ask-marcel-office-cli 2.1.0 → 2.2.0

package/docs/COMMANDS.md

@@ -1,6 +1,6 @@
 # Command reference
 
-All 180 commands across 11 categories, grouped by category. Each row shows the command name, a one-line summary, the required parameters, and the underlying Microsoft Graph endpoint. The five lifecycle commands (`login`, `logout`, `update`, `docs`, `help-json`) are listed separately in the **Authentication & lifecycle** section below — `help-json` counts them in its manifest total (184), so a 179-vs-184 gap is those five, not a drift.
+All 184 commands across 11 categories, grouped by category. Each row shows the command name, a one-line summary, the required parameters, and the underlying Microsoft Graph endpoint. The six lifecycle commands (`login`, `logout`, `update`, `docs`, `help-json`, `mcp`) are listed separately in the **Authentication & lifecycle** section below — `help-json` counts them in its manifest total (190), so a 184-vs-190 gap is those six, not a drift.
 
 The auto-generated tables below are rebuilt from the live command registry on every `bun run docs:gen` / `bun run build` — they cannot drift from the actual surface.
 
@@ -20,6 +20,7 @@ For everything else:
 | `update` | Update ask-marcel-office to the latest version on npm (auto-detects npm vs bun) |
 | `docs <cmd>` | Print Markdown docs for a single command |
 | `help-json [--terse] [--category <name>]` | Machine-readable manifest, full or filtered |
+| `mcp` | Serve the command registry to an MCP client over stdio (five gateway tools) |
 | `next-page --url <link>` | Walk a paginated response by feeding the previous `nextLink` |
 
 <!-- AUTO-GENERATED-COMMANDS:BEGIN -->
@@ -28,15 +29,15 @@ For everything else:
 
