AGENT SKILL

orchestrator

Multi-agent workflow coordination, lead-dev operating model (codex 67-68), and task delegation. Default when suit is worn — no keywords required.

License 许可
MIT
所属仓库
0xRayAI/xray
最近更新
2026-11-13
Stars
★ 2
Forks
3

01 前言 Front Matter

mcp
{ "args": [ "node_modules/0xray/dist/mcps/orchestrator.server.js" ], "command": "node", "orchestrator": null }
name
orchestrator
tags
[ "orchestration", "orchestrator", "lead-dev" ]
author
Xray Framework
source
framework
version
1.0.0
description
Multi-agent workflow coordination, lead-dev operating model (codex 67-68), and task delegation. Default when suit is worn — no keywords required.
capabilities
[ "coordinate_agents", "delegate_tasks", "manage_workflow" ]
dependencies
[]
schema_version
1.0

03 SKILL.md 正文

Orchestrator Skill — includes lead-dev operating model

When multi_agent_orchestration.lead_dev_mode is true in features.json (default when suit is worn), this is the default OS behavior — not a separate skill or MCP.

Seven rules (codex 67–68)

  1. Phased plan + detailed todos; assign best subagent; monitor output
  2. Lead dev loops test→fix until green; no permission pings
  3. Per-suite test triage after major changes; full suite last
  4. Lead stays main thread; subagents execute; update todos
  5. Read all console and test output; triage fix rerun
  6. Never defer errors as "pre-existing" — add todo and resolve
  7. Resolve all errors before phase completion

MCP loop (one tool surface — no autonomy-intake)

Step MCP tool Notes
Intake + classify analyze-complexity Pass tasks array; persists .xray/state/lead-dev-plan.json when mode is on
Delegate (consult) orchestrate-task Invokes MCP consult skills (code-review, researcher, security-audit, etc.)
Delegate (implement) host Task / spawn_subagent orchestrate-task defers backend-engineer, frontend-engineer, bug-triage — lead must spawn
Monitor get-orchestration-status Lead updates todos
Major work researcher + architect-tools + code-review Auto-listed in plan when complexity > threshold

Subagent routing

Task Subagent
Phasing strategist
Architecture architect-tools
Research researcher
Implementation backend-engineer / frontend-engineer
Test failures bug-triage
Review code-review

Goggles

The command is 0xray look. 0xray goggles is the same organ. A pipe keeps the line card. A terminal shows that card as a pane. Chat seats without hooks use the goggles MCP tools look and status_lens. Same organ.

A session starts on dev. The boot payload includes that card from .xray/state/LENS.md. The planes are dev, integration, domain, and ecosystem. A search inside the current card runs. A search of another plane returns the current card and does not run. A path on no plane runs. A path outside this checkout runs. Opening a file on the current card adds nothing. Opening a file on another plane switches to that card and the open proceeds. A file on no card stays on the current plane and the open proceeds. Compact writes that same card back. A thing on that plane has its own card. One card has its own plate and can name more. A missing plate names the stamp-plate skill. Stamp one plate for that card. The cards and plates that stay are what the next agent reads. Saying goggles clears a stale card line on the station.

Hooks (rewired, not new MCPs)

  • SessionStart → session-start.js boots lead_dev_mode
  • PreToolUse → hints per_suite_triage_required on full npm test

Config: features.json → multi_agent_orchestration.lead_dev_mode

Confer

The TUI that is wearing the suit already spawned the MCP servers. Confer uses those servers. It does not start a private server and it does not run hermes -z.

When .xray/state/confer-ask-s.1.json, confer-ask-s.2.json, and confer-ask-s.3.json are on disk:

  1. Call the server and tool named in each ask. Researcher and code-review return evidence. Architect returns the assessment with Verdict: UNREVIEWED.
  2. Write one verdict line from this TUI's model: Verdict: PASS, Verdict: CONDITIONAL, or Verdict: FAIL, plus top risks and a hardening note.
  3. Call record_confer_receipt on xray-governance with todoId, subagent, sessionId, and that text.

Only Verdict: PASS completes the todo. FAIL, CONDITIONAL, and UNREVIEWED stay open. A session id that does not match the plan does not complete the todo. Grok, OpenCode, Hermes, and OpenClaw use these same three steps. Hermes is a voter only when Hermes is the TUI, and then only as that TUI's own model.

Lead cadence (syncopation)

The seven rules are the engine. The cadence is the time signature. Disk SSOT: grok-bot/OP-PROC.md (team specifics live in house/).

Jargon is not the job. Review and cadence ask whether the change works. Do not spend the loop reciting dest names, mill liturgy, or host mantras.

  • Feat rebase: when the mandated work turns into mantra, or the job has split across planes that no longer share one present, stop. Go up a level and answer the original message again from this seat. Do not continue the side thread and call that the job.

  • Peer test: another instance with only worn 0xRay continues the ticket. No command stream.

