@bevel-software/platform-core-backend 0.23.0 → 0.25.0

package/src/modules/workspace/workspace.tools.ts

@@ -1,6 +1,7 @@
 import { spawn } from 'node:child_process';
 import nodeFs from 'node:fs/promises';
 import { join } from 'node:path';
+import AdmZip from 'adm-zip';
 import type { Router, RequestHandler } from 'express';
 import type { LocalFilesystem } from '@mastra/core/workspace';
 import type { IToolRegistry, JsonSchema } from '../tool-registry/tool.contract.js';
@@ -16,11 +17,17 @@ import type { IRoutineWritePolicy } from './routine-write-policy.js';
 import type { ToolHandlerFactory } from '../tool-helpers/tool-handler.js';
 import { requireInternalSource, requireExternalSource } from '../tool-auth/tool-auth.middleware.js';
 import { workspaceIdForBranch } from '../../shared/workspace-id.js';
-import { assertBranchProvided } from '../../shared/domain-errors.js';
+import { assertBranchProvided, GitInternalsError, WorkflowValidationError } from '../../shared/domain-errors.js';
 // Leaf-level shared primitive (same exception `workspace.service.ts` already
 // relies on) — not a workflow service, so this stays inside the module boundary.
 import { assertValidBranchName } from '../kb-fs/branch-name.js';
-import { assertInsideRepo, assertRepoRootNameFreeArgs, normalizePathArgs } from '../kb-fs/repo-path.js';
+import {
+  assertInsideRepo,
+  assertRepoRootNameFree,
+  assertRepoRootNameFreeArgs,
+  isInsideRepo,
+  normalizePathArgs,
+} from '../kb-fs/repo-path.js';
 import { GitGuardedFilesystem } from '../kb-fs/git-guarded-filesystem.js';
 import { assertNoGitInternalsSegment, assertNotGitInternals, hasGitInternalsSegment } from '../../shared/git-internals.js';
 import { isRolesYamlPath } from '../access-model/roles-yaml-guard.js';
@@ -38,28 +45,36 @@ import { createFileReaderRegistry } from './file-readers/file-reader.registry.js
 import { DocumentReader } from './file-readers/document-reader.js';
 import { mcpImageResult } from '@bevel-software/platform-mcp-core';
 import {
-  LEGACY_AGENTS_FILE,
   folderPlaceholderPath,
   isFolderPlaceholder,
   isPlatformFile,
   isPlatformFolder,
   platformFileCreationRefusal,
-  platformFileNames,
   platformFileRefusal,
+  platformFileUploadRefusal,
   platformFolderRefusal,
   entryExistsMessage,
   type ExistingEntryKind,
-  type KbLayout,
 } from '@bevel-software/platform-shared';
 import type { KbContext } from '../../shared/kb-context.js';
 import { AccessDeniedError } from '../access-model/access-errors.js';
 import { removeEmptyDirs } from './empty-dirs.js';
-import { PROPOSAL_ROUTE_NOTE, rethrowAsWriteDenial } from './write-denial.js';
+import { rethrowAsWriteDenial } from './write-denial.js';
+import { sharedRulesPointer } from '../agent-instructions/shared-file-rules.js';
 import type { IChangeReadGate } from '../access-model/change-gate.js';
 import { notFound, orDeclaredNotFound, orNotFound } from './not-found.js';
 import { logger } from '../../shared/logging.js';
 import { printable } from '../../shared/printable.js';
 import { DestinationTakenError, inspectDestination } from '../../shared/rename-no-replace.js';
+import { AgentUploadStore, type ClaimedUpload } from './agent-upload.store.js';
+import {
+  isSymlinkZipEntry,
+  isZipNoiseEntry,
+  readZipEntry,
+  zipEntryName,
+  zipEntryNameRefusal,
+  zipEntrySegments,
+} from './zip-entry-rules.js';
 
 const log = logger('workspace-tools');
 
@@ -193,60 +208,26 @@ async function keepFolderOf(
   }
 }
 
-/**
- * Appended (centrally, in `mount`) to EVERY workspace tool description. The
- * platform's managed agent guide sits at the workspace root and documents the
- * conventions of that knowledge base; agents (ours and external) should consult
- * it before touching files. It rides on every entrypoint — reads (grep/
- * list_files/file_stat) included — because any of them can be a session's first
- * touch.
- *
- * `CLAUDE.md` is named as a fallback because knowledge bases seeded before the
- * rename still carry one, and the seeder never deletes a file it did not
- * expect. Naming both means an agent finds the conventions either way, instead
- * of reading none because it looked for the newer name and stopped.
- *
- * WHEN THE GUIDE HAS BEEN RENAMED the sentence names two files, ours first. The
- * second is the organisation's OWN `AGENTS.md`, which on such a deployment is
- * ordinary content the platform never touches — and which no harness reads for a
- * remote agent, because a remote agent has no checkout. Telling it to read both
- * is the only way the conventions the customer actually wrote reach the agent
- * working in their knowledge base. Under the default name the wording collapses
- * to the one file it has always named.
- *
- * A FUNCTION of the layout, called when a description is built: the name is a
- * deployment setting, and a module-scope string would snapshot the default.
- */
-function kbConventionsNote(layout: KbLayout): string {
-  const agentsFile = layout.agentsFile ?? LEGACY_AGENTS_FILE;
-  if (agentsFile === LEGACY_AGENTS_FILE) {
-    return ' Before your first read or change in a workspace, read `AGENTS.md` at the KB root — or `CLAUDE.md` on a knowledge base seeded before it was renamed — if either exists: it holds the author\'s conventions for this knowledge base, and you should follow them.';
-  }
-  return (
-    ` Before your first read or change in a workspace, read \`${agentsFile}\` at the KB root, then ` +
-    '`AGENTS.md` if it also exists (the organisation\'s own conventions) — or `CLAUDE.md` on a knowledge base seeded before it was renamed: together they hold the conventions for this knowledge base, and you should follow them.'
-  );
-}
-
-/** The platform files as a tool description lists them — the guide under its own name. */
-function platformFileList(layout: KbLayout): string {
-  return platformFileNames(layout)
-    .map((name) => `\`${name}\``)
-    .join(', ');
-}
-
 const int = (description: string): JsonSchema => ({ type: 'integer', description });
 
 const str = (description: string): JsonSchema => ({ type: 'string', description });
 
