@ixo/oracle-runtime-workers 0.29.1 → 0.31.0

package/src/attachments/pipeline.ts

@@ -22,7 +22,15 @@
  *
  * Both lanes share one `MAX_TOTAL_SIZE` download budget (`DownloadBudget`):
  * declared sizes are checked against what is left before a download, and
- * each download is cut off at what is left.
+ * each download is cut off at what is left. A file the budget cannot take
+ * is noted; the turn always runs.
+ *
+ * A channel file (`channelRef`, see `channel.ts`) is private: it is read
+ * through the owner object's channel source and always goes to a model as
+ * the bytes already loaded, never by URL; the model, the notes and the
+ * stored metadata know it only by its `channel:` reference. It may be up to
+ * the gateway's 50 MiB: one above `MAX_FILE_SIZE` is archived to the
+ * sandbox and noted, never handed to a model.
  */
 import type { MessageContent } from '@langchain/core/messages';
 import type { ModelInputCapabilities } from '../core/llm';
@@ -34,6 +42,12 @@ import {
   buildUserMessageContent,
   type NativeAttachment,
 } from './content-blocks';
+import {
+  CHANNEL_MAX_FILE_BYTES,
+  channelArchiveName,
+  ChannelFileUnavailableError,
+  type ChannelMediaSource,
+} from './channel';
 import {
   ALLOWED_URI_SCHEMES,
   AttachmentRejectedError,
@@ -59,6 +73,7 @@ import {
 } from './sandbox-archive';
 import {
   NOTE_EVENT_ID_MAX,
+  NOTE_FILENAME_MAX,
   NOTE_MIMETYPE_MAX,
   NOTE_URI_MAX,
   sanitizeAttachmentFilename,
@@ -73,6 +88,8 @@ export const SANDBOX_TRUNCATE_LIMIT = 500;
 
 export interface AttachmentPipelineDeps {
   source: MatrixMediaSource;
+  /** Where `channelRef` files are read; absent = none can be read (noted). */
+  channel?: ChannelMediaSource;
   /** Helper-model provider for the extract lane; null → media cannot be extracted here. */
   extraction: ExtractionProvider | null;
   /** Sandbox archive target; null → originals are not archived (logged once per turn). */
@@ -115,21 +132,34 @@ export interface PrepareAttachmentsParams {
   caps: ModelInputCapabilities;
 }
 
-function metaFor(attachment: AttachmentInput): AttachmentMeta {
+/** The metadata stored on the turn's messages (a channel file: its reference, never a link). */
+export function attachmentMetaFor(attachment: AttachmentInput): AttachmentMeta {
   return {
     filename: attachment.filename,
     mimetype: attachment.mimetype,
     ...(attachment.size !== undefined ? { size: attachment.size } : {}),
     ...(attachment.mxcUri ? { mxcUri: attachment.mxcUri } : {}),
     ...(attachment.eventId ? { eventId: attachment.eventId } : {}),
+    ...(attachment.channelRef ? { channelRef: attachment.channelRef } : {}),
     category: classifyAttachment(attachment),
   };
 }
 
