feat(mcp): byte-exact SIE upload path + brevity and memory-first onboarding (#1954)

From the fourth E2E run: the agent correctly refused to reproduce a
104 KB SIE file token by token (silent mid-verifikat truncation) and
dead-ended to the web wizard, and its replies were walls of compliance
prose.

1. gnubok_create_sie_upload: signed same-origin upload URL (reuses the
   pending-document infra; .se/.sie/.si only, 50 MB HTTP cap).
   gnubok_sie_preflight and gnubok_import_sie accept upload_id as the
   byte-exact source, plus optional sha256 (hex of the raw bytes)
   verified on the upload_id/base64 paths so truncation is DETECTED,
   never silent. Inline content above 120k chars is refused with a
   pointer to the upload flow. Scope bookkeeping:write (same intent as
   import_sie).

2. Skill: brevity rule (max ~8 short lines per reply, one warning per
   step, no legal essays), memory-first rule (check what is already
   known before asking the opening questions), the upload-first SIE
   step, and gnubok_explain_voucher_gap after import for skipped
   voucher numbers.

3. CONNECTORS.md starter prompt rewritten memory-first so it stays
   copy-paste ready without the user's own data in it. Plugin v1.2.2.

tools/list ceiling 62K to 62.4K documented in the bench.

Co-authored-by: Jakob Wennberg <311770904+jakobwennberg-oss@users.noreply.github.com>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>

claude-plugin/.claude-plugin/plugin.json

@@ -2,7 +2,7 @@
   "name": "accounted",
   "displayName": "Accounted",
   "description": "Official Accounted plugin: Swedish double-entry bookkeeping flows for Claude. Connects your ledger over MCP and ships short workflow skills (daily bookkeeping, health check, month close, VAT, payroll, year-end) that work from the company's live data and load Swedish accounting knowledge from the product when needed. Every write is staged for your approval; nothing is booked on its own.",
-  "version": "1.2.1",
+  "version": "1.2.2",
   "author": {
     "name": "Accounted (erp-mafia)"
   },

claude-plugin/CONNECTORS.md

@@ -14,14 +14,9 @@ One-click add on claude.ai (opens the Add custom connector dialog prefilled; the
 https://claude.ai/customize/connectors?modal=add-custom-connector&connectorName=Accounted&connectorUrl=https%3A%2F%2Fapp.accounted.se%2Fapi%2Fextensions%2Fext%2Fmcp-server%2Fmcp%3Ftool_namespace%3Daccounted
 ```
 
-Pair the link with a starter prompt the user pastes as their first message. The short form works because the server-side onboarding skill carries the whole flow; the long form pre-answers the three opening questions and saves a round-trip:
+Pair the link with a starter prompt the user pastes as their first message. It stays copy-paste ready for everyone because it points the agent at what it already knows (Claude memory, earlier chats) instead of containing the user's own data; the server-side onboarding skill carries the rest of the flow:
 
-> Sätt upp mitt företag i Accounted.
+> Sätt upp mitt företag i Accounted. Utgå från det du redan vet om mig och mitt bolag (organisationsnummer, bank, tidigare bokföringssystem) och fråga bara efter det som saknas. Håll det kort.
-
-> Sätt upp mitt företag i Accounted.
-> Organisationsnummer: `<orgnr>`
-> Tidigare bokföringssystem: `<t.ex. Fortnox, eller inget>`
-> Bank: `<t.ex. Swedbank>`
 
 ## How the connection works
 

extensions/general/mcp-server/__tests__/payload-size.bench.test.ts

@@ -230,9 +230,13 @@ describe('tools/list payload size guard', () => {
     //     the scan-before-import step the skill instructs for shared SIE
     //     files, so default-catalog for the same Claude.ai reason; ~390
     //     tokens (schema carries the mappings-passthrough contract).