 /**
- * Where pictures go, on the two tools that write pages. An agent in core cannot
- * upload bytes yet (TODOS.md), but it can write the page with the link a person
- * will satisfy, and this sentence is what keeps every page it writes on the
- * README's convention: images beside the page, linked relatively.
+ * The upload route, named on every tool that takes content as a JSON string.
+ *
+ * ONE sentence, and the tools' own: it is what stops the three failures the
+ * route was built for. An agent landing 27 files read each one and typed it
+ * out again as a tool argument: a 37 KB write was truncated mid-answer, a page
+ * of regex backslashes failed to parse as a JSON parameter, and a PNG could not
+ * be sent at all. None of that is discoverable from a refusal — a truncated
+ * write reports success — so the tools that invite it name where the bytes
+ * should go instead, in the description itself, for a client that reads
+ * nothing else. WHY, and how the route is used, is one of the shared rules
+ * (`agent-instructions/shared-file-rules.ts`): said in full on three
+ * descriptions it took each of them past the length a client cuts at.
  */
-const IMAGE_CONVENTION_NOTE =
-  ' Images: keep them in an `assets/` folder next to the page that uses them and link them with a relative path, e.g. `![Approval screen](./assets/approval-screen.png)`; the page renders them inline.';
+const UPLOAD_ROUTE_NOTE =
+  ' Large, escape-heavy or binary content does not go through here: use `request_file_upload` + `apply_file_upload`.';
 
 /**
  * A path input that names the clone folder, and says what happens when it does
@@ -285,6 +266,9 @@ const RESERVED_ROOT_NAME_TARGETS: Readonly<Record<string, readonly string[]>> =
   copy_file: ['dest'],
   move_file: ['dest'],
   unzip: ['destination'],
+  // The folder the upload lands in. Each of its own paths is checked again
+  // inside the handler — an archive chooses its entry names, not the caller.
+  apply_file_upload: ['destination'],
 };
 
 function asText(content: string | Buffer): string {
@@ -313,16 +297,6 @@ interface DocGrepState {
   skippedUncached: number;
 }
 
-/**
- * THE binary capability contract, stated once and appended (in `mount`) to
- * every file tool's description — which is also what `tools_info` returns.
- * The split it states is enforced by the reader registry: the text tools
- * refuse what their reader marks not `textEditable` (and binary content under
- * any name) with a `binary_not_writable` refusal; the byte tools never look.
- */
-export const CONTENT_RULE =
-  ' Content rule (the same on every file tool): read_file returns text for text files and extracted text for documents (.docx/.pptx/.xlsx/.odt/.odp/.ods/.pdf, .eml/.msg); write_file, write_files and edit_file accept TEXT only — they refuse documents, images, archives and other binary files (legacy .doc/.ppt/.xls included) with kind `binary_not_writable`, naming the file\'s kind and the tool to use instead; copy_file, move_file, delete_file and unzip act on bytes of any kind; new binary content arrives through upload (`request_upload_token` + `apply_upload` where offered, otherwise Upload in the app). file_stat reports `contentMode` (`text` | `document` | `binary`) so you can decide before acting.';
-
 /** What a `binary_not_writable` refusal points to, in the order to try them. */
 const BINARY_USE_INSTEAD = ['upload', 'copy_file', 'move_file'] as const;
 
@@ -335,7 +309,7 @@ const BINARY_USE_INSTEAD = ['upload', 'copy_file', 'move_file'] as const;
 function binaryNotWritable(fileKind: FileKind, explanation: string): ToolError {
   return new ToolError(
     `${explanation} [binary_not_writable: this file's kind is ${fileKind}; write_file, write_files and edit_file accept text only. ` +
-      'Use upload for new bytes (`request_upload_token` + `apply_upload` where offered, otherwise Upload in the app), ' +
+      'Use upload for new bytes (`request_file_upload` + `apply_file_upload`, or Upload in the app), ' +
       'or copy_file / move_file to place bytes that are already in the workspace.]',
     415,
     { kind: 'binary_not_writable', fileKind, useInstead: [...BINARY_USE_INSTEAD] },
@@ -379,17 +353,17 @@ function assertNotDocumentEdit(readers: FileReaderRegistry, path: string): void
  *
  * Costs one read of the existing file, and only for readers that ask the
  * question. A path with nothing at it is a CREATE: there is nothing to destroy.
- * Returns the bytes it read (so a caller that needs the content next —
- * `edit_file` — does not read the file a second time), or undefined when it
- * had no reason to read or nothing existed.
+ * For the tools that REPLACE a file without needing what it held (`write_file`,
+ * `write_files`); `edit_file` holds the bytes already and asks
+ * `assertBytesTextEditable` of each reading it takes.
  */
 async function assertNotBinaryOverwrite(
   readers: FileReaderRegistry,
   path: string,
   fs: { readFile(p: string): Promise<string | Buffer> },
-): Promise<Buffer | undefined> {
+): Promise<void> {
   const reader = readers.readerFor(path);
-  if (reader.editRefusalForExisting === undefined) return undefined;
+  if (reader.editRefusalForExisting === undefined) return;
   let existing: Buffer;
   try {
     existing = asBytes(await fs.readFile(path));
@@ -398,12 +372,21 @@ async function assertNotBinaryOverwrite(
     // FileNotFoundError carry the disk's absence codes). Any other failure —
     // permissions, I/O — means the existing content could not be inspected:
     // propagate it rather than let the write destroy bytes the gate never saw.
-    if (isAbsence(err)) return undefined; // nothing there yet
+    if (isAbsence(err)) return; // nothing there yet
     throw err;
   }
-  const refusal = reader.editRefusalForExisting(existing, path);
+  assertBytesTextEditable(readers, path, existing);
+}
+
+/**
+ * The same refusal, over bytes the caller already holds. Split out so a tool
+ * that reads the file more than once — `edit_file`, before the lock and again
+ * under it — judges EVERY reading with the one rule, and the bytes it replaces
+ * are always bytes this gate has seen.
+ */
+function assertBytesTextEditable(readers: FileReaderRegistry, path: string, existing: Buffer): void {
+  const refusal = readers.readerFor(path).editRefusalForExisting?.(existing, path) ?? null;
   if (refusal !== null) throw binaryNotWritable('binary', refusal);
-  return existing;
 }
 
 /** What a write is ALLOWED to do at a path. `create` is the default everywhere. */
@@ -427,34 +410,6 @@ const WRITE_MODE_INPUT: JsonSchema = {
     'EXISTING file and refuses (`missing`) a path that holds nothing.',
 };
 
-/** The same three modes, said once, for both tool descriptions. */
-const WRITE_MODE_NOTE =
-  ' `mode` decides what may happen at a path and DEFAULTS TO `create`: `create` writes a new file and refuses a path that ' +
-  'already exists (`exists`, with the path — pass `mode: overwrite` to replace it), `overwrite` replaces what is there ' +
-  '(creating it if there is nothing), `update` replaces an existing file and refuses a path that does not exist (`missing`). ' +
-  'A refused path is left exactly as it was.';
-
-/**
- * What an agent needs to know about escape sequences in the content it sends,
- * on the three tools that take content as a JSON string.
- *
- * The three write routes — the MCP endpoint, the `/api/agent/tools/<name>`
- * route and `call_tool_chain` — were measured end to end against raw requests
- * and a byte-level read of the stored file (see
- * `__tests__/escape-sequences.routes.test.ts`): each stores content exactly as
- * the JSON string value decodes ONCE. So when an escape arrives already
- * decoded, the decoding happened in the client that built the request, and no
- * tool here can tell that content from content that was meant to be decoded.
- * Hence a warning rather than a fix, and the pointer to the one route whose
- * payload is bytes rather than a JSON string.
- */
-const ESCAPE_SEQUENCE_NOTE =
-  ' Escape sequences: some clients decode them in arguments before sending, so content meant to CONTAIN an escape rather ' +
-  'than what it stands for (the six characters backslash, `u`, `0`, `0`, `4`, `1`, say, rather than the letter `A`) can ' +
-  'reach this tool already decoded — what arrives is stored byte for byte, so when that distinction matters, verify what ' +
-  'landed (`read_file`, or a hash) and send such content through the upload route (`request_upload_token` + `apply_upload` ' +
-  'where offered, otherwise Upload in the app), which lands it unchanged.';
-
 /** The refusal `create` gives on a path that already holds something. */
 function pathExists(path: string): ToolError {
   return new ToolError(
@@ -486,6 +441,169 @@ function decideWrite(mode: WriteMode, path: string, exists: boolean): WriteOutco
   return exists ? 'replaced' : 'created';
 }
 
