@bitkyc08/opencodex 2.82.0 → 2.84.0

package/src/providers/quota/vendor-probes-oauth.ts

@@ -327,9 +327,9 @@ async function readAnthropicUsageQuota(accessToken: string): Promise<ProviderQuo
  * This endpoint reports quota only. Its body carries `five_hour`, `seven_day`, the
  * model-scoped weekly buckets (`seven_day_fable`/`_opus`/`_sonnet`) and a `limits` array,
  * and **no subscription or tier field** — nor does the OAuth token response, which yields only
- * `account.uuid` and `account.email_address` (`src/oauth/anthropic.ts`). That is why
- * `OAuthAccountSummary.plan` is `null` for Anthropic rather than populated here (#3777); it is
- * a missing upstream field, not an unfinished mapping.
+ * `account.uuid` and `account.email_address` (`src/oauth/anthropic.ts`). Subscription labels
+ * instead come from the authenticated profile in `src/oauth/anthropic-identity.ts`, never
+ * from this quota response.
  *
  * A tier must not be inferred from what is here. Percentages are normalized per account, so a
  * Max x5 seat at 50% is byte-identical to a Max x20 seat at 50%, and the presence of a

package/src/providers/quota.ts

@@ -15,6 +15,7 @@ import {
 } from "./kiro-usage";
 import { mapQuotaRoster, readProviderApiKeyQuotas, type ProviderApiKeyQuota } from "./quota-key-accounts";
 import type { OcxConfig, OcxProviderConfig } from "../types";
+import { isPassiveRead } from "../lib/passive-read";
 import type { ProviderQuota, QuotaFailureCode } from "./quota-types";
 import {
   accountReportCurrent,
@@ -23,7 +24,7 @@ import {
   captureProviderReportEpochs, providerReportEpochChanged,
   cacheKeyWithAggregationState,
   getProviderQuotaReportCache,
-  hasCodexPoolProvider,
+  hasCodexPoolProvider, heldProviderQuotaReports,
   inflight,
   invalidationEpoch,
   isBuiltInChatGptForwardProvider,
@@ -229,6 +230,8 @@ export function flushProviderQuotaObservationsForTests(): Promise<void> {
 }
 
 export async function fetchProviderQuotaReports(config: OcxConfig, forceRefresh = false): Promise<ProviderQuotaResponse> {
+  // A passive read (a Remote Link Child's, #6959) shows the held reports and starts nothing.
+  if (isPassiveRead()) return heldProviderQuotaReports(Object.keys(config.providers));
   // A Pool report's cache signature and provider fetch must share one account snapshot.
   // Preserve force semantics when deciding whether that snapshot refreshes upstream data.
   const prefetchedCodexSnapshot = hasCodexPoolProvider(config)

