@stratta/mcp 1.16.0 → 1.17.0

package/README.md

@@ -46,6 +46,21 @@ This package is what you want when:
   uploads a file from your disk to a project dossier;
 - you would simply rather run the server yourself.
 
+## Lighter setup: `--simple`
+
+By default the server announces all 59 tools (about 67,000 characters of
+schemas). `--simple` announces only the 10 norm-reading tools (about 11,000),
+which is enough for an agent that just reads norms:
+
+```bash
+claude mcp add stratta --scope user -- npx -y @stratta/mcp --simple
+```
+
+`--toolsets dossiers,sites` adds families (`norms` is always on; `dossiers`,
+`sites`, `ingest`), `--full` restores everything, and the `STRATTA_TOOLSETS`
+environment variable does the same. An unknown family is ignored. For the
+remote connector, use the **Tools** menu on https://stratta.ch/connections.
+
 ## Install
 
 ### Claude Code (recommended)
@@ -209,42 +224,57 @@ serves the same 58, everything minus `add_attachment`).
 | `list_site_photos`         | The photos uploaded on the sheet, with position, direction, instant and note. No image, no link.                                    |
 | `list_site_neighbours`     | The organisation's other sheets within a radius, with the distance and the geological unit their scan established.                  |
 
-**Dossier** (25 tools — keep what was decided on a project, and everything the dossier page can do to a question):
-
-| Tool                   | Purpose                                                                                                       |
-| ---------------------- | ------------------------------------------------------------------------------------------------------------- |
-| `list_dossiers`        | Your organisation's dossiers, most recently touched first, with open-question counts and site id.             |
-| `open_dossier`         | Open a project's dossier, creating it if needed. Idempotent on the name.                                      |
-| `open_question`        | Open one question to settle, with optional named options. Idempotent on the title.                            |
-| `save_finding`         | Record one piece of evidence: a cited article, a retained value and why, an observation.                      |
-| `record_decision`      | Settle a question with a decision the engineer has confirmed, and the retained option.                        |
-| `load_dossier`         | Reload everything: questions with their evidence and decisions, open ones first.                              |
-| `resolve_question`     | Close a question without a decision, or reopen one. The evidence stays.                                       |
-| `list_attachments`     | The project attachments of a dossier: site reports, borehole logs, minutes, data sheets.                      |
-| `read_attachment`      | Read an attachment's text as Markdown, page by page.                                                          |
-| `search_in_dossier`    | Full-text search over a dossier's attachments, with the page of each hit.                                     |
-| `add_attachment`       | Upload a file from the user's machine to a dossier (PDF, DOCX, XLSX, images, text). Local server only.        |
-| `list_templates`       | The checklists the organisation wrote for its types of structure.                                             |
-| `apply_template`       | Open a template's questions in a dossier and file the clauses that resolve in the corpus.                     |
-| `draft_deliverable`    | Open the document a dossier produces (project basis, use agreement, report) and get its plan and sources.     |
-| `write_section`        | Write one section of a deliverable with what it rests on; a section without a source is refused.              |
-| `list_skills`          | The procedures the agent follows for a repeatable task: the ones Stratta ships and the ones the office wrote. |
-| `get_skill`            | Read one procedure in full; the office version wins over the Stratta one of the same name.                    |
-| `list_questions`       | The questions of a dossier, paginated, with status, assignee, due date, options and decision title.           |
-| `get_question`         | One question in full: options, decision, evidence with every citation field, comments with authors.           |
-| `get_dossier_activity` | The history of a dossier, most recent first, paginated: who did what, from the app or an agent.               |
-| `list_exports`         | The verification notes and journals exported from a dossier, with hash and trusted timestamp.                 |
-| `update_question`      | Reword a question, assign it by e-mail address, set or clear its due date.                                    |
-| `add_option`           | Add one way of settling a question. Idempotent on the name.                                                   |
