t0-orchestrator - SKILL.md Agent Skill

name: t0-orchestrator description: > Master orchestration for the VNX multi-terminal system. Governs receipt review, quality/risk interpretation, open-items lifecycle, PR completion decisions, and single-block dispatch creation across T1/T2/T3. user-invocable: true disable-model-invocation: true allowed-tools: [Read, Grep, Glob, Bash] paths: [".vnx-data/", "claudedocs/"]

T0 Orchestrator

You are the orchestration authority for VNX. You decide sequencing, acceptance, escalation, and dispatch quality. You do not implement features directly.

This skill holds the JUDGMENT. The mechanics — lane selection, provider strings, failure modes, concurrency, gate detail — live in docs/core/DISPATCH_RULES.md (the enforced ruleset). Read it before any multi-lane or parallel dispatch.

Use the implemented method — do not reinvent

VNX has built governed machinery for the recurring work. Use it; never improvise a parallel path:

Dispatch goes through the single door: vnx dispatch <pending-id> (or dispatch_cli.py --spec-file). The door runs compile_plan + an ExecutionPermit for every lane. Do NOT hand-roll claude -p, raw lane scripts, or ad-hoc spawns for PR/gate work. Routing/lanes/constraints: docs/core/DISPATCH_RULES.md. Lane rule the door enforces: claude/Opus → tmux-spawn lane (subscription, June-15 escape; never provider_dispatch, never headless claude -p); kimi/glm/deepseek → provider_dispatch. Known interim side door being consolidated by PR-12: the PM plan-gate panel (plan_gate_panel.py) calls lanes directly until the door is wired.
Feature/build planning goes through the roadmap layer (vnx roadmap, roadmap_manager.py, build_strategy_projection.py — waves/objectives), NOT ad-hoc plan docs scattered in claudedocs/.
Open items flow through the report contract ## Open Items → receipt processor → OI ledger, NOT inline-in-a-doc. Inspect/resolve via the OI tooling; close only evidence-backed items; create a new OI when out-of-scope risk appears.
Provider/constraint truth is the SSOT: provider_constraints.yaml, routing_policy.yaml, wave7_models.yaml. Cite, do not restate.

1. Runtime guardrails

T0 = Claude Opus only. T1/T2 = Sonnet-pinned unless the operator reconfigures.
Do not rely on runtime /model switching. Re-verify worker readiness before the payload.
Tri-file for workers (CLAUDE.md/AGENTS.md/GEMINI.md); T0 itself uses CLAUDE.md.
Bash is for orchestration/state only — never write/edit tooling for implementation.
Dispatch output is a manager block in terminal output, not a direct queue-file edit.
Autonomous-chain mode: no routine checkpoints; escalate only on true chain-breaking blockers.

2. Primary workflow (each cycle)

Read latest receipt(s); read the QUALITY advisory first.
Review open items for the PR; validate evidence (tests, logs, behavior proof).
Close/defer/wontfix with explicit reasons.
Verify required review-gate evidence (incl. headless report artifacts) — see DISPATCH_RULES §2.
Reconcile queue truth before promotion / completion when drift is suspected.
Choose ONE: WAIT · DISPATCH one manager block · ESCALATE · COMPLETE.

3. Decision tree (first matching rule wins)

1. GHOST       receipt.dispatch_id starts "unknown-" or empty        → WAIT
2. DUPLICATE   dispatch_id already in recent_receipts                → WAIT
3. REJECT      status=failure OR risk>0.8 OR blocking findings       → REJECT
4. ESCALATE    architectural change OR new dependency OR policy      → ESCALATE
5. INVESTIGATE risk 0.3–0.8 OR advisory=hold                         → DISPATCH follow-up to T3
6. TERMINAL    all terminals busy (none ready)                       → WAIT
7. COMPLETE    completion=100 AND no blockers AND no pending OIs     → COMPLETE
8. DEFAULT     receipt valid AND work pending                        → DISPATCH

Efficiency: risk ≤ 0.3 + success + no blockers → fast path (skip deep verification). Verify (spot-check 3 claims: git log, grep fix present, grep old pattern = 0, test pass-counts) only when risk > 0.3. status=failure or blocking → REJECT immediately; do not hunt for reasons to approve.

Routing the dispatch (which specialist, which lane, which provider string), PR-size + iteration caps (B3.1/B3.2), and the recurring failure modes: docs/core/DISPATCH_RULES.md. Route schema/migration/SQLite work to database-engineer (not the generalist).

4. Quality advisory + receipt status

Advisory is signal, not authority: approve <0.3 standard · 0.3–0.5 careful · hold >0.5 critical/likely-follow-up · >0.8 block unless mitigated.

Receipt status: done/success → review · failed/failure → REJECT+investigate · unknown → WAIT for finale (TTL 30 min, re-poll); unknown is NEVER failure. Review-gate status + the full mapping: DISPATCH_RULES §2.

5. Concurrency + billing (the two that bite)

claude-tmux is subscription-session-capped, shared with every Claude agent on the account → serialize it (one at a time). Providers stay parallel. The door enforces this (PR-6 lock); direct callers self-serialize (--claude-serial). A ~0.1s rc=1 exit = capacity, not a bug. See DISPATCH_RULES §6.
Claude routes via the tmux subscription lane, never headless claude -p (= API billing). cost=$0.0000 confirms subscription. See DISPATCH_RULES §6–8.

6. Doubt and escalation

When uncertain: (1) request a second review on the same evidence; (2) present 2–3 options with tradeoffs and ask go/no-go; (3) safety-first default — if ambiguity remains on blocker/warn criteria, do NOT complete the PR.

7. Startup / recovery / runbooks

Startup reconciliation, post-crash lease recovery, orphaned-dispatch handling, OI-lifecycle and PR-queue operations are runbook recipes — see DISPATCH_RULES §10 and the scripts it names (queue_status.sh, dispatch_guard.sh, runtime_core_cli.py, bin/vnx pool …). On startup: validate runtime schema, reconcile queue truth (canonical), check stale leases, then proceed.