Fix storage growth from oversized conversations and recording artifacts

apps/desktop/src/main/builtin-tool-definitions.ts

@@ -240,6 +240,10 @@ export const builtinToolDefinitions: BuiltinToolDefinition[] = [
           type: "number",
           description: "Command timeout in milliseconds (default: 30000). Set to 0 for no timeout.",
         },
+        maxOutputChars: {
+          type: "number",
+          description: "Maximum characters to return per stream (stdout/stderr). Default 20000, minimum 1000, capped at 200000. Values below 1000 are raised to 1000.",
+        },
       },
       required: ["command"],
     },
@@ -445,4 +449,3 @@ export const builtinToolDefinitions: BuiltinToolDefinition[] = [
 export function getBuiltinToolNames(): string[] {
   return builtinToolDefinitions.map((tool) => tool.name)
 }
-

apps/desktop/src/main/builtin-tools.ts

@@ -27,6 +27,37 @@ import path from "path"
 import type { AgentMemory } from "../shared/types"
 
 const execAsync = promisify(exec)
+const DEFAULT_MAX_TOOL_OUTPUT_CHARS = 20_000
+const MAX_ALLOWED_TOOL_OUTPUT_CHARS = 200_000
+const DEFAULT_MAX_SKILL_INSTRUCTIONS_CHARS = 20_000
+
+function clampMaxOutputChars(value: unknown, fallback: number): number {
+  if (typeof value !== "number" || !Number.isFinite(value)) {
+    return fallback
+  }
+  const rounded = Math.floor(value)
+  return Math.min(Math.max(rounded, 1_000), MAX_ALLOWED_TOOL_OUTPUT_CHARS)
+}
+
+function truncateToolText(
+  input: string,
+  maxChars: number,
+): { text: string; truncated: boolean; originalLength: number } {
+  const originalLength = input.length
+  if (originalLength <= maxChars) {
+    return { text: input, truncated: false, originalLength }
+  }
+
+  const omitted = originalLength - maxChars
+  const footer = `\n\n[truncated ${omitted} characters]`
+  // Slice enough that the content + footer together don't exceed maxChars.
+  const contentChars = Math.max(0, maxChars - footer.length)
+  return {
+    text: `${input.slice(0, contentChars)}${footer}`,
+    truncated: true,
+    originalLength,
+  }
+}
 
 // Re-export from the dependency-free definitions module for backward compatibility
 // This breaks the circular dependency: profile-service -> builtin-tool-definitions (no cycle)
@@ -853,6 +884,7 @@ const toolHandlers: Record<string, ToolHandler> = {
     const timeout = (typeof rawTimeout === "number" && Number.isFinite(rawTimeout) && rawTimeout >= 0) 
       ? rawTimeout 
       : 30000
+    const maxOutputChars = clampMaxOutputChars(args.maxOutputChars, DEFAULT_MAX_TOOL_OUTPUT_CHARS)
 
     // Determine the working directory
     let cwd: string | undefined
@@ -909,6 +941,8 @@ const toolHandlers: Record<string, ToolHandler> = {
       }
 
       const { stdout, stderr } = await execAsync(command, execOptions)
+      const truncatedStdout = truncateToolText(stdout || "", maxOutputChars)
+      const truncatedStderr = truncateToolText(stderr || "", maxOutputChars)
 
       return {
         content: [
@@ -919,8 +953,12 @@ const toolHandlers: Record<string, ToolHandler> = {
               command,
               cwd: cwd || process.cwd(),
               skillName,
-              stdout: stdout || "",
-              stderr: stderr || "",
+              stdout: truncatedStdout.text,
+              stderr: truncatedStderr.text,
+              maxOutputChars,
+              stdoutOriginalLength: truncatedStdout.originalLength,
+              stderrOriginalLength: truncatedStderr.originalLength,
+              outputTruncated: truncatedStdout.truncated || truncatedStderr.truncated,
             }, null, 2),
           },
         ],
@@ -932,6 +970,8 @@ const toolHandlers: Record<string, ToolHandler> = {
       const stderr = error.stderr || ""
       const errorMessage = error.message || String(error)
       const exitCode = error.code
+      const truncatedStdout = truncateToolText(stdout, maxOutputChars)
+      const truncatedStderr = truncateToolText(stderr, maxOutputChars)
 