-| `update_option`        | Rename or describe an option, or mark it retained (the others are released).                                  |
-| `delete_option`        | Remove an option; its evidence stays on the question. Asks the user first.                                    |
-| `delete_question`      | Delete a question opened by mistake; its evidence stays, unfiled. Asks the user first.                        |
-| `add_comment`          | Leave a signed remark on a dossier, a question or an entry.                                                   |
-| `update_entry`         | Correct an entry's wording, value, confidence or citation in place.                                           |
-| `attach_entry`         | File an entry under a question and optionally an option, or unfile it.                                        |
+**Dossier** (27 tools — what a project reads, retains and produces: questions, evidence, decisions, attachments, templates, and the deliverable written section by section):
+
+| Tool                   | Purpose                                                                                                   |
+| ---------------------- | --------------------------------------------------------------------------------------------------------- |
+| `list_dossiers`        | Your organisation's dossiers, most recently touched first, with open-question counts and site id.         |
+| `open_dossier`         | Open a project's dossier, creating it if needed. Idempotent on the name.                                  |
+| `open_question`        | Open one question to settle, with optional named options. Idempotent on the title.                        |
+| `save_finding`         | Record one piece of evidence: a cited article, a retained value and why, an observation.                  |
+| `record_decision`      | Settle a question with a decision the engineer has confirmed, and the retained option.                    |
+| `load_dossier`         | Reload everything: questions with their evidence and decisions, open ones first.                          |
+| `resolve_question`     | Close a question without a decision, or reopen one. The evidence stays.                                   |
+| `list_attachments`     | The project attachments of a dossier: site reports, borehole logs, minutes, data sheets.                  |
+| `read_attachment`      | Read an attachment's text as Markdown, page by page.                                                      |
+| `search_in_dossier`    | Full-text search over a dossier's attachments, with the page of each hit.                                 |
+| `add_attachment`       | Upload a file from the user's machine to a dossier (PDF, DOCX, XLSX, images, text). Local server only.    |
+| `list_templates`       | The checklists the organisation wrote for its types of structure.                                         |
+| `apply_template`       | Open a template's questions in a dossier and file the clauses that resolve in the corpus.                 |
+| `draft_deliverable`    | Open the document a dossier produces (project basis, use agreement, report) and get its plan and sources. |
+| `write_section`        | Write one section of a deliverable with what it rests on; a section without a source is refused.          |
+| `list_questions`       | The questions of a dossier, paginated, with status, assignee, due date, options and decision title.       |
+| `get_question`         | One question in full: options, decision, evidence with every citation field, comments with authors.       |
+| `get_dossier_activity` | The history of a dossier, most recent first, paginated: who did what, from the app or an agent.           |
+| `list_exports`         | The verification notes and journals exported from a dossier, with hash and trusted timestamp.             |
+| `update_question`      | Reword a question, assign it by e-mail address, set or clear its due date.                                |
+| `add_option`           | Add one way of settling a question. Idempotent on the name.                                               |
+| `update_option`        | Rename or describe an option, or mark it retained (the others are released).                              |
+| `delete_option`        | Remove an option; its evidence stays on the question. Asks the user first.                                |
+| `delete_question`      | Delete a question opened by mistake; its evidence stays, unfiled. Asks the user first.                    |
+| `add_comment`          | Leave a signed remark on a dossier, a question or an entry.                                               |
+| `update_entry`         | Correct an entry's wording, value, confidence or citation in place.                                       |
+| `attach_entry`         | File an entry under a question and optionally an option, or unfile it.                                    |
 
