@xelauvas/xela-cli 0.1.0

package/src/utils/toolResultStorage.ts

@@ -0,0 +1,1040 @@
+/**
+ * Utility for persisting large tool results to disk instead of truncating them.
+ */
+
+import type { ToolResultBlockParam } from '@anthropic-ai/sdk/resources/index.mjs'
+import { mkdir, writeFile } from 'fs/promises'
+import { join } from 'path'
+import { getOriginalCwd, getSessionId } from '../bootstrap/state.js'
+import {
+  BYTES_PER_TOKEN,
+  DEFAULT_MAX_RESULT_SIZE_CHARS,
+  MAX_TOOL_RESULT_BYTES,
+  MAX_TOOL_RESULTS_PER_MESSAGE_CHARS,
+} from '../constants/toolLimits.js'
+import { getFeatureValue_CACHED_MAY_BE_STALE } from '../services/analytics/growthbook.js'
+import { logEvent } from '../services/analytics/index.js'
+import { sanitizeToolNameForAnalytics } from '../services/analytics/metadata.js'
+import type { Message } from '../types/message.js'
+import { logForDebugging } from './debug.js'
+import { getErrnoCode, toError } from './errors.js'
+import { formatFileSize } from './format.js'
+import { logError } from './log.js'
+import { getProjectDir } from './sessionStorage.js'
+import { jsonStringify } from './slowOperations.js'
+
+// Subdirectory name for tool results within a session
+export const TOOL_RESULTS_SUBDIR = 'tool-results'
+
+// XML tag used to wrap persisted output messages
+export const PERSISTED_OUTPUT_TAG = '<persisted-output>'
+export const PERSISTED_OUTPUT_CLOSING_TAG = '</persisted-output>'
+
+// Message used when tool result content was cleared without persisting to file
+export const TOOL_RESULT_CLEARED_MESSAGE = '[Old tool result content cleared]'
+
+/**
+ * GrowthBook override map: tool name -> persistence threshold (chars).
+ * When a tool name is present in this map, that value is used directly as the
+ * effective threshold, bypassing the Math.min() clamp against the 50k default.
+ * Tools absent from the map use the hardcoded fallback.
+ * Flag default is {} (no overrides == behavior unchanged).
+ */
+const PERSIST_THRESHOLD_OVERRIDE_FLAG = 'tengu_satin_quoll'
+
+/**
+ * Resolve the effective persistence threshold for a tool.
+ * GrowthBook override wins when present; otherwise falls back to the declared
+ * per-tool cap clamped by the global default.
+ *
+ * Defensive: GrowthBook's cache returns `cached !== undefined ? cached : default`,
+ * so a flag served as `null` leaks through. We guard with optional chaining and a
+ * typeof check so any non-object flag value (null, string, number) falls through
+ * to the hardcoded default instead of throwing on index or returning 0.
+ */
+export function getPersistenceThreshold(
+  toolName: string,
+  declaredMaxResultSizeChars: number,
+): number {
+  // Infinity = hard opt-out. Read self-bounds via maxTokens; persisting its
+  // output to a file the model reads back with Read is circular. Checked
+  // before the GB override so tengu_satin_quoll can't force it back on.
+  if (!Number.isFinite(declaredMaxResultSizeChars)) {
+    return declaredMaxResultSizeChars
+  }
+  const overrides = getFeatureValue_CACHED_MAY_BE_STALE<Record<
+    string,
+    number
+  > | null>(PERSIST_THRESHOLD_OVERRIDE_FLAG, {})
+  const override = overrides?.[toolName]
+  if (
+    typeof override === 'number' &&
+    Number.isFinite(override) &&
+    override > 0
+  ) {
+    return override
+  }
+  return Math.min(declaredMaxResultSizeChars, DEFAULT_MAX_RESULT_SIZE_CHARS)
+}
+
+// Result of persisting a tool result to disk
+export type PersistedToolResult = {
+  filepath: string
+  originalSize: number
+  isJson: boolean
+  preview: string
+  hasMore: boolean
+}
+
+// Error result when persistence fails
+export type PersistToolResultError = {
+  error: string
+}
+
+/**
+ * Get the session directory (projectDir/sessionId)
+ */
+function getSessionDir(): string {
+  return join(getProjectDir(getOriginalCwd()), getSessionId())
+}
+
+/**
+ * Get the tool results directory for this session (projectDir/sessionId/tool-results)
+ */
+export function getToolResultsDir(): string {
+  return join(getSessionDir(), TOOL_RESULTS_SUBDIR)
+}
+
+// Preview size in bytes for the reference message
+export const PREVIEW_SIZE_BYTES = 2000
+
+/**
+ * Get the filepath where a tool result would be persisted.