 | Command | Description | Required params | Graph endpoint |
 |---------|-------------|-----------------|----------------|
-| `convert-drive-item-zip` | Unzip a `.zip` from a OneDrive / SharePoint item and convert every contained file in one call — so "read the handover archive" doesn't need a separate unzip + per-file conversion. Office files (docx/xlsx/pptx/odt/ods/odp and their macro-enabled / template variants) are converted to markdown via the local pipelines; plain-text entries (txt/md/csv/json/yaml/…) are decoded inline; legacy OLE .xls (sheetjs) and .doc (word-extractor, text only) are extracted; an Outlook .msg entry is rendered to markdown (headers + body, with its own attachments converted recursively); PDFs have their text layer extracted (text/plain); images, binaries, nested archives, legacy .ppt, and scanned/image-only PDFs (no text layer) are listed with a note (not unpacked) so one unsupported entry never fails the whole archive. Pass `--include-metadata true` to append each Office file's side-channel metadata block. Capped at 100 entries (the archive is buffered in memory); beyond that the response is flagged `truncated`. | `--drive-id`, `--item-id`, `--include-metadata` | `GET /drives/{drive-id}/items/{item-id}/content` |
-| `download-drive-item-as-markdown` | Download a OneDrive / SharePoint file converted to markdown via local conversion pipelines. Supported: docx (mammoth → turndown; embedded images are replaced with `[image]` placeholders by default — pass `--inline-images true` to embed them as base64 `data:` URIs, or pull the full-resolution originals with `extract-drive-item-images`; tables become GFM pipe tables), xlsx (one markdown table per sheet via sheetjs; any sheet whose used range exceeds the `--max-cells` cap, default 50 000, is summarised with a band-by-band read hint instead of a multi-hundred-MB table), csv (rendered as a markdown table), odt/ods/odp (OpenDocument body walked from content.xml — headings, lists, tables, named sheets, per-slide text, including style-hidden content), plus plain-text passthrough (txt/md/html/json/yaml/log/xml/etc.) — the bytes are followed through any CDN redirect and returned inline as `{ contentType: "text/plain", size, text }` so the LLM never needs a separate fetch step. Loop/Fluid/Whiteboard files use Graph `?format=html` (the four inputs Microsoft documents — https://learn.microsoft.com/en-us/graph/api/driveitem-get-content-format). To locate `.loop` pages to convert, run `microsoft-search-query` with `filetype:loop` and pass the `resource.id` and `parentReference.driveId` from each hit here. pptx is flattened to per-slide text (titles, bullets, text-box / table text, with speaker notes inline) as `## Slide N` sections — document order, not guaranteed visual order; for layout / images / charts, `download-drive-item-as-pdf` + a vision-capable LLM reads the rendered deck better. pdf has its text layer extracted locally (via unpdf) and returned as `{ contentType: "text/plain", size, text }`; a scanned / image-only PDF with no text layer points you at `download-drive-item-as-pdf` + a vision model. Legacy OLE Office: .xls (Excel 97-2003) is read by sheetjs (same markdown-table path as .xlsx); .doc (Word 97-2003) is extracted by word-extractor as plain text (body only, no structure); .ppt (PowerPoint 97-2003) has no pure-JS path — use `download-drive-item-as-pdf` first, then a vision model. Outlook .msg (a saved email, also an OLE container) is rendered to markdown — an H1 subject, a From/To/Cc/Date header block, the message body, then an `## Attachments` section where each attachment is converted recursively through this same pipeline (the same way a .zip is unpacked); inline / image attachments are listed with a note. For rtf/etc. use `download-drive-item-as-pdf` — Graph `?format=pdf` accepts 38 input extensions. | `--drive-id`, `--item-id`, `--include-metadata`, `--inline-images`, `--max-cells` | `GET /drives/{drive-id}/items/{item-id}/content?format=html` |
-| `download-drive-item-as-pdf` | Download a OneDrive / SharePoint file converted to PDF on the fly by Graph (`?format=pdf`). Source must be one of the Office formats Graph supports — doc, docx, ppt, pptx, xls, xlsx, rtf, csv, odp, ods, odt, etc. The command pre-fetches the filename and short-circuits to a raw download in two cases: plain-text source extensions (txt, md, html, json, …) where conversion is meaningless, and `pdf` sources where the source IS already a PDF (Graph’s `?format=pdf` does not list `pdf` in its supported input set — the CDN responds 406 InputFormatNotSupported on `pdf → pdf`). Worst-case wall-clock is two back-to-back Graph round-trips; the `?format=pdf` transform can run up to the 5-minute request timeout on large or complex sources. | `--drive-id`, `--item-id` | `GET /drives/{drive-id}/items/{item-id}/content?format=pdf` |
-| `download-drive-item-content` | Download the binary content of a file stored in OneDrive / SharePoint, with the bytes inlined. The CLI follows the Graph 302 → SharePoint media-transform redirect internally so the LLM never has to fetch an external URL. The bytes are CONTENT-SNIFFED, not judged by extension: if they decode as valid UTF-8 they come back as `{contentType: "text/plain", size, text}` (avoids ~33% base64 bloat, works for any text file regardless of name); otherwise as `{contentType, size, base64}`. A binary file that happens to be named `.txt` is returned faithfully as base64 — never silently corrupted into `�` by a forced text decode. | `--drive-id`, `--item-id` | `GET /drives/{drive-id}/items/{item-id}/content` |
+| `convert-drive-item-zip-to-markdown` | Unzip a `.zip` from a OneDrive / SharePoint item and convert every contained file in one call — so "read the handover archive" doesn't need a separate unzip + per-file conversion. Office files (docx/xlsx/pptx/odt/ods/odp and their macro-enabled / template variants) are converted to markdown via the local pipelines; plain-text entries (txt/md/csv/json/yaml/…) are decoded inline; legacy OLE .xls (sheetjs) and .doc (word-extractor, text only) are extracted; an Outlook .msg entry is rendered to markdown (headers + body, with its own attachments converted recursively); PDFs have their text layer extracted (text/plain); images, binaries, nested archives, legacy .ppt, and scanned/image-only PDFs (no text layer) are listed with a note (not unpacked) so one unsupported entry never fails the whole archive. Pass `--include-metadata true` to append each Office file's side-channel metadata block. Capped at 100 entries (the archive is buffered in memory); beyond that the response is flagged `truncated`. | `--drive-id`, `--item-id`, `--tenant-id`, `--include-metadata`, `--keep-quoted` | `GET /drives/{drive-id}/items/{item-id}/content` |
+| `download-drive-item-as-markdown` | Download a OneDrive / SharePoint file converted to markdown via local conversion pipelines. Supported: docx (mammoth → turndown; embedded images are replaced with `[image]` placeholders by default — pass `--inline-images true` to embed them as base64 `data:` URIs, or pull the full-resolution originals with `extract-drive-item-images`; tables become GFM pipe tables), xlsx (one markdown table per sheet via sheetjs; any sheet whose used range exceeds the `--max-cells` cap, default 50 000, is summarised with a band-by-band read hint instead of a multi-hundred-MB table), csv (rendered as a markdown table), odt/ods/odp (OpenDocument body walked from content.xml — headings, lists, tables, named sheets, per-slide text, including style-hidden content), plus plain-text passthrough (txt/md/html/json/yaml/log/xml/etc.) — the bytes are followed through any CDN redirect and returned inline as `{ contentType: "text/plain", size, text }` so the LLM never needs a separate fetch step. Loop/Fluid/Whiteboard files use Graph `?format=html` (the four inputs Microsoft documents — https://learn.microsoft.com/en-us/graph/api/driveitem-get-content-format). To locate `.loop` pages to convert, run `microsoft-search-query` with `filetype:loop` and pass the `resource.id` and `parentReference.driveId` from each hit here. pptx is flattened to per-slide text (titles, bullets, text-box / table text, with speaker notes inline) as `## Slide N` sections — document order, not guaranteed visual order; for layout / images / charts, `download-drive-item-as-pdf` + a vision-capable LLM reads the rendered deck better. pdf has its text layer extracted locally (via unpdf) and returned as `{ contentType: "text/plain", size, text }`; a scanned / image-only PDF with no text layer points you at `download-drive-item-as-pdf` + a vision model. Legacy OLE Office: .xls (Excel 97-2003) is read by sheetjs (same markdown-table path as .xlsx); .doc (Word 97-2003) is extracted by word-extractor as plain text (body only, no structure); .ppt (PowerPoint 97-2003) has no pure-JS path — use `download-drive-item-as-pdf` first, then a vision model. Outlook .msg (a saved email, also an OLE container) is rendered to markdown — an H1 subject, a From/To/Cc/Date header block, the message body, then an `## Attachments` section where each attachment is converted recursively through this same pipeline (the same way a .zip is unpacked); inline / image attachments are listed with a note. For rtf/etc. use `download-drive-item-as-pdf` — Graph `?format=pdf` accepts 38 input extensions. | `--drive-id`, `--item-id`, `--tenant-id`, `--include-metadata`, `--keep-quoted`, `--inline-images`, `--max-cells` | `GET /drives/{drive-id}/items/{item-id}/content?format=html` |
+| `download-drive-item-as-pdf` | Download a OneDrive / SharePoint file converted to PDF on the fly by Graph (`?format=pdf`). Source must be one of the Office formats Graph supports — doc, docx, ppt, pptx, xls, xlsx, rtf, csv, odp, ods, odt, etc. The command pre-fetches the filename and short-circuits to a raw download in two cases: plain-text source extensions (txt, md, html, json, …) where conversion is meaningless, and `pdf` sources where the source IS already a PDF (Graph’s `?format=pdf` does not list `pdf` in its supported input set — the CDN responds 406 InputFormatNotSupported on `pdf → pdf`). Worst-case wall-clock is two back-to-back Graph round-trips; the `?format=pdf` transform can run up to the 5-minute request timeout on large or complex sources. | `--drive-id`, `--item-id`, `--tenant-id` | `GET /drives/{drive-id}/items/{item-id}/content?format=pdf` |
+| `download-drive-item-content` | Download the binary content of a file stored in OneDrive / SharePoint, with the bytes inlined. The CLI follows the Graph 302 → SharePoint media-transform redirect internally so the LLM never has to fetch an external URL. The bytes are CONTENT-SNIFFED, not judged by extension: if they decode as valid UTF-8 they come back as `{contentType: "text/plain", size, text}` (avoids ~33% base64 bloat, works for any text file regardless of name); otherwise as `{contentType, size, base64}`. A binary file that happens to be named `.txt` is returned faithfully as base64 — never silently corrupted into `�` by a forced text decode. | `--drive-id`, `--item-id`, `--tenant-id` | `GET /drives/{drive-id}/items/{item-id}/content` |
 | `download-drive-item-version` | Download a *non-current* historical version of a OneDrive / SharePoint file. `--format original` (default) returns the raw bytes — Graph refuses to serve the current version through this endpoint with "You cannot get the content of the current version"; for the current version use `download-drive-item-content`. `--format pdf` runs Graph `?format=pdf` for Office docs; plain-text and `pdf` sources short-circuit to raw bytes with `passthrough: true` + a note (Graph rejects `pdf → pdf` with InputFormatNotSupported). `--format markdown` runs the local conversion pipeline (mammoth for docx, sheetjs for xlsx, csv → table, odt/ods/odp via content.xml, plain-text passthrough). All three formats use an M365ChatClient-elevated Graph token (captured at login from m365.cloud.microsoft) — the Teams web client token returns 403 logicalPermissionAccessDenied on historical-version stream content. The CLI follows the SharePoint streamContent redirect internally so the LLM never has to fetch an external URL. caveat for `--format pdf`: Graph sometimes silently falls back to raw source bytes for the current version (which Graph occasionally serves through this endpoint) — when the response carries `passthrough: true`, save with the source extension, not `.pdf` (the global output-path flag refuses the mismatch). | `--drive-id`, `--item-id`, `--version-id`, `--format`, `--include-metadata` | `GET /drives/{drive-id}/items/{item-id}/versions/{version-id}/content` |
-| `extract-drive-item-images` | Extract the embedded images from a OneDrive / SharePoint document. For docx / xlsx / pptx (and their macro-enabled / template variants) it reads the OOXML media parts directly (png/jpg/gif/bmp/tiff/webp/svg) — including original full-resolution / un-cropped originals and images on hidden slides the rendered view never shows. For a pdf it walks every page via unpdf and re-encodes each painted image as PNG (note: page-oriented — it captures images as painted on each page, but NOT layer-hidden/unpainted XObjects or the full uncropped original behind a clipped image). Pair with the global output-dir flag to write every image to a folder; otherwise the bytes ride back base64-encoded in the response. svg rides back as its XML source (which carries the diagram text labels); legacy vector (emf/wmf) and audio/video are skipped. For any other format the command returns a 415 pointing at `download-drive-item-content`. | `--drive-id`, `--item-id` | `GET /drives/{drive-id}/items/{item-id}/content` |
+| `extract-drive-item-images` | Extract the embedded images from a OneDrive / SharePoint document. For docx / xlsx / pptx (and their macro-enabled / template variants) it reads the OOXML media parts directly (png/jpg/gif/bmp/tiff/webp/svg) — including original full-resolution / un-cropped originals and images on hidden slides the rendered view never shows. For a pdf it walks every page via unpdf and re-encodes each painted image as PNG (note: page-oriented — it captures images as painted on each page, but NOT layer-hidden/unpainted XObjects or the full uncropped original behind a clipped image). Pair with the global output-dir flag to write every image to a folder; otherwise the bytes ride back base64-encoded in the response. svg rides back as its XML source (which carries the diagram text labels); legacy vector (emf/wmf) and audio/video are skipped. For any other format the command returns a 415 pointing at `download-drive-item-content`. | `--drive-id`, `--item-id`, `--tenant-id` | `GET /drives/{drive-id}/items/{item-id}/content` |
 | `extract-sharepoint-links-in-documents` | Find every `*.sharepoint.com` URL embedded in a Word / Excel / PowerPoint or OpenDocument file on OneDrive or SharePoint and resolve each one to its driveItem (driveId, itemId, name, webUrl) so the agent can feed those into `download-drive-item-as-pdf` / `-as-markdown` etc. The document sibling of `extract-sharepoint-links-in-mail`. For OOXML (.docx/.xlsx/.pptx) it reads external hyperlinks from the package’s relationship parts (`_rels/*.rels`, `TargetMode="External"`); for OpenDocument (.odt/.ods/.odp) it reads the inline `xlink:href` links in content.xml / styles.xml — either way it catches links wherever they live (body text, headers/footers, cell formulas, slide shapes). Read-only — no conversion happens here. Capped at 25 unique URLs per call (returns `truncated: true` and `skippedCount` when there are more); duplicates are deduplicated; per-link errors are captured inside each entry instead of failing the whole call. Non-zip inputs (pdf/images) return an api_error. | `--drive-id`, `--item-id` | `GET /drives/{drive-id}/items/{item-id}/content` |
 | `get-drive-delta` | Get the incremental change set (added / modified / deleted items) under a OneDrive / SharePoint folder. Use the `@odata.deltaLink` from a previous response to resume. | `--drive-id`, `--item-id`, `--top`, `--select`, `--filter`, `--orderby`, `--expand` | `GET /drives/{drive-id}/items/{item-id}/delta()` |
-| `get-drive-item` | Get the metadata (driveItem resource) of a single file or folder in OneDrive / SharePoint. Use `--select` to slim the response — a full driveItem can run >10 KB with all the optional facets. | `--drive-id`, `--item-id`, `--select`, `--expand` | `GET /drives/{drive-id}/items/{item-id}` |
+| `get-drive-item` | Get the metadata (driveItem resource) of a single file or folder in OneDrive / SharePoint. Use `--select` to slim the response — a full driveItem can run >10 KB with all the optional facets. | `--drive-id`, `--item-id`, `--tenant-id`, `--select`, `--expand` | `GET /drives/{drive-id}/items/{item-id}` |
 | `get-drive-item-analytics` | Return view / activity analytics for a OneDrive / SharePoint file — `allTime` totals (views, viewers) and `lastSevenDays` rollup. Useful for ranking files by attention or detecting stale content. **Known empty case**: returns `{ allTime: null, lastSevenDays: null }` on low-traffic items, or when the calling identity (the Teams web client basic token) lacks the analytics scope on the tenant. Do not interpret nulls as "no views" — interpret as "not available for this caller". For active files where you expect data and see nulls, escalate to a token with `Reports.Read.All`. | `--drive-id`, `--item-id` | `GET /drives/{drive-id}/items/{item-id}/analytics` |
 | `get-drive-item-created-by-user` | Return the `user` resource for whoever created a OneDrive / SharePoint file — full profile, not just the truncated `createdBy.user` summary embedded in the parent driveItem. Useful when you need title / department / mail of the author. Use `--select` to fetch only the fields you care about (e.g. `--select id,displayName,jobTitle,department,mail`). | `--drive-id`, `--item-id`, `--select`, `--expand` | `GET /drives/{drive-id}/items/{item-id}/createdByUser` |
 | `get-drive-item-last-modified-by-user` | Return the full `user` resource for whoever last modified a OneDrive / SharePoint file — sibling to `get-drive-item-created-by-user`. Use `--select` to fetch only specific fields. | `--drive-id`, `--item-id`, `--select`, `--expand` | `GET /drives/{drive-id}/items/{item-id}/lastModifiedByUser` |
@@ -44,18 +45,19 @@ For everything else:
 | `get-drive-root-item` | Get the root folder (driveItem) of a OneDrive / SharePoint drive. Use `--select` to slim the response (e.g. `--select id,name,folder`). | `--drive-id`, `--select`, `--expand` | `GET /drives/{drive-id}/root` |
 | `get-drive-special-folder` | Resolve a OneDrive well-known folder via `--folder-name` (one of `documents`, `photos`, `cameraroll`, `approot`, `music`, `attachments`) without having to navigate from the root. Returns the folder's driveItem (id, name, parentReference, etc.) ready to feed into `list-folder-files` or `download-drive-item-content`. | `--folder-name`, `--select`, `--expand` | `GET /me/drive/special/{folder-name}` |
 | `list-accessible-drives` | Enumerate every drive (document library) the signed-in user can reach — personal OneDrive(s), Teams libraries, SharePoint M365-group sites, drives behind files shared with the user, private/shared Teams channel sites, drives behind recently-used / followed / trending items (activity signals), AND every NON-default document library of each discovered SharePoint site — by unioning `/me/drives`, `/me/joinedTeams`, `/me/memberOf` (Unified groups → `/groups/{id}/drive`), `/me/drive/sharedWithMe`, per-team `/teams/{id}/channels` → `/channels/{ch}/filesFolder` (private/shared channels only — their files live in their own site, not the team default drive), `/me/drive/recent` + `/me/drive/following` + `/me/insights/{trending,used,shared}`, and a path-addressed `/sites/{host}:/sites/{name}:/drives` per discovered site (catches secondary libraries like "Teams Wiki Data" the default-drive vectors miss). Unlike `search-sharepoint-sites-by-name` (which relies on the tenant search index and misses direct-link-only sites + OneDrives), these vectors surface drives the search index never returns; the index in turn returns sites you can open but are not a member of, so the *union of both commands* is the practical maximum on a delegated token. Each drive is tagged with the `sources[]` that found it (a drive can have several). Per-resource "can't reach this one" failures are dropped silently (404 no drive, 403 access-denied / non-member private channel, 423 admin-locked site, 400 stale/unresolvable id); only actionable failures (auth, throttling, 5xx, network) appear in `partialErrors[]`, so it stays signal-only. Fans out one `/groups/{id}/drive` + one `/teams/{id}/channels` call per joined team + member group, a `filesFolder` call per private/shared channel, five fixed activity calls, and one `/sites/{id}/drives` call per discovered site (all capped by `--max-groups`, default 100; raise carefully — large memberships can hit 429 throttling). `/me/followedSites` is not used — it 403s on this token. | `--max-groups`, `--count-files` | `GET /me/drives + /me/joinedTeams + /me/memberOf + /me/drive/sharedWithMe + per-group /groups/<id>/drive + per-team /teams/<id>/channels/<ch>/filesFolder + /me/drive/recent + /me/drive/following + /me/insights/<trending|used|shared> + per-site /sites/<host>:/sites/<name>:/drives` |
-| `list-drive-item-permissions` | List the sharing permissions on a OneDrive / SharePoint file or folder. | `--drive-id`, `--item-id`, `--top`, `--select`, `--filter`, `--orderby`, `--expand` | `GET /drives/{drive-id}/items/{item-id}/permissions` |
-| `list-drive-item-thumbnails` | List thumbnail URLs (small / medium / large) for a OneDrive / SharePoint file. Each thumbnail set has pre-signed CDN URLs you can render in a UI without further auth. | `--drive-id`, `--item-id`, `--top`, `--select`, `--filter`, `--orderby`, `--expand` | `GET /drives/{drive-id}/items/{item-id}/thumbnails` |
-| `list-drive-item-versions` | List the historical versions of a OneDrive / SharePoint file (each save creates a new version). Note: each version's `id` is a stringified float like `"79.0"` (NOT an integer like `79`) — pass it literally to the `download-drive-item-version` command (it accepts an `original | pdf | markdown` format selector); numeric coercion silently fails because Graph rejects `79` against a path templated for stringified floats. | `--drive-id`, `--item-id`, `--top`, `--select`, `--filter`, `--orderby`, `--expand` | `GET /drives/{drive-id}/items/{item-id}/versions` |
+| `list-drive-item-permissions` | List the sharing permissions on a OneDrive / SharePoint file or folder. | `--drive-id`, `--item-id`, `--top`, `--select`, `--filter`, `--orderby`, `--expand`, `--tenant-id` | `GET /drives/{drive-id}/items/{item-id}/permissions` |
+| `list-drive-item-thumbnails` | List thumbnail URLs (small / medium / large) for a OneDrive / SharePoint file. Each thumbnail set has pre-signed CDN URLs you can render in a UI without further auth. | `--drive-id`, `--item-id`, `--top`, `--select`, `--filter`, `--orderby`, `--expand`, `--tenant-id` | `GET /drives/{drive-id}/items/{item-id}/thumbnails` |
+| `list-drive-item-versions` | List the historical versions of a OneDrive / SharePoint file (each save creates a new version). Note: each version's `id` is a stringified float like `"79.0"` (NOT an integer like `79`) — pass it literally to the `download-drive-item-version` command (it accepts an `original | pdf | markdown` format selector); numeric coercion silently fails because Graph rejects `79` against a path templated for stringified floats. | `--drive-id`, `--item-id`, `--top`, `--select`, `--filter`, `--orderby`, `--expand`, `--tenant-id` | `GET /drives/{drive-id}/items/{item-id}/versions` |
 | `list-drives` | List all OneDrive / SharePoint drives the signed-in user has access to. On personal accounts this returns only the user's primary OneDrive (single entry in `value[]`); on tenanted accounts it includes every drive the user can reach including delegated mailboxes and shared SharePoint document libraries. | `--top`, `--skip`, `--select`, `--filter`, `--orderby`, `--expand` | `GET /me/drives` |
-| `list-folder-files` | List the children (files and subfolders) of a folder in OneDrive / SharePoint. | `--drive-id`, `--item-id`, `--top`, `--select`, `--filter`, `--orderby`, `--expand` | `GET /drives/{drive-id}/items/{item-id}/children` |
+| `list-folder-files` | List the children (files and subfolders) of a folder in OneDrive / SharePoint. | `--drive-id`, `--item-id`, `--top`, `--select`, `--filter`, `--orderby`, `--expand`, `--tenant-id` | `GET /drives/{drive-id}/items/{item-id}/children` |
 | `list-followed-drive-items` | List driveItems the signed-in user has explicitly followed (the OneDrive star). A small, hand-curated set of frequently-revisited files, distinct from the algorithmic `list-recent-files` and `list-recently-used-insights`. | `--top`, `--select`, `--filter`, `--orderby`, `--expand` | `GET /me/drive/following` |
 | `list-recent-files` | List the signed-in user's most recently used / opened OneDrive and SharePoint files, ranked by Microsoft's recency signal. The strongest single answer to "what is this user working on right now?". Note: Graph's recent-files feed is signal-driven and can lag the underlying drive by 24-48 hours — `lastModifiedDateTime` here may be older than the file's true mtime. For "what is the actual latest version?" call `list-drive-item-versions` on a specific item. | `--top`, `--select`, `--filter`, `--orderby`, `--expand` | `GET /me/drive/recent` |
 | `list-recently-used-insights` | List documents the signed-in user has *personally* used recently (Microsoft's machine-learning recency signal — distinct from `list-recent-files` which is the OneDrive recency feed). Each item carries a `lastUsed` (a `usageDetails` object) with `lastAccessedDateTime` + `lastModifiedDateTime`. | `--top`, `--skip`, `--select`, `--filter`, `--orderby`, `--expand` | `GET /me/insights/used` |
 | `list-shared-insights` | List documents *shared with* the signed-in user, scored by Microsoft's relevance ranking — sibling to `list-shared-with-me` but with sharing-context details (`sharingHistory[]`, `lastShared.sharedBy`, `lastShared.sharingReference`). | `--top`, `--skip`, `--select`, `--filter`, `--orderby`, `--expand` | `GET /me/insights/shared` |
 | `list-shared-with-me` | List driveItems shared with the signed-in user (typically by colleagues). Each entry includes the original drive + item ID under `remoteItem` so you can chain into `get-drive-item`, `download-drive-item-content`, etc. Note: Graph does NOT honor any OData query parameters on this endpoint (top/select/filter/etc. are all silently ignored), so the CLI does not advertise them. The full collection (~500 items in a typical tenant) is always returned; slice client-side or pair with the global output-path flag to land the raw JSON on disk. | _(none)_ | `GET /me/drive/sharedWithMe` |
 | `list-trending-insights` | List documents trending around the signed-in user — files popular in their working network (colleagues' recent edits, shares, opens). Microsoft's relevance ranking, useful for surfacing unfamiliar but related work. | `--top`, `--skip`, `--select`, `--filter`, `--orderby`, `--expand` | `GET /me/insights/trending` |
-| `resolve-drive-share-link` | Encode a OneDrive / SharePoint sharing URL into the Graph `/shares/{token}` share token (`u!<base64url>` per [shares-get](https://learn.microsoft.com/en-us/graph/api/shares-get)). Pure transformation — no Graph call. Pipe the returned `graphPath` (`/shares/{token}/driveItem`) into a sibling lookup (`get-drive-item`, `download-drive-item-content`, `convert-mail-attachment-to-pdf`, etc.) once the file has been resolved to a `driveItem`. Accepts any `*.sharepoint.com` URL (tenant + `*-my.sharepoint.com` personal OneDrive) and Microsoft's short-link host `1drv.ms`. | `--url` | `GET {url}` |
+| `resolve-drive-share-link` | Resolve a OneDrive / SharePoint sharing URL (a "Copy link" address someone sent you) to the file it points at, returning `driveId` + `itemId` ready to feed `get-drive-item`, `download-drive-item-content`, `convert-drive-item-*`, `extract-drive-item-images`, and the rest of the `*-drive-item` family. It encodes the URL to the Graph `/shares/{token}` share token (`u!<base64url>` per [shares-get](https://learn.microsoft.com/en-us/graph/api/shares-get)) and fetches `/shares/{token}/driveItem` in ONE call (basic token, `Files.Read.All`) — a raw sharing URL carries no ids, so this is the entry point into the drive-item family from a shared link. Accepts any `*.sharepoint.com` URL (tenant + `*-my.sharepoint.com` personal OneDrive) and Microsoft's short-link host `1drv.ms`. **Cross-tenant links work too**: when the URL belongs to a tenant you are only a GUEST in, your home token cannot read its SharePoint (Graph answers `invalidAudienceUri` — it cannot mint a SharePoint token for a foreign tenant), so this command identifies the owning tenant from the URL host and retries with a guest token automatically, then returns that tenant as `tenantId` — see `responseShape` for how to pass it on to the follow-up calls. | `--url` | `GET /shares/(u!<base64url> of {url})/driveItem` |
+| `search-all-files` | Search EVERY file the signed-in user can access — their personal OneDrive, files shared with them, and every SharePoint and Teams (channel) document library they can open — for a free-text query. Unlike `search-my-documents` (personal OneDrive only) or `search-onedrive-files` (one drive by id), this reaches across all accessible drives via the security-trimmed Microsoft Search index; unlike `microsoft-search-query` it returns FILES ONLY (`entityTypes: ["driveItem"]`), not mail/calendar/people/sites. It deep-pages `POST /search/query` with `from`/`size` (200 per page) following the index's `moreResultsAvailable` flag until exhausted, or the ceiling of 25×200 = 5000 files is reached (`truncated: true` — narrow with `--query` to see the rest). Hits are deduped by `hitId`. Each returned `driveItem` carries `id` + `parentReference.driveId`, the pair `download-drive-item-content` / `download-drive-item-as-markdown` need to open it. | `--query` | `POST /search/query` |
 | `search-my-documents` | Search the signed-in user’s default OneDrive for documents matching a free-text query (filename, content, metadata). | `--query`, `--top`, `--select`, `--filter`, `--orderby`, `--expand` | `GET /me/drive/search(q='{query}')` |
 | `search-onedrive-files` | Search a single OneDrive / SharePoint drive for files and folders matching a free-text query. | `--drive-id`, `--query`, `--top`, `--select`, `--filter`, `--orderby`, `--expand` | `GET /drives/{drive-id}/search(q='{query}')` |
 