package/src/providers/registry/entries-core.ts

@@ -32,12 +32,12 @@ import {
   OPENAI_API_GPT56_MAX_INPUT_TOKENS,
   OPENAI_API_GPT56_VIRTUAL_MODELS,
   OPENAI_API_GPT56_REASONING_EFFORTS,
-  META_MUSE_REASONING_EFFORTS,
-  META_MUSE_REASONING_EFFORT_MAP,
+  META_MODEL_API_REASONING_EFFORTS,
+  META_MODEL_API_REASONING_EFFORT_MAPS,
   META_MUSE_CODE_REASONING_EFFORTS,
   META_MUSE_CODE_REASONING_EFFORT_MAP,
   META_MUSE_CONTEXT_WINDOW,
-  META_MUSE_MODELS,
+  META_MODEL_API_MODELS,
   OPENAI_DAYBREAK_MODELS,
   OPENAI_DAYBREAK_CONTEXT_WINDOWS,
   OPENAI_DAYBREAK_MAX_INPUT_TOKENS,
@@ -94,7 +94,7 @@ import {
  * Only the models Anthropic documents for the lane are classified; Opus 4.6 silently runs
  * standard and Opus 4.7, Sonnet, Haiku and Fable reject `speed`, so they and future ids stay
  * unclassified. Source: https://platform.claude.com/docs/en/build-with-claude/fast-mode
- * (2026-09-23) and the live probe in devlog/_plan/260923_anthropic_fast_speed.
+ * (2026-09-23) and the live probe in devlog/_fin/260923_anthropic_fast_speed.
  */
 const ANTHROPIC_FAST_WIRE = Object.freeze({
   kind: "anthropic-speed" as const,
@@ -209,7 +209,7 @@ export const PROVIDER_REGISTRY_CORE: readonly ProviderRegistryEntry[] = [
   },
   {
     // The canonical Cognition account provider, after absorbing `devin-cli`
-    // (devlog/_plan/260913_devin_provider_merge). The two ids were the same
+    // (devlog/_fin/260913_devin_provider_merge). The two ids were the same
     // `devin` adapter, the same server.codeium.com api-server, and the same
     // `devin-session-token$<JWT>` credential — only the account source
     // differed: this entry did an Auth0 browser sign-in while `devin-cli`
@@ -273,7 +273,7 @@ export const PROVIDER_REGISTRY_CORE: readonly ProviderRegistryEntry[] = [
     // the field but answers service_tier "default" — a live downgrade, not a fast tier.
     // Unlisted and future-discovered ids stay unclassified.
     // grok-4.7 applied and confirmed priority on OAuth Responses in the 2026-09-23
-    // live probe: devlog/_plan/260923_grok47_parity/010_probe-evidence.md.
+    // live probe: devlog/_fin/260923_grok47_parity/010_probe-evidence.md.
     modelSupportsServiceTier: {
       "grok-4.7": true,
       "grok-4.6": true,
@@ -320,7 +320,7 @@ export const PROVIDER_REGISTRY_CORE: readonly ProviderRegistryEntry[] = [
     models: XAI_MODELS,
     // grok-4.7-build-fast arrives only through OAuth discovery: xAI's Grok 4.7 Fast, "the same model
     // served on faster infrastructure" (docs.x.ai/developers/grok-4-7). Probed 2026-09-30
-    // (devlog/_plan/260930_grok47_build_unify/010_probe-evidence.md): identical effort ladder, image
+    // (devlog/_fin/260930_grok47_build_unify/010_probe-evidence.md): identical effort ladder, image
     // input, 500k limit and advertised defaults, 1.5-1.7x faster, ~2x the ticks per output token. It is
     // not a second model row: shouldExposeProviderModel hides it, and grok-4.7's Fast selection on OAuth
     // serializes it (src/providers/xai-fast-model.ts). It keeps grok-4.7's facts in the lists below so an
@@ -450,7 +450,7 @@ export const PROVIDER_REGISTRY_CORE: readonly ProviderRegistryEntry[] = [
     // documents high as the 4.6 default but no multi-agent default, so do not invent one.
     modelReasoningEfforts: {
       // 2026-09-23 live probe accepted low..xhigh and rejected max on both wires;
-      // devlog/_plan/260923_grok47_parity/010_probe-evidence.md.
+      // devlog/_fin/260923_grok47_parity/010_probe-evidence.md.
       "grok-4.7": ["low", "medium", "high", "xhigh"],
       "grok-4.7-build-fast": ["low", "medium", "high", "xhigh"],
       "grok-4.6": ["low", "medium", "high", "xhigh"],
@@ -460,7 +460,7 @@ export const PROVIDER_REGISTRY_CORE: readonly ProviderRegistryEntry[] = [
     modelDefaultReasoningEfforts: { "grok-4.7": "high", "grok-4.7-build-fast": "high", "grok-4.6": "high" },
     modelContextWindows: {
       // 500k confirmed by context_length_exceeded:
-      // devlog/_plan/260923_grok47_parity/010_probe-evidence.md.
+      // devlog/_fin/260923_grok47_parity/010_probe-evidence.md.
       "grok-4.7": 500_000,
       "grok-4.7-build-fast": 500_000,
       "grok-4.6": 500_000,
@@ -736,7 +736,7 @@ export const PROVIDER_REGISTRY_CORE: readonly ProviderRegistryEntry[] = [
     authKind: "key",
     dashboardUrl: "https://dev.meta.ai/docs/authentication",
     defaultModel: "muse-spark-1.3",
-    models: META_MUSE_MODELS,
+    models: META_MODEL_API_MODELS,
     // Static roster: no authenticated /v1/models payload was ever observed (the only
     // contact was an unauthenticated GET returning 401 invalid_api_key), and Meta serves
     // non-agent families on this same base URL. Turning discovery on would publish an
@@ -746,20 +746,20 @@ export const PROVIDER_REGISTRY_CORE: readonly ProviderRegistryEntry[] = [
     // without this, registry transport canonicalization would retarget it and send their
     // saved key to Meta.
     preserveCustomDestination: true,
-    modelContextWindows: Object.fromEntries(META_MUSE_MODELS.map(id => [id, META_MUSE_CONTEXT_WINDOW])),
+    modelContextWindows: Object.fromEntries(META_MODEL_API_MODELS.map(id => [id, META_MUSE_CONTEXT_WINDOW])),
     // text+image only. Meta also documents video, audio (degraded on 1.3), and PDF, but
     // the catalog modality enum is text/image and over-advertising poisons the exported
     // client config (see tests/codex-integration/catalog-input-modality-enum.test.ts).
-    modelInputModalities: Object.fromEntries(META_MUSE_MODELS.map(id => [id, ["text", "image"] as ["text", "image"]])),
-    modelReasoningEfforts: Object.fromEntries(META_MUSE_MODELS.map(id => [id, META_MUSE_REASONING_EFFORTS])),
