@heripo/ledger-extractor 0.2.1 → 0.2.3

package/dist/index.js

@@ -1139,6 +1139,72 @@ function detectProvider(model) {
   if (providerId.includes("lmstudio")) return "lmstudio";
   return "unknown";
 }
+function withSystemPromptCache(model, systemPrompt, providerOptions) {
+  const unchanged = {
+    ...systemPrompt === void 0 ? {} : { instructions: systemPrompt },
+    ...providerOptions === void 0 ? {} : { providerOptions }
+  };
+  const provider = detectProvider(model);
+  if (provider === "anthropic")
+    return systemPrompt ? {
+      ...unchanged,
+      instructions: cachedSystem(systemPrompt, {
+        anthropic: { cacheControl: { type: "ephemeral" } }
+      })
+    } : unchanged;
+  if (provider !== "openai" || !supportsOpenAICacheBreakpoints(modelIdOf(model)) || providerOptions?.openai?.promptCacheOptions !== void 0)
+    return unchanged;
+  return {
+    ...systemPrompt ? {
+      instructions: cachedSystem(systemPrompt, {
+        openai: { promptCacheBreakpoint: { mode: "explicit" } }
+      })
+    } : unchanged,
+    providerOptions: {
+      ...providerOptions,
+      openai: {
+        ...providerOptions?.openai,
+        promptCacheOptions: { mode: "explicit" }
+      }
+    }
+  };
+}
+function withRepeatedRequestCache(model, systemPrompt, userPrompt, providerOptions) {
+  const anthropicBreakpoint = {
+    anthropic: { cacheControl: { type: "ephemeral" } }
+  };
+  if (detectProvider(model) === "anthropic")
+    return {
+      ...providerOptions === void 0 ? {} : { providerOptions },
+      instructions: cachedSystem(systemPrompt, anthropicBreakpoint),
+      messages: [
+        {
+          role: "user",
+          content: userPrompt,
+          providerOptions: anthropicBreakpoint
+        }
+      ]
+    };
+  return {
+    instructions: systemPrompt,
+    prompt: userPrompt,
+    ...providerOptions === void 0 ? {} : { providerOptions }
+  };
+}
+function cachedSystem(content, providerOptions) {
+  return { role: "system", content, providerOptions };
+}
+function modelIdOf(model) {
+  const modelId = model.modelId;
+  return typeof modelId === "string" ? modelId : "";
+}
+function supportsOpenAICacheBreakpoints(modelId) {
+  const version = /^(?:openai\/)?gpt-(\d+)(?:\.(\d+))?(?:-.+)?$/u.exec(modelId);
+  if (!version) return false;
+  const major = Number(version[1]);
+  const minor = Number(version[2] ?? 0);
+  return major > 5 || major === 5 && minor >= 6;
+}
 function cacheWrite1hTokens(steps) {
   let total;
   for (const step of steps ?? []) {
@@ -1340,7 +1406,8 @@ var LLMCaller = class {
    * 2. If all fail and fallbackModel provided, try fallback up to maxRetries times
    * 3. Throw error if all attempts exhausted
    *
-   * Provider-aware strategy is automatically applied based on the model's provider field.
+   * Provider-aware strategy is automatically applied based on the model's provider field, and the prompt cache
+   * breakpoint is placed only at the end of the system prompt (see `withSystemPromptCache`).
    *
    * @template TOutput - Output type from schema validation
    * @param config - LLM call configuration
@@ -1351,17 +1418,12 @@ var LLMCaller = class {
     return this.executeWithFallback(
       config,
       (model) => this.generateStructuredOutput(model, config.schema, {
-        instructions: config.cacheSystemPrompt ? {
-          role: "system",
-          content: config.systemPrompt,
-          providerOptions: {
-            anthropic: { cacheControl: { type: "ephemeral" } }
-          }
-        } : config.systemPrompt,
+        ...withSystemPromptCache(
+          model,
+          config.systemPrompt,
+          config.providerOptions
+        ),
         prompt: config.userPrompt,
-        ...config.providerOptions && {
-          providerOptions: config.providerOptions
-        },
         temperature: config.temperature,
         maxRetries: config.maxRetries,
         abortSignal: config.abortSignal
@@ -1384,6 +1446,7 @@ var LLMCaller = class {
       config,
       (model) => this.generateStructuredOutput(model, config.schema, {
         messages: config.messages,
+        ...withSystemPromptCache(model, void 0, void 0),
         temperature: config.temperature,
         maxRetries: config.maxRetries,
         abortSignal: config.abortSignal
@@ -1906,7 +1969,6 @@ async function discoverCandidateUnit(unit, options) {
     schema: request.schema,
     systemPrompt: request.systemPrompt,
     userPrompt: request.userPrompt,
-    cacheSystemPrompt: true,
     primaryModel: options.model,
     maxRetries: options.maxRetries,
     abortSignal: options.abortSignal,
@@ -2099,7 +2161,6 @@ async function deduplicateCandidateBlocks(plan, discovery, options) {
       systemPrompt: request.systemPrompt,
       userPrompt: request.userPrompt,
       // every block shares the system prompt, so later blocks read it from the provider's prompt cache
-      cacheSystemPrompt: true,
       providerOptions: CANDIDATE_DEDUPLICATION_BLOCK_PROVIDER_OPTIONS,
       primaryModel: options.model,
       maxRetries: options.maxRetries,
@@ -8831,11 +8892,15 @@ async function attachContentThinUnit(plan, unit, options) {
             request.providerOutputSchema
           )
         }),
-        system: request.systemPrompt,
-        prompt: request.userPrompt,
+        // r2 and later runs repeat r1's request once its answer returns
+        ...withRepeatedRequestCache(
+          audited.model,
+          request.systemPrompt,
+          request.userPrompt,
+          providerOptions
+        ),
         maxRetries: options.maxRetries ?? 0,
-        abortSignal: options.abortSignal,
-        providerOptions
+        abortSignal: options.abortSignal
       });
       result = recordContentAttachmentThinCall(plan, unit, {
         run,
@@ -9894,7 +9959,7 @@ async function routeReport(index, signals, options) {
           request.providerOutputSchema
         )
       }),
-      system: request.systemPrompt,
+      ...withSystemPromptCache(audited.model, request.systemPrompt, void 0),
       prompt: request.userPrompt,
       maxRetries: options.maxRetries ?? 0,
       abortSignal: options.abortSignal
@@ -13217,8 +13282,13 @@ async function judgeRelationParentBatch(request, options) {
             request.providerOutputSchema
           )
         }),
-        system: request.systemPrompt,
-        prompt: request.userPrompt,
+        // r2 and later runs repeat r1's request once its answer returns
+        ...withRepeatedRequestCache(
+          audited.model,
+          request.systemPrompt,
+          request.userPrompt,
+          void 0
+        ),
         maxRetries: options.maxRetries ?? 0,
         abortSignal: options.abortSignal
       });
@@ -13498,7 +13568,12 @@ async function proposePlaceAliases(plan, ledger, request, options) {
             request.providerOutputSchema
           )
         }),
-        system: request.systemPrompt,
+        // r1 and r2 are sent together, so neither can read the other's user prompt from the cache
+        ...withSystemPromptCache(
+          audited.model,
+          request.systemPrompt,
+          void 0
+        ),
         prompt: request.userPrompt,
         maxRetries: options.maxRetries ?? 0,
         abortSignal: options.abortSignal