@@ -79,7 +81,7 @@ For everything else:
 
 | Command | Description | Required params | Graph endpoint |
 |---------|-------------|-----------------|----------------|
-| `get-drive-item-list-item` | Return the SharePoint listItem projection of a OneDrive / SharePoint file — exposes the file's library-defined column values (custom metadata: status, due-date, classification, taxonomy tags, etc.) which are NOT present on the plain `driveItem`. Combine with `list-sharepoint-list-columns` to interpret the column schema. | `--drive-id`, `--item-id`, `--select`, `--expand` | `GET /drives/{drive-id}/items/{item-id}/listItem` |
+| `get-drive-item-list-item` | Return the SharePoint listItem projection of a OneDrive / SharePoint file — exposes the file's library-defined column values (custom metadata: status, due-date, classification, taxonomy tags, etc.) which are NOT present on the plain `driveItem`. Combine with `list-sharepoint-list-columns` to interpret the column schema. | `--drive-id`, `--item-id`, `--select`, `--expand`, `--tenant-id` | `GET /drives/{drive-id}/items/{item-id}/listItem` |
 | `get-sharepoint-list-column` | Return a single column definition from a SharePoint list. | `--site-id`, `--list-id`, `--column-id`, `--select`, `--expand` | `GET /sites/{site-id}/lists/{list-id}/columns/{column-id}` |
 | `get-sharepoint-site` | Get the metadata of a single SharePoint site by its site ID. | `--site-id`, `--select`, `--expand` | `GET /sites/{site-id}` |
 | `get-sharepoint-site-by-path` | Resolve a SharePoint site by its hostname + server-relative path. Use this when you have a SharePoint URL (e.g. `https://contoso.sharepoint.com/sites/Marketing`) but no site ID. | `--hostname`, `--path` | `GET /sites/{hostname}:{path}` |
@@ -122,19 +124,21 @@ For everything else:
 
