| name | thinkroom |
| description | Collaborate with people through Thinkroom documents using the Thinkroom CLI. Use when an agent needs to start or share a document, hand substantial thinking to a human, inspect or revise an existing Thinkroom share link, propose a targeted suggestion, leave a comment, or load a repository's durable agent context with thinkroom prime. |
Thinkroom
Use Thinkroom as the durable handoff surface for substantial work that benefits from human reading, judgment, suggestions, or endorsement. Keep local files authoritative while drafting; publish or revise the Thinkroom document at clear handoff points.
Prime the repository
Run this before substantial work in a repository:
thinkroom prime
Read the relevant paths it reports, especially AGENTS.md, CONCEPTS.md, and matching entries under docs/solutions/. Priming is offline and must not block work when the CLI is not logged in.
Start a document
Write the complete draft to a file or pipe it on standard input. Identify the actual agent name so provenance is useful:
thinkroom new draft.md --title "Decision memo" --agent "Codex"
Always pass --agent (or set THINKROOM_AGENT) on writes — it is required. Without it the CLI stops with an error instead of misattributing the write to a generic identity, because per-agent provenance is the point.
Use --format html only when semantic HTML is genuinely needed. Return the printed share URL to the user.
If the CLI asks for authentication, run thinkroom login and let the person approve the browser prompt. Never ask them to paste a token into chat.
Work with an existing document
Read live state before changing anything:
thinkroom show SHARE_URL --json
The content field is the canonical Markdown source and may carry
<span data-provenance data-kind="ai" data-author="…" data-state="pending">
attribution marks; plain_markdown is the same text with marks stripped.
When you update, the server stores your content verbatim: round-tripping
content preserves everyone's existing attribution, but text you add inside
someone else's span inherits their name. Wrap your new or changed passages in
your own span (data-author = your --agent name) so provenance stays
truthful; unwrapped text reads as unattributed. Never build an update from
plain_markdown — that erases all existing attribution.
For a document that is still an untouched seed, revise it in place:
thinkroom update SHARE_URL revision.md --title "Updated title" --agent "Codex"
If you are logged in (thinkroom login) and own the document, you can update it in place this way even after it is claimed or after a live editing session exists. Owner updates are full replacements of the document source at the same share URL. Replacing a live document auto-rejects any pending suggestions whose targeted text no longer exists in the new content — thinkroom update warns when this happens; check thinkroom show --json afterward if you or others have pending suggestions on that document.
If you do not own the document, do not try to overwrite a claimed or live document from the CLI. Propose the smallest exact replacement and include intent:
thinkroom suggest SHARE_URL \
--replaces "Exact current text" \
--body "Proposed replacement" \
--intent "Why this improves the document" \
--agent "Codex"
Use comments for questions or review notes that should not directly replace prose:
thinkroom comment SHARE_URL --body "Could we verify this assumption?" --agent "Codex"
When an API conflict provides a next action, follow that guidance instead of retrying the same write.
Handoff
Finish by giving the person the share URL and one sentence describing what judgment or action is needed. Do not expose CLI config files, bearer tokens, browser cookies, or raw API responses containing credentials.