  • Wear first: Read .xray/state/STATION.md. Same cloud. Same critic (resume only). Dummy Reads Station.

  • Dummy tests SKILLS.md: a suited dummy proves the worn catalog. Same-sess bodies keep context and hold a Station card.

  • Ship track: always branch + worktree/aside + PR. Docs as you go. Builds. Rollback = close the PR. Subject review. Fix n ship. After critic PASS, review dest/domain, close leftovers, then mill-gate D. PASS is not ship. Groover is not Repertoire.

  • Publish: stay logged in (do not npm logout). Every time run npm publish --access public in the live TTY. Paste the clickable auth/cli URL. No Enter, no VM browser, no subagent, no 6-digit code in chat. Poll npm view until live. Then prove fresh and upgrade registry installs. npx 0xray validate is the wear check (not leftover init.sh). npm view is not an install.

  • Peer boot: four-line dispatch. They Read Station + this cadence and install without a command novel. Not a twin critic.

  • Live tick, not pacer: /loop runs work now. Keep workstreams moving. Do not idle-wait 12 minutes. Clean ticks every cycle — unsubscribe then resubscribe so the prompt matches metal. Name-dedupe does not rewrite. A stale prompt (waiting URL after live, Hold after strip) is dirty: rewrite, do not act. If the card and WAVEBOARD are idle, unsubscribe and do not resubscribe.

  • Board clock: PR events + resume the same critic. Not two idle pacers. Merge after PASS+CI (gate A). Publish only after mill-gate D + foundry gate.

  • Dispatch: ticket + PR URLs + “Wear, then review.” The critic Reads grok-bot/OP-PROC.md (team specifics live in house/). Do not paste a command novel. When a law is spoken twice, encode it this wake.

  • Cascade: Chat is not the brain. A cascade that must survive compaction is a named repertoire signal on the stack overlay, hydrated onto the project law list. NOTES holds the same cascade for the second Read. Station stays the ticket. A reflection is not this capture. Memory scoring rules: docs/memory-scoring.md.

  • Cold brief: a new critic or any new subagent that does not already have this thread gets the latest commit SHA on the branch under review (fetch it; do not reuse an earlier SHA from a previous pass) and the Station card (.xray/state/STATION.md: intent, plan, git). If the card is missing, give branch + HEAD + what just changed. If the branch moves after you brief them, send the new SHA before they verdict.

  • Control: The lead keeps control. Do not assume the subagent has the context. Give the duty, the latest SHA, the Station card, and the laws that apply. The subagent wears the applicable 0xRay suit before it works: the tree under review, built and installed. An older published package is not that suit. The lead checks the result.

Engine cadence

This suite is the engine that builds the engine. When a seat sees OP-PROC, write it onto these existing surfaces in the same wake. Do not open a new skill.

  1. Change the engine with precision.
  2. Ship it local: build and install the tree. A live npm version stays live. Once a version is live, the stamper strips every present-tense patch pin from shipped guides and shipped OP-PROC, or reconcile advances the next cut. The stamper does not edit Station, NOTES, dest, or node_modules. It does not publish.
  3. Test the change.
  4. Run another deep review of the diff you just made.
  5. Monitor long-term memory (dest name count and whether observations are the only growth) and short-term memory (Station: intent, plan, git).
  6. Ship the PR.

Planes

Operating planes: code, OP-PROC, model, suit, mill, host, and test/ship. Test/ship is the critical OP-PROC plane. Test then ship is the hero. Ship without that proof is the catastrophe. Long-term memory is the project law list. Each wake refreshes a changed stack definition and keeps observation stats. Heat writes the matched names onto Station, the short-term card that survives compaction. Reactive decisions move out of TypeScript into OP-PROC as the model improves. That move is why the mill and the suit exist. It is not fully proved. Once a version is live, the stamper strips every present-tense patch pin from shipped guides and shipped OP-PROC, or reconcile advances the next cut. The stamper does not edit Station, NOTES, dest, or node_modules. It does not publish. Do not freeze the live cut.

  • Close: loop until Station is done. Compact: Station survives the cut. Overlay dest names are OP-PROC — reload repertoire-working.json opProcNames, not Station.
  • Board: critic Strict comments; the implementer implements; lead stays main thread.
  • Codex 69: rewire this skill and existing ops. Do not add a new SKILL.md to hold the cadence.
  • Ship matrix: grok-bot/OP-PROC.md Decision matrix (team specifics live in house/) — A PR+CI · B pack · C release:docs-check (reconcile --check; no patch stamps on guides) · D foundry release --i-mean-it. Green CI is not ship. Spend / credentials stay Ask-first. Patch number lives in package.json + CHANGELOG.
  • Metamorphosis: LLM → OP-PROC is in-context worn seats (same critic, same dummy, resume only). Not more bots.

04 同仓库更多 Skills