 | Command | Description | Required params | Graph endpoint |
 |---------|-------------|-----------------|----------------|
-| `convert-mail-attachment-to-markdown` | Convert an Outlook mail attachment to markdown. Polymorphic on the attachment’s `@odata.type`: fileAttachment decodes the inline bytes and runs them through the local conversion pipeline (docx via mammoth, xlsx via sheetjs, csv as markdown table, odt/ods/odp via content.xml, pptx as per-slide text (titles + bullets + speaker notes inline), pdf via text-layer extraction (unpdf → text/plain), legacy .xls via sheetjs and legacy .doc via word-extractor (text only), an Outlook .msg attachment rendered to markdown — headers + body with its own attachments converted recursively — plus plain-text passthrough); referenceAttachment resolves via /shares/{token}/driveItem and routes through the same dispatcher; itemAttachment (embedded mail / event / contact) is rendered locally via dedicated renderers. For pptx layout / images, `convert-mail-attachment-to-pdf` + a vision model reads the rendered deck better. A scanned / image-only PDF (no text layer), legacy .ppt, and rtf/etc. point to the PDF sibling. Loop/Fluid/Whiteboard reference-attachments use Graph `?format=html` (the four inputs Microsoft documents). | `--message-id`, `--attachment-id`, `--include-metadata` | `GET /me/messages/{message-id}/attachments/{attachment-id}` |
+| `convert-mail-attachment-to-markdown` | Convert an Outlook mail attachment to markdown. Polymorphic on the attachment’s `@odata.type`: fileAttachment decodes the inline bytes and runs them through the local conversion pipeline (docx via mammoth, xlsx via sheetjs, csv as markdown table, odt/ods/odp via content.xml, pptx as per-slide text (titles + bullets + speaker notes inline), pdf via text-layer extraction (unpdf → text/plain), legacy .xls via sheetjs and legacy .doc via word-extractor (text only), an Outlook .msg attachment rendered to markdown — headers + body (quoted reply chain stripped unless `--keep-quoted true`; inline `cid:` images shown as placeholders) with its own attachments converted recursively — plus plain-text passthrough); referenceAttachment resolves via /shares/{token}/driveItem and routes through the same dispatcher; itemAttachment (embedded mail / event / contact) is rendered locally via dedicated renderers. For pptx layout / images, `convert-mail-attachment-to-pdf` + a vision model reads the rendered deck better. A scanned / image-only PDF (no text layer), legacy .ppt, and rtf/etc. point to the PDF sibling. Loop/Fluid/Whiteboard reference-attachments use Graph `?format=html` (the four inputs Microsoft documents). | `--message-id`, `--attachment-id`, `--include-metadata`, `--keep-quoted` | `GET /me/messages/{message-id}/attachments/{attachment-id}` |
 | `convert-mail-attachment-to-pdf` | Convert an Outlook mail attachment to PDF on the fly. Polymorphic on the attachment’s `@odata.type`: fileAttachment uploads the bytes to a temp folder under /me/drive (large files use Graph’s chunked upload session — no 4 MB ceiling), runs ?format=pdf, then deletes the temp item; referenceAttachment resolves via /shares/{token}/driveItem and runs ?format=pdf in place; plain-text source extensions and `pdf` sources short-circuit to a raw-bytes envelope on either path (Graph’s `?format=pdf` does not accept `pdf` as an input format — pdf attachments are returned as-is). itemAttachment (embedded mail/event/contact) is unsupported here — Graph rejects those source types — use convert-mail-attachment-to-markdown instead. Worst-case wall-clock for huge attachments is ~22 minutes (1 metadata GET + up-to-20 chunk PUTs + 1 convert GET + 1 cleanup DELETE, each capped at 60s). | `--message-id`, `--attachment-id` | `GET /me/messages/{message-id}/attachments/{attachment-id}` |
-| `convert-mail-attachment-zip` | Unzip a `.zip` Outlook mail attachment and convert every contained file in one call — the mail-side mirror of `convert-drive-item-zip`, so reading a zipped vendor deck doesn't need `get-mail-attachment` + manual `unzip` + per-file conversion. Pulls the fileAttachment bytes, unzips them (legacy GBK / CP437 entry names — Chinese vendor archives written by WinRAR / Windows Explorer — are decoded correctly, not mojibaked), and runs each file through the local pipelines: Office files (docx/xlsx/pptx/odt/ods/odp and macro-enabled / template variants) → markdown; plain-text entries decoded inline; legacy OLE .xls (sheetjs) and .doc (word-extractor, text only) extracted; an inner Outlook .msg rendered; PDFs have their text layer extracted; images, binaries, nested archives, legacy .ppt, and scanned/image-only PDFs are listed with a note (not unpacked) so one unsupported entry never fails the whole archive. Pass `--include-metadata true` to append each Office file's side-channel metadata block. Capped at 100 entries; beyond that the response is flagged `truncated`. itemAttachment / referenceAttachment are rejected (no inline zip payload). | `--message-id`, `--attachment-id`, `--include-metadata` | `GET /me/messages/{message-id}/attachments/{attachment-id}` |
-| `convert-mail-to-markdown` | Render a single Outlook email as markdown — headers (`**Subject:**`, `**From:**`, `**To:**`, `**Cc:**` only when present, `**Date:**`), followed by the body run through turndown. By default, inline images (`isInline:true` + `image/*` content-type, size ≤ 2 MB) are embedded as base64 `data:` URIs so the output is self-contained (non-image inline attachments are NOT embedded; oversize inline images are replaced with a placeholder note). For LLM callers that only want the text body, pass `--inline-images false` to skip the per-image bytes fetch entirely — the body keeps raw `cid:<contentId>` references and the inline images surface in the file-attachments list so you can decide whether to fetch them separately via `get-mail-attachment`. File attachments are always listed below the body by name + size + id; their bytes are NOT fetched here — call `convert-mail-attachment-to-pdf` or `get-mail-attachment` with the id when you actually need them. Staged-fetch design: one call for the body, one for the attachments-metadata list (only if `hasAttachments:true`), and one per small inline image — replaces the old `?$expand=attachments` which timed out / truncated on messages with multi-MB attachments. | `--message-id`, `--inline-images`, `--keep-quoted` | `GET /me/messages/{message-id}` |
-| `create-forward-draft` | Create an UNSENT forward draft of an existing message. POST /me/messages/{id}/createForward mints the draft (FW: subject, quoted original) with your comment placed above the quote and the recipients set, in one call. Redirects a thread to the right owner without leaving the CLI. The draft is saved in Drafts and can be reviewed, edited, and sent from any Outlook client; the CLI still cannot send. | `--forward-message-id`, `--to-recipients`, `--cc-recipients`, `--body-content`, `--subject` | `POST /me/messages/{forward-message-id}/createForward (+ optional body-free PATCH for cc / subject)` |
-| `create-mail-draft` | Create a new mail draft. POST /me/messages (or /me/mailFolders/{id}/messages when --mail-folder-id is set). The draft is saved in the Drafts folder (or the specified folder) and can be sent later via the Outlook client or Graph sendMail. Recipients are comma-separated email addresses. Returns the created message object with its id — use this id with update-mail-draft to modify the draft before sending. | `--subject`, `--body-content`, `--body-content-type`, `--to-recipients`, `--cc-recipients`, `--bcc-recipients`, `--importance`, `--mail-folder-id` | `POST /me/messages (or /me/mailFolders/{mail-folder-id}/messages)` |
-| `create-reply-draft` | Create an UNSENT reply-all draft threaded on an existing message. POST /me/messages/{id}/createReplyAll mints the draft (inherited recipients, RE: subject, quoted history) with your reply text placed above the quote, in one call. Reply-all by design - dropping recipients is a deliberate act for the human in Outlook, not a default. The draft is saved in Drafts and can be reviewed, edited, and sent from any Outlook client; the CLI still cannot send. | `--reply-to-message-id`, `--body-content`, `--subject` | `POST /me/messages/{reply-to-message-id}/createReplyAll (+ optional body-free PATCH for subject)` |
+| `convert-mail-attachment-zip-to-markdown` | Unzip a `.zip` Outlook mail attachment and convert every contained file in one call — the mail-side mirror of `convert-drive-item-zip-to-markdown`, so reading a zipped vendor deck doesn't need `get-mail-attachment` + manual `unzip` + per-file conversion. Pulls the fileAttachment bytes, unzips them (legacy GBK / CP437 entry names — Chinese vendor archives written by WinRAR / Windows Explorer — are decoded correctly, not mojibaked), and runs each file through the local pipelines: Office files (docx/xlsx/pptx/odt/ods/odp and macro-enabled / template variants) → markdown; plain-text entries decoded inline; legacy OLE .xls (sheetjs) and .doc (word-extractor, text only) extracted; an inner Outlook .msg rendered; PDFs have their text layer extracted; images, binaries, nested archives, legacy .ppt, and scanned/image-only PDFs are listed with a note (not unpacked) so one unsupported entry never fails the whole archive. Pass `--include-metadata true` to append each Office file's side-channel metadata block. Capped at 100 entries; beyond that the response is flagged `truncated`. itemAttachment / referenceAttachment are rejected (no inline zip payload). | `--message-id`, `--attachment-id`, `--include-metadata`, `--keep-quoted` | `GET /me/messages/{message-id}/attachments/{attachment-id}` |
+| `convert-mail-to-markdown` | Render a single Outlook email as markdown — headers (`**Subject:**`, `**From:**`, `**To:**`, `**Cc:**` only when present, `**Date:**`), followed by the body run through turndown. By default NO image bytes are fetched: every inline `cid:` image renders as a readable `[inline image: <name>]` placeholder and the images surface in the file-attachments list, so the output stays close to the text size (an email whose 6 KB body carried 30 KB of signature-image base64 now ships at ~6 KB). Pass `--inline-images true` to embed inline images (`isInline:true` + `image/*` content-type, size ≤ 2 MB) as base64 `data:` URIs for self-contained output (non-image inline attachments are never embedded; oversize inline images keep a placeholder note; a cid whose per-image fetch fails degrades to the placeholder too). File attachments are always listed below the body by name + size + id; their bytes are NOT fetched here — call `convert-mail-attachment-to-pdf` or `get-mail-attachment` with the id when you actually need them. Staged-fetch design: one call for the body, one for the attachments-metadata list (only if `hasAttachments:true`), and with `--inline-images true` one per small inline image — replaces the old `?$expand=attachments` which timed out / truncated on messages with multi-MB attachments. | `--message-id`, `--inline-images`, `--keep-quoted` | `GET /me/messages/{message-id}` |
+| `create-forward-draft` | Create an UNSENT forward draft of an existing message. POST /me/messages/{id}/createForward mints the draft (FW: subject, quoted original) with your comment placed above the quote and the recipients set, in one call. Redirects a thread to the right owner without leaving the CLI. The draft is saved in Drafts and can be reviewed, edited, and sent from any Outlook client; the CLI still cannot send. | `--forward-message-id`, `--to-recipients`, `--cc-recipients`, `--body-content`, `--subject`, `--body-content-type` | `POST /me/messages/{forward-message-id}/createForward (+ optional body-free PATCH for cc / subject)` |
+| `create-mail-draft` | Create a new mail draft. POST /me/messages (or /me/mailFolders/{id}/messages when --mail-folder-id is set). The draft is saved in the Drafts folder (or the specified folder) and can be sent later via the Outlook client or Graph sendMail. Recipients are comma-separated email addresses. Returns a slim confirmation (id, subject, recipients, importance, bodyPreview, …) - NOT the body you just wrote; read the full draft back with get-mail-message if you need it. Use the returned id with update-mail-draft to modify before sending. | `--subject`, `--body-content`, `--body-content-type`, `--to-recipients`, `--cc-recipients`, `--bcc-recipients`, `--importance`, `--mail-folder-id` | `POST /me/messages (or /me/mailFolders/{mail-folder-id}/messages)` |
+| `create-reply-draft` | Create an UNSENT reply draft threaded on an existing message. POST /me/messages/{id}/createReplyAll mints the draft (inherited recipients, RE: subject, quoted history) with your reply text placed above the quote, in one call. Reply-all by default - dropping recipients is a deliberate act, so pass --reply-all false to reply to the sender only, which switches the action to createReply. The draft is saved in Drafts and can be reviewed, edited, and sent from any Outlook client; the CLI still cannot send. | `--reply-to-message-id`, `--body-content`, `--body-content-type`, `--subject`, `--reply-all` | `POST /me/messages/{reply-to-message-id}/createReplyAll, or /createReply when {reply-all} is false (+ optional body-free PATCH for subject)` |
 | `extract-mail-attachment-images` | Extract the embedded images from an Outlook mail attachment that is a pdf or a docx / xlsx / pptx (and their macro-enabled / template variants). OOXML reads the media parts directly (png/jpg/gif/bmp/tiff/webp/svg), including full-resolution / un-cropped originals and images on hidden slides; pdf walks every page via unpdf and re-encodes each painted image as PNG (page-oriented — not layer-hidden/unpainted/uncropped originals). fileAttachment decodes the inline bytes; referenceAttachment resolves via /shares/{token}/driveItem and fetches the content. Pair with the global output-dir flag to write every image to a folder; otherwise the bytes ride back base64-encoded. svg rides back as its XML source (which carries the diagram text labels); legacy vector (emf/wmf) and audio/video are skipped. itemAttachment and unsupported formats return a 415. | `--message-id`, `--attachment-id` | `GET /me/messages/{message-id}/attachments/{attachment-id}` |
 | `extract-sharepoint-links-in-mail` | Find every `*.sharepoint.com` URL in the body of a single Outlook email and resolve each one to its driveItem (driveId, itemId, name, webUrl) so the agent can feed those into `download-drive-item-as-pdf` / `-as-markdown` etc. Read-only — no conversion happens here. Capped at 25 unique URLs per call to bound fan-out (returns `truncated: true` and `skippedCount` when the body has more); duplicate URLs are deduplicated. Per-link errors are captured inside each entry instead of failing the whole call. | `--message-id` | `GET /me/messages/{message-id}` |
+| `find-mail-drafts` | Find existing drafts on a mail thread WITHOUT trusting a conversationId $filter. Reply and forward drafts do not always inherit the inbound message conversationId (a thread can split across several), and Graph $filter on the Drafts folder is not read-your-writes consistent, so filtering Drafts by conversationId misses drafts. This command instead scans the 50 most recently modified drafts and matches them client-side on a normalized subject (stripping RE:/FW: and localized reply/forward prefixes) plus, optionally, a shared recipient. Use it before create-reply-draft to avoid creating a duplicate: if a match comes back, revise it with update-mail-draft instead of making a new one. Read-only. | `--subject`, `--to-recipients` | `GET /me/mailFolders/drafts/messages?$top=50&$orderby=lastModifiedDateTime desc&$select=id,subject,toRecipients,ccRecipients,conversationId,lastModifiedDateTime,webLink (a read-only scan of the 50 most recently modified drafts; each is matched CLIENT-SIDE on a normalized {subject} and, when given, a shared {to-recipients} address, so neither value is sent to Graph)` |
 | `get-mail-attachment` | Get a single attachment on an Outlook message (metadata, plus the base64 `contentBytes` for file attachments). For fileAttachments, the response also carries a `base64` mirror of `contentBytes` so the global output-path flag can land the bytes on disk in one call — and when an output-path is set the CLI strips BOTH `contentBytes` and `base64` from stdout, leaving a compact metadata envelope with `savedTo` (the file is on disk; no multi-MB base64 in the terminal). When you only want metadata, use `--select id,name,contentType,size` to skip the `contentBytes` payload. | `--message-id`, `--attachment-id`, `--select`, `--expand` | `GET /me/messages/{message-id}/attachments/{attachment-id}` |
