@illuma-ai/agents 1.4.0-alpha.6 → 1.5.0

package/dist/cjs/tools/toolOutputReferences.cjs.map

@@ -0,0 +1 @@
+{"version":3,"file":"toolOutputReferences.cjs","sources":["../../../src/tools/toolOutputReferences.ts"],"sourcesContent":["/**\n * Tool output reference registry.\n *\n * When enabled via `RunConfig.toolOutputReferences.enabled`, ToolNode\n * stores each successful tool output under a stable key\n * (`tool<idx>turn<turn>`) where `idx` is the tool's position within a\n * ToolNode batch and `turn` is the batch index within the run\n * (incremented once per ToolNode invocation).\n *\n * Subsequent tool calls can pipe a previous output into their args by\n * embedding `{{tool<idx>turn<turn>}}` inside any string argument;\n * {@link ToolOutputReferenceRegistry.resolve} walks the args and\n * substitutes the placeholders immediately before invocation.\n *\n * The registry stores the *raw, untruncated* tool output so a later\n * `{{…}}` substitution pipes the full payload into the next tool —\n * even when the LLM only saw a head+tail-truncated preview in\n * `ToolMessage.content`. Outputs are stored without any annotation\n * (the `_ref` key or the `[ref: ...]` prefix seen by the LLM is\n * strictly a UX signal attached to `ToolMessage.content`). Keeping the\n * registry pristine means downstream bash/jq piping receives the\n * complete, verbatim output with no injected fields.\n */\n\nimport { ToolMessage } from '@langchain/core/messages';\nimport type { BaseMessage } from '@langchain/core/messages';\nimport {\n  calculateMaxTotalToolOutputSize,\n  HARD_MAX_TOOL_RESULT_CHARS,\n  HARD_MAX_TOTAL_TOOL_OUTPUT_SIZE,\n} from '@/utils/truncation';\n\n/**\n * Non-global matcher for a single `{{tool<i>turn<n>}}` placeholder.\n * Exported for consumers that want to detect references (e.g., syntax\n * highlighting, docs). The stateful `g` variant lives inside the\n * registry so nobody trips on `lastIndex`.\n */\nexport const TOOL_OUTPUT_REF_PATTERN = /\\{\\{(tool\\d+turn\\d+)\\}\\}/;\n\n/** Object key used when a parsed-object output has `_ref` injected. */\nexport const TOOL_OUTPUT_REF_KEY = '_ref';\n\n/**\n * Object key used to carry unresolved reference warnings on a parsed-\n * object output. Using a dedicated field instead of a trailing text\n * line keeps the annotated `ToolMessage.content` parseable as JSON for\n * downstream consumers that rely on the object shape.\n */\nexport const TOOL_OUTPUT_UNRESOLVED_KEY = '_unresolved_refs';\n\n/** Single-line prefix prepended to non-object tool outputs so the LLM sees the reference key. */\nexport function buildReferencePrefix(key: string): string {\n  return `[ref: ${key}]`;\n}\n\n/** Stable registry key for a tool output. */\nexport function buildReferenceKey(toolIndex: number, turn: number): string {\n  return `tool${toolIndex}turn${turn}`;\n}\n\nexport type ToolOutputReferenceRegistryOptions = {\n  /** Maximum characters stored per registered output. */\n  maxOutputSize?: number;\n  /** Maximum total characters retained across all registered outputs. */\n  maxTotalSize?: number;\n  /**\n   * Upper bound on the number of concurrently-tracked runs. When\n   * exceeded, the oldest run bucket is evicted (FIFO). Defaults to 32.\n   */\n  maxActiveRuns?: number;\n};\n\n/**\n * Result of resolving placeholders in tool args.\n */\nexport type ResolveResult<T> = {\n  /** Arguments with placeholders replaced. Same shape as the input. */\n  resolved: T;\n  /** Reference keys that were referenced but had no stored value. */\n  unresolved: string[];\n};\n\n/**\n * Read-only view over a frozen registry snapshot. Returned by\n * {@link ToolOutputReferenceRegistry.snapshot} for callers that need\n * to resolve placeholders against the registry state at a specific\n * point in time, ignoring any subsequent registrations.\n */\nexport interface ToolOutputResolveView {\n  resolve<T>(args: T): ResolveResult<T>;\n}\n\n/**\n * Pre-resolved arg map keyed by `toolCallId`. Used by the mixed\n * direct+event dispatch path to feed event calls' resolved args\n * (captured pre-batch) into the dispatcher without re-resolving\n * against the now-stale live registry.\n */\nexport type PreResolvedArgsMap = Map<\n  string,\n  { resolved: Record<string, unknown>; unresolved: string[] }\n>;\n\n/**\n * Per-call sink for resolved args, keyed by `toolCallId`. Threaded\n * as a per-batch local map so concurrent `ToolNode.run()` calls do\n * not race on shared sink state.\n */\nexport type ResolvedArgsByCallId = Map<string, Record<string, unknown>>;\n\nconst EMPTY_ENTRIES: ReadonlyMap<string, string> = new Map<string, string>();\n\n/**\n * Per-run state bucket held inside the registry. Each distinct\n * `run_id` gets its own bucket so overlapping concurrent runs on a\n * shared registry cannot leak outputs, turn counters, or warn-memos\n * into one another.\n */\nclass RunStateBucket {\n  entries: Map<string, string> = new Map();\n  totalSize: number = 0;\n  turnCounter: number = 0;\n  warnedNonStringTools: Set<string> = new Set();\n}\n\n/**\n * Anonymous (`run_id` absent) bucket key. Anonymous batches are\n * treated as fresh runs on every invocation — see `nextTurn`.\n */\nconst ANON_RUN_KEY = '\\0anon';\n\n/**\n * Default upper bound on the number of concurrently-tracked runs per\n * registry. When exceeded, the oldest run's bucket (by insertion\n * order) is evicted. Keeps memory bounded when a ToolNode is reused\n * across many runs without explicit `releaseRun` calls.\n */\nconst DEFAULT_MAX_ACTIVE_RUNS = 32;\n\n/**\n * Ordered map of reference-key → stored output, partitioned by run so\n * concurrent / interleaved runs sharing one registry cannot leak\n * outputs between each other.\n *\n * Each public method takes a `runId` which selects the run's bucket.\n * Hosts typically get one registry per run via `Graph`, in which\n * case only a single bucket is ever populated; the partitioning\n * exists so the registry also behaves correctly when a single\n * instance is reused directly.\n */\nexport class ToolOutputReferenceRegistry {\n  private runStates: Map<string, RunStateBucket> = new Map();\n  private readonly maxOutputSize: number;\n  private readonly maxTotalSize: number;\n  private readonly maxActiveRuns: number;\n  /**\n   * Local stateful matcher used only by `replaceInString`. Kept\n   * off-module so callers of the exported `TOOL_OUTPUT_REF_PATTERN`\n   * never see a stale `lastIndex`.\n   */\n  private static readonly PLACEHOLDER_MATCHER = /\\{\\{(tool\\d+turn\\d+)\\}\\}/g;\n\n  constructor(options: ToolOutputReferenceRegistryOptions = {}) {\n    /**\n     * Per-output default is the same ~400 KB budget as the standard\n     * tool-result truncation (`HARD_MAX_TOOL_RESULT_CHARS`). This\n     * keeps a single `{{…}}` substitution at a size that is safe to\n     * pass through typical shell `ARG_MAX` limits and matches what\n     * the LLM would otherwise have seen. Hosts that want larger per-\n     * output payloads (API consumers, long JSON streams) can raise\n     * the cap explicitly up to the 5 MB total budget.\n     */\n    const perOutput =\n      options.maxOutputSize != null && options.maxOutputSize > 0\n        ? options.maxOutputSize\n        : HARD_MAX_TOOL_RESULT_CHARS;\n    /**\n     * Clamp a caller-supplied `maxTotalSize` to\n     * `HARD_MAX_TOTAL_TOOL_OUTPUT_SIZE` (5 MB) so the documented\n     * absolute cap is enforced regardless of host config —\n     * `calculateMaxTotalToolOutputSize` already applies the same\n     * upper bound on its computed default, but the user-provided\n     * branch was bypassing it.\n     */\n    const totalRaw =\n      options.maxTotalSize != null && options.maxTotalSize > 0\n        ? Math.min(options.maxTotalSize, HARD_MAX_TOTAL_TOOL_OUTPUT_SIZE)\n        : calculateMaxTotalToolOutputSize(perOutput);\n    this.maxTotalSize = totalRaw;\n    /**\n     * The per-output cap can never exceed the per-run aggregate cap:\n     * if a single entry were allowed to be larger than `maxTotalSize`,\n     * the eviction loop would either blow the cap (to keep the entry)\n     * or self-evict a just-stored value. Clamping here turns\n     * `maxTotalSize` into a hard upper bound on *any* state the\n     * registry retains per run.\n     */\n    this.maxOutputSize = Math.min(perOutput, totalRaw);\n    this.maxActiveRuns =\n      options.maxActiveRuns != null && options.maxActiveRuns > 0\n        ? options.maxActiveRuns\n        : DEFAULT_MAX_ACTIVE_RUNS;\n  }\n\n  private keyFor(runId: string | undefined): string {\n    return runId ?? ANON_RUN_KEY;\n  }\n\n  private getOrCreate(runId: string | undefined): RunStateBucket {\n    const key = this.keyFor(runId);\n    let state = this.runStates.get(key);\n    if (state == null) {\n      state = new RunStateBucket();\n      this.runStates.set(key, state);\n      if (this.runStates.size > this.maxActiveRuns) {\n        const oldest = this.runStates.keys().next().value;\n        if (oldest != null && oldest !== key) {\n          this.runStates.delete(oldest);\n        }\n      }\n    }\n    return state;\n  }\n\n  /** Registers (or replaces) the output stored under `key` for `runId`. */\n  set(runId: string | undefined, key: string, value: string): void {\n    const bucket = this.getOrCreate(runId);\n    const clipped =\n      value.length > this.maxOutputSize\n        ? value.slice(0, this.maxOutputSize)\n        : value;\n    const existing = bucket.entries.get(key);\n    if (existing != null) {\n      bucket.totalSize -= existing.length;\n      bucket.entries.delete(key);\n    }\n    bucket.entries.set(key, clipped);\n    bucket.totalSize += clipped.length;\n    this.evictWithinBucket(bucket);\n  }\n\n  /** Returns the stored value for `key` in `runId`'s bucket, or `undefined`. */\n  get(runId: string | undefined, key: string): string | undefined {\n    return this.runStates.get(this.keyFor(runId))?.entries.get(key);\n  }\n\n  /**\n   * Returns `true` when `key` is currently stored in `runId`'s bucket.\n   * Used by {@link annotateMessagesForLLM} to gate transient annotation\n   * on whether the registry still owns the referenced output (a stale\n   * `_refKey` from a prior run silently no-ops here).\n   */\n  has(runId: string | undefined, key: string): boolean {\n    return this.runStates.get(this.keyFor(runId))?.entries.has(key) ?? false;\n  }\n\n  /** Total number of registered outputs across every run bucket. */\n  get size(): number {\n    let n = 0;\n    for (const bucket of this.runStates.values()) {\n      n += bucket.entries.size;\n    }\n    return n;\n  }\n\n  /** Maximum characters retained per output (post-clip). */\n  get perOutputLimit(): number {\n    return this.maxOutputSize;\n  }\n\n  /** Maximum total characters retained *per run*. */\n  get totalLimit(): number {\n    return this.maxTotalSize;\n  }\n\n  /** Drops every run's state. */\n  clear(): void {\n    this.runStates.clear();\n  }\n\n  /**\n   * Explicitly release `runId`'s state. Safe to call when a run has\n   * finished. Hosts sharing one registry across runs should call this\n   * to reclaim memory deterministically; otherwise LRU eviction kicks\n   * in when `maxActiveRuns` runs accumulate.\n   */\n  releaseRun(runId: string | undefined): void {\n    this.runStates.delete(this.keyFor(runId));\n  }\n\n  /**\n   * Claims the next batch turn synchronously from `runId`'s bucket.\n   *\n   * Must be called once at the start of each ToolNode batch before\n   * any `await`, so concurrent invocations within the same run see\n   * distinct turn values (reads are effectively atomic by JS's\n   * single-threaded execution of the sync prefix).\n   *\n   * If `runId` is missing the anonymous bucket is dropped and a\n   * fresh one created so each anonymous call behaves as its own run.\n   */\n  nextTurn(runId: string | undefined): number {\n    if (runId == null) {\n      this.runStates.delete(ANON_RUN_KEY);\n    }\n    const bucket = this.getOrCreate(runId);\n    return bucket.turnCounter++;\n  }\n\n  /**\n   * Records that `toolName` has been warned about in `runId` (returns\n   * `true` on the first call per run, `false` after). Used by\n   * ToolNode to emit one log line per offending tool per run when a\n   * `ToolMessage.content` isn't a string.\n   */\n  claimWarnOnce(runId: string | undefined, toolName: string): boolean {\n    const bucket = this.getOrCreate(runId);\n    if (bucket.warnedNonStringTools.has(toolName)) {\n      return false;\n    }\n    bucket.warnedNonStringTools.add(toolName);\n    return true;\n  }\n\n  /**\n   * Walks `args` and replaces every `{{tool<i>turn<n>}}` placeholder in\n   * string values with the stored output *from `runId`'s bucket*. Non-\n   * string values and object keys are left untouched. Unresolved\n   * references are left in-place and reported so the caller can\n   * surface them to the LLM. When no placeholder appears anywhere in\n   * the serialized args, the original input is returned without\n   * walking the tree.\n   */\n  resolve<T>(runId: string | undefined, args: T): ResolveResult<T> {\n    if (!hasAnyPlaceholder(args)) {\n      return { resolved: args, unresolved: [] };\n    }\n    const bucket = this.runStates.get(this.keyFor(runId));\n    return this.resolveAgainst(bucket?.entries ?? EMPTY_ENTRIES, args);\n  }\n\n  /**\n   * Captures a frozen snapshot of `runId`'s current entries and\n   * returns a view that resolves placeholders against *only* that\n   * snapshot. The snapshot is decoupled from the live registry, so\n   * subsequent `set()` calls (for example, same-turn direct outputs\n   * registering while an event branch is still in flight) are\n   * invisible to the snapshot's `resolve`. Used by the mixed\n   * direct+event dispatch path to preserve same-turn isolation when\n   * a `PreToolUse` hook rewrites event args after directs have\n   * completed.\n   */\n  snapshot(runId: string | undefined): ToolOutputResolveView {\n    const bucket = this.runStates.get(this.keyFor(runId));\n    const entries: ReadonlyMap<string, string> = bucket\n      ? new Map(bucket.entries)\n      : EMPTY_ENTRIES;\n    return {\n      resolve: <T>(args: T): ResolveResult<T> =>\n        this.resolveAgainst(entries, args),\n    };\n  }\n\n  private resolveAgainst<T>(\n    entries: ReadonlyMap<string, string>,\n    args: T\n  ): ResolveResult<T> {\n    if (!hasAnyPlaceholder(args)) {\n      return { resolved: args, unresolved: [] };\n    }\n    const unresolved = new Set<string>();\n    const resolved = this.transform(entries, args, unresolved) as T;\n    return { resolved, unresolved: Array.from(unresolved) };\n  }\n\n  private transform(\n    entries: ReadonlyMap<string, string>,\n    value: unknown,\n    unresolved: Set<string>\n  ): unknown {\n    if (typeof value === 'string') {\n      return this.replaceInString(entries, value, unresolved);\n    }\n    if (Array.isArray(value)) {\n      return value.map((item) => this.transform(entries, item, unresolved));\n    }\n    if (value !