-    modelReasoningEffortMap: Object.fromEntries(META_MUSE_MODELS.map(id => [id, META_MUSE_REASONING_EFFORT_MAP])),
+    modelInputModalities: Object.fromEntries(META_MODEL_API_MODELS.map(id => [id, ["text", "image"] as ["text", "image"]])),
+    modelReasoningEfforts: META_MODEL_API_REASONING_EFFORTS,
+    modelReasoningEffortMap: META_MODEL_API_REASONING_EFFORT_MAPS,
     // No defaultMaxOutputTokens: Meta publishes none. The only number in its docs
     // (131072) appears inside a third-party config sample, and the protocol pages call
     // the real limit "model-dependent".
     // Meta names its variable MODEL_API_KEY, but the env var opencodex reads is derived
     // from the provider id (META_MODEL_API_KEY). Saying only Meta's name would send a
     // user to export a variable this proxy never reads.
-    note: "Pay-as-you-go Meta Model API. Get a key at https://dev.meta.ai (Meta calls it MODEL_API_KEY; export it here as META_MODEL_API_KEY) — a Meta developer account needs a payment method before it can serve requests, and every call is metered per token. A Muse Code subscription does NOT work here: Meta scopes that credential to the Muse Code CLI and bills any other key pay-as-you-go (dev.meta.ai/docs/muse-code/subscriptions). The Contributor tier (muse-spark-1.3-contributor) is cheap because Meta trains on your prompts — about 92% off input, 95% off output, 99% off cached input; do not send confidential material through it. Muse Spark is also reachable through resellers: command-code carries both tiers, opencode-go serves only muse-spark-1.3-contributor.",
+    note: "Pay-as-you-go Meta Model API. Get a key at https://dev.meta.ai (Meta calls it MODEL_API_KEY; export it here as META_MODEL_API_KEY) — a Meta developer account needs a payment method before it can serve requests, and every call is metered per token. A Muse Code subscription does NOT work here: Meta scopes that credential to the Muse Code CLI and bills any other key pay-as-you-go (dev.meta.ai/docs/muse-code/subscriptions). The Contributor tier (muse-spark-1.3-contributor) is cheap because Meta trains on your prompts — about 92% off input, 95% off output, 99% off cached input; do not send confidential material through it. Muse Wonder (muse-wonder-1.0, preview) is Standard mode only on this API ($4 input / $20 output per 1M tokens, no Contributor tier) and accepts reasoning effort up to max; the Spark rows stop at xhigh. Muse Spark is also reachable through resellers: command-code carries both tiers, opencode-go serves only muse-spark-1.3-contributor.",
   },
   /* [Decision Log]
   - 목적과 의도: Let an operator who already signed the Muse Code CLI in reach Muse Spark with that credential, instead of provisioning a second key.
@@ -787,14 +787,14 @@ export const PROVIDER_REGISTRY_CORE: readonly ProviderRegistryEntry[] = [
     oauthId: "meta-muse",
     dashboardUrl: "https://dev.meta.ai",
     defaultModel: "muse-spark-1.3",
-    models: META_MUSE_MODELS,
+    models: META_MODEL_API_MODELS,
     // Same reason as meta-model: the authenticated roster carries muse-image-1.0 and
     // muse-voice-transcribe-1.0, which this Responses-agent provider cannot drive.
     liveModels: false,
-    modelContextWindows: Object.fromEntries(META_MUSE_MODELS.map(id => [id, META_MUSE_CONTEXT_WINDOW])),
-    modelInputModalities: Object.fromEntries(META_MUSE_MODELS.map(id => [id, ["text", "image"] as ["text", "image"]])),
-    modelReasoningEfforts: Object.fromEntries(META_MUSE_MODELS.map(id => [id, META_MUSE_CODE_REASONING_EFFORTS])),
-    modelReasoningEffortMap: Object.fromEntries(META_MUSE_MODELS.map(id => [id, META_MUSE_CODE_REASONING_EFFORT_MAP])),
+    modelContextWindows: Object.fromEntries(META_MODEL_API_MODELS.map(id => [id, META_MUSE_CONTEXT_WINDOW])),
+    modelInputModalities: Object.fromEntries(META_MODEL_API_MODELS.map(id => [id, ["text", "image"] as ["text", "image"]])),
+    modelReasoningEfforts: Object.fromEntries(META_MODEL_API_MODELS.map(id => [id, META_MUSE_CODE_REASONING_EFFORTS])),
+    modelReasoningEffortMap: Object.fromEntries(META_MODEL_API_MODELS.map(id => [id, META_MUSE_CODE_REASONING_EFFORT_MAP])),
     note: "Signs in to Meta with a browser device code on any platform, then mints the Muse Code subscription key. That grant is reimplemented from the one the Muse Code CLI performs and has NOT been exercised against Meta from OpenCodex, so treat the first login as unverified. If the Muse Code CLI is already signed in on macOS, the existing key is imported instead of starting a new grant. A pasted key from https://dev.meta.ai still works as a fallback when a device login cannot complete, and faces the same format check and live validation. A device login authenticates as Meta own Muse Code client, which is a stronger claim than reusing a key the CLI already minted. Meta scopes that credential to the Muse Code CLI, so this is an UNSUPPORTED use: Meta does not authorize subscription coverage outside its own CLI, how these calls settle is not observable from the API, and you should treat every call as billable against your account. The key, imported or pasted, is copied into OpenCodex's auth store. For an account signed in with the device login, OpenCodex refreshes Meta's subscription windows on demand from the same key endpoint the login uses, at most once every five minutes. For an imported or pasted key there is no endpoint to query them on demand, so OpenCodex reads them from streaming responses and shows the last observed value with its age; refreshing one then requires another streaming turn, and translated (non-passthrough) turns report none. Rate limits apply per team, not per key. For a supported path use the meta-model provider with your own key (export it as META_MODEL_API_KEY).",
   },
   {