       return {
         content: [
@@ -944,8 +984,12 @@ const toolHandlers: Record<string, ToolHandler> = {
               skillName,
               error: errorMessage,
               exitCode,
-              stdout,
-              stderr,
+              stdout: truncatedStdout.text,
+              stderr: truncatedStderr.text,
+              maxOutputChars,
+              stdoutOriginalLength: truncatedStdout.originalLength,
+              stderrOriginalLength: truncatedStderr.originalLength,
+              outputTruncated: truncatedStdout.truncated || truncatedStderr.truncated,
             }, null, 2),
           },
         ],
@@ -1800,10 +1844,14 @@ const toolHandlers: Record<string, ToolHandler> = {
       const skillByName = allSkills.find(s => s.name.toLowerCase() === skillId.toLowerCase())
 
       if (skillByName) {
+        const truncated = truncateToolText(
+          `# ${skillByName.name}\n\n${skillByName.instructions}`,
+          DEFAULT_MAX_SKILL_INSTRUCTIONS_CHARS,
+        )
         return {
           content: [{
             type: "text",
-            text: `# ${skillByName.name}\n\n${skillByName.instructions}`,
+            text: truncated.text,
           }],
           isError: false,
         }
@@ -1821,10 +1869,15 @@ const toolHandlers: Record<string, ToolHandler> = {
       }
     }
 
+    const truncated = truncateToolText(
+      `# ${skill.name}\n\n${skill.instructions}`,
+      DEFAULT_MAX_SKILL_INSTRUCTIONS_CHARS,
+    )
+
     return {
       content: [{
         type: "text",
-        text: `# ${skill.name}\n\n${skill.instructions}`,
+        text: truncated.text,
       }],
       isError: false,
     }

apps/desktop/src/main/config.ts

@@ -83,6 +83,11 @@ const getConfig = () => {
     conversationsEnabled: true,
     maxConversationsToKeep: 100,
     autoSaveConversations: true,
+    recordingsCleanupEnabled: true,
+    recordingHistoryMaxItems: 2000,
+    recordingHistoryRetentionDays: 90,
+    recordingFileRetentionDays: 90,
+    recordingOrphanGracePeriodMinutes: 5,
     // Settings hotkey defaults
     settingsHotkeyEnabled: true,
     settingsHotkey: "ctrl-shift-s",

apps/desktop/src/main/context-budget.ts

@@ -622,36 +622,50 @@ export async function shrinkMessagesForLLM(opts: ShrinkOptions): Promise<ShrinkR
 
   let messages = [...opts.messages]
   let tokens = estimateTokensFromMessages(messages)
+  const initialTokens = tokens
 
   if (isDebugLLM()) {
     logLLM("ContextBudget: initial", { providerId, model, maxTokens, targetTokens, estTokens: tokens, count: messages.length })
   }
 
   if (tokens <= targetTokens) {
-    return { messages, appliedStrategies: applied, estTokensBefore: tokens, estTokensAfter: tokens, maxTokens }
+    return { messages, appliedStrategies: applied, estTokensBefore: initialTokens, estTokensAfter: tokens, maxTokens }
   }
 