-| `get-mail-message` | Get a single Outlook message by ID. The CLI ships a slim default `--select=id,subject,from,toRecipients,ccRecipients,receivedDateTime,hasAttachments,isRead,importance,bodyPreview` so an LLM caller doesn't pull a 41 KB resource just to read a subject line. Pass `--select id,subject,body` (or any other comma-separated field list) to override; for the raw RFC-822 source use `get-mail-message-mime` instead. | `--message-id`, `--select`, `--expand` | `GET /me/messages/{message-id}` |
+| `get-mail-message` | Get a single Outlook message by ID. The CLI ships a slim default `--select=id,subject,from,toRecipients,ccRecipients,receivedDateTime,hasAttachments,isRead,importance,bodyPreview,conversationId` (`conversationId` gives you the thread, e.g. for `list-conversation-messages`) so an LLM caller doesn't pull a 41 KB resource just to read a subject line. Pass `--select id,subject,body` (or any other comma-separated field list) to override; for the raw RFC-822 source use `get-mail-message-mime` instead. | `--message-id`, `--select`, `--expand` | `GET /me/messages/{message-id}` |
 | `get-mail-message-mime` | Return the raw RFC 5322 MIME source of a single Outlook message — full headers, every attachment encoded inline. Useful for archiving, full-fidelity forensic inspection, or feeding into a tool that reads MIME directly. For human-readable content prefer `get-mail-message` or `convert-mail-to-markdown`. | `--message-id` | `GET /me/messages/{message-id}/$value` |
 | `get-mail-rule` | Return a single Outlook message rule by ID, including its conditions and actions. Sibling to `list-mail-rules`. `--mail-folder-id` defaults to `inbox` (the only folder where rules actually live in Graph); the flag is preserved for callers that want to pass a resolved Inbox ID explicitly. | `--mail-folder-id`, `--message-rule-id` | `GET /me/mailFolders/{mail-folder-id}/messageRules/{message-rule-id}` |
+| `get-mail-signature` | Read your own email signature as HTML, lifted from a message you already sent. Graph-created drafts carry NO signature (create-mail-draft, create-reply-draft, and create-forward-draft all produce unsigned bodies), so this is where you get one: take the `text` this returns, append it to your reply text, and hand the result to `update-mail-draft` in comment mode (HTML body-content-type) to place it above the quoted history. Scans your last 10 sent messages newest-first and returns the first `<div id="Signature">` block it finds, stopping there, with any logo the block references embedded as a base64 data: URI so the HTML renders on its own. Read-only. NOTE: the marker is written by Outlook on the web and new Outlook; mail composed in Outlook desktop does not carry it, so pin a webmail-sent message with --message-id if the scan finds nothing. | `--message-id` | `GET /me/mailFolders/sentitems/messages (scan, skipped when {message-id} is given) then /me/messages/{message-id}?$select=body,sentDateTime,hasAttachments (+ /attachments per referenced logo)` |
 | `get-mailbox-settings` | Get the signed-in user's Outlook mailbox settings (timezone, working hours, automatic replies). Note: Graph silently ignores `$select` / `$expand` on this endpoint, so the CLI does NOT expose them — the full payload (including the auto-reply HTML body) is always returned. Slim client-side if you only need a subset. | _(none)_ | `GET /me/mailboxSettings` |
 | `get-shared-mailbox-message` | Return a single message from a shared / delegated mailbox. Use `--select` to fetch only specific fields (e.g. `--select id,subject,from,receivedDateTime`) — sibling to `get-mail-message` for /me. | `--user-id`, `--message-id`, `--select`, `--expand` | `GET /users/{user-id}/messages/{message-id}` |
 | `list-conversation-messages` | List every message in a single Outlook conversation (thread) using `$filter=conversationId eq '...'`. Reconstructs a complete thread regardless of which subject lines or folders the replies landed in. Accepts the OData passthrough flags top/skip/select/expand — the filter and orderby passthroughs are intentionally omitted (the path already pins a `$filter`, and Graph rejects this filter combined with `$orderby` as `InefficientFilter` since `conversationId` is not a sortable index). The caller can sort by `receivedDateTime` client-side. KQL `$search` does not index `conversationId`, so `$filter` is the only documented Graph idiom for whole-thread retrieval. | `--conversation-id`, `--top`, `--skip`, `--select`, `--expand` | `GET /me/messages?$filter=conversationId eq '{conversation-id}'` |
@@ -147,15 +151,15 @@ For everything else:
 | `list-mail-folder-messages-delta` | Track incremental changes (added / updated / deleted messages) within a single mail folder using Microsoft Graph delta tokens. The first call returns the current snapshot plus a `@odata.deltaLink`; subsequent calls with that link return only what has changed since. | `--mail-folder-id`, `--top`, `--skip`, `--select`, `--filter`, `--orderby`, `--expand` | `GET /me/mailFolders/{mail-folder-id}/messages/delta()` |
 | `list-mail-folders` | List the top-level mail folders in the signed-in user’s Outlook mailbox (Inbox, Sent Items, etc.). | `--top`, `--skip`, `--select`, `--filter`, `--orderby`, `--expand` | `GET /me/mailFolders` |
 | `list-mail-folders-delta` | Track incremental changes to the mail-folder tree itself (folders added / renamed / deleted). The first call returns the current snapshot plus a `@odata.deltaLink`; subsequent calls with that link return only what has changed. Companion to `list-mail-folder-messages-delta` which tracks message changes inside one folder. Note: Graph explicitly rejects `$top`, `$filter`, `$orderby`, and `$search` on this delta endpoint (`ErrorInvalidUrlQuery: not supported with change tracking over the 'Folders' resource`), so the OData passthrough is intentionally NOT exposed here. | _(none)_ | `GET /me/mailFolders/delta()` |
-| `list-mail-messages` | List the most recent messages from across the signed-in user's entire Outlook mailbox (every folder including Sent, Archive, Junk; default sort `receivedDateTime` desc). The CLI ships a slim default `--select=id,subject,from,toRecipients,ccRecipients,receivedDateTime,hasAttachments,isRead,importance,bodyPreview` so a page of 25 messages stays ~30-60 KB instead of ~1 MB. Pass `--select id,subject,body` (or any other comma-separated field list) to override. Use `list-mail-folder-messages` to scope to a single folder such as Inbox. | `--top`, `--skip`, `--select`, `--filter`, `--orderby`, `--expand` | `GET /me/messages` |
+| `list-mail-messages` | List the most recent messages from across the signed-in user's entire Outlook mailbox (every folder including Sent, Archive, Junk; default sort `receivedDateTime` desc). The CLI ships a slim default `--select=id,subject,from,toRecipients,ccRecipients,receivedDateTime,hasAttachments,isRead,importance,bodyPreview,conversationId` (`conversationId` groups messages into a thread and can be handed to `list-conversation-messages`) so a page of 25 messages stays ~30-60 KB instead of ~1 MB. Pass `--select id,subject,body` (or any other comma-separated field list) to override. Use `list-mail-folder-messages` to scope to a single folder such as Inbox. | `--top`, `--skip`, `--select`, `--filter`, `--orderby`, `--expand` | `GET /me/messages` |
 | `list-mail-rules` | List the message rules on the Outlook Inbox. Microsoft Graph only supports message rules on the Inbox folder; passing any other folder ID (drafts, sentitems, archive, a custom folder) returns `MailFolderNotSupportedError` from Graph. `--mail-folder-id` defaults to `inbox` because that is the only value Graph accepts; the flag is kept (optional) for callers that want to pass a resolved Inbox ID explicitly. Note: Graph silently ignores every OData passthrough on this endpoint, so the CLI does NOT expose them — the full rule set is always returned. | `--mail-folder-id` | `GET /me/mailFolders/{mail-folder-id}/messageRules` |
 | `list-outlook-categories` | List the signed-in user's Outlook color categories — the named tags that can be applied to mail, calendar items, and contacts. Each entry has `displayName` and a `color` from Outlook's preset palette. Note: Graph silently ignores every OData passthrough on this endpoint (`$top`, `$skip`, `$select`, `$filter`, `$orderby`, `$expand`), so the CLI does not expose any of those flags — the full collection is always returned. Slice client-side. | _(none)_ | `GET /me/outlook/masterCategories` |
 | `list-shared-mailbox-folder-messages` | List messages in a single folder of a shared / delegated mailbox. | `--user-id`, `--mail-folder-id`, `--top`, `--skip`, `--select`, `--filter`, `--orderby`, `--expand` | `GET /users/{user-id}/mailFolders/{mail-folder-id}/messages` |
 | `list-shared-mailbox-messages` | List messages from a shared or delegated mailbox the signed-in user has read access to. Same shape as `list-mail-messages` but scoped to a specific mailbox owner. 403 if the signed-in user does not have shared access to that mailbox. | `--user-id`, `--top`, `--skip`, `--select`, `--filter`, `--orderby`, `--expand` | `GET /users/{user-id}/messages` |
-| `read-mail-attachment` | Read an Outlook mail attachment whatever it is — one command that auto-routes by file type, preferring the content-type when the filename extension is misleading (a real `.jpg` that is actually a spreadsheet still converts), so a caller never has to choose between the convert-mail-attachment-* siblings. A `.zip` fileAttachment is unpacked and every entry converted (mirrors `convert-mail-attachment-zip`, returning the `{ count, files }` envelope; legacy GBK/CP437 names decoded). Any other attachment — docx/xlsx/pptx/odt/ods/odp + macro/template variants → markdown, csv → table, pdf → text layer (with `pageCount`), legacy .xls/.doc extracted, an inner Outlook .msg rendered recursively, plain text passed through, referenceAttachment resolved via `/shares`, and itemAttachment (embedded mail/event/contact) rendered — goes through the same dispatch as `convert-mail-attachment-to-markdown` (returning its `{ contentType, size, text }` envelope). Images, scanned/image-only PDFs, and legacy .ppt return an actionable 415 pointing at `convert-mail-attachment-to-pdf` + a vision model or `get-mail-attachment` for the raw bytes. Pass `--include-metadata true` to append Office side-channel metadata. Use the explicit `convert-mail-attachment-to-markdown` / `-to-pdf` / `-zip` siblings only when you need to force a specific output format. | `--message-id`, `--attachment-id`, `--include-metadata` | `GET /me/messages/{message-id}/attachments/{attachment-id}` |
+| `read-mail-attachment` | Read an Outlook mail attachment whatever it is — one command that auto-routes by file type, preferring the content-type when the filename extension is misleading (a real `.jpg` that is actually a spreadsheet still converts), so a caller never has to choose between the convert-mail-attachment-* siblings. A `.zip` fileAttachment is unpacked and every entry converted (mirrors `convert-mail-attachment-zip-to-markdown`, returning the `{ count, files }` envelope; legacy GBK/CP437 names decoded). Any other attachment — docx/xlsx/pptx/odt/ods/odp + macro/template variants → markdown, csv → table, pdf → text layer (with `pageCount`), legacy .xls/.doc extracted, an inner Outlook .msg rendered recursively (quoted chain stripped unless `--keep-quoted true`), plain text passed through, referenceAttachment resolved via `/shares`, and itemAttachment (embedded mail/event/contact) rendered — goes through the same dispatch as `convert-mail-attachment-to-markdown` (returning its `{ contentType, size, text }` envelope). Images, scanned/image-only PDFs, and legacy .ppt return an actionable 415 pointing at `convert-mail-attachment-to-pdf` + a vision model or `get-mail-attachment` for the raw bytes. Pass `--include-metadata true` to append Office side-channel metadata. Use the explicit `convert-mail-attachment-to-markdown` / `-to-pdf` / `-zip` siblings only when you need to force a specific output format. | `--message-id`, `--attachment-id`, `--include-metadata`, `--keep-quoted` | `GET /me/messages/{message-id}/attachments/{attachment-id}` |
 | `resolve-mail-link` | Parse a Microsoft Outlook web mail link (the URL emitted by the "Copy link" / address-bar share of an email) into its `messageId`. Pure transformation — no Graph call. Pipe the result into `get-mail-message` to fetch the body, or `convert-mail-to-markdown` to render it. For Outlook calendar links use `resolve-calendar-link` instead — this command rejects them with a pointer. | `--url` | `GET {url}` |
-| `search-mail-messages` | Search the signed-in user's entire Outlook mailbox using KQL or free text. Results are ranked by Graph relevance. The CLI ships a slim default `--select=id,subject,from,toRecipients,ccRecipients,receivedDateTime,hasAttachments,isRead,importance,bodyPreview` (same as `list-mail-messages`) so a 3-result page stays ~3 KB instead of ~30 KB. Pass `--select id,subject,body` to widen, or override entirely. Note: Graph does not allow `$search` and `$filter` together — the CLI rejects `--filter` client-side with a pointer to `list-mail-messages` (which supports OData filtering). For sorting, server-side `$orderby` is also not allowed with `$search`; use the relevance ranking Graph returns. **KQL quoting gotcha**: pass the raw KQL expression, e.g. `--query 'subject:invoice from:alice'`; do NOT wrap your terms in extra double-quotes (Graph then rejects with `BadRequest: An identifier was expected at position 0` because it sees `"..."` after the `$search=` interpolation). The CLI already wraps the entire `--query` value in `"..."` on the wire. | `--query`, `--top`, `--skip`, `--select`, `--filter`, `--orderby`, `--expand` | `GET /me/messages?$search="{query}"` |
-| `update-mail-draft` | Update an existing mail draft. PATCH /me/messages/{id} — modifies a draft created by create-mail-draft (or any existing draft in the Drafts folder). Only the fields you pass are updated; omitted fields are left unchanged. At least one field must be provided. Returns the updated message object. Use get-mail-message to verify the final state before sending. | `--message-id`, `--subject`, `--body-content`, `--body-content-type`, `--to-recipients`, `--cc-recipients`, `--bcc-recipients`, `--importance` | `PATCH /me/messages/{message-id}` |
+| `search-mail-messages` | Search the signed-in user's entire Outlook mailbox using KQL or free text. Results are ranked by Graph relevance. The CLI ships a slim default `--select=id,subject,from,toRecipients,ccRecipients,receivedDateTime,hasAttachments,isRead,importance,bodyPreview,conversationId` (same as `list-mail-messages`; `conversationId` is included so you can group hits into a thread or feed one to `list-conversation-messages`) so a 3-result page stays ~3 KB instead of ~30 KB. Pass `--select id,subject,body` to widen, or override entirely. Note: Graph does not allow `$search` and `$filter` together — the CLI rejects `--filter` client-side with a pointer to `list-mail-messages` (which supports OData filtering). For sorting, server-side `$orderby` is also not allowed with `$search`; use the relevance ranking Graph returns. **Exact-phrase search works**: `--query '"budget allocation"'` and embedded field phrases like `--query 'subject:"Contoso A2 & B7 timeline"'` are supported — the CLI escapes your double quotes into KQL phrase quotes, wraps the whole expression in the `"…"` Graph requires, and percent-encodes the value so `&`, `#`, and `+` are wire-safe. Pass raw KQL otherwise, e.g. `--query 'subject:invoice from:alice'`. | `--query`, `--top`, `--skip`, `--select`, `--filter`, `--orderby`, `--expand` | `GET /me/messages?$search="{query}"` |
+| `update-mail-draft` | Update an existing mail draft. PATCH /me/messages/{id} — modifies a draft created by create-mail-draft (or any existing draft in the Drafts folder). Only the fields you pass are updated; omitted fields are left unchanged. At least one field must be provided. On a THREADED draft (one made by create-reply-draft / create-forward-draft), revise your text with --comment, which rewrites only what sits above the quoted history and leaves the quote byte-identical; --body-content would replace the whole body and drop the thread. Passing an EMPTY string to a recipient flag clears that list, which is how you drop recipients a reply-all or forward inherited; omitting the flag leaves the list alone. Returns a slim confirmation (id, subject, recipients, importance, bodyPreview, …) - NOT the full body, which you just wrote; read it back with get-mail-message if you need the whole draft before sending. | `--message-id`, `--subject`, `--body-content`, `--comment`, `--body-content-type`, `--to-recipients`, `--cc-recipients`, `--bcc-recipients`, `--importance` | `PATCH /me/messages/{message-id} (+ a GET of body,isDraft first when {comment} is used)` |
 
