agentleFS
Vinix24Vinix24/vnx-orchestration/.claude/skills/t0-orchestrator/SKILL.md

t0-orchestrator

Skill63 starsChanged 5 days agoMITView on GitHub

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.

The file

We split this skill into the rules it gives an agent

5 rules. About 0 are common, 3 are distinctive, and 2 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈3

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 2

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Very specific:

    Do not rely on runtime /model switching.

    line 53
  2. Distinctive:

    Read latest receipt(s); read the QUALITY advisory first.

    line 61
  3. Very specific:

    Verify required review-gate evidence (incl.

    line 64
  4. Distinctive:

    Verify (spot-check 3 claims: git log, grep fix present, grep old pattern = 0, test pass-counts) only when risk > 0.3.

    line 82
  5. Distinctive:

    When uncertain: (1) request a second review on the same evidence.

    line 111

≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

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/Sonnet → claude_headless (claude -p, subscription, isolated worktree per dispatch), the only claude lane since the tmux-spawn lane was removed on 2026-09-18; never provider_dispatch. kimi/glm/deepseek → provider_dispatch. Mechanism and failure modes: DISPATCH_RULES §8. Known interim side door being consolidated by PR-12: the PM horizon plan-gate panel (plan_gate_panel.py) calls lanes directly until the door is wired.
  • Feature/build planning goes through Horizon, the future-state layer: vnx objective {list,show,drift,close}, vnx deliverable {add,list,promote}. The tracks DB is the SSOT; ROADMAP.yaml is a static example. Use the horizon skill for planning and plan-first gate panel decisions. See vnx_cli/commands/horizon.py. NOT ad-hoc plan docs 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 (t0-opus-only, pin_semantics=floor). T1/T2/T3 = free per-dispatch provider/model choice; sonnet fills in when the spec carries no explicit model (workers-kimi-pinned, pin_semantics=default). kimi-k3 is an explicit choice, not the default.
  • 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)

  1. Read latest receipt(s); read the QUALITY advisory first.
  2. Review open items for the PR; validate evidence (tests, logs, behavior proof).
  3. Close/defer/wontfix with explicit reasons.
  4. Verify required review-gate evidence (incl. headless report artifacts) — see DISPATCH_RULES §2.
  5. Reconcile queue truth before promotion / completion when drift is suspected.
  6. 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. SERIAL      dispatch needs claude-tmux lane AND lane occupied      → 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.

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 (N-slot, default 10). Providers stay parallel. The door enforces this (PR-6 lock + #1017 concurrency config); 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 by default; headless claude -p is opened (2026-08-11 operator directive) but usable-not-yet-governed (ignores isolation=worktree, report-gate gap — see DISPATCH_RULES §6). Both lanes run on the subscription, not API billing — cost=$0.0000 confirms it either way. 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.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.