+    //   * 62K to 62.4K with gnubok_create_sie_upload + upload_id/sha256 on the
+    //     two SIE tools: the byte-exact upload path after a real 104 KB file
+    //     dead-ended in chat (a model cannot reproduce 30k tokens verbatim
+    //     without silent-truncation risk); skill-instructed, so default catalog.
     // Long-term answer to growth is leaning harder on gnubok_search_tools: if this
     // fires again, prefer trimming descriptions or making a tool opt-in via search
     // before bumping further.
-    expect(approxTokens).toBeLessThan(62_000)
+    expect(approxTokens).toBeLessThan(62_400)
   })
 })

extensions/general/mcp-server/__tests__/sie-preflight.test.ts

@@ -184,4 +184,45 @@ describe('gnubok_sie_preflight', () => {
   it('rejects a call with neither content field', async () => {
     await expect(run({})).rejects.toMatchObject({ code: 'VALIDATION_ERROR' })
   })
+
+  it('refuses oversized inline content instead of risking silent mid-verifikat truncation', async () => {
+    const huge = VALID_SIE + '\n' + '#KONTO 9999 "x"\n'.repeat(10_000)
+    await expect(run({ file_content: huge })).rejects.toMatchObject({
+      code: 'VALIDATION_ERROR',
+      message: expect.stringContaining('gnubok_create_sie_upload'),
+    })
+  })
+
+  it('verifies sha256 on the base64 path and rejects a mismatch as truncation', async () => {
+    const bytes = Buffer.from(VALID_SIE, 'utf8')
+    const { createHash } = await import('node:crypto')
+    const good = createHash('sha256').update(bytes).digest('hex')
+
+    const ok = await run({ file_content_base64: bytes.toString('base64'), sha256: good })
+    expect(ok.verdict).toBe('ok')
+
+    await expect(
+      run({ file_content_base64: bytes.toString('base64'), sha256: 'a'.repeat(64) })
+    ).rejects.toMatchObject({
+      code: 'VALIDATION_ERROR',
+      message: expect.stringContaining('sha256 mismatch'),
+    })
+  })
+})
+
+describe('gnubok_create_sie_upload', () => {
+  const uploadTool = tools.find((t) => t.name === 'gnubok_create_sie_upload')!
+
+  it('is registered with the filename-only contract', () => {
+    expect(uploadTool).toBeDefined()
+    expect(
+      (uploadTool.inputSchema as { required: string[] }).required
+    ).toEqual(['filename'])
+  })
+
+  it('rejects a filename that is not a SIE file', async () => {
+    await expect(
+      uploadTool.execute({ filename: 'export.xlsx' }, COMPANY_ID, 'user-1', {} as never)
+    ).rejects.toMatchObject({ code: 'VALIDATION_ERROR' })
+  })
 })

extensions/general/mcp-server/server.ts

@@ -257,8 +257,11 @@ import {
   createPendingDocumentUpload,
   uploadDocument,
   MAX_DOCUMENT_SIZE,
+  buildPendingDocumentStoragePath,
+  DOCUMENTS_BUCKET,
 } from '@/lib/core/documents/document-service'
 import { toSameOriginStorageUrl } from '@/lib/core/documents/storage-proxy'
+import { createHash } from 'node:crypto'
 import { extractInvoiceFields, ExtractionSchema as InvoiceExtractionSchema, AgentExtractionSchema } from '@/extensions/general/invoice-inbox/lib/extract-invoice-fields'
 import { mirrorExtractionToDocument } from '@/extensions/general/invoice-inbox/lib/mirror-extraction'
 // Skatteverket filing tools (PR5). Cross-extension lib import, same sanctioned
@@ -1445,22 +1448,101 @@ export function isDefaultCatalogTool(tool: { catalogVisibility?: 'default' | 'se
 }
 
 /**
- * Resolve SIE file content from tool args: plain text (the model read the
+ * Inline SIE content above this length is refused: a model reproducing tens
- * attachment) or base64 (exact bytes, e.g. from a code-execution sandbox).