+/**
+ * How many of an upload's paths the answer NAMES before it stops and says how
+ * many there were. A 300-file zip's full outcome list is pages of text an agent
+ * pays for on every call; 25 is enough to see the shape of what happened, and
+ * `total` plus `truncated` say that there is more. `all: true` asks for the
+ * rest, for a caller that really does have to read each one.
+ */
+const APPLY_ANSWER_CAP = 25;
+
+/**
+ * How many entries of one uploaded archive are landed, and how many bytes of
+ * uncompressed content in total.
+ *
+ * Tighter than `unzip`'s own caps on purpose. An apply lands its whole set as
+ * ONE commit, which means every entry's bytes are held in memory at once —
+ * the property that makes the commit atomic is the one that makes a zip bomb
+ * expensive. The upload itself is already bounded by the deployment's upload
+ * limit; these bound what that upload is allowed to expand into.
+ */
+const APPLY_MAX_ENTRIES = 5_000;
+const APPLY_MAX_TOTAL_BYTES = 128 * 1024 * 1024; // 128 MB uncompressed
+
+/**
+ * One path of an upload, as `apply_file_upload` plans it: the bytes to write,
+ * or the reason this path is refused before any gate is asked. A refused path
+ * carries the name the archive held rather than a workspace path, because for
+ * those the whole problem is that no workspace path can be built from it.
+ */
+interface PlannedUploadPath {
+  path: string;
+  content?: Buffer;
+  error?: string;
+  message?: string;
+}
+
+/**
+ * Turn a stored upload into one planned path per file.
+ *
+ * A single file is one path: the destination plus the name it was sent with.
+ * A zip is one path per member, with the member's folder structure kept under
+ * the destination — judged by the same entry rules `unzip` applies
+ * (`zip-entry-rules.ts`), plus one `unzip` does not have: an entry that is a
+ * symbolic LINK is refused outright. A zip stores a link as a member whose
+ * bytes are its target text, so a reader that ignored the mode bits would
+ * write that text out as a file — content nobody sent, under a name that was
+ * meant to point elsewhere.
+ */
+/** The refusal an entry gets when the archive would expand past what one commit lands. */
+function tooLargeToApply(): string {
+  return `This archive expands past the ${APPLY_MAX_TOTAL_BYTES} byte total the apply lands in one commit; this entry was not applied.`;
+}
+
+async function planUpload(
+  upload: ClaimedUpload,
+  destination: string,
+  kbDirName: string,
+): Promise<PlannedUploadPath[]> {
+  const bytes = await nodeFs.readFile(upload.absolutePath);
+  if (upload.kind !== 'zip') {
+    return [{ path: `${destination}/${upload.filename}`, content: bytes }];
+  }
+  let zip: AdmZip;
+  try {
+    zip = new AdmZip(bytes);
+  } catch (err) {
+    throw new ToolError(
+      `"${upload.filename}" could not be opened as a .zip archive: ${err instanceof Error ? err.message : String(err)}`,
+      422,
+      { code: 'unreadable_archive' },
+    );
+  }
+  const planned: PlannedUploadPath[] = [];
+  let seen = 0;
+  let totalBytes = 0;
+  for (const entry of zip.getEntries()) {
+    const rawName = zipEntryName(entry.entryName);
+    if (isZipNoiseEntry(rawName)) continue;
+    if (seen >= APPLY_MAX_ENTRIES) {
+      planned.push({
+        path: rawName || '(empty)',
+        error: 'too_many_entries',
+        message: `This archive holds more than ${APPLY_MAX_ENTRIES} entries; the rest were not applied.`,
+      });
+      continue;
+    }
+    seen++;
+    const nameRefusal = zipEntryNameRefusal(rawName);
+    if (nameRefusal !== null) {
+      planned.push({ path: rawName || '(empty)', error: 'invalid_entry', message: nameRefusal });
+      continue;
+    }
+    if (isSymlinkZipEntry(entry)) {
+      planned.push({
+        path: rawName,
+        error: 'link',
+        message: `"${rawName}" is a symbolic link, not a file; an upload lands files, never links.`,
+      });
+      continue;
+    }
+    // A folder comes into being with the files under it (the write path mkdirs
+    // each parent), so a directory member has nothing of its own to land.
+    if (entry.isDirectory) continue;
+    const segments = zipEntrySegments(rawName);
+    const target = [destination, ...segments].join('/');
+    // Belt and braces: `zipEntryNameRefusal` already refuses a `..` segment
+    // and a root-anchored name, so nothing should reach here that climbs out.
+    // The check stays because the cost of being wrong about that is bytes
+    // landing outside the folder the caller named.
+    if (!target.startsWith(`${destination}/`) || !isInsideRepo(target, kbDirName)) {
+      planned.push({ path: rawName, error: 'invalid_entry', message: 'Path escapes destination' });
+      continue;
+    }
+    // Through the one bounded reader `unzip` uses too, capped at what is left
+    // of the budget: a deflate stream can expand a thousandfold, and the
+    // header's declared size is the archive's claim, not a fact — an entry
+    // declaring ZERO would otherwise be inflated with no cap at all (see
+    // `readZipEntry`). A read that fails is this entry's outcome and no more.
+    const read = readZipEntry(entry, APPLY_MAX_TOTAL_BYTES - totalBytes);
+    if (!read.ok) {
+      planned.push(
+        read.reason === 'too_large'
+          ? { path: rawName, error: 'too_large', message: tooLargeToApply() }
+          : { path: rawName, error: 'unreadable_entry', message: `"${rawName}" could not be read: ${read.detail}.` },
+      );
+      continue;
+    }
+    totalBytes += read.data.byteLength;
+    planned.push({ path: target, content: read.data });
+  }
+  return planned;
+}
+
+/**
+ * Record on `entry` that this path was refused, saying what the gate that
+ * refused it said. Four kinds of refusal count as one path's outcome: a
+ * typed tool refusal (`exists`, `platform_file`, the mode gate), a permission
+ * refusal (the caller may not write this path, where the DESTINATION was
+ * writable), the git folder in any spelling, and a path-shape refusal from the
+ * repository rules. Anything else
+ * is not a verdict about this path — it is a gate failing — so it travels on
+ * and the whole apply fails loudly, exactly as it does in `write_files`.
+ */
+function refuseEntry(entry: Record<string, unknown>, err: unknown): void {
+  entry.outcome = 'refused';
+  if (err instanceof ToolError) {
+    const details = (err.details ?? {}) as { code?: string; kind?: string };
+    entry.error = details.code ?? details.kind ?? 'refused';
+    entry.message = err.message;
+    return;
+  }
+  if (err instanceof AccessDeniedError) {
+    entry.error = 'write-denied';
+    entry.message = err.message;
+    return;
+  }
+  if (err instanceof GitInternalsError || err instanceof WorkflowValidationError) {
+    entry.error = (err.payload as { kind?: string } | undefined)?.kind ?? 'refused';
+    entry.message = err.message;
+    return;
+  }
+  throw err;
+}
+
 /** The three modes, as a set the handler can check a raw argument against. */
 const WRITE_MODES: readonly WriteMode[] = ['create', 'overwrite', 'update'];
 
