@bitkyc08/opencodex 2.52.0 → 2.53.0-preview.20260913

package/src/types/accounts.ts

@@ -29,6 +29,8 @@ export interface CodexAccountCredentialRecord {
   credential?: CodexAccountCredentials;
   generation: number;
   refreshGrantFingerprint?: string;
+  /** Private non-secret publication identity, stable across same-account token refresh. */
+  quotaHistoryIdentity?: string;
   deletedAt?: number;
   replacedAt?: number;
   lastCodexValidatedAt?: number;

package/src/types/config.ts

@@ -380,6 +380,8 @@ export interface OcxConfig {
   privacy?: OcxPrivacyConfig;
   /** Opt in to one identical-turn retry when a Responses completion has no text or tool call. */
   emptyCompletionRetry?: boolean;
+  /** Suppress allowlisted client-facing Codex transport hints; provider enforcement is unchanged. */
+  dropCodexSafetyBuffering?: boolean;
   /**
    * Whether a login may open a browser on the machine running the proxy.
    *
@@ -637,11 +639,18 @@ export interface OcxConfig {
    * - "v2": force ALL models to v2 surface (override upstream pins)
    */
   multiAgentMode?: "v1" | "default" | "v2";
+  /**
+   * Which revision of the sub-agent surface advisory this install has answered.
+   * Absent means it has answered none. Written by the dashboard, never by a mode change.
+   */
+  multiAgentSurfaceAdvisoryVersion?: number;
   /**
    * When `multiAgentMode` is `"v2"`, keep ChatGPT-native catalog rows on v1.
    * Routed parents get v2 tools; Sol/Terra can still spawn Grok/Claude (issue #92).
    */
   keepNativeChatGptOnV1?: boolean;
+  /** Experimental plaintext delivery for native v2 collaboration messages; disabled unless true. */
+  plaintextV2AgentMessages?: boolean;
   /** Experimental, default-off ChatGPT recovery for encrypted V2 routed tasks. */
   agentTaskRecovery?: {
     enabled?: boolean;
@@ -822,15 +831,6 @@ export interface OcxConfig {
    * selector map remains visible for compatibility with hand-written configurations.
    */
   codexAccountPickerEnabled?: boolean;
-  /**
-   * Show the GPT-5.3-Codex-Spark 5-hour and weekly windows on Codex quota surfaces. Default false.
-   *
-   * Spark is a single-model window that reads 0% for most operators, and on a multi-account
-   * pool it doubles the bar count for information almost nobody acts on. Hidden by default and
-   * revealed by an explicit `true`; a malformed value reads as hidden rather than rejecting the
-   * whole config.
-   */
-  showCodexSparkQuota?: boolean;
   /**
    * Opt-in auto-redemption of a main-account Codex reset credit shortly before it expires
    * (#822). Default off. `leadTimeMinutes` (1–60, default 10) is how long before
@@ -866,7 +866,7 @@ export interface OcxConfig {
   /** Auto-switch threshold (0-100). Default 80. 0 = disabled. */
   autoSwitchThreshold?: number;
   /** New-session account rotation strategy for the Codex pool. Default quota (today's behaviour). */
-  accountPoolStrategy?: OcxAccountPoolRotationStrategy;
+  accountPoolStrategy?: OcxAccountPoolRotationStrategy | "reset-first";
   /** Successful new-session binds retained on one round-robin selection. Default 1; range 1..100. */
   accountPoolStickyLimit?: number;
   /** Consecutive non-2xx upstream responses before switching future new threads. Default 3. 0 = disabled. */
@@ -967,8 +967,9 @@ export type OcxComboDefaultEffort = "low" | "medium" | "high" | "xhigh" | "max"
  * advertises no effort control (`reasoningEfforts: []`) empties the combo's picker.
  * `adaptive` excludes those empty ladders from the published intersection, keeping the
  * control usable for a mixed-capability group. Unknown (`undefined`) ladders stay
- * wildcards in both modes. Dispatch is unchanged: each concrete target still resolves
- * its own effort at request time.
+ * wildcards in both modes. An explicit empty ladder removes unsupported effort controls
+ * in either mode; adaptive dispatch also removes them before sending to an unknown target,
+ * while each known target still resolves its own effort.
  */
 export type OcxComboReasoningEffortMode = "strict" | "adaptive";
 

package/src/types/provider.ts

@@ -227,6 +227,14 @@ export type TierDecision =
  * One configured provider entry. `authMode` (default `"key"`) decides whether same-target 429
  * retries are allowed; OAuth/forward credentials and local runtimes are never replayed.
  */
+/** Explicit per-model operator declarations; absent axes keep legacy behavior. */
+export interface ModelCapabilities {
+  inputModalities?: Array<"text" | "image" | "audio" | "video">;
+  /** Requested tier only; does not imply an upstream window or activate an unverified wire. */
+  contextTier?: "default" | "long_context";
+  video?: { processing?: "static" | "agentic" };
+}
+
 export interface OcxProviderConfig {
   /** Optional short provider namespace used only at request/catalog presentation time. */
   alias?: string;
@@ -478,6 +486,7 @@ export interface OcxProviderConfig {
   modelContextWindows?: Record<string, number>;
   /** Model-specific Codex catalog input modalities, e.g. ["text"] or ["text", "image"]. */
   modelInputModalities?: Record<string, string[]>;
+  modelCapabilities?: Record<string, ModelCapabilities>;
   /** Model-specific max input token limits. Values cap auto_compact_token_limit. */
   modelMaxInputTokens?: Record<string, number>;
   /**
@@ -501,6 +510,28 @@ export interface OcxProviderConfig {
    * all-zero entry means "not billable here" and falls through to the catalogs.
    */
   modelCosts?: Record<string, ProviderCostOverlay>;
+  /**
+   * Provider-wide auto-review (approval) model for routed models of this provider.
+   *
+   * The value is a catalog selector: either a bare model id of this provider
+   * (for example `deepseek-v4-flash`) or a full public slug (for example
+   * `opencode-go/deepseek-v4-flash`). During catalog synchronization the
+   * selector is resolved against the final catalog and stamped as
+   * `auto_review_model_override` on each routed row of this provider that has
+   * no per-model override. The root Codex `auto_review_model` remains the
+   * fallback for every row without a provider stamp. Null or blank clears the
+   * provider-wide stamp; see `autoReviewModelOverrides` for per-model targets.
+   */
+  autoReviewModel?: string;
+  /**
+   * Per-model auto-review (approval) overrides for routed models of this
+   * provider. Keys are exact upstream model ids under this provider (either
+   * spelling of a slash-containing id is accepted). Each value is a catalog
+   * selector with the same meaning as `autoReviewModel`; an entry wins over
+   * the provider-wide value for its model. Null or blank entries remove the
+   * model from the map while preserving other entries.
+   */
+  autoReviewModelOverrides?: Record<string, string>;
   headers?: Record<string, string>;
   /** Default provider-routing preferences for models sent through the canonical OpenRouter API. */
   openRouterRouting?: OpenRouterProviderRouting;
@@ -773,6 +804,12 @@ export interface OcxProviderConfig {
    * out explicitly (e.g. MiniMax, where low effort disables thinking).
    */
   requiresReasoningPlaceholderModels?: string[];
+  /**
+   * Default to displaying provider-authored summaries when Responses summary is omitted.
+   * Explicit wire summary:"none" wins; false disables a seeded provider default.
+   * Raw reasoning is never relabeled as a summary.
+   */
+  showThinkingSummary?: boolean;
   /**
    * Opt-in same-target 429 retry policy. Codex itself never retries 429 (it retries 5xx only,
    * openai/codex#30471), and single-key pools have no failover, so the proxy waits and replays