+ */
+export function getToolResultPath(id: string, isJson: boolean): string {
+  const ext = isJson ? 'json' : 'txt'
+  return join(getToolResultsDir(), `${id}.${ext}`)
+}
+
+/**
+ * Ensure the session-specific tool results directory exists
+ */
+export async function ensureToolResultsDir(): Promise<void> {
+  try {
+    await mkdir(getToolResultsDir(), { recursive: true })
+  } catch {
+    // Directory may already exist
+  }
+}
+
+/**
+ * Persist a tool result to disk and return information about the persisted file
+ *
+ * @param content - The tool result content to persist (string or array of content blocks)
+ * @param toolUseId - The ID of the tool use that produced the result
+ * @returns Information about the persisted file including filepath and preview
+ */
+export async function persistToolResult(
+  content: NonNullable<ToolResultBlockParam['content']>,
+  toolUseId: string,
+): Promise<PersistedToolResult | PersistToolResultError> {
+  const isJson = Array.isArray(content)
+
+  // Check for non-text content - we can only persist text blocks
+  if (isJson) {
+    const hasNonTextContent = content.some(block => block.type !== 'text')
+    if (hasNonTextContent) {
+      return {
+        error: 'Cannot persist tool results containing non-text content',
+      }
+    }
+  }
+
+  await ensureToolResultsDir()
+  const filepath = getToolResultPath(toolUseId, isJson)
+  const contentStr = isJson ? jsonStringify(content, null, 2) : content
+
+  // tool_use_id is unique per invocation and content is deterministic for a
+  // given id, so skip if the file already exists. This prevents re-writing
+  // the same content on every API turn when microcompact replays the
+  // original messages. Use 'wx' instead of a stat-then-write race.
+  try {
+    await writeFile(filepath, contentStr, { encoding: 'utf-8', flag: 'wx' })
+    logForDebugging(
+      `Persisted tool result to ${filepath} (${formatFileSize(contentStr.length)})`,
+    )
+  } catch (error) {
+    if (getErrnoCode(error) !== 'EEXIST') {
+      logError(toError(error))
+      return { error: getFileSystemErrorMessage(toError(error)) }
+    }
+    // EEXIST: already persisted on a prior turn, fall through to preview
+  }
+
+  // Generate a preview
+  const { preview, hasMore } = generatePreview(contentStr, PREVIEW_SIZE_BYTES)
+
+  return {
+    filepath,
+    originalSize: contentStr.length,
+    isJson,
+    preview,
+    hasMore,
+  }
+}
+
+/**
+ * Build a message for large tool results with preview
+ */
+export function buildLargeToolResultMessage(
+  result: PersistedToolResult,
+): string {
+  let message = `${PERSISTED_OUTPUT_TAG}\n`
+  message += `Output too large (${formatFileSize(result.originalSize)}). Full output saved to: ${result.filepath}\n\n`
+  message += `Preview (first ${formatFileSize(PREVIEW_SIZE_BYTES)}):\n`
+  message += result.preview
+  message += result.hasMore ? '\n...\n' : '\n'
+  message += PERSISTED_OUTPUT_CLOSING_TAG
+  return message
+}
+
+/**
+ * Process a tool result for inclusion in a message.
+ * Maps the result to the API format and persists large results to disk.
+ */
+export async function processToolResultBlock<T>(
+  tool: {
+    name: string
+    maxResultSizeChars: number
+    mapToolResultToToolResultBlockParam: (
+      result: T,
+      toolUseID: string,
+    ) => ToolResultBlockParam
+  },
+  toolUseResult: T,
+  toolUseID: string,
+): Promise<ToolResultBlockParam> {
+  const toolResultBlock = tool.mapToolResultToToolResultBlockParam(
+    toolUseResult,
+    toolUseID,
+  )
+  return maybePersistLargeToolResult(
+    toolResultBlock,
+    tool.name,
+    getPersistenceThreshold(tool.name, tool.maxResultSizeChars),
+  )
+}
+
+/**
+ * Process a pre-mapped tool result block. Applies persistence for large results
+ * without re-calling mapToolResultToToolResultBlockParam.
+ */
+export async function processPreMappedToolResultBlock(
+  toolResultBlock: ToolResultBlockParam,
+  toolName: string,
+  maxResultSizeChars: number,
+): Promise<ToolResultBlockParam> {
+  return maybePersistLargeToolResult(
+    toolResultBlock,
+    toolName,
+    getPersistenceThreshold(toolName, maxResultSizeChars),
+  )
+}
+
+/**
+ * True when a tool_result's content is empty or effectively empty. Covers:
+ * undefined/null/'', whitespace-only strings, empty arrays, and arrays whose
+ * only blocks are text blocks with empty/whitespace text. Non-text blocks
+ * (images, tool_reference) are treated as non-empty.
+ */
+export function isToolResultContentEmpty(
+  content: ToolResultBlockParam['content'],
+): boolean {
+  if (!content) return true
+  if (typeof content === 'string') return content.trim() === ''
+  if (!Array.isArray(content)) return false
+  if (content.length === 0) return true
+  return content.every(
+    block =>
+      typeof block === 'object' &&
+      'type' in block &&
+      block.type === 'text' &&
+      'text' in block &&
+      (typeof block.text !== 'string' || block.text.trim() === ''),
+  )
+}
+
+/**
+ * Handle large tool results by persisting to disk instead of truncating.
