Skip to main content
Each example is a request you can type to a connected agent, followed by the tools it runs. Names, ids and dates are fictional. An agent with the Sidecars skills installed runs these the careful way by default: read before write, exact excerpts, human edits kept.

​
Where the ids come from

list_folders turns a folder name into the folder_id the other tools take: { "query": "launch" } finds the folder, with its path so two folders with one name can be told apart. Opening it ({ "folder_id": "<id>" }) shows its subfolders and sessions. Ids also arrive on list_documents rows; the full table is in IDs for other things.

​
Review everything in a folder

Go through the Launch folder and tell me what’s out of date.
  1. list_folders with query: "Launch", then again with the matching folder_id to see its subfolders, sessions and document count.
  2. list_documents with that folder_id, following next_cursor until it is null.
  3. read_document each one, and add_comment where something is stale instead of editing.
Subfolders are not expanded: open each with its own folder_id.

​
Summarise what we know about Acme into a doc in the launch folder

Pull together what we have on the Acme sandbox timeline and write it up as a document in the launch folder.
  1. list_folders with query: "launch" to resolve the folder’s folder_id.
  2. ask_knowledge for a sourced answer, then search_knowledge for the passages worth quoting.
  3. list_documents with folder_id to check nothing similar exists.
  4. create_document with folder_id, content_markdown and an idempotency_key. The folder’s members can reach it straight away.
Watch for: the document_id values the knowledge tools return identify ingested sources, not necessarily Sidecars documents. Pass only ids from list_documents or a create to the document tools. To write the doc into a session instead, pass the session’s session_id (the last part of its URL, /sessions/<id>) in place of folder_id: the card is posted to the session, and everyone who can open it can reach the document.

​
Update the Risks section of the Q3 plan

In the Q3 launch plan, update the Risks section: the vendor SDK now has a GA date of 1 October, and add the Northwind contract as a risk.
  1. read_document; keep version.
  2. edit_document with mode: "replace_section", a section_excerpt quoted exactly from the read, and expected_version.
  3. Read the result: kept_blocks means someone edited nearby while the agent wrote, and their text was kept.
Watch for: a stale excerpt is refused with excerpt_not_found and the nearest candidates. The agent re-reads and quotes one exactly; a failed replace is never turned into an append.

​
Review this doc and leave comments, don’t edit it

Review the Q3 launch plan and leave comments where the dates don’t line up with the kickoff notes. Don’t change the text.
  1. read_document and read_comments. The role in the comments result says whether commenting is allowed at all.
  2. add_comment with an anchor_excerpt for each finding, so the thread attaches to the passage.
  3. resolve_comment later, once a thread is addressed.
Watch for: a quote that appears more than once, or not at all, still creates the thread, unanchored. Check anchored in the result.

​
Rebuild the rollout page with December GA

The rollout timeline page still says GA in November. Change it to 2 December and keep everything else.
  1. read_page; keep version.
  2. write_page with the full new HTML and expected_version.
Watch for: if someone changed the page after the read, the write is refused with version_conflict, the current version and a line diff. The agent merges onto the current page and writes again. There is no force flag.

​
Export the plan as a PDF for the board

Give me the Q3 launch plan as a PDF.
  1. export_document with format: "pdf".
  2. Fetch download_url within 15 minutes.
Watch for: the file captures source_version. After further edits, export again; an unchanged version is served from cache.

​
Put the doc I just copied into context

Here’s a doc from Sidecars. Use the latest saved version if you can reach it. (pasted: a Copy for agent handoff)
  1. Read the Artifact ID line from the handoff.
  2. read_document with that id for the latest saved content and version.
  3. If the handoff carries the UNSAVED CHANGES block, the snapshot holds edits the server never saw, and the server copy may have moved on too. Merge the snapshot’s edits onto the live version and write with the live version; on version_conflict, read and merge again.
Watch for: without an MCP connection the agent works from the snapshot and should say the original was not updated. The handoff format is described in IDs and URLs.