package/src/providers/registry/entries-extended.ts

@@ -1176,6 +1176,10 @@ export const PROVIDER_REGISTRY_EXTENDED: readonly ProviderRegistryEntry[] = [
       "openai/o3-pro": ["text"],
     },
     noVisionModels: ["openai/o3-pro"],
+    // #6877: on the Responses wire OpenGateway answers 400 once history replays native
+    // custom_tool_call/custom_tool_call_output items (Codex code-mode exec), and accepts the
+    // same request after the custom tool is lowered to a function. Same contract as xAI.
+    supportsResponsesCustomTools: false,
     // Chat is the broad wire (77/80 active rows). o3-pro is Responses-only, on all
     // inbounds. Only the two live-verified Sionic models bypass translation for Codex.
     // Other OpenAI/Anthropic/Gemini rows stay on Chat until their Responses tool replay
@@ -1464,7 +1468,6 @@ export const PROVIDER_REGISTRY_EXTENDED: readonly ProviderRegistryEntry[] = [
     models: [...QODER_GLOBAL_MODELS],
     liveModels: true,
     reasoningEfforts: [...QODER_REASONING_EFFORTS],
-    noVisionModels: [...QODER_GLOBAL_MODELS],
     note: "Official Qoder Global CLI using QODER_PERSONAL_ACCESS_TOKEN. Models are discovered per account with `qoder --list-models`; the documented roster is a degraded fallback. The CLI runs single-turn with tools, MCP, settings hooks, and session persistence disabled. Requires `npm install -g @qoder-ai/qodercli`.",
   },
   {