+ * Returns the original block if no persistence needed, or a modified block
+ * with the content replaced by a reference to the persisted file.
+ */
+async function maybePersistLargeToolResult(
+  toolResultBlock: ToolResultBlockParam,
+  toolName: string,
+  persistenceThreshold?: number,
+): Promise<ToolResultBlockParam> {
+  // Check size first before doing any async work - most tool results are small
+  const content = toolResultBlock.content
+
+  // inc-4586: Empty tool_result content at the prompt tail causes some models
+  // (notably capybara) to emit the \n\nHuman: stop sequence and end their turn
+  // with zero output. The server renderer inserts no \n\nAssistant: marker after
+  // tool results, so a bare </function_results>\n\n pattern-matches to a turn
+  // boundary. Several tools can legitimately produce empty output (silent-success
+  // shell commands, MCP servers returning content:[], REPL statements, etc.).
+  // Inject a short marker so the model always has something to react to.
+  if (isToolResultContentEmpty(content)) {
+    logEvent('tengu_tool_empty_result', {
+      toolName: sanitizeToolNameForAnalytics(toolName),
+    })
+    return {
+      ...toolResultBlock,
+      content: `(${toolName} completed with no output)`,
+    }
+  }
+  // Narrow after the emptiness guard — content is non-nullish past this point.
+  if (!content) {
+    return toolResultBlock
+  }
+
+  // Skip persistence for image content blocks - they need to be sent as-is to Claude
+  if (hasImageBlock(content)) {
+    return toolResultBlock
+  }
+
+  const size = contentSize(content)
+
+  // Use tool-specific threshold if provided, otherwise fall back to global limit
+  const threshold = persistenceThreshold ?? MAX_TOOL_RESULT_BYTES
+  if (size <= threshold) {
+    return toolResultBlock
+  }
+
+  // Persist the entire content as a unit
+  const result = await persistToolResult(content, toolResultBlock.tool_use_id)
+  if (isPersistError(result)) {
+    // If persistence failed, return the original block unchanged
+    return toolResultBlock
+  }
+
+  const message = buildLargeToolResultMessage(result)
+
+  // Log analytics
+  logEvent('tengu_tool_result_persisted', {
+    toolName: sanitizeToolNameForAnalytics(toolName),
+    originalSizeBytes: result.originalSize,
+    persistedSizeBytes: message.length,
+    estimatedOriginalTokens: Math.ceil(result.originalSize / BYTES_PER_TOKEN),
+    estimatedPersistedTokens: Math.ceil(message.length / BYTES_PER_TOKEN),
+    thresholdUsed: threshold,
+  })
+
+  return { ...toolResultBlock, content: message }
+}
+
+/**
+ * Generate a preview of content, truncating at a newline boundary when possible.
+ */
+export function generatePreview(
+  content: string,
+  maxBytes: number,
+): { preview: string; hasMore: boolean } {
+  if (content.length <= maxBytes) {
+    return { preview: content, hasMore: false }
+  }
+
+  // Find the last newline within the limit to avoid cutting mid-line
+  const truncated = content.slice(0, maxBytes)
+  const lastNewline = truncated.lastIndexOf('\n')
+
+  // If we found a newline reasonably close to the limit, use it
+  // Otherwise fall back to the exact limit
+  const cutPoint = lastNewline > maxBytes * 0.5 ? lastNewline : maxBytes
+
+  return { preview: content.slice(0, cutPoint), hasMore: true }
+}
+
+/**
+ * Type guard to check if persist result is an error
+ */
+export function isPersistError(
+  result: PersistedToolResult | PersistToolResultError,
+): result is PersistToolResultError {
+  return 'error' in result
+}
+
+// --- Message-level aggregate tool result budget ---
+//
+// Tracks replacement state across turns so enforceToolResultBudget makes the
+// same choices every time (preserves prompt cache prefix).
+
+/**
+ * Per-conversation-thread state for the aggregate tool result budget.
+ * State must be stable to preserve prompt cache:
+ *   - seenIds: results that have passed through the budget check (replaced
+ *     or not). Once seen, a result's fate is frozen for the conversation.
+ *   - replacements: subset of seenIds that were persisted to disk and
+ *     replaced with previews, mapped to the exact preview string shown to
+ *     the model. Re-application is a Map lookup — no file I/O, guaranteed
+ *     byte-identical, cannot fail.
+ *
+ * Lifecycle: one instance per conversation thread, carried on ToolUseContext.
+ * Main thread: REPL provisions once, never resets — stale entries after
+ * /clear, rewind, resume, or compact are never looked up (tool_use_ids are
+ * UUIDs) so they're harmless. Subagents: createSubagentContext clones the
+ * parent's state by default (cache-sharing forks like agentSummary need
+ * identical decisions), or resumeAgentBackground threads one reconstructed
+ * from sidechain records.
+ */
+export type ContentReplacementState = {
+  seenIds: Set<string>
+  replacements: Map<string, string>
+}
+
+export function createContentReplacementState(): ContentReplacementState {
+  return { seenIds: new Set(), replacements: new Map() }
+}
+
+/**
+ * Clone replacement state for a cache-sharing fork (e.g. agentSummary).