 A dossier is read, annotated, reviewed and exported from
-[stratta.ch/dossiers](https://stratta.ch/dossiers).
+[stratta.ch/dossiers](https://stratta.ch/dossiers): four tabs (Terrain,
+Lectures, Décisions, Livrables). Every section the agent reads for a named
+project is logged under Lectures without a gesture; a decision can be recorded
+in one sentence (`record_decision` with `dossierId` and `question`); the
+deliverable is draft, in review or signed, exported to Word inside the office's
+own report, and signing never goes through an agent.
+
+**Library** (2 tools — the written procedures an agent follows for a repeatable task):
+
+| Tool          | Purpose                                                                                                       |
+| ------------- | ------------------------------------------------------------------------------------------------------------- |
+| `list_skills` | The procedures the agent follows for a repeatable task: the ones Stratta ships and the ones the office wrote. |
+| `get_skill`   | Read one procedure in full; the office version wins over the Stratta one of the same name.                    |
+
+An office skill with the same name as a shipped one replaces it for every agent
+of the organisation. A report template learned from an office Word file on
+[stratta.ch/bibliotheque](https://stratta.ch/bibliotheque) is served as a skill
+too, and its file is the layout the Word export is rendered into.
 
 **Ingest** (10 tools — add YOUR licensed norms; driven by the bundled `ingest-norm` skill; owner or admin role):
 
@@ -273,18 +303,25 @@ For querying:
 6. Follow `crossRefs` for compound questions (e.g. SIA 261 → SIA 263 → EC).
 7. Call `get_figure` when the section references a figure relevant to the answer.
 
-Two skills ship in the package: **`consult-stratta`** (the working method an agent
-follows to answer from the norms, generated from the same text the server serves
-through `get_methodology`) and **`ingest-norm`**.
+Nine skills ship in the package, under `skills/`: **`consult-stratta`** (the
+working method an agent follows to answer from the norms, generated from the
+same text the server serves through `get_methodology`), **`ingest-norm`**,
+**`project-basis`** and **`use-agreement`** (the two SIA 260 documents, written
+with `draft_deliverable` and `write_section`), **`site-chapter`**,
+**`instruct-structure`**, **`review-open-questions`**, **`handoff-to-word`** and
+**`verification-note`**. The same nine are served to every client by
+`list_skills` and `get_skill`, so a client that installs no skills folder reads
+them anyway; an office skill of the same name replaces the shipped one.
 
 ## Resources and prompts
 
-Four `stratta://` resources can be pinned to a conversation or read without a
+Five `stratta://` resources can be pinned to a conversation or read without a
 tool call, each scoped to your workspace: `stratta://norms` (JSON),
 `stratta://methodology` (Markdown), `stratta://norm/{code}/toc` (the table of
-contents to depth 2, code URL-encoded, e.g. `stratta://norm/SIA%20267/toc`) and
-`stratta://dossier/{id}` (a project dossier as Markdown). The same four are
-served by the remote connector.
+contents to depth 2, code URL-encoded, e.g. `stratta://norm/SIA%20267/toc`),
+`stratta://dossier/{id}` (a project dossier as Markdown) and
+`stratta://skills/{name}` (one skill as Markdown, the office version when there
+is one). The same five are served by the remote connector.
 
 Three prompts appear as slash commands in clients that support them:
 `investigate-question` (question, project?), `resume-dossier` (name) and
@@ -325,17 +362,22 @@ Requires Python ≥ 3.10 with PyMuPDF (`python -m pip install --user pymupdf`).
 Your organization reached one of its limits. The error names the dimension, your
 current count and the plan limit. Retrying will fail identically.
 
-| Limit           | Free  | Pro    | Max    | Team   |
-| --------------- | ----- | ------ | ------ | ------ |
-| Norms           | 1     | 15     | 60     | 75     |
-| Sections        | 1,000 | 15,000 | 60,000 | 75,000 |
-| Figures         | 60    | 750    | 3,000  | 3,750  |
-| Queries / month | 500   | 15,000 | 60,000 | 75,000 |
-| Members         | 1     | 1      | 1      | 5      |
+| Limit           | Free  | Pro    | Max    |
+| --------------- | ----- | ------ | ------ |
+| Norms           | 1     | 15     | 60     |
+| Sections        | 1,000 | 15,000 | 60,000 |
+| Figures         | 60    | 750    | 3,000  |
+| Queries / month | 500   | 10,000 | 40,000 |
+| Members         | 1     | 1      | 1      |
+| Site sheet      | –     | –      | ✓      |
 
 Beyond the included queries, paid plans bill the overage per thousand. An
-Enterprise contract scales seats, norms and queries further; the calculator is
-at https://stratta.ch/tarifs
+Enterprise contract scales seats, norms and queries further and includes the
+site sheet; the calculator is at https://stratta.ch/tarifs. Dossiers,
+deliverables and the library are on every plan; `scan_site` and the site tools
+answer `PLAN_REQUIRED` below Max. A tool call weighs a whole number of queries:
+a light read 1, a section or a sheet 2, a search 5, a borehole PDF read 10, a
+site scan 100.
 
