@tryarcanist/cli 0.1.287 → 0.1.288

package/dist/index.js

@@ -8593,6 +8593,10 @@ var OpenAIModel = {
   GPT52ChatLatest: "gpt-5.2-chat-latest",
   GPT52Codex: "gpt-5.2-codex"
 };
+var DeepSeekModel = {
+  V4Flash: "deepseek-v4-flash",
+  V4Pro: "deepseek-v4-pro"
+};
 var MODEL_PROVIDERS_SET = /* @__PURE__ */ new Set(["openai", "anthropic"]);
 var BACKEND_DESKTOP_IMAGE_FEEDBACK_CONFIGS = {
   [CODEX_AGENT_RUNTIME_BACKEND]: {
@@ -8839,6 +8843,67 @@ var MODEL_REGISTRY = [
     contextWindow: 4e5,
     reasoning: { efforts: ["low", "medium", "high", "xhigh"], default: "high" },
     pricing: { inputPerMillion: 1.75, outputPerMillion: 14, cacheReadPerMillion: 0.175 }
+  },
+  {
+    id: DeepSeekModel.V4Flash,
+    name: "DeepSeek V4 Flash",
+    // `provider` names the WIRE PROTOCOL the gateway speaks (OpenAI Responses),
+    // not the vendor. DeepSeek V4 is served by Baseten: the gateway maps these
+    // registry ids to Baseten wire ids and swaps the upstream host, while every
+    // request/response stays OpenAI-shaped. Do NOT "fix" this to a new
+    // "baseten" provider — `ModelProvider` is what review-model validation, the
+    // credential gate (`PROVIDER_ENV_VAR` / spawn credential resolution), and
+    // the sandbox bridge's model guard all key off, so a new provider value
+    // silently fails those closed instead of routing anywhere.
+    provider: "openai",
+    backends: [CODEX_AGENT_RUNTIME_BACKEND],
+    // Required for codex session-start eligibility: the codex backend is the
+    // sole consumer of this flag and drops any model without it from the
+    // session-start set (see modelSupportsRequiredSessionStartCapabilities).
+    capabilities: { codexToolSearch: true },
+    contextWindow: 1048576,
+    // Only the effort our live Baseten probes actually proved. Registry efforts
+    // are treated as supported capabilities and are selectable by other paths
+    // (isValidReasoningEffort, review config, CLI overrides), so widen this list
+    // only after probing each added value end-to-end against Baseten.
+    reasoning: { efforts: ["high"], default: "high" },
+    // Verified against https://www.baseten.co/pricing on 2026-08-11 (Model APIs
+    // token pricing, row `deepseek-ai/DeepSeek-V4-Flash-0731`): $0.13/M input,
+    // $0.028/M cached input, $0.26/M output. Baseten publishes a cached-input
+    // rate for this model, so cacheReadPerMillion carries it rather than
+    // mirroring the input rate. No longContext and no flex: both are OpenAI-only
+    // pricing axes and Baseten prices a single flat tier.
+    pricing: { inputPerMillion: 0.13, outputPerMillion: 0.26, cacheReadPerMillion: 0.028 },
+    sessionStart: { eligible: true },
+    visibility: "internal_probe"
+    // Deliberately no overloadFallback. A Baseten 429 must not silently re-run
+    // an A/B arm on a different model; with no fallback the loop retries the
+    // same model, which is what an A/B comparison requires.
+  },
+  {
+    id: DeepSeekModel.V4Pro,
+    name: "DeepSeek V4 Pro",
+    // See the DeepSeek V4 Flash entry above: "openai" is the wire protocol
+    // (OpenAI Responses), not the vendor. Changing it to a "baseten" provider
+    // breaks review-model validation, the credential gate, and the sandbox
+    // bridge's model guard.
+    provider: "openai",
+    backends: [CODEX_AGENT_RUNTIME_BACKEND],
+    // Required for codex session-start eligibility (sole consumer of the flag).
+    capabilities: { codexToolSearch: true },
+    contextWindow: 262144,
+    // Only the probed effort; widen only after probing each value against
+    // Baseten, since registry efforts are selectable capabilities.
+    reasoning: { efforts: ["high"], default: "high" },
+    // Verified against https://www.baseten.co/pricing on 2026-08-11 (Model APIs
+    // token pricing, row `DeepSeek V4 Pro`): $1.74/M input, $0.145/M cached
+    // input, $3.48/M output. Cached-input rate is published, so it is carried
+    // verbatim. No longContext and no flex (OpenAI-only axes).
+    pricing: { inputPerMillion: 1.74, outputPerMillion: 3.48, cacheReadPerMillion: 0.145 },
+    sessionStart: { eligible: true },
+    visibility: "internal_probe"
+    // Deliberately no overloadFallback — see DeepSeek V4 Flash: an A/B arm must
+    // retry the same model on a Baseten 429, never silently swap models.
   }
 ];
 var MODEL_PROVIDER_NAMES = {
@@ -8931,6 +8996,7 @@ function extractModelId(raw) {
 function getProviderForModel(modelId) {
   return MODEL_PROVIDERS[modelId] ?? "openai";
 }
+var BASETEN_SERVED_MODEL_IDS = new Set(Object.values(DeepSeekModel));
 function toModelSelection(raw) {
   const modelID = extractModelId(raw);
   if (!modelID) return void 0;
@@ -11273,6 +11339,7 @@ var ERROR_CODES = [
   "api_error",
   "model_overloaded",
   "config_error",
+  "github_pr_context",
   "github_app_permission",
   "failed_edits",
   // No producer since the memory subsystem was removed. Retained because the code is
@@ -11319,6 +11386,7 @@ var ERROR_CODE_LABELS = {
   api_error: "Model service error",
   model_overloaded: "Model at capacity",
   config_error: "Configuration error",
+  github_pr_context: "GitHub pull request context unavailable",
   github_app_permission: "GitHub App permissions missing",
   failed_edits: "Edit failure",
   memory_enforcement_failed: "Memory enforcement failed",

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tryarcanist/cli",
-  "version": "0.1.287",
+  "version": "0.1.288",
   "description": "CLI for Arcanist - create and manage coding agent sessions",
   "type": "module",
   "bin": {