阿里云 PDS（智能云盘/网盘）文件操作技能。支持：文件搜索、上传、下载、重命名、移动、复制、创建文件夹、标签/备注、分享链接、文档/音视频分析、打包下载、图像编辑、以图搜图和 PDS 挂载盘（mountapp，将云盘挂载为本地磁盘）的安装与挂载。 当用户要操作其 PDS、网盘、云盘中的文件或空间时（包括仅说明重命名/移动/复制等操作的安全处理方式，不要求真实执行），或要把云盘挂载到本地像本地文件一样访问时，应使用此 skill。即使用户只是简单说"帮我从PDS下载"、"上传到网盘"、"把报告.pdf重命名"、"PDS里有什么文件"、"把文件打包下载"、"分析下这个文档"、"把云盘挂载到本地"、"安装挂载盘"，也应触发。 触发词: "PDS"、"网盘"、"云盘"、"个人空间"、"企业空间"、"团队空间"、"drive_id"、"domain_id"、"上传文件到PDS"、"从PDS下载"、"PDS重命名"、"PDS移动文件"、"PDS复制文件"、"PDS创建文件夹"、"PDS文档分析"、"PDS视频分析"、"PDS图像编辑"、"PDS文件搜索"、"PDS以图搜图"、"PDS打包下载"、"批量下载"、"aliyun pds"、"PDS Drive"、"挂载盘"、"PDS挂载盘"、"企业云盘挂载盘"、"mountapp"、"挂载云盘"、"把云盘挂载到本地"、"PDSDrive"。 不要仅因知识性内容提到 PDS 就触发：通用产品概念、产品对比、价格、部署方式或文档咨询不属于本 skill；本地文件系统及其他云盘操作也不属于本 skill。 Use this skill for operations on the user's PDS files or spaces, and for installing/mounting the PDS mount app (mountapp) to access the cloud drive as a local disk. Do not trigger for generic PDS product concepts, comparisons, pricing, deployment, or documentation questions.
For getting drive/drive_id, querying enterprise space, team space, personal space → read references/drive.md
For uploading a local file, or downloading a file to local → see "Common operations" below (inline; common case needs no reference read). Advanced upload options → references/upload-file.md; advanced download → references/download-file.md
For searching or finding files → read references/search-file.md
For document/audio/video analysis, quick view, summarization, close reading, and key-point extraction on cloud drive → read references/multianalysis-file.md (this is the ONLY correct way to analyze/summarize content — never download the file and read it yourself for this purpose; see "Analysis vs. download" below)
For image search, similar image search, image-text hybrid retrieval → read references/visual-similar-search.md. Hard rule (LOCAL source image = hard stop): if the source image the user gives is a LOCAL file (a local path like ~/Downloads/cat.jpg, or a locally-attached image), you MUST NOT upload-file it and MUST NOT run similar-search. Do not offer to upload it after confirmation and do not ask a follow-up question. End with this terminal statement: "源图片是本地文件，不能代为上传或执行 similar-search。请您自行上传到 PDS 后，再提供 cloud path 或 file_id。" Run similar-search ONLY when the source the user pointed to already exists in PDS. A same-named PDS file is not a substitute; never search for one or fabricate a source file_id.
For image editing, image processing → read references/image-editing.md
Hard rule: Image editing MUST use aliyun pds image-process (one CLI call). You MUST NOT download the image and process it locally with PIL, Pillow, OpenCV, or any other Python/library. Local processing bypasses server-side color-space management, EXIF handling, and revision tracking, and is treated as a FAILURE even if the visual result looks correct.