@@ -730,6 +848,15 @@ export function registerWorkspaceTools(
    * verdict alone then decides, which differs only at a root.
    */
   changeGate?: IChangeReadGate,
+  /**
+   * The upload-token store behind `request_file_upload` / `apply_file_upload`
+   * — the route an agent lands bytes by, without their content passing
+   * through the model. Optional for the same reason the two above are: a tool
+   * harness that is about the file primitives need not stand one up. Every
+   * real composition wires it (`create-core-server.ts`), and without it the
+   * two tools are not mounted at all rather than mounted and broken.
+   */
+  uploads?: AgentUploadStore,
 ): void {
   const { kbDirName } = kb;
   /**
@@ -1347,10 +1474,14 @@ export function registerWorkspaceTools(
     handler: ToolHandler;
   }): void => {
     const path = `/api/agent/tools/${spec.name}`;
-    // Every workspace entrypoint carries the agent-guide reminder, every file
-    // tool the one content rule, and every tool a permission can refuse the
-    // proposal route — appended once here so no tool (especially the
-    // read-only ones a session hits first) can miss them.
+    // Every description ends with ONE sentence pointing at the rules these
+    // tools share — the content rule, the agent guide, the write modes, the
+    // dry-run protocol, the proposal route. They used to be appended here in
+    // FULL, which made a description several thousand characters of text the
+    // agent had already read on the tool above, and clients cut a long
+    // description from the END, where what is specific to the tool sits. The
+    // rules themselves are in the handshake instructions and in the managed
+    // guide (see `shared-file-rules.ts`), stated once and from one text.
     // Whether a call to this tool MUST name a branch, read off the tool's own
     // declaration rather than assumed of the family. Every tool mounted here
     // requires `branch` today; keying on the schema means a tool that declares
@@ -1359,10 +1490,8 @@ export function registerWorkspaceTools(
     const requiresBranch = ((spec.inputs as { required?: string[] }).required ?? []).includes('branch');
     const describe = (): string =>
       (typeof spec.description === 'function' ? spec.description() : spec.description) +
-      (spec.proposable ? PROPOSAL_ROUTE_NOTE : '') +
-      (spec.fileTool === false ? '' : CONTENT_RULE) +
-      kbConventionsNote(kb.layout) +
-      (spec.gated ? agentAccessGate.notes.gatedToolNote() : '');
+      (spec.gated ? agentAccessGate.notes.gatedToolNote() : '') +
+      sharedRulesPointer(kb.layout);
     const def = toolDef({
       name: spec.name,
       description: describe(),
@@ -1528,7 +1657,13 @@ export function registerWorkspaceTools(
     name: 'read_file',
     gated: true,
     description:
-      'Read a workspace file as text. Returns `{ path, content }`. Images (.png/.jpg/.jpeg/.gif/.webp) return the IMAGE ITSELF as native MCP image content (plus a one-line text note naming the file), so you can look at the picture — up to 3.5 MB of raw image data; a larger image gets an honest refusal asking for a locally downscaled copy or a smaller export (`.svg` is text and reads as text). Images come back only on a DIRECT call: inside `call_tool_chain` an image read yields an `{ image_omitted, note }` stub instead. Office and OpenDocument files (.docx/.pptx/.xlsx, .odt/.odp/.ods) and PDFs return their EXTRACTED text under an honest `[extracted text of …]` header, with `[slide N]`/`[sheet: Name]`/`[page N]` markers — the extraction is READ-ONLY (layout/images omitted; such files cannot be edited as text, only replaced by uploading a new version). Email files (.eml/.msg) return their EXTRACTED text the same way: a `[from]`/`[to]`/`[subject]`/`[date]` header block, the body (plain-text part preferred; an HTML-only body is stripped to text), and an `[attachments]` name list — attachments are listed, never extracted. Other binary files return a one-line description instead of raw bytes. Optional `offset`/`limit` slice the content (characters for a file, bytes for a `__tool_chain_spill__/…` ref; ignored for an image) — use them to page through large files or a `call_tool_chain` spill rather than reading multi-MB in full. A spill ref is workspace-independent: `branch` is ignored for it.',
+      'Read a workspace file as text. Returns `{ path, content }`. What comes back for a document, an email file, an image ' +
+      'or any other binary file is the content rule\'s business (see the shared rules): text files as text, documents and ' +
+      'email files as extracted text, an image as the picture itself, anything else as a one-line description. ' +
+      'Optional `offset`/`limit` slice the content (characters for a file, bytes for a `__tool_chain_spill__/…` ref; ignored ' +
+      'for an image) — use them to page through large files or a `call_tool_chain` spill rather than reading multi-MB in full. ' +
+      'It also reads a `__tool_chain_spill__/…` ref back from a truncated `call_tool_chain`: such a ref belongs to no ' +
+      'workspace, so `branch` is ignored for it.',
     inputs: {
       type: 'object',
       properties: {
@@ -1635,13 +1770,18 @@ export function registerWorkspaceTools(
   mount({
     name: 'file_stat',
     gated: true,
-    description: () =>
-      'Get a file/directory\'s metadata (name, type, size, …) without returning content. A file also reports `contentMode`: `text` (read, write and edit it as text), `document` (read returns an extraction; replace it by upload) or `binary` (bytes: copy, move, delete, or replace by upload), plus `kind` (`text` | `document` | `image` | `binary`), `mime`, `mimeSource` and `textEditable` — decided by the same file readers read_file, grep and the write tools use, so an extensionless text file is `text/plain`.' +
-      ' Every entry also reports what you may DO with it. ' +
-      `\`managed\` is true for a platform item — a platform file (${platformFileList(kb.layout)}) or a platform folder (the repository root or a reserved root folder such as \`KnowledgeBase/\`); managed items are never movable or deletable through these tools. ` +
-      '`access: { read, write, download, owner }` is your own verdict under the access rules; pass `explainAccess: true` to learn why, and who else holds each verb. `movable` and `deletable` say whether `move_file` / `delete_file` / `delete_folder` would be allowed for you, judged like their dry runs: not managed, no symbolic link, and on a protected branch you hold write on the item AND on every file under a folder (on a draft branch writes are not gated). `movable` judges the source side only; the destination is judged by a `move_file` dry run. ' +
