Goal

Let projects talk to each other through GitHub issues. When something outside a repository's own T3 agents opens an issue, comments, or edits or closes an issue in a managed repository, that repository's orchestrator finds out. Waking it should cost almost nothing.

Example: an Amine-Thermodynamics agent opens an issue in ePC-SAFT, or comments on one. The ePC-SAFT orchestrator should notice without the owner relaying it.

Owner request and decisions (2026-10-14).

What exists

• Writing across projects already works. In full access, agents can run gh issue create -R <other repo> and gh issue comment.
• A background job already polls each managed repository. IssueWorkService runs repository-issue-work every 60 s and reconciles every active repository root. It lists issues in the selected milestones and admits eligible ones. It doesn't notice external comments, edits or new issues outside that admission flow.
• T3 knows its own writes, through issue_changes and the comment receipts and effects. That lets it tell "ours" apart from "external".

Design (decided)

• Scope: managed repositories only, meaning repositories with a repository orchestrator.
• Watcher, with no language model:
  • Every ~2 minutes per managed repository, inside the existing scheduler, read issues and issue comments changed since a stored cursor: GET /repos/{o}/{r}/issues?since=…&state=all and /issues/comments?since=…, with ETag/conditional requests where supported.
  • Drop T3's own writes for that repository (known comment and issue ids, plus the T3 actor) and bot accounts.
• Inbox: a fork table, one row per external event: repository, issue number, kind (issue_opened, comment, edited, closed/reopened, labeled/milestoned), author, URL, a short snippet, tagged, created time and read state.
• Tagged: the issue has the label cross-project, or the comment or body mentions @orchestrator.
• Delivery to the repository orchestrator:
  • Untagged events: stay in the inbox and are mentioned in one line at the orchestrator's next natural turn. They never start a turn.
  • Tagged events while the orchestrator is idle: at most one short digest message every 15 minutes, and only when new tagged events exist. For example: "2 tagged external issue events since 14:00 (epc-saft fix(web): show sidebar counts in the caret slot, caret on hover #61 comment by amine-agent, fix(web): flip the sidebar's Don't ask toggle to Ask mode #62 opened). Read them with t3_issue_inbox." It holds no bodies, to stay token-light.
  • Tagged events while the orchestrator is running: the same one-line note is delivered at the next safe point. Queue it rather than steering, because Claude's steering interrupts tools.
• MCP tool: t3_issue_inbox (list unread, read by id, mark read), available to the repository orchestrator and its children.
• UI:
  • an unread badge on the repository orchestrator's sidebar row;
  • an "Inbox" section on the repository's Issues page that lists events with links and marks them read.
• Modularity: fork-owned files behind small hooks. A fork migration numbered after the current ledger.

Not in scope

Webhooks: polling is enough at this scale, and webhooks need a public endpoint. Luna triage: add it only if the rules prove noisy. Mobile. Writing to other repositories, which already works.

Done when

• An issue or comment created in managed repository B by something other than B's own agents appears in B's inbox within a few minutes.
• A tagged event wakes an idle B orchestrator with one short digest at most every 15 minutes.
• Untagged events never start a turn.
• B's own agents' writes never appear.
• The badge and the Issues page Inbox show unread events.
• Tests cover the filter, the cursor, digest throttling and tagging. Live GitHub is verified on a mock repository, not the owner's real ones.