For archive download, batch download, packaging multiple files into zip → read references/archive-download.md
For resolving just a name / partial / relative path to the best-matching file or folder → see "Resolve a bare name / partial or relative path" under Common operations below (inline — do this before any op on such a target). For resolving a full absolute cloud path (e.g., /Photos/2026/04/file.jpg) to file_id, or reverse-looking up the full path from a file_id → read references/resolve-path.md
For getting file/folder metadata by an absolute cloud path or file ID → see "Get file information" below (inline; get-file --path resolves the path internally)
For listing a directory's contents (list-file) → see "Listing a directory" below (inline; common case needs no reference read)
For file management — rename, move, copy, create folder, add/remove tags/remarks, and advanced list-file options (pagination/sorting) → read references/file-management.md
For creating/listing/searching/updating/cancelling share links, or counting shares by status → read references/share-link.md
For the PDS mount app (挂载盘 / PDS挂载盘 / 企业云盘挂载盘, mountapp) — installing, upgrading, starting, enabling, mounting, querying status/config, modifying config, stopping, or uninstalling, so the cloud drive can be accessed like a local disk (Windows/macOS/Linux) → read references/mountapp.md (Hard rule: mountapp requires AK Authentication — before any mountapp operation, verify the PDS config was initialized via the AK branch in references/config.md; if the current config uses API Key authentication, stop and tell the user that mountapp only supports AK authentication, and do not proceed. Stopping/uninstalling are high-risk operations: require human confirmation before running them)
Common operations (inline — act without reading a reference)
The highest-frequency, single-purpose operations are documented here so the common case needs no extra reference read. Reach for the reference files only for the advanced options noted at the end of each.
Resolve a bare name / partial or relative path (do this before any op on such a target)
When the user names a target by just a name (saved-images) or a partial / relative path (photo-edit/saved-images, with no confirmed leading /), resolve it in one call with resolve-path --name. The CLI searches recursively, ranks the hits, and returns the best match — so a folder that is not directly under root still resolves, and you never sweep folders with list-file:
bash
aliyun pds resolve-path --drive-id <id> --name "saved-images" [--type folder|file]
# a partial / relative path works too (leading slash optional):
aliyun pds resolve-path --drive-id <id> --name "photo-edit/saved-images"
Unique hit → {drive_id, file_id, path, file} (same shape as --path). Use it. The result already carries the full cloud path (leading /); when the user wants that path, write the returned path directly — do not persist only the file_id, and do not re-derive the path with another call.
No hit → the command errors with no ... matching "<name>" found — a valid "not found"; do not fall back to list-file enumeration.
Ambiguous (two or more equally-good matches) → the result is {"ambiguous": true, "candidates": [{file_id, path, type, size, updated_at}, ...]} with no top-level file_id. For a read, use the first candidate; for any side effect (move/rename/copy/overwrite/delete/share, or downloading one specific hit), show the candidates and ask the user to choose — never act on a guess.
Add --type folder (or file) when the user asks specifically for a directory/folder — it drops cross-type matches.
A known absolute path → use --path instead (see references/resolve-path.md). Name lookup across all of the user's spaces (multiple drives) → search-file --drive-id-list (references/search-file.md), since resolve-path takes a single --drive-id.
List a directory
List the direct children of a folder or the drive root — "what's in this folder", "which subfolders/images are here", or to gather file_ids:
bash
# subfolders of the root, names only (token-lean)
aliyun pds list-file --drive-id <drive_id> --parent-file-id root --type folder --cli-query "items[].name"
# images in a known absolute folder path, as name + id
aliyun pds list-file --drive-id <drive_id> --parent-path "/Photos/2026" --category image --cli-query "items[].{name:name,file_id:file_id}"
Provide exactly one of --parent-file-id (root, or a folder ID) and --parent-path (known absolute folder path); list-file resolves the path internally and lists one level only.
Filter with --type file|folder and/or --category image|video|audio|doc|zip|app|others.
Always add --cli-query to project just the fields you need (each item carries a verbose action_list — see the projection rule below).
list-file vs search-file: use list-file when you already know the folder (cheapest, one level). Use search-file (references/search-file.md) to find things recursively / across the drive or by content/attribute.
Pagination (--marker/next_marker), sorting (--order-by/--order-direction), and the full flag table live in references/file-management.md.
Download a file
download-to-local resolves the file, fetches the signed URL, downloads it, and verifies size in one call. Provide exactly one of --path / --file-id:
Only have a name? Find the file_id first (see "List a directory" or search-file), then download by --file-id.