== null && typeof value === 'object') {\n      const source = value as Record<string, unknown>;\n      const next: Record<string, unknown> = {};\n      for (const [key, item] of Object.entries(source)) {\n        next[key] = this.transform(entries, item, unresolved);\n      }\n      return next;\n    }\n    return value;\n  }\n\n  private replaceInString(\n    entries: ReadonlyMap<string, string>,\n    input: string,\n    unresolved: Set<string>\n  ): string {\n    if (input.indexOf('{{tool') === -1) {\n      return input;\n    }\n    return input.replace(\n      ToolOutputReferenceRegistry.PLACEHOLDER_MATCHER,\n      (match, key: string) => {\n        const stored = entries.get(key);\n        if (stored == null) {\n          unresolved.add(key);\n          return match;\n        }\n        return stored;\n      }\n    );\n  }\n\n  private evictWithinBucket(bucket: RunStateBucket): void {\n    if (bucket.totalSize <= this.maxTotalSize) {\n      return;\n    }\n    for (const key of bucket.entries.keys()) {\n      if (bucket.totalSize <= this.maxTotalSize) {\n        return;\n      }\n      const entry = bucket.entries.get(key);\n      if (entry == null) {\n        continue;\n      }\n      bucket.totalSize -= entry.length;\n      bucket.entries.delete(key);\n    }\n  }\n}\n\n/**\n * Cheap pre-check: returns true if any string value in `args` contains\n * the `{{tool` substring. Lets `resolve()` skip the deep tree walk (and\n * its object allocations) for the common case of plain args.\n */\nfunction hasAnyPlaceholder(value: unknown): boolean {\n  if (typeof value === 'string') {\n    return value.indexOf('{{tool') !== -1;\n  }\n  if (Array.isArray(value)) {\n    for (const item of value) {\n      if (hasAnyPlaceholder(item)) {\n        return true;\n      }\n    }\n    return false;\n  }\n  if (value !== null && typeof value === 'object') {\n    for (const item of Object.values(value as Record<string, unknown>)) {\n      if (hasAnyPlaceholder(item)) {\n        return true;\n      }\n    }\n    return false;\n  }\n  return false;\n}\n\n/**\n * Annotates `content` with a reference key and/or unresolved-ref\n * warnings so the LLM sees both alongside the tool output.\n *\n * Behavior:\n *  - If `content` parses as a plain (non-array, non-null) JSON object\n *    and the object does not already have a conflicting `_ref` key,\n *    the reference key and (when present) `_unresolved_refs` array\n *    are injected as object fields, preserving JSON validity for\n *    downstream consumers that parse the output.\n *  - Otherwise (string output, JSON array/primitive, parse failure,\n *    or `_ref` collision), a `[ref: <key>]\\n` prefix line is\n *    prepended and unresolved refs are appended as a trailing\n *    `[unresolved refs: …]` line.\n *\n * The annotated string is what the LLM sees as `ToolMessage.content`.\n * The *original* (un-annotated) value is what gets stored in the\n * registry, so downstream piping remains pristine.\n *\n * @param content     Raw (post-truncation) tool output.\n * @param key         Reference key for this output, or undefined when\n *                    there is nothing to register (errors etc.).\n * @param unresolved  Reference keys that failed to resolve during\n *                    argument substitution. Surfaced so the LLM can\n *                    self-correct its next tool call.\n */\nexport function annotateToolOutputWithReference(\n  content: string,\n  key: string | undefined,\n  unresolved: string[] = []\n): string {\n  const hasRefKey = key != null;\n  const hasUnresolved = unresolved.length > 0;\n  if (!hasRefKey && !hasUnresolved) {\n    return content;\n  }\n  const trimmed = content.trimStart();\n  if (trimmed.startsWith('{')) {\n    const annotated = tryInjectRefIntoJsonObject(content, key, unresolved);\n    if (annotated != null) {\n      return annotated;\n    }\n  }\n  const prefix = hasRefKey ? `${buildReferencePrefix(key!)}\\n` : '';\n  const trailer = hasUnresolved\n    ? `\\n[unresolved refs: ${unresolved.join(', ')}]`\n    : '';\n  return `${prefix}${content}${trailer}`;\n}\n\nfunction tryInjectRefIntoJsonObject(\n  content: string,\n  key: string | undefined,\n  unresolved: string[]\n): string | null {\n  let parsed: unknown;\n  try {\n    parsed = JSON.parse(content);\n  } catch {\n    return null;\n  }\n\n  if (parsed === null || typeof parsed !== 'object' || Array.isArray(parsed)) {\n    return null;\n  }\n\n  const obj = parsed as Record<string, unknown>;\n  const injectingRef = key != null;\n  const injectingUnresolved = unresolved.length > 0;\n\n  /**\n   * Reject the JSON-injection path (fall back to prefix form) when\n   * either of our keys collides with real payload data:\n   *  - `_ref` collision: existing value is non-null and differs from\n   *    the key we're about to inject.\n   *  - `_unresolved_refs` collision: existing value is non-null and\n   *    is not a deep-equal match for the array we'd inject.\n   * This keeps us from silently overwriting legitimate tool output.\n   */\n  if (\n    injectingRef &&\n    TOOL_OUTPUT_REF_KEY in obj &&\n    obj[TOOL_OUTPUT_REF_KEY] !== key &&\n    obj[TOOL_OUTPUT_REF_KEY] != null\n  ) {\n    return null;\n  }\n  if (\n    injectingUnresolved &&\n    TOOL_OUTPUT_UNRESOLVED_KEY in obj &&\n    obj[TOOL_OUTPUT_UNRESOLVED_KEY] != null &&\n    !arraysShallowEqual(obj[TOOL_OUTPUT_UNRESOLVED_KEY], unresolved)\n  ) {\n    return null;\n  }\n\n  /**\n   * Only strip the framework-owned key we're actually injecting —\n   * leave everything else (including a pre-existing `_ref` on the\n   * unresolved-only path, or a pre-existing `_unresolved_refs` on a\n   * plain-annotation path) untouched so we annotate rather than\n   * mutate downstream payload data. Our injected keys land first in\n   * the serialized JSON so the LLM sees them before the body.\n   */\n  const omitKeys = new Set<string>();\n  if (injectingRef) omitKeys.add(TOOL_OUTPUT_REF_KEY);\n  if (injectingUnresolved) omitKeys.add(TOOL_OUTPUT_UNRESOLVED_KEY);\n  const rest: Record<string, unknown> = {};\n  for (const [k, v] of Object.entries(obj)) {\n    if (!omitKeys.has(k)) {\n      rest[k] = v;\n    }\n  }\n  const injected: Record<string, unknown> = {};\n  if (injectingRef) {\n    injected[TOOL_OUTPUT_REF_KEY] = key;\n  }\n  if (injectingUnresolved) {\n    injected[TOOL_OUTPUT_UNRESOLVED_KEY] = unresolved;\n  }\n  Object.assign(injected, rest);\n\n  const pretty = /^\\{\\s*\\n/.test(content);\n  return pretty ? JSON.stringify(injected, null, 2) : JSON.stringify(injected);\n}\n\nfunction arraysShallowEqual(a: unknown, b: readonly string[]): boolean {\n  if (!Array.isArray(a) || a.length !== b.length) {\n    return false;\n  }\n  for (let i = 0; i < a.length; i++) {\n    if (a[i] !== b[i]) {\n      return false;\n    }\n  }\n  return true;\n}\n\n/**\n * Lazy projection that, given a registry and a runId, returns a new\n * `messages` array where each `ToolMessage` carrying ref metadata is\n * projected into a transient copy with annotated content (when the ref\n * is live in the registry) and with the framework-owned `additional_\n * kwargs` keys (`_refKey`, `_refScope`, `_unresolvedRefs`) stripped\n * regardless of whether annotation applied. The original input array\n * and its messages are never mutated.\n *\n * Annotation is gated on registry presence: a stale `_refKey` from a\n * prior run (e.g. one that survived in persisted history) silently\n * no-ops on the *content* side. The strip-metadata side still runs so\n * stale framework keys never leak onto the wire under any custom or\n * future provider serializer that might transmit `additional_kwargs`.\n * `_unresolvedRefs` is always meaningful and is not gated.\n *\n * **Feature-disabled fast path:** when the host hasn't enabled the\n * tool-output-reference feature, the registry is `undefined` and this\n * function returns the input array reference-equal *without iterating\n * a single message*. The loop is exclusive to the feature-enabled\n * code path.\n */\nexport function annotateMessagesForLLM(\n  messages: BaseMessage[],\n  registry: ToolOutputReferenceRegistry | undefined,\n  runId: string | undefined\n): BaseMessage[] {\n  if (registry == null) return messages;\n\n  /**\n   * Lazy-allocate the output array so the common case (no ToolMessage\n   * carries framework metadata) returns the input reference-equal with\n   * zero allocations beyond the per-message predicate checks.\n   */\n  let out: BaseMessage[] | undefined;\n  for (let i = 0; i < messages.length; i++) {\n    const m = messages[i];\n    if (m._getType() !== 'tool') continue;\n    /**\n     * `additional_kwargs` is untyped at the LangChain layer\n     * (`Record<string, unknown>`), so persisted or client-supplied\n     * ToolMessages can carry arbitrary shapes — including primitives\n     * (a malformed serializer might write a string, or `null`).\n     * Guard with a runtime object check before the `in` probes\n     * because the `in` operator throws `TypeError` on primitives.\n     * A single malformed message must never crash the provider call\n     * path; skip its annotation/strip and continue.\n     */\n    const rawMeta = m.additional_kwargs as unknown;\n    if (rawMeta == null || typeof rawMeta !== 'object') continue;\n    const meta = rawMeta as Record<string, unknown>;\n    const hasRefKey = '_refKey' in meta;\n    const hasRefScope = '_refScope' in meta;\n    const hasUnresolvedField = '_unresolvedRefs' in meta;\n    if (!hasRefKey && !hasRefScope && !hasUnresolvedField) continue;\n\n    const refKey = readRefKey(meta);\n    const unresolved = readUnresolvedRefs(meta);\n\n    /**\n     * Prefer the message-stamped `_refScope` for the registry lookup.\n     * For named runs it equals the current `runId`; for anonymous\n     * invocations it carries the per-batch synthetic scope minted by\n     * ToolNode (`\\0anon-<n>`), which `runId` from config cannot\n     * recover. Falling back to `runId` keeps backward compatibility\n     * with messages stamped before this field existed.\n     */\n    const lookupScope = readRefScope(meta) ?? runId;\n    const liveRef =\n      refKey != null && registry.has(lookupScope, refKey) ? refKey : undefined;\n    const annotates = liveRef != null || unresolved.length > 0;\n\n    const tm = m as ToolMessage;\n    let nextContent: ToolMessage['content'] = tm.content;\n\n    if (annotates && typeof tm.content === 'string') {\n      nextContent = annotateToolOutputWithReference(\n        tm.content,\n        liveRef,\n        unresolved\n      );\n    } else if (\n      annotates &&\n      Array.isArray(tm.content) &&\n      unresolved.length > 0\n    ) {\n      const warningBlock = {\n        type: 'text' as const,\n        text: `[unresolved refs: ${unresolved.join(', ')}]`,\n      };\n      /**\n       * `as unknown as ToolMessage['content']` is unavoidable here:\n       * LangChain's content union (`MessageContentComplex[] |\n       * DataContentBlock[] | string`) does not accept a freshly built\n       * mixed array literal even though the structural shape is valid\n       * at runtime. The double-cast is structurally safe — we\n       * preserve every block from `tm.content` and prepend a single\n       * `{ type: 'text', text }` block that all providers accept.\n       */\n      nextContent = [\n        warningBlock,\n        ...tm.content,\n      ] as unknown as ToolMessage['content'];\n    }\n\n    /**\n     * Project unconditionally: even when no annotation applies (stale\n     * `_refKey` or non-annotatable content), `cloneToolMessageWithContent`\n     * runs `stripFrameworkRefMetadata` on `additional_kwargs` so the\n     * framework-owned keys never reach the wire.\n     */\n    out ??= messages.slice();\n    out[i] = cloneToolMessageWithContent(tm, nextContent);\n  }\n\n  return out ?? messages;\n}\n\n/**\n * Reads `_refKey` defensively from untyped `additional_kwargs`. Returns\n * undefined for non-string values so a malformed field cannot poison\n * the registry lookup or downstream string operations.\n */\nfunction readRefKey(\n  meta: Record<string, unknown> | undefined\n): string | undefined {\n  const v = meta?._refKey;\n  return typeof v === 'string' ? v : undefined;\n}\n\n/**\n * Reads `_refScope` defensively from untyped `additional_kwargs`.\n * Mirrors {@link readRefKey} — non-string scopes are dropped (the\n * caller falls back to the run-derived scope) rather than passed into\n * the registry as a malformed key.\n */\nfunction readRefScope(\n  meta: Record<string, unknown> | undefined\n): string | undefined {\n  const v = meta?._refScope;\n  return typeof v === 'string' ? v : undefined;\n}\n\n/**\n * Reads `_unresolvedRefs` defensively from untyped `additional_kwargs`.\n * Returns an empty array for any non-array value, and filters out\n * non-string entries from a real array. Without this guard, a hydrated\n * ToolMessage carrying e.g. `_unresolvedRefs: 'tool0turn0'` would crash\n * `attemptInvoke` on the eventual `.length` / `.join(...)` call.\n */\nfunction readUnresolvedRefs(\n  meta: Record<string, unknown> | undefined\n): string[] {\n  const v = meta?._unresolvedRefs;\n  if (!Array.isArray(v)) return [];\n  const out: string[] = [];\n  for (const item of v) {\n    if (typeof item === 'string') out.push(item);\n  }\n  return out;\n}\n\n/**\n * Builds a fresh `ToolMessage` that mirrors `tm`'s identity fields with\n * the supplied `content`. Every `ToolMessage` field but `content` is\n * carried over so the projection is structurally identical to the\n * original from a LangChain serializer's perspective.\n *\n * `additional_kwargs` is rebuilt with the framework-owned ref keys\n * stripped. Defensive: LangChain's standard provider serializers do not\n * transmit `additional_kwargs` to provider HTTP APIs, but a custom\n * adapter or future LangChain change could. Stripping keeps the\n * implementation correct under any serializer behavior at the cost of a\n * shallow object spread per annotated message.\n */\nfunction cloneToolMessageWithContent(\n  tm: ToolMessage,\n  content: ToolMessage['content']\n): ToolMessage {\n  return new ToolMessage({\n    id: tm.id,\n    name: tm.name,\n    status: tm.status,\n    artifact: tm.artifact,\n    tool_call_id: tm.tool_call_id,\n    response_metadata: tm.response_metadata,\n    additional_kwargs: stripFrameworkRefMetadata(tm.additional_kwargs),\n    content,\n  });\n}\n\n/**\n * Returns a copy of `kwargs` with `_refKey`, `_refScope`, and\n * `_unresolvedRefs` removed. Returns the input reference-equal when\n * none of those keys are present so the no-strip path stays cheap;\n * returns `undefined` when stripping leaves the object empty so the\n * caller can drop the field entirely.\n */\nfunction stripFrameworkRefMetadata(\n  kwargs: Record<string, unknown> | undefined\n): Record<string, unknown> | undefined {\n  if (kwargs == null) return undefined;\n  if (\n    !('_refKey' in kwargs) &&\n    !('_refScope' in kwargs) &&\n    !