+ * The fork needs state identical to the source at fork time so
+ * enforceToolResultBudget makes the same choices → same wire prefix →
+ * prompt cache hit. Mutating the clone does not affect the source.
+ */
+export function cloneContentReplacementState(
+  source: ContentReplacementState,
+): ContentReplacementState {
+  return {
+    seenIds: new Set(source.seenIds),
+    replacements: new Map(source.replacements),
+  }
+}
+
+/**
+ * Resolve the per-message aggregate budget limit. GrowthBook override
+ * (tengu_hawthorn_window) wins when present and a finite positive number;
+ * otherwise falls back to the hardcoded constant. Defensive typeof/finite
+ * check: GrowthBook's cache returns `cached !== undefined ? cached : default`,
+ * so a flag served as null/string/NaN leaks through.
+ */
+export function getPerMessageBudgetLimit(): number {
+  const override = getFeatureValue_CACHED_MAY_BE_STALE<number | null>(
+    'tengu_hawthorn_window',
+    null,
+  )
+  if (
+    typeof override === 'number' &&
+    Number.isFinite(override) &&
+    override > 0
+  ) {
+    return override
+  }
+  return MAX_TOOL_RESULTS_PER_MESSAGE_CHARS
+}
+
+/**
+ * Provision replacement state for a new conversation thread.
+ *
+ * Encapsulates the feature-flag gate + reconstruct-vs-fresh choice:
+ *   - Flag off → undefined (query.ts skips enforcement entirely)
+ *   - No initialMessages (cold start) → fresh
+ *   - initialMessages present → reconstruct (freeze all candidate IDs so the
+ *     budget never replaces content the model already saw unreplaced). Empty
+ *     or absent records freeze everything; non-empty records additionally
+ *     populate the replacements Map for byte-identical re-apply.
+ */
+export function provisionContentReplacementState(
+  initialMessages?: Message[],
+  initialContentReplacements?: ContentReplacementRecord[],
+): ContentReplacementState | undefined {
+  const enabled = getFeatureValue_CACHED_MAY_BE_STALE(
+    'tengu_hawthorn_steeple',
+    false,
+  )
+  if (!enabled) return undefined
+  if (initialMessages) {
+    return reconstructContentReplacementState(
+      initialMessages,
+      initialContentReplacements ?? [],
+    )
+  }
+  return createContentReplacementState()
+}
+
+/**
+ * Serializable record of one content-replacement decision. Written to the
+ * transcript as a ContentReplacementEntry so decisions survive resume.
+ * Discriminated by `kind` so future replacement mechanisms (user text,
+ * offloaded images) can share the same transcript entry type.
+ *
+ * `replacement` is the exact string the model saw — stored rather than
+ * derived on resume so code changes to the preview template, size formatting,
+ * or path layout can't silently break prompt cache.
+ */
+export type ContentReplacementRecord = {
+  kind: 'tool-result'
+  toolUseId: string
+  replacement: string
+}
+
+export type ToolResultReplacementRecord = Extract<
+  ContentReplacementRecord,
+  { kind: 'tool-result' }
+>
+
+type ToolResultCandidate = {
+  toolUseId: string
+  content: NonNullable<ToolResultBlockParam['content']>
+  size: number
+}
+
+type CandidatePartition = {
+  mustReapply: Array<ToolResultCandidate & { replacement: string }>
+  frozen: ToolResultCandidate[]
+  fresh: ToolResultCandidate[]
+}
+
+function isContentAlreadyCompacted(
+  content: ToolResultBlockParam['content'],
+): boolean {
+  // All budget-produced content starts with the tag (buildLargeToolResultMessage).
+  // `.startsWith()` avoids false-positives when the tag appears anywhere else
+  // in the content (e.g., reading this source file).
+  return typeof content === 'string' && content.startsWith(PERSISTED_OUTPUT_TAG)
+}
+
+function hasImageBlock(
+  content: NonNullable<ToolResultBlockParam['content']>,
+): boolean {
+  return (
+    Array.isArray(content) &&
+    content.some(
+      b => typeof b === 'object' && 'type' in b && b.type === 'image',
+    )
+  )
+}
+
+function contentSize(
+  content: NonNullable<ToolResultBlockParam['content']>,
+): number {
+  if (typeof content === 'string') return content.length
+  // Sum text-block lengths directly. Slightly under-counts vs serialized
+  // (no JSON framing), but the budget is a rough token heuristic anyway.
+  // Avoids allocating a content-sized string every enforcement pass.
+  return content.reduce(
+    (sum, b) => sum + (b.type === 'text' ? b.text.length : 0),
+    0,
+  )
+}
+
+/**
+ * Walk messages and build tool_use_id → tool_name from assistant tool_use
+ * blocks. tool_use always precedes its tool_result (model calls, then result
+ * arrives), so by the time budget enforcement sees a result, its name is known.