+/** A filename as the hidden notes quote it (no controls, brackets or quotes). */
+function noteName(attachment: AttachmentInput): string {
+  return sanitizeNoteField(attachment.filename, NOTE_FILENAME_MAX);
+}
+
+/** A mimetype as the hidden notes show it. */
+function noteType(attachment: AttachmentInput): string {
+  return sanitizeNoteField(attachment.mimetype, NOTE_MIMETYPE_MAX);
+}
+
 /**
  * The line that names an attachment's reference to the model, for both lanes:
- * `[source: eventId="…"]` (the media event id) or `[source: url="…"]` (the
- * mxc / http URI), the same two values `attachmentRef` accepts.
+ * `[source: eventId="…"]` (the media event id), `[source: ref="channel:…"]`
+ * (a channel file) or `[source: url="…"]` (the mxc / http URI), the values
+ * `attachmentRef` accepts.
  */
 function sourceRef(attachment: AttachmentInput): string {
   // A reference the cleaner would alter is not the reference tools match on,
@@ -142,6 +172,10 @@ function sourceRef(attachment: AttachmentInput): string {
     ? clean(attachment.eventId, NOTE_EVENT_ID_MAX)
     : '';
   if (eventId) return `[source: eventId="${eventId}"]`;
+  const channelRef = attachment.channelRef
+    ? clean(attachment.channelRef, NOTE_EVENT_ID_MAX)
+    : '';
+  if (channelRef) return `[source: ref="${channelRef}"]`;
   const url = attachment.mxcUri ? clean(attachment.mxcUri, NOTE_URI_MAX) : '';
   if (url) return `[source: url="${url}"]`;
   return '';
@@ -151,7 +185,24 @@ function sourceRef(attachment: AttachmentInput): string {
 export function nativeAttachmentNote(attachment: AttachmentInput): string {
   const ref = sourceRef(attachment);
   if (!ref) return '';
-  return `${ref} "${sanitizeAttachmentFilename(attachment.filename)}" (${sanitizeNoteField(attachment.mimetype, NOTE_MIMETYPE_MAX)}) is attached to this message.`;
+  return `${ref} "${noteName(attachment)}" (${noteType(attachment)}) is attached to this message.`;
+}
+
+/** The sandbox archive name: a channel file's carries its digest (`photo-<12 hex>.jpg`). */
+function archiveName(attachment: AttachmentInput): string {
+  const safeName = sanitizeAttachmentFilename(attachment.filename);
+  return attachment.channelRef
+    ? channelArchiveName(safeName, attachment.channelRef)
+    : safeName;
+}
+
+/** The largest file of this source the turn downloads. */
+function fileCap(attachment: AttachmentInput): number {
+  return attachment.channelRef ? CHANNEL_MAX_FILE_BYTES : MAX_FILE_SIZE;
+}
+
+function megabytes(bytes: number): string {
+  return `${Math.round(bytes / 1024 / 1024)} MB`;
 }
 
 function errorMessage(error: unknown): string {
@@ -179,24 +230,30 @@ class DownloadBudget {
     params: PrepareAttachmentsParams,
     deps: AttachmentPipelineDeps,
   ): Promise<{ bytes: Uint8Array; mimetype: string; sniffed: boolean }> {
-    if (attachment.size && attachment.size > MAX_FILE_SIZE)
+    const perFile = fileCap(attachment);
+    if (attachment.size && attachment.size > perFile)
       throw new AttachmentTooLargeError('File exceeds maximum size');
     const remaining = this.remaining;
     if (remaining === 0 || (attachment.size && attachment.size > remaining))
       throw new AttachmentRejectedError(BUDGET_EXCEEDED);
-    const cap = Math.min(MAX_FILE_SIZE, remaining);
+    const cap = Math.min(perFile, remaining);
     try {
       const loaded = await loadAttachmentBytes(
         attachment,
         params.roomId,
         deps.source,
-        { fetchImpl: deps.fetchImpl, signal: deps.signal, maxBytes: cap },
+        {
+          fetchImpl: deps.fetchImpl,
+          signal: deps.signal,
+          maxBytes: cap,
+          ...(deps.channel ? { channel: deps.channel } : {}),
+        },
       );
       this.used += loaded.bytes.length;
       return loaded;
     } catch (error) {
       // Cut off by the budget rather than by the per-file cap.
-      if (error instanceof AttachmentTooLargeError && cap < MAX_FILE_SIZE)
+      if (error instanceof AttachmentTooLargeError && cap < perFile)
         throw new AttachmentRejectedError(BUDGET_EXCEEDED);
       throw error;
     }
@@ -215,7 +272,7 @@ async function extractOne(
   preloaded?: Uint8Array,
 ): Promise<{ text: string; usage?: ExtractionUsage }> {
   const safeName = sanitizeAttachmentFilename(attachment.filename);
-  if (!attachment.eventId && !attachment.mxcUri)
+  if (!attachment.eventId && !attachment.mxcUri && !attachment.channelRef)
     throw new Error('Either mxcUri or eventId must be provided');
   if (attachment.mxcUri && !ALLOWED_URI_SCHEMES.test(attachment.mxcUri))
     throw new Error('Invalid URI scheme');
@@ -226,7 +283,7 @@ async function extractOne(
       `[attachments] unsupported file type ${attachment.mimetype} for ${attachment.filename}`,
     );
     return {
-      text: `[File "${safeName}" (${attachment.mimetype}) is not a supported file type and could not be processed]`,
+      text: `[File "${noteName(attachment)}" (${noteType(attachment)}) is not a supported file type and could not be processed]`,
     };
   }
 