('_unresolvedRefs' in kwargs)\n  ) {\n    return kwargs;\n  }\n  const { _refKey, _refScope, _unresolvedRefs, ...rest } = kwargs as Record<\n    string,\n    unknown\n  > & {\n    _refKey?: unknown;\n    _refScope?: unknown;\n    _unresolvedRefs?: unknown;\n  };\n  void _refKey;\n  void _refScope;\n  void _unresolvedRefs;\n  return Object.keys(rest).length === 0 ? undefined : rest;\n}\n"],"names":["HARD_MAX_TOOL_RESULT_CHARS","HARD_MAX_TOTAL_TOOL_OUTPUT_SIZE","calculateMaxTotalToolOutputSize","ToolMessage"],"mappings":";;;;;AAAA;;;;;;;;;;;;;;;;;;;;;;AAsBG;AAkBH;AACO,MAAM,mBAAmB,GAAG;AAEnC;;;;;AAKG;AACI,MAAM,0BAA0B,GAAG;AAE1C;AACM,SAAU,oBAAoB,CAAC,GAAW,EAAA;IAC9C,OAAO,CAAA,MAAA,EAAS,GAAG,CAAA,CAAA,CAAG;AACxB;AAEA;AACM,SAAU,iBAAiB,CAAC,SAAiB,EAAE,IAAY,EAAA;AAC/D,IAAA,OAAO,CAAA,IAAA,EAAO,SAAS,CAAA,IAAA,EAAO,IAAI,EAAE;AACtC;AAoDA,MAAM,aAAa,GAAgC,IAAI,GAAG,EAAkB;AAE5E;;;;;AAKG;AACH,MAAM,cAAc,CAAA;AAClB,IAAA,OAAO,GAAwB,IAAI,GAAG,EAAE;IACxC,SAAS,GAAW,CAAC;IACrB,WAAW,GAAW,CAAC;AACvB,IAAA,oBAAoB,GAAgB,IAAI,GAAG,EAAE;AAC9C;AAED;;;AAGG;AACH,MAAM,YAAY,GAAG,QAAQ;AAE7B;;;;;AAKG;AACH,MAAM,uBAAuB,GAAG,EAAE;AAElC;;;;;;;;;;AAUG;MACU,2BAA2B,CAAA;AAC9B,IAAA,SAAS,GAAgC,IAAI,GAAG,EAAE;AACzC,IAAA,aAAa;AACb,IAAA,YAAY;AACZ,IAAA,aAAa;AAC9B;;;;AAIG;AACK,IAAA,OAAgB,mBAAmB,GAAG,2BAA2B;AAEzE,IAAA,WAAA,CAAY,UAA8C,EAAE,EAAA;AAC1D;;;;;;;;AAQG;AACH,QAAA,MAAM,SAAS,GACb,OAAO,CAAC,aAAa,IAAI,IAAI,IAAI,OAAO,CAAC,aAAa,GAAG;cACrD,OAAO,CAAC;cACRA,qCAA0B;AAChC;;;;;;;AAOG;AACH,QAAA,MAAM,QAAQ,GACZ,OAAO,CAAC,YAAY,IAAI,IAAI,IAAI,OAAO,CAAC,YAAY,GAAG;cACnD,IAAI,CAAC,GAAG,CAAC,OAAO,CAAC,YAAY,EAAEC,0CAA+B;AAChE,cAAEC,0CAA+B,CAAC,SAAS,CAAC;AAChD,QAAA,IAAI,CAAC,YAAY,GAAG,QAAQ;AAC5B;;;;;;;AAOG;QACH,IAAI,CAAC,aAAa,GAAG,IAAI,CAAC,GAAG,CAAC,SAAS,EAAE,QAAQ,CAAC;AAClD,QAAA,IAAI,CAAC,aAAa;YAChB,OAAO,CAAC,aAAa,IAAI,IAAI,IAAI,OAAO,CAAC,aAAa,GAAG;kBACrD,OAAO,CAAC;kBACR,uBAAuB;IAC/B;AAEQ,IAAA,MAAM,CAAC,KAAyB,EAAA;QACtC,OAAO,KAAK,IAAI,YAAY;IAC9B;AAEQ,IAAA,WAAW,CAAC,KAAyB,EAAA;QAC3C,MAAM,GAAG,GAAG,IAAI,CAAC,MAAM,CAAC,KAAK,CAAC;QAC9B,IAAI,KAAK,GAAG,IAAI,CAAC,SAAS,CAAC,GAAG,CAAC,GAAG,CAAC;AACnC,QAAA,IAAI,KAAK,IAAI,IAAI,EAAE;AACjB,YAAA,KAAK,GAAG,IAAI,cAAc,EAAE;YAC5B,IAAI,CAAC,SAAS,CAAC,GAAG,CAAC,GAAG,EAAE,KAAK,CAAC;YAC9B,IAAI,IAAI,CAAC,SAAS,CAAC,IAAI,GAAG,IAAI,CAAC,aAAa,EAAE;AAC5C,gBAAA,MAAM,MAAM,GAAG,IAAI,CAAC,SAAS,CAAC,IAAI,EAAE,CAAC,IAAI,EAAE,CAAC,KAAK;gBACjD,IAAI,MAAM,IAAI,IAAI,IAAI,MAAM,KAAK,GAAG,EAAE;AACpC,oBAAA,IAAI,CAAC,SAAS,CAAC,MAAM,CAAC,MAAM,CAAC;gBAC/B;YACF;QACF;AACA,QAAA,OAAO,KAAK;IACd;;AAGA,IAAA,GAAG,CAAC,KAAyB,EAAE,GAAW,EAAE,KAAa,EAAA;QACvD,MAAM,MAAM,GAAG,IAAI,CAAC,WAAW,CAAC,KAAK,CAAC;QACtC,MAAM,OAAO,GACX,KAAK,CAAC,MAAM,GAAG,IAAI,CAAC;cAChB,KAAK,CAAC,KAAK,CAAC,CAAC,EAAE,IAAI,CAAC,aAAa;cACjC,KAAK;QACX,MAAM,QAAQ,GAAG,MAAM,CAAC,OAAO,CAAC,GAAG,CAAC,GAAG,CAAC;AACxC,QAAA,IAAI,QAAQ,IAAI,IAAI,EAAE;AACpB,YAAA,MAAM,CAAC,SAAS,IAAI,QAAQ,CAAC,MAAM;AACnC,YAAA,MAAM,CAAC,OAAO,CAAC,MAAM,CAAC,GAAG,CAAC;QAC5B;QACA,MAAM,CAAC,OAAO,CAAC,GAAG,CAAC,GAAG,EAAE,OAAO,CAAC;AAChC,QAAA,MAAM,CAAC,SAAS,IAAI,OAAO,CAAC,MAAM;AAClC,QAAA,IAAI,CAAC,iBAAiB,CAAC,MAAM,CAAC;IAChC;;IAGA,GAAG,CAAC,KAAyB,EAAE,GAAW,EAAA;QACxC,OAAO,IAAI,CAAC,SAAS,CAAC,GAAG,CAAC,IAAI,CAAC,MAAM,CAAC,KAAK,CAAC,CAAC,EAAE,OAAO,CAAC,GAAG,CAAC,GAAG,CAAC;IACjE;AAEA;;;;;AAKG;IACH,GAAG,CAAC,KAAyB,EAAE,GAAW,EAAA;QACxC,OAAO,IAAI,CAAC,SAAS,CAAC,GAAG,CAAC,IAAI,CAAC,MAAM,CAAC,KAAK,CAAC,CAAC,EAAE,OAAO,CAAC,GAAG,CAAC,GAAG,CAAC,IAAI,KAAK;IAC1E;;AAGA,IAAA,IAAI,IAAI,GAAA;QACN,IAAI,CAAC,GAAG,CAAC;QACT,KAAK,MAAM,MAAM,IAAI,IAAI,CAAC,SAAS,CAAC,MAAM,EAAE,EAAE;AAC5C,YAAA,CAAC,IAAI,MAAM,CAAC,OAAO,CAAC,IAAI;QAC1B;AACA,QAAA,OAAO,CAAC;IACV;;AAGA,IAAA,IAAI,cAAc,GAAA;QAChB,OAAO,IAAI,CAAC,aAAa;IAC3B;;AAGA,IAAA,IAAI,UAAU,GAAA;QACZ,OAAO,IAAI,CAAC,YAAY;IAC1B;;IAGA,KAAK,GAAA;AACH,QAAA,IAAI,CAAC,SAAS,CAAC,KAAK,EAAE;IACxB;AAEA;;;;;AAKG;AACH,IAAA,UAAU,CAAC,KAAyB,EAAA;AAClC,QAAA,IAAI,CAAC,SAAS,CAAC,MAAM,CAAC,IAAI,CAAC,MAAM,CAAC,KAAK,CAAC,CAAC;IAC3C;AAEA;;;;;;;;;;AAUG;AACH,IAAA,QAAQ,CAAC,KAAyB,EAAA;AAChC,QAAA,IAAI,KAAK,IAAI,IAAI,EAAE;AACjB,YAAA,IAAI,CAAC,SAAS,CAAC,MAAM,CAAC,YAAY,CAAC;QACrC;QACA,MAAM,MAAM,GAAG,IAAI,CAAC,WAAW,CAAC,KAAK,CAAC;AACtC,QAAA,OAAO,MAAM,CAAC,WAAW,EAAE;IAC7B;AAEA;;;;;AAKG;IACH,aAAa,CAAC,KAAyB,EAAE,QAAgB,EAAA;QACvD,MAAM,MAAM,GAAG,IAAI,CAAC,WAAW,CAAC,KAAK,CAAC;QACtC,IAAI,MAAM,CAAC,oBAAoB,CAAC,GAAG,CAAC,QAAQ,CAAC,EAAE;AAC7C,YAAA,OAAO,KAAK;QACd;AACA,QAAA,MAAM,CAAC,oBAAoB,CAAC,GAAG,CAAC,QAAQ,CAAC;AACzC,QAAA,OAAO,IAAI;IACb;AAEA;;;;;;;;AAQG;IACH,OAAO,CAAI,KAAyB,EAAE,IAAO,EAAA;AAC3C,QAAA,IAAI,CAAC,iBAAiB,CAAC,IAAI,CAAC,EAAE;YAC5B,OAAO,EAAE,QAAQ,EAAE,IAAI,EAAE,UAAU,EAAE,EAAE,EAAE;QAC3C;AACA,QAAA,MAAM,MAAM,GAAG,IAAI,CAAC,SAAS,CAAC,GAAG,CAAC,IAAI,CAAC,MAAM,CAAC,KAAK,CAAC,CAAC;AACrD,QAAA,OAAO,IAAI,CAAC,cAAc,CAAC,MAAM,EAAE,OAAO,IAAI,aAAa,EAAE,IAAI,CAAC;IACpE;AAEA;;;;;;;;;;AAUG;AACH,IAAA,QAAQ,CAAC,KAAyB,EAAA;AAChC,QAAA,MAAM,MAAM,GAAG,IAAI,CAAC,SAAS,CAAC,GAAG,CAAC,IAAI,CAAC,MAAM,CAAC,KAAK,CAAC,CAAC;QACrD,MAAM,OAAO,GAAgC;AAC3C,cAAE,IAAI,GAAG,CAAC,MAAM,CAAC,OAAO;cACtB,aAAa;QACjB,OAAO;AACL,YAAA,OAAO,EAAE,CAAI,IAAO,KAClB,IAAI,CAAC,cAAc,CAAC,OAAO,EAAE,IAAI,CAAC;SACrC;IACH;IAEQ,cAAc,CACpB,OAAoC,EACpC,IAAO,EAAA;AAEP,QAAA,IAAI,CAAC,iBAAiB,CAAC,IAAI,CAAC,EAAE;YAC5B,OAAO,EAAE,QAAQ,EAAE,IAAI,EAAE,UAAU,EAAE,EAAE,EAAE;QAC3C;AACA,QAAA,MAAM,UAAU,GAAG,IAAI,GAAG,EAAU;AACpC,QAAA,MAAM,QAAQ,GAAG,IAAI,CAAC,SAAS,CAAC,OAAO,EAAE,IAAI,EAAE,UAAU,CAAM;AAC/D,QAAA,OAAO,EAAE,QAAQ,EAAE,UAAU,EAAE,KAAK,CAAC,IAAI,CAAC,UAAU,CAAC,EAAE;IACzD;AAEQ,IAAA,SAAS,CACf,OAAoC,EACpC,KAAc,EACd,UAAuB,EAAA;AAEvB,QAAA,IAAI,OAAO,KAAK,KAAK,QAAQ,EAAE;YAC7B,OAAO,IAAI,CAAC,eAAe,CAAC,OAAO,EAAE,KAAK,EAAE,UAAU,CAAC;QACzD;AACA,QAAA,IAAI,KAAK,CAAC,OAAO,CAAC,KAAK,CAAC,EAAE;YACxB,OAAO,KAAK,CAAC,GAAG,CAAC,CAAC,IAAI,KAAK,IAAI,CAAC,SAAS,CAAC,OAAO,EAAE,IAAI,EAAE,UAAU,CAAC,CAAC;QACvE;QACA,IAAI,KAAK,KAAK,IAAI,IAAI,OAAO,KAAK,KAAK,QAAQ,EAAE;YAC/C,MAAM,MAAM,GAAG,KAAgC;YAC/C,MAAM,IAAI,GAA4B,EAAE;AACxC,YAAA,KAAK,MAAM,CAAC,GAAG,EAAE,IAAI,CAAC,IAAI,MAAM,CAAC,OAAO,CAAC,MAAM,CAAC,EAAE;AAChD,gBAAA,IAAI,CAAC,GAAG,CAAC,GAAG,IAAI,CAAC,SAAS,CAAC,OAAO,EAAE,IAAI,EAAE,UAAU,CAAC;YACvD;AACA,YAAA,OAAO,IAAI;QACb;AACA,QAAA,OAAO,KAAK;IACd;AAEQ,IAAA,eAAe,CACrB,OAAoC,EACpC,KAAa,EACb,UAAuB,EAAA;QAEvB,IAAI,KAAK,CAAC,OAAO,CAAC,QAAQ,CAAC,KAAK,EAAE,EAAE;AAClC,YAAA,OAAO,KAAK;QACd;AACA,QAAA,OAAO,KAAK,CAAC,OAAO,CAClB,2BAA2B,CAAC,mBAAmB,EAC/C,CAAC,KAAK,EAAE,GAAW,KAAI;YACrB,MAAM,MAAM,GAAG,OAAO,CAAC,GAAG,CAAC,GAAG,CAAC;AAC/B,YAAA,IAAI,MAAM,IAAI,IAAI,EAAE;AAClB,gBAAA,UAAU,CAAC,GAAG,CAAC,GAAG,CAAC;AACnB,gBAAA,OAAO,KAAK;YACd;AACA,YAAA,OAAO,MAAM;AACf,QAAA,CAAC,CACF;IACH;AAEQ,IAAA,iBAAiB,CAAC,MAAsB,EAAA;QAC9C,IAAI,MAAM,CAAC,SAAS,IAAI,IAAI,CAAC,YAAY,EAAE;YACzC;QACF;QACA,KAAK,MAAM,GAAG,IAAI,MAAM,CAAC,OAAO,CAAC,IAAI,EAAE,EAAE;YACvC,IAAI,MAAM,CAAC,SAAS,IAAI,IAAI,CAAC,YAAY,EAAE;gBACzC;YACF;YACA,MAAM,KAAK,GAAG,MAAM,CAAC,OAAO,CAAC,GAAG,CAAC,GAAG,CAAC;AACrC,YAAA,IAAI,KAAK,IAAI,IAAI,EAAE;gBACjB;YACF;AACA,YAAA,MAAM,CAAC,SAAS,IAAI,KAAK,CAAC,MAAM;AAChC,YAAA,MAAM,CAAC,OAAO,CAAC,MAAM,CAAC,GAAG,CAAC;QAC5B;IACF;;AAGF;;;;AAIG;AACH,SAAS,iBAAiB,CAAC,KAAc,EAAA;AACvC,IAAA,IAAI,OAAO,KAAK,KAAK,QAAQ,EAAE;QAC7B,OAAO,KAAK,CAAC,OAAO,CAAC,QAAQ,CAAC,KAAK,EAAE;IACvC;AACA,IAAA,IAAI,KAAK,CAAC,OAAO,CAAC,KAAK,CAAC,EAAE;AACxB,QAAA,KAAK,MAAM,IAAI,IAAI,KAAK,EAAE;AACxB,YAAA,IAAI,iBAAiB,CAAC,IAAI,CAAC,EAAE;AAC3B,gBAAA,OAAO,IAAI;YACb;QACF;AACA,QAAA,OAAO,KAAK;IACd;IACA,IAAI,KAAK,KAAK,IAAI,IAAI,OAAO,KAAK,KAAK,QAAQ,EAAE;QAC/C,KAAK,MAAM,IAAI,IAAI,MAAM,CAAC,MAAM,CAAC,KAAgC,CAAC,EAAE;AAClE,YAAA,IAAI,iBAAiB,CAAC,IAAI,CAAC,EAAE;AAC3B,gBAAA,OAAO,IAAI;YACb;QACF;AACA,QAAA,OAAO,KAAK;IACd;AACA,IAAA,OAAO,KAAK;AACd;AAEA;;;;;;;;;;;;;;;;;;;;;;;;;AAyBG;AACG,SAAU,+BAA+B,CAC7C,OAAe,EACf,GAAuB,EACvB,aAAuB,EAAE,EAAA;AAEzB,IAAA,MAAM,SAAS,GAAG,GAAG,IAAI,IAAI;AAC7B,IAAA,MAAM,aAAa,GAAG,UAAU,CAAC,MAAM,GAAG,CAAC;AAC3C,IAAA,IAAI,CAAC,SAAS,IAAI,CAAC,aAAa,EAAE;AAChC,QAAA,OAAO,OAAO;IAChB;AACA,IAAA,MAAM,OAAO,GAAG,OAAO,CAAC,SAAS,EAAE;AACnC,IAAA,IAAI,OAAO,CAAC,UAAU,CAAC,GAAG,CAAC,EAAE;QAC3B,MAAM,SAAS,GAAG,0BAA0B,CAAC,OAAO,EAAE,GAAG,EAAE,UAAU,CAAC;AACtE,QAAA,IAAI,SAAS,IAAI,IAAI,EAAE;AACrB,YAAA,OAAO,SAAS;QAClB;IACF;AACA,IAAA,MAAM,MAAM,GAAG,SAAS,GAAG,CAAA,EAAG,oBAAoB,CAAC,GAAI,CAAC,CAAA,EAAA,CAAI,GAAG,EAAE;IACjE,MAAM,OAAO,GAAG;UACZ,uBAAuB,UAAU,CAAC,IAAI,CAAC,IAAI,CAAC,CAAA,CAAA;UAC5C,EAAE;AACN,IAAA,OAAO,GAAG,MAAM,CAAA,EAAG,OAAO,CAAA,EAAG,OAAO,EAAE;AACxC;AAEA,SAAS,0BAA0B,CACjC,OAAe,EACf,GAAuB,EACvB,UAAoB,EAAA;AAEpB,IAAA,IAAI,MAAe;AACnB,IAAA,IAAI;AACF,QAAA,MAAM,GAAG,IAAI,CAAC,KAAK,CAAC,OAAO,CAAC;IAC9B;AAAE,IAAA,MAAM;AACN,QAAA,OAAO,IAAI;IACb;AAEA,IAAA,IAAI,MAAM,KAAK,IAAI,IAAI,OAAO,MAAM,KAAK,QAAQ,IAAI,KAAK,CAAC,OAAO,CAAC,MAAM,CAAC,EAAE;AAC1E,QAAA,OAAO,IAAI;IACb;IAEA,MAAM,GAAG,GAAG,MAAiC;AAC7C,IAAA,MAAM,YAAY,GAAG,GAAG,IAAI,IAAI;AAChC,IAAA,MAAM,mBAAmB,GAAG,UAAU,CAAC,MAAM,GAAG,CAAC;AAEjD;;;;;;;;AAQG;AACH,IAAA,IACE,YAAY;AACZ,QAAA,mBAAmB,IAAI,GAAG;AAC1B,QAAA,GAAG,CAAC,mBAAmB,CAAC,KAAK,GAAG;AAChC,QAAA,GAAG,CAAC,mBAAmB,CAAC,IAAI,IAAI,EAChC;AACA,QAAA,OAAO,IAAI;IACb;AACA,IAAA,IACE,mBAAmB;AACnB,QAAA,0BAA0B,IAAI,GAAG;AACjC,QAAA,GAAG,CAAC,0BAA0B,CAAC,IAAI,IAAI;QACvC,CAAC,kBAAkB,CAAC,GAAG,CAAC,0BAA0B,CAAC,EAAE,UAAU,CAAC,EAChE;AACA,QAAA,OAAO,IAAI;IACb;AAEA;;;;;;;AAOG;AACH,IAAA,MAAM,QAAQ,GAAG,IAAI,GAAG,EAAU;AAClC,IAAA,IAAI,YAAY;AAAE,QAAA,QAAQ,CAAC,GAAG,CAAC,mBAAmB,CAAC;AACnD,IAAA,IAAI,mBAAmB;AAAE,QAAA,QAAQ,CAAC,GAAG,CAAC,0BAA0B,CAAC;IACjE,MAAM,IAAI,GAA4B,EAAE;AACxC,IAAA,KAAK,MAAM,CAAC,CAAC,EAAE,CAAC,CAAC,IAAI,MAAM,CAAC,OAAO,CAAC,GAAG,CAAC,EAAE;QACxC,IAAI,CAAC,QAAQ,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE;AACpB,YAAA,IAAI,CAAC,CAAC,CAAC,GAAG,CAAC;QACb;IACF;IACA,MAAM,QAAQ,GAA4B,EAAE;IAC5C,IAAI,YAAY,EAAE;AAChB,QAAA,QAAQ,CAAC,mBAAmB,CAAC,GAAG,GAAG;IACrC;IACA,IAAI,mBAAmB,EAAE;AACvB,QAAA,QAAQ,CAAC,0BAA0B,CAAC,GAAG,UAAU;IACnD;AACA,IAAA,MAAM,CAAC,MAAM,CAAC,QAAQ,EAAE,IAAI,CAAC;IAE7B,MAAM,MAAM,GAAG,UAAU,CAAC,IAAI,CAAC,OAAO,CAAC;IACvC,OAAO,MAAM,GAAG,IAAI,CAAC,SAAS,CAAC,QAAQ,EAAE,IAAI,EAAE,CAAC,CAAC,GAAG,IAAI,CAAC,SAAS,CAAC,QAAQ,CAAC;AAC9E;AAEA,SAAS,kBAAkB,CAAC,CAAU,EAAE,CAAoB,EAAA;AAC1D,IAAA,IAAI,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,MAAM,KAAK,CAAC,CAAC,MAAM,EAAE;AAC9C,QAAA,OAAO,KAAK;IACd;AACA,IAAA,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE;QACjC,IAAI,CAAC,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,EAAE;AACjB,YAAA,OAAO,KAAK;QACd;IACF;AACA,IAAA,OAAO,IAAI;AACb;AAEA;;;;;;;;;;;;;;;;;;;;;AAqBG;SACa,sBAAsB,CACpC,QAAuB,EACvB,QAAiD,EACjD,KAAyB,EAAA;IAEzB,IAAI,QAAQ,IAAI,IAAI;AAAE,QAAA,OAAO,QAAQ;AAErC;;;;AAIG;AACH,IAAA,IAAI,GAA8B;AAClC,IAAA,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,QAAQ,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE;AACxC,QAAA,MAAM,CAAC,GAAG,QAAQ,CAAC,CAAC,CAAC;AACrB,QAAA,IAAI,CAAC,CAAC,QAAQ,EAAE,KAAK,MAAM;YAAE;AAC7B;;;;;;;;;AASG;AACH,QAAA,MAAM,OAAO,GAAG,CAAC,CAAC,iBAA4B;AAC9C,QAAA,IAAI,OAAO,IAAI,IAAI,IAAI,OAAO,OAAO,KAAK,QAAQ;YAAE;QACpD,MAAM,IAAI,GAAG,OAAkC;AAC/C,QAAA,MAAM,SAAS,GAAG,SAAS,IAAI,IAAI;AACnC,QAAA,MAAM,WAAW,GAAG,WAAW,IAAI,IAAI;AACvC,QAAA,MAAM,kBAAkB,GAAG,iBAAiB,IAAI,IAAI;AACpD,QAAA,IAAI,CAAC,SAAS,IAAI,CAAC,WAAW,IAAI,CAAC,kBAAkB;YAAE;AAEvD,QAAA,MAAM,MAAM,GAAG,UAAU,CAAC,IAAI,CAAC;AAC/B,QAAA,MAAM,UAAU,GAAG,kBAAkB,CAAC,IAAI,CAAC;AAE3C;;;;;;;AAOG;QACH,MAAM,WAAW,GAAG,YAAY,CAAC,IAAI,CAAC,IAAI,KAAK;QAC/C,MAAM,OAAO,GACX,MAAM,IAAI,IAAI,IAAI,QAAQ,CAAC,GAAG,CAAC,WAAW,EAAE,MAAM,CAAC,GAAG,MAAM,GAAG,SAAS;QAC1E,MAAM,SAAS,GAAG,OAAO,IAAI,IAAI,IAAI,UAAU,CAAC,MAAM,GAAG,CAAC;QAE1D,MAAM,EAAE,GAAG,CAAgB;AAC3B,QAAA,IAAI,WAAW,GAA2B,EAAE,CAAC,OAAO;QAEpD,IAAI,SAAS,IAAI,OAAO,EAAE,CAAC,OAAO,KAAK,QAAQ,EAAE;YAC/C,WAAW,GAAG,+BAA+B,CAC3C,EAAE,CAAC,OAAO,EACV,OAAO,EACP,UAAU,CACX;QACH;AAAO,aAAA,IACL,SAAS;AACT,YAAA,KAAK,CAAC,OAAO,CAAC,EAAE,CAAC,OAAO,CAAC;AACzB,YAAA,UAAU,CAAC,MAAM,GAAG,CAAC,EACrB;AACA,YAAA,MAAM,YAAY,GAAG;AACnB,gBAAA,IAAI,EAAE,MAAe;gBACrB,IAAI,EAAE,qBAAqB,UAAU,CAAC,IAAI,CAAC,IAAI,CAAC,CAAA,CAAA,CAAG;aACpD;AACD;;;;;;;;AAQG;AACH,YAAA,WAAW,GAAG;gBACZ,YAAY;gBACZ,GAAG,EAAE,CAAC,OAAO;aACuB;QACxC;AAEA;;;;;AAKG;AACH,QAAA,GAAG,KAAK,QAAQ,CAAC,KAAK,EAAE;QACxB,GAAG,CAAC,CAAC,CAAC,GAAG,2BAA2B,CAAC,EAAE,EAAE,WAAW,CAAC;IACvD;IAEA,OAAO,GAAG,IAAI,QAAQ;AACxB;AAEA;;;;AAIG;AACH,SAAS,UAAU,CACjB,IAAyC,EAAA;AAEzC,IAAA,MAAM,CAAC,GAAG,IAAI,EAAE,OAAO;AACvB,IAAA,OAAO,OAAO,CAAC,KAAK,QAAQ,GAAG,CAAC,GAAG,SAAS;AAC9C;AAEA;;;;;AAKG;AACH,SAAS,YAAY,CACnB,IAAyC,EAAA;AAEzC,IAAA,MAAM,CAAC,GAAG,IAAI,EAAE,SAAS;AACzB,IAAA,OAAO,OAAO,CAAC,KAAK,QAAQ,GAAG,CAAC,GAAG,SAAS;AAC9C;AAEA;;;;;;AAMG;AACH,SAAS,kBAAkB,CACzB,IAAyC,EAAA;AAEzC,IAAA,MAAM,CAAC,GAAG,IAAI,EAAE,eAAe;AAC/B,IAAA,IAAI,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC;AAAE,QAAA,OAAO,EAAE;IAChC,MAAM,GAAG,GAAa,EAAE;AACxB,IAAA,KAAK,MAAM,IAAI,IAAI,CAAC,EAAE;QACpB,IAAI,OAAO,IAAI,KAAK,QAAQ;AAAE,YAAA,GAAG,CAAC,IAAI,CAAC,IAAI,CAAC;IAC9C;AACA,IAAA,OAAO,GAAG;AACZ;AAEA;;;;;;;;;;;;AAYG;AACH,SAAS,2BAA2B,CAClC,EAAe,EACf,OAA+B,EAAA;IAE/B,OAAO,IAAIC,oBAAW,CAAC;QACrB,EAAE,EAAE,EAAE,CAAC,EAAE;QACT,IAAI,EAAE,EAAE,CAAC,IAAI;QACb,MAAM,EAAE,EAAE,CAAC,MAAM;QACjB,QAAQ,EAAE,EAAE,CAAC,QAAQ;QACrB,YAAY,EAAE,EAAE,CAAC,YAAY;QAC7B,iBAAiB,EAAE,EAAE,CAAC,iBAAiB;AACvC,QAAA,iBAAiB,EAAE,yBAAyB,CAAC,EAAE,CAAC,iBAAiB,CAAC;QAClE,OAAO;AACR,KAAA,CAAC;AACJ;AAEA;;;;;;AAMG;AACH,SAAS,yBAAyB,CAChC,MAA2C,EAAA;IAE3C,IAAI,MAAM,IAAI,IAAI;AAAE,QAAA,OAAO,SAAS;AACpC,IAAA,IACE,EAAE,SAAS,IAAI,MAAM,CAAC;AACtB,QAAA,EAAE,WAAW,IAAI,MAAM,CAAC;AACxB,QAAA,EAAE,iBAAiB,IAAI,MAAM,CAAC,EAC9B;AACA,QAAA,OAAO,MAAM;IACf;AACA,IAAA,MAAM,EAAE,OAAO,EAAE,SAAS,EAAE,eAAe,EAAE,GAAG,IAAI,EAAE,GAAG,MAOxD;AAID,IAAA,OAAO,MAAM,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC,MAAM,KAAK,CAAC,GAAG,SAAS,GAAG,IAAI;AAC1D;;;;;;;;;;"}

package/dist/cjs/types/agent-cache.cjs

@@ -0,0 +1,53 @@
+'use strict';
+
+/**
+ * Agent system-message prompt-cache types.
+ * ==========================================
+ *
+ * `@illuma-ai/agents` exposes a generic primitive — `system_cache_blocks` —
+ * for composing a system message that gets explicit cache markers
+ * (cachePoint on Bedrock, cache_control on Anthropic) at consumer-defined
+ * boundaries. The library encodes each entry into the right wire format
+ * for the active provider; consumers decide what content to put in each
+ * block.
+ *
+ * Why this is generic rather than ranger-specific
+ * ------------------------------------------------
+ * Earlier iterations exposed `platform_instructions` and `cache_ttl:
+ * { platform, agent }` — vocabulary borrowed from a multi-tenant chat
+ * platform's mental model. That leaked a specific consumer's domain into
+ * the library and made it confusing for other consumers (CLI tools,
+ * desktop apps, gateways) which have no notion of "platforms" or "agents".
+ *
+ * The current shape lets every consumer define their own cache-tier
+ * organization without the library knowing or caring.
+ *
+ * Cache key composition
+ * ---------------------
+ * Both Anthropic and Bedrock hash the message bytes leading up to each
+ * cache marker. So entry N's cache key is the concatenation of bytes
+ * 0..N. Stable content (universal rules, framework docs) goes earlier;
+ * volatile content (per-agent identity, per-tool docs) goes later. The
+ * earlier entries get cross-tenant cache hits; the later entries get
+ * narrower hits.
+ *
+ * Provider budget
+ * ---------------
+ * Bedrock's Converse API supports up to 4 cachePoint blocks per request.
+ * The default tools-array cache strategy (see `IllumaBedrockConverse`)
+ * uses up to 2 of those. So at most 2 entries should appear in
+ * `system_cache_blocks` to leave budget for the trailing
+ * `instructions` block. The library throws on more than 2.
+ *
+ * Anthropic supports up to 4 cache breakpoints per request workspace-wide;
+ * the same constraint applies to keep tools + system within budget.
+ */
+/**
+ * Maximum number of `system_cache_blocks` entries the library accepts.
+ * Higher than this would exceed Bedrock's 4-cachePoint budget once the
+ * tools array and trailing `instructions` block are counted.
+ */
+const MAX_SYSTEM_CACHE_BLOCKS = 2;
+
+exports.MAX_SYSTEM_CACHE_BLOCKS = MAX_SYSTEM_CACHE_BLOCKS;
+//# sourceMappingURL=agent-cache.cjs.map

package/dist/cjs/types/agent-cache.cjs.map

@@ -0,0 +1 @@
