AGENT SKILL

fleet

Orchestrate bounded GPT-only Codex fleets across Sol, Terra, and Luna Max lanes. Use when a user asks for Fleet, parallel agents, deep multi-agent work, broad classification, independent verification, or a cron-friendly batch that must use ChatGPT-plan Codex authentication without third-party providers.

License 许可
MIT
所属仓库
Babinci/copernicus
最近更新
2026-09-30
Stars
★ 0
Forks
0

01 前言 Front Matter

name
fleet
description
Orchestrate bounded GPT-only Codex fleets across Sol, Terra, and Luna Max lanes. Use when a user asks for Fleet, parallel agents, deep multi-agent work, broad classification, independent verification, or a cron-friendly batch that must use ChatGPT-plan Codex authentication without third-party providers.

02 资源文件 6

03 SKILL.md 正文

Copernicus Fleet

Route each seat to the cheapest GPT lane that can meet a checkable contract. The lead owns decomposition, width, evidence, and synthesis; workers never create more workers.

Resolve <fleet-skill-dir> to the directory containing this SKILL.md. Resolve all bundled scripts, references, and model policy files from that directory, never from the user's current working directory.

Read guide.md when explaining Fleet, choosing between an interactive mission and a batch, designing a first mission, or changing this skill. It describes the purpose, execution model, outputs, examples, privacy boundary, and limitations.

Open a fleet mission with:

Using Copernicus Fleet: GPT-only.
Mission: <one sentence>.
Execute vs validate: <what runs now>.

Route

Read fleet.yaml before routing. It contains the GPT model bindings, named presets, and the active default_preset. The shipped default is terra-first: an omitted manifest lane resolves to Terra. Use luna-breadth when repeatable width should be the default, or name any lane explicitly to override a preset.

  • Sol (gpt-5.6-sol): ambiguous, high-value synthesis, architecture, and final verification. Use 3-4 seats only for an explicitly deep mission.
  • Terra (gpt-5.6-terra): default everyday implementation, grounded review, mapping, and reducers.
  • Luna Max (gpt-5.6-luna, max): clear, repeatable extraction, classification, and bounded shards. Run waves of at most five.

Read model-policy.md before changing lane names, presets, concurrency, model IDs, or reasoning efforts. If a configured model is unavailable to the signed-in account, report it as unavailable; never fall back to a non-GPT provider. For a different account catalog, copy fleet.yaml, bind only observed GPT IDs, and pass the copy with --config-file.

Execute

  1. Run deterministic commands first. Do not spend a seat on a claim a test, parser, grep, or measurement can settle.
  2. Give every seat a unique role, bounded material, allowed paths, edit policy, typed output, evidence standard, falsifier, and stop rule. State Do not spawn subagents.
  3. Keep verification independent: an author does not certify its own result. Prefer a deterministic gate, then a fresh Terra or Sol review.
  4. Launch only independent seats in parallel. Use Luna waves for breadth, then deduplicate before synthesis.
  5. Treat timeout, quota, auth, and unavailable-model results as missing evidence, never as a vote.
  6. Verify every load-bearing claim against source or a runnable check before persisting it.

Use Codex collaboration tools for interactive missions. For local scheduled or shell-driven batches, use the bundled runner:

python3 <fleet-skill-dir>/scripts/fleet.py list
python3 <fleet-skill-dir>/scripts/fleet.py list --preset luna-breadth
python3 <fleet-skill-dir>/scripts/fleet.py batch manifest.jsonl --run-dir .copernicus/runs/example --workdir . --dry-run
python3 <fleet-skill-dir>/scripts/fleet.py batch manifest.jsonl --run-dir .copernicus/runs/example --workdir .

Manifest rows contain id, prompt, and an optional lane. Omitting lane uses the active preset's default; an explicit lane always wins:

{"id":"review-api","prompt":"prompts/review-api.md"}
{"id":"map-api","lane":"luna","prompt":"prompts/map-api.md"}

The runner is read-only by default, invokes codex exec without a shell, clamps concurrency to lane limits, stores one private answer/events/status directory per seat, and returns nonzero if any seat fails. Use --sandbox workspace-write only for an explicitly authorized implementation batch. It never supports danger-full-access.

Seat contract

ROLE: <one bounded function>. Read-only unless exact writable paths follow. Do not spawn subagents.
GOAL: <user-visible destination>.
SUCCESS: <checkable result and verification>.
MATERIAL: <exact paths or evidence slice>.
CONSTRAINTS: GPT-only; no secrets; no destructive or external writes; cite evidence read.
OUTPUT: <schema or headings>.
FALSIFIER: <observation that disproves the claim>.
STOP: emit UNVERIFIED when evidence is missing; stop after the bounded result.

Hard rules

  • Use only model IDs beginning with gpt-; the runner enforces this again at execution.
  • Use the user's existing Codex sign-in. Never read, copy, print, or commit auth caches, API keys, or tokens.
  • Do not use provider proxies, OpenRouter, or third-party model fallbacks.
  • Do not let workers deploy, purchase, sign, publish, message people, or mutate external systems without the user's explicit authorization.
  • Do not create immortal agents. Durable state belongs in files, tests, receipts, or scheduled-task history.
  • Do not count correlated Luna outputs as independent validation.

Keep seat provenance in the private run trace. In a normal conversational response, lead with the verified outcome and omit the roster. Show a compact roster only when the user requests execution provenance, or when a failed or unavailable seat or material dissent changes confidence. Then give verified claims, rejected or unverified claims, and useful next commands.

Fleet is self-contained: interactive work uses Codex's collaboration surface; batch work uses the bundled Python standard-library runner and the user's existing codex executable. It needs no provider proxy, API-key router, database, daemon, or private project files.

04 同仓库更多 Skills