package/src/types/request.ts

@@ -79,6 +79,8 @@ export interface OcxParsedRequest {
    * prepareOpaqueBlobRecovery after an authoritative rejection; consumers strip replayed blobs.
    */
   _stripReasoningEncryptedContent?: boolean;
+  /** Final-route opt-in: emit v2 collaboration message arguments as plaintext on ChatGPT. */
+  _plaintextV2AgentMessages?: boolean;
   /**
    * Optional authenticated tenant/operator namespace for Cursor thread→conversation derivation.
    * When absent (single-operator local proxy), derivation stays local-scoped.

package/src/types/tools.ts

@@ -47,16 +47,17 @@ export function dottedToolName(namespace: string | undefined, name: string): str
  *
  * Codex's code-mode shell tool is declared as `exec` (a freeform custom tool whose own
  * description mentions the nested `await tools.exec_command(...)` helper). Some routed providers
- * echo that helper name as the tool-call name, emitting `exec_command`, `write_stdin`, or
- * `apply_patch` instead of the declared `exec`. Accept these nested helper names only when the
- * request catalog actually declares `exec` and does not itself declare the emitted name (an MCP
- * server may legitimately advertise one under its own namespace).
+ * echo that helper name as the tool-call name, emitting `exec_command`, `write_stdin`,
+ * `apply_patch`, or `view_image` instead of the declared `exec`. Accept these nested helper names
+ * only when the request catalog actually declares `exec` and does not itself declare the emitted
+ * name (an MCP server may legitimately advertise one under its own namespace).
  */
 const LEGACY_SHELL_BRIDGE_TOOL_NAMES = ["exec_command", "shell_command"] as const;
 const CODE_MODE_HELPER_TOOL_NAMES = [
   ...LEGACY_SHELL_BRIDGE_TOOL_NAMES,
   "write_stdin",
   "apply_patch",
+  "view_image",
 ] as const;
 
 /**
@@ -71,7 +72,7 @@ export const CODE_MODE_EXEC_TOOL_NAME = "exec";
  *
  * Rewrites invented `default.<name>` prefixes back to a declared bare tool when that bare tool
  * is declared and neither `default.<name>` nor `default__<name>` was explicitly declared (#4176).
- * Also normalizes legacy helper names (`exec_command`, `shell_command`, `apply_patch`) to
+ * Also normalizes legacy helper names (`exec_command`, `shell_command`, `apply_patch`, `view_image`) to
  * `exec` when code-mode `exec` is declared in the request catalog.
  *
  * @param name - The tool name emitted on the wire by the provider.
@@ -98,6 +99,17 @@ export function normalizeDeclaredToolName(
       && !declared.has("default__" + bare)
     ) {
       candidate = bare;
+    } else if (
+      // Code mode never declares bare helper names; a provider that invents `default.`
+      // for one still means the nested helper. Strip the prefix so the helper list
+      // below can rewrite it to `exec` (#4412).
+      bare.length > 0
+      && declared.has(CODE_MODE_EXEC_TOOL_NAME)
+      && (CODE_MODE_HELPER_TOOL_NAMES as readonly string[]).includes(bare)
+      && !declared.has("default." + bare)
+      && !declared.has("default__" + bare)
+    ) {
+      candidate = bare;
     }
   }
   if (!declared.has(CODE_MODE_EXEC_TOOL_NAME)) return candidate;

package/src/types.ts

@@ -111,6 +111,7 @@ export type {
   TierObservationContext,
   TierDecision,
   OcxProviderConfig,
+  ModelCapabilities,
 } from "./types/provider";
 
 export { PROVIDER_WEB_SEARCH_BRIDGE_BACKENDS } from "./types/provider";