   // Tier 0: Aggressive truncation of very large tool responses (>5000 chars)
   // This happens BEFORE summarization to avoid expensive LLM calls on huge payloads
   const AGGRESSIVE_TRUNCATE_THRESHOLD = 5000
+  const tokensBeforeAggressiveTruncate = tokens
   for (let i = 0; i < messages.length; i++) {
     const msg = messages[i]
-    if (msg.role === "user" && msg.content && msg.content.length > AGGRESSIVE_TRUNCATE_THRESHOLD) {
-      // Check if this looks like a tool result (contains JSON arrays/objects)
-      if (msg.content.includes('"url":') || msg.content.includes('"id":')) {
-        // Truncate aggressively and add note
-        messages[i] = {
-          ...msg,
-          content: msg.content.substring(0, AGGRESSIVE_TRUNCATE_THRESHOLD) +
-                   '\n\n... (truncated ' + (msg.content.length - AGGRESSIVE_TRUNCATE_THRESHOLD) +
-                   ' characters for context management. Key information preserved above.)'
-        }
-        applied.push("aggressive_truncate")
-        tokens = estimateTokensFromMessages(messages)
-        if (tokens <= targetTokens) {
-          if (isDebugLLM()) logLLM("ContextBudget: after aggressive_truncate", { estTokens: tokens })
-          return { messages, appliedStrategies: applied, estTokensBefore: tokens, estTokensAfter: tokens, maxTokens }
-        }
+    if (!msg.content || msg.content.length <= AGGRESSIVE_TRUNCATE_THRESHOLD) {
+      continue
+    }
+
+    // Heuristic: aggressively truncate oversized tool payloads first.
+    // JSON-key heuristics are only applied to non-user messages to avoid
+    // accidentally truncating user prompts that happen to contain JSON.
+    const looksLikeToolPayload =
+      msg.role === "tool" ||
+      (msg.role !== "user" && (
+        msg.content.includes('"stdout"') ||
+        msg.content.includes('"stderr"') ||
+        msg.content.includes('"url":') ||
+        msg.content.includes('"id":')
+      ))
+
+    if (looksLikeToolPayload) {
+      messages[i] = {
+        ...msg,
+        content: msg.content.substring(0, AGGRESSIVE_TRUNCATE_THRESHOLD) +
+                 '\n\n... (truncated ' + (msg.content.length - AGGRESSIVE_TRUNCATE_THRESHOLD) +
+                 ' characters for context management. Key information preserved above.)'
+      }
+      applied.push("aggressive_truncate")
+      tokens = estimateTokensFromMessages(messages)
+      if (tokens <= targetTokens) {
+        if (isDebugLLM()) logLLM("ContextBudget: after aggressive_truncate", { estTokens: tokens })
+        return { messages, appliedStrategies: applied, estTokensBefore: tokensBeforeAggressiveTruncate, estTokensAfter: tokens, maxTokens }
       }
     }
   }
@@ -692,7 +706,7 @@ export async function shrinkMessagesForLLM(opts: ShrinkOptions): Promise<ShrinkR
 
   if (tokens <= targetTokens) {
     if (isDebugLLM()) logLLM("ContextBudget: after summarize", { estTokens: tokens })
-    return { messages, appliedStrategies: applied, estTokensBefore: tokens, estTokensAfter: tokens, maxTokens }
+    return { messages, appliedStrategies: applied, estTokensBefore: initialTokens, estTokensAfter: tokens, maxTokens }
   }
 
   // Tier 2: Remove middle messages (keep system, first user, last N)
@@ -760,7 +774,7 @@ export async function shrinkMessagesForLLM(opts: ShrinkOptions): Promise<ShrinkR
 
   if (tokens <= targetTokens) {
     if (isDebugLLM()) logLLM("ContextBudget: after drop_middle", { estTokens: tokens, kept: messages.length })
-    return { messages, appliedStrategies: applied, estTokensBefore: tokens, estTokensAfter: tokens, maxTokens, toolResultsSummarized }
+    return { messages, appliedStrategies: applied, estTokensBefore: initialTokens, estTokensAfter: tokens, maxTokens, toolResultsSummarized }
   }
 
   // Tier 3: Minimal system prompt
@@ -780,6 +794,5 @@ export async function shrinkMessagesForLLM(opts: ShrinkOptions): Promise<ShrinkR
 
   if (isDebugLLM()) logLLM("ContextBudget: after minimal_system_prompt", { estTokens: tokens })
 
-  return { messages, appliedStrategies: applied, estTokensBefore: tokens, estTokensAfter: tokens, maxTokens, toolResultsSummarized }
+  return { messages, appliedStrategies: applied, estTokensBefore: initialTokens, estTokensAfter: tokens, maxTokens, toolResultsSummarized }
 }
-