+ * of thousands of tokens verbatim WILL eventually truncate mid-#VER, and a
- * The base64 path runs the same encoding detection as the HTTP upload route,
+ * truncated file that still parses imports silently incomplete bookkeeping.
- * so CP437 exports keep their åäö instead of arriving pre-mangled through a
+ * Larger files go through gnubok_create_sie_upload (raw bytes, no model in
- * host's UTF-8 read. Returns null when neither field is usable.
+ * the path). ~120k chars ≈ a few thousand verifikat rows.
  */
-async function decodeSieToolContent(args: Record<string, unknown>): Promise<string | null> {
+const MAX_INLINE_SIE_CHARS = 120_000
-  if (typeof args.file_content === 'string' && args.file_content.length > 0) {
+
-    return args.file_content
+/**
+ * Resolve SIE file content from tool args, three sources in priority order:
+ *
+ *   - upload_id: raw bytes PUT to a gnubok_create_sie_upload URL. The only
+ *     path with no model in the loop; required for large files.
+ *   - file_content_base64: exact bytes base64-encoded (e.g. from a
+ *     code-execution sandbox).
+ *   - file_content: plain text as the model read the attachment.
+ *
+ * Byte paths run the same encoding detection as the HTTP upload route, so
+ * CP437 exports keep their åäö. An optional sha256 (hex, of the RAW BYTES)
+ * is verified on the two byte-exact paths and proves the content was not
+ * truncated or altered in transit; it cannot be checked for plain text
+ * (the model's read may legitimately differ from the file's bytes).
+ * Returns null when no source is usable.
+ */
+async function resolveSieToolContent(
+  args: Record<string, unknown>,
+  companyId: string,
+  userId: string
+): Promise<string | null> {
+  const sha256 = typeof args.sha256 === 'string' ? args.sha256.trim().toLowerCase() : null
+  const verifyBytes = (buffer: Buffer, source: string) => {
+    if (!sha256) return
+    const actual = createHash('sha256').update(buffer).digest('hex')
+    if (actual !== sha256) {
+      throw Object.assign(
+        new Error(
+          `sha256 mismatch on ${source}: expected ${sha256}, got ${actual}. The content was truncated or altered; re-send it.`
+        ),
+        { code: 'VALIDATION_ERROR' }
+      )
+    }
   }
-  if (typeof args.file_content_base64 === 'string' && args.file_content_base64.length > 0) {
+
-    const { detectEncoding, decodeBuffer } = await import('@/lib/import/sie-parser')
+  const { detectEncoding, decodeBuffer } = await import('@/lib/import/sie-parser')
-    const buffer = Buffer.from(args.file_content_base64, 'base64')
+
+  if (typeof args.upload_id === 'string' && args.upload_id.length > 0) {
+    const fileName = typeof args.filename === 'string' ? args.filename : ''
+    if (!fileName) {
+      throw Object.assign(new Error('filename is required together with upload_id'), {
+        code: 'VALIDATION_ERROR',
+      })
+    }
+    const service = createServiceClientNoCookies()
+    const pendingPath = buildPendingDocumentStoragePath(companyId, userId, args.upload_id, fileName)
+    const { data, error } = await service.storage.from(DOCUMENTS_BUCKET).download(pendingPath)
+    if (error || !data) {
+      throw Object.assign(
+        new Error(
+          'No uploaded file found for this upload_id. PUT the raw file bytes to the upload_url from gnubok_create_sie_upload first (same upload_id and filename).'
