orchestrator-refresh

name	orchestrator-refresh
description	Re-assert orchestrator discipline on a one-shot delayed wake: ledger reconcile, advisor routing, warm executor reuse, single-pending re-arm via status_gate. Terminates when the gate says stop. Triggers: '/orchestrator-refresh', orchestrator-refresh, refresh the orchestrator loop, re-arm orchestrator.

Orchestrator Refresh

Detect the host profile first (see Host profiles in _shared/advisor/advisor-protocol.md). Re-assert the discipline for that host only — do not invent an Agent-tool Claude session-advisor spawn on a third-party host.

0. status_gate first (deterministic)

Script home (prefer install path, else package checkout):

• %USERPROFILE%/.claude/skills/orchestrator/scripts/status_gate.py
• ~/.claude/skills/orchestrator/scripts/status_gate.py
• skills/orchestrator/scripts/status_gate.py

Pass the same --run-slug used at activate, if any.

0a. begin-firing (consume prior re-arm)

python <status_gate.py> begin-firing [--run-slug SLUG]

Exit	Action
1	Stop. Cancel matching host schedules for /orchestrator-refresh if the host allows. Report inactive/done. Do not re-arm. Do not spawn.
0	Latch cleared. Continue with steps 1–6.

0b. Done after ledger (step 1)

After ledger reconcile: if every task is completed/cancelled and no executor is running:

python <status_gate.py> set --status done [--run-slug SLUG]

Cancel matching host schedules; stop without re-arming.

Discipline steps

1. Reconcile the task ledger first. Call TaskList after the gate. The ledger is stale when any of these holds: a running or finished executor has no in_progress task naming it as owner; a finished executor's task is still open (or was closed without its result merged); the next phase you will dispatch has no pending task; a blockedBy link contradicts the actual run order. Fix every mismatch with TaskCreate / TaskUpdate in this same firing — never defer.

2. You are the orchestrator. Orchestrate and hold the user conversation; spawn executor subagents for every code edit and build or test run.

   • Worker model. Every coding spawn sets model: sonnet on a Claude host, or the resolver-printed sonnet-equivalent id on a third-party host, and fails closed — see Workflow Agent Routing in skills/orchestrator/SKILL.md.
   • Focused tickets. One mechanical done-check per ticket; resume a warm agent with a thin next-slice ticket rather than a fresh cold spawn, and keep thick context in the assignment file.

3. Hard decisions go to the shared advisor.

   • Claude host: executors consult the warm session-advisor via SendMessage (ENDORSE / CORRECTION / PLAN / STOP). This session routes the same way; keep tool use to orchestration and light verification reads.
   • Third-party host: advisor is a max-tier Claude CLI bind owned by this session (claude_chain_runner.py, Fable high then Opus max). Do not spawn session-advisor via Agent. Executors report blockers here; consult the CLI advisor and relay signals. If the CLI bind is unreachable, fail closed — do not answer the four signals as this third-party session.

4. Resume before you spawn. SendMessage an existing executor by name or agentId before a cold spawn. (Third-party: executor reuse only — advisor stays on the CLI chain.)

5. Fresh spawn only for a genuine task switch. Never tell an agent to compact for a clean context.

6. Single-pending re-arm only. Same protocol as the orchestrator skill (host-agnostic; create then claim):

   1. Cancel matching schedules only when the host can list/cancel by prompt (/orchestrator-refresh + --run-slug if used). On Claude, skip selective cancel — the latch is the sole stacking enforcement.
   2. python <status_gate.py> should-reschedule [--run-slug SLUG]
      • Exit 1 → stop; do not schedule.
      • Exit 0 → continue.
   3. Create exactly one non-recurring delayed wake (~1200–2700s) with prompt /orchestrator-refresh (plus --run-slug when used). Host one-shot tool only (on Claude: ScheduleWakeup). Never recurring / never cadence / never a second create this firing.
   4. python <status_gate.py> claim-rearm [--run-slug SLUG] right after a successful create. Exit 1 → cancel that schedule and stop.
   5. If create fails: do not claim; stop or retry once from cancel.

Gotchas

• Stacking loops. A second schedule create while one is already queued multiplies firings. Cancel → should-reschedule → one create → claim-rearm. Never claim before create on Claude (PreToolUse denies ScheduleWakeup when pending).
• Skipping begin-firing. Prior rearm_pending stays set; re-arm stays denied. Always run step 0a first.
• Create without claim. Skip claim after create and a second create can stack. Claim immediately after success.

File Index

File	Purpose
SKILL.md	Refresh firing steps; points at orchestrator status_gate.py.

Folder Map

• SKILL.md — this skill (thin); gate implementation lives under skills/orchestrator/scripts/.
• Advisor policy: _shared/advisor/advisor-protocol.md.