package/src/usage/expected-prices.ts

@@ -70,6 +70,29 @@ const KIMI_K27_CODE: Cost4 = { input: 0.95, output: 4, cacheRead: 0.19, cacheWri
 const KIMI_K27_CODE_HIGHSPEED: Cost4 = { input: 1.9, output: 8, cacheRead: 0.38, cacheWrite: 1.9 };
 const KIMI_K26: Cost4 = { input: 0.95, output: 4, cacheRead: 0.16, cacheWrite: 0.95 };
 const KIMI_K25: Cost4 = { input: 0.6, output: 3, cacheRead: 0.1, cacheWrite: 0.6 };
+/*
+ * Z.AI GLM list prices (USD / 1M tokens), verified 2026-09-13 against
+ * https://docs.z.ai/guides/overview/pricing. Neither z.ai nor bigmodel.cn
+ * publishes a cache-write rate — both list cache storage as limited-time free,
+ * an open-beta promotion the vendor may change or end — so cacheWrite is 0 as a
+ * 2026-09-13 snapshot, not a guaranteed rate; re-check the pricing page before
+ * relying on it long-term. glm-4.5-flash and glm-4.7-flash are officially
+ * "Free" and deliberately get no rows: a zero-cost overlay is inert in the
+ * resolver, which requires a nonzero tuple. glm-5-turbo / glm-5v-turbo are
+ * published only in CNY on bigmodel.cn and stay unregistered — the same hold
+ * the xiaomi CNY rows took in devlog/_fin/260720_toks_speed_price_columns/003.
+ * glm-4.5 (0.6/2.2/0.11), glm-4.5-air (0.2/1.1/0.03) and glm-4.5v (0.6/1.8/0.11)
+ * are verified on the same page but no registered provider exposes them, so
+ * they have no constants here.
+ */
+const GLM_46: Cost4 = { input: 0.6, output: 2.2, cacheRead: 0.11, cacheWrite: 0 };
+const GLM_46V: Cost4 = { input: 0.3, output: 0.9, cacheRead: 0.05, cacheWrite: 0 };
+const GLM_47: Cost4 = { input: 0.6, output: 2.2, cacheRead: 0.11, cacheWrite: 0 };
+const GLM_5: Cost4 = { input: 1, output: 3.2, cacheRead: 0.2, cacheWrite: 0 };
+const GLM_51: Cost4 = { input: 1.4, output: 4.4, cacheRead: 0.26, cacheWrite: 0 };
+const GLM_52: Cost4 = { input: 1.4, output: 4.4, cacheRead: 0.26, cacheWrite: 0 };
+const GLM_53: Cost4 = { input: 1.4, output: 4.4, cacheRead: 0.26, cacheWrite: 0 };
+const GLM_53_FLASH: Cost4 = { input: 0.15, output: 0.5, cacheRead: 0.03, cacheWrite: 0 };
 const QWEN38_MAX: Cost4 = { input: 2, output: 6, cacheRead: 0, cacheWrite: 0 };
 // Anthropic official list prices (USD / 1M tokens). Cache write uses the published 5-minute rate.
 const CLAUDE_SONNET_46: Cost4 = { input: 3, output: 15, cacheRead: 0.3, cacheWrite: 3.75 };
@@ -104,6 +127,13 @@ const DEEPSEEK_PRICING = "https://api-docs.deepseek.com/quick_start/pricing-deta
 // Kimi official tables publish input/output/cache-hit only; cacheWrite is mapped to the
 // cache-miss input price (Kimi auto-caches with no separate write billing). 2026-07-20 re-verified.
 const KIMI_PRICING = "https://platform.kimi.ai/docs/pricing (official table; cacheWrite derived = input, Kimi auto-cache has no write billing)";
+// Z.AI publishes one USD table for the international surface; the Coding Plan
+// subscription and the domestic bigmodel.cn endpoints bill differently
+// (subscription quota / CNY tiers), so every GLM row below is verified-derived:
+// the numbers are the verified z.ai list prices shown as estimates.
+const ZAI_PRICING = "https://docs.z.ai/guides/overview/pricing (official USD table, 2026-09-13; cacheWrite=0 — cache storage is limited-time free on both z.ai and bigmodel.cn)";
+const ZAI_CODING_PLAN_NOTE = "z.ai list price shown as estimate; GLM Coding Plan is subscription-billed";
+const BIGMODEL_NOTE = "z.ai international list price shown as estimate; domestic bigmodel.cn billing is CNY tiered (docs.bigmodel.cn/cn/guide/start/pricing)";
 // 260804: Qwen3.8-Max shipped as a stable model and Qwen published a per-token rate, which
 // is the exit condition the previous Routeway reseller overlay named. Two caveats are
 // deliberately in the source string rather than dropped: the figure comes from Qwen's own
@@ -113,6 +143,31 @@ const KIMI_PRICING = "https://platform.kimi.ai/docs/pricing (official table; cac
 // under a vendor-price label would be a wrong value wearing a verified badge.
 const QWEN38_MAX_PRICING = "https://qwen.ai/blog?id=qwen3.8 (Qwen release announcement; no Model Studio billing row yet; cache rates unpublished -> 0)";
 