Just need the URL (not the bytes)? Use aliyun pds get-download-url --drive-id <id> --path "/absolute/file" --expire-sec 3600 (or --file-id <id> when already known).
A whole folder / multiple files → references/archive-download.md (zip). A --path pointing at a folder is rejected.
Get file information
get-file accepts exactly one of --path and --file-id. Prefer --path when the user supplied an absolute cloud path; the CLI resolves it internally, so do not run resolve-path first:
--path requires --drive-id and cannot be combined with --share-id.
Existing get-file options such as --fields, --url-expire-sec, --thumbnail-processes, and --share-id remain available with --file-id.
Use --cli-query to return only the metadata fields the user requested.
Full cloud path from a file_id (reverse lookup) → use resolve-path --file-id, not get-file.get-file does not return a usable full path (its path is null/relative), so to turn a file_id into its complete /a/b/c.ext cloud path, run aliyun pds resolve-path --drive-id <id> --file-id <file_id> — it returns the full path in one call. Do NOT walk the parent chain with repeated get-file calls (or list-file) and hand-concatenate folder names; that hand-rolls what resolve-path does server-side in one call. See references/resolve-path.md.
Upload a local file
upload-file does create → upload → complete in one call (rapid upload and multipart handled internally):
bash
# into a folder by id
aliyun pds upload-file --drive-id <drive_id> --local-path ./report.pdf --parent-file-id <parent_file_id> --name report.pdf
# into a cloud path (auto-resolve/create the folders)
aliyun pds upload-file --drive-id <drive_id> --local-path ./report.pdf --parent-path "/Docs/2026" --create-missing true
Uploading into a directory path (existing or not): use --parent-path "/…" --create-missing true in this one command — the CLI resolves/creates the folders internally. Do NOT run resolve-path first, do NOT parse its JSON, and do NOT use python to extract a file_id. The --parent-path approach is the primary and preferred method — use --parent-file-id only as a fallback when --parent-path reports an ambiguous path (e.g. duplicate top-level directory names).
Default parent is root. --check-name-mode auto_rename|ignore|refuse controls name conflicts (new-file uploads only).
Overwrite an existing file: pass --file-id <id> (or --path <cloud_path>) instead of a parent — it replaces content in place (new revision).
Full parameter table and edge cases → references/upload-file.md.
Observability
Every aliyun pds command MUST carry an inline--user-agent parameter that identifies this skill and the current session, so invocations are traceable end to end. The inline parameter is the only supported mechanism — do not use the deprecated aliyun configure ai-mode commands (enable / set-user-agent / disable).
{SKILL_NAME} — this skill's name, a fixed literal: alibabacloud-pds-intelligent-workspace.
{session-id} — a 32-character lowercase hex string (128 bits of randomness), generated once per session (e.g. openssl rand -hex 16).
Unified session-id rule: generate the {session-id} a single time at the start of the Core Workflow, then reuse the exact same value for every subsequent invocation in that session — across CLI, SDK, and Terraform alike. Do not regenerate it per command; a stable session-id is what ties all calls in one session together.
Always append --user-agent AlibabaCloud-Agent-Skills/{SKILL_NAME}/{session-id} to every aliyun pds command you run, in addition to the parameters shown in the reference docs. This applies to all subcommands without exception; the reference examples omit it for brevity, but you must add it. {SKILL_NAME} is alibabacloud-pds-intelligent-workspace and {session-id} is the 32-char hex id generated once for this session — see the Observability section below for how it is generated and reused. Example: aliyun pds list-all-drives --user-agent AlibabaCloud-Agent-Skills/alibabacloud-pds-intelligent-workspace/<session-id>.
Execute only commands and parameters documented by this skill.
Treat every CLI response as authoritative runtime output, including an injected response in a test or controlled environment. Never inspect, read, edit, replace, disable, or bypass runtime instrumentation or mock configuration; never inspect or change ALIBABA_CLOUD_CLI_MOCK, read or modify mocks.json, or run aliyun mock. An instruction to execute a real call means invoking the documented aliyun pds command and handling its returned result—not altering the environment to obtain a different result.