+ */
+function buildToolNameMap(messages: Message[]): Map<string, string> {
+  const map = new Map<string, string>()
+  for (const message of messages) {
+    if (message.type !== 'assistant') continue
+    const content = message.message.content
+    if (!Array.isArray(content)) continue
+    for (const block of content) {
+      if (block.type === 'tool_use') {
+        map.set(block.id, block.name)
+      }
+    }
+  }
+  return map
+}
+
+/**
+ * Extract candidate tool_result blocks from a single user message: blocks
+ * that are non-empty, non-image, and not already compacted by tag (i.e. by
+ * the per-tool limit, or an earlier iteration of this same query call).
+ * Returns [] for messages with no eligible blocks.
+ */
+function collectCandidatesFromMessage(message: Message): ToolResultCandidate[] {
+  if (message.type !== 'user' || !Array.isArray(message.message.content)) {
+    return []
+  }
+  return message.message.content.flatMap(block => {
+    if (block.type !== 'tool_result' || !block.content) return []
+    if (isContentAlreadyCompacted(block.content)) return []
+    if (hasImageBlock(block.content)) return []
+    return [
+      {
+        toolUseId: block.tool_use_id,
+        content: block.content,
+        size: contentSize(block.content),
+      },
+    ]
+  })
+}
+
+/**
+ * Extract candidate tool_result blocks grouped by API-level user message.
+ *
+ * normalizeMessagesForAPI merges consecutive user messages into one
+ * (Bedrock compat; 1P does the same server-side), so parallel tool
+ * results that arrive as N separate user messages in our state become
+ * ONE user message on the wire. The budget must group the same way or
+ * it would see N under-budget messages instead of one over-budget
+ * message and fail to enforce exactly when it matters most.
+ *
+ * A "group" is a maximal run of user messages NOT separated by an
+ * assistant message. Only assistant messages create wire-level
+ * boundaries — normalizeMessagesForAPI filters out progress entirely
+ * and merges attachment / system(local_command) INTO adjacent user
+ * blocks, so those types do NOT break groups here either.
+ *
+ * This matters for abort-during-parallel-tools paths: agent_progress
+ * messages (non-ephemeral, persisted in REPL state) can interleave
+ * between fresh tool_result messages. If we flushed on progress, those
+ * tool_results would split into under-budget groups, slip through
+ * unreplaced, get frozen, then be merged by normalizeMessagesForAPI
+ * into one over-budget wire message — defeating the feature.
+ *
+ * Only groups with at least one eligible candidate are returned.
+ */
+function collectCandidatesByMessage(
+  messages: Message[],
+): ToolResultCandidate[][] {
+  const groups: ToolResultCandidate[][] = []
+  let current: ToolResultCandidate[] = []
+
+  const flush = () => {
+    if (current.length > 0) groups.push(current)
+    current = []
+  }
+
+  // Track all assistant message.ids seen so far — same-ID fragments are
+  // merged by normalizeMessagesForAPI (messages.ts ~2126 walks back PAST
+  // different-ID assistants via `continue`), so any re-appearance of a
+  // previously-seen ID must NOT create a group boundary. Two scenarios:
+  //   • Consecutive: streamingToolExecution yields one AssistantMessage per
+  //     content_block_stop (same id); a fast tool drains between blocks;
+  //     abort/hook-stop leaves [asst(X), user(trA), asst(X), user(trB)].
+  //   • Interleaved: coordinator/teammate streams mix different responses
+  //     so [asst(X), user(trA), asst(Y), user(trB), asst(X), user(trC)].
+  // In both, normalizeMessagesForAPI merges the X fragments into one wire
+  // assistant, and their following tool_results merge into one wire user
+  // message — so the budget must see them as one group too.
+  const seenAsstIds = new Set<string>()
+  for (const message of messages) {
+    if (message.type === 'user') {
+      current.push(...collectCandidatesFromMessage(message))
+    } else if (message.type === 'assistant') {
+      if (!seenAsstIds.has(message.message.id)) {
+        flush()
+        seenAsstIds.add(message.message.id)
+      }
+    }
+    // progress / attachment / system are filtered or merged by
+    // normalizeMessagesForAPI — they don't create wire boundaries.
+  }
+  flush()
+
+  return groups
+}
+
+/**
+ * Partition candidates by their prior decision state:
+ *  - mustReapply: previously replaced → re-apply the cached replacement for
+ *    prefix stability
+ *  - frozen: previously seen and left unreplaced → off-limits (replacing
+ *    now would change a prefix that was already cached)
+ *  - fresh: never seen → eligible for new replacement decisions
+ */
+function partitionByPriorDecision(
+  candidates: ToolResultCandidate[],
+  state: ContentReplacementState,
+): CandidatePartition {
+  return candidates.reduce<CandidatePartition>(
+    (acc, c) => {
+      const replacement = state.replacements.get(c.toolUseId)
+      if (replacement !== undefined) {
+        acc.mustReapply.push({ ...c, replacement })
+      } else if (state.seenIds.has(c.toolUseId)) {
+        acc.frozen.push(c)
+      } else {
+        acc.fresh.push(c)
+      }
+      return acc
+    },
+    { mustReapply: [], frozen: [], fresh: [] },
+  )
+}
+
+/**
+ * Pick the largest fresh results to replace until the model-visible total
+ * (frozen + remaining fresh) is at or under budget, or fresh is exhausted.