+/*
+ * Cognition/Devin list prices (USD / 1M tokens), verified 2026-09-13 against the
+ * official "AI Models" page — its embedded modelCostData table publishes
+ * input / cache-read / cache-write / output per model uid. Self-serve extra
+ * usage and enterprise ACU conversion both bill at these list rates, so the
+ * tuples are the vendor's own published numbers; every row still stays
+ * verified-derived because the surface itself is subscription/ACU, not a
+ * per-token API.
+ * Time-boxed promos are NOT baked in: SWE-2 shows $0 self-serve through
+ * 2026-10-08 and 75%-off enterprise through 2026-12-31, and the doc states the
+ * list rate is what applies afterward, so the list rate is the durable catalog
+ * value. swe-1-7 keeps its list rate for the same reason even though the
+ * self-serve column currently shows 0. gemini-3-8-flash is absent from the
+ * table entirely, so its row derives from Google's published rate instead.
+ */
+const DEVIN_SWE_2: Cost4 = { input: 3, output: 15, cacheRead: 0.3, cacheWrite: 0 };
+const DEVIN_SWE_17: Cost4 = { input: 0.5, output: 2.5, cacheRead: 0.2, cacheWrite: 0 };
+const DEVIN_SWE_17_LIGHTNING: Cost4 = { input: 2.5, output: 12.5, cacheRead: 1, cacheWrite: 0 };
+const DEVIN_SONNET_5: Cost4 = { input: 2, output: 10, cacheRead: 0.2, cacheWrite: 2.5 };
+const DEVIN_KIMI_K3: Cost4 = { input: 3, output: 15, cacheRead: 0.3, cacheWrite: 0 };
+const DEVIN_KIMI_K27: Cost4 = { input: 0.95, output: 4, cacheRead: 0.19, cacheWrite: 0 };
+const DEVIN_GROK: Cost4 = { input: 2, output: 6, cacheRead: 0.3, cacheWrite: 0 };
+const DEVIN_PRICING = "https://docs.devin.ai/desktop/models (official modelCostData table, 2026-09-13; list rates for self-serve overage / enterprise ACU conversion on a subscription surface)";
+const DEVIN_SWE2_NOTE = "list rate; $0 self-serve through 2026-10-08 and 75%-off enterprise through 2026-12-31 are time-boxed promos, not baked in";
+
 export const EXPECTED_PRICE_OVERLAYS: readonly ExpectedPriceOverlay[] = [
   { provider: "openai-apikey", modelId: "gpt-6-astra", cost4: GPT6_ASTRA, source: ASTRA_API_PRICING, verifiedAt: "2026-09-05", status: "verified" },
   // Display estimates use API prices for both login and API-key routes, including cache writes.
@@ -236,6 +291,78 @@ export const EXPECTED_PRICE_OVERLAYS: readonly ExpectedPriceOverlay[] = [
   { provider: "alibaba-token-plan-intl", modelId: "qwen3.8-max", cost4: QWEN38_MAX, source: QWEN38_MAX_PRICING, verifiedAt: "2026-08-04", status: "verified" },
   // Cursor Auto router — Cursor's published fixed token price (verified).
   { provider: "cursor", modelId: "auto", cost4: { input: 1.25, output: 6, cacheRead: 0.25, cacheWrite: 1.25 }, source: "https://docs.cursor.com/account/pricing + https://cursor.com/blog/aug-2025-pricing", verifiedAt: "2026-07-20", status: "verified" },
+  // Z.AI GLM family — the zai bundle's rows are all-zero upstream, and the four
+  // provider surfaces below resolve overlays by exact provider id, so each one
+  // needs its own rows (same pattern as kimi/moonshot/kimi-code). All rows are
+  // verified-derived: the tuples are the verified z.ai USD list prices, while
+  // the Coding Plan rows are subscription products and zhipu-bigmodel is the
+  // domestic CNY-tiered PAYG — see ZAI_CODING_PLAN_NOTE / BIGMODEL_NOTE.
+  // zai (api.z.ai Coding Plan) exposes: glm-5.3, glm-5.3[1m], glm-5.3-flash,
+  // glm-5.2, glm-5.2[1m], glm-5.1, glm-5, glm-4.6.
+  { provider: "zai", modelId: "glm-5.3", cost4: GLM_53, source: `${ZAI_CODING_PLAN_NOTE}; ${ZAI_PRICING}`, verifiedAt: "2026-09-13", status: "verified-derived" },
+  { provider: "zai", modelId: "glm-5.3[1m]", cost4: GLM_53, source: `derived: glm-5.3 1M-context compat notation; ${ZAI_CODING_PLAN_NOTE}; ${ZAI_PRICING}`, verifiedAt: "2026-09-13", status: "verified-derived" },
+  { provider: "zai", modelId: "glm-5.3-flash", cost4: GLM_53_FLASH, source: `${ZAI_CODING_PLAN_NOTE}; ${ZAI_PRICING}`, verifiedAt: "2026-09-13", status: "verified-derived" },
+  { provider: "zai", modelId: "glm-5.2", cost4: GLM_52, source: `${ZAI_CODING_PLAN_NOTE}; ${ZAI_PRICING}`, verifiedAt: "2026-09-13", status: "verified-derived" },
+  { provider: "zai", modelId: "glm-5.2[1m]", cost4: GLM_52, source: `derived: glm-5.2 1M-context compat notation; ${ZAI_CODING_PLAN_NOTE}; ${ZAI_PRICING}`, verifiedAt: "2026-09-13", status: "verified-derived" },
+  { provider: "zai", modelId: "glm-5.1", cost4: GLM_51, source: `${ZAI_CODING_PLAN_NOTE}; ${ZAI_PRICING}`, verifiedAt: "2026-09-13", status: "verified-derived" },
+  { provider: "zai", modelId: "glm-5", cost4: GLM_5, source: `${ZAI_CODING_PLAN_NOTE}; ${ZAI_PRICING}`, verifiedAt: "2026-09-13", status: "verified-derived" },
+  { provider: "zai", modelId: "glm-4.6", cost4: GLM_46, source: `${ZAI_CODING_PLAN_NOTE}; ${ZAI_PRICING}`, verifiedAt: "2026-09-13", status: "verified-derived" },
+  // zhipu-bigmodel (open.bigmodel.cn PAYG) exposes: glm-4.6, glm-4.7,
+  // glm-4.7-flash (officially free — no row), glm-5, glm-5.1, glm-5.2, glm-5.3,
+  // glm-4.6v.
+  { provider: "zhipu-bigmodel", modelId: "glm-4.6", cost4: GLM_46, source: `${BIGMODEL_NOTE}; ${ZAI_PRICING}`, verifiedAt: "2026-09-13", status: "verified-derived" },
+  { provider: "zhipu-bigmodel", modelId: "glm-4.6v", cost4: GLM_46V, source: `${BIGMODEL_NOTE}; ${ZAI_PRICING}`, verifiedAt: "2026-09-13", status: "verified-derived" },
+  { provider: "zhipu-bigmodel", modelId: "glm-4.7", cost4: GLM_47, source: `${BIGMODEL_NOTE}; ${ZAI_PRICING}`, verifiedAt: "2026-09-13", status: "verified-derived" },
+  { provider: "zhipu-bigmodel", modelId: "glm-5", cost4: GLM_5, source: `${BIGMODEL_NOTE}; ${ZAI_PRICING}`, verifiedAt: "2026-09-13", status: "verified-derived" },
+  { provider: "zhipu-bigmodel", modelId: "glm-5.1", cost4: GLM_51, source: `${BIGMODEL_NOTE}; ${ZAI_PRICING}`, verifiedAt: "2026-09-13", status: "verified-derived" },
+  { provider: "zhipu-bigmodel", modelId: "glm-5.2", cost4: GLM_52, source: `${BIGMODEL_NOTE}; ${ZAI_PRICING}`, verifiedAt: "2026-09-13", status: "verified-derived" },
+  { provider: "zhipu-bigmodel", modelId: "glm-5.3", cost4: GLM_53, source: `${BIGMODEL_NOTE}; ${ZAI_PRICING}`, verifiedAt: "2026-09-13", status: "verified-derived" },
+  // zhipu-bigmodel-coding exposes the same roster as the zai Coding Plan row.
+  { provider: "zhipu-bigmodel-coding", modelId: "glm-5.3", cost4: GLM_53, source: `${ZAI_CODING_PLAN_NOTE}; ${ZAI_PRICING}`, verifiedAt: "2026-09-13", status: "verified-derived" },
+  { provider: "zhipu-bigmodel-coding", modelId: "glm-5.3[1m]", cost4: GLM_53, source: `derived: glm-5.3 1M-context compat notation; ${ZAI_CODING_PLAN_NOTE}; ${ZAI_PRICING}`, verifiedAt: "2026-09-13", status: "verified-derived" },
