Commentary draft review

Commentary Draft Review

Use this skill to collaborate with a human reviewer on local text artifacts through Commentary draft reviews. Keep the local artifact as the source of truth unless the user explicitly asks to pull reviewed content back from Commentary.

Command Runner

Use the Commentary CLI by resolving a command runner in this order:

1. Use commentary when commentary --help succeeds.
2. Use the project-local binary when present:
   • POSIX: ./node_modules/.bin/commentary
   • Windows PowerShell: .\node_modules\.bin\commentary.cmd
3. Otherwise use npx -y @commentary-dev/cli.

In examples below, replace commentary with the resolved runner. Prefer --json when the output will be parsed by the agent.

Authentication

Check auth before creating or reading reviews:

commentary whoami --json

If authentication is missing or invalid, help the user authenticate without storing secrets in the repo:

commentary login --no-open

Use commentary login --token <token> only when the user provides a token directly for that purpose. For automation, prefer an environment variable:

COMMENTARY_TOKEN=<token> commentary whoami --json

Never write tokens into project files, .commentary/session.json, source files, docs, shell history snippets, or skill memory.

Alias And Memory

Give yourself a stable, human-readable alias before replying or resolving comments, for example Docs agent, Draft agent, or a name based on the artifact. If the platform supports memory, remember the alias for future Commentary sessions. If memory is unavailable, keep it in the current conversation and pass it explicitly:

commentary reply <thread-id> "Updated this in revision 2." --alias "Docs agent"
commentary resolve <thread-id> --message "Addressed in revision 2." --alias "Docs agent"

For repeated shell commands in one session, set COMMENTARY_AGENT_ALIAS instead of repeating --alias.

New Review Flow

When the user asks to create an artifact and review it with Commentary:

1. Create or update the local Markdown, MDX, HTML, or text artifact first.

2. If git is available and the workspace is inside a GitHub-backed git repository, default single-file reviews to a real Commentary GitHub base by using --git-base auto. This infers the GitHub owner/repo from the local origin remote, uses HEAD, and uses the repository-relative file path.

   commentary review ./docs/spec.md --title "Product spec" --git-base auto
   

   If auto-detection is unavailable but the user provides exact GitHub metadata, use --git-base-repo <owner/repo> and --git-base-sha <sha>. If no real GitHub base can be resolved, omit git base options and state that no real GitHub base was set. Do not fake this with description text. Do not create branches, commits, pull requests, or provider reviews.

3. Create the review:

   commentary review ./docs/spec.md --title "Product spec" --git-base auto
   

4. Share the review URL with the user if the CLI prints one or if commentary open is needed in a headless environment.

5. For ongoing sync, either start with --watch:

   commentary review ./docs/spec.md --title "Product spec" --git-base auto --watch
   

   or create the review once and run:

   commentary watch
   

Use --root, --include, and --exclude when reviewing a directory so the review contains only intended supported files.

If a reviewer asks for a new file to be included in an existing review, create the local file and add it to the same review:

commentary track ./docs/new-page.md --message "Add requested page"

Use track instead of creating a second review. It uploads a full revision containing the existing tracked files and the new files, then updates .commentary/session.json.

Sharing Reviews

When the user asks to share a review, use the CLI share command against the current local review session. If there is no local .commentary/session.json, use --session <id> when the user provides a session id; otherwise create or restore the review first.

For anyone-with-link sharing:

commentary share --anyone --json

For a specific user:

commentary share --user [email protected] --json

If the user asks to inspect current sharing state:

commentary share --list --json

Return the share URL to the user when the CLI response includes a shareLink.url or shareLink.shareUrl. For user-specific grants, return the recipient/access-grant confirmation and the draft review URL from local session metadata or commentary status --json if no share URL is present.

Do not write share links, recipients, or access grants into .commentary/session.json, source files, docs, comments, or skill memory. Sharing state is stored by Commentary.

After sharing a review, ask the current user how to handle comment threads authored by other reviewers:

• Default to requiring approval before resolving other-authored threads.
• Accept approval from the current user in the agent conversation, or from a reply in the Commentary thread approving the requested change.
• If the user explicitly authorizes automatic resolution for other-authored threads, continue the normal edit, sync, and resolve loop for those threads.