+        ),
+        { code: 'NOT_FOUND' }
+      )
+    }
+    const buffer = Buffer.from(await data.arrayBuffer())
+    verifyBytes(buffer, 'the uploaded file')
     const arrayBuffer = buffer.buffer.slice(buffer.byteOffset, buffer.byteOffset + buffer.byteLength)
     return decodeBuffer(arrayBuffer, detectEncoding(arrayBuffer))
   }
+
+  if (typeof args.file_content_base64 === 'string' && args.file_content_base64.length > 0) {
+    if (args.file_content_base64.length > MAX_INLINE_SIE_CHARS) {
+      throw Object.assign(
+        new Error(
+          'File too large to pass inline safely. Use gnubok_create_sie_upload, PUT the raw bytes to its upload_url, and pass the upload_id here instead.'
+        ),
+        { code: 'VALIDATION_ERROR' }
+      )
+    }
+    const buffer = Buffer.from(args.file_content_base64, 'base64')
+    verifyBytes(buffer, 'file_content_base64')
+    const arrayBuffer = buffer.buffer.slice(buffer.byteOffset, buffer.byteOffset + buffer.byteLength)
+    return decodeBuffer(arrayBuffer, detectEncoding(arrayBuffer))
+  }
+
+  if (typeof args.file_content === 'string' && args.file_content.length > 0) {
+    if (args.file_content.length > MAX_INLINE_SIE_CHARS) {
+      throw Object.assign(
+        new Error(
+          'File too large to pass inline safely (silent mid-verifikat truncation risk). Use gnubok_create_sie_upload, PUT the raw bytes to its upload_url, and pass the upload_id here instead.'
+        ),
+        { code: 'VALIDATION_ERROR' }
+      )
+    }
+    return args.file_content
+  }
   return null
 }
 
@@ -16170,6 +16252,60 @@ export const tools: McpTool[] = [
     },
   },
 
+  {
+    name: 'gnubok_create_sie_upload',
+    title: 'Create SIE Upload',
+    description:
+      'Short-lived URL for a model-free SIE upload: PUT the raw .se/.sie bytes (max 50 MB) to upload_url, then pass upload_id (+ same filename) to gnubok_sie_preflight and gnubok_import_sie. Required for files too large to pass inline; add sha256 of the bytes there to prove integrity.',
+    inputSchema: {
+      type: 'object',
+      additionalProperties: false,
+      properties: {
+        filename: {
+          type: 'string',
+          minLength: 1,
+          maxLength: 255,
+          description: 'File name with extension, for example "export.se"',
+        },
+      },
+      required: ['filename'],
+    },
+    outputSchema: {
+      type: 'object',
+      additionalProperties: false,
+      properties: {
+        upload_id: { type: 'string' },
+        upload_url: { type: 'string' },
+        expires_at: { type: 'string' },
+      },
+      required: ['upload_id', 'upload_url', 'expires_at'],
+    },
+    annotations: {
+      readOnlyHint: false,
+      destructiveHint: false,
+      idempotentHint: false,
+      openWorldHint: false,
+    },
+    async execute(args, companyId, userId, supabase) {
+      const fileName = args.filename as string
+      const lower = fileName.toLowerCase()
+      if (!lower.endsWith('.se') && !lower.endsWith('.sie') && !lower.endsWith('.si')) {
+        throw Object.assign(new Error('filename must end in .se, .sie or .si'), {
+          code: 'VALIDATION_ERROR',
+        })
+      }
+      const uploadId = crypto.randomUUID()
+      const reservation = await createPendingDocumentUpload(supabase, companyId, userId, uploadId, fileName)
+      // Served from the app origin: agent sandboxes only reach the MCP host,
+      // not <project>.supabase.co. See storage-proxy.ts.