 ### Notes (OneNote)
 
@@ -182,6 +186,7 @@ For everything else:
 | `get-my-manager` | Return the signed-in user's manager (a single `user` resource). When no manager is set in the directory, Graph returns 404 `Request_ResourceNotFound`; this command maps that one specific 404 to `{ ok: true, data: { manager: null, note: '...' } }` so an LLM can distinguish 'no manager' from a permission failure without parsing prose. Use `--select` to slim the response (e.g. `--select id,displayName,mail`). | `--select`, `--expand` | `GET /me/manager` |
 | `get-my-profile-photo` | Download the signed-in user's profile photo (largest available size), inlined. The CLI follows the Graph 302 → CDN redirect internally so the LLM never has to fetch an external URL. | _(none)_ | `GET /me/photo/$value` |
 | `get-organization` | Return the tenant's organization metadata — display name, country, verified domains, business phones, technical / security notification contacts, assigned Microsoft 365 SKUs / licensing. Graph wraps the single organization resource under `value[]` (— even though only one tenant exists, the endpoint returns a collection). The full resource is ~57 KB; use `--select` to slim it (e.g. `--select id,displayName,verifiedDomains`). | `--select`, `--expand` | `GET /organization` |
+| `get-user` | Look up a directory user. Pass an Azure AD id, UPN, or email as --user-id and get that user's FULL profile (displayName, mail, jobTitle, department, officeLocation, phones) via GET /users/{id} on the elevated M365 token — preflight it with `ask-marcel-office scopes-check` (no Graph call) and run `ask-marcel-office login` first if it is cold. An email resolves even when it is the user's `mail` rather than their sign-in UPN: guest / B2B users carry a `#EXT#` UPN whose local part is NOT their email address, so when the direct lookup 404s the command falls back to `GET /users?$filter=mail eq '<email>'` and returns the single match. Only THIS tenant's directory is queried: a person's home-tenant object id (e.g. a cross-tenant Teams `8:orgid:<home-id>` participant, whose id lives in their own tenant) and any email that is not their `mail`/UPN here are unresolvable by design — reach an external person via their LOCAL guest projection (by name, or by their real `mail`), never by their home id. Pass a NAME instead and it searches your relevant-people graph (GET /me/people) and returns candidate matches (id, displayName, mail, jobTitle, department) so you can pick the right person and re-query. Re-query by the candidate's `id` when it is a directory GUID; an EXTERNAL contact's candidate carries a base64-ish People-API id instead, which nothing can resolve — re-query those by the candidate's `mail` (the CLI rejects an opaque contact id with that remedy rather than returning empty matches). Name search covers colleagues in your people graph, not the whole tenant directory; use `microsoft-search-query` for a broader tenant-wide person search. Without `--select`, the profile ships a default projection of id, displayName, userPrincipalName, mail, jobTitle, department, officeLocation, businessPhones, mobilePhone. | `--user-id`, `--select`, `--expand` | `GET /users/{user-id}  (id / UPN; an email that misses falls back to /users?$filter=mail eq)  OR /me/people?$search="{user-id}"  (a bare name)` |
 | `get-user-manager` | Return a specific user's manager (a single `user` resource). When the user has no manager set in the directory, Graph returns 404 `Request_ResourceNotFound`; this command maps that one specific 404 to `{ ok: true, data: { manager: null, note: '...' } }` (same shape as `get-my-manager`) so an LLM can distinguish 'no manager' from 'unknown user' with a single discriminator across both commands. Use `--select` to slim the response. | `--user-id`, `--select`, `--expand` | `GET /users/{user-id}/manager` |
 | `list-group-members` | List members of an Azure AD / Microsoft 365 group. Returns users, groups, and other directoryObjects depending on the group's membership. | `--group-id`, `--top`, `--skip`, `--select`, `--filter`, `--orderby`, `--expand` | `GET /groups/{group-id}/members` |
 | `list-group-owners` | List the owners of an Azure AD / Microsoft 365 group. | `--group-id`, `--top`, `--skip`, `--select`, `--filter`, `--orderby`, `--expand` | `GET /groups/{group-id}/owners` |
@@ -197,7 +202,7 @@ For everything else:
 