+{"version":3,"file":"agent-cache.cjs","sources":["../../../src/types/agent-cache.ts"],"sourcesContent":["/**\n * Agent system-message prompt-cache types.\n * ==========================================\n *\n * `@illuma-ai/agents` exposes a generic primitive — `system_cache_blocks` —\n * for composing a system message that gets explicit cache markers\n * (cachePoint on Bedrock, cache_control on Anthropic) at consumer-defined\n * boundaries. The library encodes each entry into the right wire format\n * for the active provider; consumers decide what content to put in each\n * block.\n *\n * Why this is generic rather than ranger-specific\n * ------------------------------------------------\n * Earlier iterations exposed `platform_instructions` and `cache_ttl:\n * { platform, agent }` — vocabulary borrowed from a multi-tenant chat\n * platform's mental model. That leaked a specific consumer's domain into\n * the library and made it confusing for other consumers (CLI tools,\n * desktop apps, gateways) which have no notion of \"platforms\" or \"agents\".\n *\n * The current shape lets every consumer define their own cache-tier\n * organization without the library knowing or caring.\n *\n * Cache key composition\n * ---------------------\n * Both Anthropic and Bedrock hash the message bytes leading up to each\n * cache marker. So entry N's cache key is the concatenation of bytes\n * 0..N. Stable content (universal rules, framework docs) goes earlier;\n * volatile content (per-agent identity, per-tool docs) goes later. The\n * earlier entries get cross-tenant cache hits; the later entries get\n * narrower hits.\n *\n * Provider budget\n * ---------------\n * Bedrock's Converse API supports up to 4 cachePoint blocks per request.\n * The default tools-array cache strategy (see `IllumaBedrockConverse`)\n * uses up to 2 of those. So at most 2 entries should appear in\n * `system_cache_blocks` to leave budget for the trailing\n * `instructions` block. The library throws on more than 2.\n *\n * Anthropic supports up to 4 cache breakpoints per request workspace-wide;\n * the same constraint applies to keep tools + system within budget.\n */\n\n/** Provider TTL hint applied to a cachePoint / cache_control marker. */\nexport type AgentCacheTTL = '5m' | '1h';\n\n/**\n * One cacheable system content block. Each entry produces a text block\n * followed by a provider-specific cache marker in the assembled system\n * message. Entries are emitted in array order — earlier = more stable\n * = wider cache key.\n */\nexport interface SystemCacheBlock {\n  /**\n   * Text payload. Caller is responsible for keeping this byte-stable\n   * across requests that should share a cache entry. ANY change to\n   * these bytes invalidates every cache key that includes this entry.\n   */\n  text: string;\n  /**\n   * Optional TTL hint for the cache marker. Defaults to '5m' when\n   * omitted. '1h' costs more to write (2× base vs 1.25×) but survives\n   * longer idle, which is the right tradeoff for low-traffic blocks.\n   */\n  ttl?: AgentCacheTTL;\n}\n\n/**\n * Maximum number of `system_cache_blocks` entries the library accepts.\n * Higher than this would exceed Bedrock's 4-cachePoint budget once the\n * tools array and trailing `instructions` block are counted.\n */\nexport const MAX_SYSTEM_CACHE_BLOCKS = 2;\n"],"names":[],"mappings":";;AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAyCG;AA0BH;;;;AAIG;AACI,MAAM,uBAAuB,GAAG;;;;"}

package/dist/cjs/types/graph.cjs.map

@@ -1 +1 @@
-{"version":3,"file":"graph.cjs","sources":["../../../src/types/graph.ts"],"sourcesContent":["// src/types/graph.ts\nimport type {\n  START,\n  StateType,\n  UpdateType,\n  StateGraph,\n  StateGraphArgs,\n  StateDefinition,\n  CompiledStateGraph,\n  BinaryOperatorAggregate,\n} from '@langchain/langgraph';\nimport type { BindToolsInput } from '@langchain/core/language_models/chat_models';\nimport type {\n  BaseMessage,\n  AIMessageChunk,\n  SystemMessage,\n} from '@langchain/core/messages';\nimport type { RunnableConfig, Runnable } from '@langchain/core/runnables';\nimport type { ChatGenerationChunk } from '@langchain/core/outputs';\nimport type { GoogleAIToolType } from '@langchain/google-common';\nimport type {\n  ToolMap,\n  ToolEndEvent,\n  GenericTool,\n  LCTool,\n  ToolApprovalConfig,\n} from '@/types/tools';\nimport type { Providers, Callback, GraphNodeKeys } from '@/common';\nimport type { StandardGraph, MultiAgentGraph } from '@/graphs';\nimport type { ClientOptions } from '@/types/llm';\nimport type {\n  RunStep,\n  RunStepDeltaEvent,\n  MessageDeltaEvent,\n  ReasoningDeltaEvent,\n} from '@/types/stream';\nimport type { TokenCounter } from '@/types/run';\n\n/** Interface for bound model with stream and invoke methods */\nexport interface ChatModel {\n  stream?: (\n    messages: BaseMessage[],\n    config?: RunnableConfig\n  ) => Promise<AsyncIterable<AIMessageChunk>>;\n  invoke: (\n    messages: BaseMessage[],\n    config?: RunnableConfig\n  ) => Promise<AIMessageChunk>;\n}\n\n/** Payload for ON_AGENT_TRANSITION events */\nexport type AgentTransitionEvent = {\n  sourceAgentId?: string;\n  sourceAgentName?: string;\n  destinationAgentId: string;\n  destinationAgentName: string;\n  edgeType: string; // 'handoff' | 'transfer' | 'sequence'\n  timestamp: number;\n  /** When true, this event signals handoff completion (child → parent return) */\n  isCompletion?: boolean;\n  /** Duration of child agent execution in milliseconds (only on completion events) */\n  durationMs?: number;\n  /** Length of child agent result text in characters (only on completion events) */\n  resultLength?: number;\n};\n\nexport type GraphNode = GraphNodeKeys | typeof START;\nexport type ClientCallback<T extends unknown[]> = (\n  graph: StandardGraph,\n  ...args: T\n) => void;\n\nexport type ClientCallbacks = {\n  [Callback.TOOL_ERROR]?: ClientCallback<[Error, string]>;\n  [Callback.TOOL_START]?: ClientCallback<unknown[]>;\n  [Callback.TOOL_END]?: ClientCallback<unknown[]>;\n};\n\nexport type SystemCallbacks = {\n  [K in keyof ClientCallbacks]: ClientCallbacks[K] extends ClientCallback<\n    infer Args\n  >\n    ? (...args: Args) => void\n    : never;\n};\n\nexport type BaseGraphState = {\n  messages: BaseMessage[];\n  /**\n   * Structured response when using structured output mode.\n   * Contains the validated JSON response conforming to the configured schema.\n   */\n  structuredResponse?: Record<string, unknown>;\n};\n\nexport type MultiAgentGraphState = BaseGraphState & {\n  agentMessages?: BaseMessage[];\n};\n\nexport type IState = BaseGraphState;\n\nexport interface EventHandler {\n  handle(\n    event: string,\n    data:\n      | StreamEventData\n      | ModelEndData\n      | RunStep\n      | RunStepDeltaEvent\n      | MessageDeltaEvent\n      | ReasoningDeltaEvent\n      | { result: ToolEndEvent },\n    metadata?: Record<string, unknown>,\n    graph?: StandardGraph | MultiAgentGraph\n  ): void | Promise<void>;\n}\n\nexport type GraphStateChannels<T extends BaseGraphState> =\n  StateGraphArgs<T>['channels'];\n\nexport type Workflow<\n  T extends BaseGraphState = BaseGraphState,\n  U extends Partial<T> = Partial<T>,\n  N extends string = string,\n> = StateGraph<T, U, N>;\n\nexport type CompiledWorkflow<\n  T extends BaseGraphState = BaseGraphState,\n  U extends Partial<T> = Partial<T>,\n  N extends string = string,\n> = CompiledStateGraph<T, U, N>;\n\nexport type CompiledStateWorkflow = CompiledStateGraph<\n  StateType<{\n    messages: BinaryOperatorAggregate<BaseMessage[], BaseMessage[]>;\n  }>,\n  UpdateType<{\n    messages: BinaryOperatorAggregate<BaseMessage[], BaseMessage[]>;\n  }>,\n  string,\n  {\n    messages: BinaryOperatorAggregate<BaseMessage[], BaseMessage[]>;\n  },\n  {\n    messages: BinaryOperatorAggregate<BaseMessage[], BaseMessage[]>;\n  },\n  StateDefinition\n>;\n\nexport type CompiledMultiAgentWorkflow = CompiledStateGraph<\n  StateType<{\n    messages: BinaryOperatorAggregate<BaseMessage[], BaseMessage[]>;\n    agentMessages: BinaryOperatorAggregate<BaseMessage[], BaseMessage[]>;\n  }>,\n  UpdateType<{\n    messages: BinaryOperatorAggregate<BaseMessage[], BaseMessage[]>;\n    agentMessages: BinaryOperatorAggregate<BaseMessage[], BaseMessage[]>;\n  }>,\n  string,\n  {\n    messages: BinaryOperatorAggregate<BaseMessage[], BaseMessage[]>;\n    agentMessages: BinaryOperatorAggregate<BaseMessage[], BaseMessage[]>;\n  },\n  {\n    messages: BinaryOperatorAggregate<BaseMessage[], BaseMessage[]>;\n    agentMessages: BinaryOperatorAggregate<BaseMessage[], BaseMessage[]>;\n  },\n  StateDefinition\n>;\n\nexport type CompiledAgentWorfklow = CompiledStateGraph<\n  {\n    messages: BaseMessage[];\n  },\n  {\n    messages?: BaseMessage[] | undefined;\n  },\n  '__start__' | `agent=${string}` | `tools=${string}`,\n  {\n    messages: BinaryOperatorAggregate<BaseMessage[], BaseMessage[]>;\n  },\n  {\n    messages: BinaryOperatorAggregate<BaseMessage[], BaseMessage[]>;\n  },\n  StateDefinition,\n  {\n    [x: `agent=${string}`]: Partial<BaseGraphState>;\n    // eslint-disable-next-line @typescript-eslint/no-explicit-any\n    [x: `tools=${string}`]: any;\n  }\n>;\n\nexport type SystemRunnable =\n  | Runnable<\n      BaseMessage[],\n      (BaseMessage | SystemMessage)[],\n      RunnableConfig<Record<string, unknown>>\n    >\n  | undefined;\n\n/**\n * Optional compile options passed to workflow.compile().\n * These are intentionally untyped to avoid coupling to library internals.\n */\nexport type CompileOptions = {\n  // A checkpointer instance (e.g., MemorySaver, SQL saver)\n  // eslint-disable-next-line @typescript-eslint/no-explicit-any\n  checkpointer?: any;\n  interruptBefore?: string[];\n  interruptAfter?: string[];\n  /**\n   * Human-in-the-loop tool approval configuration.\n   * When set, tools matching the policy will trigger an interrupt()\n   * before execution, pausing the graph for human approval.\n   * Requires a checkpointer to be set for interrupt/resume to work.\n   */\n  toolApprovalConfig?: ToolApprovalConfig;\n};\n\nexport type EventStreamCallbackHandlerInput =\n  Parameters<CompiledWorkflow['streamEvents']>[2] extends Omit<\n    infer T,\n    'autoClose'\n  >\n    ? T\n    : never;\n\nexport type StreamChunk =\n  | (ChatGenerationChunk & {\n      message: AIMessageChunk;\n    })\n  | AIMessageChunk;\n\n/**\n * Data associated with a StreamEvent.\n */\nexport type StreamEventData = {\n  /**\n   * The input passed to the runnable that generated the event.\n   * Inputs will sometimes be available at the *START* of the runnable, and\n   * sometimes at the *END* of the runnable.\n   * If a runnable is able to stream its inputs, then its input by definition\n   * won't be known until the *END* of the runnable when it has finished streaming\n   * its inputs.\n   */\n  input?: unknown;\n  /**\n   * The output of the runnable that generated the event.\n   * Outputs will only be available at the *END* of the runnable.\n   * For most runnables, this field can be inferred from the `chunk` field,\n   * though there might be some exceptions for special cased runnables (e.g., like\n   * chat models), which may return more information.\n   */\n  output?: unknown;\n  /**\n   * A streaming chunk from the output that generated the event.\n   * chunks support addition in general, and adding them up should result\n   * in the output of the runnable that generated the event.\n   */\n  chunk?: StreamChunk;\n  /**\n   * Runnable config for invoking other runnables within handlers.\n   */\n  config?: RunnableConfig;\n  /**\n   * Custom result from the runnable that generated the event.\n   */\n  result?: unknown;\n  /**\n   * Custom field to indicate the event was manually emitted, and may have been handled already\n   */\n  emitted?: boolean;\n};\n\n/**\n * A streaming event.\n *\n * Schema of a streaming event which is produced from the streamEvents method.\n */\nexport type StreamEvent = {\n  /**\n   * Event names are of the format: on_[runnable_type]_(start|stream|end).\n   *\n   * Runnable types are one of:\n   * - llm - used by non chat models\n   * - chat_model - used by chat models\n   * - prompt --  e.g., ChatPromptTemplate\n   * - tool -- LangChain tools\n   * - chain - most Runnables are of this type\n   *\n   * Further, the events are categorized as one of:\n   * - start - when the runnable starts\n   * - stream - when the runnable is streaming\n   * - end - when the runnable ends\n   *\n   * start, stream and end are associated with slightly different `data` payload.\n   *\n   * Please see the documentation for `EventData` for more details.\n   */\n  event: string;\n  /** The name of the runnable that generated the event. */\n  name: string;\n  /**\n   * An randomly generated ID to keep track of the execution of the given runnable.\n   *\n   * Each child runnable that gets invoked as part of the execution of a parent runnable\n   * is assigned its own unique ID.\n   */\n  run_id: string;\n  /**\n   * Tags associated with the runnable that generated this event.\n   * Tags are always inherited from parent runnables.\n   */\n  tags?: string[];\n  /** Metadata associated with the runnable that generated this event. */\n  metadata: Record<string, unknown>;\n  /**\n   * Event data.\n   *\n   * The contents of the event data depend on the event type.\n   */\n  data: StreamEventData;\n};\n\nexport type GraphConfig = {\n  provider: string;\n  thread_id?: string;\n  run_id?: string;\n};\n\nexport type PartMetadata = {\n  progress?: number;\n  asset_pointer?: string;\n  status?: string;\n  action?: boolean;\n  output?: string;\n};\n\nexport type ModelEndData =\n  | (StreamEventData & { output: AIMessageChunk | undefined })\n  | undefined;\nexport type GraphTools = GenericTool[] | BindToolsInput[] | GoogleAIToolType[];\nexport type StandardGraphInput = {\n  runId?: string;\n  signal?: AbortSignal;\n  agents: AgentInputs[];\n  tokenCounter?: TokenCounter;\n  indexTokenCountMap?: Record<string, number>;\n};\n\n/**\n * Configuration for an approval gate placed on a sequence edge.\n * When present, the graph inserts an approval gate node between the source\n * and destination agents. The gate ALWAYS fires (regardless of ExecutionContext)\n * and calls interrupt() to pause the graph for human approval.\n */\nexport type ApprovalGateConfig = {\n  /** Unique identifier for this gate (used as node ID suffix) */\n  gateId: string;\n  /**\n   * Approval channel — where the approval UI is rendered.\n   * - 'chat': SSE-based chat UI (default)\n   * - 'outlook': MS Graph Actionable Messages\n   * - 'telegram': Telegram Bot inline keyboard\n   */\n  channel?: 'chat' | 'outlook' | 'telegram';\n  /** Optional human-readable prompt shown to the approver */\n  prompt?: string;\n  /** Optional approver identifier (e.g., email, user ID) */\n  approver?: string;\n  /** Timeout in ms before the gate auto-expires (default: 5 minutes) */\n  timeoutMs?: number;\n  /** What to do on denial: 'stop' ends the graph, 'skip' skips the destination agent */\n  onDeny?: 'stop' | 'skip';\n};\n\nexport type GraphEdge = {\n  /** Agent ID, use a list for multiple sources */\n  from: string | string[];\n  /** Agent ID, use a list for multiple destinations */\n  to: string | string[];\n  description?: string;\n  /** Can return boolean or specific destination(s) */\n  condition?: (state: BaseGraphState) => boolean | string | string[];\n  /**\n   * EdgeType.HANDOFF — true agent handoff, parent calls child inline and gets result back.\n   * EdgeType.TRANSFER — one-way transfer, parent exits and child takes over.\n   * EdgeType.SEQUENCE — fixed graph edges for sequential/parallel transitions.\n   */\n  edgeType?: import('@/common').EdgeType;\n  /**\n   * For sequence edges: Optional prompt to add when transitioning through this edge.