Load /unslop before the first progress update and keep it active for every question, issue comment, ticket body, decision, and summary.
Accept /supervise-this <task> for a new run or /supervise-this #<spec> to resume. Reject invocation outside an AO project orchestrator session.
Read AGENTS.md, ARCHITECTURE.md, the supervise-this leaf, CONTEXT.md, the parent issue when resuming, and the AO project configuration.
Set REPO from the AO project's configured repository. Pass --repo "$REPO" to every gh command so reconciliation survives a broken worktree gitdir.
Verify ao status --json, the current AO_PROJECT_ID, the orchestrator session, project role profiles, ao agent ls --refresh --json, GitHub access, and an explicit reviewer policy.
Require an orchestrator profile for planning and final review and a worker profile for implementation. Resolve each profile's agent, model, and supported session modes from AO. Choose chat or tui from the worker's capabilities.
Fetch origin/<default-branch>. Synchronize the project base or prove it is an ancestor before planning or spawning. A worker repeats that proof before its first edit and reports BASE_CURRENT.
Feed these observed facts to scripts/workflow.ts preflight --input <path>. Stop before a spawn on any returned error. Read the exported input types when preparing JSON. Every helper command requires --input or --json, so a missing input fails instead of waiting on stdin.
Ask one ELI18 decision with a recommendation when preflight cannot pass.

Plan

For a new task, delegate the complete planning workflow to /plan-this. Planning runs in this same persistent orchestrator session, not a child Agent Manager session. Honor every planning approval gate.
After /plan-this publishes the parent specification and tickets, record one structured parent comment containing the AO project, orchestrator and worker role models, review round, and fixed review base.
Treat planning as complete only when the parent issue and every planned child ticket are published and their native dependency edges are readable.
Continue the same orchestrator run after recording that checkpoint. Do not give the user a conclusive summary yet.

Schedule

Reconstruct the durable state from GitHub before reading live AO sessions: parent issue, child and follow-up issues, native blockers, labels, assignees, comments, pull requests, checks, and merged commits.
Add AO sessions and local or remote branches to that snapshot. Feed open PRs, sessions, branches, assignees, and issue links to scripts/workflow.ts reconcile. Its review, resume, or spawn result is the scheduling gate.
Before each spawn, rerun preflight with ownershipClear true only for the helper's spawn result. DUPLICATE_OWNERSHIP is a stop signal.

Preserve parent order and keep at most three implementation workers active. A merged blocker makes its descendants eligible for the next wave. Start each worker with the mode returned by preflight:

ao spawn --project "$AO_PROJECT_ID" --kind worker --name "issue-<n>" --issue <n> --mode "$WORKER_MODE" --prompt "Run /implement-this #<n> in AO pull-request delivery mode. Prove origin/<default-branch> is an ancestor before editing."

Use names of twenty characters or fewer. The project worker profile chooses the model. Do not pass an unverified --model flag.
A worker spawn is an intermediate checkpoint. Keep the orchestrator active, consume AO completion or handoff events, and reconcile GitHub before the next spawn. If AO presents a worker completion message, process it instead of summarizing the whole run.

Worker contract

Each worker receives exactly one issue and runs /implement-this #<n>.
AO workers use pull-request delivery. The worker commits and opens or updates its PR, but does not push directly to main or close the issue before merge.
AO owns the worker session, worktree, CI feedback, review feedback, merge conflict routing, and session recovery. The supervisor owns the dependency graph, acceptance evidence, issue labels, and parent state.
Track only these delivery states: READY, CLAIMED, BASE_CURRENT, EDITING, PR_OPEN, REVIEWED, MERGED, EVIDENCED, and CLOSED. Derive them with scripts/workflow.ts; AO activity labels never advance work.
Use ao session get <id> --json and ao send --session <id> --message "..." for targeted recovery. A blocked or permission-waiting worker receives no automatic prompt that would bypass its approval.
A worker is complete at CLOSED, after merge and acceptance evidence.
After a merge, re-read native blockers and refill free worker slots in parent order. Descendants remain unscheduled until every blocker reaches MERGED or a later delivery state.

Resume and recovery

/supervise-this #<spec> rebuilds the ownership snapshot and runs scripts/workflow.ts reconcile before any write. Repeated evidence collapses to one action, so resume cannot duplicate issues, sessions, branches, or PRs.
Evaluate an idle worker with scripts/workflow.ts idle-signal. An open issue with no matching PR and no tracked change is red, regardless of lastActivityAt.
Classify each failure as infrastructure, task, or implementation. Track retry counts per class with the helper. Finish infrastructure recovery before sending task continuation; send implementation correction only for the last class.
Add needs-info, record the concrete blocker, and ask one ELI18 decision when a class exceeds its cap. Preserve the AO session and issue history.
Record replacement decisions in a new parent comment. Never silently change the AO project role model or use a cheaper replacement.

Review and closure

After every planned child PR merges, record a fixed base and run full repository verification. Then run /code-review in the AO orchestrator against that base and the parent specification.
Apply the configured reviewer policy through scripts/workflow.ts review-decision. Same-account review may record a verdict but cannot satisfy an approval-only policy.
Record the reviewed head SHA. Feed current head, reviewed head, repository, and AO manageability to scripts/workflow.ts merge-decision. Stop if the head changed. Use its AO command for managed PRs. Use its explicit-repository GitHub fallback only when AO cannot manage a legacy unclaimed PR.
Turn each confirmed in-scope finding into the smallest independently verifiable follow-up ticket with native blockers where needed. Spawn follow-up workers through the same AO worker profile and PR contract.
Allow at most two automatic follow-up and review rounds. After the second, mark unresolved findings needs-info and ask the user for one decision.
Close the parent only after every planned and follow-up PR is merged, every issue has acceptance evidence, all required checks pass, and final review has no confirmed finding. Post one final evidence comment with role models, AO sessions, PRs, commits, checks, review base, recovery decisions, and outcome.
Finish with an ELI18 Why / What / Where / How summary only at that terminal condition or when a human decision blocks all safe progress.

supervise-this

Start