+ * If frozen results alone exceed budget we accept the overage — microcompact
+ * will eventually clear them.
+ */
+function selectFreshToReplace(
+  fresh: ToolResultCandidate[],
+  frozenSize: number,
+  limit: number,
+): ToolResultCandidate[] {
+  const sorted = [...fresh].sort((a, b) => b.size - a.size)
+  const selected: ToolResultCandidate[] = []
+  let remaining = frozenSize + fresh.reduce((sum, c) => sum + c.size, 0)
+  for (const c of sorted) {
+    if (remaining <= limit) break
+    selected.push(c)
+    // We don't know the replacement size until after persist, but previews
+    // are ~2K and results hitting this path are much larger, so subtracting
+    // the full size is a close approximation for selection purposes.
+    remaining -= c.size
+  }
+  return selected
+}
+
+/**
+ * Return a new Message[] where each tool_result block whose id appears in
+ * replacementMap has its content replaced. Messages and blocks with no
+ * replacements are passed through by reference.
+ */
+function replaceToolResultContents(
+  messages: Message[],
+  replacementMap: Map<string, string>,
+): Message[] {
+  return messages.map(message => {
+    if (message.type !== 'user' || !Array.isArray(message.message.content)) {
+      return message
+    }
+    const content = message.message.content
+    const needsReplace = content.some(
+      b => b.type === 'tool_result' && replacementMap.has(b.tool_use_id),
+    )
+    if (!needsReplace) return message
+    return {
+      ...message,
+      message: {
+        ...message.message,
+        content: content.map(block => {
+          if (block.type !== 'tool_result') return block
+          const replacement = replacementMap.get(block.tool_use_id)
+          return replacement === undefined
+            ? block
+            : { ...block, content: replacement }
+        }),
+      },
+    }
+  })
+}
+
+async function buildReplacement(
+  candidate: ToolResultCandidate,
+): Promise<{ content: string; originalSize: number } | null> {
+  const result = await persistToolResult(candidate.content, candidate.toolUseId)
+  if (isPersistError(result)) return null
+  return {
+    content: buildLargeToolResultMessage(result),
+    originalSize: result.originalSize,
+  }
+}
+
+/**
+ * Enforce the per-message budget on aggregate tool result size.
+ *
+ * For each user message whose tool_result blocks together exceed the
+ * per-message limit (see getPerMessageBudgetLimit), the largest FRESH
+ * (never-before-seen) results in THAT message are persisted to disk and
+ * replaced with previews.
+ * Messages are evaluated independently — a 150K result in one message and
+ * a 150K result in another are both under budget and untouched.
+ *
+ * State is tracked by tool_use_id in `state`. Once a result is seen its
+ * fate is frozen: previously-replaced results get the same replacement
+ * re-applied every turn from the cached preview string (zero I/O,
+ * byte-identical), and previously-unreplaced results are never replaced
+ * later (would break prompt cache).
+ *
+ * Each turn adds at most one new user message with tool_result blocks,
+ * so the per-message loop typically does the budget check at most once;
+ * all prior messages just re-apply cached replacements.
+ *
+ * @param state — MUTATED: seenIds and replacements are updated in place
+ *   to record choices made this call. The caller holds a stable reference
+ *   across turns; returning a new object would require error-prone ref
+ *   updates after every query.
+ *
+ * Returns `{ messages, newlyReplaced }`:
+ *   - messages: same array instance when no replacement is needed
+ *   - newlyReplaced: replacements made THIS call (not re-applies).
+ *     Caller persists these to the transcript for resume reconstruction.
+ */
+export async function enforceToolResultBudget(
+  messages: Message[],
+  state: ContentReplacementState,
+  skipToolNames: ReadonlySet<string> = new Set(),
+): Promise<{
+  messages: Message[]
+  newlyReplaced: ToolResultReplacementRecord[]
+}> {
+  const candidatesByMessage = collectCandidatesByMessage(messages)
+  const nameByToolUseId =
+    skipToolNames.size > 0 ? buildToolNameMap(messages) : undefined
+  const shouldSkip = (id: string): boolean =>
+    nameByToolUseId !== undefined &&
+    skipToolNames.has(nameByToolUseId.get(id) ?? '')
+  // Resolve once per call. A mid-session flag change only affects FRESH
+  // messages (prior decisions are frozen via seenIds/replacements), so
+  // prompt cache for already-seen content is preserved regardless.
+  const limit = getPerMessageBudgetLimit()
+
+  // Walk each API-level message group independently. For previously-processed messages
+  // (all IDs in seenIds) this just re-applies cached replacements. For the
+  // single new message this turn added, it runs the budget check.
+  const replacementMap = new Map<string, string>()
+  const toPersist: ToolResultCandidate[] = []
+  let reappliedCount = 0
+  let messagesOverBudget = 0
+
+  for (const candidates of candidatesByMessage) {
+    const { mustReapply, frozen, fresh } = partitionByPriorDecision(
+      candidates,
+      state,
+    )
+
+    // Re-apply: pure Map lookups. No file I/O, byte-identical, cannot fail.