@@ -236,14 +293,19 @@ async function extractOne(
     deps.logger.warn(`[attachments] ${m}`),
   );
 
-  // Public http(s) images/videos: let the provider fetch the URL itself.
+  // Public http(s) images/videos: let the provider fetch the URL itself. A
+  // channel file has no URL here (only its reference): it goes as bytes.
   const isHttpUrl =
     Boolean(attachment.mxcUri) &&
     !attachment.eventId &&
+    !attachment.channelRef &&
     /^https?:\/\//i.test(attachment.mxcUri ?? '');
   let text: string | undefined;
   let usage: ExtractionUsage | undefined;
-  if (
+  if (bytes.length > MAX_FILE_SIZE) {
+    // A channel file larger than any model is handed here: kept, not read.
+    text = `[File "${noteName(attachment)}" (${noteType(attachment)}, ${megabytes(bytes.length)}) is larger than the ${megabytes(MAX_FILE_SIZE)} a model reads here, so its content was not read]`;
+  } else if (
     isHttpUrl &&
     deps.extraction &&
     (category === 'image' || category === 'video')
@@ -280,23 +342,28 @@ async function extractOne(
   if (!deps.sandbox) {
     return { text, ...(usage ? { usage } : {}) };
   }
-  const destPath = `${SANDBOX_OUTPUT_PREFIX}/${safeName}`;
+  const storedName = archiveName(attachment);
+  const destPath = `${SANDBOX_OUTPUT_PREFIX}/${storedName}`;
   try {
-    await uploadToSandbox(
+    // The path the sandbox reports is where the file is (it may differ).
+    const archived = await uploadToSandbox(
       bytes,
-      safeName,
+      storedName,
       destPath,
       deps.sandbox,
       attachment.mimetype,
       deps.fetchImpl,
     );
-    const actualPath = sanitizeSandboxPath(destPath);
+    const actualPath = sanitizeSandboxPath(archived.path);
     let analysisPath: string | undefined;
-    if (category === 'image' || category === 'video' || category === 'audio') {
-      const analysisFilename = `${safeName.replace(/\.[^.]+$/, '')}-analysis.md`;
+    if (
+      bytes.length <= MAX_FILE_SIZE &&
+      (category === 'image' || category === 'video' || category === 'audio')
+    ) {
+      const analysisFilename = `${storedName.replace(/\.[^.]+$/, '')}-analysis.md`;
       const analysisDest = `${SANDBOX_OUTPUT_PREFIX}/${analysisFilename}`;
       try {
-        await uploadToSandbox(
+        const analysis = await uploadToSandbox(
           new TextEncoder().encode(
             buildAnalysisMarkdown(
               safeName,
@@ -312,7 +379,7 @@ async function extractOne(
           'text/markdown',
           deps.fetchImpl,
         );
-        analysisPath = sanitizeSandboxPath(analysisDest);
+        analysisPath = sanitizeSandboxPath(analysis.path);
       } catch (error) {
         deps.logger.warn(
           `[attachments] analysis .md upload failed for "${attachment.filename}": ${errorMessage(error)}`,
@@ -350,7 +417,7 @@ export async function prepareAttachments(
   params: PrepareAttachmentsParams,
   deps: AttachmentPipelineDeps,
 ): Promise<PreparedAttachments> {
-  const metas = params.attachments.map(metaFor);
+  const metas = params.attachments.map((a) => attachmentMetaFor(a));
   const usage = { cost: 0, promptTokens: 0, completionTokens: 0 };
   if (params.attachments.length === 0) {
     return {
@@ -362,16 +429,9 @@ export async function prepareAttachments(
     };
   }
 
