@tryarcanist/cli 0.1.287 → 0.1.289

package/dist/index.js

@@ -8524,6 +8524,19 @@ async function egressValidateCommand(path = EGRESS_ALLOWLIST_SOURCE_PATH, option
   });
 }
 
+// src/commands/hades.ts
+async function hadesCommand(options = {}, command) {
+  assertArcanistSessionMutationAllowed("hades");
+  const { config } = resolveBusinessContext(command, options);
+  const result = await apiFetch(config, "/api/hades-runs", {
+    method: "POST",
+    body: "{}"
+  });
+  emit(command, options, result, (payload) => {
+    console.log(`Started Hades run ${payload.runId} (${payload.jobId}).`);
+  });
+}
+
 // src/commands/login.ts
 async function loginCommand(options, command) {
   const runtime = getRuntimeOptions(command, options);
@@ -8593,6 +8606,10 @@ 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 +8856,67 @@ 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 +9009,7 @@ 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;
@@ -10286,7 +10365,7 @@ function getRawSessionEventPromptId(event) {
   if (isCanonicalRawSessionEvent(event)) {
     return canonicalPromptId(event);
   }
-  return typeof event.data?.promptId === "string" && event.data.promptId.length > 0 ? event.data.promptId : void 0;
+  return isRecord(event.data) && typeof event.data.promptId === "string" && event.data.promptId.length > 0 ? event.data.promptId : void 0;
 }
 function getRawSessionEventKind(event) {
   if (!isCanonicalRawSessionEvent(event)) return event.type;
@@ -10332,7 +10411,7 @@ function getRawSessionEventTimestamp(event) {
   if (isCanonicalRawSessionEvent(event)) {
     return Number.isFinite(event.timestampMs) && Number.isFinite(new Date(event.timestampMs).getTime()) ? new Date(event.timestampMs).toISOString() : void 0;
   }
-  const dataTimestamp = event.data?.timestamp;
+  const dataTimestamp = isRecord(event.data) ? event.data.timestamp : void 0;
   if (typeof dataTimestamp === "string" && dataTimestamp.length > 0 || typeof dataTimestamp === "number") {
     const parsed = new Date(dataTimestamp);
     if (Number.isFinite(parsed.getTime())) return parsed.toISOString();
@@ -10411,6 +10490,7 @@ function getRawSessionEventData(event) {
         ...base,
         error: payload.message,
         ...typeof payload.code === "string" ? { code: payload.code } : {},
+        ...typeof payload.command === "string" ? { command: payload.command } : {},
         ...payload.details !== void 0 ? { errorDetails: payload.details } : {}
       });
     case "bridge.event":
@@ -10695,21 +10775,20 @@ function mergeMemoryRefs(ids, rawRefs) {
   const byId = /* @__PURE__ */ new Map();
   if (Array.isArray(rawRefs)) {
     for (const entry of rawRefs) {
-      if (!entry || typeof entry !== "object") continue;
-      const record = entry;
-      const id = typeof record.id === "string" ? record.id.trim() : "";
+      if (!isRecord(entry)) continue;
+      const id = typeof entry.id === "string" ? entry.id.trim() : "";
       if (!id) continue;
       const ref = { id };
-      if (typeof record.path === "string" && record.path.trim()) ref.path = record.path.trim();
-      if (typeof record.title === "string" && record.title.trim()) ref.title = record.title.trim();
-      if (typeof record.selectionRank === "number") ref.selectionRank = record.selectionRank;
-      if (typeof record.selectionScore === "number") ref.selectionScore = record.selectionScore;
-      if (typeof record.reason === "string" && record.reason.trim()) ref.reason = record.reason.trim();
-      if (typeof record.expectedEffect === "string" && record.expectedEffect.trim()) {
-        ref.expectedEffect = record.expectedEffect.trim();
+      if (typeof entry.path === "string" && entry.path.trim()) ref.path = entry.path.trim();
+      if (typeof entry.title === "string" && entry.title.trim()) ref.title = entry.title.trim();
+      if (typeof entry.selectionRank === "number") ref.selectionRank = entry.selectionRank;
+      if (typeof entry.selectionScore === "number") ref.selectionScore = entry.selectionScore;
+      if (typeof entry.reason === "string" && entry.reason.trim()) ref.reason = entry.reason.trim();
+      if (typeof entry.expectedEffect === "string" && entry.expectedEffect.trim()) {
+        ref.expectedEffect = entry.expectedEffect.trim();
       }
-      if (typeof record.observedEffect === "string" && record.observedEffect.trim()) {
-        ref.observedEffect = record.observedEffect.trim();
+      if (typeof entry.observedEffect === "string" && entry.observedEffect.trim()) {
+        ref.observedEffect = entry.observedEffect.trim();
       }
       byId.set(id, ref);
     }
@@ -10758,7 +10837,8 @@ function projectSessionError(data, index) {
     id: resolveEventId(data, "err", index),
     error: typeof data?.error === "string" ? data.error : "Unknown error",
     ...resolvePromptId(data) ? { promptId: resolvePromptId(data) } : {},
-    ...typeof data?.code === "string" ? { code: data.code } : {}
+    ...typeof data?.code === "string" ? { code: data.code } : {},
+    ...typeof data?.command === "string" ? { command: data.command } : {}
   };
 }
 function recordMalformedSearchBlock(data, state) {
@@ -11273,6 +11353,7 @@ 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 +11400,7 @@ 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",
@@ -12329,6 +12411,13 @@ The result is posted as a PR comment marked arcanist-anubis:v1.
 Anubis requires a write-scoped CLI token and repository write access.
 `
 ).action((prUrl, options, command) => anubisCommand(prUrl, options, command));
+program.command("hades").description("Start an internal Hades adversarial Zeus probe").addHelpText(
+  "after",
+  `
+Hades is internal-only. It creates one proved adversarial patch in tryarcanist/arcanist;
+the control plane opens the draft pull request and triggers Zeus.
+`
+).action((options, command) => hadesCommand(options, command));
 sessions.command("stop").description("Stop the active run for a session").argument("<session-id>", "Session ID").addHelpText(
   "after",
   `

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tryarcanist/cli",
-  "version": "0.1.287",
+  "version": "0.1.289",
   "description": "CLI for Arcanist - create and manage coding agent sessions",
   "type": "module",
   "bin": {