package/src/providers/registry/model-seeds.ts

@@ -129,7 +129,7 @@ export const ZAI_GLM_5X_SIDECAR_VISION_MODELS = ZAI_GLM_5X_MODELS.filter(id => i
  *
  * Authoritative source: `GET https://api.z.ai/api/v1/models` returns `input_modalities:
  * ["text"]` for glm-5.3 and `["text", "image"]` for glm-5.3-flash (captured in
- * devlog/_plan/260912_zcode_protocol_and_catalog/evidence/zai-responses-models.json).
+ * devlog/_fin/260912_zcode_protocol_and_catalog/evidence/zai-responses-models.json).
  * docs.z.ai/devpack/latest-model says the same in prose: "GLM-5.3 is a text-only model...
  * GLM-5.3-FLASH is a multimodal model". Upstream also lists video and file for Flash;
  * neither the internal vocabulary nor the export vocabulary can express them, so `image`
@@ -158,7 +158,7 @@ export const ZAI_GLM_5X_REASONING_EFFORTS: Record<string, string[]> = {
 // 260930 MiniMax-M3.1-Flash-Preview: Token Plan / MiniMax Code only, 1M context, thinking
 // always on (effort none or thinking disabled answers 400 code 2013), omitted effort = max.
 // It returns thinking as reasoning_content and ignores reasoning_split. The live /v1/models
-// roster does not list it yet. Evidence: devlog/_plan/260930_minimax_m31_flash_preview/.
+// roster does not list it yet. Evidence: devlog/_fin/260930_minimax_m31_flash_preview/.
 export const MINIMAX_M31_FLASH_PREVIEW = "MiniMax-M3.1-Flash-Preview";
 export const MINIMAX_MODELS = [
   MINIMAX_M31_FLASH_PREVIEW,
@@ -255,6 +255,23 @@ export const META_MUSE_CODE_REASONING_EFFORT_MAP: Record<string, string> = Objec
 /** Both Muse Spark 1.3 tiers publish a 1,048,576-token window (dev.meta.ai/docs/models). */
 export const META_MUSE_CONTEXT_WINDOW = 1_048_576;
 export const META_MUSE_MODELS = ["muse-spark-1.3", "muse-spark-1.3-contributor"];
+/*
+ * Muse Wonder (preview), announced 2026-10-12 (research.meta.ai/blog/muse-wonder). Per
+ * dev.meta.ai/docs/models and dev.meta.ai/docs/pricing-rate-limits, verified 2026-10-13:
+ * id muse-wonder-1.0, a 1,048,576-token window, reasoning effort "minimal" through "max",
+ * and Standard mode only on the Model API (no Contributor variant). Muse Wonder Full is
+ * reachable only inside Muse Code subscriptions and has no Model API id, so it is absent.
+ */
+export const META_MUSE_WONDER_MODEL = "muse-wonder-1.0";
+/** Roster both direct Meta providers serve: the Spark 1.3 tiers plus Muse Wonder. */
+export const META_MODEL_API_MODELS = [...META_MUSE_MODELS, META_MUSE_WONDER_MODEL];
+/** Model API ladders: Spark stops at xhigh; Wonder documents minimal..max. */
+export const META_MODEL_API_REASONING_EFFORTS: Record<string, string[]> = Object.fromEntries(
+  META_MODEL_API_MODELS.map(id => [id, id === META_MUSE_WONDER_MODEL ? META_MUSE_CODE_REASONING_EFFORTS : META_MUSE_REASONING_EFFORTS]),
+);
+export const META_MODEL_API_REASONING_EFFORT_MAPS: Record<string, Record<string, string>> = Object.fromEntries(
+  META_MODEL_API_MODELS.map(id => [id, id === META_MUSE_WONDER_MODEL ? META_MUSE_CODE_REASONING_EFFORT_MAP : META_MUSE_REASONING_EFFORT_MAP]),
+);
 /**
  * Daybreak program aliases. These `-latest` ids are the stable contract: OpenAI repoints
  * them at newer snapshots over time (red -> gpt-5.6-cyber, blue -> gpt-5.6-sol as of
@@ -416,7 +433,7 @@ export const COMMAND_CODE_IMAGE_MODELS = [
   // Live 2026-09-23 3x3 random-color grid (180x180) via ocx 2.62.0:
   // 4.7 read 9/9 in user messages and tool results; 4.6 read 9/9 and 8/9.
   // Neither route requested a vision sidecar. Evidence:
-  // devlog/_plan/260923_grok47_parity/010_probe-evidence.md.
+  // devlog/_fin/260923_grok47_parity/010_probe-evidence.md.
   "xai/grok-4.6",
   "xai/grok-4.7",
   // Native Z.AI VLM (docs.z.ai/guides/vlm/glm-5.3-flash). This exact id is already

package/src/providers/request-pacing.ts

@@ -33,6 +33,12 @@ export const REQUEST_PACING_CONCURRENCY_RETRY_AFTER_SECONDS = 5;
 let maxQueueDepth = REQUEST_PACING_MAX_QUEUE_DEPTH;
 let maxQueueAgeMs = REQUEST_PACING_MAX_QUEUE_AGE_MS;
 