-  const reportedTotal = params.attachments.reduce(
-    (sum, a) => sum + (a.size ?? 0),
-    0,
-  );
-  if (reportedTotal > MAX_TOTAL_SIZE) {
-    throw new Error(
-      `Total attachment size (${Math.round(reportedTotal / 1024 / 1024)} MB) exceeds budget (${Math.round(MAX_TOTAL_SIZE / 1024 / 1024)} MB)`,
-    );
-  }
-
+  // Declared sizes over the budget are not refused here as a whole: each
+  // file is checked against what is left (`DownloadBudget`) and one that
+  // does not fit is noted, so the turn always runs.
   const nativeCandidates: AttachmentInput[] = [];
   // A file the native lane already refused (empty, over budget) carries the
   // reason and is noted; one it downloaded but will not send carries its
@@ -387,7 +447,12 @@ export async function prepareAttachments(
     deps.logger.log(
       `[attachments] "${attachment.filename}" (${attachment.mimetype}) kind=${kind} model=${params.model} → ${strategy}`,
     );
-    if (strategy === 'send-native') nativeCandidates.push(attachment);
+    // A file larger than a model is handed is never sent as a block.
+    if (
+      strategy === 'send-native' &&
+      !(attachment.size && attachment.size > MAX_FILE_SIZE)
+    )
+      nativeCandidates.push(attachment);
     else extractQueue.push({ attachment });
   }
 
@@ -409,6 +474,10 @@ export async function prepareAttachments(
         mimetype,
         filename: sniffed ? '' : attachment.filename,
       });
+      if (bytes.length > MAX_FILE_SIZE) {
+        extractQueue.push({ attachment, preloaded: bytes });
+        continue;
+      }
       if (kind !== classifyAttachment(attachment)) {
         deps.logger.warn(
           `[attachments] "${attachment.filename}" claims ${attachment.mimetype} but is ${mimetype} — not sent natively`,
@@ -420,20 +489,22 @@ export async function prepareAttachments(
         kind: kind === 'image' ? 'image' : 'file',
         mimeType: mimetype,
         base64: bytesToBase64(bytes),
-        filename: attachment.filename,
+        filename: sanitizeAttachmentFilename(attachment.filename),
       });
       const note = nativeAttachmentNote(attachment);
-      if (note) nativeNotes.push({ text: note, meta: metaFor(attachment) });
+      if (note)
+        nativeNotes.push({ text: note, meta: attachmentMetaFor(attachment) });
       deps.logger.log(
         `[attachments] NATIVE → "${attachment.filename}" sent directly to ${params.model} (${bytes.length} bytes, ${kind})`,
       );
       if (deps.sandbox) {
-        const safeName = sanitizeAttachmentFilename(attachment.filename);
-        // Off the hot path — the agent's file tools can reach the original later.
+        const storedName = archiveName(attachment);
+        // Off the hot path — the agent's file tools can reach the original
+        // later. The upload holds the bytes until it ends; nothing else does.
         void uploadToSandbox(
           bytes,
-          safeName,
-          `${SANDBOX_OUTPUT_PREFIX}/${safeName}`,
+          storedName,
+          `${SANDBOX_OUTPUT_PREFIX}/${storedName}`,
           deps.sandbox,
           mimetype,
           deps.fetchImpl,
@@ -468,8 +539,11 @@ export async function prepareAttachments(
         '[attachments] sandbox archive unavailable for this turn — originals are not archived',
       );
     }
-    for (const { attachment, refused, preloaded } of extractQueue) {
-      const meta = metaFor(attachment);
+    for (const entry of extractQueue) {
+      const { attachment, refused, preloaded } = entry;
+      // The queue lets go of a file's bytes as soon as its turn comes.
+      delete entry.preloaded;
+      const meta = attachmentMetaFor(attachment);
       let text: string;
       try {
         if (refused) throw refused;
@@ -491,7 +565,11 @@ export async function prepareAttachments(
         deps.logger.error(
           `[attachments] failed to process ${attachment.filename}: ${errorMessage(error)}`,
         );
-        text = `[File "${sanitizeAttachmentFilename(attachment.filename)}" (${sanitizeNoteField(attachment.mimetype, NOTE_MIMETYPE_MAX)}) failed to process: ${errorMessage(error)}]`;
+        // A channel file's error never names its link (`channel.ts`).
+        text =
+          error instanceof ChannelFileUnavailableError
+            ? `[File "${noteName(attachment)}" (${noteType(attachment)}) could not be downloaded: ${errorMessage(error)}]`
+            : `[File "${noteName(attachment)}" (${noteType(attachment)}) failed to process: ${errorMessage(error)}]`;
       }
       const ref = sourceRef(attachment);
       extracted.push({ text: ref ? `${ref}\n${text}` : text, meta });