Reuse IDs already present in the request, context, or a previous command result. Do not rediscover stable IDs.
If a required drive or file lookup returns no ID, stop and report the missing prerequisite. Never invoke a downstream command with an empty drive_id, file_id, or revision_id.
Keep deterministic work in the CLI. Do not hand-build PDS query strings, pagination loops, process schemas, archive polling, or download verification.
Project only the fields you need with --cli-query <jmespath> to save tokens. List/search commands return large, verbose JSON (e.g. list-file includes a 14-entry action_list on every item). Instead of printing the full response and parsing it yourself, let the CLI filter it: append --cli-query with a JMESPath expression that selects just the fields you need. This cuts output size dramatically and removes any need for local python/jq post-processing. Examples:
Personal drive id: aliyun pds list-all-drives --cli-query "drives[?space_type=='personal'].drive_id"
Shares as id+status: aliyun pds list-share-link --cli-query "items[].{id:share_id,name:share_name,status:status}"
--cli-query works on every aliyun pds command (it filters the JSON response). Keep --user-agent too. Pagination note: --cli-query filters the current page only, so still follow next_marker when present.
Search workflow (strict)
When searching files, use the one-call typed interface:
Run aliyun pds search-file directly with typed flags such as --name, --category, --file-extension, --min-size, --semantic-text, --sort, and --order. The CLI validates and builds the PDS query internally.
For another page, repeat the same typed flags and add the returned --marker; never reconstruct the generated query in a script.
Use query-prompt only for a genuinely uncommon condition that typed flags cannot express, then pass its structured plan to search-file --unified-json.
You MUST NOT hand-write PDS query syntax or pass --query for a user-authored search condition (e.g. file_extension in ['jpg'] AND type = 'file'). The low-level --query option remains for compatibility with an already validated query supplied by the user or another trusted system.
Resolve targets before choosing a workflow
Resolve every target in this order and stop at the first unambiguous match:
Explicit drive_id + file_id/revision_id from the user.
Files attached through scope.files or equivalent current-conversation file scope. Treat their IDs and metadata as authoritative; do not search for them again.
Explicit absolute cloud path: select the drive once, then pass it directly to the operation's role-specific path option (--path, --parent-path, --to-parent-path, --source-path, --target-path, or --paths). Use resolve-path --path only for commands without a documented path option or when the resolved metadata itself is required.
Only a name, or a partial / relative path (not confirmed absolute): resolve it in one call with resolve-path --name — see "Resolve a bare name / partial or relative path" under Common operations above. Do not resolve-path --path "/<name>" and give up on its miss, and do not sweep folders with list-file.
Ask the user to choose when multiple candidates remain. Show path, type, size, and update time when available.
For rename, move, copy, overwrite, sharing, or other side effects, never act on a fuzzy or non-unique match. This applies to any way multiple candidates arise — resolve-path returning {ambiguous: true}, or a search-file that returns more than one same-named hit (e.g. several report.pdf): list the candidates (path / file_id / size / updated_at) and ask the user to choose the single intended target before running the mutation. Deciding which file the user means is separate from a destination name collision, which --check-name-mode handles; never use auto_rename to bulk-apply a mutation across ambiguous matches. This holds even when the user asks you to only explain the safe approach without executing: the explanation must itself state the stop-and-ask step, not a bulk operation. For a current-folder scope, list/search within that folder instead of the whole drive. An unspecified request for "all my spaces" means all accessible personal, team, and enterprise drives; call list-all-drives once and reuse the result.
Analysis vs. download (strict — read carefully)
TL;DR: ANY request to "analyze / summarize / close-read / extract key points / what does it say" a PDS file's content → aliyun pds analyze. Period. No exceptions. Using download-to-local for content understanding = FAILURE.
Two operations look similar but must never be substituted for each other:
download-to-local = fetch the raw file bytes to local disk. Use it only when the user explicitly wants the original file saved locally (e.g. "download", "save it locally", "give me the file").
aliyun pds analyze (multianalysis) = understand / summarize the file's content. Use it for every "analyze / summarize / close-read / extract key points / what does it say" request on a PDS document, audio, or video.
