fix(capabilities): publish real GPT-6/GPT-5.4+ context windows and combo token limits

open-sse/providers/capabilities.js

@@ -6,15 +6,18 @@
 //   3. PATTERN_CAPABILITIES                     — glob match, ordered specific -> generic
 //   4. DEFAULT_CAPABILITIES                     — safe floor (always returned)
 //
-// Two extra layers then refine the result, and neither can override the hand
+// Two extra layers then refine the result:
-// written tables above (steps 1-2 short-circuit before they are consulted):
 //   • the synced catalog — modalities keyed by model, limits keyed by provider
 //     + model, refreshed from models.dev in the background. It reads a file, so
 //     the server installs it via setCatalogSource(); this module stays free of
 //     node:fs because the dashboard bundles it into the browser too.
 //   • visionPatterns.js — name-based vision detection, last resort so a model
 //     nobody has catalogued yet still accepts images.
-// Both only ever turn a capability ON.
+// Modalities only ever turn a capability ON. Limits from the catalog overlay
+// the canonical exact entry (step 2) so a gateway-specific models.dev delta
+// (Copilot's 32k Claude output, etc.) actually publishes. Step 1 still
+// short-circuits: a hand-written PROVIDER_CAPABILITIES truncation is the
+// gateway's own number and must not be overwritten.
 //
 // ── HOW TO ADD / UPDATE A MODEL ──────────────────────────────────────
 // Authoritative data source: https://models.dev/api.json (145 providers, 4000+
@@ -165,6 +168,11 @@ const CODEX_GPT_56_SOL_CAPS  = { vision: true, reasoning: true, search: true, th
 const CODEX_GPT_56_DEFAULT_CAPS = { vision: true, reasoning: true, search: true, thinkingFormat: "openai", contextWindow: 272000, maxOutput: 128000 };
 const CODEX_EXTENDED_CAPS = { ...CODEX_GPT_56_DEFAULT_CAPS, contextWindow: 872000 };
 
+// Devin CLI's registry declares a 200k context window for these GPT variants.
+// Keep the GPT feature/output fields because provider overrides short-circuit
+// the generic pattern rather than merging with it.
+const DEVIN_CLI_GPT_CAPS = { vision: true, reasoning: true, search: true, thinkingFormat: "openai", contextWindow: 200000, maxOutput: 128000 };
+
 /**
  * Provider-specific capability overrides. Keyed by provider alias/id.
  */
@@ -215,6 +223,15 @@ export const PROVIDER_CAPABILITIES = {
     "gpt-5.6-terra-thinking-agentic": KIRO_GPT_5_6_CAPABILITIES,
     "gpt-5.6-luna-thinking-agentic": KIRO_GPT_5_6_CAPABILITIES,
   },
+  "devin-cli": {
+    "gpt-5.4-high": DEVIN_CLI_GPT_CAPS,
+    "gpt-5.4-medium": DEVIN_CLI_GPT_CAPS,
+    "gpt-5.4-low": DEVIN_CLI_GPT_CAPS,
+    "gpt-5.5-xhigh": DEVIN_CLI_GPT_CAPS,
+    "gpt-5.5-high": DEVIN_CLI_GPT_CAPS,
+    "gpt-5.5-medium": DEVIN_CLI_GPT_CAPS,
+    "gpt-5.5-low": DEVIN_CLI_GPT_CAPS,
+  },
   // CodeBuddy.cn — authoritative per-model metadata from the gateway's model
   // config (contextWindow=maxInputTokens, maxOutput=maxOutputTokens, vision=
   // supportsImages). Every model reasons via OpenAI-style reasoning_effort
@@ -278,6 +295,8 @@ export const PROVIDER_CAPABILITIES = {
 // the intl Qoder capability table verbatim (vision/reasoning/contextWindow).
 PROVIDER_CAPABILITIES["qoder-cn"] = PROVIDER_CAPABILITIES["qoder"];
 PROVIDER_CAPABILITIES.cx = PROVIDER_CAPABILITIES.codex;
+PROVIDER_CAPABILITIES.dv = PROVIDER_CAPABILITIES["devin-cli"];
+PROVIDER_CAPABILITIES.devin = PROVIDER_CAPABILITIES["devin-cli"];
 
