@outputai/cli 0.7.1-next.ae5bab4.0 → 0.7.1-next.bd6bd49.0

package/dist/services/cost_calculator.js

@@ -5,6 +5,43 @@ const ARRAY_ACCESS_PATTERN = /^(\w+)\[(\d+)\]$/;
 function tokenCost(tokens, pricePerMillion) {
     return (tokens / 1_000_000) * pricePerMillion;
 }
+function hostFromUrl(url) {
+    if (!url) {
+        return 'unknown';
+    }
+    return url.replace(/^https?:\/\//, '').split('/')[0] || 'unknown';
+}
+function lineRate(type, pricing) {
+    const rates = {
+        input: pricing.input ?? 0,
+        input_cached: pricing.cached_input ?? 0,
+        output: pricing.output ?? 0,
+        reasoning: pricing.reasoning ?? pricing.output ?? 0
+    };
+    return rates[type];
+}
+// Re-prices the event's usage lines at costs.yml rates. A line type without a
+// configured mapping degrades to its as-charged total rather than $0, so new
+// producer line types never silently vanish from the adjusted figure.
+function priceLines(lines, pricing) {
+    return lines.reduce((sum, line) => {
+        const rate = lineRate(line.type, pricing);
+        return sum + (rate === undefined ? line.total : tokenCost(line.amount, rate));
+    }, 0);
+}
+// Token counts for display. The producer's 'input' line excludes cached tokens
+// (sdk/llm emits input − cached), while AI SDK reports inputTokens as the
+// total — add cached back so the columns match what providers report.
+function eventTokenUsage(lines) {
+    const sumOf = (type) => lines.filter(l => l.type === type).reduce((s, l) => s + l.amount, 0);
+    const cached = sumOf('input_cached');
+    return {
+        inputTokens: sumOf('input') + cached,
+        cachedInputTokens: cached,
+        outputTokens: sumOf('output'),
+        reasoningTokens: sumOf('reasoning')
+    };
+}
 export function extractValue(obj, path) {
     if (!path || !obj) {
         return obj;
@@ -69,42 +106,37 @@ function findCalls(node, match, extract, parentStepName = null, seenIds = new Se
     }
     return calls;
 }
+// Only nodes carrying an llm:usage event are priced — the event holds the
+// as-charged cost and the per-token-type amounts. Traces from SDKs that
+// predate cost attributes report no LLM costs.
 export function findLLMCalls(node, parentStepName = null, seenIds = new Set()) {
-    return findCalls(node, n => n.kind === 'llm' && !!n.output?.usage, (n, stepName) => {
-        const loadedPrompt = n.input?.loadedPrompt;
-        const outputRecord = n.output;
-        const inputRecord = n.input;
-        const model = loadedPrompt?.config?.model ||
-            outputRecord?.model ||
-            inputRecord?.model ||
-            'unknown';
+    return findCalls(node, n => n.kind === 'llm' && !!n.attributes?.['llm:usage'], (n, stepName) => {
+        const event = n.attributes['llm:usage'];
         return {
             stepName: stepName || n.name || 'unknown',
             llmName: n.name || 'llm',
-            model,
-            usage: n.output.usage
+            model: event.modelId || 'unknown',
+            usage: eventTokenUsage(event.usage ?? []),
+            originalCost: event.total,
+            lines: event.usage ?? []
         };
     }, parentStepName, seenIds);
 }
 export function findHTTPCalls(node, parentStepName = null, seenIds = new Set()) {
-    return findCalls(node, n => n.kind === 'http', (n, stepName) => ({
-        stepName: stepName || 'unknown',
-        url: n.input?.url || '',
-        method: n.input?.method || 'GET',
-        input: n.input || {},
-        output: n.output || {},
-        status: n.output?.status
-    }), parentStepName, seenIds);
-}
-export function calculateLLMCallCost(usage, modelPricing) {
-    if (!modelPricing) {
-        return { cost: 0, warning: 'unknown model' };
-    }
-    const inputCost = tokenCost(usage.inputTokens ?? 0, modelPricing.input ?? 0);
-    const outputCost = tokenCost(usage.outputTokens ?? 0, modelPricing.output ?? 0);
-    const cachedCost = tokenCost(usage.cachedInputTokens ?? 0, modelPricing.cached_input ?? 0);
-    const reasoningCost = tokenCost(usage.reasoningTokens ?? 0, modelPricing.reasoning || modelPricing.output || 0);
-    return { cost: inputCost + outputCost + cachedCost + reasoningCost };
+    return findCalls(node, n => n.kind === 'http', (n, stepName) => {
+        const costEvent = n.attributes?.['http:request:cost'];
+        const url = costEvent?.url || n.input?.url || '';
+        return {
+            stepName: stepName || 'unknown',
+            url,
