nemoris 0.1.0

package/src/utils/token-budget.js

@@ -0,0 +1,74 @@
+/**
+ * Token budget utilities — rough estimation and context pruning.
+ * Uses character-count heuristic (length / 4) to avoid tiktoken dependency.
+ */
+
+/**
+ * Estimate token count for an array of messages.
+ * @param {Array} messages
+ * @returns {number}
+ */
+export function estimateTokens(messages) {
+  return Math.ceil(JSON.stringify(messages).length / 4);
+}
+
+/**
+ * Prune stale tool_use + tool_result pairs from the start of a message list
+ * when the estimated token count exceeds 80% of maxContextTokens.
+ * Always preserves the last 4 turns.
+ *
+ * @param {Array<{role: string, content: any}>} messages
+ * @param {number} maxContextTokens
+ * @returns {{ pruned: Array, removed: number, estimatedTokens: number }}
+ */
+export function pruneContext(messages, maxContextTokens) {
+  const threshold = maxContextTokens * 0.8;
+
+  if (estimateTokens(messages) <= threshold) {
+    return { pruned: messages, removed: 0, estimatedTokens: estimateTokens(messages) };
+  }
+
+  // Identify tool_use + tool_result pair indices, from oldest to newest.
+  // A pair is: assistant turn with tool_use content followed immediately by
+  // user turn with tool_result content.
+  const pairs = [];
+  for (let i = 0; i < messages.length - 1; i++) {
+    const msg = messages[i];
+    const next = messages[i + 1];
+    if (msg.role === "assistant" && isToolUse(msg.content) &&
+        next.role === "user" && isToolResult(next.content)) {
+      pairs.push([i, i + 1]);
+    }
+  }
+
+  const protectedStart = Math.max(0, messages.length - 4);
+  // Only consider pairs where both indices are before the protected zone
+  const pruneable = pairs.filter(([a, b]) => a < protectedStart && b < protectedStart);
+
+  let working = [...messages];
+  let removed = 0;
+
+  for (const [a, b] of pruneable) {
+    if (estimateTokens(working) <= threshold) break;
+    // Adjust indices for previously removed items
+    const adj = a - removed * 2;
+    working.splice(adj, 2);
+    removed++;
+  }
+
+  return { pruned: working, removed, estimatedTokens: estimateTokens(working) };
+}
+
+function isToolUse(content) {
+  if (Array.isArray(content)) {
+    return content.some((c) => c.type === "tool_use");
+  }
+  return false;
+}
+
+function isToolResult(content) {
+  if (Array.isArray(content)) {
+    return content.some((c) => c.type === "tool_result");
+  }
+  return false;
+}

package/src/utils/usage-cost.js

@@ -0,0 +1,25 @@
+// Rough cost per 1M tokens (USD) — update as pricing changes
+const COSTS = {
+  "anthropic/claude-sonnet-4-6":  { in: 3.00, out: 15.00 },
+  "anthropic/claude-haiku-4-5":   { in: 0.80, out: 4.00 },
+  "anthropic/claude-opus-4-6":    { in: 15.00, out: 75.00 },
+  "openrouter/anthropic/claude-sonnet-4-6": { in: 3.00, out: 15.00 },
+  "openrouter/anthropic/claude-haiku-4-5":  { in: 0.80, out: 4.00 },
+  "openrouter/anthropic/claude-opus-4-6":   { in: 15.00, out: 75.00 },
+};
+
+/**
+ * Estimate the cost of an LLM turn in USD.
+ * @param {string} model The model ID (e.g. "anthropic/claude-sonnet-4-6")
+ * @param {number} tokensIn Input tokens
+ * @param {number} tokensOut Output tokens
+ * @returns {number|null} Estimated cost in USD, or null if model unknown
+ */
+export function estimateCost(model, tokensIn, tokensOut) {
+  if (typeof model === "string" && model.startsWith("ollama/")) {
+    return 0;
+  }
+  const rates = COSTS[model];
+  if (!rates) return null; // Unknown or unpriced remote models are intentionally left unresolved.
+  return (tokensIn / 1_000_000) * rates.in + (tokensOut / 1_000_000) * rates.out;
+}

package/src/utils/usage-metrics.js

@@ -0,0 +1,14 @@
+export function extractUsageMetrics(rawResponse) {
+  return {
+    tokensIn: rawResponse?.usage?.input_tokens
+      || rawResponse?.usage?.prompt_tokens
+      || 0,
+    tokensOut: rawResponse?.usage?.output_tokens
+      || rawResponse?.usage?.completion_tokens
+      || 0,
+    cacheIn: rawResponse?.usage?.cache_read_input_tokens
+      || rawResponse?.usage?.prompt_tokens_details?.cached_tokens
+      || 0,
+    cacheCreation: rawResponse?.usage?.cache_creation_input_tokens || 0,
+  };
+}