("Seedance only: output aspect ratio, e.g. '9:16' for vertical/mobile, '16:9' for landscape. Defaults to the model's own default. Ignored by xAI/Sora."),
+        resolution: z.enum(["480p", "720p", "1080p", "4K"]).optional().describe("Seedance only: output resolution. Defaults to 720p. Higher resolutions cost more (token-priced upstream, ~2.25x at 1080p and ~9x at 4K). Per-model sets from token360's published schema: seedance-2.0 480p/720p/1080p/4K · 1.5-pro 480p/720p/1080p · 2.0-fast and 2.5 480p/720p only. Ignored by xAI/Sora (dropped from the request)."),
+        aspect_ratio: z.enum(["adaptive", "16:9", "9:16", "1:1", "4:3", "3:4", "21:9"]).optional().describe("Output aspect ratio. Seedance honors the full set; Sora uses it only to pick portrait vs landscape (9:16 / 3:4 -> portrait); Grok ignores it (the gateway never forwards it to xAI). Defaults to the model's own default. (9:21 removed 2026-08-07 — no Seedance model offers it; use 9:16 for vertical.)"),
         last_frame_url: z.string().url().optional().describe("Seedance 1.5-pro / 2.0 / 2.0-fast only (NOT 2.5): first-and-last-frame interpolation. A second image URL that seeds the FINAL frame so the model tweens from image_url (first frame) → last_frame_url (last frame). Requires image_url; mutually exclusive with real_face_asset_id."),
         model: z.enum(["azure/sora-2", "xai/grok-imagine-video", "bytedance/seedance-1.5-pro", "bytedance/seedance-2.0-fast", "bytedance/seedance-2.0", "bytedance/seedance-2.5"]).optional().default("xai/grok-imagine-video").describe("Video model to use"),
         agent_id: z.string().optional().describe("Agent identifier for budget tracking and enforcement."),
@@ -272,21 +275,17 @@ Returns a permanent blockrun-hosted MP4 URL (the gateway mirrors the asset to GC
           }
         }
 
-        // Resolution ceilings, per model and per generation mode. Only Seedance
+        // Resolution ceilings, per model. Only Seedance
         // is checked: Sora and Grok bill per second and ignore the parameter,
         // which is what the schema promises — so for them it is DROPPED from the
         // body below rather than rejected. (Forwarding it earned a gateway 400,
         // which is the opposite of "ignored".)
         const seedanceRes = SEEDANCE_RESOLUTIONS[selectedModel];
-        if (resolution && seedanceRes) {
-          const imageConditioned = Boolean(image_url || real_face_asset_id || last_frame_url);
-          const allowed = imageConditioned ? seedanceRes.i2v : seedanceRes.t2v;
-          if (!allowed.has(resolution)) {
-            return {
-              content: [{ type: "text", text: formatError(`${selectedModel} does not render ${resolution}${imageConditioned ? " for image-conditioned video" : ""}. ${seedanceRes.note}. Supported here: ${[...allowed].join(", ")}.`) }],
-              isError: true,
-            };
-          }
+        if (resolution && seedanceRes && !seedanceRes.resolutions.has(resolution)) {
+          return {
+            content: [{ type: "text", text: formatError(`${selectedModel} does not render ${resolution}. ${seedanceRes.note}. Supported: ${[...seedanceRes.resolutions].join(", ")}.`) }],
+            isError: true,
+          };
         }
 
         const billedSeconds = duration_seconds ?? VIDEO_DEFAULT_DURATION[selectedModel] ?? 8;
@@ -359,7 +358,7 @@ Returns a permanent blockrun-hosted MP4 URL (the gateway mirrors the asset to GC
         const settledUsd = amountToUsd(details.amount);
 
         // The 402 carries the REAL price; Seedance/Sora are token-priced, so a
-        // 1080p/2K/4K render can far exceed the per-second estimate reserved at
+        // 1080p/4K render can far exceed the per-second estimate reserved at
         // the gate. Re-reserve against the cap BEFORE paying so a single high-res
         // call can't settle past the budget (and concurrent jobs hold the true
         // amount, not the low estimate, for the whole polling window).