package/src/attachments/retention.ts

@@ -16,6 +16,11 @@
  */
 import { HumanMessage, type BaseMessage } from '@langchain/core/messages';
 import type { AttachmentMeta } from '../sqlite/serialization';
+import {
+  NOTE_FILENAME_MAX,
+  NOTE_MIMETYPE_MAX,
+  sanitizeNoteField,
+} from './sanitize';
 
 /** Newest user turns whose inline attachment payloads stay in the messages. */
 export const ATTACHMENT_PAYLOAD_TURNS = 2;
@@ -90,9 +95,12 @@ export function isAttachmentViewMessage(message: BaseMessage): boolean {
   );
 }
 
-/** The reference `view_attachment` takes: the media event id, else the URI. */
+/**
+ * The reference `view_attachment` takes: the media event id, else a channel
+ * file's opaque reference, else the URI.
+ */
 export function attachmentRef(meta: AttachmentMeta): string | undefined {
-  return meta.eventId ?? meta.mxcUri;
+  return meta.eventId ?? meta.channelRef ?? meta.mxcUri;
 }
 
 /** The attachment metadata a message carries (`attachments`, else `attachment`). */
@@ -115,7 +123,7 @@ function formatSize(size: number | undefined): string {
 export function buildPlaceholderText(metas: AttachmentMeta[]): string {
   const lines = metas.map((meta, i) => {
     const ref = attachmentRef(meta);
-    return `${i + 1}. "${meta.filename}" (${meta.mimetype}${formatSize(meta.size)})${
+    return `${i + 1}. "${sanitizeNoteField(meta.filename, NOTE_FILENAME_MAX)}" (${sanitizeNoteField(meta.mimetype, NOTE_MIMETYPE_MAX)}${formatSize(meta.size)})${
       ref ? ` — ref: ${ref}` : ' — no reference; it cannot be fetched again'
     }`;
   });

package/src/attachments/sanitize.ts

@@ -6,33 +6,45 @@
 /** Where every attachment a user sends is archived in their sandbox. */
 export const SANDBOX_OUTPUT_PREFIX = '/workspace/output';
 