 Stock limits free up when you delete a norm (`ingest_delete`). The monthly query
 counter resets on its own. Gauges live on the Workspace page of your dashboard,

package/dist/errors.d.ts

@@ -1,3 +1,11 @@
+/**
+ * The code and the sentence the usage report carries for a failed call, so
+ * the admin console can say why (the server bounds and scrubs both again).
+ */
+export declare function digestToolError(err: unknown): {
+    code: string;
+    message: string;
+};
 /**
  * Turns a server error into something an agent can act on. A quota rejection
  * is not a bug: the agent should stop retrying and tell the user which limit

package/dist/errors.js

@@ -57,6 +57,38 @@ function withHelp(code, lines) {
     const url = HELP[code];
     return (url ? [...lines, `Détails : ${url}`] : lines).join('\n');
 }
+const CODE_RE = /^[A-Z][A-Z0-9_]{1,63}$/;
+const MESSAGE_MAX = 300;
+/**
+ * The code and the sentence the usage report carries for a failed call, so
+ * the admin console can say why (the server bounds and scrubs both again).
+ */
+export function digestToolError(err) {
+    const clamp = (text) => text.replace(/\s+/g, ' ').trim().slice(0, MESSAGE_MAX);
+    if (err instanceof ConvexError) {
+        const data = err.data;
+        if (typeof data === 'string') {
+            return {
+                code: CODE_RE.test(data) ? data : 'APPLICATION_ERROR',
+                message: clamp(data),
+            };
+        }
+        if (data && typeof data === 'object') {
+            const { code, message } = data;
+            return {
+                code: typeof code === 'string' && CODE_RE.test(code)
+                    ? code
+                    : 'APPLICATION_ERROR',
+                message: clamp(typeof message === 'string' ? message : JSON.stringify(data)),
+            };
+        }
+    }
+    const message = err instanceof Error ? err.message : String(err);
+    return {
+        code: CODE_RE.test(message) ? message : 'UNEXPECTED',
+        message: clamp(message),
+    };
+}
 /**
  * Turns a server error into something an agent can act on. A quota rejection
  * is not a bug: the agent should stop retrying and tell the user which limit

package/dist/index.js

@@ -5,7 +5,7 @@ import { fileURLToPath } from 'node:url';
 import { McpServer } from '@modelcontextprotocol/sdk/server/mcp.js';
 import { StdioServerTransport } from '@modelcontextprotocol/sdk/server/stdio.js';
 import { createConvexClient, api } from './client.js';
-import { formatToolError } from './errors.js';
+import { digestToolError, formatToolError } from './errors.js';
 import { ensureAuthenticated, resolveApiKey } from './auth.js';
 import { elicitApiKey } from './elicit.js';
 import { confirmWithUser } from './confirm.js';
@@ -18,6 +18,8 @@ import { SESSION_ID, usageDigest } from './usage.js';
 import { SERVER_INSTRUCTIONS } from './tools/catalog.gen.js';
 import { registerResources } from './resources.js';
 import { registerPrompts } from './prompts.js';
+import { TOOLSET_CHOICE } from './mode.js';
+import { toolInChoice } from './tools/toolsets.gen.js';
 /**
  * Read from package.json rather than repeated in a literal. The previous
  * version announced `0.3.0` while the published package was `0.6.0`, so every
@@ -62,6 +64,7 @@ async function call(def, args) {
     let resolvedKey = null;
     let authenticated = false;
     let status = 'success';
+    let failure;
     const reportUsage = () => {
         if (!authenticated || !resolvedKey)
             return;
@@ -73,6 +76,8 @@ async function call(def, args) {
             durationMs: Date.now() - startedAt,
             sessionId: SESSION_ID,
             argsDigest: usageDigest(args),
+            errorCode: failure?.code,
+            errorMessage: failure?.message,
         })
             .catch((e) => console.error('[stratta-mcp] usage report failed:', e));
     };
@@ -97,6 +102,7 @@ async function call(def, args) {
     }
     catch (err) {
         status = 'error';
+        failure = digestToolError(err);
         reportUsage();
         return {
             content: [{ type: 'text', text: `Error: ${formatToolError(err)}` }],
@@ -104,13 +110,15 @@ async function call(def, args) {
         };
     }
 }
+// `--simple` or `--toolsets` (ADR 57): an agent that only reads norms is
+// shown ten tools instead of fifty-nine.
 for (const def of [
     ...readTools,
     ...siteTools,
     ...dossierTools,
     ...libraryTools,
     ...ingestTools,
-]) {
+].filter((d) => toolInChoice(d.name, TOOLSET_CHOICE))) {
     server.registerTool(def.name, {
         title: def.title,
         description: def.description,
@@ -150,10 +158,17 @@ COMMANDES
   --version        Version installée
   --help           Cet écran
 
+OPTIONS DU SERVEUR
+  --simple                 Normes seules : dix outils, démarrage plus léger
+  --toolsets dossiers,sites  Normes plus les familles choisies
+                           (dossiers, sites, ingest)
+  STRATTA_TOOLSETS=...     Même réglage par variable d'environnement
+
 Sans commande, le serveur MCP démarre sur stdio : c'est ce que fait votre
 agent, vous n'avez pas à le lancer vous-même.
 