+            method: n.input?.method || 'GET',
+            input: n.input || {},
+            output: n.output || {},
+            status: n.output?.status,
+            host: hostFromUrl(url),
+            originalCost: costEvent?.total
+        };
+    }, parentStepName, seenIds);
 }
 export function identifyService(httpCall, services) {
     if (!services) {
@@ -119,31 +151,51 @@ export function identifyService(httpCall, services) {
 }
 function calculateTokenServiceCost(httpCall, config) {
     if (!config.usage_path) {
-        return { step: httpCall.stepName, cost: 0, usage: 'no usage data', warning: 'no usage data' };
+        return {
+            step: httpCall.stepName, cost: 0, usage: 'no usage data',
+            kind: 'failed', warning: 'no usage data'
+        };
     }
     const usage = extractValue(httpCall.output, config.usage_path);
     if (config.input_field && config.output_field) {
         const usageObj = usage;
-        const inputTokens = usageObj?.[config.input_field] ?? 0;
-        const outputTokens = usageObj?.[config.output_field] ?? 0;
+        if (!usageObj) {
+            return {
+                step: httpCall.stepName, cost: 0, usage: 'no usage data',
+                kind: 'failed', warning: 'no usage data'
+            };
+        }
+        const inputTokens = usageObj[config.input_field] ?? 0;
+        const outputTokens = usageObj[config.output_field] ?? 0;
         const inputCost = tokenCost(inputTokens, config.input_per_million ?? 0);
         const outputCost = tokenCost(outputTokens, config.output_per_million ?? 0);
         return {
             step: httpCall.stepName,
             cost: inputCost + outputCost,
-            usage: `${(inputTokens + outputTokens).toLocaleString('en-US')} tokens`
+            usage: `${(inputTokens + outputTokens).toLocaleString('en-US')} tokens`,
+            kind: 'computed'
         };
     }
     const tokens = typeof usage === 'number' ? usage : 0;
     if (tokens === 0) {
-        return { step: httpCall.stepName, cost: 0, usage: 'no usage data', warning: 'no usage data' };
+        return {
+            step: httpCall.stepName, cost: 0, usage: 'no usage data',
+            kind: 'failed', warning: 'no usage data'
+        };
     }
     const cost = tokenCost(tokens, config.per_million ?? 0);
-    return { step: httpCall.stepName, cost, usage: `${tokens.toLocaleString('en-US')} tokens` };
+    return {
+        step: httpCall.stepName,
+        cost,
+        usage: `${tokens.toLocaleString('en-US')} tokens`,
+        kind: 'computed'
+    };
 }
+// units: undefined means the call couldn't be measured (no endpoint match, or
+// a units_per_line endpoint without a string body) — distinct from a measured 0.
 function resolveUnitEndpoint(url, httpCall, config) {
     if (!config.endpoints) {
-        return { units: 0, endpoint: 'unknown' };
+        return { units: undefined, endpoint: 'unknown' };
     }
     for (const [endpointName, endpointConfig] of Object.entries(config.endpoints)) {
         if (!url.includes(endpointConfig.pattern)) {
@@ -160,25 +212,40 @@ function resolveUnitEndpoint(url, httpCall, config) {
                 return { units, endpoint: endpointName };
             }
         }
-        return { units: 0, endpoint: endpointName };
+        return { units: undefined, endpoint: endpointName };
     }
-    return { units: 0, endpoint: 'unknown' };
+    return { units: undefined, endpoint: 'unknown' };
 }
 function calculateUnitServiceCost(httpCall, config) {
     const { units, endpoint } = resolveUnitEndpoint(httpCall.url, httpCall, config);
+    if (units === undefined) {
+        return {
+            step: httpCall.stepName, cost: 0, usage: '0 units',
+            kind: 'failed', warning: 'unknown endpoint', endpoint
+        };
+    }
     const cost = units * (config.price_per_unit || 0);
     return {
         step: httpCall.stepName,
         cost,
         usage: `${units.toLocaleString('en-US')} units`,
+        kind: 'computed',
         endpoint
     };
 }
 function calculateRequestServiceCost(httpCall, config) {
     if (config.models && config.model_path) {
         const model = extractValue(httpCall.input, config.model_path);
-        const price = (model && config.models[model]) || config.default_price || 0;
-        return { step: httpCall.stepName, cost: price, usage: '1 request', model };
+        // ?? rather than || so a configured price of 0 (free tier) is honored.