-      'For a folder, `descendants` is the number of files under it at any depth; counting stops at 10000 and `descendantsTruncated` says so, and past that point `movable` and `deletable` are false because a folder that large was not judged in full — run the `move_file` or `delete_folder` dry run for the real verdict. ' +
-      'Call this before a move or delete to see what it would touch.',
+    description:
+      'Get a file/directory\'s metadata (name, type, size, …) without returning content, and what you may DO with it. ' +
+      'A file also reports `contentMode`, `kind`, `mime`, `mimeSource` and `textEditable` — decided by the same readers ' +
+      'read_file, grep and the write tools use, so an extensionless text file is `text/plain`. ' +
+      '`access: { read, write, download, owner }` is your own verdict under the access rules; pass `explainAccess: true` to ' +
+      'learn why, and who else holds each verb. ' +
+      'Call this before a move or delete: `managed`, `movable` and `deletable` answer the shared rules on what these tools ' +
+      'never move or delete, judged like the dry runs (on a draft branch writes are not gated); `movable` judges the SOURCE ' +
+      'side only, so the destination still wants a `move_file` dry run. ' +
+      'For a folder, `descendants` counts the files under it at any depth; counting stops at 10000 and ' +
+      '`descendantsTruncated` says so, past which `movable` and `deletable` are false — a folder that large was not judged ' +
+      'in full, so run the `move_file` or `delete_folder` dry run for the real verdict.',
     inputs: {
       type: 'object',
       properties: {
@@ -1949,9 +2089,7 @@ export function registerWorkspaceTools(
     description:
       'Write a workspace TEXT file. The change is committed + pushed as you. Returns `{ path, bytes, outcome }`, where `outcome` is ' +
       '`created`, `replaced` or `updated`.' +
-      WRITE_MODE_NOTE +
-      IMAGE_CONVENTION_NOTE +
-      ESCAPE_SEQUENCE_NOTE,
+      UPLOAD_ROUTE_NOTE,
     inputs: {
       type: 'object',
       properties: {
@@ -2039,9 +2177,7 @@ export function registerWorkspaceTools(
       '`created` / `replaced` / `updated` for a path it wrote, or `refused` with `error` (the code) and `message` (why) for a ' +
       'path it could not. `count` is how many were written. A path it refuses — the mode said no, or the file is not text — ' +
       'does not stop the others; read `files` to see what landed.' +
-      WRITE_MODE_NOTE +
-      IMAGE_CONVENTION_NOTE +
-      ESCAPE_SEQUENCE_NOTE,
+      UPLOAD_ROUTE_NOTE,
     inputs: {
       type: 'object',
       properties: {
@@ -2211,7 +2347,7 @@ export function registerWorkspaceTools(
     gated: true,
     description:
       'Replace an exact string in a workspace TEXT file. `old_string` must appear exactly once unless `replace_all`. Committed + pushed as you.' +
-      ESCAPE_SEQUENCE_NOTE,
+      UPLOAD_ROUTE_NOTE,
     inputs: {
       type: 'object',
       properties: {
@@ -2240,29 +2376,57 @@ export function registerWorkspaceTools(
       const path = a.path as string;
       const oldStr = a.old_string as string;
       const newStr = a.new_string as string;
-      // The overwrite gate already read the file when its reader asked the
-      // binary question — reuse those bytes instead of reading twice.
-      const content = await orNotFound(path, async () => {
-        const existing = await assertNotBinaryOverwrite(readers, path, fs);
-        return asText(existing ?? (await fs.readFile(path)));
-      });
-      const count = oldStr ? content.split(oldStr).length - 1 : 0;
-      if (count === 0) throw new ToolError('old_string not found in the file.', 400);
-      if (count > 1 && a.replace_all !== true) {
-        throw new ToolError(`old_string appears ${count} times — add more context to make it unique, or set replace_all.`, 400);
+      // Everything the tool decides about ONE reading of the file: may these
+      // bytes be edited as text at all, is `old_string` there, is it unique,
+      // and what the file becomes. One function, because the file is read
+      // twice — before the lock and under it — and a reading that skipped any
+      // of these questions would let bytes land that were never judged.
+      // `split`/`join`, not `String.replace`, which reads `$&`, `$'`, `` $` ``
+      // and `$$` in `new_string` as patterns and writes something the caller
+      // never sent.
+      const edit = (existing: Buffer): { updated: string; replaced: number } => {
+        assertBytesTextEditable(readers, path, existing);
+        const text = asText(existing);
+        const pieces = oldStr ? text.split(oldStr) : [text];
+        const count = pieces.length - 1;
+        if (count === 0) throw new ToolError('old_string not found in the file.', 400);
+        if (count > 1 && a.replace_all !== true) {
+          throw new ToolError(`old_string appears ${count} times — add more context to make it unique, or set replace_all.`, 400);
+        }
+        return { updated: pieces.join(newStr), replaced: count };
+      };
+      // A first verdict before any lock is taken, so an ordinary refusal costs
+      // no lock cycle. It is a verdict about a file anyone may still change.
+      let result = edit(await orNotFound(path, async () => asBytes(await fs.readFile(path))));
+      // The one the answer carries is taken again with the path's lock HELD,
+      // over the bytes read there (`write: true` guarantees the locking
+      // filesystem): read, verdict and write are one step nobody can get
+      // between. Taken before the lock only, two callers replacing the same
+      // text — two runners claiming a work item by filling its empty owner
+      // field — were BOTH told their edit landed, and the second silently
+      // overwrote the first. A filesystem without the method has no lock to
+      // read under, so the first verdict stands.
+      const locking = fs as unknown as {
+        rewriteFile?(path: string, rewrite: (current: Buffer | null) => string): Promise<void>;
+      };
+      if (typeof locking.rewriteFile === 'function') {
+        await locking.rewriteFile(path, (current) => {
+          if (current === null) throw notFound(path);
+          result = edit(current);
+          return result.updated;
+        });
+      } else {
+        await fs.writeFile(path, result.updated);
       }
-      const updated = a.replace_all === true ? content.split(oldStr).join(newStr) : content.replace(oldStr, newStr);
-      await fs.writeFile(path, updated);
-      return { path, replaced: a.replace_all === true ? count : 1, ...(await saveWarnings(ctx, path, updated)) };
+      return { path, replaced: result.replaced, ...(await saveWarnings(ctx, path, result.updated)) };
     },
   });
 