+      return {
+        upload_id: reservation.uploadId,
+        upload_url: toSameOriginStorageUrl(reservation.signedUrl),
+        expires_at: reservation.expiresAt,
+      }
+    },
+  },
+
   {
     name: 'gnubok_sie_preflight',
     title: 'SIE Preflight Scan',
@@ -16179,8 +16315,10 @@ export const tools: McpTool[] = [
       type: 'object',
       additionalProperties: false,
       properties: {
-        file_content: { type: 'string', description: 'Full SIE file contents as text' },
+        file_content: { type: 'string', description: 'Full SIE file contents as text (small files only)' },
-        file_content_base64: { type: 'string', description: 'Exact file bytes base64-encoded (preferred when available: preserves CP437 åäö)' },
+        file_content_base64: { type: 'string', description: 'Exact file bytes base64-encoded (preserves CP437 åäö)' },
+        upload_id: { type: 'string', description: 'From gnubok_create_sie_upload after PUTting the bytes; the only safe path for large files' },
+        sha256: { type: 'string', description: 'Hex sha256 of the raw file bytes; verified on upload_id/base64 paths to prove nothing was truncated' },
         filename: { type: 'string' },
       },
       required: ['filename'],
@@ -16205,11 +16343,13 @@ export const tools: McpTool[] = [
       idempotentHint: true,
       openWorldHint: false,
     },
-    async execute(args, companyId, _userId, supabase) {
+    async execute(args, companyId, userId, supabase) {
-      const content = await decodeSieToolContent(args)
+      const content = await resolveSieToolContent(args, companyId, userId)
       if (!content) {
         throw Object.assign(
-          new Error('Provide the SIE file as file_content (text) or file_content_base64 (exact bytes).'),
+          new Error(
+            'Provide the SIE file as upload_id (from gnubok_create_sie_upload), file_content_base64, or file_content.'
+          ),
           { code: 'VALIDATION_ERROR' }
         )
       }
@@ -16341,8 +16481,10 @@ export const tools: McpTool[] = [
       type: 'object',
       additionalProperties: false,
       properties: {
-        file_content: { type: 'string', description: 'Full SIE file contents' },
+        file_content: { type: 'string', description: 'Full SIE file contents (small files only)' },
         file_content_base64: { type: 'string', description: 'Exact file bytes base64-encoded; alternative to file_content (preserves CP437 åäö)' },
+        upload_id: { type: 'string', description: 'From gnubok_create_sie_upload after PUTting the bytes; the only safe path for large files' },
+        sha256: { type: 'string', description: 'Hex sha256 of the raw file bytes; verified on upload_id/base64 paths' },
         filename: { type: 'string', description: 'Original filename' },
         mappings: {
           type: 'array',
@@ -16363,12 +16505,12 @@ export const tools: McpTool[] = [
     async execute(args, companyId, userId, supabase, actor) {
       // Decoded once here; the staged params carry the decoded text so
       // commitImportSie re-parses exactly what was previewed.
-      const fileContent = await decodeSieToolContent(args)
+      const fileContent = await resolveSieToolContent(args, companyId, userId)
       const filename = args.filename as string
       const mappings = args.mappings as unknown[] | undefined
 
       if (!fileContent || !filename || !Array.isArray(mappings)) {
-        throw new Error('file_content (or file_content_base64), filename, and mappings are required')
+        throw new Error('file content (upload_id, file_content_base64 or file_content), filename, and mappings are required')
       }
 
       // Parse + validate at stage time so the approver sees real content (which

extensions/general/mcp-server/skills/onboarding.ts

@@ -16,6 +16,17 @@ continue straight into import or categorization. Staged writes still go
 through their normal approval, but that approval IS the conversation, not an
 extra pause around it.
 
+**Brevity rule: this is onboarding, not a course.** Keep every reply under
+~8 short lines. At most ONE warning per step, one line, and only when it
+changes what the user should do right now. No legal essays, no deadline
+tables, no cross-checks the user did not ask about (bolagsstämma dates,
+EU-moms edge cases, K-regelverk). The user can always ask for depth; the
+flow must never make them scroll past it.
+
+**Memory first.** Before asking the opening questions, check what you
+already know about the user (memory, earlier conversation): orgnr, company
+name, bank, previous system. Ask only for what is genuinely missing.
