fix(devin): report the context windows Cognition actually serves

scripts/test-layout/layout.json

@@ -856,6 +856,7 @@
     "model-pinned-effort.test.ts": "codex-integration",
     "model-presets.test.ts": "providers",
     "model-rename-migration.test.ts": "providers",
+    "context-window-seed-repair.test.ts": "providers",
     "model-selection-guidance.test.ts": "cli",
     "model-visibility-management-api.test.ts": "codex-integration",
     "models-feedback-callback.test.ts": "gui",

src/adapters/devin-cli/models.ts

@@ -21,3 +21,37 @@ export const DEVIN_CLI_MODELS = [
   "glm-5-3-low",
   "kimi-k3-high",
 ] as const;
+
+/**
+ * Context windows for the CLI roster, in the same effort-suffixed ids the CLI
+ * accepts.
+ *
+ * Without this the picker fell back to the 128k default for every Devin CLI
+ * model, including `swe-2`, which is the roster's own default — so the one
+ * model most sessions ran reported less than half its real window.
+ *
+ * The numbers come from Cognition's `GetCascadeModelConfigs` catalog
+ * (`ClientModelConfig` field #18), which is the only first-party source: the
+ * Devin CLI and Desktop model pages, the SWE-2 announcement, and the Windsurf
+ * model reference all list these models without a window. The CLI is a separate
+ * product from the cloud, but Cognition documents the same models on both and
+ * describes no per-surface difference — the SWE-2 announcement ships it to
+ * Desktop, CLI, Web, and Fusion in one sentence — so the catalog's figure is
+ * used for both rather than inventing a second table.
+ *
+ * ACP has no discovery call, so unlike the cloud provider this cannot be
+ * refreshed live; it needs updating when the roster above does.
+ */
+export const DEVIN_CLI_MODEL_CONTEXT_WINDOWS: Record<string, number> = {
+  "swe-2": 262_000,
+  "swe-2-high": 262_000,
+  "claude-opus-5-medium": 1_000_000,
+  "claude-fable-5-1-medium": 1_000_000,
+  "claude-sonnet-5-medium": 1_000_000,
+  "gpt-6-astra-medium": 1_000_000,
+  "gpt-5-6-sol-medium": 1_000_000,
+  "gemini-3-8-flash-medium": 1_048_576,
+  "glm-5-3-high": 1_048_576,
+  "glm-5-3-low": 1_048_576,
+  "kimi-k3-high": 1_048_576,
+};

src/adapters/devin/cloud-direct/catalog.ts

@@ -63,6 +63,21 @@ export interface ModelCatalogEntry {
   label: string;
   /** True when the caller's account tier cannot use this UID for chat. */
   disabled: boolean;
+  /**
+   * Maximum input tokens the account may send this model, from
+   * `ClientModelConfig` field #18.
+   *
+   * Cognition publishes no context-window numbers anywhere: not in the Devin
+   * CLI or Desktop model pages, not in the SWE-2 or SWE-1.7 announcements, and
+   * not in the Windsurf model reference, which has no such table. The only
+   * numbers on those pages are long-context PRICING thresholds, which are a
+   * different quantity. That makes this field the single first-party source,
+   * and it is per account rather than per model id.
+   *
+   * Absent when the entry omits the field, which is how a future schema change
+   * degrades: the caller keeps its static fallback instead of reporting zero.
+   */
+  contextWindow?: number;
 }
 