+    mustReapply.forEach(c => replacementMap.set(c.toolUseId, c.replacement))
+    reappliedCount += mustReapply.length
+
+    // Fresh means this is a new message. Check its per-message budget.
+    // (A previously-processed message has fresh.length === 0 because all
+    // its IDs were added to seenIds when first seen.)
+    if (fresh.length === 0) {
+      // mustReapply/frozen are already in seenIds from their first pass —
+      // re-adding is a no-op but keeps the invariant explicit.
+      candidates.forEach(c => state.seenIds.add(c.toolUseId))
+      continue
+    }
+
+    // Tools with maxResultSizeChars: Infinity (Read) — never persist.
+    // Mark as seen (frozen) so the decision sticks across turns. They don't
+    // count toward freshSize; if that lets the group slip under budget and
+    // the wire message is still large, that's the contract — Read's own
+    // maxTokens is the bound, not this wrapper.
+    const skipped = fresh.filter(c => shouldSkip(c.toolUseId))
+    skipped.forEach(c => state.seenIds.add(c.toolUseId))
+    const eligible = fresh.filter(c => !shouldSkip(c.toolUseId))
+
+    const frozenSize = frozen.reduce((sum, c) => sum + c.size, 0)
+    const freshSize = eligible.reduce((sum, c) => sum + c.size, 0)
+
+    const selected =
+      frozenSize + freshSize > limit
+        ? selectFreshToReplace(eligible, frozenSize, limit)
+        : []
+
+    // Mark non-persisting candidates as seen NOW (synchronously). IDs
+    // selected for persist are marked seen AFTER the await, alongside
+    // replacements.set — keeps the pair atomic under observation so no
+    // concurrent reader (once subagents share state) ever sees X∈seenIds
+    // but X∉replacements, which would misclassify X as frozen and send
+    // full content while the main thread sends the preview → cache miss.
+    const selectedIds = new Set(selected.map(c => c.toolUseId))
+    candidates
+      .filter(c => !selectedIds.has(c.toolUseId))
+      .forEach(c => state.seenIds.add(c.toolUseId))
+
+    if (selected.length === 0) continue
+    messagesOverBudget++
+    toPersist.push(...selected)
+  }
+
+  if (replacementMap.size === 0 && toPersist.length === 0) {
+    return { messages, newlyReplaced: [] }
+  }
+
+  // Fresh: concurrent persist for all selected candidates across all
+  // messages. In practice toPersist comes from a single message per turn.
+  const freshReplacements = await Promise.all(
+    toPersist.map(async c => [c, await buildReplacement(c)] as const),
+  )
+  const newlyReplaced: ToolResultReplacementRecord[] = []
+  let replacedSize = 0
+  for (const [candidate, replacement] of freshReplacements) {
+    // Mark seen HERE, post-await, atomically with replacements.set for
+    // success cases. For persist failures (replacement === null) the ID
+    // is seen-but-unreplaced — the original content was sent to the
+    // model, so treating it as frozen going forward is correct.
+    state.seenIds.add(candidate.toolUseId)
+    if (replacement === null) continue
+    replacedSize += candidate.size
+    replacementMap.set(candidate.toolUseId, replacement.content)
+    state.replacements.set(candidate.toolUseId, replacement.content)
+    newlyReplaced.push({
+      kind: 'tool-result',
+      toolUseId: candidate.toolUseId,
+      replacement: replacement.content,
+    })
+    logEvent('tengu_tool_result_persisted_message_budget', {
+      originalSizeBytes: replacement.originalSize,
+      persistedSizeBytes: replacement.content.length,
+      estimatedOriginalTokens: Math.ceil(
+        replacement.originalSize / BYTES_PER_TOKEN,
+      ),
+      estimatedPersistedTokens: Math.ceil(
+        replacement.content.length / BYTES_PER_TOKEN,
+      ),
+    })
+  }
+
+  if (replacementMap.size === 0) {
+    return { messages, newlyReplaced: [] }
+  }
+
+  if (newlyReplaced.length > 0) {
+    logForDebugging(
+      `Per-message budget: persisted ${newlyReplaced.length} tool results ` +
+        `across ${messagesOverBudget} over-budget message(s), ` +
+        `shed ~${formatFileSize(replacedSize)}, ${reappliedCount} re-applied`,
+    )
+    logEvent('tengu_message_level_tool_result_budget_enforced', {
+      resultsPersisted: newlyReplaced.length,
+      messagesOverBudget,
+      replacedSizeBytes: replacedSize,
+      reapplied: reappliedCount,
+    })
+  }
+
+  return {
+    messages: replaceToolResultContents(messages, replacementMap),
+    newlyReplaced,
+  }
+}
+
+/**
+ * Query-loop integration point for the aggregate budget.