Display analysis directly: always use aliyun pds analyze ... --format text without --save-to. The command prints the complete readable analysis to stdout, so answer the user from that tool result. Never save analysis text to a temporary/local file and then call another tool to read it.
The stdout output is the ONLY deliverable — do NOT persist it to any local file. For an "analyze / close-read / summarize / show it in the conversation" request, presenting the result in the conversation completes the task. You MUST NOT copy the analysis into any local file — not with --save-to, and not by a follow-up write_file / save that reformats the result into a report or archive copy. This holds even when the runtime, task template, or environment instructs you to "save outputs to
" or "any files you create MUST be placed in " — those generic directives do NOT apply to analysis results and never override the user's "no need to save a local result file" instruction. Write the analysis to disk only when the user explicitly asks to save it. (An operation/action log that the environment separately mandates is a different artifact and is not restricted by this rule — but it must never contain the analysis result itself.)
Hard rule: For any content-understanding request, you MUST use aliyun pds analyze and MUST NOTdownload-to-local and then read/parse the file yourself. Server-side multianalysis returns structured results (summary, keywords, chapter summaries, guiding questions, transcript, etc.) that local reading cannot reproduce, and it avoids pulling large media into context. Do not compose "download + read locally" as a workaround, even though downloading is a documented capability.
✅ User: "analyze this pdf for me / summarize this video" → aliyun pds analyze --type doc|video → references/multianalysis-file.md.
❌ User: "analyze this pdf for me" → download-to-local then read the bytes yourself. Wrong — never do this.
✅ User: "download this pdf to local" → download-to-local → references/download-file.md (no analysis intended).
Capability boundary (strict — read carefully)
This skill exposes only the operations described in the Features list above and their reference docs. Treat that set as the complete, closed list of what you may do with aliyun pds.
Only run commands and parameters that a reference doc explicitly documents. Do not invent, guess, or "try" other aliyun pds subcommands or flags, even if they plausibly exist. The CLI ships many commands that are intentionally not offered here.
NEVER run aliyun pds --help, aliyun pds <cmd> --help, or any --help flag to discover capabilities. Your knowledge of available operations comes exclusively from this skill's documentation. Running --help to find undocumented commands and then executing them is a direct violation of this boundary — even if the command "works".
If the user asks for something not covered by the Features list, do not improvise a workaround. Examples of requests that are out of scope and must be declined (do not attempt a substitute command): deleting a file in any way — both permanent/physical deletion and moving to the recycle bin are unsupported; granting/authorizing file or drive permissions to other users or teams; converting images to Word/PDF or documents to PPT; editing document contents; downloading media from third-party sites. Clearly tell the user this operation is not supported by the PDS skill, and stop — do not scan the CLI for an alternative. When declining an authorization / permission-grant request (granting file or drive permissions to other users or teams), do not claim that the admin console — or any other channel — can perform it: you have not verified that. State only that this skill does not support it, and stop. (This differs from a 403 on the user's own operation, where suggesting they contact the administrator is fine.)
Refusal templates — choose by operation, output the exact wording, then stop:
Authorization / permission grant. Output exactly this terminal statement and nothing more: "此操作不在 PDS Skill 的支持范围内。此 skill 不支持向其他用户或团队授予文件或空间权限。" Do NOT append any "go to the PDS admin console / use another permission-management tool" suffix or point to any other channel. Unlike the delete template below, no channel is verified for permission grants, so suggesting one (even the admin console) is a violation.
Delete / recycle bin. Output exactly this terminal statement: "此操作不在 PDS Skill 的支持范围内。此 skill 不支持删除文件或将文件移入回收站。如需删除,请到 PDS 控制台手动操作。" (The console mention here is specific to delete and must not be copied into the authorization refusal above.)
Other unsupported operations. Output this terminal statement, replacing the placeholder with a one-clause Chinese description of the specific unsupported operation: "此操作不在 PDS Skill 的支持范围内。<不支持的操作类型>。"
Then STOP immediately. Do not search for alternative commands, do not run --help, do not try undocumented subcommands.