 /**
  * Pattern fallback — glob (* = wildcard), matched case-insensitively and
@@ -315,11 +334,24 @@ export const PATTERN_CAPABILITIES = [
   { pattern: "*nanobanana*",    caps: { vision: true, imageOutput: true } },
 
   // ── OpenAI GPT-6.x (vision + thinking + web search) ──────────────
-  { pattern: "*gpt-6*",         caps: { vision: true, reasoning: true, search: true, thinkingFormat: "openai", contextWindow: 272000, maxOutput: 128000 } },
+  // 1.05M is the API window for the whole gpt-6 family (astra, luna, sol alike).
+  // A gateway that truncates lower records its own number in
+  // PROVIDER_CAPABILITIES, which wins over this pattern — Kiro at 272k, Codex
+  // OAuth at 272k/372k (see CODEX_GPT_56_* above). This entry used to carry
+  // Kiro's 272k, so every other provider's gpt-6 models inherited one gateway's
+  // limit and were published at 3.9x under their real window.
+  { pattern: "*gpt-6*",         caps: { vision: true, reasoning: true, search: true, thinkingFormat: "openai", contextWindow: 1050000, maxOutput: 128000 } },
 
   // ── OpenAI GPT-5.x (vision + thinking + web search) ──────────────
   { pattern: "*gpt-5*image*",   caps: { imageOutput: true } },
   { pattern: "*gpt-5*codex*",   caps: { reasoning: true, search: true, thinkingFormat: "openai", contextWindow: 400000, maxOutput: 128000 } },
+  // gpt-5.4 is where the 1.05M window starts, but the mini and nano tiers stayed
+  // at 400k — first match wins, so those two have to be listed ahead of it.
+  { pattern: "*gpt-5.4-mini*",  caps: { vision: true, reasoning: true, search: true, thinkingFormat: "openai", contextWindow: 400000, maxOutput: 128000 } },
+  { pattern: "*gpt-5.4-nano*",  caps: { vision: true, reasoning: true, search: true, thinkingFormat: "openai", contextWindow: 400000, maxOutput: 128000 } },
+  { pattern: "*gpt-5.4*",       caps: { vision: true, reasoning: true, search: true, thinkingFormat: "openai", contextWindow: 1050000, maxOutput: 128000 } },
+  { pattern: "*gpt-5.5*",       caps: { vision: true, reasoning: true, search: true, thinkingFormat: "openai", contextWindow: 1050000, maxOutput: 128000 } },
+  { pattern: "*gpt-5.6*",       caps: { vision: true, reasoning: true, search: true, thinkingFormat: "openai", contextWindow: 1050000, maxOutput: 128000 } },
   { pattern: "*gpt-5*",         caps: { vision: true, reasoning: true, search: true, thinkingFormat: "openai", contextWindow: 400000, maxOutput: 128000 } },
   { pattern: "*gpt-4o*",        caps: { vision: true, search: true, contextWindow: 128000, maxOutput: 16384 } },
   { pattern: "*gpt-4.1*",       caps: { vision: true, contextWindow: 1000000, maxOutput: 32768 } },
@@ -618,9 +650,10 @@ export function getCapabilitiesForModel(provider, model) {
     if (providerCaps?.[baseModel]) return { ...DEFAULT_CAPABILITIES, ...providerCaps[baseModel] };
   }
 
-  // 2. Canonical exact
+  // 2. Canonical exact, then catalog overlay so provider-scoped models.dev
-  if (MODEL_CAPABILITIES[baseModel]) return { ...DEFAULT_CAPABILITIES, ...MODEL_CAPABILITIES[baseModel] };
+  //    deltas still apply. Step 1 above still short-circuits.
-  if (MODEL_CAPABILITIES[model]) return { ...DEFAULT_CAPABILITIES, ...MODEL_CAPABILITIES[model] };
+  if (MODEL_CAPABILITIES[baseModel]) return refine(MODEL_CAPABILITIES[baseModel], provider, model);
+  if (MODEL_CAPABILITIES[model]) return refine(MODEL_CAPABILITIES[model], provider, model);
 