   Installation   claude mcp add stratta --scope user -- npx -y @stratta/mcp
+  Mode simple    claude mcp add stratta --scope user -- npx -y @stratta/mcp --simple
   Raccourci      npm install -g @stratta/mcp   puis   stratta doctor
   Documentation  https://stratta.ch/docs`);
 }

package/dist/mode.d.ts

@@ -0,0 +1,16 @@
+import { type ToolsetChoice } from './tools/toolsets.gen.js';
+/**
+ * Which tool families this server shows (ADR 57).
+ *
+ *     npx -y @stratta/mcp --simple            norms only, ten tools
+ *     npx -y @stratta/mcp --toolsets dossiers norms and dossiers
+ *     STRATTA_TOOLSETS=sites npx -y @stratta/mcp
+ *
+ * The flag wins over the variable. Nothing set: every tool, as before 1.17.
+ * An unknown word is dropped, never fatal: a server that refuses to start on
+ * a typo looks, from the agent, exactly like a server that hangs.
+ */
+export declare function readToolsetChoice(argv: ReadonlyArray<string>, env: Readonly<Record<string, string | undefined>>): ToolsetChoice;
+/** The choice as the server's `get_methodology` reads it; absent for full. */
+export declare function toolsetsParam(choice: ToolsetChoice): string | undefined;
+export declare const TOOLSET_CHOICE: ToolsetChoice;

package/dist/mode.js

@@ -0,0 +1,32 @@
+import { parseToolsets } from './tools/toolsets.gen.js';
+/**
+ * Which tool families this server shows (ADR 57).
+ *
+ *     npx -y @stratta/mcp --simple            norms only, ten tools
+ *     npx -y @stratta/mcp --toolsets dossiers norms and dossiers
+ *     STRATTA_TOOLSETS=sites npx -y @stratta/mcp
+ *
+ * The flag wins over the variable. Nothing set: every tool, as before 1.17.
+ * An unknown word is dropped, never fatal: a server that refuses to start on
+ * a typo looks, from the agent, exactly like a server that hangs.
+ */
+export function readToolsetChoice(argv, env) {
+    if (argv.includes('--simple'))
+        return parseToolsets('simple');
+    if (argv.includes('--full'))
+        return null;
+    for (let i = 0; i < argv.length; i++) {
+        const arg = argv[i];
+        if (arg.startsWith('--toolsets=')) {
+            return parseToolsets(arg.slice('--toolsets='.length));
+        }
+        if (arg === '--toolsets')
+            return parseToolsets(argv[i + 1]);
+    }
+    return parseToolsets(env.STRATTA_TOOLSETS);
+}
+/** The choice as the server's `get_methodology` reads it; absent for full. */
+export function toolsetsParam(choice) {
+    return choice === null ? undefined : choice.join(',');
+}
+export const TOOLSET_CHOICE = readToolsetChoice(process.argv.slice(2), process.env);

package/dist/prompts.js

@@ -1,5 +1,7 @@
 import { z } from 'zod';
 import { PROMPTS } from './tools/catalog.gen.js';
+import { choiceHas } from './tools/toolsets.gen.js';
+import { TOOLSET_CHOICE } from './mode.js';
 /**
  * Prompts: the three ways a user starts a conversation with Stratta, as
  * slash commands in the clients that support them (Claude Code, Claude
@@ -20,6 +22,10 @@ export function renderPrompt(p, args) {
 }
 export function registerPrompts(server) {
     for (const p of PROMPTS) {
+        // A prompt that drives dossier tools is not offered without them.
+        if (p.name === 'resume-dossier' && !choiceHas(TOOLSET_CHOICE, 'dossiers')) {
+            continue;
+        }
         const shape = {};
         for (const a of p.args) {
             const s = z.string().describe(a.description);