\n   * String prompts can include variables like {results} which will be replaced with\n   * messages from startIndex onwards. When {results} is used, excludeResults defaults to true.\n   *\n   * For transfer edges: Description for the input parameter that the transfer tool accepts,\n   * allowing the supervisor to pass specific instructions/context to the transferred agent.\n   */\n  prompt?:\n    | string\n    | ((\n        messages: BaseMessage[],\n        runStartIndex: number\n      ) => string | Promise<string> | undefined);\n  /**\n   * When true, excludes messages from startIndex when adding prompt.\n   * Automatically set to true when {results} variable is used in prompt.\n   */\n  excludeResults?: boolean;\n  /**\n   * For transfer edges: Customizes the parameter name for the transfer input.\n   * Defaults to \"instructions\" if not specified.\n   * Only applies when prompt is provided for transfer edges.\n   *\n   * For handoff edges: Customizes the parameter name for the handoff instruction input.\n   */\n  promptKey?: string;\n  /**\n   * For handoff edges: Maximum characters for the result returned to the parent.\n   * Uses head/tail truncation (60/40 split) to preserve key findings and conclusions.\n   * Defaults to DEFAULT_HANDOFF_MAX_RESULT_CHARS (32768 chars, ~8192 tokens).\n   */\n  maxResultChars?: number;\n  /**\n   * For handoff edges: When true, the child agent receives the full parent\n   * conversation history plus the orchestrator's instructions appended.\n   * When false (default), the child only receives the orchestrator's scoped\n   * instructions — isolated from the parent conversation.\n   */\n  passthrough?: boolean;\n  /**\n   * Approval gate configuration for sequence edges.\n   * When set, inserts an approval gate node between source and destination.\n   * The gate ALWAYS fires regardless of ExecutionContext (unlike tool approval).\n   */\n  approvalGate?: ApprovalGateConfig;\n};\n\nexport type MultiAgentGraphInput = StandardGraphInput & {\n  edges: GraphEdge[];\n  /**\n   * When set, the graph routes START to this agent instead of the default starting nodes.\n   * Used for multi-turn resumption: the caller reads `lastActiveAgentId` from the\n   * previous turn's metadata and passes it here so follow-up messages route to the\n   * agent that last handled the conversation.\n   *\n   * If the agent ID is invalid (not in the graph), falls back to default starting nodes.\n   */\n  resumeFromAgentId?: string;\n};\n\n/**\n * Structured output mode determines how the agent returns structured data.\n * - 'tool': Uses tool calling to return structured output (works with all tool-calling models)\n * - 'provider': Uses provider-native structured output via LangChain's jsonMode (OpenAI, Anthropic, etc.)\n * - 'native': Uses provider's constrained decoding API directly for guaranteed schema compliance\n *   (Anthropic output_config.format, OpenAI response_format.json_schema). Falls back to 'tool' for unsupported providers.\n * - 'auto': Automatically selects the best strategy — 'native' for supported providers, 'tool' for others\n */\nexport type StructuredOutputMode = 'tool' | 'provider' | 'native' | 'auto';\n\n/**\n * Resolved method used internally after mode resolution.\n * Maps to LangChain's withStructuredOutput method parameter plus our native path.\n */\nexport type ResolvedStructuredOutputMethod =\n  | 'functionCalling'\n  | 'jsonMode'\n  | 'jsonSchema'\n  | 'native'\n  | undefined;\n\n/**\n * Error thrown when the model refuses to produce structured output due to safety policies.\n */\nexport class StructuredOutputRefusalError extends Error {\n  constructor(public refusalText: string) {\n    super(`Model refused to produce structured output: ${refusalText}`);\n    this.name = 'StructuredOutputRefusalError';\n  }\n}\n\n/**\n * Error thrown when the structured output response was truncated due to max_tokens.\n */\nexport class StructuredOutputTruncatedError extends Error {\n  constructor(public stopReason: string) {\n    super(\n      `Structured output was truncated (stop_reason: ${stopReason}). ` +\n        'Increase max_tokens to allow the full JSON response to be generated.'\n    );\n    this.name = 'StructuredOutputTruncatedError';\n  }\n}\n\n/**\n * Configuration for structured JSON output from agents.\n * When configured, the agent will return a validated JSON response\n * instead of streaming text.\n */\nexport interface StructuredOutputConfig {\n  /**\n   * JSON Schema defining the output structure.\n   * The model will be forced to return data conforming to this schema.\n   */\n  schema: Record<string, unknown>;\n  /**\n   * Name for the structured output format (used in tool mode).\n   * @default 'StructuredResponse'\n   */\n  name?: string;\n  /**\n   * Description of what the structured output represents.\n   * Helps the model understand the expected format.\n   */\n  description?: string;\n  /**\n   * Output mode strategy.\n   * @default 'auto'\n   */\n  mode?: StructuredOutputMode;\n  /**\n   * Enable strict schema validation.\n   * When true, the response must exactly match the schema.\n   * @default true\n   */\n  strict?: boolean;\n  /**\n   * Error handling configuration.\n   * - true: Auto-retry on validation errors (default)\n   * - false: Throw error on validation failure\n   * - string: Custom error message for retry\n   */\n  handleErrors?: boolean | string;\n  /**\n   * Maximum number of retry attempts on validation failure.\n   * @default 2\n   */\n  maxRetries?: number;\n  /**\n   * Include the raw AI message along with structured response.\n   * Useful for debugging.\n   * @default false\n   */\n  includeRaw?: boolean;\n}\n\n/**\n * Database/API structured output format (snake_case with enabled flag).\n * This matches the format stored in MongoDB and sent from frontends.\n */\nexport interface StructuredOutputInput {\n  /** Whether structured output is enabled */\n  enabled?: boolean;\n  /** JSON Schema defining the expected response structure */\n  schema?: Record<string, unknown>;\n  /** Name identifier for the structured output */\n  name?: string;\n  /** Description of what the structured output represents */\n  description?: string;\n  /** Mode for structured output: 'tool' | 'provider' | 'native' | 'auto' */\n  mode?: StructuredOutputMode;\n  /** Whether to enforce strict schema validation */\n  strict?: boolean;\n}\n\n/**\n * Trigger strategy for when summarization should activate.\n * - 'contextPercentage': Trigger when context utilization exceeds a threshold percentage\n * - 'messageCount': Trigger when pruned message count exceeds a threshold\n * - 'tokenThreshold': Trigger when total token count exceeds a raw threshold\n */\nexport type SummarizationTriggerType =\n  | 'contextPercentage'\n  | 'messageCount'\n  | 'tokenThreshold';\n\n/**\n * Configuration for summarization behavior within the agent pipeline.\n * All fields are optional — sensible defaults are provided via constants.\n *\n * @see SUMMARIZATION_CONTEXT_THRESHOLD, SUMMARIZATION_RESERVE_RATIO, PRUNING_EMA_ALPHA\n */\nexport interface SummarizationConfig {\n  /**\n   * Strategy for when summarization triggers.\n   * @default 'contextPercentage'\n   */\n  triggerType?: SummarizationTriggerType;\n\n  /**\n   * Threshold value interpreted based on triggerType:\n   * - contextPercentage: 0-100 (percentage of context window)\n   * - messageCount: absolute count of messages pruned\n   * - tokenThreshold: absolute token count\n   * @default 80 (for contextPercentage)\n   */\n  triggerThreshold?: number;\n\n  /**\n   * Fraction of context window (0-1) reserved for recent messages.\n   * Prevents over-pruning by ensuring at least this fraction of the\n   * context budget is preserved as recent conversation history.\n   * @default 0.3\n   */\n  reserveRatio?: number;\n\n  /**\n   * Whether context pruning is enabled (can be disabled for debugging).\n   * @default true\n   */\n  contextPruning?: boolean;\n\n  /**\n   * Initial summary text to seed across runs.\n   * Different from persistedSummary: this is provided by the caller as a\n   * cross-conversation seed (e.g., agent personality or recurring context),\n   * while persistedSummary is loaded from the conversation's own history.\n   */\n  initialSummary?: string;\n}\n\n/**\n * Runtime state for EMA-based pruning calibration.\n * Maintained across iterations within a single run to smooth pruning decisions.\n */\nexport interface PruneCalibrationState {\n  /** Current EMA calibration ratio */\n  ratio: number;\n  /** Number of calibration updates applied */\n  iterations: number;\n}\n\n/**\n * Lightweight file metadata entry for conversation-level file awareness.\n * Contains only IDs and names — NOT full content — so the agent always knows\n * what files exist in the conversation even after compaction pushes old messages\n * behind the summary window. The agent can retrieve full content on-demand\n * via file_search (RAG) or content_tool read (by contentId).\n */\nexport interface FileManifestEntry {\n  /** Unique file identifier (e.g., MongoDB ObjectId or UUID) */\n  fileId: string;\n  /** Original filename (e.g., \"quarterly-report.pdf\") */\n  filename: string;\n  /** Content identifier for on-demand retrieval via content_tool read */\n  contentId?: string;\n  /** File source (e.g., \"local\", \"sharepoint\", \"onedrive\") */\n  source?: string;\n  /** Index of the message that introduced this file (0-based in the original message array) */\n  messageIndex?: number;\n}\n\nexport interface AgentInputs {\n  agentId: string;\n  /** Human-readable name for the agent (used in handoff context). Defaults to agentId if not provided. */\n  name?: string;\n  /** Description of what this agent does (used to enrich handoff tool descriptions). */\n  description?: string;\n  toolEnd?: boolean;\n  toolMap?: ToolMap;\n  tools?: GraphTools;\n  provider: Providers;\n  instructions?: string;\n  streamBuffer?: number;\n  maxContextTokens?: number;\n  clientOptions?: ClientOptions;\n  additional_instructions?: string;\n  reasoningKey?: 'reasoning_content' | 'reasoning';\n  /** Format content blocks as strings (for legacy compatibility i.e. Ollama/Azure Serverless) */\n  useLegacyContent?: boolean;\n  /**\n   * Tool definitions for all tools, including deferred and programmatic.\n   * Used for tool search and programmatic tool calling.\n   * Maps tool name to LCTool definition.\n   */\n  toolRegistry?: Map<string, LCTool>;\n  /**\n   * Dynamic context that changes per-request (e.g., current time, user info).\n   * This is injected as a user message rather than system prompt to preserve cache.\n   * Keeping this separate from instructions ensures the system message stays static\n   * and can be cached by Bedrock/Anthropic prompt caching.\n   */\n  dynamicContext?: string;\n  /**\n   * Structured output configuration (camelCase).\n   * When set, disables streaming and returns a validated JSON response\n   * conforming to the specified schema.\n   */\n  structuredOutput?: StructuredOutputConfig;\n  /**\n   * Structured output configuration (snake_case - database/API format).\n   * Alternative to structuredOutput for compatibility with MongoDB/frontend.\n   * Uses an `enabled` flag to control activation.\n   * @deprecated Use structuredOutput instead when possible\n   */\n  structured_output?: StructuredOutputInput;\n  /**\n   * Serializable tool definitions for event-driven execution.\n   * When provided, ToolNode operates in event-driven mode, dispatching\n   * ON_TOOL_EXECUTE events instead of invoking tools directly.\n   */\n  toolDefinitions?: LCTool[];\n  /**\n   * Tool names discovered from previous conversation history.\n   * These tools will be pre-marked as discovered so they're included\n   * in tool binding without requiring tool_search.\n   */\n  discoveredTools?: string[];\n  /**\n   * Optional callback for summarizing messages that were pruned from context.\n   * When provided, discarded messages are summarized by the host caller\n   * using a cheap LLM call, and the summary is prepended to the context.\n   */\n  summarizeCallback?: (\n    messagesToRefine: import('@langchain/core/messages').BaseMessage[]\n  ) => Promise<string | undefined>;\n  /**\n   * Pre-existing summary text loaded from persistent storage (MongoDB/Redis).\n   * When provided, this summary is injected into the initial message context\n   * so the agent has prior conversation history even on new turns.\n   * Set by the host's summary store when resuming a conversation.\n   */\n  persistedSummary?: string;\n  /**\n   * Summarization configuration controlling trigger strategy, reserve ratio,\n   * and EMA calibration for pruning. When omitted, sensible defaults apply.\n   * @see SummarizationConfig\n   */\n  summarizationConfig?: SummarizationConfig;\n  /**\n   * Lightweight file manifest for the conversation.\n   * Contains file IDs, names, and metadata — NOT full content.\n   *\n   * Used by the compaction engine to inject a [Conversation Files] block\n   * into the windowed view, ensuring the LLM always knows what files exist\n   * even when old messages (with full file content) are behind the summary.\n   *\n   * The agent can retrieve full content on-demand via:\n   *   - file_search (RAG semantic search over embedded files)\n   *   - content_tool read (by contentId for exact file retrieval)\n   *\n   * Built by the host orchestrator from message_file_map\n   * and metadata.context_files across all conversation messages.\n   */\n  fileManifest?: FileManifestEntry[];\n}\n"],"names":[],"mappings":";;AAwdA;;AAEG;AACG,MAAO,4BAA6B,SAAQ,KAAK,CAAA;AAClC,IAAA,WAAA;AAAnB,IAAA,WAAA,CAAmB,WAAmB,EAAA;AACpC,QAAA,KAAK,CAAC,CAAA,4CAAA,EAA+C,WAAW,CAAA,CAAE,CAAC;QADlD,IAAA,CAAA,WAAW,GAAX,WAAW;AAE5B,QAAA,IAAI,CAAC,IAAI,GAAG,8BAA8B;IAC5C;AACD;AAED;;AAEG;AACG,MAAO,8BAA+B,SAAQ,KAAK,CAAA;AACpC,IAAA,UAAA;AAAnB,IAAA,WAAA,CAAmB,UAAkB,EAAA;QACnC,KAAK,CACH,CAAA,8CAAA,EAAiD,UAAU,CAAA,GAAA,CAAK;AAC9D,YAAA,sEAAsE,CACzE;QAJgB,IAAA,CAAA,UAAU,GAAV,UAAU;AAK3B,QAAA,IAAI,CAAC,IAAI,GAAG,gCAAgC;IAC9C;AACD;;;;;"}
