Make file-picker free

Author: jahooma

common/src/constants/free-agents.ts

@@ -0,0 +1,30 @@
+import { parseAgentId } from '../util/agent-id-parsing'
+
+/**
+ * Agents that don't charge credits.
+ *
+ * These are typically lightweight utility agents that:
+ * - Use cheap models (e.g., Gemini Flash Lite)
+ * - Have limited, programmatic capabilities
+ * - Are frequently spawned as subagents
+ *
+ * Making them free avoids user confusion when they connect their own
+ * Claude subscription (BYOK) but still see credit charges for non-Claude models.
+ */
+export const FREE_TIER_AGENTS = new Set([
+  'file-picker',
+  'file-picker-max',
+  'file-lister',
+])
+
+/**
+ * Check if an agent should be free (no credit charge).
+ * Handles all agent ID formats:
+ * - 'file-picker'
+ * - 'file-picker@1.0.0'
+ * - 'codebuff/file-picker@0.0.2'
+ */
+export function isFreeAgent(fullAgentId: string): boolean {
+  const { agentId } = parseAgentId(fullAgentId)
+  return agentId ? FREE_TIER_AGENTS.has(agentId) : false
+}

web/src/llm-api/helpers.ts

@@ -1,5 +1,6 @@
 import { setupBigQuery } from '@codebuff/bigquery'
 import { consumeCreditsAndAddAgentStep } from '@codebuff/billing'
+import { isFreeAgent } from '@codebuff/common/constants/free-agents'
 import { PROFIT_MARGIN } from '@codebuff/common/old-constants'
 
 import type { InsertMessageBigqueryFn } from '@codebuff/common/types/contracts/bigquery'
@@ -101,7 +102,7 @@ export async function consumeCreditsForMessage(params: {
   usageData: UsageData
   byok: boolean
   logger: Logger
-}) {
+}): Promise<number> {
   const {
     messageId,
     userId,
@@ -118,6 +119,11 @@ export async function consumeCreditsForMessage(params: {
     logger,
   } = params
 
+  // Free tier agents (like file-picker) don't charge credits to avoid confusion
+  // when users connect their Claude subscription but subagents use other models
+  const initialCredits = Math.round(usageData.cost * 100 * (1 + PROFIT_MARGIN))
+  const credits = isFreeAgent(agentId) && initialCredits < 5 ? 0 : initialCredits
+
   await consumeCreditsAndAddAgentStep({
     messageId,
     userId,
@@ -130,7 +136,7 @@ export async function consumeCreditsForMessage(params: {
     reasoningText,
     response: responseText,
     cost: usageData.cost,
-    credits: Math.round(usageData.cost * 100 * (1 + PROFIT_MARGIN)),
+    credits,
     inputTokens: usageData.inputTokens,
     cacheCreationInputTokens: null,
     cacheReadInputTokens: usageData.cacheReadInputTokens,
@@ -140,4 +146,6 @@ export async function consumeCreditsForMessage(params: {
     byok,
     logger,
   })
+
+  return credits
 }

web/src/llm-api/openrouter.ts

@@ -1,3 +1,4 @@
+import { PROFIT_MARGIN } from '@codebuff/common/constants/limits'
 import { getErrorObject } from '@codebuff/common/util/error'
 import { env } from '@codebuff/internal/env'
 
@@ -17,6 +18,13 @@ import type { InsertMessageBigqueryFn } from '@codebuff/common/types/contracts/b
 import type { Logger } from '@codebuff/common/types/contracts/logger'
 
 type StreamState = { responseText: string; reasoningText: string }
+
+/** Result from processing a line, including optional billed credits for final chunk */
+type LineResult = {
+  state: StreamState
+  billedCredits?: number
+}
+
 function createOpenRouterRequest(params: {
   body: any
   openrouterApiKey: string | null
@@ -143,8 +151,8 @@ export async function handleOpenRouterNonStream({
       logger.error({ error }, 'Failed to insert message into BigQuery')
     })
 
-    // Consume credits
-    await consumeCreditsForMessage({
+    // Consume credits and get the actual billed amount
+    const billedCredits = await consumeCreditsForMessage({
       messageId: firstData.id,
       userId,
       stripeCustomerId,
@@ -175,7 +183,9 @@ export async function handleOpenRouterNonStream({
         completion_tokens: aggregatedUsage.outputTokens,
         total_tokens:
           aggregatedUsage.inputTokens + aggregatedUsage.outputTokens,
-        cost: aggregatedUsage.cost,
+        // Overwrite cost so SDK calculates exact credits we charged
+        cost: creditsToFakeCost(billedCredits),
+        cost_details: { upstream_inference_cost: 0 },
       },
     }
   }
@@ -211,8 +221,8 @@ export async function handleOpenRouterNonStream({
     logger.error({ error }, 'Failed to insert message into BigQuery')
   })
 
-  // Consume credits
-  await consumeCreditsForMessage({
+  // Consume credits and get the actual billed amount
+  const billedCredits = await consumeCreditsForMessage({
     messageId: data.id,
     userId,
     stripeCustomerId,
@@ -228,6 +238,12 @@ export async function handleOpenRouterNonStream({
     logger,
   })
 