+// Optional owners register internal response markers; only this owned body rewrap carries them.
+const responseProvenance = new Set<WeakMap<Response, unknown>>();
+export function registerProviderResponseProvenance(markers: WeakMap<Response, unknown>): void {
+  responseProvenance.add(markers);
+}
+
 export type RequestPacingQueueOverloadReason = "queue_full" | "queue_expired";
 
 export class RequestPacingQueueOverloadError extends Error {
@@ -588,6 +594,9 @@ export function trackProviderRequestSlotBody(
   // Retry helpers mark the response they return from, and recovery decisions key on these
   // identity markers; the rewrap must not make a non-replayable response look replayable.
   if (isNonReplayableResponse(response)) markResponseNonReplayable(wrapped);
+  for (const markers of responseProvenance) {
+    if (markers.has(response)) markers.set(wrapped, markers.get(response));
+  }
   return carryReplayRefusal(response, wrapped);
 }
 

package/src/providers/stale-no-vision-seed-migration.ts

@@ -0,0 +1,72 @@
+/**
+ * Drop a saved `noVisionModels` list that is still, byte for byte, a retired registry seed.
+ *
+ * `enrichProviderFromRegistry` fills `noVisionModels` all-or-nothing, so a provider row saved
+ * while the registry seeded a blanket list keeps that list after the registry stops seeding it.
+ * Qoder Global seeded its whole static discovery roster as text-only until the Global adapter
+ * began forwarding image input to the Qoder CLI (#6906). The saved copy would keep every image
+ * on the vision sidecar path, stripped or captioned before the adapter could forward it.
+ *
+ * This projection removes one thing: a saved list identical, in order, to a seed this file names,
+ * on a row the registry still owns (same adapter, and the registry transport rule enrichment uses),
+ * and only while the registry no longer seeds the list at all. Any edited list does not match and
+ * is left alone. `modelCapabilities` and `modelInputModalities` are untouched; they are where a
+ * deliberate per-model text-only declaration belongs.
+ */
+import { PROVIDER_REGISTRY, providerMatchesRegistryTransport } from "./registry";
+import type { OcxConfig } from "../types";
+
+export interface StaleNoVisionSeed {
+  /** Registry provider id whose saved rows may carry the retired list. */
+  provider: string;
+  /** Every list the registry ever seeded, in seed order; only an exact match is removed. */
+  seeds: readonly (readonly string[])[];
+}
+
+export interface StaleNoVisionSeedProjection {
+  config: OcxConfig;
+  changed: boolean;
+  warnings: string[];
+}
+
+export const STALE_NO_VISION_SEEDS: readonly StaleNoVisionSeed[] = [
+  {
+    provider: "qoder",
+    seeds: [
+      // 2026-09-03 (Qoder Global PAT provider) until the DeepSeek V4.1 transition (#4282).
+      ["Qwen3.8-Max", "Qwen3.7-Max", "Qwen3.7-Plus", "Kimi-K3", "Kimi-K2.7-Code", "GLM-5.3", "GLM-5.2", "DeepSeek-V4-Pro"],
+      // #4282 until Global image input (#6906).
+      ["Qwen3.8-Max", "Qwen3.7-Max", "Qwen3.7-Plus", "Kimi-K3", "Kimi-K2.7-Code", "GLM-5.3", "GLM-5.2"],
+    ],
+  },
+];
+
+function sameList(current: readonly unknown[], expected: readonly string[]): boolean {
+  return current.length === expected.length && current.every((value, index) => value === expected[index]);
+}
+
+/** Pure projection. The caller decides whether to persist. */
+export function projectStaleNoVisionSeeds(
+  config: OcxConfig,
+  entries: readonly StaleNoVisionSeed[] = STALE_NO_VISION_SEEDS,
+): StaleNoVisionSeedProjection {
+  const warnings: string[] = [];
+  let changed = false;
+  for (const entry of entries) {
+    const prov = config.providers?.[entry.provider];
+    if (!prov || !Array.isArray(prov.noVisionModels)) continue;
+    const registry = PROVIDER_REGISTRY.find(row => row.id === entry.provider);
+    // While the registry still seeds a list, the saved one is not stale by this file's rule.
+    if (!registry || registry.noVisionModels !== undefined || registry.adapter !== prov.adapter) continue;
+    if (!providerMatchesRegistryTransport(entry.provider, prov)) continue;
+    const saved = prov.noVisionModels;
+    if (!entry.seeds.some(seed => sameList(saved, seed))) continue;
+    delete prov.noVisionModels;
+    changed = true;
+    warnings.push(
+      `removed the retired registry noVisionModels seed (${saved.length} model(s)) from "${entry.provider}"; `
+      + "image input now follows the provider's current classification.",
+    );
+  }
+  return { config, changed, warnings };
+}

