@stratta/mcp 0.9.1 → 0.9.2

package/README.md

@@ -4,7 +4,7 @@ package.json exactly, and it must be present in the *published* tarball —
 the registry reads the README from npm, not from this repository. Without it
 `mcp-publisher publish` rejects the server as unverified.
 
-mcp-name: io.github.hugogebel-boop/stratta
+mcp-name: ch.stratta/mcp
 -->
 
 # @stratta/mcp

package/dist/login.js

@@ -52,15 +52,15 @@ function openBrowser(url) {
     }
 }
 function page(title, body) {
-    return `<!doctype html><html lang="fr"><meta charset="utf-8">
-<title>${title}</title>
-<style>
-  body{font:16px/1.6 ui-sans-serif,system-ui,-apple-system,"Segoe UI",sans-serif;
-       background:#12110f;color:#f5f1e8;display:grid;place-items:center;
-       min-height:100vh;margin:0;padding:2rem;text-align:center}
-  .c{max-width:26rem} h1{font-size:1.25rem;margin:0 0 .5rem}
-  p{margin:0;color:#a8a196}
-</style>
+    return `<!doctype html><html lang="fr"><meta charset="utf-8">
+<title>${title}</title>
+<style>
+  body{font:16px/1.6 ui-sans-serif,system-ui,-apple-system,"Segoe UI",sans-serif;
+       background:#12110f;color:#f5f1e8;display:grid;place-items:center;
+       min-height:100vh;margin:0;padding:2rem;text-align:center}
+  .c{max-width:26rem} h1{font-size:1.25rem;margin:0 0 .5rem}
+  p{margin:0;color:#a8a196}
+</style>
 <div class="c"><h1>${title}</h1><p>${body}</p></div>`;
 }
 /**

package/dist/tools/dossier.js

@@ -59,16 +59,16 @@ export const openDossier = defineTool({
 export const saveFinding = defineTool({
     name: 'save_finding',
     title: 'Save a finding to a dossier',
-    description: `Record ONE decision in a dossier. Call it as you work, not in a batch at the end — a finding saved when it is made carries the reasoning that produced it.
-
-WHAT TO RECORD, by kind:
-- reference: what a norm says, that the project relies on. ALWAYS fill normCode + sectionPath (+ page). If you cannot cite it, it is not a reference.
-- hypothesis: a value the engineer RETAINS, and why. This is the one that matters most in geotechnics, where a retained value is a judgement between disagreeing measurements rather than the output of a formula. Put the number in \`value\` and the reasoning in \`detail\`.
-- observation: what the site, a borehole, or a survey showed. Include the date in \`detail\` when known.
-- question: something not settled. These surface FIRST when the dossier is reloaded, so record them even when you cannot answer — especially then.
-
-WHAT NOT TO RECORD: your own prose, intermediate steps, anything the user did not treat as a decision. A dossier of forty entries where three mattered is worse than a dossier of three.
-
+    description: `Record ONE decision in a dossier. Call it as you work, not in a batch at the end — a finding saved when it is made carries the reasoning that produced it.
+
+WHAT TO RECORD, by kind:
+- reference: what a norm says, that the project relies on. ALWAYS fill normCode + sectionPath (+ page). If you cannot cite it, it is not a reference.
+- hypothesis: a value the engineer RETAINS, and why. This is the one that matters most in geotechnics, where a retained value is a judgement between disagreeing measurements rather than the output of a formula. Put the number in \`value\` and the reasoning in \`detail\`.
+- observation: what the site, a borehole, or a survey showed. Include the date in \`detail\` when known.
+- question: something not settled. These surface FIRST when the dossier is reloaded, so record them even when you cannot answer — especially then.
+
+WHAT NOT TO RECORD: your own prose, intermediate steps, anything the user did not treat as a decision. A dossier of forty entries where three mattered is worse than a dossier of three.
+
 Set \`confidence\` whenever the entry is a value: established (computed or read directly), judgement (the engineer chose it), to_confirm (provisional, needs a test or a check). Confusing those three is the professional fault this field exists to prevent.`,
     inputSchema: {
         dossierId,

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "@stratta/mcp",
-  "mcpName": "io.github.hugogebel-boop/stratta",
-  "version": "0.9.1",
+  "mcpName": "ch.stratta/mcp",
+  "version": "0.9.2",
   "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/publish-release.ps1

