+
+ 01 / First release
+ Mark and copy
+ Personal “For agent” marks on existing threads. Copy individually, by selection, or across a document or workspace. Paths, quotes, versions, requests and replies travel together.
+ Done when: native, hosted and reviewer exports agree; marks survive reload; Markdown, HTML and stale selections retain useful context.
+
+
+ 02 / Connect
+ Give agents the tools
+ A common service to read feedback, wait for changes, acknowledge receipt, reply to a thread and propose a diff. A CLI and MCP adapter expose the same contract.
+ Done when: a shell consumer and an MCP client can complete the same feedback loop, with replies attached to their original threads.
+
+
+ 03 / Optional live mode
+ Keep the conversation moving
+ Connect a workspace and agent session. Queue new comments, show delivery status, pause or resume. Native first; hosted use pairs with a local bridge or agent runner.
+ Done when: two independent agent hosts handle new comments, follow-ups and reconnects without duplicate replies or proposals.
+
+
+
+
Defaults to review
+
Your routing marks are personal and persist on this device. Comments keep their normal room visibility. V1 uses existing review rooms; unshared local comments and cross-device routing sync are later extensions.
+
Copying never means “resolved.” Source edits follow permissions in the agent conversation; attn suggestions remain yours to accept. A hosted document needs source mapping or a proposed patch.
+
+
+
What “live” can promise
+
attn delivers new feedback immediately. The agent host chooses when to act: at a tool boundary, while waiting, or through supported mid-run steering. A connection alone cannot make an arbitrary chat app run a model.
+
Real delivery needs durable cursors, revision-aware retries, cancellation and accurate status. Agent replies must never trigger themselves. MCP interaction model ↗
+
+