Comment Loop

For existing comments, read open threads:

commentary comments --format markdown --open

For live collaboration, use next-comment. It starts the live event stream, checks currently open threads, and waits only when no open threads exist:

commentary next-comment --timeout 60s --json

Use narrower waits when helpful:

commentary next-comment --file docs/spec.md --timeout 60s --json
commentary next-comment --no-include-replies --json

Use short bounded waits for interactive agent loops so the agent can stop promptly when the user says to stop. Use longer waits such as --timeout 15m only when the user explicitly asks for unattended listening.

wait-comment is future-event-only by default. It is not a substitute for checking open threads and should be used only when the workflow specifically needs a future live event cursor.

For a dedicated background-style listener, use cooperative stop-file support:

commentary comments --watch --jsonl
commentary comments --stop

Run comments --stop instead of killing processes manually. If using a custom stop file, pass the same --stop-file <path> to both commands.

When a comment arrives:

1. Inspect the thread id, file path, selected text, line hints, and comment body.

2. Determine whether the thread appears to be authored by the current agent alias or by another reviewer. Use agentAlias, authorLogin, and author fields when present; if authorship is ambiguous after a review has been shared, treat the thread as other-authored.

3. Edit the local file to address the request.

4. Upload a new revision:

   commentary sync --message "Address review comments"
   

5. Resolve only when the requested change is fully addressed, and include the final visible response in the resolve command:

   commentary resolve <thread-id> --message "Updated the introduction and synced revision 2." --alias "Docs agent"
   

For other-authored threads after sharing, do not resolve by default after syncing. Ask the current user for approval, or wait for an approving reply on the thread. A reply created by another reviewer can arrive through the default next-comment behavior because replies are included unless --no-include-replies is used. If approval is missing, leave the thread open and optionally reply with a concise progress update.

Use commentary reply only for non-final progress updates or follow-up discussion, because replies can reopen resolved threads.

Keep repeating short next-comment --timeout 60s --json waits, edit, sync or track, and resolve --message until the user says the review is done or there are no open threads.

Pulling Reviewed Content

Prefer editing local files directly and syncing revisions. Use pull only when the user asks to bring Commentary-side content back to disk. Start safely:

commentary pull --dry-run

Then choose a non-destructive option unless the user explicitly wants overwrites:

commentary pull --output reviewed
commentary pull --backup --yes

CLI Versus MCP

The Commentary CLI is the default for file-backed draft review workflows because it manages .commentary/session.json, tracks local files, syncs revisions, watches file changes, and streams live comment events.

If a Commentary MCP server is already installed and exposes equivalent draft-review tools, prefer MCP only for typed, in-process operations that do not need local file watching or project metadata, such as listing or replying to comments. Before using MCP against a CLI-created review, verify that MCP can read the exact draft review session id. If MCP reports "Draft review not found" while the CLI can access the same session, treat it as an auth/resource mismatch and continue CLI-only. Do not mix surfaces for that session. Do not require, install, or configure MCP for this skill. Fall back to the CLI whenever MCP is unavailable, incomplete, ambiguous, or not explicitly connected to the current local files.

Operational Guardrails

• Keep the CLI thin: use Commentary for review transport, not for git branching or commits.
• Use --git-base auto or explicit git base flags for real GitHub base metadata; do not put Base commit: <sha> in the description as a substitute.
• Use --json for automation and stable parsing.
• Use --dry-run before potentially overwriting local files.
• Do not print spinners or rely on interactive prompts in non-TTY automation.
• Prefer short bounded waits for interactive agent sessions; use comments --watch --jsonl plus comments --stop for long-running listeners.
• Preserve .commentary/session.json; it stores review metadata but must never contain secrets.
• Do not persist share links, recipients, access grants, or reviewer instructions in local repo files unless the user explicitly asks to document them.
• Respect draft-review limits: at most 20 files, 512 KiB per file, and 2 MiB total per revision.