1. Confirm the installed client and server capabilities
CLI 0.2.0, Agent SDK 0.2.0 and Tracker SDK 0.1.0 are tested implementation candidates; npm, CDN and marketplace publication are not implied. Use the installation and authentication instructions in the public documentation for an actually available release.
Select an authorized workspace and scoped agent grant. Inspect the server's current OpenAPI or MCP discovery before using new operations. Old-server capability errors should stop the new workflow with a useful diagnostic.
2. Create optional Project context
Save a project.json request with workspaceId, name and objectivesMarkdown. The selected workspace must match your grant. Reuse an idempotency key only for an exact retry; use a new key when the intent changes.
The equivalent MCP call is commentary_projects with action create and the same input object. Do not put credentials, reviewer identities or provider secrets in the payload.
commentary project create --workspace WORKSPACE_ID --file project.json --idempotency-key export-project-1 --json
{
"workspaceId": "WORKSPACE_ID",
"name": "Reviewable export",
"objectivesMarkdown": "Review the exact export behavior and source boundaries."
}3. Validate and publish a static walkthrough
Create README.md and details.md in ./walkthrough alongside the manifest below. Use the dedicated package collector; it never runs customer build scripts. Publication is resumable and preserves immutable revisions.
The MCP review_package family supports publication and upload workflows for clients without a filesystem CLI. Use current discovery for upload fields and limits.
commentary package validate ./walkthrough --manifest ./walkthrough/commentary.package.json
commentary package publish ./walkthrough --manifest ./walkthrough/commentary.package.json --idempotency-key export-package-1 --no-open --json
{
"kind": "commentary_review_package",
"manifestVersion": 1,
"title": "Reviewable export walkthrough",
"entryPageId": "overview",
"pages": [
{ "id": "overview", "path": "README.md" },
{ "id": "details", "path": "details.md" }
],
"trustTier": "static"
}4. Compose exact targets and ask for review
Read the ready package receipt, then create a Session using its Resource id, exact revision id and manifest digest. Add authorized source and evidence targets as needed. The MCP family is project_review_session; create and acceptance_request call the same authoritative services.
Use review-session get to read the current ETag before a conditional mutation. Request review through review-session request-review with the exact server-validated scope. Agents have no human acceptance operation. Canonical request examples and schemas are in the OpenAPI reference.
5. Correct, reconnect and keep authority current
Read native discussion and typed feedback; revise the original artifacts and exact Session manifest. On a cursor gap, refetch bounded current authorized state. Retry an indeterminate mutation with its original body, ETag and key before changing intent.
Do not infer acceptance from execution, fulfillment or evidence. Provider and linked-Resource authority are rechecked at server boundaries. A custom renderer uses the Tracker SDK broker and isolated origin without agent or app credentials.
MCP alternative: create the same Project context
For an MCP-only client, connect the authenticated Commentary endpoint and discover commentary_projects. Send this tools/call request with a real authorized workspace id. It uses the same canonical create schema as the CLI request above.
Retain the returned identity and transport metadata. Continue with discovered review_package and project_review_session actions; upload and exact revision fields come from their current schemas. No client can manufacture a human Decision.
{
"jsonrpc": "2.0",
"id": "project-create-1",
"method": "tools/call",
"params": {
"name": "commentary_projects",
"arguments": {
"action": "create",
"input": {
"workspaceId": "WORKSPACE_ID",
"name": "Reviewable export",
"objectivesMarkdown": "Review the exact export behavior and source boundaries."
},
"idempotencyKey": "export-project-1"
}
}
}Continue with the tested surface
Public technical documentation covers installation, current scopes, canonical request fixtures, compatibility and recovery. Read OpenAPI or discover MCP tools before enabling a client. The skills integration guide describes supported installation targets and their limits.
Projects, Trackers and Agentic Engineering Review explain the candidate product surfaces.