 export interface CacheEntry {
@@ -89,26 +104,38 @@ function flightKey(apiKey: string, host: string): string {
  * Parse a GetCascadeModelConfigsResponse buffer into a UID-keyed map.
  * A malformed catalog returns an empty map.
  */
-function parseCatalogBuffer(buf: Buffer, apiKey: string, host: string): CacheEntry {
+export function parseCatalogBuffer(buf: Buffer, apiKey: string, host: string): CacheEntry {
   // GetCascadeModelConfigsResponse #1 (repeated ClientModelConfig)
   const byUid = new Map<string, ModelCatalogEntry>();
   for (const f of iterFields(buf)) {
     if (f.num !== 1 || f.wire !== 2 || !Buffer.isBuffer(f.value)) continue;
     let label = '';
     let modelUid = '';
     let disabled = false;
+    let contextWindow = 0;
     for (const sf of iterFields(f.value as Buffer)) {
       if (sf.num === 1 && sf.wire === 2 && Buffer.isBuffer(sf.value)) {
         label = (sf.value as Buffer).toString('utf8');
       } else if (sf.num === 4 && sf.wire === 0) {
         // #4 = disabled (bool, varint 0/1)
         disabled = sf.value === 1n;
+      } else if (sf.num === 18 && sf.wire === 0) {
+        // #18 = max input tokens. Identified by dumping a live catalog and
+        // reading the varints back against models whose windows are known from
+        // their upstream vendors: 1000000 on the Claude and GPT rows, 1048576
+        // on Gemini/Kimi/GLM, 500000 on Grok, 262000 on swe-2.
+        contextWindow = Number(sf.value);
       } else if (sf.num === 22 && sf.wire === 2 && Buffer.isBuffer(sf.value)) {
         modelUid = (sf.value as Buffer).toString('utf8');
       }
     }
     if (modelUid.length > 0) {
-      byUid.set(modelUid, { modelUid, label: label || modelUid, disabled });
+      byUid.set(modelUid, {
+        modelUid,
+        label: label || modelUid,
+        disabled,
+        ...(contextWindow > 0 ? { contextWindow } : {}),
+      });
     }
   }
   return { byUid, fetchedAt: Date.now(), apiKey, host };

src/adapters/devin/live-models.ts

@@ -29,19 +29,43 @@ export const DEVIN_STATIC_MODELS = [
   "grok-4-5",
 ] as const;
 
-/** Per-model context windows for Devin/Cognition models. Source: Cognition model catalog. */
+/**
+ * Degraded-mode context windows, used only when live discovery cannot run.
+ *
+ * Every number here was read from a live `GetCascadeModelConfigs` response
+ * (`ClientModelConfig` field #18) rather than from documentation, because
+ * Cognition publishes none: the Devin CLI and Desktop model pages, the SWE-2
+ * and SWE-1.7 announcements, and the Windsurf model reference all state model
+ * names without a context window. The only published numbers are long-context
+ * pricing thresholds, which are a different quantity and were not used.
+ *
+ * The previous copy of this table was wrong for nine of its eleven rows — the
+ * three Claude models were listed at 200k against an actual 1M, `grok-4-5` at
+ * 256k against 500k, and the GPT rows at 1.05M against 1M — because it was
+ * assembled from each model's upstream vendor window instead of what Cognition
+ * actually serves. Measure the catalog when updating this; do not carry a
+ * number over from the model's original vendor.
+ */
 export const DEVIN_MODEL_CONTEXT_WINDOWS: Record<string, number> = {
-  "swe-1-7": 256_000,
-  "swe-1-7-lightning": 256_000,
-  "gpt-5-6-sol": 1_050_000,
-  "gpt-5-6-luna": 1_050_000,
-  "gpt-5-6-terra": 1_050_000,
-  "claude-opus-4-8": 200_000,
-  "claude-fable-5-1": 200_000,
-  "claude-sonnet-5": 200_000,
+  "swe-2": 262_000,
+  "swe-1-7": 262_000,
+  "swe-1-7-lightning": 202_752,
+  "swe-1-6": 200_000,
+  "gpt-5-6-sol": 1_000_000,
+  "gpt-5-6-luna": 1_000_000,
+  "gpt-5-6-terra": 1_000_000,
+  "gpt-6-astra": 1_000_000,
+  "claude-opus-4-8": 1_000_000,
+  "claude-opus-5": 1_000_000,
+  "claude-fable-5-1": 1_000_000,
+  "claude-sonnet-5": 1_000_000,
   "glm-5-2": 200_000,
-  "kimi-k2-7": 256_000,
-  "grok-4-5": 256_000,
+  "glm-5-3": 1_048_576,
+  "kimi-k2-7": 262_144,
+  "kimi-k3": 1_048_576,
+  "gemini-3-8-flash": 1_048_576,
+  "grok-4-5": 500_000,
+  "grok-4-6": 500_000,
 };
 
