@pushpalsdev/cli 1.0.22 → 1.0.24

package/runtime/sandbox/apps/workerpals/src/execute_job.ts

@@ -2022,12 +2022,29 @@ export function shouldUseCodexCliForExecutor(executor: string): boolean {
   return executor.trim().toLowerCase() === "openai_codex";
 }
 
-function normalizeCodexReasoningEffort(value: unknown): "low" | "medium" | "high" {
+function normalizeCodexReasoningEffort(
+  value: unknown,
+  model = "",
+): "low" | "medium" | "high" | "xhigh" {
   const normalized = String(value ?? "")
     .trim()
     .toLowerCase();
-  if (normalized === "low" || normalized === "medium" || normalized === "high") {
-    return normalized;
+  const supportsExtraHigh = !/^(gpt-5\.4(?:$|-)|codex-1p(?:$|-))/i.test(String(model ?? "").trim());
+  if (
+    normalized === "low" ||
+    normalized === "medium" ||
+    normalized === "high" ||
+    normalized === "xhigh"
+  ) {
+    return normalized === "xhigh" && !supportsExtraHigh ? "high" : normalized;
+  }
+  if (
+    normalized === "extra high" ||
+    normalized === "extra-high" ||
+    normalized === "extrahigh" ||
+    normalized === "x-high"
+  ) {
+    return supportsExtraHigh ? "xhigh" : "high";
   }
   return "high";
 }
@@ -2090,6 +2107,7 @@ async function generateCommitMessageFromDiffViaCodex(
   })();
   const reasoningEffort = normalizeCodexReasoningEffort(
     runtimeConfig.workerpals.llm.reasoningEffort,
+    model,
   );
   const tmpOutputPath = resolve(
     Bun.env.TEMP || Bun.env.TMP || Bun.env.TMPDIR || "/tmp",

package/runtime/sandbox/apps/workerpals/src/workerpals_main.ts

@@ -91,6 +91,99 @@ function workerLlmConfig(runtimeConfig: ReturnType<typeof loadPushPalsConfig>):
   };
 }
 
+function estimateTokensFromText(text: string): number {
+  return Math.max(0, Math.ceil(String(text ?? "").length / 3));
+}
+
+function buildWorkerLlmUsageEvent(
+  job: {
+    kind: string;
+    sessionId?: string | null;
+    params?: Record<string, unknown> | null;
+  },
+  result: WorkerJobResult,
+): Record<string, unknown> | null {
+  const sessionId = String(job.sessionId ?? CONFIG.sessionId ?? "").trim();
+  if (!sessionId) return null;
+  const llmConfig = workerLlmConfig(CONFIG);
+  const explicitUsage = result.usage;
+  if (
+    explicitUsage &&
+    Number.isFinite(explicitUsage.promptTokens) &&
+    explicitUsage.promptTokens >= 0 &&
+    Number.isFinite(explicitUsage.completionTokens) &&
+    explicitUsage.completionTokens >= 0
+  ) {
+    const promptTokens = Math.round(explicitUsage.promptTokens);
+    const completionTokens = Math.round(explicitUsage.completionTokens);
+    const totalTokens =
+      Number.isFinite(explicitUsage.totalTokens) && (explicitUsage.totalTokens ?? 0) >= 0
+        ? Math.round(explicitUsage.totalTokens ?? promptTokens + completionTokens)
+        : promptTokens + completionTokens;
+    return {
+      service: "workerpals",
+      sessionId,
+      backend: String(explicitUsage.backend ?? resolveExecutor(CONFIG)).trim() || resolveExecutor(CONFIG),
+      modelId: String(explicitUsage.modelId ?? llmConfig.model).trim() || llmConfig.model,
+      promptTokens,
+      completionTokens,
+      totalTokens,
+      estimated: explicitUsage.estimated === true,
+    };
+  }
+
+  const promptSource = (() => {
+    try {
+      return JSON.stringify({
+        kind: job.kind,
+        params: job.params ?? {},
+      });
+    } catch {
+      return `${job.kind}\n${String(job.params?.instruction ?? job.params?.prompt ?? "")}`.trim();
+    }
+  })();
+  const completionSource = [result.summary, result.stdout ?? "", result.stderr ?? ""]
+    .filter(Boolean)
+    .join("\n\n");
+  const promptTokens = estimateTokensFromText(promptSource);
+  const completionTokens = estimateTokensFromText(completionSource);
+  return {
+    service: "workerpals",
+    sessionId,
+    backend: resolveExecutor(CONFIG),
+    modelId: llmConfig.model,
+    promptTokens,
+    completionTokens,
+    totalTokens: promptTokens + completionTokens,
+    estimated: true,
+  };
+}
+
+async function reportWorkerLlmUsage(
+  server: string,
+  headers: Record<string, string>,
+  job: {
+    kind: string;
+    sessionId?: string | null;
+    params?: Record<string, unknown> | null;
+  },
+  result: WorkerJobResult,
+): Promise<void> {
+  const payload = buildWorkerLlmUsageEvent(job, result);
+  if (!payload) return;
+  const response = await fetch(`${server}/telemetry/llm-usage`, {
+    method: "POST",
+    headers,
+    body: JSON.stringify(payload),
+  });
+  if (!response.ok) {
+    const detail = await response.text().catch(() => "");
+    throw new Error(
+      `usage telemetry rejected (${response.status})${detail ? `: ${detail.trim()}` : ""}`,
+    );
+  }
+}
+
 function integrationBranchName(): string {
   const configuredBaseRef = CONFIG.workerpals.baseRef.trim();
   if (!configuredBaseRef) return "main_agents";
@@ -1047,6 +1140,15 @@ async function workerLoop(
             const jobDurationMs = Math.max(0, Date.now() - jobStartedAtMs);
 