+{"version":3,"file":"graph.cjs","sources":["../../../src/types/graph.ts"],"sourcesContent":["// src/types/graph.ts\nimport type {\n  START,\n  StateType,\n  UpdateType,\n  StateGraph,\n  StateGraphArgs,\n  StateDefinition,\n  CompiledStateGraph,\n  BinaryOperatorAggregate,\n} from '@langchain/langgraph';\nimport type { BindToolsInput } from '@langchain/core/language_models/chat_models';\nimport type {\n  BaseMessage,\n  AIMessageChunk,\n  SystemMessage,\n} from '@langchain/core/messages';\nimport type { RunnableConfig, Runnable } from '@langchain/core/runnables';\nimport type { ChatGenerationChunk } from '@langchain/core/outputs';\nimport type { GoogleAIToolType } from '@langchain/google-common';\nimport type {\n  ToolMap,\n  ToolEndEvent,\n  GenericTool,\n  LCTool,\n  ToolApprovalConfig,\n  ToolExecuteBatchRequest,\n} from '@/types/tools';\nimport type { Providers, Callback, GraphNodeKeys } from '@/common';\nimport type { StandardGraph, MultiAgentGraph } from '@/graphs';\nimport type { ClientOptions } from '@/types/llm';\nimport type {\n  RunStep,\n  RunStepDeltaEvent,\n  MessageDeltaEvent,\n  ReasoningDeltaEvent,\n} from '@/types/stream';\nimport type { TokenCounter } from '@/types/run';\nimport type { SystemCacheBlock, AgentCacheTTL } from '@/types/agent-cache';\n\n/** Interface for bound model with stream and invoke methods */\nexport interface ChatModel {\n  stream?: (\n    messages: BaseMessage[],\n    config?: RunnableConfig\n  ) => Promise<AsyncIterable<AIMessageChunk>>;\n  invoke: (\n    messages: BaseMessage[],\n    config?: RunnableConfig\n  ) => Promise<AIMessageChunk>;\n}\n\n/** Payload for ON_AGENT_TRANSITION events */\nexport type AgentTransitionEvent = {\n  sourceAgentId?: string;\n  sourceAgentName?: string;\n  destinationAgentId: string;\n  destinationAgentName: string;\n  edgeType: string; // 'handoff' | 'transfer' | 'sequence'\n  timestamp: number;\n  /** When true, this event signals handoff completion (child → parent return) */\n  isCompletion?: boolean;\n  /** Duration of child agent execution in milliseconds (only on completion events) */\n  durationMs?: number;\n  /** Length of child agent result text in characters (only on completion events) */\n  resultLength?: number;\n};\n\nexport type GraphNode = GraphNodeKeys | typeof START;\nexport type ClientCallback<T extends unknown[]> = (\n  graph: StandardGraph,\n  ...args: T\n) => void;\n\nexport type ClientCallbacks = {\n  [Callback.TOOL_ERROR]?: ClientCallback<[Error, string]>;\n  [Callback.TOOL_START]?: ClientCallback<unknown[]>;\n  [Callback.TOOL_END]?: ClientCallback<unknown[]>;\n};\n\nexport type SystemCallbacks = {\n  [K in keyof ClientCallbacks]: ClientCallbacks[K] extends ClientCallback<\n    infer Args\n  >\n    ? (...args: Args) => void\n    : never;\n};\n\nexport type BaseGraphState = {\n  messages: BaseMessage[];\n  /**\n   * Structured response when using structured output mode.\n   * Contains the validated JSON response conforming to the configured schema.\n   */\n  structuredResponse?: Record<string, unknown>;\n};\n\nexport type MultiAgentGraphState = BaseGraphState & {\n  agentMessages?: BaseMessage[];\n};\n\nexport type IState = BaseGraphState;\n\nexport interface EventHandler {\n  handle(\n    event: string,\n    data:\n      | StreamEventData\n      | ModelEndData\n      | RunStep\n      | RunStepDeltaEvent\n      | MessageDeltaEvent\n      | ReasoningDeltaEvent\n      | SubagentUpdateEvent\n      | ToolExecuteBatchRequest\n      | { result: ToolEndEvent },\n    metadata?: Record<string, unknown>,\n    graph?: StandardGraph | MultiAgentGraph\n  ): void | Promise<void>;\n}\n\nexport type GraphStateChannels<T extends BaseGraphState> =\n  StateGraphArgs<T>['channels'];\n\nexport type Workflow<\n  T extends BaseGraphState = BaseGraphState,\n  U extends Partial<T> = Partial<T>,\n  N extends string = string,\n> = StateGraph<T, U, N>;\n\nexport type CompiledWorkflow<\n  T extends BaseGraphState = BaseGraphState,\n  U extends Partial<T> = Partial<T>,\n  N extends string = string,\n> = CompiledStateGraph<T, U, N>;\n\nexport type CompiledStateWorkflow = CompiledStateGraph<\n  StateType<{\n    messages: BinaryOperatorAggregate<BaseMessage[], BaseMessage[]>;\n  }>,\n  UpdateType<{\n    messages: BinaryOperatorAggregate<BaseMessage[], BaseMessage[]>;\n  }>,\n  string,\n  {\n    messages: BinaryOperatorAggregate<BaseMessage[], BaseMessage[]>;\n  },\n  {\n    messages: BinaryOperatorAggregate<BaseMessage[], BaseMessage[]>;\n  },\n  StateDefinition\n>;\n\nexport type CompiledMultiAgentWorkflow = CompiledStateGraph<\n  StateType<{\n    messages: BinaryOperatorAggregate<BaseMessage[], BaseMessage[]>;\n    agentMessages: BinaryOperatorAggregate<BaseMessage[], BaseMessage[]>;\n  }>,\n  UpdateType<{\n    messages: BinaryOperatorAggregate<BaseMessage[], BaseMessage[]>;\n    agentMessages: BinaryOperatorAggregate<BaseMessage[], BaseMessage[]>;\n  }>,\n  string,\n  {\n    messages: BinaryOperatorAggregate<BaseMessage[], BaseMessage[]>;\n    agentMessages: BinaryOperatorAggregate<BaseMessage[], BaseMessage[]>;\n  },\n  {\n    messages: BinaryOperatorAggregate<BaseMessage[], BaseMessage[]>;\n    agentMessages: BinaryOperatorAggregate<BaseMessage[], BaseMessage[]>;\n  },\n  StateDefinition\n>;\n\nexport type CompiledAgentWorfklow = CompiledStateGraph<\n  {\n    messages: BaseMessage[];\n  },\n  {\n    messages?: BaseMessage[] | undefined;\n  },\n  '__start__' | `agent=${string}` | `tools=${string}`,\n  {\n    messages: BinaryOperatorAggregate<BaseMessage[], BaseMessage[]>;\n  },\n  {\n    messages: BinaryOperatorAggregate<BaseMessage[], BaseMessage[]>;\n  },\n  StateDefinition,\n  {\n    [x: `agent=${string}`]: Partial<BaseGraphState>;\n    // eslint-disable-next-line @typescript-eslint/no-explicit-any\n    [x: `tools=${string}`]: any;\n  }\n>;\n\nexport type SystemRunnable =\n  | Runnable<\n      BaseMessage[],\n      (BaseMessage | SystemMessage)[],\n      RunnableConfig<Record<string, unknown>>\n    >\n  | undefined;\n\n/**\n * Optional compile options passed to workflow.compile().\n * These are intentionally untyped to avoid coupling to library internals.\n */\nexport type CompileOptions = {\n  // A checkpointer instance (e.g., MemorySaver, SQL saver)\n  // eslint-disable-next-line @typescript-eslint/no-explicit-any\n  checkpointer?: any;\n  interruptBefore?: string[];\n  interruptAfter?: string[];\n  /**\n   * Human-in-the-loop tool approval configuration.\n   * When set, tools matching the policy will trigger an interrupt()\n   * before execution, pausing the graph for human approval.\n   * Requires a checkpointer to be set for interrupt/resume to work.\n   */\n  toolApprovalConfig?: ToolApprovalConfig;\n};\n\nexport type EventStreamCallbackHandlerInput =\n  Parameters<CompiledWorkflow['streamEvents']>[2] extends Omit<\n    infer T,\n    'autoClose'\n  >\n    ? T\n    : never;\n\nexport type StreamChunk =\n  | (ChatGenerationChunk & {\n      message: AIMessageChunk;\n    })\n  | AIMessageChunk;\n\n/**\n * Data associated with a StreamEvent.\n */\nexport type StreamEventData = {\n  /**\n   * The input passed to the runnable that generated the event.\n   * Inputs will sometimes be available at the *START* of the runnable, and\n   * sometimes at the *END* of the runnable.\n   * If a runnable is able to stream its inputs, then its input by definition\n   * won't be known until the *END* of the runnable when it has finished streaming\n   * its inputs.\n   */\n  input?: unknown;\n  /**\n   * The output of the runnable that generated the event.\n   * Outputs will only be available at the *END* of the runnable.\n   * For most runnables, this field can be inferred from the `chunk` field,\n   * though there might be some exceptions for special cased runnables (e.g., like\n   * chat models), which may return more information.\n   */\n  output?: unknown;\n  /**\n   * A streaming chunk from the output that generated the event.\n   * chunks support addition in general, and adding them up should result\n   * in the output of the runnable that generated the event.\n   */\n  chunk?: StreamChunk;\n  /**\n   * Runnable config for invoking other runnables within handlers.\n   */\n  config?: RunnableConfig;\n  /**\n   * Custom result from the runnable that generated the event.\n   */\n  result?: unknown;\n  /**\n   * Custom field to indicate the event was manually emitted, and may have been handled already\n   */\n  emitted?: boolean;\n};\n\n/**\n * A streaming event.\n *\n * Schema of a streaming event which is produced from the streamEvents method.\n */\nexport type StreamEvent = {\n  /**\n   * Event names are of the format: on_[runnable_type]_(start|stream|end).\n   *\n   * Runnable types are one of:\n   * - llm - used by non chat models\n   * - chat_model - used by chat models\n   * - prompt --  e.g., ChatPromptTemplate\n   * - tool -- LangChain tools\n   * - chain - most Runnables are of this type\n   *\n   * Further, the events are categorized as one of:\n   * - start - when the runnable starts\n   * - stream - when the runnable is streaming\n   * - end - when the runnable ends\n   *\n   * start, stream and end are associated with slightly different `data` payload.\n   *\n   * Please see the documentation for `EventData` for more details.\n   */\n  event: string;\n  /** The name of the runnable that generated the event. */\n  name: string;\n  /**\n   * An randomly generated ID to keep track of the execution of the given runnable.\n   *\n   * Each child runnable that gets invoked as part of the execution of a parent runnable\n   * is assigned its own unique ID.\n   */\n  run_id: string;\n  /**\n   * Tags associated with the runnable that generated this event.\n   * Tags are always inherited from parent runnables.\n   */\n  tags?: string[];\n  /** Metadata associated with the runnable that generated this event. */\n  metadata: Record<string, unknown>;\n  /**\n   * Event data.\n   *\n   * The contents of the event data depend on the event type.\n   */\n  data: StreamEventData;\n};\n\nexport type GraphConfig = {\n  provider: string;\n  thread_id?: string;\n  run_id?: string;\n};\n\nexport type PartMetadata = {\n  progress?: number;\n  asset_pointer?: string;\n  status?: string;\n  action?: boolean;\n  output?: string;\n};\n\nexport type ModelEndData =\n  | (StreamEventData & { output: AIMessageChunk | undefined })\n  | undefined;\nexport type GraphTools = GenericTool[] | BindToolsInput[] | GoogleAIToolType[];\nexport type StandardGraphInput = {\n  runId?: string;\n  signal?: AbortSignal;\n  agents: AgentInputs[];\n  tokenCounter?: TokenCounter;\n  indexTokenCountMap?: Record<string, number>;\n};\n\n/**\n * Configuration for an approval gate placed on a sequence edge.\n * When present, the graph inserts an approval gate node between the source\n * and destination agents. The gate ALWAYS fires (regardless of ExecutionContext)\n * and calls interrupt() to pause the graph for human approval.\n */\nexport type ApprovalGateConfig = {\n  /** Unique identifier for this gate (used as node ID suffix) */\n  gateId: string;\n  /**\n   * Approval channel — where the approval UI is rendered.\n   * - 'chat': SSE-based chat UI (default)\n   * - 'outlook': MS Graph Actionable Messages\n   * - 'telegram': Telegram Bot inline keyboard\n   */\n  channel?: 'chat' | 'outlook' | 'telegram';\n  /** Optional human-readable prompt shown to the approver */\n  prompt?: string;\n  /** Optional approver identifier (e.g., email, user ID) */\n  approver?: string;\n  /** Timeout in ms before the gate auto-expires (default: 5 minutes) */\n  timeoutMs?: number;\n  /** What to do on denial: 'stop' ends the graph, 'skip' skips the destination agent */\n  onDeny?: 'stop' | 'skip';\n};\n\nexport type GraphEdge = {\n  /** Agent ID, use a list for multiple sources */\n  from: string | string[];\n  /** Agent ID, use a list for multiple destinations */\n  to: string | string[];\n  description?: string;\n  /** Can return boolean or specific destination(s) */\n  condition?: (state: BaseGraphState) => boolean | string | string[];\n  /**\n   * EdgeType.HANDOFF — one-way routing: parent emits an `lc_transfer_to_*`\n   *   tool call and exits, child takes over and responds directly to the\n   *   user. Aligns with upstream's handoff semantics.\n   * EdgeType.DIRECT — fixed graph edges for automatic sequential / parallel\n   *   transitions. Ranger preserves its enriched wiring (fan-in with prompt,\n   *   parallel groups, ApprovalGateNode, excludeResults / agentMessages).\n   */\n  edgeType?: import('@/common').EdgeType;\n  /**\n   * For sequence edges: Optional prompt to add when transitioning through this edge.\n   * String prompts can include variables like {results} which will be replaced with\n   * messages from startIndex onwards. When {results} is used, excludeResults defaults to true.\n   *\n   * For transfer edges: Description for the input parameter that the transfer tool accepts,\n   * allowing the supervisor to pass specific instructions/context to the transferred agent.\n   */\n  prompt?:\n    | string\n    | ((\n        messages: BaseMessage[],\n        runStartIndex: number\n      ) => string | Promise<string> | undefined);\n  /**\n   * When true, excludes messages from startIndex when adding prompt.\n   * Automatically set to true when {results} variable is used in prompt.\n   */\n  excludeResults?: boolean;\n  /**\n   * For transfer edges: Customizes the parameter name for the transfer input.\n   * Defaults to \"instructions\" if not specified.\n   * Only applies when prompt is provided for transfer edges.\n   *\n   * For handoff edges: Customizes the parameter name for the handoff instruction input.\n   */\n  promptKey?: string;\n  /**\n   * Approval gate configuration for sequence edges.\n   * When set, inserts an approval gate node between source and destination.\n   * The gate ALWAYS fires regardless of ExecutionContext (unlike tool approval).\n   */\n  approvalGate?: ApprovalGateConfig;\n};\n\nexport type MultiAgentGraphInput = StandardGraphInput & {\n  edges: GraphEdge[];\n  /**\n   * When set, the graph routes START to this agent instead of the default\n   * starting nodes. Used for multi-turn resumption: the caller reads the\n   * last active agent id from the previous turn (e.g. via\n   * `Run.getLastActiveAgentId()`, which derives it from the run's content\n   * data trail) and passes it here so follow-up messages route to the\n   * agent that last handled the conversation.\n   *\n   * If the agent ID is invalid (not in the graph), falls back to default\n   * starting nodes.\n   */\n  resumeFromAgentId?: string;\n};\n\n/**\n * Structured output mode determines how the agent returns structured data.\n * - 'tool': Uses tool calling to return structured output (works with all tool-calling models)\n * - 'provider': Uses provider-native structured output via LangChain's jsonMode (OpenAI, Anthropic, etc.)\n * - 'native': Uses provider's constrained decoding API directly for guaranteed schema compliance\n *   (Anthropic output_config.format, OpenAI response_format.json_schema). Falls back to 'tool' for unsupported providers.\n * - 'auto': Automatically selects the best strategy — 'native' for supported providers, 'tool' for others\n */\nexport type StructuredOutputMode = 'tool' | 'provider' | 'native' | 'auto';\n\n/**\n * Resolved method used internally after mode resolution.\n * Maps to LangChain's withStructuredOutput method parameter plus our native path.\n */\nexport type ResolvedStructuredOutputMethod =\n  | 'functionCalling'\n  | 'jsonMode'\n  | 'jsonSchema'\n  | 'native'\n  | undefined;\n\n/**\n * Error thrown when the model refuses to produce structured output due to safety policies.\n */\nexport class StructuredOutputRefusalError extends Error {\n  constructor(public refusalText: string) {\n    super(`Model refused to produce structured output: ${refusalText}`);\n    this.name = 'StructuredOutputRefusalError';\n  }\n}\n\n/**\n * Error thrown when the structured output response was truncated due to max_tokens.\n */\nexport class StructuredOutputTruncatedError extends Error {\n  constructor(public stopReason: string) {\n    super(\n      `Structured output was truncated (stop_reason: ${stopReason}). ` +\n        'Increase max_tokens to allow the full JSON response to be generated.'\n    );\n    this.name = 'StructuredOutputTruncatedError';\n  }\n}\n\n/**\n * Configuration for structured JSON output from agents.\n * When configured, the agent will return a validated JSON response\n * instead of streaming text.\n */\nexport interface StructuredOutputConfig {\n  /**\n   * JSON Schema defining the output structure.\n   * The model will be forced to return data conforming to this schema.\n   */\n  schema: Record<string, unknown>;\n  /**\n   * Name for the structured output format (used in tool mode).