 /**
@@ -62,7 +86,7 @@ export function collapseDevinModelUid(uid: string): string {
 }
 
 export type DevinUsableModelsResult =
-  | { ok: true; models: string[] }
+  | { ok: true; models: string[]; contextWindows: Record<string, number> }
   | { ok: false; error: "auth" | "http" | "empty" | "unknown"; detail?: string };
 
 /**
@@ -80,15 +104,27 @@ export async function fetchDevinUsableModels(opts: {
     const catalog = await getCachedCatalog(opts.apiKey, host, opts.signal);
     if (!catalog) return { ok: false, error: "empty" };
     const bases = new Set<string>();
+    const contextWindows: Record<string, number> = {};
     for (const entry of catalog.byUid.values()) {
       if (entry.disabled) continue;
       // Skip internal enum constants (e.g. MODEL_GPT_5_2_LOW, MODEL_PRIVATE_*).
       // Real chat model UIDs are lowercase dashed strings (swe-1-7, gpt-5-6-sol).
       if (entry.modelUid.startsWith("MODEL_")) continue;
-      bases.add(collapseDevinModelUid(entry.modelUid));
+      const base = collapseDevinModelUid(entry.modelUid);
+      bases.add(base);
+      if (entry.contextWindow && entry.contextWindow > 0) {
+        // Variants of one base can disagree: the opt-in `-1m` rows report a
+        // larger window than the plain row of the same base, and both collapse
+        // here because `1m` is an effort token. Keep the smallest, because the
+        // base id routes to the plain variant — advertising the long-context
+        // number would promise a window the request the picker actually sends
+        // cannot use.
+        const seen = contextWindows[base];
+        contextWindows[base] = seen === undefined ? entry.contextWindow : Math.min(seen, entry.contextWindow);
+      }
     }
     if (bases.size === 0) return { ok: false, error: "empty" };
-    return { ok: true, models: [...bases].sort() };
+    return { ok: true, models: [...bases].sort(), contextWindows };
   } catch (error) {
     const message = error instanceof Error ? error.message : String(error);
     if (/unauth|401|invalid token|login/i.test(message)) return { ok: false, error: "auth", detail: message };

src/codex/catalog/provider-fetch.ts

@@ -1722,11 +1722,23 @@ async function fetchProviderModelsWithAuth(
     if (liveResult.ok) {
       // Live catalog is the source of truth — use the discovered base models
       // directly, not a filtered subset of the static seed.
-      const result = liveResult.models.map((id) => ({
-        id,
-        provider: name,
-        ...catalogHintsFromProviderConfig(name, prov, id, contextCap, metadataModelIdCaseFold, captured.effectiveAlias),
-      }) as CatalogModel);
+      //
+      // That extends to the context window. Cognition publishes no window
+      // anywhere, so the per-account catalog is the only first-party number,
+      // and the shipped static table is a degraded-mode guess that was wrong
+      // for nine of its eleven rows. The live value is applied first and the
+      // config hints run after it, so an explicit per-model override and an
+      // enabled Context cap still win — this only replaces the number nobody
+      // chose.
+      const result = liveResult.models.map((id) => {
+        const liveWindow = liveResult.contextWindows[id];
+        return {
+          id,
+          provider: name,
+          ...(liveWindow ? { contextWindow: liveWindow } : {}),
+          ...catalogHintsFromProviderConfig(name, prov, id, contextCap, metadataModelIdCaseFold, captured.effectiveAlias),
+        } as CatalogModel;
+      });
       const forCache = withConfiguredRetention(result, { retainComboTargets: false });
       if (!setCached(name, forCache, Date.now(), cacheGeneration)) {
         return observed(withConfiguredRetention(configured), "degraded");

src/providers/model-rename-startup.ts

@@ -1,7 +1,25 @@
 import { mutatePersistedConfig } from "../config";
 import { projectModelRenames } from "./model-rename-migration";
+import { projectStaleContextWindows } from "./stale-context-window-migration";
 import type { OcxConfig } from "../types";
 