+  { provider: "zhipu-bigmodel-coding", modelId: "glm-5.3-flash", cost4: GLM_53_FLASH, source: `${ZAI_CODING_PLAN_NOTE}; ${ZAI_PRICING}`, verifiedAt: "2026-09-13", status: "verified-derived" },
+  { provider: "zhipu-bigmodel-coding", modelId: "glm-5.2", cost4: GLM_52, source: `${ZAI_CODING_PLAN_NOTE}; ${ZAI_PRICING}`, verifiedAt: "2026-09-13", status: "verified-derived" },
+  { provider: "zhipu-bigmodel-coding", modelId: "glm-5.2[1m]", cost4: GLM_52, source: `derived: glm-5.2 1M-context compat notation; ${ZAI_CODING_PLAN_NOTE}; ${ZAI_PRICING}`, verifiedAt: "2026-09-13", status: "verified-derived" },
+  { provider: "zhipu-bigmodel-coding", modelId: "glm-5.1", cost4: GLM_51, source: `${ZAI_CODING_PLAN_NOTE}; ${ZAI_PRICING}`, verifiedAt: "2026-09-13", status: "verified-derived" },
+  { provider: "zhipu-bigmodel-coding", modelId: "glm-5", cost4: GLM_5, source: `${ZAI_CODING_PLAN_NOTE}; ${ZAI_PRICING}`, verifiedAt: "2026-09-13", status: "verified-derived" },
+  { provider: "zhipu-bigmodel-coding", modelId: "glm-4.6", cost4: GLM_46, source: `${ZAI_CODING_PLAN_NOTE}; ${ZAI_PRICING}`, verifiedAt: "2026-09-13", status: "verified-derived" },
+  // zhipu-bigmodel-responses exposes glm-5.3, glm-5.3-flash, glm-5-turbo; the
+  // turbo id is CNY-only upstream and stays unregistered (see the GLM_* note).
+  { provider: "zhipu-bigmodel-responses", modelId: "glm-5.3", cost4: GLM_53, source: `${ZAI_CODING_PLAN_NOTE}; ${ZAI_PRICING}`, verifiedAt: "2026-09-13", status: "verified-derived" },
+  { provider: "zhipu-bigmodel-responses", modelId: "glm-5.3-flash", cost4: GLM_53_FLASH, source: `${ZAI_CODING_PLAN_NOTE}; ${ZAI_PRICING}`, verifiedAt: "2026-09-13", status: "verified-derived" },
+  // Cognition/Devin — the two OAuth surfaces resolve by exact provider id, so
+  // each carries the roster its liveModels discovery can surface. swe-2 and
+  // swe-1-6 are listed on both even though each static seed names only one
+  // side: the live catalog is authoritative and drifts between them.
+  // gpt-5-6-sol uses the enterprise list column — the same table's self-serve
+  // column shows a discounted 1.2/6, and the doc calls the list rate the
+  // billing rate for overage. glm-5-2 likewise takes the nonzero list column.
+  { provider: "devin-cli", modelId: "swe-2", cost4: DEVIN_SWE_2, source: `${DEVIN_SWE2_NOTE}; ${DEVIN_PRICING}`, verifiedAt: "2026-09-13", status: "verified-derived" },
+  { provider: "devin-cli", modelId: "swe-1-7", cost4: DEVIN_SWE_17, source: `list rate; self-serve column currently shows 0 (unannounced promo); ${DEVIN_PRICING}`, verifiedAt: "2026-09-13", status: "verified-derived" },
+  { provider: "devin-cli", modelId: "swe-1-7-lightning", cost4: DEVIN_SWE_17_LIGHTNING, source: DEVIN_PRICING, verifiedAt: "2026-09-13", status: "verified-derived" },
+  { provider: "devin-cli", modelId: "swe-1-6", cost4: DEVIN_SWE_17, source: DEVIN_PRICING, verifiedAt: "2026-09-13", status: "verified-derived" },
+  { provider: "devin-cli", modelId: "gpt-5-6-sol", cost4: GPT56_SOL, source: `enterprise list column (self-serve shows discounted 1.2/6); ${DEVIN_PRICING}`, verifiedAt: "2026-09-13", status: "verified-derived" },
+  { provider: "devin-cli", modelId: "gpt-6-astra", cost4: GPT6_ASTRA, source: DEVIN_PRICING, verifiedAt: "2026-09-13", status: "verified-derived" },
+  { provider: "devin-cli", modelId: "claude-opus-5", cost4: CLAUDE_OPUS_46, source: DEVIN_PRICING, verifiedAt: "2026-09-13", status: "verified-derived" },
+  { provider: "devin-cli", modelId: "claude-fable-5-1", cost4: CLAUDE_FABLE_51, source: DEVIN_PRICING, verifiedAt: "2026-09-13", status: "verified-derived" },
+  { provider: "devin-cli", modelId: "claude-sonnet-5", cost4: DEVIN_SONNET_5, source: DEVIN_PRICING, verifiedAt: "2026-09-13", status: "verified-derived" },
+  { provider: "devin-cli", modelId: "glm-5-3", cost4: GLM_53, source: DEVIN_PRICING, verifiedAt: "2026-09-13", status: "verified-derived" },
+  { provider: "devin-cli", modelId: "kimi-k3", cost4: DEVIN_KIMI_K3, source: DEVIN_PRICING, verifiedAt: "2026-09-13", status: "verified-derived" },
+  { provider: "devin-cli", modelId: "gemini-3-8-flash", cost4: GEMINI_38_FLASH, source: `derived: absent from Devin's modelCostData table; Google published promotional rate through 2026-12-31 shown as estimate ${GEMINI_38_PRICING}`, verifiedAt: "2026-09-13", status: "verified-derived" },
+  { provider: "devin-cli", modelId: "grok-4-6", cost4: DEVIN_GROK, source: DEVIN_PRICING, verifiedAt: "2026-09-13", status: "verified-derived" },
+  { provider: "devin", modelId: "swe-2", cost4: DEVIN_SWE_2, source: `${DEVIN_SWE2_NOTE}; ${DEVIN_PRICING}`, verifiedAt: "2026-09-13", status: "verified-derived" },
+  { provider: "devin", modelId: "swe-1-7", cost4: DEVIN_SWE_17, source: `list rate; self-serve column currently shows 0 (unannounced promo); ${DEVIN_PRICING}`, verifiedAt: "2026-09-13", status: "verified-derived" },
+  { provider: "devin", modelId: "swe-1-7-lightning", cost4: DEVIN_SWE_17_LIGHTNING, source: DEVIN_PRICING, verifiedAt: "2026-09-13", status: "verified-derived" },
+  { provider: "devin", modelId: "swe-1-6", cost4: DEVIN_SWE_17, source: DEVIN_PRICING, verifiedAt: "2026-09-13", status: "verified-derived" },
+  { provider: "devin", modelId: "gpt-5-6-sol", cost4: GPT56_SOL, source: `enterprise list column (self-serve shows discounted 1.2/6); ${DEVIN_PRICING}`, verifiedAt: "2026-09-13", status: "verified-derived" },
+  { provider: "devin", modelId: "gpt-5-6-luna", cost4: GPT56_LUNA, source: DEVIN_PRICING, verifiedAt: "2026-09-13", status: "verified-derived" },
+  { provider: "devin", modelId: "gpt-5-6-terra", cost4: GPT56_TERRA, source: DEVIN_PRICING, verifiedAt: "2026-09-13", status: "verified-derived" },
+  { provider: "devin", modelId: "claude-opus-4-8", cost4: CLAUDE_OPUS_46, source: DEVIN_PRICING, verifiedAt: "2026-09-13", status: "verified-derived" },
+  { provider: "devin", modelId: "claude-fable-5-1", cost4: CLAUDE_FABLE_51, source: DEVIN_PRICING, verifiedAt: "2026-09-13", status: "verified-derived" },
+  { provider: "devin", modelId: "claude-sonnet-5", cost4: DEVIN_SONNET_5, source: DEVIN_PRICING, verifiedAt: "2026-09-13", status: "verified-derived" },
+  { provider: "devin", modelId: "glm-5-2", cost4: GLM_52, source: `enterprise list column (self-serve shows an unannounced 0 promo); ${DEVIN_PRICING}`, verifiedAt: "2026-09-13", status: "verified-derived" },
+  { provider: "devin", modelId: "kimi-k2-7", cost4: DEVIN_KIMI_K27, source: DEVIN_PRICING, verifiedAt: "2026-09-13", status: "verified-derived" },
+  { provider: "devin", modelId: "grok-4-5", cost4: DEVIN_GROK, source: DEVIN_PRICING, verifiedAt: "2026-09-13", status: "verified-derived" },
 ];
 
