@tryarcanist/cli 0.1.286 → 0.1.288

package/dist/index.js

@@ -8804,47 +8804,6 @@ var MODEL_REGISTRY = [
     sessionStart: { eligible: true },
     visibility: "internal_probe"
   },
-  {
-    // Served by Fireworks through Ramp Router, never by OpenAI. Pricing is Fireworks' published
-    // per-million rate, confirmed against billed dashboard rows to the cent on 2026-08-12 --
-    // Ramp's own `GET /v1/models` catalog misprices the OpenAI family and is not a source here.
-    id: DeepSeekModel.V4Flash,
-    name: "DeepSeek V4 Flash",
-    provider: "openai",
-    backends: [CODEX_AGENT_RUNTIME_BACKEND],
-    capabilities: { codexToolSearch: true },
-    contextWindow: 1048576,
-    reasoning: { efforts: ["high"], default: "high" },
-    pricing: { inputPerMillion: 0.14, outputPerMillion: 0.28, cacheReadPerMillion: 0.028 },
-    // Startable but hidden, the same two-axis shape gpt-5.3-codex-spark uses. `sessionStart` is
-    // what lets an explicitly authorized caller actually start a session on this model --
-    // createSessionState resolves through extractSessionStartModelIdForBackend, which reads this
-    // block. `visibility` is what keeps it out of every customer surface: the public provider
-    // groups filter on it, and extractSessionStartModelIdAnyBackend refuses a non-subscription
-    // probe outright, so no stored default or automatic route can reach it.
-    sessionStart: { eligible: true },
-    visibility: "internal_probe"
-  },
-  {
-    // The second candidate on the Ramp fallback list, so a Flash capacity 429 does not strand a
-    // review. Roughly 12x Flash's rate, which is why the budget reservation prices against Pro.
-    id: DeepSeekModel.V4Pro,
-    name: "DeepSeek V4 Pro",
-    provider: "openai",
-    backends: [CODEX_AGENT_RUNTIME_BACKEND],
-    capabilities: { codexToolSearch: true },
-    contextWindow: 1048576,
-    reasoning: { efforts: ["high"], default: "high" },
-    pricing: { inputPerMillion: 1.74, outputPerMillion: 3.48, cacheReadPerMillion: 0.348 },
-    // Startable but hidden, the same two-axis shape gpt-5.3-codex-spark uses. `sessionStart` is
-    // what lets an explicitly authorized caller actually start a session on this model --
-    // createSessionState resolves through extractSessionStartModelIdForBackend, which reads this
-    // block. `visibility` is what keeps it out of every customer surface: the public provider
-    // groups filter on it, and extractSessionStartModelIdAnyBackend refuses a non-subscription
-    // probe outright, so no stored default or automatic route can reach it.
-    sessionStart: { eligible: true },
-    visibility: "internal_probe"
-  },
   {
     id: OpenAIModel.GPT53Codex,
     name: "GPT-5.3 Codex",
@@ -8884,6 +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 = {
@@ -8954,14 +8974,12 @@ function splitModelIdentifier(value) {
 }
 function getRawModelValue(raw) {
   if (typeof raw === "string") return asNonEmptyString(raw);
-  if (!raw || typeof raw !== "object") return void 0;
-  const value = raw;
-  return asNonEmptyString(value.modelID) ?? asNonEmptyString(value.modelId) ?? asNonEmptyString(value.id);
+  if (!isRecord(raw)) return void 0;
+  return asNonEmptyString(raw.modelID) ?? asNonEmptyString(raw.modelId) ?? asNonEmptyString(raw.id);
 }
 function getExplicitProvider(raw) {
-  if (raw && typeof raw === "object") {
-    const value = raw;
-    const providerID = asNonEmptyString(value.providerID) ?? asNonEmptyString(value.providerId);
+  if (isRecord(raw)) {
+    const providerID = asNonEmptyString(raw.providerID) ?? asNonEmptyString(raw.providerId);
     if (providerID && MODEL_PROVIDERS_SET.has(providerID)) {
       return providerID;
     }
@@ -8978,6 +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;
@@ -9273,7 +9292,7 @@ function yamlLineForPath(doc, keyPath) {
 }
 function assertKnownKeys(path, doc, value) {
   const issues = [];
-  if (!value || typeof value !== "object" || Array.isArray(value)) {
+  if (!isRecord(value)) {
     return [issue(path, "manifest must be a YAML object", { code: "manifest_type" })];
   }
   const root = value;
@@ -9292,7 +9311,7 @@ function assertKnownKeys(path, doc, value) {
       );
     }
   }
-  if (root.layer && typeof root.layer === "object" && !Array.isArray(root.layer)) {
+  if (isRecord(root.layer)) {
     for (const key of Object.keys(root.layer)) {
       if (key !== "dockerfile") {
         issues.push(
@@ -9305,7 +9324,7 @@ function assertKnownKeys(path, doc, value) {
       }
     }
   }
-  if (root.smoke && typeof root.smoke === "object" && !Array.isArray(root.smoke)) {
+  if (isRecord(root.smoke)) {
     for (const key of Object.keys(root.smoke)) {
       if (key !== "commands") {
         issues.push(
@@ -9335,7 +9354,7 @@ function parseSandboxLayerManifest(path, text) {
   }
   const parsed = doc.toJSON();