+        const price = (model !== undefined ? config.models[model] : undefined) ??
+            config.default_price;
+        if (price === undefined) {
+            return {
+                step: httpCall.stepName, cost: 0, usage: '1 request',
+                kind: 'failed', warning: 'unknown model price', model
+            };
+        }
+        return { step: httpCall.stepName, cost: price, usage: '1 request', kind: 'computed', model };
     }
     if (config.endpoints) {
         for (const [endpointName, endpointConfig] of Object.entries(config.endpoints)) {
@@ -188,27 +255,40 @@ function calculateRequestServiceCost(httpCall, config) {
                         step: httpCall.stepName,
                         cost: endpointConfig.price,
                         usage: '1 request',
+                        kind: 'computed',
                         endpoint: endpointName
                     };
                 }
                 if (endpointConfig.price_per_item && endpointConfig.items_path) {
                     const items = extractValue(httpCall.input, endpointConfig.items_path);
-                    const count = Array.isArray(items) ? items.length : 0;
+                    // Missing/un-captured request body is not the same as zero items —
+                    // only a real array is a measured count.
+                    if (!Array.isArray(items)) {
+                        return {
+                            step: httpCall.stepName, cost: 0, usage: 'items not captured',
+                            kind: 'failed', warning: 'items not captured', endpoint: endpointName
+                        };
+                    }
                     return {
                         step: httpCall.stepName,
-                        cost: count * endpointConfig.price_per_item,
-                        usage: `${count} items`,
+                        cost: items.length * endpointConfig.price_per_item,
+                        usage: `${items.length} items`,
+                        kind: 'computed',
                         endpoint: endpointName
                     };
                 }
             }
         }
     }
-    return { step: httpCall.stepName, cost: 0, usage: 'unknown endpoint', warning: 'unknown endpoint' };
+    return {
+        step: httpCall.stepName, cost: 0, usage: 'unknown endpoint',
+        kind: 'failed', warning: 'unknown endpoint'
+    };
 }
 function calculateResponseCostService(httpCall, config) {
     const cost = extractValue(httpCall, config.cost_path);
-    if (typeof cost === 'number' && cost > 0) {
+    // A provider-reported cost — including a legitimate $0 — is an exact figure.
+    if (typeof cost === 'number') {
         const costDollars = extractValue(httpCall, 'output.body.costDollars');
         const model = extractValue(httpCall, 'output.body.model');
         const numSearches = costDollars?.numSearches ?? 0;
@@ -217,6 +297,7 @@ function calculateResponseCostService(httpCall, config) {
             step: httpCall.stepName,
             cost,
             usage: `${numSearches} searches, ${Math.round(numPages)} pages`,
+            kind: 'computed',
             model: model || 'unknown',
             details: costDollars
         };
@@ -226,27 +307,37 @@ function calculateResponseCostService(httpCall, config) {
             extractValue(httpCall, 'output.body.model') ||
             'unknown';
         const fallbackPrice = config.fallback_models[model];
-        if (fallbackPrice) {
+        if (fallbackPrice !== undefined) {
             return {
                 step: httpCall.stepName,
                 cost: fallbackPrice,
                 usage: '1 request (estimated)',
+                kind: 'estimated',
                 model,
                 warning: 'using fallback estimate'
             };
         }
-        if (config.default_fallback) {
+        if (config.default_fallback !== undefined) {
             return {
                 step: httpCall.stepName,
                 cost: config.default_fallback,
                 usage: '1 request (estimated)',
+                kind: 'estimated',
                 model: 'unknown',
                 warning: 'using default estimate'
             };
         }
     }
-    return { step: httpCall.stepName, cost: 0, usage: 'no cost data', warning: 'no cost data' };
+    return {
+        step: httpCall.stepName, cost: 0, usage: 'no cost data',
+        kind: 'failed', warning: 'no cost data'
+    };
 }
+// Body-dependent service rules (token usage paths, response_cost, per-item
+// counts, units_per_line) only produce a 'computed' result on traces recorded
+// with OUTPUT_TRACE_HTTP_VERBOSE=true (the docker-compose-dev default), since
+// production traces omit HTTP bodies — there they fall back to the as-charged
+// event cost.