   mount({
     name: 'delete_file',
     gated: true,
-    description: () =>
-      'Delete ONE workspace file (a symbolic link is refused: links are never followed or removed). Committed + pushed as you. Its folder stays, even when this was its last file. Files only: a folder is refused with a pointer to `delete_folder`. ' +
-      `A platform file (\`access.md\` or \`.bevelignore\` in any folder, \`roles.yaml\` or \`${kb.layout.agentsFile}\` at the repository root) and git metadata are refused.`,
+    description:
+      'Delete ONE workspace file. Committed + pushed as you. Its folder stays, even when this was its last file. Files only: a folder is refused with a pointer to `delete_folder`.',
     inputs: {
       type: 'object',
       properties: {
@@ -2315,10 +2479,12 @@ export function registerWorkspaceTools(
     gated: true,
     description:
       'Delete a workspace FOLDER and every file under it, at any depth; the whole folder lands as ONE committed + pushed change as you — all of it or none of it — then the empty folder is removed. This is the one way a folder goes away: the folder that held it stays, even if this was all it had, and a folder holding nothing but its empty-folder placeholder counts as empty. ' +
-      'Preflight first: `dryRun: true` changes nothing and answers `{ path, kind: "folder", descendants, files, filesTruncated, allowed, reason? }` — `descendants` is the file count, `files` names up to 100 of them. ' +
-      'A non-empty folder is deleted only with `confirm: true`; without it the call deletes nothing and returns the same impact with `confirmationRequired: true`. Do NOT set `confirm: true` on your first call — dry-run, check the impact, then confirm. ' +
-      'Refused (in a dry run as `allowed: false` with the `reason`): a platform folder (the repository root or a reserved root folder such as `KnowledgeBase/`), git metadata, a folder holding a symbolic link (links are never removed), and a folder holding any file you may not write. A path that is a file is refused with a pointer to `delete_file`, and a path through a symbolic link is refused (links are never followed). ' +
-      'The folder\'s own platform files (`access.md`, `.bevelignore`) go with it in that same one change, so its files are never left ungoverned part-way; you must be able to write those platform files too.',
+      'The dry run answers `{ path, kind: "folder", descendants, files, filesTruncated, allowed, reason? }` — `descendants` is ' +
+      'the file count, `files` names up to 100 of them — and a non-empty folder wants `confirm: true`. ' +
+      'Beyond what the shared rules refuse, a folder HOLDING a symbolic link, or any file you may not write, is refused ' +
+      '(the link itself is never removed), and a path that is a FILE ' +
+      'is refused with a pointer to `delete_file`. You must be able to write the folder\'s own platform files too: they go ' +
+      'with it in that same one change, so its files are never left ungoverned part-way.',
     inputs: {
       type: 'object',
       properties: {
@@ -2481,11 +2647,15 @@ export function registerWorkspaceTools(
   mount({
     name: 'move_file',
     gated: true,
-    description: () =>
+    // A plain string again: what refuses a move names the guide, and that is in
+    // the shared rules now, which are rebuilt from the layout where they live.
+    description:
       'Move or rename a workspace FILE or FOLDER; a folder moves recursively, with everything under it. `dest` is the full new path, not the folder to move into. Lands as a delete + create, committed + pushed as you. ' +
-      `Rules: the destination must not exist — a move never overwrites a file or merges into a folder; a platform file (\`access.md\` or \`.bevelignore\` in any folder, \`roles.yaml\` or \`${kb.layout.agentsFile}\` at the repository root) is refused with "<name> is a platform file and stays in its folder." — a folder that moves takes its own platform files along, still in their folder; a platform folder (the repository root or a reserved root folder such as \`KnowledgeBase/\`) and git metadata are refused; a move cannot create a platform file or folder at \`dest\` either (renaming a note to \`access.md\` is refused); a path through a symbolic link is refused, since links are never followed; on a protected branch you must be able to write both ends — for a folder, every file under it at its old and its new path. ` +
-      'Access follows the destination folder. Preflight first: `dryRun: true` changes nothing and answers `{ src, dest, kind, descendants, access: { before, after }, accessChanges, allowed, reason? }` — `access` is your own `{ read, write, download, owner }` at the source and at the destination AS IT WILL BE once the move has landed, with every `access.md` inside a moved folder counted at its new place. ' +
-      'A move whose `accessChanges` is true runs only with `confirm: true`; without it the call moves nothing and returns the same impact with `confirmationRequired: true`. Do NOT set `confirm: true` on your first call — dry-run, check the impact, then confirm.',
+      'The destination must not exist — a move never overwrites a file or merges into a folder. Access follows the ' +
+      'DESTINATION folder, so a move can change what you (and others) may do with the file: the dry run answers ' +
+      '`{ src, dest, kind, descendants, access: { before, after }, accessChanges, allowed, reason? }`, where `access` is your ' +
+      'own `{ read, write, download, owner }` at the source and at the destination AS IT WILL BE once the move has landed, ' +
+      'with every `access.md` inside a moved folder counted at its new place, and a move whose `accessChanges` is true wants `confirm: true`.',
     inputs: {
       type: 'object',
       properties: {
@@ -2830,6 +3000,335 @@ export function registerWorkspaceTools(
     },
   });
 