@@ -0,0 +1,99 @@
+<#
+.SYNOPSIS
+  Publish @stratta/mcp to npm, then register it with the official MCP registry.
+
+.DESCRIPTION
+  The two publications are ordered, not independent: the registry reads
+  `mcpName` from the tarball npm serves, so a version that is not on npm yet
+  cannot be registered. Doing it by hand in the wrong order fails with
+
+      NPM package '@stratta/mcp' exists, but version 'x.y.z' was not found
+
+  which reads like a transient error and is not one.
+
+  Run this from an interactive terminal. `npm publish` triggers a 2FA prompt
+  that opens a browser, and `npm login` needs one too if the session has
+  lapsed. Neither works from a non-interactive shell -- by design, so that
+  access to a machine is not access to the package.
+
+  ASCII only, on purpose: Windows PowerShell 5.1 reads a UTF-8 file without a
+  BOM as ANSI, and a single arrow or em dash in a comment breaks the parser
+  with an error that points at the wrong line.
+
+.PARAMETER PrivateKey
+  The Ed25519 key proving control of stratta.ch, 64 hex characters. Lives in
+  Doppler under MCP_REGISTRY_PRIVATE_KEY. Its public half is served from
+  /.well-known/mcp-registry-auth -- deleting that file breaks publishing.
+
+.PARAMETER PublisherPath
+  Path to mcp-publisher. Download it from the registry releases if absent:
+  https://github.com/modelcontextprotocol/registry/releases
+
+.EXAMPLE
+  .\publish-release.ps1 -PrivateKey (doppler secrets get MCP_REGISTRY_PRIVATE_KEY --plain)
+#>
+param(
+  [Parameter(Mandatory = $true)][string]$PrivateKey,
+  [string]$PublisherPath = "mcp-publisher"
+)
+
+$ErrorActionPreference = "Stop"
+Push-Location (Join-Path $PSScriptRoot "..")
+
+try {
+  $pkg = Get-Content package.json -Raw | ConvertFrom-Json
+  $srv = Get-Content server.json -Raw | ConvertFrom-Json
+
+  Write-Host "-> @stratta/mcp $($pkg.version) as $($pkg.mcpName)" -ForegroundColor Cyan
+
+  # Fail before publishing, not after. The three names and the three versions
+  # have to agree, or the registry rejects the server as unverified once npm
+  # has already gone out, and a version number is burnt.
+  if ($pkg.mcpName -ne $srv.name) {
+    throw "package.json mcpName ($($pkg.mcpName)) does not match server.json name ($($srv.name))"
+  }
+  if ($pkg.version -ne $srv.version -or $pkg.version -ne $srv.packages[0].version) {
+    throw "version mismatch: package.json $($pkg.version), server.json $($srv.version), packages[0] $($srv.packages[0].version)"
+  }
+  if (-not (Select-String -Path README.md -Pattern ("mcp-name: " + [regex]::Escape($pkg.mcpName)) -Quiet)) {
+    throw "README.md is missing the marker 'mcp-name: $($pkg.mcpName)'"
+  }
+  Write-Host "   manifest is coherent" -ForegroundColor DarkGray
+
+  Write-Host "-> npm whoami" -ForegroundColor Cyan
+  npm whoami
+  if ($LASTEXITCODE -ne 0) { throw "Not logged in to npm. Run 'npm login' and retry." }
+
+  Write-Host "-> npm publish (a 2FA prompt may open a browser)" -ForegroundColor Cyan
+  npm publish
+  if ($LASTEXITCODE -ne 0) { throw "npm publish failed" }
+
+  # npm's CDN needs a moment before the registry can see the new version.
+  Write-Host "-> waiting for npm to serve $($pkg.version)" -ForegroundColor Cyan
+  $seen = $false
+  foreach ($i in 1..20) {
+    Start-Sleep -Seconds 3
+    try {
+      $served = Invoke-RestMethod "https://registry.npmjs.org/@stratta/mcp/$($pkg.version)"
+      if ($served.version -eq $pkg.version) { $seen = $true; break }
+    } catch { }
+  }
+  if (-not $seen) { throw "npm has not served $($pkg.version) after 60s; rerun the registry steps by hand" }
+  Write-Host "   npm serves $($pkg.version)" -ForegroundColor DarkGray
+
+  # Domain proof rather than the GitHub device flow: nothing waits on a human
+  # typing a code. The JWT it returns expires in under an hour, which is why
+  # login and publish run back to back with nothing in between.
+  Write-Host "-> mcp-publisher login (domain proof)" -ForegroundColor Cyan
+  & $PublisherPath login http --domain stratta.ch --private-key $PrivateKey
+  if ($LASTEXITCODE -ne 0) { throw "registry login failed. Check that https://stratta.ch/.well-known/mcp-registry-auth is still served." }
+
+  Write-Host "-> mcp-publisher publish" -ForegroundColor Cyan
+  & $PublisherPath publish
+  if ($LASTEXITCODE -ne 0) { throw "registry publish failed" }
+
+  Write-Host "OK  $($pkg.version) published to npm and to registry.modelcontextprotocol.io" -ForegroundColor Green
+}
+finally {
+  Pop-Location
+}

package/skills/ingest-norm/SKILL.md

@@ -16,6 +16,7 @@ writes go through the Stratta MCP `ingest_*` tools, scoped to your org.
 > for. You are responsible for your usage rights (see Stratta's Terms).
 
 ## Prerequisites
+
 - The Stratta MCP server is installed and your `STRATTA_API_KEY` resolves
   (via env, `~/.stratta/config.json`, or first-call elicitation).
 - **Python ≥ 3.10 with PyMuPDF**. Install once:
@@ -23,12 +24,14 @@ writes go through the Stratta MCP `ingest_*` tools, scoped to your org.
 - The norm PDF is available locally.
 