\n   * @default 'StructuredResponse'\n   */\n  name?: string;\n  /**\n   * Description of what the structured output represents.\n   * Helps the model understand the expected format.\n   */\n  description?: string;\n  /**\n   * Output mode strategy.\n   * @default 'auto'\n   */\n  mode?: StructuredOutputMode;\n  /**\n   * Enable strict schema validation.\n   * When true, the response must exactly match the schema.\n   * @default true\n   */\n  strict?: boolean;\n  /**\n   * Error handling configuration.\n   * - true: Auto-retry on validation errors (default)\n   * - false: Throw error on validation failure\n   * - string: Custom error message for retry\n   */\n  handleErrors?: boolean | string;\n  /**\n   * Maximum number of retry attempts on validation failure.\n   * @default 2\n   */\n  maxRetries?: number;\n  /**\n   * Include the raw AI message along with structured response.\n   * Useful for debugging.\n   * @default false\n   */\n  includeRaw?: boolean;\n}\n\n/**\n * Database/API structured output format (snake_case with enabled flag).\n * This matches the format stored in MongoDB and sent from frontends.\n */\nexport interface StructuredOutputInput {\n  /** Whether structured output is enabled */\n  enabled?: boolean;\n  /** JSON Schema defining the expected response structure */\n  schema?: Record<string, unknown>;\n  /** Name identifier for the structured output */\n  name?: string;\n  /** Description of what the structured output represents */\n  description?: string;\n  /** Mode for structured output: 'tool' | 'provider' | 'native' | 'auto' */\n  mode?: StructuredOutputMode;\n  /** Whether to enforce strict schema validation */\n  strict?: boolean;\n}\n\n/**\n * Trigger strategy for when summarization should activate.\n * - 'contextPercentage': Trigger when context utilization exceeds a threshold percentage\n * - 'messageCount': Trigger when pruned message count exceeds a threshold\n * - 'tokenThreshold': Trigger when total token count exceeds a raw threshold\n */\nexport type SummarizationTriggerType =\n  | 'contextPercentage'\n  | 'messageCount'\n  | 'tokenThreshold';\n\n/**\n * Configuration for summarization behavior within the agent pipeline.\n * All fields are optional — sensible defaults are provided via constants.\n *\n * @see SUMMARIZATION_CONTEXT_THRESHOLD, SUMMARIZATION_RESERVE_RATIO, PRUNING_EMA_ALPHA\n */\nexport interface SummarizationConfig {\n  /**\n   * Strategy for when summarization triggers.\n   * @default 'contextPercentage'\n   */\n  triggerType?: SummarizationTriggerType;\n\n  /**\n   * Threshold value interpreted based on triggerType:\n   * - contextPercentage: 0-100 (percentage of context window)\n   * - messageCount: absolute count of messages pruned\n   * - tokenThreshold: absolute token count\n   * @default 80 (for contextPercentage)\n   */\n  triggerThreshold?: number;\n\n  /**\n   * Fraction of context window (0-1) reserved for recent messages.\n   * Prevents over-pruning by ensuring at least this fraction of the\n   * context budget is preserved as recent conversation history.\n   * @default 0.3\n   */\n  reserveRatio?: number;\n\n  /**\n   * Whether context pruning is enabled (can be disabled for debugging).\n   * @default true\n   */\n  contextPruning?: boolean;\n\n  /**\n   * Initial summary text to seed across runs.\n   * Different from persistedSummary: this is provided by the caller as a\n   * cross-conversation seed (e.g., agent personality or recurring context),\n   * while persistedSummary is loaded from the conversation's own history.\n   */\n  initialSummary?: string;\n\n  /**\n   * Upstream-aligned optional fields. When the host wants the summarization\n   * pass to run on a different LLM than the agent's own (e.g. a cheaper\n   * model for compaction), these provider/model overrides flow through to\n   * the summarize callback.\n   */\n  provider?: Providers;\n  model?: string;\n  parameters?: Record<string, unknown>;\n  prompt?: string;\n  updatePrompt?: string;\n  trigger?: { type: string; value: number };\n  maxSummaryTokens?: number;\n}\n\n/**\n * Runtime state for EMA-based pruning calibration.\n * Maintained across iterations within a single run to smooth pruning decisions.\n */\nexport interface PruneCalibrationState {\n  /** Current EMA calibration ratio */\n  ratio: number;\n  /** Number of calibration updates applied */\n  iterations: number;\n}\n\n/**\n * Lightweight file metadata entry for conversation-level file awareness.\n * Contains only IDs and names — NOT full content — so the agent always knows\n * what files exist in the conversation even after compaction pushes old messages\n * behind the summary window. The agent can retrieve full content on-demand\n * via file_search (RAG) or content_tool read (by contentId).\n */\nexport interface FileManifestEntry {\n  /** Unique file identifier (e.g., MongoDB ObjectId or UUID) */\n  fileId: string;\n  /** Original filename (e.g., \"quarterly-report.pdf\") */\n  filename: string;\n  /** Content identifier for on-demand retrieval via content_tool read */\n  contentId?: string;\n  /** File source (e.g., \"local\", \"sharepoint\", \"onedrive\") */\n  source?: string;\n  /** Index of the message that introduced this file (0-based in the original message array) */\n  messageIndex?: number;\n}\n\n/** Configuration for a subagent type that can be spawned by a parent agent. */\nexport type SubagentConfig = {\n  /** Identifier used in the tool's `subagent_type` enum (e.g. 'researcher', 'coder'). */\n  type: string;\n  /** Human-readable display name. */\n  name: string;\n  /** What this subagent specializes in — shown to the LLM. */\n  description: string;\n  /** Full agent config for the child graph. Omit when `self` is true. */\n  agentInputs?: AgentInputs;\n  /** When true, reuse the parent's AgentInputs (context isolation without separate config). */\n  self?: boolean;\n  /** Max AGENT→TOOLS cycles before forced stop (default: 25). */\n  maxTurns?: number;\n  /** Allow this subagent to spawn its own subagents (default: false). */\n  allowNested?: boolean;\n};\n\n/** SubagentConfig with agentInputs guaranteed present (self-spawn resolved). */\nexport type ResolvedSubagentConfig = SubagentConfig & {\n  agentInputs: AgentInputs;\n};\n\n/** Lifecycle phase carried on {@link SubagentUpdateEvent}. */\nexport type SubagentUpdatePhase =\n  | 'start'\n  | 'run_step'\n  | 'run_step_delta'\n  | 'run_step_completed'\n  | 'message_delta'\n  | 'reasoning_delta'\n  | 'stop'\n  | 'error';\n\n/**\n * Wrapper event emitted when a subagent's child graph dispatches activity.\n * Lets hosts show subagent progress in a UI surface separate from the parent\n * conversation without having to untangle events by agent ID.\n */\nexport interface SubagentUpdateEvent {\n  /** Parent run ID. */\n  runId: string;\n  /** Child run ID (unique per subagent execution). */\n  subagentRunId: string;\n  /**\n   * Parent-side `tool_call_id` for the `subagent` tool invocation that\n   * triggered this run. Stable for the duration of the child; lets hosts\n   * correlate updates deterministically instead of inferring by ordering.\n   * Omitted when the executor was invoked outside of a tool-call context.\n   */\n  parentToolCallId?: string;\n  /** Subagent `type` identifier from the SubagentConfig. */\n  subagentType: string;\n  /** Child agent ID assigned to this subagent execution. */\n  subagentAgentId: string;\n  /** Parent agent ID that spawned this subagent. */\n  parentAgentId?: string;\n  /** Lifecycle phase carried by this update. */\n  phase: SubagentUpdatePhase;\n  /** Underlying event payload (shape depends on phase). */\n  data?: unknown;\n  /** Short human-readable description. Hosts can render this directly. */\n  label?: string;\n  /** ISO timestamp for ordering / display. */\n  timestamp: string;\n}\n\nexport interface AgentInputs {\n  agentId: string;\n  /** Human-readable name for the agent (used in handoff context). Defaults to agentId if not provided. */\n  name?: string;\n  /** Description of what this agent does (used to enrich handoff tool descriptions). */\n  description?: string;\n  toolEnd?: boolean;\n  toolMap?: ToolMap;\n  tools?: GraphTools;\n  provider: Providers;\n  /**\n   * Ordered list of cacheable system content blocks emitted BEFORE\n   * `instructions` in the system message. Each block gets its own cache\n   * marker (cachePoint on Bedrock, cache_control on Anthropic) so the\n   * cache key for each block includes only the bytes up to and including\n   * its own marker. Entries are emitted in array order — earlier =\n   * stabler = wider cache key (best for cross-tenant sharing).\n   *\n   * See `src/types/agent-cache.ts` for full semantics. Capped at\n   * `MAX_SYSTEM_CACHE_BLOCKS` (2) entries to stay within Bedrock's\n   * 4-cachePoint budget after the tools array consumes 2.\n   */\n  system_cache_blocks?: SystemCacheBlock[];\n  /**\n   * Stable/cacheable system instructions for this agent. Always emitted\n   * with its own trailing cache marker (when caching is supported by\n   * the provider/model). Use `instructions_cache_ttl` to override the\n   * default '5m' TTL.\n   */\n  instructions?: string;\n  /** TTL for the trailing `instructions` cache marker. Defaults to '5m'. */\n  instructions_cache_ttl?: AgentCacheTTL;\n  streamBuffer?: number;\n  maxContextTokens?: number;\n  clientOptions?: ClientOptions;\n  /**\n   * Dynamic system tail appended after the cacheable instructions\n   * without any cache marker. Per-user / per-message context belongs\n   * here so it does not invalidate the cacheable prefix.\n   */\n  additional_instructions?: string;\n  reasoningKey?: 'reasoning_content' | 'reasoning';\n  /**\n   * Subagent types this agent may spawn. When non-empty, Graph injects the\n   * `subagent` tool into the agent's toolset with the per-config enum and\n   * description populated.\n   */\n  subagentConfigs?: SubagentConfig[];\n  /**\n   * Maximum subagent depth allowed below this agent. Used by SubagentExecutor\n   * to abort runaway nesting. Defaults to a small constant.\n   */\n  maxSubagentDepth?: number;\n  /**\n   * Pre-existing summary injected into the system message via\n   * AgentContext.buildInstructionsString. Populated by SubagentExecutor when\n   * spawning a self-config subagent that should inherit the parent's prior\n   * compaction summary; otherwise host-supplied at run time.\n   */\n  initialSummary?: { text: string; tokenCount: number };\n  /**\n   * Upstream-aligned opt-in for summarization. When false, the agent never\n   * invokes the summarize callback regardless of context utilization.\n   * Defaults to undefined (treated as enabled in Ranger to preserve prior\n   * behaviour); hosts that want strict opt-in semantics should set it\n   * explicitly.\n   */\n  summarizationEnabled?: boolean;\n  /** Format content blocks as strings (for legacy compatibility i.e. Ollama/Azure Serverless) */\n  useLegacyContent?: boolean;\n  /**\n   * Tool definitions for all tools, including deferred and programmatic.\n   * Used for tool search and programmatic tool calling.\n   * Maps tool name to LCTool definition.\n   */\n  toolRegistry?: Map<string, LCTool>;\n  /**\n   * Dynamic context that changes per-request (e.g., current time, user info).\n   * This is injected as a user message rather than system prompt to preserve cache.\n   * Keeping this separate from instructions ensures the system message stays static\n   * and can be cached by Bedrock/Anthropic prompt caching.\n   */\n  dynamicContext?: string;\n  /**\n   * Structured output configuration (camelCase).\n   * When set, disables streaming and returns a validated JSON response\n   * conforming to the specified schema.\n   */\n  structuredOutput?: StructuredOutputConfig;\n  /**\n   * Structured output configuration (snake_case - database/API format).\n   * Alternative to structuredOutput for compatibility with MongoDB/frontend.\n   * Uses an `enabled` flag to control activation.\n   * @deprecated Use structuredOutput instead when possible\n   */\n  structured_output?: StructuredOutputInput;\n  /**\n   * Serializable tool definitions for event-driven execution.\n   * When provided, ToolNode operates in event-driven mode, dispatching\n   * ON_TOOL_EXECUTE events instead of invoking tools directly.\n   */\n  toolDefinitions?: LCTool[];\n  /**\n   * Tool names discovered from previous conversation history.\n   * These tools will be pre-marked as discovered so they're included\n   * in tool binding without requiring tool_search.\n   */\n  discoveredTools?: string[];\n  /**\n   * Optional callback for summarizing messages that were pruned from context.\n   * When provided, discarded messages are summarized by the host caller\n   * using a cheap LLM call, and the summary is prepended to the context.\n   */\n  summarizeCallback?: (\n    messagesToRefine: import('@langchain/core/messages').BaseMessage[]\n  ) => Promise<string | undefined>;\n  /**\n   * Pre-existing summary text loaded from persistent storage (MongoDB/Redis).\n   * When provided, this summary is injected into the initial message context\n   * so the agent has prior conversation history even on new turns.\n   * Set by the host's summary store when resuming a conversation.\n   */\n  persistedSummary?: string;\n  /**\n   * Summarization configuration controlling trigger strategy, reserve ratio,\n   * and EMA calibration for pruning. When omitted, sensible defaults apply.\n   * @see SummarizationConfig\n   */\n  summarizationConfig?: SummarizationConfig;\n  /**\n   * Lightweight file manifest for the conversation.\n   * Contains file IDs, names, and metadata — NOT full content.\n   *\n   * Used by the compaction engine to inject a [Conversation Files] block\n   * into the windowed view, ensuring the LLM always knows what files exist\n   * even when old messages (with full file content) are behind the summary.\n   *\n   * The agent can retrieve full content on-demand via:\n   *   - file_search (RAG semantic search over embedded files)\n   *   - content_tool read (by contentId for exact file retrieval)\n   *\n   * Built by the host orchestrator from message_file_map\n   * and metadata.context_files across all conversation messages.\n   */\n  fileManifest?: FileManifestEntry[];\n}\n\n/**\n * Tunable knobs for position-based content degradation.\n * See `messages/contextPruning.ts` and `messages/contextPruningSettings.ts`.\n */\nexport interface ContextPruningConfig {\n  enabled?: boolean;\n  keepLastAssistants?: number;\n  softTrimRatio?: number;\n  hardClearRatio?: number;\n  minPrunableToolChars?: number;\n  softTrim?: {\n    maxChars?: number;\n    headChars?: number;\n    tailChars?: number;\n  };\n  hardClear?: {\n    enabled?: boolean;\n    placeholder?: string;\n  };\n}\n"],"names":[],"mappings":";;AAqdA;;AAEG;AACG,MAAO,4BAA6B,SAAQ,KAAK,CAAA;AAClC,IAAA,WAAA;AAAnB,IAAA,WAAA,CAAmB,WAAmB,EAAA;AACpC,QAAA,KAAK,CAAC,CAAA,4CAAA,EAA+C,WAAW,CAAA,CAAE,CAAC;QADlD,IAAA,CAAA,WAAW,GAAX,WAAW;AAE5B,QAAA,IAAI,CAAC,IAAI,GAAG,8BAA8B;IAC5C;AACD;AAED;;AAEG;AACG,MAAO,8BAA+B,SAAQ,KAAK,CAAA;AACpC,IAAA,UAAA;AAAnB,IAAA,WAAA,CAAmB,UAAkB,EAAA;QACnC,KAAK,CACH,CAAA,8CAAA,EAAiD,UAAU,CAAA,GAAA,CAAK;AAC9D,YAAA,sEAAsE,CACzE;QAJgB,IAAA,CAAA,UAAU,GAAV,UAAU;AAK3B,QAAA,IAAI,CAAC,IAAI,GAAG,gCAAgC;IAC9C;AACD;;;;;"}

package/dist/cjs/utils/truncation.cjs

@@ -0,0 +1,135 @@
+'use strict';
+
+/**
+ * Ingestion-time and pre-flight truncation utilities for tool results.