+/**
+ * The startup projection: registry model renames, then the context-window
+ * repair. Both fix a saved row the registry can no longer reach on its own —
+ * `enrichProviderFromRegistry` backfills a missing field and never rewrites a
+ * present one — so they share this pass rather than adding a second boot step
+ * with its own persistence, adopt, and failure handling.
+ */
+export function projectStartupConfigRepairs(config: OcxConfig): ReturnType<typeof projectModelRenames> {
+  const renames = projectModelRenames(config);
+  const windows = projectStaleContextWindows(renames.config);
+  return {
+    config: windows.config,
+    changed: renames.changed || windows.changed,
+    warnings: [...renames.warnings, ...windows.warnings],
+  };
+}
+
 export interface ModelRenameStartupDeps {
   project: typeof projectModelRenames;
   /** Injected writer for tests and callers that own their own persistence. */
@@ -59,7 +77,7 @@ function adoptConfig(target: OcxConfig, source: OcxConfig): void {
  */
 export function runModelRenameStartupMigration(
   config: OcxConfig,
-  deps: ModelRenameStartupDeps = { project: projectModelRenames },
+  deps: ModelRenameStartupDeps = { project: projectStartupConfigRepairs },
 ): OcxConfig {
   const projection = deps.project(structuredClone(config));
   if (!projection.changed) {

src/providers/registry.ts

@@ -1,7 +1,7 @@
 import type { CodexAccountMode, FastWire, OcxProviderConfig } from "../types";
 import { fastWireDeclarationError } from "./fastwire";
 import { KIRO_MODELS, KIRO_MODEL_CONTEXT_WINDOWS, KIRO_MODEL_REASONING_EFFORTS } from "./kiro-models";
-import { DEVIN_CLI_DEFAULT_MODEL, DEVIN_CLI_MODELS } from "../adapters/devin-cli/models";
+import { DEVIN_CLI_DEFAULT_MODEL, DEVIN_CLI_MODEL_CONTEXT_WINDOWS, DEVIN_CLI_MODELS } from "../adapters/devin-cli/models";
 import { DEVIN_MODEL_CONTEXT_WINDOWS } from "../adapters/devin/live-models";
 import { ANTIGRAVITY_MODELS, ANTIGRAVITY_MODEL_CONTEXT_WINDOWS, ANTIGRAVITY_MODEL_EFFORTS, ANTIGRAVITY_MODEL_INPUT_MODALITIES } from "./antigravity-models";
 import type { ProviderBaseUrlChoice } from "./base-url-choices";
@@ -1343,6 +1343,7 @@ export const PROVIDER_REGISTRY: readonly ProviderRegistryEntry[] = [
     note: "Drives the locally installed Devin CLI over the Agent Client Protocol (`devin acp`, newline-delimited JSON-RPC on stdio). Requires the CLI on PATH and a completed `devin auth login`; no API key is stored by opencodex. Set OPENCODEX_DEVIN_CLI_BIN to point at a specific build, and OPENCODEX_DEVIN_CLI_ALLOW_TOOLS=1 to let the CLI read and write files — the default is to refuse.",
     models: [...DEVIN_CLI_MODELS],
     defaultModel: DEVIN_CLI_DEFAULT_MODEL,
+    modelContextWindows: DEVIN_CLI_MODEL_CONTEXT_WINDOWS,
   },
   {
     id: "devin",