package/dist/tools/read.js

@@ -1,3 +1,4 @@
+import { TOOLSET_CHOICE, toolsetsParam } from '../mode.js';
 import { z } from 'zod';
 import { api } from '../client.js';
 import { requireApiKey } from '../auth.js';
@@ -31,7 +32,11 @@ const edition = (tool) => z.string().optional().describe(param(tool, 'edition'))
 export const getMethodology = defineTool({
     ...headOf('get_methodology'),
     inputSchema: {},
-    run: (client) => client.action(api._mcp.getMethodology, { apiKey: requireApiKey() }),
+    run: (client) => client.action(api._mcp.getMethodology, {
+        apiKey: requireApiKey(),
+        // The dossier policy only when the dossier tools are shown (ADR 57).
+        toolsets: toolsetsParam(TOOLSET_CHOICE),
+    }),
     // One Markdown document, read at the start of every conversation.
     toContent: (result) => ({
         content: [{ type: 'text', text: result.text }],

package/dist/tools/toolsets.gen.d.ts

@@ -0,0 +1,49 @@
+/**
+ * Tool families: which MCP tools an agent is shown (ADR 57).
+ *
+ * Fifty-nine tools cost a client that loads every schema up front about
+ * 13 500 tokens before the first question, and a long menu makes a model
+ * pick worse. Most conversations only read norms. So the catalogue is cut
+ * into families, the way GitHub's MCP server cuts its own into toolsets:
+ *
+ * - `norms` is always there: the method, the account, and the norm reader.
+ * - `dossiers` adds the project work: questions, findings, decisions,
+ *   attachments, templates, deliverables and the office's methods.
+ * - `sites` adds the site scan and its readings.
+ * - `ingest` adds the ingestion tools (stdio only: they read a local PDF).
+ *
+ * "Simple" is `norms` alone. "Full" is everything the transport serves, and
+ * stays the default: the published ChatGPT plugin is a snapshot of the full
+ * list, and changing what a connection sees is the user's choice, never ours.
+ *
+ * ⚠️ Read by the browser (the connections page) and by the generator of the
+ * npm package's catalogue: imports nothing, erasable TypeScript only. The
+ * npm package gets a verbatim copy (`toolsets.gen.ts`).
+ * `mcpToolsets.test.ts` checks that every catalogue tool has a family.
+ */
+export type Toolset = 'norms' | 'dossiers' | 'sites' | 'ingest';
+export declare const TOOLSETS: ReadonlyArray<Toolset>;
+/** The families a user can turn on or off; `norms` cannot be turned off. */
+export declare const OPTIONAL_TOOLSETS: ReadonlyArray<Toolset>;
+export declare const TOOLSET_OF: Readonly<Record<string, Toolset>>;
+/** A tool with no family is shown only in full mode, never hidden by mistake. */
+export declare function toolsetOf(name: string): Toolset | null;
+/**
+ * What a connection was set to. `null` means full: every tool the transport
+ * serves, the behaviour of every connection made before families existed.
+ */
+export type ToolsetChoice = ReadonlyArray<Toolset> | null;
+export declare const SIMPLE_TOOLSETS: ReadonlyArray<Toolset>;
+/**
+ * Read a choice written by a person or a flag: `simple`, `full`, `all`, or a
+ * comma list such as `dossiers,sites`. Unknown words are dropped rather than
+ * refused, so a typo degrades to fewer tools, never to a server that does not
+ * start. `norms` is always added back.
+ */
+export declare function parseToolsets(raw: string | null | undefined): ToolsetChoice;
+/** Whether a tool is shown under a choice. */
+export declare function toolInChoice(name: string, choice: ToolsetChoice): boolean;
+/** Whether the choice carries a family (a full choice carries all of them). */
+export declare function choiceHas(choice: ToolsetChoice, family: Toolset): boolean;
+/** The stored form: a sorted list, or `undefined` for full. */
+export declare function normalizeChoice(choice: ToolsetChoice): Toolset[] | undefined;