+  // ── uploads (bytes that never pass through the model) ───────────────────
+  //
+  // The pair exists because MCP tool arguments are JSON. Every byte an agent
+  // sends through `write_file` is first typed out by the model, which
+  // truncates long files, mangles backslash and `\u` escapes, and cannot carry
+  // a PNG at all. `request_file_upload` answers an address; the agent POSTs
+  // the file (or one zip holding many) there with any HTTP client;
+  // `apply_file_upload` lands it on a branch in one commit. The bytes go from
+  // the agent's disk to the server's and never enter a prompt.
+  //
+  /**
+   * Why one of an upload's paths may not be landed, judged on the path ALONE —
+   * or undefined when nothing about the name itself refuses it.
+   *
+   * The platform files are the whole of it. `access.md` governs who may read
+   * and write the folder it sits in, `roles.yaml` says which roles exist, and
+   * the agent guide is read as instructions: each is configuration the platform
+   * obeys, and each has a write path that CHECKS the change (the roles gate
+   * refuses an edit that would lock every admin out; a folder's access rules
+   * are judged against who is asking). Bytes arriving by upload meet none of
+   * those gates — they are a buffer the sender chose — so an upload never
+   * lands one, whatever else the caller may write. `unzip` has refused
+   * `roles.yaml` from an archive for the same reason; this is that rule, over
+   * all four names.
+   */
+  const platformFileReason = (wsPath: string): string | undefined => {
+    const rel = toKbRelative(wsPath, kbDirName);
+    return rel !== null && isPlatformFile(rel, kb.layout) ? platformFileUploadRefusal(rel) : undefined;
+  };
+
+  /**
+   * `apply_file_upload`'s handler: resolve the stored bytes into one path per
+   * file, judge each path the way `write_files` judges its own, and land the
+   * survivors as ONE commit.
+   *
+   * The judging is deliberately the same shape as `write_files`, down to the
+   * second verdict under the lock, because the promise the ticket makes is
+   * that an upload is judged "exactly as `write_file` would judge it". Three
+   * gates run per path and a path that fails one is that path's outcome and no
+   * more: the deployment's write hook, the platform-file rule above, and the
+   * `mode`. What is judged ONCE for the whole call is the destination — a
+   * caller who may not write the folder at all gets one refusal naming the
+   * change-request route, rather than the same refusal repeated per entry.
+   */
+  const applyFileUpload = async (
+    a: Record<string, unknown>,
+    ctx: ToolContext,
+    uploads: AgentUploadStore,
+  ): Promise<unknown> => {
+    const branch = a.branch as string;
+    const token = a.token;
+    if (typeof token !== 'string' || token === '') {
+      throw new ToolError(
+        'Name the `token` `request_file_upload` answered with, after POSTing the file to its `uploadUrl`.',
+        400,
+        { code: 'token-required' },
+      );
+    }
+    const mode = modeOf(a);
+    const destination = (a.destination as string).replace(/\/+$/, '');
+    assertInsideRepo(destination, kbDirName);
+    // CLAIMED, not consumed: an apply refused whole (a protected destination,
+    // an archive that will not open) leaves the token alive so the caller can
+    // retry somewhere else rather than send the bytes again. The claim is what
+    // keeps it single-use meanwhile — a second apply finds the token in use.
+    const upload = uploads.claim(token, ctx.user.id);
+    let spent = false;
+    try {
+      const fs = await ctx.getFilesystem(branch);
+      const root = await workspaceRoot(branch, ctx);
+      // The destination, once, for the whole call. On a protected branch a
+      // caller who may not write the folder gets the lock gate's own refusal —
+      // which `rethrowAsWriteDenial` turns into `write-denied` with the
+      // change-request steps — and nothing lands.
+      const blockedDest = await writeBlocked(branch, ctx, [destination]);
+      if (blockedDest.length > 0) throw await writeRefusal(branch, blockedDest[0], 'dir');
+      if ((await kindOf(fs, destination)) === 'file') {
+        throw new ToolError(
+          `"${displayPath(destination)}" is a file, not a folder — \`destination\` names the folder the upload lands in.`,
+          409,
+          { code: 'not_a_folder' },
+        );
+      }
+
+      const planned = await planUpload(upload, destination, kbDirName);
+      const paths = planned.filter((p) => p.content !== undefined).map((p) => p.path as string);
+      // One batched access read for every path, like `write_files` — empty on
+      // a draft branch, where changes reach a protected branch only through a
+      // change request.
+      const blocked = new Set(await writeBlocked(branch, ctx, paths));
+
+      const writes: { path: string; content: Buffer }[] = [];
+      const outcomes: Record<string, unknown>[] = [];
+      /** The `files` entry for `writes[i]`, so the under-lock verdict can revise it. */
+      const entryOf: Record<string, unknown>[] = [];
+      for (const item of planned) {
+        const entry: Record<string, unknown> = { path: item.path };
+        outcomes.push(entry);
+        if (item.content === undefined) {
+          entry.outcome = 'refused';
+          entry.error = item.error;
+          entry.message = item.message;
+          continue;
+        }
+        const wsPathOf = item.path;
+        try {
+          if (blocked.has(wsPathOf)) throw await writeRefusal(branch, wsPathOf);
+          const platform = platformFileReason(wsPathOf);
+          if (platform !== undefined) throw new ToolError(platform, 422, { code: 'platform_file' });
+          // The git folder is never a workspace path, in any spelling. A ZIP
+          // entry's name has already met this rule in `zipEntryNameRefusal`; a
+          // SINGLE uploaded file's has not — `.git` is a name the upload
+          // route's `validateFilename` accepts — and the preflight that reads
+          // the caller's own arguments never sees it either, because the name
+          // came from the upload, not from the call. Asked here so that path
+          // is REFUSED like any other, with the rest of the upload landing,
+          // rather than failing the whole apply from inside `writeFiles`.
+          assertNoGitInternalsSegment(wsPathOf);
+          assertRepoRootNameFree(wsPathOf, kbDirName);
+          // A link already on disk under the destination must not redirect
+          // these bytes — the rule `unzip` applies per entry, applied here on
+          // the path the write will take.
+          const link = await symlinkOnPath(root, wsPathOf);
+          if (link !== undefined) {
+            throw new ToolError(
+              `"${wsPathOf}" goes through the symbolic link "${link}"; an upload never follows links.`,
+              400,
+              { code: 'symlink' },
+            );
+          }
+          writePolicy.assertPathWritable(ctx.sessionId, wsPathOf);
+          await assertAgentWriteAllowed(agentAccessGate, ctx, branch, wsPathOf);
+          // An earlier entry of this same upload counts as existing, as it
+          // does in `write_files`: two `create` entries for one path are a
+          // mistake the commit would otherwise hide.
+          const exists =
+            writes.some((w) => w.path === wsPathOf) || (await kindOf(fs, wsPathOf)) !== null;
+          entry.outcome = decideWrite(mode, wsPathOf, exists);
+          writes.push({ path: wsPathOf, content: item.content });
+          entryOf.push(entry);
+        } catch (err) {
+          refuseEntry(entry, err);
+        }
+      }
+
+      // The mode gate again, with every path's lock held — the verdict the
+      // answer carries, for the reason `write_file` states at length. A path
+      // whose verdict changed under the lock is dropped from the batch and
+      // reported refused, leaving the rest to land.
+      const recheck = async (
+        pending: readonly { path: string; content: Buffer }[],
+      ): Promise<{ path: string; content: Buffer }[]> => {
+        const kept: { path: string; content: Buffer }[] = [];
+        for (let i = 0; i < pending.length; i++) {
+          const entry = entryOf[i];
+          try {
+            const exists =
+              kept.some((k) => k.path === pending[i].path) || (await kindOf(fs, pending[i].path)) !== null;
+            entry.outcome = decideWrite(mode, pending[i].path, exists);
+            kept.push(pending[i]);
+          } catch (err) {
+            refuseEntry(entry, err);
+          }
+        }
+        return kept;
+      };
+      if (writes.length > 0) {
+        // `write: true` guarantees a LockingFilesystem here; `writeFiles` lands
+        // the whole set as ONE commit and takes a Buffer as content, so bytes
+        // reach disk exactly as they were sent — no text decode anywhere on
+        // the way, which is what makes a PNG and a backslash-heavy page land