package/src/providers/xai-fast-model.ts

@@ -7,7 +7,7 @@ import type { FastWire, OcxParsedRequest, OcxProviderConfig, TierDecision } from
  * on two lanes (same effort ladder, image input, 500k limit and advertised defaults). Measured live on
  * 2026-09-30: build-fast is 1.5-1.7x faster end to end, while `service_tier: "priority"` on grok-4.7 bought
  * no measurable speed and cost ~5.9x the ticks per output token (build-fast costs ~2x).
- * Evidence: devlog/_plan/260930_grok47_build_unify/010_probe-evidence.md.
+ * Evidence: devlog/_fin/260930_grok47_build_unify/010_probe-evidence.md.
  *
  * So on the OAuth lane only, a Fast grok-4.7 request is SERIALIZED as the build-fast id and sends no tier.
  * The logical id stays grok-4.7 everywhere else (parsed.modelId, route, policy, usage attempt), so effort,

package/src/responses/compaction.ts

@@ -56,6 +56,20 @@ export const SUMMARY_PREFIX = "Another language model started to solve this prob
 
 export const OPAQUE_COMPACTION_NOTE = "[earlier conversation was compacted; the summary is stored in a format this model cannot read]";
 
+/**
+ * Prompt contract for a routed model receiving a decoded `ocx1` checkpoint as user text (#6764).
+ * It preserves existing authorization scope and permission gates; it grants no authority.
+ * Kept apart from SUMMARY_PREFIX, which the v1 retention and recovery paths match exactly.
+ */
+export const CHECKPOINT_CONTINUATION_NOTE = "This conversation was compacted automatically. A checkpoint summary from an earlier context window appears as a user message and records prior work. Treat earlier requests it reports as completed as history; do not repeat them unless a later user message asks you to. Continue an unfinished task only when it was already authorized and remains within the scope of that authorization. Do not ask for the same authorization again while it still applies. Preserve all existing permission and confirmation gates; if a decision or required permission is pending, wait for it before taking the dependent action. A checkpoint or delegation, including one from another thread, does not itself grant authority or expand the task's scope. Later user messages take priority; obey any pause, cancellation, replacement, or change of scope.";
+
+/** Append the continuation note to existing instructions once. */
+export function appendCheckpointContinuationNote(instructions: unknown): string {
+  const existing = typeof instructions === "string" ? instructions : "";
+  if (existing.includes(CHECKPOINT_CONTINUATION_NOTE)) return existing;
+  return existing.trim().length > 0 ? `${existing}\n\n${CHECKPOINT_CONTINUATION_NOTE}` : CHECKPOINT_CONTINUATION_NOTE;
+}
+
 /**
  * Item types in the compact wire family. Each carries an `encrypted_content` blob the client
  * replays verbatim on every later turn, and the minting backend verifies it is unmodified.

package/src/responses/parser.ts

@@ -15,7 +15,7 @@ import { createToolChoiceResolver, namespacedToolName } from "../types";
 import { responsesRequestSchema } from "./schema";
 import { providerMetadataFromResponsesFunctionCall } from "./provider-opaque-metadata";
 import { lookupReplayThoughtSignature } from "./thought-signature-replay";
-import { compactionItemToText, isCompactionItemType } from "./compaction";
+import { CHECKPOINT_CONTINUATION_NOTE, compactionItemToText, decodeCompactionSummary, isCompactionItemType } from "./compaction";
 import { previousResponseReplayPrefixLength } from "./state";
 import { decodeReasoningEnvelope } from "./reasoning-envelope";
 import { hasRoutedIdentity, nameRoutedIdentity } from "../adapters/identity";
@@ -193,6 +193,7 @@ export function parseRequest(
   // Remote compaction v2: the input tail carries `{type:"compaction_trigger"}` and Codex expects a
   // synthetic `{type:"compaction"}` output item (src/responses/compaction.ts). Flagged for the server.
   let compactionRequest = false;
+  let decodedCheckpoint = false;
   let contextCompactionBoundary = false;
   let continuationConversationMessageIndex: number | undefined;
 