+
 ## When to use
 
 - "Sätt upp bokföring för mitt AB / min enskilda firma"
@@ -94,19 +105,28 @@ reaches far enough back anyway.
    Exportera data → SIE, **Björn Lundén / Briox / Wint** under Export.
    Every Swedish system exports SIE4 (.se/.sie); ask them to attach the
    file here in the chat.
-2. When the file arrives, call \`gnubok_sie_preflight\` with its content
+2. When the file arrives, get its BYTES to the server without retyping
-   (\`file_content\` as read, or \`file_content_base64\` when exact bytes are
+   them. Preferred (and REQUIRED for anything beyond a small file): call
-   available: that preserves åäö in CP437 exports). Summarize the scan:
+   \`gnubok_create_sie_upload\`, PUT the raw file bytes to the returned
-   source system, fiscal years, verifikat count, balance status, org-number
+   \`upload_url\` (from your code sandbox when you have one), compute the
-   match, warnings. This is the "does it look correct" moment: surface
+   file's sha256, then call \`gnubok_sie_preflight\` with \`upload_id\` +
-   problems BEFORE anything is written.
+   \`sha256\` + the same \`filename\`. Small files may go inline
-3. On their go-ahead: \`gnubok_import_sie\` with the same file and the
+   (\`file_content_base64\` + \`sha256\` preferred over plain
-   preflight's \`mappings\`. It stages for approval; after commit verify with
+   \`file_content\`). NEVER reproduce a large file token by token: the
-   \`gnubok_get_trial_balance\`.
+   tools refuse oversized inline content because a mid-verifikat
+   truncation imports silently incomplete bookkeeping.
+3. Summarize the preflight in a few lines: source system, fiscal years,
+   verifikat count, balance status, org-number match, the one warning that
+   matters. On the user's go-ahead: \`gnubok_import_sie\` with the same
+   source (\`upload_id\` or content) and the preflight's \`mappings\`. It
+   stages for approval; after commit verify with
+   \`gnubok_get_trial_balance\`, and explain any skipped voucher numbers
+   with \`gnubok_explain_voucher_gap\` (BFNAR 2013:2; an unexplained gap
+   blocks year-end).
 4. Multiple fiscal years = multiple files: import oldest first so IB/UB
-   chains. If the file is very large for chat, the web wizard at
+   chains. The web wizard at \`/import?mode=sie\` is the fallback when no
-   \`/import?mode=sie\` is the fallback; Fortnox users can also run the full
+   upload path works; Fortnox users can also run the full API migration
-   API migration (invoices, customers, documents) at
+   (invoices, customers, documents) at
    \`/import?mode=migration&provider=fortnox\`.
 
 ## Step 4: connect bank and Skatteverket (together, no pause)
@@ -136,6 +156,7 @@ message instead of making them ask.
 
 - \`gnubok_lookup_company\`: registry facts + prefill from the orgnr; call first
 - \`gnubok_create_company\`: preview (no confirm) then create (confirm=true)
+- \`gnubok_create_sie_upload\`: byte-exact upload URL for the SIE file
 - \`gnubok_sie_preflight\`: scan a shared SIE file, nothing written
 - \`gnubok_import_sie\`: staged import; use the preflight's mappings
 - \`gnubok_connect_bank\` / \`gnubok_connect_skatteverket\`: status + connect links

lib/auth/scope-catalog.ts

@@ -340,6 +340,8 @@ export const TOOL_SCOPE_MAP: Record<string, ApiKeyScope> = {
   gnubok_export_sie:                      'reports:read',
   gnubok_audit_package:                   'reports:read',
   gnubok_import_sie:                      'bookkeeping:write',
+  // Byte-exact SIE upload URL feeding gnubok_import_sie (same write intent).
+  gnubok_create_sie_upload:               'bookkeeping:write',
   // Rot/rut begäran om utbetalning (records a payout request on generate)
   gnubok_generate_rot_rut_file:           'invoices:write',
   // Supplier CRUD