+        // with the checksum they were uploaded with.
+        const batching = fs as unknown as {
+          writeFiles(
+            writes: { path: string; content: Buffer }[],
+            summary: string,
+            deletes: string[],
+            check: (
+              pending: readonly { path: string; content: Buffer }[],
+            ) => Promise<{ path: string; content: Buffer }[]>,
+          ): Promise<void>;
+        };
+        // In the DESTINATION folder's TURN, which `delete_folder` takes over the
+        // same subtree (and `keepFolderOf` with it). `writeFiles` creates the
+        // destination, and any folder above a zip entry on the way to it, as
+        // part of landing the batch — and a folder delete running between that
+        // creation and the commit enumerates the folder's files BEFORE these
+        // exist and then removes the folder they are landing in, which is an
+        // answer saying `created` for bytes that are already gone. The turn is
+        // taken OUTSIDE `writeFiles`, so it is held across the under-lock
+        // recheck and the commit both, and in the same order the delete takes
+        // its own (the folder's turn first, then each path's lock), which is
+        // what keeps two callers from waiting on each other's half.
+        await ctx.workspaceService.withFolderTurn(workspaceIdForBranch(branch), destination, async () => {
+          await batching.writeFiles(writes, `Apply upload of ${writes.length} file(s)`, [], recheck);
+        });
+      }
+      // The token is spent once an ANSWER exists, even an answer in which
+      // every path was refused: the apply ran and said what happened at each
+      // path, and re-running it would say the same. Only a refusal that landed
+      // nothing AND answered nothing (thrown above) gives the token back.
+      spent = true;
+      await uploads.consume(token);
+      const listed = a.all === true ? outcomes : outcomes.slice(0, APPLY_ANSWER_CAP);
+      return {
+        destination,
+        count: outcomes.filter((o) => o.outcome !== 'refused').length,
+        total: outcomes.length,
+        files: listed,
+        ...(listed.length < outcomes.length ? { truncated: true } : {}),
+      };
+    } finally {
+      if (!spent) uploads.release(token);
+    }
+  };
+
+  // Mounted only when the composition supplied a store — see the `uploads`
+  // parameter. Core always does.
+  if (uploads) {
+    mount({
+      name: 'request_file_upload',
+      fileTool: false,
+      description:
+        // Within the description cap (`tool-registry/description-length.ts`):
+        // why a file goes this way is one of the shared rules, and the header
+        // spelling of the token is on the `uploadUrl` output, where the
+        // address it changes is.
+        'Ask for a one-time address to send FILE BYTES to, so their content never passes through this conversation. ' +
+        'Use it for anything `write_file` cannot carry faithfully: a large file, a file full of backslashes or `\\u` ' +
+        'escapes, a binary file (a PNG, a PDF, a zip), or many files at once (zip them). ' +
+        'Returns `{ uploadUrl, token, expiresAt, expiresInSeconds, maxBytes }`. THEN: ' +
+        '(1) POST the file as the raw request body to `uploadUrl` with `?filename=<name>` — ' +
+        '`curl -X POST --data-binary @skill.zip "<uploadUrl>?filename=skill.zip"` — which answers what it received; ' +
+        '(2) call `apply_file_upload` with the same `token`, a `branch` and a destination folder. ' +
+        'One token carries one file or one zip, is bound to you and expires at `expiresAt`: an upload nobody applies ' +
+        'by then is deleted, and one over `maxBytes` is refused when you send it, naming the limit.',
+      inputs: { type: 'object', properties: {}, additionalProperties: false },
+      outputs: {
+        type: 'object',
+        properties: {
+          uploadUrl: str(
+            'The absolute URL to POST the bytes to. Carries the token; add `?filename=<name>`. To keep the token out ' +
+              'of a URL — when the command line you send from is logged or shared — POST to this address without its ' +
+              'last (token) segment and send the token in an `x-upload-token` header instead.',
+          ),
+          token: str(
+            'The token itself — what `apply_file_upload` takes, and what an `x-upload-token` header carries when you ' +
+              'would rather it not sit in a URL. Treat it as a credential.',
+          ),
+          expiresAt: str('ISO-8601 instant after which the token, and any bytes sent with it, are gone.'),
+          expiresInSeconds: int('Seconds from now until `expiresAt`.'),
+          maxBytes: int('The largest upload this deployment accepts, in bytes.'),
+        },
+        required: ['uploadUrl', 'token', 'expiresAt', 'expiresInSeconds', 'maxBytes'],
+      },
+      // A read-scoped caller has nothing to do with an upload token: the only
+      // thing it unlocks is a write. Refused at the handler factory, by scope,
+      // before the token is minted.
+      write: true,
+      handler: async (_a, ctx: ToolContext) => uploads.issue(ctx.user),
+    });
+
+    mount({
+      name: 'apply_file_upload',
+      gated: true,
+      description:
+        'Land a file you have already uploaded (see `request_file_upload`) in a folder on a branch, in ONE commit, as you. ' +
+        'A single file lands under the name it was sent with; a zip lands as its entries, keeping their folder structure. ' +
+        'Returns `{ destination, count, total, files }`: one entry per path, each `{ path, outcome }` — `created` / ' +
+        '`replaced` / `updated`, or `refused` with `error` (the code) and `message` (why). `count` is how many landed and ' +
+        '`total` how many paths there were; `files` is cut to the first 25 unless you pass `all: true`. ' +
+        'Every path is judged one by one — by your write access, the platform-file rules and what is already there — ' +
+        'exactly as `write_file` judges it, and a refused path does not stop the others. ' +
+        'The token is single-use: it is spent by the apply that lands it, and refused if you use it twice, let it ' +
+        'expire, or present one issued to somebody else. `mode` means what it means on `write_file`.',
+      inputs: {
+        type: 'object',
+        properties: {
+          branch: BRANCH_INPUT,
+          token: str('The `token` from `request_file_upload`, after you have POSTed the file to its `uploadUrl`.'),
+          destination: wsPath(kbDirName, 'Folder the upload lands in (created if it is not there yet)'),
+          mode: WRITE_MODE_INPUT,
+          all: {
+            type: 'boolean',
+            description:
+              'List EVERY path in `files` instead of the first 25. `total` always says how many there were, so ask for ' +
+              'all only when you need to read each outcome.',
+          },
+          sessionId: SESSION_ID_INPUT,
+        },
+        required: ['branch', 'token', 'destination'],
+        additionalProperties: false,
+      },
+      outputs: {
+        type: 'object',
+        properties: {
+          destination: str('The folder the upload was applied to (echoes the input).'),
+          count: int('How many paths landed — the entries in `files` whose `outcome` is not `refused`.'),
+          total: int('How many paths the upload held, whether or not `files` lists them all.'),
+          files: {
+            type: 'array',
+            description: 'One entry per path, in the order the upload held them. Cut to 25 unless `all` was true.',
+            items: {
+              type: 'object',
+              properties: {
+                path: str('The workspace path this entry was judged at.'),
+                outcome: {
+                  type: 'string',
+                  enum: ['created', 'replaced', 'updated', 'refused'],
+                  description: 'What happened at this path. `refused` means nothing was written there.',
+                },
+                error: str('Present when `outcome` is `refused`: the refusal code — e.g. `exists`, `missing`, `invalid_entry`, `platform_file`, `write-denied`.'),
+                message: str('Present when `outcome` is `refused`: the full refusal, the same one `write_file` would have given.'),
+              },
+              required: ['path', 'outcome'],
+            },
+          },
+          truncated: { type: 'boolean', description: 'True when `files` was cut: `total` is larger than what it lists. Pass `all: true` for the rest.' },
+        },
+        required: ['destination', 'count', 'total', 'files'],
+      },
+      write: true,
+      // So a protected-branch refusal arrives as `write-denied`, with the
+      // change-request steps, exactly as it does from write_file.
+      proposable: true,
+      handler: async (a, ctx: ToolContext) => applyFileUpload(a, ctx, uploads),
+    });
+  }
+
   // ── shell (internal-only) ───────────────────────────────────────────────
   mount({
     name: 'execute_command',