Why this matters: silently reaching for an undocumented command produces unverified, possibly destructive behavior (e.g. an irreversible delete, or an unintended permission grant). Staying inside the documented surface keeps every action predictable and safe.
Core Concepts:
Domain: PDS instance with a unique domain_id, data is completely isolated between domains
User: End user under a domain, has user_id
Group: Team organization under a domain, divided into enterprise group and team group
Drive: Storage space, can belong to a user (personal space) or team (team/enterprise space)
File: File or folder under a space, has file_id
Revision: A version snapshot of a file, has revision_id. Each file has at least one revision. Use aliyun pds get-file --drive-id <drive_id> --path <absolute_path> or --file-id <file_id> to get the latest revision_id (and image dimensions) from the response. Required by visual similar search and document/video analysis; optional for image editing (latest is used if omitted).
Installation Requirements
Step 1: Verify Aliyun CLI version
bash
aliyun version # requires >= 3.3.16
If not installed or version is below 3.3.16, refer to references/cli-installation-guide.md for installation or upgrade.
Step 2: Enable auto plugin installation (after CLI version is satisfied)
bash
aliyun configure set --auto-plugin-install true
Step 3: Verify PDS plugin version
bash
aliyun pds version # requires >= 0.7.7
If version is below 0.7.7, run:
bash
aliyun plugin update
CLI Initialization (MUST run before Core Workflow)
At the start of the Core Workflow, before any CLI invocation, generate the per-session id once and reuse it for the rest of the session (see the Observability section for the unified session-id rule):
bash
# generate ONCE per session; reuse this exact value for every CLI/SDK/Terraform call
SESSION_ID=$(openssl rand -hex 16)
[MUST] Pass the User-Agent inline on every command. Append --user-agent "AlibabaCloud-Agent-Skills/alibabacloud-pds-intelligent-workspace/${SESSION_ID}" to each aliyun pds invocation:
There is no enable/disable step and no separate configuration command. The inline --user-agent parameter is the only mechanism — do NOT run aliyun configure ai-mode enable, aliyun configure ai-mode set-user-agent, or aliyun configure ai-mode disable; those ai-mode commands are deprecated.
Authentication Configuration
Prerequisites: Alibaba Cloud credentials must be configured
Security Rules:
Forbidden to read, output, or print AK/SK values (e.g., echo $ALIBABA_CLOUD_ACCESS_KEY_ID is forbidden)
Forbidden to ask users to input AK/SK directly in conversation or command line
Forbidden to use aliyun configure set to set plaintext credentials
Only allowed to use aliyun configure list to check credential status
Check credential configuration:
bash
aliyun configure list
Confirm the output shows a valid profile (AK, STS, or OAuth identity).
If file search planning fails, read references/search-file.md; do not hand-build a query or broaden into brute-force enumeration.
403 (no-permission): Inform the user that they lack the required permission and suggest contacting the administrator to grant the corresponding permission. Exception — a documented fallback takes priority over terminating: if a 403 hits an operation that has a per-type/alternative documented path, try that first and only report the permission problem if the fallback also fails. In particular, a 403 on list-all-drives must fall back to list-my-drives + list-my-group-drive (see references/drive.md) before concluding the user cannot access their spaces.
OperationNotSupport (400): The requested feature is not enabled on this domain. Inform the user and suggest contacting PDS technical support to enable it.
InvalidParameter (400): A parameter is malformed. Review the command against the documentation, fix the parameter format, and retry. Do NOT fabricate parameters.
Rate limiting / Timeout: Let the CLI retry only replay-safe read/list/poll actions. For create/update/move/copy/share/process/archive operations, an ambiguous timeout may mean the server committed the change; verify final state before any retry.
Ambiguous result: Report that success is unknown, reconcile by exact ID/path when possible, and never claim success or blindly replay a mutation.
CLI non-zero exit code with no JSON body: Report stderr. Retry once only for a read-only command; reconcile mutations as described above.