+/**
+ * `value` in at most `max` UTF-16 units, cut before a surrogate pair rather
+ * than through it (a multi-byte filename never ends in half a character).
+ */
+function cutUnits(value: string, max: number): string {
+  if (value.length <= max) return value;
+  const end = /[\ud800-\udbff]/.test(value[max - 1] ?? '') ? max - 1 : max;
+  return value.slice(0, end);
+}
+
 /**
  * Strip control characters and bracket sequences from a filename to prevent
  * prompt injection when interpolated into LLM context. The archive writes
  * under this exact name.
  */
 export function sanitizeAttachmentFilename(filename: string): string {
-  return (
+  return cutUnits(
     filename
       // eslint-disable-next-line no-control-regex
       .replace(/[\x00-\x1f\x7f]/g, '')
-      .replace(/[[\]]/g, '')
-      .slice(0, 255)
+      .replace(/[[\]]/g, ''),
+    255,
   );
 }
 
-/** Longest mimetype / event id / mxc URI written into a hidden note. */
+/** Longest filename / mimetype / event id / mxc URI written into a hidden note. */
+export const NOTE_FILENAME_MAX = 255;
 export const NOTE_MIMETYPE_MAX = 100;
 export const NOTE_EVENT_ID_MAX = 255;
 export const NOTE_URI_MAX = 512;
 
 /**
- * Clean a client-supplied value (mimetype, media event id, mxc URI) for the
- * hidden notes the model reads: the filename cleaner plus quotes removed
- * (the notes wrap refs in `"…"`), cut to `max` characters.
+ * Clean a client-supplied value (filename, mimetype, media event id, mxc
+ * URI) for the hidden notes the model reads: the filename cleaner plus
+ * quotes removed (the notes wrap names and refs in `"…"`), cut to `max`
+ * UTF-16 units without splitting a surrogate pair.
  */
 export function sanitizeNoteField(value: string, max: number): string {
-  return sanitizeAttachmentFilename(value).replace(/"/g, '').slice(0, max);
+  return cutUnits(sanitizeAttachmentFilename(value).replace(/"/g, ''), max);
 }
 
 /**

package/src/attachments/types.ts

@@ -11,6 +11,13 @@ export interface AttachmentInput {
   mxcUri?: string;
   /** Matrix event id (`$…`) — encrypted media is fetched and decrypted via the event. */
   eventId?: string;
+  /**
+   * A file the person sent on a channel (`channel:<16 hex>`, see
+   * `channel.ts`): private, read only through the owner object's channel
+   * source, never by URL. Set by the runtime for channel turns, never taken
+   * from a request body.
+   */
+  channelRef?: string;
   filename: string;
   mimetype: string;
   size?: number;

package/src/attachments/view.ts

@@ -7,6 +7,7 @@
  * are cached per session (`deps.cache`); native views are not (they are the
  * bytes themselves).
  */
+import type { ChannelMediaSource } from './channel';
 import { classifyAttachment } from './classify';
 import type { NativeAttachment } from './content-blocks';
 import { loadAttachmentBytes, type MatrixMediaSource } from './download';
@@ -20,7 +21,11 @@ import { categorizeFile, isPlainTextType, verifyMagicBytes } from './magic';
 import { extractPdfText } from './pdf-text';
 import { attachmentMetas, attachmentRef } from './retention';
 import { routeAttachment } from './route';
-import { sanitizeAttachmentFilename } from './sanitize';
+import {
+  NOTE_FILENAME_MAX,
+  NOTE_MIMETYPE_MAX,
+  sanitizeNoteField,
+} from './sanitize';
 import {
   LOCAL_TEXT_EXTRACTION,
   type AttachmentTextCache,
@@ -39,6 +44,8 @@ export type AttachmentView =
 
 export interface ViewAttachmentDeps {
   source: MatrixMediaSource;
+  /** Where `channel:` references are read (the owner object's links). */
+  channel?: ChannelMediaSource;
   extraction: ExtractionProvider | null;
   /** Capabilities of the model that will read the result. */
   caps: ModelInputCapabilities;
@@ -103,6 +110,16 @@ function toInput(meta: AttachmentMeta): AttachmentInput {
     ...(meta.size !== undefined ? { size: meta.size } : {}),
     ...(meta.mxcUri ? { mxcUri: meta.mxcUri } : {}),
     ...(meta.eventId ? { eventId: meta.eventId } : {}),
+    ...(meta.channelRef ? { channelRef: meta.channelRef } : {}),
+  };
+}
+
+/** How `deps` reads bytes: the turn's signal and fetch, and its channel source. */
+function downloadOptions(deps: ViewAttachmentDeps) {
+  return {
+    fetchImpl: deps.fetchImpl,
+    signal: deps.signal,
+    ...(deps.channel ? { channel: deps.channel } : {}),
   };
 }
 