 | Command | Description | Required params | Graph endpoint |
 |---------|-------------|-----------------|----------------|
-| `convert-calendar-event-attachment-to-markdown` | Convert an attachment on an Outlook calendar event to markdown. Polymorphic on the attachment’s `@odata.type` (shares the mail-attachment pipeline): fileAttachment decodes the inline bytes and runs them through the local conversion pipeline (docx via mammoth, xlsx via sheetjs, csv as markdown table, odt/ods/odp via content.xml, plus plain-text passthrough); referenceAttachment resolves via /shares/{token}/driveItem and routes through the same dispatcher; itemAttachment (embedded mail / event / contact) is rendered locally. For pptx decks attached to a meeting, `convert-calendar-event-attachment-to-pdf` preserves slide layout (a pptx here yields only its speaker notes / titles / comments via `## PPTX metadata` with `--include-metadata true`). For pdf/rtf/etc. also use the PDF sibling. | `--event-id`, `--attachment-id`, `--include-metadata` | `GET /me/events/{event-id}/attachments/{attachment-id}` |
+| `convert-calendar-event-attachment-to-markdown` | Convert an attachment on an Outlook calendar event to markdown. Polymorphic on the attachment’s `@odata.type` (shares the mail-attachment pipeline): fileAttachment decodes the inline bytes and runs them through the local conversion pipeline (docx via mammoth, xlsx via sheetjs, csv as markdown table, odt/ods/odp via content.xml, plus plain-text passthrough); referenceAttachment resolves via /shares/{token}/driveItem and routes through the same dispatcher; itemAttachment (embedded mail / event / contact) is rendered locally. For pptx decks attached to a meeting, `convert-calendar-event-attachment-to-pdf` preserves slide layout (a pptx here yields only its speaker notes / titles / comments via `## PPTX metadata` with `--include-metadata true`). For pdf/rtf/etc. also use the PDF sibling. | `--event-id`, `--attachment-id`, `--include-metadata`, `--keep-quoted` | `GET /me/events/{event-id}/attachments/{attachment-id}` |
 | `convert-calendar-event-attachment-to-pdf` | Convert an attachment on an Outlook calendar event to PDF on the fly (shares the mail-attachment pipeline). fileAttachment uploads the bytes to a temp folder under /me/drive, runs Graph `?format=pdf`, then deletes the temp item; referenceAttachment resolves via /shares/{token}/driveItem and converts in place; plain-text and `pdf` sources short-circuit to a raw-bytes envelope (Graph’s `?format=pdf` does not accept `pdf` as an input). image attachments are rejected (Graph rejects image inputs); itemAttachment (embedded mail/event/contact) is unsupported — use convert-calendar-event-attachment-to-markdown. Best for the deck attached to a meeting, where PDF preserves slide layout for a vision-capable LLM. | `--event-id`, `--attachment-id` | `GET /me/events/{event-id}/attachments/{attachment-id}` |
 | `get-calendar-event` | Fetch a single calendar event by ID from the signed-in user’s default calendar. Pass `--select` to project only the fields you need (the full event body can be large with HTML body and attendee lists). | `--event-id`, `--select`, `--expand` | `GET /me/events/{event-id}` |
 | `get-my-calendar` | Return metadata for the signed-in user's *primary* calendar — `id`, `name`, `color`, `owner`, `canShare`, `canViewPrivateItems`, `canEdit`, `defaultOnlineMeetingProvider`. Sibling to `list-calendars` which returns every calendar (incl. shared / subscribed). Use `--select` to fetch only the fields you need. | `--select`, `--expand` | `GET /me/calendar` |
@@ -252,11 +257,11 @@ For everything else:
 
 | Command | Description | Required params | Graph endpoint |
 |---------|-------------|-----------------|----------------|
-| `convert-local-file` | Convert a file ON DISK to markdown — the only command that never calls Microsoft Graph (works offline, no login). Runs the same local pipelines as `download-drive-item-as-markdown`: docx (mammoth → turndown), xlsx (sheetjs tables, `--max-cells` OOM cap), pptx (per-slide text), odt/ods/odp, csv, pdf (text layer via unpdf), legacy OLE .xls / .doc, Outlook .msg (headers + body, attachments converted recursively), plain-text passthrough — and a `.zip` is unpacked with every contained file converted in one call (legacy GBK / CP437 entry names decoded, not mojibaked). What it canNOT do locally: convert TO pdf, and Loop/Fluid/Whiteboard sources — both need a Graph server round-trip (upload to OneDrive and use the drive-item siblings). Pass `--include-metadata true` for the Office side-channel metadata blocks; `--inline-images true` to embed docx images as base64 data URIs. | `--path`, `--include-metadata`, `--inline-images`, `--include-images`, `--max-cells` | `GET (local) reads {path} from the local filesystem; not a Graph endpoint` |
-| `extract-local-file-images` | Extract the embedded images from a file ON DISK — the local sibling of `extract-drive-item-images`, and like `convert-local-file` it never calls Microsoft Graph (works offline, no login). Same per-extension dispatch: docx / xlsx / pptx (and their macro-enabled / template variants) have their OOXML media parts read directly (png/jpg/gif/bmp/tiff/webp/svg — full-resolution originals, including images on hidden slides); a pdf is walked page by page via unpdf with each painted image re-encoded as PNG. Two flows only this command completes: a Graph-rendered PDF saved locally (legacy `.ppt` → `download-drive-item-as-pdf` with the global output-path flag → this command pulls the slide images for OCR), and Office files unpacked from a local archive. Pair with the global output-dir flag to write every image to a folder; otherwise the bytes ride back base64-encoded. Any other extension returns a 415 naming the local ways out. | `--path` | `GET (local) reads {path} from the local filesystem; not a Graph endpoint` |
+| `convert-local-file-to-markdown` | Convert a file ON DISK to markdown — the only command that never calls Microsoft Graph (works offline, no login). Runs the same local pipelines as `download-drive-item-as-markdown`: docx (mammoth → turndown), xlsx (sheetjs tables, `--max-cells` OOM cap), pptx (per-slide text), odt/ods/odp, csv, pdf (text layer via unpdf), legacy OLE .xls / .doc, Outlook .msg (headers + body with the quoted reply chain stripped — `--keep-quoted true` restores it — and inline `cid:` images shown as placeholders, attachments converted recursively), plain-text passthrough — and a `.zip` is unpacked with every contained file converted in one call (legacy GBK / CP437 entry names decoded, not mojibaked). What it canNOT do locally: convert TO pdf, and Loop/Fluid/Whiteboard sources — both need a Graph server round-trip (upload to OneDrive and use the drive-item siblings). Pass `--include-metadata true` for the Office side-channel metadata blocks; `--inline-images true` to embed docx images as base64 data URIs. | `--path`, `--include-metadata`, `--keep-quoted`, `--inline-images`, `--include-images`, `--max-cells` | `GET (local) reads {path} from the local filesystem; not a Graph endpoint` |
+| `extract-local-file-images` | Extract the embedded images from a file ON DISK — the local sibling of `extract-drive-item-images`, and like `convert-local-file-to-markdown` it never calls Microsoft Graph (works offline, no login). Same per-extension dispatch: docx / xlsx / pptx (and their macro-enabled / template variants) have their OOXML media parts read directly (png/jpg/gif/bmp/tiff/webp/svg — full-resolution originals, including images on hidden slides); a pdf is walked page by page via unpdf with each painted image re-encoded as PNG. Two flows only this command completes: a Graph-rendered PDF saved locally (legacy `.ppt` → `download-drive-item-as-pdf` with the global output-path flag → this command pulls the slide images for OCR), and Office files unpacked from a local archive. Pair with the global output-dir flag to write every image to a folder; otherwise the bytes ride back base64-encoded. Any other extension returns a 415 naming the local ways out. | `--path` | `GET (local) reads {path} from the local filesystem; not a Graph endpoint` |
 | `microsoft-search-query` | Run a federated KQL search across the signed-in user's mail, files, list items, sites, calendar events, and people. Microsoft Graph v1.0 rejects multi-entity search bodies on most tenants (`Multiple entity search is not supported in v1.0`), so this command issues SIX parallel POSTs — one per entityType — and merges the per-entity `searchHits` containers into a single `value[]`. Each container is identifiable by the resource type inside `hits[].resource`. If a sub-request fails (e.g. tenant lacks the scope for one entity), the others still return; failures show up in `partialErrors[]`. Page size is fixed at 25 per sub-request and `top` is NOT exposed (Graph rejects $top in /search/query bodies). `chatMessage` is excluded since `Chat.Read*` is unavailable. To find Microsoft Loop pages (`.loop`) for markdown conversion, query `filetype:loop`: each `driveItem` hit carries `resource.id` plus `resource.parentReference.driveId`, the exact pair `download-drive-item-as-markdown` needs to render the page via Graph `?format=html`. (`filetype:fluid` returns nothing on this corpus; Loop pages index as `.loop`.) | `--query` | `POST /search/query` |
 | `my-quick-context` | One-shot discovery for the IDs every other command needs, plus the user's job title and tenant timezone / locale / working-hours. Issues 9 Graph calls in parallel and returns what each succeeded for. Partial-result mode: only `/me` is load-bearing — if any other sub-call fails (missing license, scope, or tenant policy) the corresponding field is `undefined` but the rest are still returned. Replaces the audit's 5-call discovery chain — feed the IDs straight into `list-mail-folder-messages`, `list-folder-files`, `list-planner-tasks`, `list-onenote-notebook-sections`, etc. For Microsoft To Do lists call `list-todo-task-lists` on demand (intentionally dropped from this command's fan-out — the array of {id, displayName, wellknownListName} entries crowded the envelope with IDs an LLM rarely needs on first contact). `tenantTimeZone` lets an LLM stop treating every datetime as UTC on first contact. | _(none)_ | `GET (meta) parallel: /me, /me/drive, /me/mailFolders/inbox, /me/calendar, /me/planner/plans, /me/onenote/notebooks, /me/joinedTeams, /me/drive/recent, /me/mailboxSettings` |
-| `next-page` | Fetch the next page of a paginated Graph response. Pass the cursor the previous command emitted — in text mode that is the `next: <url>` value in the `---` footer; in JSON mode it is the top-level `nextLink` field. Never reach into `data["@odata.nextLink"]`; the CLI strips that and surfaces it as a first-class envelope/footer field. Automatically signs `/me/chats` and `/chats/...` cursors with the M365ChatClient elevated token to match the chat-metadata commands. | `--url` | `GET {url}` |
-| `scopes-check` | Decode the cached Teams web client access token and return its scopes, audience, and expiry without making a Graph call. Use this as a self-test before running a command an LLM expects to fail with `accessDenied` — if the required scope isn't in the returned list, the call will reject regardless of tenant config. Each command's `scopesRequired` field in `help-json` lists the scopes that command needs; intersect with the array returned here for a pre-flight check (pipe both through `jq` and diff). The `expiresInSeconds` field lets an LLM decide pre-emptively to `login` again — typically worth doing under ~5 minutes (300 s) so a long-running session doesn't hit the wall mid-command. The `elevated` block reports whether the *separate* M365ChatClient-elevated token (needed by the historical-version download / convert commands) is cached and still usable — so a fresh process can pre-flight `deep-scan`-style workloads instead of discovering a 403 mid-run; `available:false` when it is absent, expired, or within the same 5-minute buffer the download path applies. The `chatsvcagg` and `ic3` blocks report the two Teams-chat substrate tokens (used by `list-teams-chat*` / `find-chats-with-user`) the same way; both self-heal from the shared refresh token, so they are informational rather than a preflight gate. | _(none)_ | `GET (meta) cached-token introspection — no Graph endpoint` |
+| `next-page` | Fetch the next page of a paginated Graph response. Pass the cursor the previous command emitted — in text mode the `---` footer prints the whole ready-to-run command (`next: ask-marcel-office next-page --url '<url>'`), so copy the line as-is (the URL is single-quoted because it contains `$`); in JSON mode use the top-level `nextLink` field. Never reach into `data["@odata.nextLink"]`; the CLI strips that and surfaces it as a first-class envelope/footer field. Automatically signs `/me/chats` and `/chats/...` cursors with the M365ChatClient elevated token to match the chat-metadata commands. | `--url` | `GET {url}` |
+| `scopes-check` | Decode the cached Teams web client access token and return its scopes, audience, and expiry without making a Graph call. Use this as a self-test before running a command an LLM expects to fail with `accessDenied` — if the required scope isn't in the returned list, the call will reject regardless of tenant config. Each command's `scopesRequired` field in `help-json` lists the scopes that command needs; intersect with the array returned here for a pre-flight check (pipe both through `jq` and diff). The `expiresInSeconds` field lets an LLM decide pre-emptively to `login` again — typically worth doing under ~5 minutes (300 s) so a long-running session doesn't hit the wall mid-command. The `elevated` block reports whether the *separate* M365ChatClient-elevated token (needed by the historical-version download / convert commands) is cached and still usable — so a fresh process can pre-flight `deep-scan`-style workloads instead of discovering a 403 mid-run; `available:false` when it is absent, expired, or within the same 5-minute buffer the download path applies. The `chatsvcagg` and `ic3` blocks report the two Teams-chat substrate tokens (used by `list-teams-chat*` / `find-chats-with-user`) the same way; both self-heal from the shared refresh token, so they are informational rather than a preflight gate. Every tier block also lists that token's OWN granted scopes (decoded from its `scp` claim, distinct per token) and its `refresh` route (`automatic` = self-heals from the shared refresh token; `interactive` = the elevated tier, which carries no refresh token of its own and is re-captured by a browser login); the `hint` field says how to refresh them. | _(none)_ | `GET (meta) cached-token introspection — no Graph endpoint` |
 