 /**

package/src/usage/log.ts

@@ -10,6 +10,7 @@ import type { AttemptTierOutcome, OcxUsage } from "../types";
 import { normalizeRouteDecisionTrace, type RouteDecisionTraceV1 } from "../routing/trace";
 import { ACCOUNT_LOG_LABEL_RE, CODEX_ACCOUNT_LOG_LABEL_RE } from "../codex/account-label";
 import { claudeCompatibilityReason, normalizeClaudeFeatureCodes, type ClaudeFeatureCode } from "../claude/compatibility";
+import type { CodexWsStageRecord } from "../server/responses/codex-ws-wire";
 
 export interface PersistedClaudeCompatibilityLog {
   decision: "shadow";
@@ -70,8 +71,10 @@ export type AttemptRecoveryKind =
   | "anthropic-oauth-429"
   | "oauth-account-429"
   | "image-413"
+  | "console-go-upload-retry"
   | "opaque-blob-rejection"
-  | "empty-completion";
+  | "empty-completion"
+  | "reasoning-effort-downgrade";
 
 /** Request-time upstream credential class, never a credential or account identifier. */
 export type UsageCredentialSource = "grok-oauth" | "xai-api-key";
@@ -118,6 +121,14 @@ export interface PersistedUsageAttempt {
   reasoningWireValue?: string | number | boolean;
   /** Adapter-produced tier fact for this physical attempt; absent on pre-B0 rows. */
   tierOutcome?: AttemptTierOutcome;
+  /**
+   * #4191: content-free stage record of a Codex WS upstream exchange that
+   * served this attempt (frame size, counters, close code, versions). Absent
+   * on HTTP-transport attempts and pre-instrumentation rows. Numbers,
+   * booleans, and semver strings only — never reason text, headers, or
+   * account identifiers.
+   */
+  codexWsStage?: CodexWsStageRecord;
 }
 