@@ -118,7 +135,7 @@ export async function viewAttachment(
       attachment,
       deps.roomId,
       deps.source,
-      { fetchImpl: deps.fetchImpl, signal: deps.signal },
+      downloadOptions(deps),
     );
     deps.logger.log(
       `[attachments] view "${attachment.filename}" (${mimetype}) kind=${kind} → ${strategy} (${bytes.length} bytes)`,
@@ -147,7 +164,7 @@ export async function viewAttachment(
   if (category === 'unsupported') {
     return {
       kind: 'text',
-      text: `[File "${sanitizeAttachmentFilename(attachment.filename)}" (${attachment.mimetype}) is not a supported file type and could not be processed]`,
+      text: `[File "${sanitizeNoteField(attachment.filename, NOTE_FILENAME_MAX)}" (${sanitizeNoteField(attachment.mimetype, NOTE_MIMETYPE_MAX)}) is not a supported file type and could not be processed]`,
     };
   }
   // Text extracted earlier in this session by the same model is final.
@@ -170,7 +187,7 @@ export async function viewAttachment(
     attachment,
     deps.roomId,
     deps.source,
-    { fetchImpl: deps.fetchImpl, signal: deps.signal },
+    downloadOptions(deps),
   );
   deps.logger.log(
     `[attachments] view "${attachment.filename}" (${mimetype}) kind=${kind} → ${strategy} (${bytes.length} bytes)`,
@@ -202,7 +219,7 @@ function megabytes(bytes: number): string {
 /** The refusal for a scanned or binary file over the helper-model limit. */
 function modelLaneTooLarge(filename: string, bytes: number): Error {
   return new Error(
-    `"${sanitizeAttachmentFilename(filename)}" cannot be read: scanned or binary files up to ${megabytes(MAX_MODEL_LANE_BYTES)} MB can be processed; this one is ${megabytes(bytes)} MB. PDFs with a text layer and text files can be larger.`,
+    `"${sanitizeNoteField(filename, NOTE_FILENAME_MAX)}" cannot be read: scanned or binary files up to ${megabytes(MAX_MODEL_LANE_BYTES)} MB can be processed; this one is ${megabytes(bytes)} MB. PDFs with a text layer and text files can be larger.`,
   );
 }
 
@@ -223,7 +240,7 @@ export async function extractAttachmentText(
   const category = categorizeFile(attachment.mimetype);
   if (category === 'unsupported') {
     throw new Error(
-      `"${sanitizeAttachmentFilename(attachment.filename)}" (${attachment.mimetype}) is not a supported file type`,
+      `"${sanitizeNoteField(attachment.filename, NOTE_FILENAME_MAX)}" (${sanitizeNoteField(attachment.mimetype, NOTE_MIMETYPE_MAX)}) is not a supported file type`,
     );
   }
   const plain = category === 'document' && isPlainTextType(attachment.mimetype);
@@ -306,7 +323,7 @@ export async function extractAttachmentText(
     attachment,
     deps.roomId,
     deps.source,
-    { fetchImpl: deps.fetchImpl, signal: deps.signal },
+    downloadOptions(deps),
   );
   verifyMagicBytes(bytes, category, attachment, (m) =>
     deps.logger.warn(`[attachments] ${m}`),

package/src/channels/contract.ts

@@ -1,6 +1,11 @@
 import type { ActionOutcomeWire } from '@ixo/ixo-transaction/topic-proposals';
 import type { ActionOutcomeV1 } from '@ixo/action-protocol';
 import { z } from 'zod';
+import {
+  CHANNEL_MAX_FILE_BYTES,
+  channelAttachmentRef,
+} from '../attachments/channel';
+import type { AttachmentInput } from '../attachments/types';
 import type { ReplyPlan } from '../delivery/types';
 