             await logChain;
+            try {
+              await reportWorkerLlmUsage(opts.server, headers, jobData, result);
+            } catch (err) {
+              console.warn(
+                `[WorkerPals] Failed to report LLM usage for job ${job.id}: ${
+                  err instanceof Error ? err.message : String(err)
+                }`,
+              );
+            }
 
             let completionCommit: CommitRef | null = null;
             if (result.ok && shouldCommit(job.kind, CONFIG)) {

package/runtime/sandbox/configs/default.toml

@@ -21,6 +21,7 @@ port = 3001
 debug_http = false
 stale_claim_ttl_ms = 120000
 stale_claim_sweep_interval_ms = 5000
+session_token_budget = 2000000
 
 [localbuddy]
 enabled = false

package/runtime/sandbox/configs/local.example.toml

@@ -8,7 +8,7 @@
 
 [localbuddy.llm]
 backend = "openai_codex"
-model = "gpt-5-codex"
+model = "gpt-5.4"
 codex_auth_mode = "chatgpt"
 codex_bin = "bun x --yes @openai/codex"
 codex_timeout_ms = 120000
@@ -16,7 +16,7 @@ reasoning_effort = "high"
 
 [remotebuddy.llm]
 backend = "openai_codex"
-model = "gpt-5-codex"
+model = "gpt-5.4"
 codex_auth_mode = "chatgpt"
 codex_bin = "bun x --yes @openai/codex"
 codex_timeout_ms = 120000
@@ -42,7 +42,7 @@ retention_days = 30
 
 [workerpals.llm]
 backend = "openai_codex"
-model = "gpt-5-codex"
+model = "gpt-5.4"
 codex_auth_mode = "chatgpt"
 codex_bin = "bun x --yes @openai/codex"
 codex_timeout_ms = 120000

package/runtime/sandbox/packages/shared/src/config.ts

@@ -82,6 +82,8 @@ export interface PushPalsConfig {
     debugHttp: boolean;
     staleClaimTtlMs: number;
     staleClaimSweepIntervalMs: number;
+    sessionTokenBudget: number;
+    sessionTokenBudgetAction: "pause";
   };
   localbuddy: {
     enabled: boolean;
@@ -682,6 +684,14 @@ export function loadPushPalsConfig(options: LoadOptions = {}): PushPalsConfig {
       5_000,
     ),
   );
+  const sessionTokenBudget = Math.max(
+    0,
+    asInt(
+      parseIntEnv("PUSHPALS_SESSION_TOKEN_BUDGET") ?? serverNode.session_token_budget,
+      1_000_000,
+    ),
+  );
+  const sessionTokenBudgetAction: "pause" = "pause";
 
   const globalStatusHeartbeatMs = parseIntEnv("PUSHPALS_STATUS_HEARTBEAT_MS");
 
@@ -1467,6 +1477,8 @@ export function loadPushPalsConfig(options: LoadOptions = {}): PushPalsConfig {
       debugHttp,
       staleClaimTtlMs,
       staleClaimSweepIntervalMs,
+      sessionTokenBudget,
+      sessionTokenBudgetAction,
     },
     localbuddy: {
       enabled: localEnabled,