   const issues = assertKnownKeys(path, doc, parsed);
-  const raw = parsed && typeof parsed === "object" && !Array.isArray(parsed) ? parsed : {};
+  const raw = isRecord(parsed) ? parsed : {};
   if (!Object.prototype.hasOwnProperty.call(raw, "version")) {
     issues.push(issue(path, "version is required", { field: "version", code: "required" }));
   } else if (raw.version !== 1) {
@@ -9368,7 +9387,7 @@ function parseSandboxLayerManifest(path, text) {
     );
   }
   let normalizedDockerfile;
-  const dockerfile = layer && typeof layer === "object" && !Array.isArray(layer) ? layer.dockerfile : void 0;
+  const dockerfile = isRecord(layer) ? layer.dockerfile : void 0;
   if (typeof dockerfile !== "string") {
     issues.push(
       issue(path, "layer.dockerfile must be a string", {
@@ -9393,7 +9412,7 @@ function parseSandboxLayerManifest(path, text) {
     }
   }
   const commandsNode = (0, import_yaml.isMap)(doc.contents) ? doc.getIn(["smoke", "commands"], true) : void 0;
-  const rawCommands = raw.smoke?.commands ?? [];
+  const rawCommands = isRecord(raw.smoke) ? raw.smoke.commands : [];
   const smokeCommands = [];
   if (!Array.isArray(rawCommands)) {
     issues.push(
@@ -9633,7 +9652,7 @@ function normalizeSandboxLayerInstructions(instructions) {
 }
 function sortJsonValue(value) {
   if (Array.isArray(value)) return value.map(sortJsonValue);
-  if (!value || typeof value !== "object") return value;
+  if (!isRecord(value)) return value;
   return Object.fromEntries(
     Object.entries(value).sort(([left], [right]) => left.localeCompare(right)).map(([key, entry]) => [key, sortJsonValue(entry)])
   );
@@ -10255,7 +10274,7 @@ function isOnboardingStageStatus(value) {
   return typeof value === "string" && ONBOARDING_STAGE_STATUSES.includes(value);
 }
 function parseOnboardingStageEvent(input) {
-  if (!input || typeof input !== "object" || Array.isArray(input)) return null;
+  if (!isRecord(input)) return null;
   const record = input;
   if (!isOnboardingStage(record.stage)) return null;
   const status = isOnboardingStageStatus(record.status) ? record.status : "started";
@@ -11320,6 +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
@@ -11366,6 +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",
@@ -11650,7 +11671,8 @@ function parseSsePayload(payload) {
   let status = null;
   const events = [];
   for (const message of messages) {
-    const data = message.data ? parseJsonObject(message.data) : {};
+    const parsed = message.data ? parseJsonObject(message.data) : {};
+    const data = isRecord(parsed) ? parsed : {};
     if (message.event === "status") {
       const phaseRaw = data.phase;
       const phase = typeof phaseRaw === "string" && VALID_PHASES.has(phaseRaw) ? phaseRaw : null;
@@ -11683,8 +11705,7 @@ function validateEnumField(value, validSet) {
 }
 function parseJsonObject(value) {
   try {
-    const parsed = JSON.parse(value);
-    return parsed && typeof parsed === "object" ? parsed : {};
+    return JSON.parse(value);
   } catch (err) {
     throw new Error(`Malformed SSE JSON payload: ${stringifyError(err)}`);
   }
@@ -11703,7 +11724,7 @@ function buildPromptLabelMap(prompts) {
   return labels;
 }
 function renderWatchEvent(event, state) {
-  const data = event.data;
+  const data = isRecord(event.data) ? event.data : {};
   switch (event.type) {
     case "text":
       return { kind: "text", text: String(data.text ?? "") };
@@ -11980,7 +12001,7 @@ async function getSessionCommand(sessionId, options, command) {
     writeJson(payload);
     return;
   }
-  const session = payload.session && typeof payload.session === "object" ? payload.session : payload;
+  const session = payload.session && typeof payload.session === "object" && !Array.isArray(payload.session) ? payload.session : payload;
   console.log(`Session: ${String(session.sessionId ?? session.id ?? sessionId)}`);
   console.log(`Status: ${String(session.phase ?? session.status ?? "unknown")}`);
   if (session.repoUrl) console.log(`Repo: ${String(session.repoUrl)}`);
@@ -12062,7 +12083,7 @@ async function usageCommand(sessionId, options, command) {
     writeJson(payload);
     return;
   }
-  const usage = payload.usage && typeof payload.usage === "object" ? payload.usage : payload;
+  const usage = payload.usage && typeof payload.usage === "object" && !Array.isArray(payload.usage) ? payload.usage : payload;
   console.log(`Input tokens: ${String(usage.inputTokens ?? 0)}`);
   console.log(`Output tokens: ${String(usage.outputTokens ?? 0)}`);
   console.log(`Total tokens: ${String(usage.totalTokens ?? 0)}`);

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tryarcanist/cli",
-  "version": "0.1.286",
+  "version": "0.1.288",
   "description": "CLI for Arcanist - create and manage coding agent sessions",
   "type": "module",
   "bin": {