   // 3. Pattern match (first match wins), refined by catalog + name heuristic
   for (const { pattern, caps } of PATTERN_CAPABILITIES) {

src/app/api/v1/models/route.js

@@ -1,5 +1,6 @@
 import { PROVIDER_MODELS, PROVIDER_ID_TO_ALIAS, getModelKind } from "@/shared/constants/models";
 import {
+  ALIAS_TO_ID,
   AI_PROVIDERS,
   getProviderAlias,
   isAnthropicCompatibleProvider,
@@ -40,6 +41,22 @@ async function resolveQoderLiveModels(conn, provider) {
   return { models: models.map((m) => ({ id: m.id, name: m.name })) };
 }
 
+// Combo seats use UI aliases; the model registry also has transport aliases.
+// Capability overrides and catalog limits are keyed by provider id.
+const ALIAS_TO_PROVIDER_ID = {
+  ...Object.fromEntries(
+    Object.entries(PROVIDER_ID_TO_ALIAS).map(([id, alias]) => [alias, id])
+  ),
+  ...ALIAS_TO_ID,
+};
+
+function comboSeatCapabilities(seat) {
+  const slash = seat.indexOf("/");
+  if (slash <= 0) return null;
+  const alias = seat.slice(0, slash);
+  return getCapabilitiesForModel(ALIAS_TO_PROVIDER_ID[alias] || alias, seat.slice(slash + 1));
+}
+
 // Per-provider live model resolvers. Each receives a connection record and
 // returns { models: [{ id, name? }, ...] } | null on failure.
 // Adding a provider here makes /v1/models prefer the live catalog for it.
@@ -254,6 +271,47 @@ function comboMatchesKinds(combo, kindFilter) {
   return kindFilter.includes(kind);
 }
 
+// Nested combo names are valid seats — the model selector exposes them and
+// chat routing resolves them recursively — but a no-slash seat is otherwise
+// treated as a literal model and publishes the 200k floor. Expand nested
+// names (cycle-guarded) so the published window is the true min across the
+// whole chain.
+function comboSeatLimits(combo, combosByName, visiting = new Set()) {
+  const name = typeof combo?.name === "string" ? combo.name : null;
+  if (name) {
+    if (visiting.has(name)) return { contextWindow: undefined, maxOutput: undefined };
+    visiting.add(name);
+  }
+
+  let contextWindow = Infinity;
+  let maxOutput = Infinity;
+  try {
+    for (const seat of Array.isArray(combo?.models) ? combo.models : []) {
+      if (typeof seat !== "string") continue;
+      const slash = seat.indexOf("/");
+      if (slash <= 0) {
+        const nested = combosByName.get(seat);
+        if (nested) {
+          const nestedLimits = comboSeatLimits(nested, combosByName, visiting);
+          if (Number.isFinite(nestedLimits.contextWindow)) contextWindow = Math.min(contextWindow, nestedLimits.contextWindow);
+          if (Number.isFinite(nestedLimits.maxOutput)) maxOutput = Math.min(maxOutput, nestedLimits.maxOutput);
+          continue;
+        }
+      }
+      const caps = comboSeatCapabilities(seat) || getCapabilitiesForModel(null, seat);
+      if (Number.isFinite(caps?.contextWindow)) contextWindow = Math.min(contextWindow, caps.contextWindow);
+      if (Number.isFinite(caps?.maxOutput)) maxOutput = Math.min(maxOutput, caps.maxOutput);
+    }
+  } finally {
+    if (name) visiting.delete(name);
+  }
+
+  return {
+    contextWindow: Number.isFinite(contextWindow) ? contextWindow : undefined,
+    maxOutput: Number.isFinite(maxOutput) ? maxOutput : undefined,
+  };
+}
+
 /**
  * Build OpenAI-format models list filtered by service kinds.
  * @param {string[]} kindFilter - List of service kinds to include (e.g. ["llm"], ["webSearch","webFetch"]).
@@ -308,6 +366,9 @@ export async function buildModelsList(kindFilter, options = {}) {
   }
 
   const models = [];
+  const combosByName = new Map(
+    combos.filter((c) => typeof c?.name === "string").map((c) => [c.name, c]),
+  );
 