+ *
+ * Prevents oversized tool outputs from entering the message array and
+ * consuming the entire context window.
+ */
+/**
+ * Absolute hard cap on tool result length (characters).
+ * Even if the model has a 1M-token context, a single tool result
+ * larger than this is almost certainly a bug (e.g., dumping a binary file).
+ */
+const HARD_MAX_TOOL_RESULT_CHARS = 400_000;
+/**
+ * Absolute hard cap on the aggregate size (characters) of all registered
+ * tool outputs kept for `{{tool<i>turn<n>}}` substitution. Set at 5 MB
+ * because the registry stores *raw, untruncated* tool output — full
+ * fidelity for piping into downstream bash/jq — so the budget needs
+ * enough headroom to keep a handful of large responses without
+ * ballooning unbounded.
+ */
+const HARD_MAX_TOTAL_TOOL_OUTPUT_SIZE = 5_000_000;
+/**
+ * Computes the dynamic max tool result size based on the model's context window.
+ * Uses 30% of the context window (in estimated characters, ~4 chars/token)
+ * capped at HARD_MAX_TOOL_RESULT_CHARS.
+ *
+ * @param contextWindowTokens - The model's max context tokens (optional).
+ * @returns Maximum allowed characters for a single tool result.
+ */
+function calculateMaxToolResultChars(contextWindowTokens) {
+    if (contextWindowTokens == null || contextWindowTokens <= 0) {
+        return HARD_MAX_TOOL_RESULT_CHARS;
+    }
+    return Math.min(Math.floor(contextWindowTokens * 0.3) * 4, HARD_MAX_TOOL_RESULT_CHARS);
+}
+/**
+ * Computes the default aggregate size (characters) for the tool output
+ * reference registry based on the per-output budget. Mirrors
+ * `calculateMaxToolResultChars`'s shape: a multiple of the per-output
+ * cap, clamped to `HARD_MAX_TOTAL_TOOL_OUTPUT_SIZE`.
+ *
+ * @param maxOutputSize - Per-output maximum characters (e.g., the
+ *   ToolNode's `maxToolResultChars`). When omitted or non-positive,
+ *   falls back to the absolute total cap.
+ * @returns Maximum total characters retained across the registry.
+ */
+function calculateMaxTotalToolOutputSize(maxOutputSize) {
+    if (maxOutputSize == null || maxOutputSize <= 0) {
+        return HARD_MAX_TOTAL_TOOL_OUTPUT_SIZE;
+    }
+    return Math.min(maxOutputSize * 2, HARD_MAX_TOTAL_TOOL_OUTPUT_SIZE);
+}
+/**
+ * Truncates a tool-call input (the arguments/payload of a tool_use block)
+ * using head+tail strategy. Returns an object with `_truncated` (the
+ * truncated string) and `_originalChars` (for diagnostics).
+ *
+ * Accepts any type — objects are JSON-serialized before truncation.
+ *
+ * @param input - The tool input (string, object, etc.).
+ * @param maxChars - Maximum allowed characters.
+ */
+function truncateToolInput(input, maxChars) {
+    const serialized = typeof input === 'string' ? input : JSON.stringify(input);
+    if (serialized.length <= maxChars) {
+        return { _truncated: serialized, _originalChars: serialized.length };
+    }
+    const indicator = `\n… [truncated: ${serialized.length} chars exceeded ${maxChars} limit] …\n`;
+    const available = maxChars - indicator.length;
+    if (available < 100) {
+        return {
+            _truncated: serialized.slice(0, maxChars) + indicator.trimEnd(),
+            _originalChars: serialized.length,
+        };
+    }
+    const headSize = Math.ceil(available * 0.7);
+    const tailSize = available - headSize;
+    return {
+        _truncated: serialized.slice(0, headSize) +
+            indicator +
+            serialized.slice(serialized.length - tailSize),
+        _originalChars: serialized.length,
+    };
+}
+/**
+ * Truncates tool result content that exceeds `maxChars` using a head+tail
+ * strategy. Keeps the beginning (structure/headers) and end (return value /
+ * conclusion) of the content so the model retains both the opening context
+ * and the final outcome.
+ *
+ * Head gets ~70% of the budget, tail gets ~30%. Falls back to head-only
+ * when the budget is too small for a meaningful tail.
+ *
+ * @param content - The tool result string content.
+ * @param maxChars - Maximum allowed characters.
+ * @returns The (possibly truncated) content string.
+ */
+function truncateToolResultContent(content, maxChars) {
+    if (content.length <= maxChars) {
+        return content;
+    }
+    const indicator = `\n\n… [truncated: ${content.length} chars exceeded ${maxChars} limit] …\n\n`;
+    const available = maxChars - indicator.length;
+    if (available <= 0) {
+        return content.slice(0, maxChars);
+    }
+    // When budget is too small for a meaningful tail, fall back to head-only
+    if (available < 200) {
+        return content.slice(0, available) + indicator.trimEnd();
+    }
+    const headSize = Math.ceil(available * 0.7);
+    const tailSize = available - headSize;
+    // Try to break at newline boundaries for cleaner output
+    let headEnd = headSize;
+    const headNewline = content.lastIndexOf('\n', headSize);
+    if (headNewline > headSize - 200 && headNewline > 0) {
+        headEnd = headNewline;
+    }
+    let tailStart = content.length - tailSize;
+    const tailNewline = content.indexOf('\n', tailStart);
+    if (tailNewline > 0 && tailNewline < tailStart + 200) {
+        tailStart = tailNewline + 1;
+    }
+    return content.slice(0, headEnd) + indicator + content.slice(tailStart);
+}
+
+exports.HARD_MAX_TOOL_RESULT_CHARS = HARD_MAX_TOOL_RESULT_CHARS;
+exports.HARD_MAX_TOTAL_TOOL_OUTPUT_SIZE = HARD_MAX_TOTAL_TOOL_OUTPUT_SIZE;
+exports.calculateMaxToolResultChars = calculateMaxToolResultChars;
+exports.calculateMaxTotalToolOutputSize = calculateMaxTotalToolOutputSize;
+exports.truncateToolInput = truncateToolInput;
+exports.truncateToolResultContent = truncateToolResultContent;
+//# sourceMappingURL=truncation.cjs.map

package/dist/cjs/utils/truncation.cjs.map

@@ -0,0 +1 @@
+{"version":3,"file":"truncation.cjs","sources":["../../../src/utils/truncation.ts"],"sourcesContent":["/**\n * Ingestion-time and pre-flight truncation utilities for tool results.\n *\n * Prevents oversized tool outputs from entering the message array and\n * consuming the entire context window.\n */\n\n/**\n * Absolute hard cap on tool result length (characters).\n * Even if the model has a 1M-token context, a single tool result\n * larger than this is almost certainly a bug (e.g., dumping a binary file).\n */\nexport const HARD_MAX_TOOL_RESULT_CHARS = 400_000;\n\n/**\n * Absolute hard cap on the aggregate size (characters) of all registered\n * tool outputs kept for `{{tool<i>turn<n>}}` substitution. Set at 5 MB\n * because the registry stores *raw, untruncated* tool output — full\n * fidelity for piping into downstream bash/jq — so the budget needs\n * enough headroom to keep a handful of large responses without\n * ballooning unbounded.\n */\nexport const HARD_MAX_TOTAL_TOOL_OUTPUT_SIZE = 5_000_000;\n\n/**\n * Computes the dynamic max tool result size based on the model's context window.\n * Uses 30% of the context window (in estimated characters, ~4 chars/token)\n * capped at HARD_MAX_TOOL_RESULT_CHARS.\n *\n * @param contextWindowTokens - The model's max context tokens (optional).\n * @returns Maximum allowed characters for a single tool result.\n */\nexport function calculateMaxToolResultChars(\n  contextWindowTokens?: number\n): number {\n  if (contextWindowTokens == null || contextWindowTokens <= 0) {\n    return HARD_MAX_TOOL_RESULT_CHARS;\n  }\n  return Math.min(\n    Math.floor(contextWindowTokens * 0.3) * 4,\n    HARD_MAX_TOOL_RESULT_CHARS\n  );\n}\n\n/**\n * Computes the default aggregate size (characters) for the tool output\n * reference registry based on the per-output budget. Mirrors\n * `calculateMaxToolResultChars`'s shape: a multiple of the per-output\n * cap, clamped to `HARD_MAX_TOTAL_TOOL_OUTPUT_SIZE`.\n *\n * @param maxOutputSize - Per-output maximum characters (e.g., the\n *   ToolNode's `maxToolResultChars`). When omitted or non-positive,\n *   falls back to the absolute total cap.\n * @returns Maximum total characters retained across the registry.\n */\nexport function calculateMaxTotalToolOutputSize(\n  maxOutputSize?: number\n): number {\n  if (maxOutputSize == null || maxOutputSize <= 0) {\n    return HARD_MAX_TOTAL_TOOL_OUTPUT_SIZE;\n  }\n  return Math.min(maxOutputSize * 2, HARD_MAX_TOTAL_TOOL_OUTPUT_SIZE);\n}\n\n/**\n * Truncates a tool-call input (the arguments/payload of a tool_use block)\n * using head+tail strategy. Returns an object with `_truncated` (the\n * truncated string) and `_originalChars` (for diagnostics).\n *\n * Accepts any type — objects are JSON-serialized before truncation.\n *\n * @param input - The tool input (string, object, etc.).\n * @param maxChars - Maximum allowed characters.\n */\nexport function truncateToolInput(\n  input: unknown,\n  maxChars: number\n): { _truncated: string; _originalChars: number } {\n  const serialized = typeof input === 'string' ? input : JSON.stringify(input);\n  if (serialized.length <= maxChars) {\n    return { _truncated: serialized, _originalChars: serialized.length };\n  }\n  const indicator = `\\n… [truncated: ${serialized.length} chars exceeded ${maxChars} limit] …\\n`;\n  const available = maxChars - indicator.length;\n\n  if (available < 100) {\n    return {\n      _truncated: serialized.slice(0, maxChars) + indicator.trimEnd(),\n      _originalChars: serialized.length,\n    };\n  }\n\n  const headSize = Math.ceil(available * 0.7);\n  const tailSize = available - headSize;\n\n  return {\n    _truncated:\n      serialized.slice(0, headSize) +\n      indicator +\n      serialized.slice(serialized.length - tailSize),\n    _originalChars: serialized.length,\n  };\n}\n\n/**\n * Truncates tool result content that exceeds `maxChars` using a head+tail\n * strategy. Keeps the beginning (structure/headers) and end (return value /\n * conclusion) of the content so the model retains both the opening context\n * and the final outcome.\n *\n * Head gets ~70% of the budget, tail gets ~30%. Falls back to head-only\n * when the budget is too small for a meaningful tail.\n *\n * @param content - The tool result string content.\n * @param maxChars - Maximum allowed characters.\n * @returns The (possibly truncated) content string.\n */\nexport function truncateToolResultContent(\n  content: string,\n  maxChars: number\n): string {\n  if (content.length <= maxChars) {\n    return content;\n  }\n\n  const indicator = `\\n\\n… [truncated: ${content.length} chars exceeded ${maxChars} limit] …\\n\\n`;\n  const available = maxChars - indicator.length;\n  if (available <= 0) {\n    return content.slice(0, maxChars);\n  }\n\n  // When budget is too small for a meaningful tail, fall back to head-only\n  if (available < 200) {\n    return content.slice(0, available) + indicator.trimEnd();\n  }\n\n  const headSize = Math.ceil(available * 0.7);\n  const tailSize = available - headSize;\n\n  // Try to break at newline boundaries for cleaner output\n  let headEnd = headSize;\n  const headNewline = content.lastIndexOf('\\n', headSize);\n  if (headNewline > headSize - 200 && headNewline > 0) {\n    headEnd = headNewline;\n  }\n\n  let tailStart = content.length - tailSize;\n  const tailNewline = content.indexOf('\\n', tailStart);\n  if (tailNewline > 0 && tailNewline < tailStart + 200) {\n    tailStart = tailNewline + 1;\n  }\n\n  return content.slice(0, headEnd) + indicator + content.slice(tailStart);\n}\n"],"names":[],"mappings":";;AAAA;;;;;AAKG;AAEH;;;;AAIG;AACI,MAAM,0BAA0B,GAAG;AAE1C;;;;;;;AAOG;AACI,MAAM,+BAA+B,GAAG;AAE/C;;;;;;;AAOG;AACG,SAAU,2BAA2B,CACzC,mBAA4B,EAAA;IAE5B,IAAI,mBAAmB,IAAI,IAAI,IAAI,mBAAmB,IAAI,CAAC,EAAE;AAC3D,QAAA,OAAO,0BAA0B;IACnC;AACA,IAAA,OAAO,IAAI,CAAC,GAAG,CACb,IAAI,CAAC,KAAK,CAAC,mBAAmB,GAAG,GAAG,CAAC,GAAG,CAAC,EACzC,0BAA0B,CAC3B;AACH;AAEA;;;;;;;;;;AAUG;AACG,SAAU,+BAA+B,CAC7C,aAAsB,EAAA;IAEtB,IAAI,aAAa,IAAI,IAAI,IAAI,aAAa,IAAI,CAAC,EAAE;AAC/C,QAAA,OAAO,+BAA+B;IACxC;IACA,OAAO,IAAI,CAAC,GAAG,CAAC,aAAa,GAAG,CAAC,EAAE,+BAA+B,CAAC;AACrE;AAEA;;;;;;;;;AASG;AACG,SAAU,iBAAiB,CAC/B,KAAc,EACd,QAAgB,EAAA;AAEhB,IAAA,MAAM,UAAU,GAAG,OAAO,KAAK,KAAK,QAAQ,GAAG,KAAK,GAAG,IAAI,CAAC,SAAS,CAAC,KAAK,CAAC;AAC5E,IAAA,IAAI,UAAU,CAAC,MAAM,IAAI,QAAQ,EAAE;QACjC,OAAO,EAAE,UAAU,EAAE,UAAU,EAAE,cAAc,EAAE,UAAU,CAAC,MAAM,EAAE;IACtE;IACA,MAAM,SAAS,GAAG,CAAA,gBAAA,EAAmB,UAAU,CAAC,MAAM,CAAA,gBAAA,EAAmB,QAAQ,CAAA,WAAA,CAAa;AAC9F,IAAA,MAAM,SAAS,GAAG,QAAQ,GAAG,SAAS,CAAC,MAAM;AAE7C,IAAA,IAAI,SAAS,GAAG,GAAG,EAAE;QACnB,OAAO;AACL,YAAA,UAAU,EAAE,UAAU,CAAC,KAAK,CAAC,CAAC,EAAE,QAAQ,CAAC,GAAG,SAAS,CAAC,OAAO,EAAE;YAC/D,cAAc,EAAE,UAAU,CAAC,MAAM;SAClC;IACH;IAEA,MAAM,QAAQ,GAAG,IAAI,CAAC,IAAI,CAAC,SAAS,GAAG,GAAG,CAAC;AAC3C,IAAA,MAAM,QAAQ,GAAG,SAAS,GAAG,QAAQ;IAErC,OAAO;QACL,UAAU,EACR,UAAU,CAAC,KAAK,CAAC,CAAC,EAAE,QAAQ,CAAC;YAC7B,SAAS;YACT,UAAU,CAAC,KAAK,CAAC,UAAU,CAAC,MAAM,GAAG,QAAQ,CAAC;QAChD,cAAc,EAAE,UAAU,CAAC,MAAM;KAClC;AACH;AAEA;;;;;;;;;;;;AAYG;AACG,SAAU,yBAAyB,CACvC,OAAe,EACf,QAAgB,EAAA;AAEhB,IAAA,IAAI,OAAO,CAAC,MAAM,IAAI,QAAQ,EAAE;AAC9B,QAAA,OAAO,OAAO;IAChB;IAEA,MAAM,SAAS,GAAG,CAAA,kBAAA,EAAqB,OAAO,CAAC,MAAM,CAAA,gBAAA,EAAmB,QAAQ,CAAA,aAAA,CAAe;AAC/F,IAAA,MAAM,SAAS,GAAG,QAAQ,GAAG,SAAS,CAAC,MAAM;AAC7C,IAAA,IAAI,SAAS,IAAI,CAAC,EAAE;QAClB,OAAO,OAAO,CAAC,KAAK,CAAC,CAAC,EAAE,QAAQ,CAAC;IACnC;;AAGA,IAAA,IAAI,SAAS,GAAG,GAAG,EAAE;AACnB,QAAA,OAAO,OAAO,CAAC,KAAK,CAAC,CAAC,EAAE,SAAS,CAAC,GAAG,SAAS,CAAC,OAAO,EAAE;IAC1D;IAEA,MAAM,QAAQ,GAAG,IAAI,CAAC,IAAI,CAAC,SAAS,GAAG,GAAG,CAAC;AAC3C,IAAA,MAAM,QAAQ,GAAG,SAAS,GAAG,QAAQ;;IAGrC,IAAI,OAAO,GAAG,QAAQ;IACtB,MAAM,WAAW,GAAG,OAAO,CAAC,WAAW,CAAC,IAAI,EAAE,QAAQ,CAAC;IACvD,IAAI,WAAW,GAAG,QAAQ,GAAG,GAAG,IAAI,WAAW,GAAG,CAAC,EAAE;QACnD,OAAO,GAAG,WAAW;IACvB;AAEA,IAAA,IAAI,SAAS,GAAG,OAAO,CAAC,MAAM,GAAG,QAAQ;IACzC,MAAM,WAAW,GAAG,OAAO,CAAC,OAAO,CAAC,IAAI,EAAE,SAAS,CAAC;IACpD,IAAI,WAAW,GAAG,CAAC,IAAI,WAAW,GAAG,SAAS,GAAG,GAAG,EAAE;AACpD,QAAA,SAAS,GAAG,WAAW,GAAG,CAAC;IAC7B;AAEA,IAAA,OAAO,OAAO,CAAC,KAAK,CAAC,CAAC,EAAE,OAAO,CAAC,GAAG,SAAS,GAAG,OAAO,CAAC,KAAK,CAAC,SAAS,CAAC;AACzE;;;;;;;;;"}