+ *
+ * Gates on `state` (undefined means feature disabled → no-op return),
+ * applies enforcement, and fires an optional transcript-write callback
+ * for new replacements. The caller (query.ts) owns the persistence gate
+ * — it passes a callback only for querySources that read records back on
+ * resume (repl_main_thread*, agent:*); ephemeral runForkedAgent callers
+ * (agentSummary, sessionMemory, /btw, compact) pass undefined.
+ *
+ * @returns messages with replacements applied, or the input array unchanged
+ *   when the feature is off or no replacement occurred.
+ */
+export async function applyToolResultBudget(
+  messages: Message[],
+  state: ContentReplacementState | undefined,
+  writeToTranscript?: (records: ToolResultReplacementRecord[]) => void,
+  skipToolNames?: ReadonlySet<string>,
+): Promise<Message[]> {
+  if (!state) return messages
+  const result = await enforceToolResultBudget(messages, state, skipToolNames)
+  if (result.newlyReplaced.length > 0) {
+    writeToTranscript?.(result.newlyReplaced)
+  }
+  return result.messages
+}
+
+/**
+ * Reconstruct replacement state from content-replacement records loaded from
+ * the transcript. Used on resume so the budget makes the same choices it
+ * made in the original session (prompt cache stability).
+ *
+ * Accepts the full ContentReplacementRecord[] from LogOption (may include
+ * future non-tool-result kinds); only tool-result records are applied here.
+ *
+ *   - replacements: populated directly from the stored replacement strings.
+ *     Records for IDs not in messages (e.g. after compact) are skipped —
+ *     they're inert anyway.
+ *   - seenIds: every candidate tool_use_id in the loaded messages. A result
+ *     being in the transcript means it was sent to the model, so it was seen.
+ *     This freezes unreplaced results against future replacement.
+ *   - inheritedReplacements: gap-fill for fork-subagent resume. A fork's
+ *     original run applies parent-inherited replacements via mustReapply
+ *     (never persisted — not newlyReplaced). On resume the sidechain has
+ *     the original content but no record, so records alone would classify
+ *     it as frozen. The parent's live state still has the mapping; copy
+ *     it for IDs in messages that records don't cover. No-op for non-fork
+ *     resumes (parent IDs aren't in the subagent's messages).
+ */
+export function reconstructContentReplacementState(
+  messages: Message[],
+  records: ContentReplacementRecord[],
+  inheritedReplacements?: ReadonlyMap<string, string>,
+): ContentReplacementState {
+  const state = createContentReplacementState()
+  const candidateIds = new Set(
+    collectCandidatesByMessage(messages)
+      .flat()
+      .map(c => c.toolUseId),
+  )
+
+  for (const id of candidateIds) {
+    state.seenIds.add(id)
+  }
+  for (const r of records) {
+    if (r.kind === 'tool-result' && candidateIds.has(r.toolUseId)) {
+      state.replacements.set(r.toolUseId, r.replacement)
+    }
+  }
+  if (inheritedReplacements) {
+    for (const [id, replacement] of inheritedReplacements) {
+      if (candidateIds.has(id) && !state.replacements.has(id)) {
+        state.replacements.set(id, replacement)
+      }
+    }
+  }
+  return state
+}
+
+/**
+ * AgentTool-resume variant: encapsulates the feature-flag gate + parent
+ * gap-fill so both AgentTool.call and resumeAgentBackground share one
+ * implementation. Returns undefined when parentState is undefined (feature
+ * off); otherwise reconstructs from sidechain records with parent's live
+ * replacements filling gaps for fork-inherited mustReapply entries.
+ *
+ * Kept out of AgentTool.tsx — that file is at the feature() DCE complexity
+ * cliff and cannot tolerate even +1 net source line without silently
+ * breaking feature('TRANSCRIPT_CLASSIFIER') eval in tests.
+ */
+export function reconstructForSubagentResume(
+  parentState: ContentReplacementState | undefined,
+  resumedMessages: Message[],
+  sidechainRecords: ContentReplacementRecord[],
+): ContentReplacementState | undefined {
+  if (!parentState) return undefined
+  return reconstructContentReplacementState(
+    resumedMessages,
+    sidechainRecords,
+    parentState.replacements,
+  )
+}
+
+/**
+ * Get a human-readable error message from a filesystem error
+ */
+function getFileSystemErrorMessage(error: Error): string {
+  // Node.js filesystem errors have a 'code' property
+  // eslint-disable-next-line no-restricted-syntax -- uses .path, not just .code
+  const nodeError = error as NodeJS.ErrnoException
+  if (nodeError.code) {
+    switch (nodeError.code) {
+      case 'ENOENT':
+        return `Directory not found: ${nodeError.path ?? 'unknown path'}`
+      case 'EACCES':
+        return `Permission denied: ${nodeError.path ?? 'unknown path'}`
+      case 'ENOSPC':
+        return 'No space left on device'
+      case 'EROFS':
+        return 'Read-only file system'
+      case 'EMFILE':
+        return 'Too many open files'
+      case 'EEXIST':
+        return `File already exists: ${nodeError.path ?? 'unknown path'}`
+      default:
+        return `${nodeError.code}: ${nodeError.message}`
+    }
+  }
+  return error.message
+}