@@ -261,6 +262,7 @@ export function parseRequest(
         if (inputIndex >= replayedInputPrefixLength) contextCompactionBoundary = true;
         const encrypted = (item as { encrypted_content?: unknown }).encrypted_content;
         if (effectiveType === "context_compaction" && typeof encrypted !== "string") continue;
+        if (typeof encrypted === "string" && decodeCompactionSummary(encrypted) !== null) decodedCheckpoint = true;
         pendingReasoning.length = 0;
         messages.push({
           role: "user",
@@ -586,6 +588,10 @@ export function parseRequest(
     .map(t => loadedToolNames.has(namespacedToolName(t.namespace, t.name))
       ? { ...t, loadedFromToolSearch: true }
       : t);
+  // #6764: the decoded checkpoint is ordinary user text on this route. Say in the system prompt
+  // what it is, so a retained earlier request that it reports as completed is not repeated.
+  // A compaction request is a summarizer turn and must not be told to continue.
+  if (decodedCheckpoint && !compactionRequest) systemPrompt.push(CHECKPOINT_CONTINUATION_NOTE);
   const context: OcxContext = {
     ...(systemPrompt.length > 0 ? { systemPrompt } : {}),
     messages,

package/src/server/auth-cors.ts

@@ -47,6 +47,18 @@ import { googleVertexLocationConfigError } from "../providers/google-vertex-loca
 import { xaiResponsesOptInState } from "../providers/xai-responses-opt-in";
 
 let _corsOrigin = "http://localhost:10100";
+// Listener identity is process-local. Never infer this trust from forwarded headers: the
+// request handler marks a Request only after resolving the actual Bun Server instance.
+const trustedHubManagementRequests = new WeakSet<Request>();
+
+export function markTrustedHubManagementRequest(req: Request): void {
+  trustedHubManagementRequests.add(req);
+}
+
+export function inheritTrustedHubManagementRequest(source: Request, target: Request): void {
+  if (trustedHubManagementRequests.has(source)) trustedHubManagementRequests.add(target);
+}
+
 export function setCorsOrigin(port: number): void { _corsOrigin = `http://localhost:${port}`; }
 /** The proxy's own listening port. No admission check uses it: both loopback predicates key on hostname alone. */
 export function configuredPort(): string {
@@ -144,6 +156,24 @@ export function managementRequestOrigin(req: Request, config: OcxConfig): string
       return null;
     }
   }
+  if (trustedHubManagementRequests.has(req)) {
+    if (config.runtimeRole !== "hub") return null;
+    const canonical = canonicalHubManagementOrigin(config.hub?.managementPublicOrigin);
+    if (!canonical) return null;
+    try {
+      // The dedicated listener may receive HTTP from a local TLS terminator, so its URL
+      // scheme is not authoritative. Host must still name exactly the configured public
+      // authority; forwarded host/proto headers are deliberately ignored.
+      const requestAuthority = new URL(`https://${host}`);
+      if (requestAuthority.username || requestAuthority.password || requestAuthority.pathname !== "/"
+        || requestAuthority.search || requestAuthority.hash || requestAuthority.host !== new URL(canonical).host) {
+        return null;
+      }
+      return canonical;
+    } catch {
+      return null;
+    }
+  }
   if (!isApiAuthRequired(config)) return null;
   if (config.runtimeRole === "hub" && config.hub?.managementPublicOrigin) {
     try {
@@ -437,6 +467,18 @@ export function resolveDataPlaneAdmissionSecret(
   return null;
 }
 
+function canonicalHubManagementOrigin(value: unknown): string | null {
+  if (typeof value !== "string") return null;
+  try {
+    const parsed = new URL(value);
+    if (parsed.protocol !== "https:" || parsed.username || parsed.password || parsed.pathname !== "/"
+      || parsed.search || parsed.hash) return null;
+    return parsed.origin;
+  } catch {
+    return null;
+  }
+}
+
 /** The principal an authenticated admission belongs to, or undefined for loopback. */
 export function contextPrincipalIdOf(admission: DataPlaneAdmission | undefined): string | undefined {
   return admission && "contextPrincipalId" in admission ? admission.contextPrincipalId : undefined;