 <!-- AUTO-GENERATED-COMMANDS:END -->

package/docs/USAGE.md

@@ -18,7 +18,7 @@ bun add -g ask-marcel-office-cli
 ask-marcel-office login
 
 # the rest is discoverable
-ask-marcel-office --help                                # ~34 KB, one-sentence summaries
+ask-marcel-office --help                                # ~38 KB, one-sentence summaries
 ask-marcel-office help-json --terse --category mail     # ~6 KB JSON for one category
 ask-marcel-office docs list-mail-messages               # full per-command Markdown
 ```
@@ -174,6 +174,53 @@ Every paginated command advertises this in three places: `ask-marcel-office <cmd
 
 `ask-marcel-office my-quick-context` returns `{ user, primaryDriveId, inboxId, todoLists, primaryCalendarId }` in a single round trip — five Graph calls in parallel. Use it as the first call in any LLM session that needs per-user IDs to feed into other commands.
 
+## MCP server (`ask-marcel-office mcp`)
+
+Serves every command to an MCP client over stdio. Register it:
+
+```bash
+claude mcp add --transport stdio --scope user ask-marcel-office -- ask-marcel-office mcp
+# from a dev clone:
+claude mcp add --transport stdio --scope user ask-marcel-office -- bun <repo>/src/main.ts mcp
+```
+
+Five gateway tools, not one per command (184 schemas per session is the bloat this CLI exists to
+avoid). Discovery is three hops:
+
+```
+list-commands { category?: string }            → terse manifest, start here
+get-command-docs { command: string }           → full docs for one command
+run-command { command, params?, outputPath?, outputDir? }        → the 180 READ commands
+run-write-command { command, params?, outputPath?, outputDir? }  → the 4 mail-draft WRITE commands
+login { force?: boolean }                      → sign in / refresh
+```
+
+`params` are the command's flags **without** the `--` prefix, keyed camelCase:
+
+```jsonc
+// CLI:  ask-marcel-office list-mail-messages --top 10
+run-command { "command": "list-mail-messages", "params": { "top": "10" } }
+```
+
+Notes:
+
+- `run-command` is annotated `readOnlyHint: true` so clients can auto-approve it. A write routed
+  through it is refused before it executes. `run-write-command` carries the 4 draft commands and is
+  marked non-destructive: each produces an UNSENT draft, and this CLI cannot send mail.
+- **Raise your client's tool timeout to ~5 minutes** (`MCP_TOOL_TIMEOUT=300000`, or the equivalent).
+  Measured live: a browser sign-in takes **37–64 s with no MFA prompt at all** (it varies with how
+  warm the persistent browser profile is), and the MCP default request timeout is **60 s** — so
+  `login` times out intermittently at the default, right on the boundary. The server keeps running
+  through a client-side timeout, so the sign-in has usually completed anyway: re-run your original
+  command, or check `scopes-check`, before calling `login` a second time.
+- **Log in from a terminal first** (`ask-marcel-office login`). A first-time MFA prompt adds minutes on top
+  of the above. After that the `login` tool covers the hourly elevated-token refresh.
+- `logout` and `update` are deliberately CLI-only.
+- Results are text (the same YAML-ish rendering the CLI prints, `hint:` / `source:` remedies
+  included). Two known v1 warts: paginated results print `next: ask-marcel-office next-page --url '...'`,
+  which maps to `run-command { "command": "next-page", "params": { "url": "..." } }`; and
+  `--output-path` rejections name the CLI flag rather than the `outputPath` param.
+
 ## Library API
 
 The package exports a typed library API for embedding inside your own CLI, agent, or MCP server.
@@ -194,11 +241,11 @@ const graph = createGraphClient({
 const me = await commands['get-current-user'].execute(graph, {});
 ```
 
-One special case: `convert-local-file` and `extract-local-file-images` read the **local filesystem**, not Graph — their registry-typed `execute` returns a redirect error, and the real entry point is the optional `executeLocal(fs, params)` on the same command object (the CLI wires this automatically; library consumers pass their own `FileSystem`):
+One special case: `convert-local-file-to-markdown` and `extract-local-file-images` read the **local filesystem**, not Graph — their registry-typed `execute` returns a redirect error, and the real entry point is the optional `executeLocal(fs, params)` on the same command object (the CLI wires this automatically; library consumers pass their own `FileSystem`):
 
 ```ts
 import { commands, createNodeFileSystem } from 'ask-marcel-office-cli';
-const md = await commands['convert-local-file'].executeLocal?.(createNodeFileSystem(), { path: './report.docx' });
+const md = await commands['convert-local-file-to-markdown'].executeLocal?.(createNodeFileSystem(), { path: './report.docx' });
 const images = await commands['extract-local-file-images'].executeLocal?.(createNodeFileSystem(), { path: './deck.pdf' });
 ```
 
@@ -216,10 +263,21 @@ src/
   domain/          — Result<T,E>, branded value-object types (AccessToken, EnvVar), JWT utilities, format-error
   infra/           — Auth recovery ladder (cache → refresh → Playwright browser), Graph API HTTP client, Winston logger
   use-cases/       — Commands (schemas + execute functions), ports
-  composition/     — CLI wiring (Commander), dependency graph
-  presenter/       — Output formatting (text YAML-ish default + JSON envelope opt-in)
+  composition/     — CLI wiring (Commander), MCP gateway, the shared command executor, dependency graph
+  presenter/       — Output formatting: pure renderers (render-to-string) + the CLI's stdout shim (output)
 ```
 
+Two front ends, one execution path. `composition/cli.ts` (Commander) and `composition/mcp.ts` (MCP
+stdio) both call `composition/run-registry-command.ts`, which owns everything between "resolved
+command + params" and "value or failure": option-alias normalization, local-filesystem routing,
+error-source classification, and `--output-path` / `--output-dir` persistence. Adding a fix there
+reaches both surfaces.
+
+The presenter split exists for the same reason. `render-to-string.ts` formats and returns a string;
+`output.ts` is the only sanctioned `process.stdout` writer. The MCP server cannot write to stdout —
+that stream carries its JSON-RPC frames — so it renders through the pure module and returns the text
+as tool content.
+
 - **Auth**: Three-rung recovery ladder — file-based cached JWT → OAuth refresh_token exchange → Playwright browser intercepting Teams login
 - **Client ID**: `5e3ce6c0-2b1f-4285-8d4b-75ee78787346` (Teams Web)
 - **Scopes**: `https://graph.microsoft.com/.default openid profile offline_access`
@@ -231,7 +289,7 @@ src/
 
 `download-drive-item-version --format <original|pdf|markdown>` needs a Graph token whose `appid` is on Microsoft's ODSP allow-list — the Teams web client token returns 403 with `logicalPermissionAccessDenied` against historical-version bytes.
 
-Login captures a *second* Graph token from `https://m365.cloud.microsoft/search` whose first-party identity is M365ChatClient (`c0ab8ce9-e9a0-42e7-b064-33d422df41f1`) — an app on the ODSP allow-list. It is stored alongside the Teams token (`elevated_access_token` / `elevated_expires_on` fields in the cache) and used only by the historical-version command. Refresh path is re-capture via a brief Edge launch — the persistent profile cookies do silent SSO when fresh; if the federated IdP session has lapsed, interactive sign-in completes inside the popup. If the elevated capture fails at login, every other command (including `list-chats` / `get-chat`, which use the regular Teams token) still works. Because the elevated token carries no refresh token of its own, a cache-hit `login` does not renew it; run `ask-marcel-office login --force` to re-capture every token (basic + elevated + the chatsvcagg / ic3 substrate tokens) in one browser pass. Both `login` and `scopes-check` report all four tokens' `{ available, expiresInSeconds? }` status so you can see which one is about to lapse before a command hits a 403.
+Login captures a *second* Graph token from `https://m365.cloud.microsoft/search` whose first-party identity is M365ChatClient (`c0ab8ce9-e9a0-42e7-b064-33d422df41f1`) — an app on the ODSP allow-list. It is stored alongside the Teams token (`elevated_access_token` / `elevated_expires_on` fields in the cache) and used only by the historical-version command. Refresh path is re-capture via a brief Edge launch — the persistent profile cookies do silent SSO when fresh; if the federated IdP session has lapsed, interactive sign-in completes inside the popup. If the elevated capture fails at login, every other command (including `list-chats` / `get-chat`, which use the regular Teams token) still works. Because the elevated token carries no refresh token of its own, a cache-hit `login` does not renew it; run `ask-marcel-office login --force` to re-capture every token (basic + elevated + the chatsvcagg / ic3 substrate tokens) in one browser pass. `login` itself prints only a slim availability summary; **`scopes-check`** is the side-effect-free detailed view — per token it reports availability, seconds-to-expiry, refresh route, and that token's own granted scopes (decoded from its `scp`; the four sets are distinct) — so you can see which token is about to lapse, or lacks a required scope, before a command hits a 403.
 
 ## Configuration
 
@@ -248,7 +306,7 @@ Environment variables read at composition time:
 ## Quality gates (atelier four-check loop)
 
 ```bash
-bun test           # full suite (3700+ tests)
+bun test           # full suite (4800+ tests)
 bun run lint       # ESLint (0 warnings, 0 errors)
 bun run typecheck  # tsc --noEmit
 bun run coverage   # per-tier gates (100% on every tier: domain, use-cases, infra, composition, presenter)

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "ask-marcel-office-cli",
-  "version": "2.1.0",
+  "version": "2.2.0",
   "description": "Microsoft Graph CLI + library \u2014 typed Bun/TypeScript wrapper around 150+ Graph operations (read + on-the-fly PDF/markdown conversion + federated Microsoft Search) reachable from a Teams browser-OAuth token.",
   "license": "MIT",
   "author": "Vincent Delacourt <vincent.delacourt@adama-development.com>",
@@ -87,6 +87,7 @@
   },
   "dependencies": {
     "@kenjiuno/msgreader": "^1.28.0",
+    "@modelcontextprotocol/sdk": "^1.29.0",
     "commander": "^14.0.3",
     "fast-xml-parser": "^5.8.0",
     "jszip": "^3.10.1",