package/dist/tools/toolsets.gen.js

@@ -0,0 +1,130 @@
+// Generated by scripts/build-mcp-catalog.mjs from convex/mcpToolsets.ts.
+// Do not edit: change the source and run the script.
+export const TOOLSETS = [
+    'norms',
+    'dossiers',
+    'sites',
+    'ingest',
+];
+/** The families a user can turn on or off; `norms` cannot be turned off. */
+export const OPTIONAL_TOOLSETS = [
+    'dossiers',
+    'sites',
+    'ingest',
+];
+export const TOOLSET_OF = {
+    get_methodology: 'norms',
+    whoami: 'norms',
+    list_norms: 'norms',
+    get_toc: 'norms',
+    get_subtree: 'norms',
+    get_section: 'norms',
+    search_in_norm: 'norms',
+    search_corpus: 'norms',
+    get_cross_refs: 'norms',
+    get_figure: 'norms',
+    list_dossiers: 'dossiers',
+    open_dossier: 'dossiers',
+    load_dossier: 'dossiers',
+    open_question: 'dossiers',
+    save_finding: 'dossiers',
+    record_decision: 'dossiers',
+    resolve_question: 'dossiers',
+    list_attachments: 'dossiers',
+    read_attachment: 'dossiers',
+    search_in_dossier: 'dossiers',
+    add_attachment: 'dossiers',
+    list_templates: 'dossiers',
+    apply_template: 'dossiers',
+    draft_deliverable: 'dossiers',
+    write_section: 'dossiers',
+    list_skills: 'dossiers',
+    get_skill: 'dossiers',
+    list_questions: 'dossiers',
+    get_question: 'dossiers',
+    get_dossier_activity: 'dossiers',
+    list_exports: 'dossiers',
+    update_question: 'dossiers',
+    add_option: 'dossiers',
+    update_option: 'dossiers',
+    delete_option: 'dossiers',
+    delete_question: 'dossiers',
+    add_comment: 'dossiers',
+    update_entry: 'dossiers',
+    attach_entry: 'dossiers',
+    scan_site: 'sites',
+    get_boreholes: 'sites',
+    get_site_context: 'sites',
+    list_sites: 'sites',
+    get_site: 'sites',
+    get_site_section: 'sites',
+    list_site_overlays: 'sites',
+    get_groundwater_contours: 'sites',
+    list_site_photos: 'sites',
+    list_site_neighbours: 'sites',
+    ingest_status: 'ingest',
+    ingest_create_document: 'ingest',
+    ingest_create_sections: 'ingest',
+    ingest_attach_formula: 'ingest',
+    ingest_attach_table: 'ingest',
+    ingest_attach_cross_ref: 'ingest',
+    ingest_upload_figure: 'ingest',
+    ingest_normalize_cross_refs: 'ingest',
+    ingest_publish: 'ingest',
+    ingest_delete: 'ingest',
+};
+/** A tool with no family is shown only in full mode, never hidden by mistake. */
+export function toolsetOf(name) {
+    return TOOLSET_OF[name] ?? null;
+}
+export const SIMPLE_TOOLSETS = ['norms'];
+/**
+ * Read a choice written by a person or a flag: `simple`, `full`, `all`, or a
+ * comma list such as `dossiers,sites`. Unknown words are dropped rather than
+ * refused, so a typo degrades to fewer tools, never to a server that does not
+ * start. `norms` is always added back.
+ */
+export function parseToolsets(raw) {
+    const text = (raw ?? '').trim().toLowerCase();
+    if (!text || text === 'full' || text === 'all')
+        return null;
+    if (text === 'simple')
+        return SIMPLE_TOOLSETS;
+    const picked = new Set(['norms']);
+    for (const word of text.split(/[\s,]+/)) {
+        const alias = {
+            norm: 'norms',
+            normes: 'norms',
+            dossier: 'dossiers',
+            deliverables: 'dossiers',
+            livrables: 'dossiers',
+            site: 'sites',
+            ingestion: 'ingest',
+        };
+        const name = (alias[word] ?? word);
+        if (TOOLSETS.includes(name))
+            picked.add(name);
+    }
+    return TOOLSETS.filter((t) => picked.has(t));
+}
+/** Whether a tool is shown under a choice. */
+export function toolInChoice(name, choice) {
+    if (choice === null)
+        return true;
+    const family = toolsetOf(name);
+    return family !== null && choice.includes(family);
+}
+/** Whether the choice carries a family (a full choice carries all of them). */
+export function choiceHas(choice, family) {
+    return choice === null || choice.includes(family);
+}
+/** The stored form: a sorted list, or `undefined` for full. */
+export function normalizeChoice(choice) {
+    if (choice === null)
+        return undefined;
+    const set = new Set(['norms', ...choice]);
+    // Every family turned on is the same as full, and full is how it is stored.
+    if (TOOLSETS.every((t) => set.has(t)))
+        return undefined;
+    return TOOLSETS.filter((t) => set.has(t));
+}