 export function calculateServiceCost(httpCall, serviceInfo) {
     const { config } = serviceInfo;
     switch (config.type) {
@@ -259,31 +350,30 @@ export function calculateServiceCost(httpCall, serviceInfo) {
         case 'response_cost':
             return calculateResponseCostService(httpCall, config);
         default:
-            return { step: httpCall.stepName, cost: 0, usage: 'unknown type', warning: 'unknown type' };
+            return {
+                step: httpCall.stepName, cost: 0, usage: 'unknown type',
+                kind: 'failed', warning: 'unknown type'
+            };
     }
 }
 function findModelPricing(model, models) {
     if (models[model]) {
-        return { pricing: models[model], matchedKey: model };
+        return models[model];
     }
-    const prefixMatch = Object.entries(models).find(([key]) => model.startsWith(key));
-    return prefixMatch ?
-        { pricing: prefixMatch[1], matchedKey: prefixMatch[0] } :
-        { pricing: undefined, matchedKey: undefined };
+    return Object.entries(models).find(([key]) => model.startsWith(key))?.[1];
 }
 function aggregateLLMCosts(llmCalls, config) {
-    const unknownModels = new Set();
     const results = [];
-    const totals = { inputTokens: 0, outputTokens: 0, cachedTokens: 0, reasoningTokens: 0, cost: 0 };
+    const totals = {
+        inputTokens: 0, outputTokens: 0, cachedTokens: 0, reasoningTokens: 0,
+        originalCost: 0, adjustedCost: 0
+    };
     for (const call of llmCalls) {
-        const { pricing, matchedKey } = findModelPricing(call.model, config.models ?? {});
-        const { cost, warning } = calculateLLMCallCost(call.usage, pricing);
-        const prefixWarning = (pricing && matchedKey !== call.model) ?
-            `priced as ${matchedKey}` :
-            undefined;
-        if (!pricing) {
-            unknownModels.add(call.model);
-        }
+        const pricing = findModelPricing(call.model, config.models ?? {});
+        // Original = as-charged from the trace event. Adjusted = the event lines
+        // re-priced at costs.yml rates, when the model is configured.
+        const originalCost = call.originalCost;
+        const adjustedCost = pricing ? priceLines(call.lines, pricing) : originalCost;
         results.push({
             step: call.stepName,
             model: call.model,
@@ -291,14 +381,15 @@ function aggregateLLMCosts(llmCalls, config) {
             output: call.usage.outputTokens ?? 0,
             cached: call.usage.cachedInputTokens ?? 0,
             reasoning: call.usage.reasoningTokens ?? 0,
-            cost,
-            warning: warning ?? prefixWarning
+            originalCost,
+            adjustedCost
         });
         totals.inputTokens += call.usage.inputTokens ?? 0;
         totals.outputTokens += call.usage.outputTokens ?? 0;
         totals.cachedTokens += call.usage.cachedInputTokens ?? 0;
         totals.reasoningTokens += call.usage.reasoningTokens ?? 0;
-        totals.cost += cost;
+        totals.originalCost += originalCost;
+        totals.adjustedCost += adjustedCost;
     }
     return {
         results,
@@ -306,58 +397,79 @@ function aggregateLLMCosts(llmCalls, config) {
         totalOutputTokens: totals.outputTokens,
         totalCachedTokens: totals.cachedTokens,
         totalReasoningTokens: totals.reasoningTokens,
-        llmTotalCost: totals.cost,
-        unknownModels: [...unknownModels]
+        llmOriginalCost: totals.originalCost,
+        llmAdjustedCost: totals.adjustedCost
     };
 }
-export function calculateCost(trace, config, traceFile = '') {
-    const llmCalls = findLLMCalls(trace);
-    const httpCalls = findHTTPCalls(trace);
-    const serviceResults = {};
+function pushHTTPResult(acc, result) {
+    if (!acc[result.host]) {
+        acc[result.host] = {
+            host: result.host, calls: [], originalTotalCost: 0, adjustedTotalCost: 0
+        };
+    }
+    acc[result.host].calls.push(result);
+    acc[result.host].originalTotalCost += result.originalCost;
+    acc[result.host].adjustedTotalCost += result.adjustedCost;
+}
+// For an event-bearing (billable) request, decide the adjusted cost: apply the
+// costs.yml recompute only when it produced an exact figure ('computed' —
+// which includes a legitimate $0). Estimates and failed recomputes never
+// replace the as-charged cost, and an errored response can't be re-priced
+// from service rules even though its event proves it was charged.