+  // Overwrite cost so SDK calculates exact credits we charged
+  if (data.usage) {
+    data.usage.cost = creditsToFakeCost(billedCredits)
+    data.usage.cost_details = { upstream_inference_cost: 0 }
+  }
+
   return data
 }
 
@@ -323,7 +339,7 @@ export async function handleOpenRouterStream({
             const line = buffer.slice(0, lineEnd + 1)
             buffer = buffer.slice(lineEnd + 1)
 
-            state = await handleLine({
+            const lineResult = await handleLine({
               userId,
               stripeCustomerId,
               agentId,
@@ -337,10 +353,15 @@ export async function handleOpenRouterStream({
               logger,
               insertMessage: insertMessageBigquery,
             })
+            state = lineResult.state
 
             if (!clientDisconnected) {
               try {
-                controller.enqueue(new TextEncoder().encode(line))
+                // Overwrite cost in final chunk so SDK calculates exact credits we charged
+                const lineToSend = lineResult.billedCredits !== undefined
+                  ? overwriteCostWithBilledCredits(line, lineResult.billedCredits)
+                  : line
+                controller.enqueue(new TextEncoder().encode(lineToSend))
               } catch (error) {
                 logger.warn(
                   'Client disconnected during stream, continuing for billing',
@@ -413,14 +434,14 @@ async function handleLine({
   state: StreamState
   logger: Logger
   insertMessage: InsertMessageBigqueryFn
-}): Promise<StreamState> {
+}): Promise<LineResult> {
   if (!line.startsWith('data: ')) {
-    return state
+    return { state }
   }
 
   const raw = line.slice('data: '.length)
   if (raw === '[DONE]\n') {
-    return state
+    return { state }
   }
 
   // Parse the string into an object
@@ -432,7 +453,7 @@ async function handleLine({
       { error: getErrorObject(error, { includeRawError: true }) },
       'Received non-JSON OpenRouter response',
     )
-    return state
+    return { state }
   }
 
   // Extract usage
@@ -442,10 +463,10 @@ async function handleLine({
       { error: getErrorObject(parsed.error, { includeRawError: true }) },
       'Unable to parse OpenRouter response',
     )
-    return state
+    return { state }
   }
 
-  return await handleResponse({
+  return handleResponse({
     userId,
     stripeCustomerId,
     agentId,
@@ -487,7 +508,7 @@ async function handleResponse({
   state: StreamState
   logger: Logger
   insertMessage: InsertMessageBigqueryFn
-}): Promise<StreamState> {
+}): Promise<LineResult> {
   const model = 'model' in data ? data.model : undefined
   state = await handleStreamChunk({
     data,
@@ -500,7 +521,7 @@ async function handleResponse({
 
   if ('error' in data || !data.usage) {
     // Stream not finished
-    return state
+    return { state }
   }
 
   const usageData = extractUsageAndCost(data.usage)
@@ -520,7 +541,8 @@ async function handleResponse({
     logger.error({ error }, 'Failed to insert message into BigQuery')
   })
 
-  await consumeCreditsForMessage({
+  // Consume credits and get the actual billed amount
+  const billedCredits = await consumeCreditsForMessage({
     messageId: data.id,
     userId,
     stripeCustomerId,
@@ -536,7 +558,7 @@ async function handleResponse({
     logger,
   })
 
-  return state
+  return { state, billedCredits }
 }
 
 async function handleStreamChunk({
@@ -727,3 +749,41 @@ async function parseOpenRouterError(
   }
   return new OpenRouterError(response.status, response.statusText, errorBody)
 }
+
+/**
+ * Convert credits (integer cents) back to a cost value that will result in the same
+ * credits when the SDK applies its formula: credits = Math.round(cost * (1 + PROFIT_MARGIN) * 100)
+ */
+function creditsToFakeCost(credits: number): number {
+  return credits / ((1 + PROFIT_MARGIN) * 100)
+}
+
+/**
+ * Overwrite the cost field in the final SSE chunk to reflect actual billed credits.
+ * This ensures the SDK calculates the exact credits value we stored in the database,
+ * making the server the single source of truth for credit tracking.
+ */
+function overwriteCostWithBilledCredits(line: string, billedCredits: number): string {
+  if (!line.startsWith('data: ')) {
+    return line
+  }
+
+  const raw = line.slice('data: '.length)
+  if (raw === '[DONE]\n' || raw === '[DONE]') {
+    return line
+  }
+
+  try {
+    const obj = JSON.parse(raw)
+    // Only modify if there's usage data (final chunk)
+    if (obj.usage) {
+      obj.usage.cost = creditsToFakeCost(billedCredits)
+      obj.usage.cost_details = { upstream_inference_cost: 0 }
+      return `data: ${JSON.stringify(obj)}\n`
+    }
+  } catch {
+    // If parsing fails, return original line
+  }
+
+  return line
+}