 /**
@@ -109,6 +114,64 @@ export type ChannelControlOutcome =
   | { ok: true; result: ChannelControlResult }
   | Exclude<ChannelTurnOutcome, { ok: true }>;
 
+/** The longest filename a channel file keeps, in code points (the gateway's own bound). */
+export const CHANNEL_FILENAME_MAX = 200;
+
+/**
+ * A filename as the runtime keeps it: at most `CHANNEL_FILENAME_MAX` code
+ * points (a longer one is cut, never refused), and the file's kind when the
+ * name is empty, so a name alone never refuses a turn.
+ */
+export function channelFilename(filename: string, kind: string): string {
+  const clipped = Array.from(filename).slice(0, CHANNEL_FILENAME_MAX).join('');
+  return clipped || (kind === 'image' ? 'photo' : kind);
+}
+
+/**
+ * A file the person sent on the channel (a WhatsApp photo, video, document
+ * or audio file). The gateway downloads it from the provider and serves it
+ * from a capability link of its own; the bytes never ride in the body.
+ * `size` is bounded by the gateway's own per-file limit (50 MiB).
+ */
+export const ChannelAttachment = z
+  .strictObject({
+    url: z
+      .string()
+      .max(2000)
+      .url()
+      .refine((value) => value.startsWith('https://'), 'must be an https URL'),
+    filename: z.string(),
+    mimetype: z.string().min(1).max(100),
+    size: z.number().int().nonnegative().max(CHANNEL_MAX_FILE_BYTES),
+    sha256: z.string().regex(/^[0-9a-f]{64}$/),
+    kind: z.enum(['image', 'video', 'document', 'audio']),
+  })
+  .transform((attachment) => ({
+    ...attachment,
+    filename: channelFilename(attachment.filename, attachment.kind),
+  }));
+export type ChannelAttachment = z.infer<typeof ChannelAttachment>;
+
+/**
+ * The run's attachment inputs for channel files: each names its file by its
+ * opaque `channel:` reference, never by its link, so the model, the hidden
+ * notes and the stored metadata never see a link. The owner object reads
+ * the bytes through its channel source (`attachments/channel.ts`).
+ */
+export function channelAttachmentInputs(
+  attachments: readonly ChannelAttachment[],
+): AttachmentInput[] {
+  return attachments.map((attachment) => ({
+    channelRef: channelAttachmentRef(attachment.sha256),
+    filename: attachment.filename,
+    mimetype: attachment.mimetype,
+    size: attachment.size,
+  }));
+}
+
+/** Same cap as the attachment pipeline (`MAX_ATTACHMENTS`). */
+export const MAX_CHANNEL_ATTACHMENTS = 10;
+
 export const ChannelTurnBody = z.strictObject({
   provider: ChannelProviderId,
   bindingId: z.string().regex(/^chb_[A-Za-z0-9_-]{1,120}$/),
@@ -119,6 +182,12 @@ export const ChannelTurnBody = z.strictObject({
   /** An explicit correction stays in its current session, even with routing enabled. */
   sessionPolicy: z.literal('continue').optional(),
   message: z.string().min(1).max(16_000),
+  /** Files sent with the message; `message` carries the caption or a placeholder. */
+  attachments: z
+    .array(ChannelAttachment)
+    .min(1)
+    .max(MAX_CHANNEL_ATTACHMENTS)
+    .optional(),
   replyTo: ChannelReplyReference.optional(),
   /**
    * Which kind of oracle the gateway believes it addresses. `companion` is