 export interface PersistedUsageEntry {
@@ -309,8 +320,10 @@ const ATTEMPT_RECOVERY_KINDS = new Set<AttemptRecoveryKind>([
   "anthropic-oauth-429",
   "oauth-account-429",
   "image-413",
+  "console-go-upload-retry",
   "opaque-blob-rejection",
   "empty-completion",
+  "reasoning-effort-downgrade",
 ]);
 const USAGE_STATUSES = new Set<UsageStatus>([
   "reported",
@@ -436,6 +449,9 @@ function normalizeUsageAttempt(raw: unknown): PersistedUsageAttempt | null {
   const tierOutcome = "tierOutcome" in attempt
     ? normalizeAttemptTierOutcome(attempt.tierOutcome)
     : undefined;
+  const codexWsStage = "codexWsStage" in attempt
+    ? normalizeCodexWsStageRecord(attempt.codexWsStage)
+    : undefined;
   const recoveryKinds = Array.isArray(attempt.recoveryKinds)
     ? [...new Set(attempt.recoveryKinds.filter(
       (value): value is AttemptRecoveryKind => typeof value === "string"
@@ -455,6 +471,7 @@ function normalizeUsageAttempt(raw: unknown): PersistedUsageAttempt | null {
     durationMs: attempt.durationMs,
     // Absent by default; only the literal `true` marker survives the round trip.
     ...(attempt.streamAborted === true ? { streamAborted: true } : {}),
+    ...(attempt.locallyAnswered === true ? { locallyAnswered: true } : {}),
     ...(isNonNegativeFiniteNumber(attempt.firstOutputMs)
       ? { firstOutputMs: attempt.firstOutputMs }
       : {}),
@@ -490,6 +507,46 @@ function normalizeUsageAttempt(raw: unknown): PersistedUsageAttempt | null {
         : { reasoningWireValue: attempt.reasoningWireValue }
       : {}),
     ...(tierOutcome ? { tierOutcome } : {}),
+    ...(codexWsStage ? { codexWsStage } : {}),
+  };
+}
+
+/**
+ * #4191: a persisted stage record is trusted only when every field matches the
+ * exchange's own shapes. Anything else — a hand-edited number as a string, an
+ * injected free-form field — drops the whole record rather than passing
+ * attacker text into the DTO.
+ */
+function normalizeCodexWsStageRecord(value: unknown): CodexWsStageRecord | undefined {
+  if (value === null || typeof value !== "object" || Array.isArray(value)) return undefined;
+  const stage = value as Record<string, unknown>;
+  for (const key of ["upstreamFrames", "controlFrames", "relayedEvents", "pings", "pongs"] as const) {
+    if (!isNonNegativeFiniteNumber(stage[key])) return undefined;
+  }
+  if (!(stage.requestBytes === null || isNonNegativeFiniteNumber(stage.requestBytes))) return undefined;
+  if (!(stage.firstFrameMs === null || isNonNegativeFiniteNumber(stage.firstFrameMs))) return undefined;
+  if (!(stage.elapsedMs === null || isNonNegativeFiniteNumber(stage.elapsedMs))) return undefined;
+  if (!(stage.closeCode === null || (typeof stage.closeCode === "number"
+    && Number.isInteger(stage.closeCode) && stage.closeCode >= 1000 && stage.closeCode <= 4999))) {
+    return undefined;
+  }
+  if (typeof stage.sent !== "boolean" || typeof stage.reused !== "boolean") return undefined;
+  if (typeof stage.ocxVersion !== "string" || !stage.ocxVersion || stage.ocxVersion.length > 32) return undefined;
+  if (typeof stage.bunVersion !== "string" || !stage.bunVersion || stage.bunVersion.length > 32) return undefined;
+  return {
+    requestBytes: stage.requestBytes as number | null,
+    sent: stage.sent,
+    upstreamFrames: stage.upstreamFrames as number,
+    controlFrames: stage.controlFrames as number,
+    relayedEvents: stage.relayedEvents as number,
+    firstFrameMs: stage.firstFrameMs as number | null,
+    elapsedMs: stage.elapsedMs as number | null,
+    pings: stage.pings as number,
+    pongs: stage.pongs as number,
+    closeCode: stage.closeCode as number | null,
+    reused: stage.reused,
+    ocxVersion: stage.ocxVersion,
+    bunVersion: stage.bunVersion,
   };
 }
 

package/src/vision/eligibility.ts

@@ -77,9 +77,12 @@ type EnrichedProviderCache = Map<string, OcxProviderConfig>;
  * not a text-only model and must not be widened to image through the vision sidecar.
  */
 export function isModelVisionSidecarConsumer(
-  provider: Pick<OcxProviderConfig, "noVisionModels" | "modelInputModalities">,
+  provider: Pick<OcxProviderConfig, "noVisionModels" | "modelInputModalities" | "modelCapabilities">,
   modelId: string,
 ): boolean {
+  const declared = Object.hasOwn(provider.modelCapabilities ?? {}, modelId)
+    ? provider.modelCapabilities?.[modelId]?.inputModalities : undefined;
+  if (declared !== undefined) return declared.includes("text") && !declared.includes("image");
   if (modelInList(provider.noVisionModels, modelId)) return true;
   const modalities = modelRecordValue(provider.modelInputModalities, modelId);
   return Array.isArray(modalities) && modalities.includes("text") && !modalities.includes("image");
@@ -151,10 +154,18 @@ function modelAcceptsImageInputWithCache(
   candidate: VisionCandidateModel,
   cache: EnrichedProviderCache,
 ): boolean | undefined {
-  if (isVisionSidecarConsumerWithCache(config, candidate.provider, candidate.id, cache)) return false;
   if (candidate.native === true || (candidate.provider === "openai" && SUPPORTED_NATIVE_OPENAI_SLUGS.has(candidate.id))) {
+    const nativeProvider = enrichedProviderForVision(config, candidate.provider, cache);
+    if (nativeProvider && isModelVisionSidecarConsumer({
+      noVisionModels: nativeProvider.noVisionModels, modelInputModalities: nativeProvider.modelInputModalities,
+    }, candidate.id)) return false;
     return advertisesImageInput(nativeInputModalities(candidate.id)) ?? true;
   }
+  if (isVisionSidecarConsumerWithCache(config, candidate.provider, candidate.id, cache)) return false;
+  const provider = enrichedProviderForVision(config, candidate.provider, cache);
+  const declared = Object.hasOwn(provider?.modelCapabilities ?? {}, candidate.id)
+    ? provider?.modelCapabilities?.[candidate.id]?.inputModalities : undefined;
+  if (declared !== undefined) return declared.includes("image");
   const fromRow = advertisesImageInput(candidate.inputModalities);
   if (fromRow !== undefined) return fromRow;
   return metadataImageInput(candidate.provider, candidate.id);

package/src/web-search/loop.ts

@@ -3,6 +3,7 @@ import type { AdapterEvent, OcxMessage, OcxParsedRequest, OcxProviderConfig, Ocx
 import { namespacedToolName, toolChoiceToolPredicate } from "../types";
 import { cloneProviderOpaqueToolCallMetadata } from "../responses/provider-opaque-metadata";
 import type { AttemptRecoveryKind } from "../usage/log";
+import { isTruncatedStopReason } from "../responses/truncated-stop-reason";
 import { bridgeToResponsesSSE } from "../bridge";
 import { runWebSearch, type SidecarOutcome, type SidecarOutcomeRecorder, type SidecarSettings } from "./executor";
 import { runAnthropicWebSearch } from "./anthropic-executor";
@@ -230,6 +231,24 @@ function forcedAnswerNudge(): OcxMessage {
   };
 }
 
