idd-skill: Instructions file for GitHub Copilot

IDD — Orchestrator Phase (fan-out)

Read this only when fan-out activates or its session restarts. Requirements: [fan-out variant][fv] and delegation.

1. Activation

A restarted session with a unique session-suffixed agent id skips this section. Any other restart neither activates fan-out nor refills: it follows recovery (section 2).

Activate fan-out only when all conditions hold:

• the harness's worker mechanism table row has Eligibility orchestrator;
• helper support is installed (instructions-only has none), so the profile-selected idd-worker-budget exists;
• the effective orchestrator.maxWorkers is at least 2 (read maxWorkers from idd-worker-budget --running 0 --startable 0);
• the session is not on the lite profile;
• the session was not started for one explicit issue (A0-T) or a single-issue request;
• GITHUB_ACTIONS is not true.

Otherwise, continue with idd-discover.instructions.md unchanged. The cap is per orchestrator session, not per host.

2. Roster

One entry per worker: issue number, claim ID, branch, worker handle, report draft path, and state (running, reporting, disposed, stalled).

Lost on restart: see recovery.

3. Dispatch

1. Run Discover A0 through A4 Step 1.5 (A0-C runs once per Discover run, before the first dispatch; refills skip it); survivors are the startable candidates.
2. Call the profile-selected idd-worker-budget with --running <count not disposed> and --startable <candidate count>. Add --harness-limit <limit> when the table gives this harness a numeric limit. Its slots is the number to start now.
3. If slots is 0, go to section 4. Otherwise pick that many candidates with discover-shared-file-overlap, passing survivors as --issues, --batch <slots>, one --in-flight per entry not disposed, and --check-overlap. When discover.selectionDesync is session-offset, also pass --desync-token (generated at Discover entry) on every refill.
4. For each pick in batch order, run A4.5, then the fresh-claim order in idd-claim.instructions.md through claim verification only (worktree first; B1 is the worker's). Start a worker with the delegation brief, a section 5 report request naming a per-worker report path, and the pre-created worktree it reuses without B1. Record a running entry whose handle is the identifier the start returned.
5. A pick that fails A4.5, or A5 before the claim marker is posted, has its worktree and branch removed (deleting its lock). A failure after the marker is contested: follow idd-claim.instructions.md and keep its claim and worktree. Either way no worker starts, so no entry exists.
6. Dispatched and failed picks leave --issues before a refill.