autopilot-issue-delivery-scheduler

User-facing ACCEPTED-ISSUE implementation queue. This skill SELECTS authorized issues and THROTTLES fan-out. It does not triage and does not review PRs.

Compose autopilot-issue-delivery-worker one issue per slot. Never borrow slots from autopilot-issue-triage-scheduler, autopilot-pr-review-scheduler, or autopilot-pr-triage-scheduler.

Activation card

activation_card: on. Before any queue read or spawn, emit this Enter card with every field filled. Missing field -> stop.

skill: autopilot-issue-delivery-scheduler
skill_path: <resolved directory of this SKILL.md>
mode: run
subject: microsoft/apm
path: delivery
intent: select accepted issues and fan out delivery workers
origin: unattended | actor-session
write: off
repo: microsoft/apm
fanout_limit: <positive integer>
invocation: agentic-workflow | actor-session

Rules:

• write is always off. This scheduler never comments, labels, assigns, or requests reviewers.
• write: on -> stop.
• origin fail-closed unknown -> unattended.
• invocation is the harness. Copilot App, local session, Cloud Agent, and Remote Agent are actor-session. agentic-workflow is only gh-aw / GitHub Actions. Do not copy origin into invocation.
• One queue. Do not nest another scheduler path.

After the run, emit this Exit receipt:

skill: autopilot-issue-delivery-scheduler
subject: microsoft/apm
path: delivery
write: off
queued: <integer>
spawned: <integer>
approved: n/a

Invocation

Works in local sessions, Copilot App automations, Cloud Agent, Remote Agent, and Agentic Workflows.

Resolve ORIGIN before any GitHub write:

• unattended -- workers still run; they must not assign.
• actor-session -- workers treat assignment as a hard gate before any implementation. Assignment is the public signal of which user is working the issue.

Ownership writes:

• Issue triage: none (no assignment needed)
• Code (accepted implementation): assign the implementing user (@me)
• PR review: request the reviewing user as reviewer, never assignee

INTENT for this scheduler is implement only as a parent signal to workers. This scheduler itself never assigns, never requests reviewers, and never writes human decision labels.

Selector

• all (default) -- every open issue that already carries status/accepted, plus any issue the caller named as a bounded accept. Type does not matter (type/bug, type/feature, type/automation, docs, refactor). Escalate everything else. Do not run triage-panel from this scheduler.
• bugs -- same gate, plus type/bug. Worker uses LEGIT / UNCLEAR / FIXED-AT-HEAD plus PRINCIPLES.md alignment.

type/bug or triage/recommended without status/accepted is not eligible unless the caller named a bounded accept.

Selection

Build the queue from the caller list or gh issue list on status/accepted (selector bugs: also require type/bug).

An issue is eligible for the queue only if it already carries status/accepted or the caller named it as a bounded accept. That membership is a communication signal, not implementation permission. triage/recommended and legacy status/triaged are advisory processing markers and are not authorization. Workers re-check scripts/governance/eligibility.cjs from the trusted default branch and require fresh responsible-human confirmation before any mutate. ORIGIN unattended never implements. Do not dispatch an unaccepted issue. Do not run triage-panel to create any marker. Do not write status/accepted.

Skip locked and closed issues unless the caller named them. Deduplicate by number. Do not skip bot-authored issues that already carry status/accepted or that the caller named. Human accept is the gate; author type is not.

Skip status/needs-design unless the caller named the issue. Skip status/needs-triage and status/deferred unless named. Skip status/shepherding and status/in-flight unless named (those belong to PR review or an in-flight owner). Skip issues assigned to another user unless the caller named them; do not steal in-progress work.

Do not open a second PR when one already addresses the issue; return that PR number.

After a worker opens a PR, do NOT fill this pool with autopilot-pr-merge-worker. Hand the PR number to the caller for autopilot-pr-review-scheduler.

Queue table (mandatory)

Before any spawn, emit the keep-set and the drop-set in plan.md and in the session report. Missing table, missing column, or blank rationale -> stop. Do not spawn.

Keep-set (items this run will schedule). One row per item. Do not truncate to FANOUT_LIMIT:

| number | kind | labels | rationale | slot |

Drop-set (considered, then not scheduled). slot is -:

| number | kind | labels | rationale | slot |

• kind: issue or pr
• labels: current GitHub labels, comma-separated; none if empty
• rationale: why queued or dropped (helper rule and labels)
• slot: 1-based spawn order, or - when dropped

Empty keep-set is success. Still emit the drop-set, or none. You are the sole table writer. Put selector and existing-PR facts in rationale.

Fan-out

Load assets/fan-out-pool.md. Default FANOUT_LIMIT=2 concurrent slots. Isolated to this run. FANOUT_LIMIT is concurrency, not queue length. Drain the full selected list; when a slot returns, fill it with the next item.

Procedure

1. Probe worker-code on disk. Missing sibling -> stop.
2. Build the queue. Emit the mandatory queue table (labels + rationale) before any spawn.
3. Drain the selected list. Concurrent slots <= FANOUT_LIMIT. One issue per slot. When a slot returns, dispatch the next queued issue. Do not stop because the pool was full. Name each worker session Issue delivery #<issue-number>.
4. Never auto-merge.
5. Print a final report from the table.

Hard nos

• Do not share this pool.
• Do not dispatch the same issue to two slots.
• Do not dispatch an unaccepted issue.
• Do not dispatch an issue assigned to another user unless named.
• Do not drop a bot-authored issue that already carries status/accepted.
• Do not triage or review PRs inside this scheduler.
• Do not contradict CODEOWNERS.
• ASCII only.