+/**
+ * Transient developer-role nudge for the ONE recovery pass after a forced answer came back empty.
+ * The recovery also removes every tool, so the model has nothing to call and can only return text;
+ * this turn says so explicitly rather than relying on the removal alone. Like {@link forcedAnswerNudge}
+ * it is iteration-local and never touches the persisted `messages`.
+ */
+function forcedAnswerRetryNudge(): OcxMessage {
+  return {
+    role: "developer",
+    content:
+      "Your previous response contained no usable answer. Web search has finished for this turn and " +
+      "no tools are available for this response. Answer the user's question now in assistant text, " +
+      "using the web search results already gathered above. If those results are insufficient, say " +
+      "what is missing instead of returning an empty response.",
+    timestamp: Date.now(),
+  };
+}
+
 function jsonError(status: number, message: string): Response {
   return new Response(JSON.stringify({ error: { message, type: "upstream_error", code: null } }), {
     status,
@@ -370,7 +389,9 @@ export async function runWithWebSearch(deps: WebSearchLoopDeps): Promise<Respons
   const signal = internalAbort.signal;
 
   // Hard iteration bound (termination safety net); forceAnswer normally ends the loop sooner.
-  const HARD_CAP = maxSearches + 2;
+  // One iteration beyond the forced answer is reserved for its empty-answer recovery below.
+  const HARD_CAP = maxSearches + 3;
+  let emptyAnswerRetries = 0;
   const connectTimeoutMs = deps.connectTimeoutMs ?? 200_000;
   const routedModelStallTimeoutMs = deps.routedModelStallTimeoutMs ?? 200_000;
 
@@ -407,12 +428,19 @@ export async function runWithWebSearch(deps: WebSearchLoopDeps): Promise<Respons
     // ignores what the search found, which reads to the user as "the search did nothing". Nudge it
     // (iteration-locally — never mutate the shared `messages`) to actually use the gathered results.
     // Only when a REAL search ran (executedSearchCount, not empty-query/limit/repeat placeholders).
-    const iterMessages: OcxMessage[] = forceAnswer && executedSearchCount > 0
+    let iterMessages: OcxMessage[] = forceAnswer && executedSearchCount > 0
       ? [...messages, forcedAnswerNudge()]
       : messages;
+    // #1001 follow-up: the recovery pass for an empty forced answer. Removing every tool leaves the
+    // model nothing to call, and the extra developer turn asks it for the text it just failed to
+    // produce. `toolChoice: "none"` is what drops those definitions in the adapter, so the retry
+    // cannot repeat the same empty or tool-shaped response.
+    const recoveringEmptyAnswer = forceAnswer && emptyAnswerRetries > 0;
+    if (recoveringEmptyAnswer) iterMessages = [...iterMessages, forcedAnswerRetryNudge()];
     const iterParsed: OcxParsedRequest = {
       ...parsed, stream: true,
-      context: { ...parsed.context, messages: iterMessages, tools: forceAnswer ? toolsNoWebSearch : allTools },
+      ...(recoveringEmptyAnswer ? { options: { ...parsed.options, toolChoice: "none" as const } } : {}),
+      context: { ...parsed.context, messages: iterMessages, tools: recoveringEmptyAnswer ? [] : forceAnswer ? toolsNoWebSearch : allTools },
     };
     // One cumulative header deadline spans every pool-key 429 rotation in this model iteration.
     // clear() stops only its timer after final headers; the direct turn signal remains attached to
@@ -847,9 +875,34 @@ export async function runWithWebSearch(deps: WebSearchLoopDeps): Promise<Respons
               // An unterminated call flushes AFTER the terminal event, so find
               // the terminal rather than assuming it is last (#1001).
               const terminalEvent = split.passthrough.find(event => event.type === "done");
+              if (terminalEvent?.type === "done" && !split.hasMalformedToolCall
+                && isTruncatedStopReason(terminalEvent.stopReason)) {
+                // A provider refusal or truncation is authoritative, even without text.
+                // Preserve it once; neither an empty-answer retry nor a generic 502 applies.
+                yield* replay(split.passthrough.slice(split.streamedPassthroughCount));
+                return;
+              }
               if (terminalEvent?.type === "done"
                 && (split.hasMalformedToolCall
                   || (!split.hasRealToolCall && !hasVisibleAssistantText(split.passthrough)))) {
+                // #1001 fixed the silent success by failing here. A malformed call still fails: it
+                // reports a protocol problem, and replaying it would only re-ask an unwell upstream.
+                // Silence is different — it is recoverable, so retry exactly once with the results
+                // already gathered before failing the turn.
+                console.warn("[web-search-loop] unusable forced answer", JSON.stringify({
+                  model: parsed.modelId,
+                  recoveryAttempt: emptyAnswerRetries,
+                  searchCalls: split.calls.length,
+                  malformed: split.hasMalformedToolCall,
+                  stopReason: terminalEvent.stopReason,
+                  eventTypes: [...new Set(split.passthrough.map(event => event.type))],
+                }));
+                if (!split.hasMalformedToolCall && !split.hasRealToolCall && emptyAnswerRetries === 0) {
+                  emptyAnswerRetries++;
+                  console.warn("[web-search-loop] empty forced answer — retrying once without tools");
+                  yield { type: "heartbeat" };
+                  continue;
+                }
                 throw new LoopError(502, "forced-answer pass produced no usable assistant output");
               }
             }