   // Lookup map so aggregateComboCapabilities can recursively resolve nested combos
   const comboByName = Object.fromEntries(combos.map((c) => [c.name, c.models]));
@@ -323,19 +384,23 @@ export async function buildModelsList(kindFilter, options = {}) {
     if (combo.kind === "webSearch" || combo.kind === "webFetch") {
       entry.kind = combo.kind;
     } else {
-      const comboCaps = aggregateComboCapabilities(combo.models, comboByName);
+      const comboCaps = aggregateComboCapabilities(combo.models, comboByName, comboSeatCapabilities);
       if (comboCaps) entry.capabilities = comboCaps;
+      // Any seat can serve the request, so the only window a combo can promise is
+      // its smallest. Combo entries were the only models on this endpoint that
+      // published no limits at all, which leaves a client to guess from the name —
+      // and it guesses high (see the snake_case note on the per-provider path).
+      const { contextWindow, maxOutput } = comboSeatLimits(combo, combosByName);
+      if (Number.isFinite(contextWindow)) entry.context_length = contextWindow;
+      if (Number.isFinite(maxOutput)) entry.max_completion_tokens = maxOutput;
     }
     models.push(entry);
   }
 
   if (connections.length === 0) {
     // DB unavailable -> return static models, filtered by per-model kind
-    const aliasToProviderId = Object.fromEntries(
-      Object.entries(PROVIDER_ID_TO_ALIAS).map(([id, alias]) => [alias, id])
-    );
     for (const [alias, providerModels] of Object.entries(PROVIDER_MODELS)) {
-      const providerId = aliasToProviderId[alias] || alias;
+      const providerId = ALIAS_TO_PROVIDER_ID[alias] || alias;
       if (!providerMatchesKinds(providerId, kindFilter)) continue;
       for (const model of providerModels) {
         if (!kindFilter.includes(modelKind(model))) continue;

src/lib/modelCatalog/sync.js

@@ -24,6 +24,7 @@ const LIMIT_TOLERANCE = 0.1;
 // while building rather than on every lookup. Providers absent here keep whatever
 // the local pattern table resolves; names that already match need no entry.
 export const PROVIDER_ALIASES = {
+  "github": "github-copilot",
   "glm": "zai",
   "glm-cn": "zhipuai",
   "claude": "anthropic",