+function resolveHTTPOverride(call, serviceInfo, originalCost) {
+    if (!serviceInfo || (call.status && call.status >= 400)) {
+        return { adjustedCost: originalCost, usage: 'as-charged' };
+    }
+    const recompute = calculateServiceCost(call, serviceInfo);
+    return recompute.kind === 'computed' ?
+        { adjustedCost: recompute.cost, usage: recompute.usage } :
+        { adjustedCost: originalCost, usage: 'as-charged' };
+}
+// Only calls carrying an http:request:cost event are billable — the event is
+// proof of a charge (counted regardless of HTTP status). Calls without one
+// (count-only webhooks, polling requests, uninstrumented clients) are not
+// priced.
+function aggregateHTTPCosts(httpCalls, config) {
+    const hosts = {};
     for (const call of httpCalls) {
-        if (call.status && call.status >= 400) {
+        if (call.originalCost === undefined) {
             continue;
         }
         const serviceInfo = identifyService(call, config.services);
-        if (!serviceInfo) {
-            continue;
-        }
-        if (serviceInfo.config.type === 'response_cost') {
-            const hasCostData = extractValue(call, serviceInfo.config.cost_path);
-            const isBillableMethod = serviceInfo.config.billable_method &&
-                call.method === serviceInfo.config.billable_method;
-            if (!hasCostData && !isBillableMethod) {
-                continue;
-            }
-        }
-        const result = calculateServiceCost(call, serviceInfo);
-        if (!serviceResults[serviceInfo.serviceName]) {
-            serviceResults[serviceInfo.serviceName] = {
-                serviceName: serviceInfo.serviceName,
-                calls: [],
-                totalCost: 0
-            };
-        }
-        serviceResults[serviceInfo.serviceName].calls.push(result);
-        serviceResults[serviceInfo.serviceName].totalCost += result.cost;
+        const { adjustedCost, usage } = resolveHTTPOverride(call, serviceInfo, call.originalCost);
+        pushHTTPResult(hosts, {
+            step: call.stepName,
+            host: call.host,
+            usage,
+            originalCost: call.originalCost,
+            adjustedCost
+        });
     }
-    const { results: llmResults, totalInputTokens, totalOutputTokens, totalCachedTokens, totalReasoningTokens, llmTotalCost, unknownModels } = aggregateLLMCosts(llmCalls, config);
-    const serviceTotalCost = Object.values(serviceResults).reduce((sum, s) => sum + s.totalCost, 0);
-    const totalCost = llmTotalCost + serviceTotalCost;
+    return hosts;
+}
+export function calculateCost(trace, config, traceFile = '') {
+    const llmCalls = findLLMCalls(trace);
+    const httpCalls = findHTTPCalls(trace);
+    const { results: llmResults, totalInputTokens, totalOutputTokens, totalCachedTokens, totalReasoningTokens, llmOriginalCost, llmAdjustedCost } = aggregateLLMCosts(llmCalls, config);
+    const httpCosts = Object.values(aggregateHTTPCosts(httpCalls, config));
+    const httpOriginalCost = httpCosts.reduce((sum, h) => sum + h.originalTotalCost, 0);
+    const httpAdjustedCost = httpCosts.reduce((sum, h) => sum + h.adjustedTotalCost, 0);
     const durationMs = trace.endedAt && trace.startedAt ? trace.endedAt - trace.startedAt : null;
     return {
         traceFile,
         workflowName: trace.name || 'unknown',
         durationMs,
         llmCalls: llmResults,
-        llmTotalCost,
+        llmOriginalCost,
+        llmAdjustedCost,
         totalInputTokens,
         totalOutputTokens,
         totalCachedTokens,
         totalReasoningTokens,
-        unknownModels,
-        services: Object.values(serviceResults),
-        serviceTotalCost,
-        totalCost
+        httpCosts,
+        httpOriginalCost,
+        httpAdjustedCost,
+        originalTotalCost: llmOriginalCost + httpOriginalCost,
+        totalCost: llmAdjustedCost + httpAdjustedCost
     };
 }