 ## Tools used (all scoped to your workspace)
+
 `ingest_status` · `ingest_create_document` · `ingest_create_sections` ·
 `ingest_attach_formula` · `ingest_attach_table` · `ingest_attach_cross_ref` ·
 `ingest_upload_figure` · `ingest_normalize_cross_refs` · `ingest_publish` ·
 `ingest_delete`.
 
 ## Quotas — read before you start
+
 The workspace has server-enforced limits on norms, sections and figures. A
 `QUOTA_EXCEEDED` error names the dimension, the current count and the limit.
 
@@ -43,29 +46,37 @@ half-ingested document behind, which the user then has to delete by hand.
 ## Workflow
 
 ### 1. Locate the pre-pass script
+
 It ships inside this package at `scripts/ingest-prepass.py`. Resolve its path:
+
 ```bash
 node -e "console.log(require.resolve('@stratta/mcp/package.json'))"
 # → <root>/package.json  →  <root>/scripts/ingest-prepass.py
 ```
+
 If the user is working in the Stratta monorepo, the script also lives at
 `packages/mcp/scripts/ingest-prepass.py`.
 
 ### 2. Check for an existing copy
+
 `ingest_status { code }` (e.g. `"SIA 261"`). To re-ingest, call
 `ingest_delete { documentId }` first.
 
 ### 3. Run the pre-pass
+
 ```bash
 python <pkg-root>/scripts/ingest-prepass.py \
   --pdf <path-to-pdf> \
   --output .stratta-ingest/<code-slug>
 ```
+
 Output under `.stratta-ingest/<code-slug>/`:
+
 - `prepass.json` — full manifest (see below).
 - `figures/page-NNN.png` — one PNG per page that contains a `Figure N` caption.
 
 `prepass.json` structure:
+
 - `doc` — `pageCount`, detected `language`, `tocSource`, raw `metadata`.
 - `stats` — `sectionCount`, `byDepth`, `figureCount`.
 - `sections[]` — full hierarchical tree (depth **0 = chapter**, 1+ = sub-sections),
@@ -79,12 +90,15 @@ The pre-pass is **exhaustive** (e.g. ~550 nodes on SIA 261). You decide what
 to keep in the next step.
 
 ### 4. Create the document
+
 Read `prepass.json`, then:
 `ingest_create_document { code, year, title, language: <doc.language>, totalPages: <doc.pageCount> }`
 → returns `documentId`. Keep it for every subsequent call.
 
 ### 5. Decide section granularity + generate summaries
+
 Iterate `sections[]` and decide what to keep. Two viable strategies:
+
 - **Keep all** — most faithful, ~500 sections on a typical SIA norm. Great
   for fine-grained navigation but verbose.
 - **Aggregate trivial leaves** — fold paragraph-level nodes (`6.1.1`...`6.1.11`)
@@ -92,6 +106,7 @@ Iterate `sections[]` and decide what to keep. Two viable strategies:
   100-150 sections. Recommended unless the user asks for max granularity.
 
 For each kept section, prepare:
+
 - `summary` — 1-3 sentences derived from `rawText` (mention formulas/values).
 - `content` — enriched text with LaTeX inline where the source has math
   (e.g. `$\sigma_d = f_{yd} \cdot \gamma$`). Open the PDF visually for pages
@@ -102,18 +117,22 @@ Keep the `nodeId` / `parentNodeId` / `path` / `pageStart` / `pageEnd` /
 `orderIndex` / `depth` from the pre-pass — those are deterministic.
 
 ### 6. Insert sections (batched)
+
 `ingest_create_sections { documentId, sections: [...] }` in batches of 30-50.
 Parent links resolve via `parentNodeId` within the batch and across prior
 batches. The call returns a `nodeId → sectionId` map — **use those `sectionId`s**
 for every enrichment call below.
 
 ### 7. Enrich
+
 - `ingest_attach_formula { sectionId, latex, description, formulaNumber }`
 - `ingest_attach_table { sectionId, data: { headers, rows }, caption, tableNumber }`
 - `ingest_attach_cross_ref { sourceSectionId, targetDocumentCode, targetSectionPath?, refText, refType }`
 
 ### 8. Upload figures
+
 For each figure in `prepass.json#figures`:
+
 - Read `.stratta-ingest/<code>/<fileName>` and base64-encode the bytes.
 - Find the owning section: the kept section whose `pageStart..pageEnd`
   range includes the figure's `page`.
@@ -121,15 +140,18 @@ For each figure in `prepass.json#figures`:
   (≤ 8 MB per image).
 
 ### 9. Auto cross-references (optional but recommended)
+
 `ingest_normalize_cross_refs { documentId }` scans every section's text for
 references to other norms (SIA / SN EN / EN / ISO / DIN …) and rebuilds the
 cross-ref index. Idempotent.
 
 ### 10. Publish
+
 `ingest_publish { documentId }`. The norm is now queryable in your workspace
 via `list_norms`, `get_toc`, `get_section`, `search_in_norm`, `get_figure`, etc.
 
 ## Quality bar
+
 - Trust the pre-pass for `path` + `pageStart`/`pageEnd` — it's deterministic
   and verified against the PDF bookmarks.
 - Summaries drive navigation — be specific (mention key formulas/values).