tests/unit/gpt-6-context-window.test.js

@@ -0,0 +1,112 @@
+import { describe, expect, it } from "vitest";
+
+import { getCapabilitiesForModel, setCatalogSource } from "../../open-sse/providers/capabilities.js";
+import { PROVIDER_ALIASES, build } from "../../src/lib/modelCatalog/sync.js";
+
+// The gpt-6 family's API window is 1.05M. The pattern table published 272,000 for
+// it — Kiro's own truncation, copied into the global glob — so every other
+// provider's gpt-6 models were advertised at 3.9x under their real window, and a
+// client reading context_length compacted (or refused) far too early.
+const API_WINDOW = 1050000;
+const LEGACY_GPT5_WINDOW = 400000;
+
+describe("gpt-6 / gpt-5.4+ context windows", () => {
+  it("reports the 1.05M API window for gpt-6 models on ordinary providers", () => {
+    for (const [provider, model] of [
+      ["github", "gpt-6-luna"],
+      ["azure", "gpt-6-luna"],
+      ["openai", "gpt-6-luna"],
+      ["github", "gpt-6-sol"],
+      ["openai", "gpt-6-astra"],
+    ]) {
+      expect(getCapabilitiesForModel(provider, model).contextWindow, `${provider}/${model}`).toBe(API_WINDOW);
+    }
+  });
+
+  // These two gateways really do truncate below the API, and their numbers live in
+  // PROVIDER_CAPABILITIES, which outranks the pattern. Correcting the pattern must
+  // leave them alone — that split is the whole point of the layering.
+  it("leaves the gateways that truncate lower on their own numbers", () => {
+    expect(getCapabilitiesForModel("kiro", "gpt-5.6-luna").contextWindow).toBe(272000);
+    expect(getCapabilitiesForModel("kiro", "gpt-5.6-luna-thinking-agentic").contextWindow).toBe(272000);
+    expect(getCapabilitiesForModel("codex", "gpt-5.6-luna").contextWindow).toBe(272000);
+    expect(getCapabilitiesForModel("codex", "gpt-5.6-sol").contextWindow).toBe(372000);
+    expect(getCapabilitiesForModel("codex", "gpt-6-astra").contextWindow).toBe(272000);
+  });
+
+  it("keeps Devin CLI's seven GPT-5.4/5.5 variants at the gateway's 200k limit", () => {
+    for (const model of [
+      "gpt-5.4-high", "gpt-5.4-medium", "gpt-5.4-low",
+      "gpt-5.5-xhigh", "gpt-5.5-high", "gpt-5.5-medium", "gpt-5.5-low",
+    ]) {
+      expect(getCapabilitiesForModel("devin-cli", model), model).toMatchObject({
+        contextWindow: 200000,
+        maxOutput: 128000,
+        vision: true,
+        reasoning: true,
+        thinkingFormat: "openai",
+      });
+    }
+    expect(getCapabilitiesForModel("dv", "gpt-5.5-high").contextWindow).toBe(200000);
+    expect(getCapabilitiesForModel("devin", "gpt-5.5-high").contextWindow).toBe(200000);
+  });
+
+  // gpt-5.4 is where the 1.05M window starts and the mini/nano tiers are the
+  // exception that stayed at 400k. Pattern resolution is first-match-wins, so this
+  // is really a guard on the ORDER of the entries: move the tier patterns above
+  // the mini/nano ones and both tiers silently report 1.05M.
+  it("splits the 1.05M tiers from the 400k ones", () => {
+    for (const model of [
+      "gpt-5.4", "gpt-5.4-pro", "gpt-5.5", "gpt-5.5-pro", "gpt-5.6", "gpt-5.6-luna", "gpt-5.6-terra",
+    ]) {
+      expect(getCapabilitiesForModel("openai", model).contextWindow, model).toBe(API_WINDOW);
+    }
+    for (const model of ["gpt-5.4-mini", "gpt-5.4-nano", "gpt-5", "gpt-5.1", "gpt-5.2", "gpt-5.3-codex"]) {
+      expect(getCapabilitiesForModel("openai", model).contextWindow, model).toBe(LEGACY_GPT5_WINDOW);
+    }
+  });
+});
+
+// Copilot is "github" locally and "github-copilot" upstream. Without that mapping
+// build() resolved no upstream provider for it and skipped every Copilot model, so
+// the daily models.dev sync could never correct a stale hand-written number — which
+// is how the gpt-6 window stayed 3.9x wrong without anything noticing.
+describe("models.dev sync reaches GitHub Copilot", () => {
+  it("maps the local github id onto the upstream github-copilot id", () => {
+    expect(PROVIDER_ALIASES.github).toBe("github-copilot");
+  });
+
+  it("records a Copilot limit that disagrees with the local tables", () => {
+    // Copilot caps Claude output at 32k where the local floor assumes 64k.
+    const upstream = {
+      "github-copilot": {
+        models: { "claude-sonnet-4.6": { limit: { context: 200000, output: 32000 } } },
+      },
+    };
+    const entries = [
+      { provider: "github", model: "claude-sonnet-4.6", current: { contextWindow: 200000, maxOutput: 64000 } },
+    ];
+
+    // Context agrees, so only the output delta is recorded — and it is filed under
+    // the local id, which is what the reader looks up.
+    expect(build(upstream, entries).providers.github).toEqual({
+      "claude-sonnet-4.6": { maxOutput: 32000 },
+    });
+  });
+
+  it("applies those Copilot deltas to an exact MODEL_CAPABILITIES id", () => {
+    // claude-sonnet-4.6 is canonical-exact (128k). Without refine() on that
+    // path the 32k Copilot delta from build() would never be read.
+    setCatalogSource({
+      getModalities: () => null,
+      getLimits: (provider, model) =>
+        provider === "github" && model === "claude-sonnet-4.6" ? { maxOutput: 32000 } : null,
+    });
+    try {
+      expect(getCapabilitiesForModel("github", "claude-sonnet-4.6").maxOutput).toBe(32000);
+      expect(getCapabilitiesForModel("claude", "claude-sonnet-4.6").maxOutput).toBe(128000);
+    } finally {
+      setCatalogSource(null);
+    }
+  });
+});