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "@stratta/mcp",
   "mcpName": "ch.stratta/mcp",
-  "version": "1.16.0",
+  "version": "1.17.0",
   "description": "MCP server exposing the engineering norms your firm is licensed for (SIA / Eurocodes) to any MCP client, via Stratta TreeRAG.",
   "license": "UNLICENSED",
   "author": "SmartFlow <hello@stratta.ch>",

package/scripts/ingest-prepass.py

@@ -77,6 +77,46 @@ NUMBER_TOKEN_RE = re.compile(r"^(?:\d{1,2}(?:\.\d{1,3}){0,4}\.?|[A-Z](?:\.\d{1,2
 OCR_DOTTED_NUMBER_RE = re.compile(r"^(\d+(?:\.\d+)*) \.(\d)")
 
 
+# Above this share of a page's words, vertical text is content, not a watermark.
+VERTICAL_WATERMARK_SHARE = 0.25
+
+
+def _vertical_lines(page: "fitz.Page") -> list[tuple[tuple[float, float, float, float], set[str]]]:
+    """Box and words of each line not written left to right.
+
+    Matched to the words by position and text, never by (block, line)
+    numbers: `get_text("dict")` counts image blocks and `get_text("words")`
+    does not, so the numbers of one extraction can point at a clause line of
+    the other.
+    """
+    out: list[tuple[tuple[float, float, float, float], set[str]]] = []
+    try:
+        blocks = page.get_text("dict").get("blocks", [])
+    except Exception:  # a page the dict extraction cannot read keeps its words
+        return out
+    for block in blocks:
+        for line in block.get("lines", []):
+            dx, dy = line.get("dir", (1.0, 0.0))
+            if abs(dy) <= 0.1 and dx >= 0:
+                continue
+            text = " ".join(span.get("text", "") for span in line.get("spans", []))
+            words = {w for w in text.split() if w}
+            if words:
+                out.append((tuple(line["bbox"]), words))
+    return out
+
+
+def _in_vertical_line(
+    word: tuple, vertical: list[tuple[tuple[float, float, float, float], set[str]]]
+) -> bool:
+    cx = (word[0] + word[2]) / 2
+    cy = (word[1] + word[3]) / 2
+    for (x0, y0, x1, y1), texts in vertical:
+        if x0 - 1 <= cx <= x1 + 1 and y0 - 1 <= cy <= y1 + 1 and word[4] in texts:
+            return True
+    return False
+
+
 def _rows(page: "fitz.Page") -> list[list[tuple]]:
     words = page.get_text("words")  # x0, y0, x1, y1, text, block, line, word
     if not words:
@@ -86,7 +126,19 @@ def _rows(page: "fitz.Page") -> list[list[tuple]]:
     # A word set vertically (a library watermark such as "Ecole Polytechnique
     # Fédérale de Lausanne" running up the margin) has a box far taller than
     # the text. Left in, its words land on every line they cross.
+    # ⚠️ The height alone lets its short words through ("Ecole", "EPFL,",
+    # "SNV", "de", "/"): 310 sections of 16 norms carried them (2026-09-30).
+    # A vertical line is also told by its writing direction.
     words = [w for w in words if (w[3] - w[1]) <= 2.5 * median]
+    # ⚠️ A watermark is a thin minority of a page. When vertical lines carry
+    # a large share of the words, the page itself is rotated (a landscape
+    # table in SIA 197-1, p. 45) and dropping them would empty it.
+    vertical = _vertical_lines(page)
+    if vertical:
+        flagged = [w for w in words if _in_vertical_line(w, vertical)]
+        if len(flagged) <= VERTICAL_WATERMARK_SHARE * len(words):
+            dropped = set(map(id, flagged))
+            words = [w for w in words if id(w) not in dropped]
     if not words:
         return []
     tol = max(2.0, 0.45 * median)

package/skills/consult-stratta/SKILL.md

@@ -68,4 +68,4 @@ Before drafting a document (project basis, use agreement, terrain chapter, verif
 - Do not expose internal identifiers unless the next Stratta tool call requires them.
 - Ingesting a new norm starts from a PDF on the user's own machine: it goes through the local `@stratta/mcp` server and its `ingest-norm` skill, never through the remote connector.
 
-methodology: v3
+methodology: v4