src/providers/stale-context-window-migration.ts

@@ -0,0 +1,92 @@
+/**
+ * Repair context windows that a saved config inherited from a wrong registry seed.
+ *
+ * `enrichProviderFromRegistry` is fill-only: it seeds `modelContextWindows` when
+ * the row has none and never rewrites it afterwards. That posture is right — a
+ * hand-tuned window must survive an upgrade — but it means a registry table that
+ * shipped WRONG numbers is frozen into every config that was saved while those
+ * numbers were current. Correcting the registry alone fixes new installs and
+ * leaves existing ones reporting the old figure forever.
+ *
+ * This rewrites one thing: a window whose saved value is still byte-for-byte the
+ * wrong number this file names, on a provider that still carries the registry's
+ * adapter. A value the user changed does not match `from` and is left alone, and
+ * nothing else in the row is touched. Same shape and the same restraint as
+ * `model-rename-migration`, for the case where the id was right and the number
+ * was not.
+ */
+import { PROVIDER_REGISTRY } from "./registry";
+import type { OcxConfig } from "../types";
+
+export interface StaleContextWindow {
+  /** Registry provider id whose saved rows may carry the wrong window. */
+  provider: string;
+  model: string;
+  /** The wrong value this migration is allowed to replace, and nothing else. */
+  from: number;
+  to: number;
+}
+
+export interface StaleContextWindowProjection {
+  config: OcxConfig;
+  changed: boolean;
+  warnings: string[];
+}
+
+/**
+ * Cognition windows corrected against a live `GetCascadeModelConfigs` response.
+ *
+ * The shipped table had been assembled from each model's ORIGINAL vendor window
+ * rather than from what Cognition serves, so the Claude rows claimed 200k against
+ * an actual 1M and Grok claimed 256k against 500k. Cognition documents no window
+ * anywhere, so the per-account catalog is the only first-party source; these are
+ * the degraded-mode figures, and live discovery supersedes them when it runs.
+ */
+export const STALE_CONTEXT_WINDOWS: readonly StaleContextWindow[] = [
+  { provider: "devin", model: "swe-1-7", from: 256_000, to: 262_000 },
+  { provider: "devin", model: "swe-1-7-lightning", from: 256_000, to: 202_752 },
+  { provider: "devin", model: "gpt-5-6-sol", from: 1_050_000, to: 1_000_000 },
+  { provider: "devin", model: "gpt-5-6-luna", from: 1_050_000, to: 1_000_000 },
+  { provider: "devin", model: "gpt-5-6-terra", from: 1_050_000, to: 1_000_000 },
+  { provider: "devin", model: "claude-opus-4-8", from: 200_000, to: 1_000_000 },
+  { provider: "devin", model: "claude-fable-5-1", from: 200_000, to: 1_000_000 },
+  { provider: "devin", model: "claude-sonnet-5", from: 200_000, to: 1_000_000 },
+  { provider: "devin", model: "kimi-k2-7", from: 256_000, to: 262_144 },
+  { provider: "devin", model: "grok-4-5", from: 256_000, to: 500_000 },
+];
+
+function providerStillMatchesRegistry(id: string, adapter: unknown): boolean {
+  const entry = PROVIDER_REGISTRY.find(row => row.id === id);
+  return entry !== undefined && entry.adapter === adapter;
+}
+
+/** Pure projection. The caller decides whether to persist. */
+export function projectStaleContextWindows(
+  config: OcxConfig,
+  entries: readonly StaleContextWindow[] = STALE_CONTEXT_WINDOWS,
+): StaleContextWindowProjection {
+  const warnings: string[] = [];
+  const repaired = new Map<string, string[]>();
+
+  for (const entry of entries) {
+    const prov = config.providers?.[entry.provider];
+    if (!prov) continue;
+    if (!providerStillMatchesRegistry(entry.provider, prov.adapter)) continue;
+    const windows = prov.modelContextWindows;
+    if (!windows || windows[entry.model] !== entry.from) continue;
+    windows[entry.model] = entry.to;
+    const list = repaired.get(entry.provider) ?? [];
+    list.push(`${entry.model} ${entry.from} -> ${entry.to}`);
+    repaired.set(entry.provider, list);
+  }
+
+  for (const [provider, list] of repaired) {
+    warnings.push(
+      `corrected ${list.length} context window(s) on "${provider}" that the saved config `
+      + `inherited from a wrong registry seed: ${list.join(", ")}.`,
+    );
+  }
+
+  return { config, changed: repaired.size > 0, warnings };
+}
+

tests/fixtures/test-layout-expected.json

@@ -690,6 +690,7 @@
   "model-pinned-effort.test.ts": "codex-integration",
   "model-presets.test.ts": "providers",
   "model-rename-migration.test.ts": "providers",
+  "context-window-seed-repair.test.ts": "providers",
   "model-selection-guidance.test.ts": "cli",
   "model-visibility-management-api.test.ts": "codex-integration",
   "models-feedback-callback.test.ts": "gui",