tests/unit/v1-models-combo-context.test.js

@@ -0,0 +1,83 @@
+import { describe, expect, it, vi } from "vitest";
+import { setCatalogSource } from "../../open-sse/providers/capabilities.js";
+
+const db = vi.hoisted(() => ({
+  getProviderConnections: vi.fn(),
+  getCombos: vi.fn(),
+  getCustomModels: vi.fn(async () => []),
+  getModelAliases: vi.fn(async () => ({})),
+}));
+
+vi.mock("@/lib/localDb", () => db);
+vi.mock("@/lib/disabledModelsDb", () => ({
+  getDisabledModels: vi.fn(async () => ({})),
+}));
+
+const { buildModelsList } = await import("../../src/app/api/v1/models/route.js");
+
+const syncedLimits = { contextWindow: 180000, maxOutput: 16000 };
+
+async function modelsWithCombo(providerId, modelId, combos) {
+  db.getProviderConnections.mockResolvedValue([{
+    id: 1,
+    provider: providerId,
+    isActive: true,
+    providerSpecificData: { enabledModels: [modelId] },
+  }]);
+  db.getCombos.mockResolvedValue(combos);
+  setCatalogSource({
+    getModalities: () => null,
+    getLimits: (provider, model) =>
+      provider === providerId && model === modelId ? syncedLimits : null,
+  });
+  try {
+    return await buildModelsList(["llm"]);
+  } finally {
+    setCatalogSource(null);
+  }
+}
+
+describe("/v1/models combo limits", () => {
+  it.each([
+    ["ocg", "opencode-go", "mimo-v2.5"],
+    ["xmtp", "xiaomi-tokenplan", "mimo-v2.5"],
+    ["ps", "poolside", "custom-model"],
+    ["ds", "deepseek", "deepseek-chat"],
+  ])("uses the real provider for a %s UI-alias seat", async (uiAlias, providerId, modelId) => {
+    const combo = { name: "ui-alias-combo", models: [`${uiAlias}/${modelId}`] };
+    const models = await modelsWithCombo(providerId, modelId, [combo]);
+    const published = models.find((model) => model.id === combo.name);
+
+    expect(published).toMatchObject({
+      context_length: 180000,
+      max_completion_tokens: 16000,
+      capabilities: { contextWindow: 180000, maxOutput: 16000 },
+    });
+  });
+
+  it("carries provider-scoped limits through a nested combo", async () => {
+    const models = await modelsWithCombo("opencode-go", "mimo-v2.5", [
+      { name: "inner-combo", models: ["ocg/mimo-v2.5"] },
+      { name: "outer-combo", models: ["inner-combo"] },
+    ]);
+    const outer = models.find((model) => model.id === "outer-combo");
+
+    expect(outer).toMatchObject({
+      context_length: 180000,
+      max_completion_tokens: 16000,
+      capabilities: { contextWindow: 180000, maxOutput: 16000 },
+    });
+  });
+
+  it("publishes Devin CLI's 200k limit for a dv combo seat", async () => {
+    const models = await modelsWithCombo("devin-cli", "gpt-5.5-high", [
+      { name: "devin-combo", models: ["dv/gpt-5.5-high"] },
+    ]);
+    const combo = models.find((model) => model.id === "devin-combo");
+
+    expect(combo).toMatchObject({
+      context_length: 200000,
+      capabilities: { contextWindow: 200000 },
+    });
+  });
+});