new-agent

スター7

フォーク2

更新日2026年6月12日 14:17

Create or update a TerMinal agent from a natural-language request — writes a script-first .agents/<id>.sh + <id>.json sidecar. Use on /new-agent, 'create an agent', or wanting a reusable button/scheduled agent.

インストール

Codex または Claude でインストールこの Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。

ソース

trevormil

trevormil/TerMinal

GitHub リポジトリを開く Creator のリポジトリを見る

ダウンロード

関連職種SOC

SOC 職業分類に基づく

その他コンピュータ職コンピュータ・数学職·SOC 15-1299

SKILL.md

readonly

このリポジトリの他の Skills

同じリポジトリ

code-review

trevormil/TerMinal

Compatibility launcher for the code-review agent. Delegates a GitHub/GitLab PR review to Codex: deterministic preflight does recon, then Codex scores six axes into one artifact at .TerMinal/reviews/<pr>/<sha>.md (v2) or .reviews/<pr>/<sha>.md (legacy v1). Use on /code-review or 'review this PR'.

2026-07-187

digest

trevormil/TerMinal

Human-review digest for a PR/MR — chunks the diff into a noise-filtered, risk-ranked, code-first review surface (green/red diffs, 🟢/🟡/🔴, design-decision callouts) WITHOUT abstracting the code away. Deterministic chunker + one bounded codex pass. Sits alongside /code-review (the gate) and /iterate (tests). Use on /digest or 'digest this PR for review'. Renders in TerMinal.

2026-07-187

document-audit

trevormil/TerMinal

Audit docs/ and ~/.claude/docs/ for rot — broken refs, stale runbooks, ADR contradictions, architecture drift, orphan CLAUDE.md. Reports only; never auto-edits. Use on /document-audit or 'check doc health'.

2026-07-187

document

trevormil/TerMinal

Sweep recent changes and propose doc updates (ADRs, runbooks, learnings, architecture, per-folder CLAUDE.md), one candidate at a time. Use on /document, 'capture decisions', or natural post-commit/post-PR pauses.

2026-07-187

emergency-fix

trevormil/TerMinal

FORCE-MODE hotfix: smallest patch → commit → push DIRECTLY to main → follow-up ticket. Needs TERMINAL_FORCE_MAIN=1 (hook-enforced). Use on /emergency-fix or 'this is breaking prod' / 'hotfix' / 'we need to revert now'.

2026-07-187

enqueue-request

trevormil/TerMinal

Enqueue a one-off TerMinal automation-inbox request (can create Activity, tickets, HITL, agent runs, or background tasks). Use on /enqueue-request or to queue a durable local automation request.

2026-07-187

name	new-agent
description	Create or update a TerMinal agent from a natural-language request — writes a script-first .agents/<id>.sh + <id>.json sidecar. Use on /new-agent, 'create an agent', or wanting a reusable button/scheduled agent.

/new-agent — Create a TerMinal agent

Create a repo-local TerMinal agent as a bash script plus JSON sidecar. Agents show up in the TerMinal Agents tab and can be run manually or scheduled.

Target files

Default target is the current repo:

.agents/<id>.sh — executable script body (chmod 755)
.agents/<id>.json — metadata sidecar

If the user explicitly asks for a global agent, write to ~/.config/TerMinal/scripts/<id>.sh and <id>.json instead.

Workflow

Read enough local context before writing:
- CLAUDE.md / AGENTS.md
- .agents/scripts.md if present
- .agents/forge.md if the agent may open a PR/MR
- existing .agents/*.sh and .agents/*.json to avoid duplicate ids
Pick a short kebab-case id, a clear title, a one-line description, and a lucide icon name.
Write the bash script. Prefer deterministic checks first, cap any context passed to an LLM, then call an LLM only when useful.
Write the sidecar JSON.
Verify the script is executable and summarize the paths.

Script contract

Use this shape:

#!/usr/bin/env bash
set -uo pipefail

repo="${TERMINAL_REPO:-$(git rev-parse --show-toplevel)}"
worktree="${TERMINAL_WORKTREE:-$repo}"
engine="${TERMINAL_ENGINE:-claude}"
# Recurring/report agents should default cheap; implementation agents may leave
# this empty so the selected engine uses its normal coding default.
model="${TERMINAL_MODEL:-haiku}"

# Do deterministic work first. For TerMinal helpers:
# terminal-cli ticket "<title>" "<body>"
# terminal-cli hitl "<title>" "<action>"
# terminal-cli activity <kind> "<title>" "<detail>"
# terminal-cli state get|set <key> [value]
# terminal-cli mcp list_agents repo="$(basename "$repo")"
# terminal-cli mcp request_agent_artifact repo="$(basename "$repo")" agentId=knowledge-base title="Question" prompt="..."
#
# Portable repo-local fallbacks available in project-template repos:
# .claude/bin/list-agents
# .claude/bin/request-agent-artifact --agent knowledge-base --title "Question" -- "Prompt..."

When calling an engine from inside the script:

case "$engine" in
  codex) model="${TERMINAL_MODEL:-gpt-5-mini}" ;;
  cursor) model="${TERMINAL_MODEL:-composer-2.5-fast}" ;;
  *) model="${TERMINAL_MODEL:-haiku}" ;;
esac

claude -p "$prompt" --permission-mode auto ${model:+--model "$model"}
codex exec -s danger-full-access -C "$worktree" ${model:+--model "$model"} "$prompt"
cursor-agent -p --force --trust --workspace "$worktree" ${model:+--model "$model"} "$prompt"

Rules:

Cap logs and file lists before interpolation (tail -200, head -100, rg -n -C2 ... | head -160). Prefer letting the engine open a specific path over pasting whole files into the prompt.
Never merge to main/master.
Open a PR/MR only when the script creates concrete code changes.
Use terminal-cli ticket for findings the agent should not fix in the same run.
Use terminal-cli mcp list_agents or .claude/bin/list-agents before filing tickets that need an owner agent.
Use terminal-cli mcp request_agent_artifact or .claude/bin/request-agent-artifact for focused cross-domain knowledge requests. Keep the returned artifact path and a short summary; do not paste long delegated transcripts into the parent prompt.
At the end of implementation-style agents, file follow-up tickets for any deferred cross-agent work. Each follow-up gets exactly one agent_id, agent_scope, and agent_kind; multi-phase work becomes linked tickets.
Use terminal-cli hitl only for true human blockers.
Treat HITL as append-only: do not edit hitl.json or resolve it from the script. After filing, exit only if that lane is blocked; otherwise continue independent work and periodically re-check the blocker or query HITL status.
Exit non-zero on real failure; exit 0 for no-op success.

Sidecar JSON

{
  "id": "kebab-case",
  "title": "Short label",
  "description": "One-line summary",
  "icon": "Bot",
  "opensPr": false,
  "engine": "claude",
  "model": "haiku",
  "modelPolicy": {
    "default": "haiku",
    "cheap": "haiku",
    "deep": "sonnet",
    "judge": "gpt-5-mini",
    "allowOverride": true
  },
  "inPlace": false,
  "outputContract": "Human-readable summary plus ticket/PR/artifact links.",
  "quality": {
    "acceptanceCriteria": [
      "The agent completes its declared task or reports why it could not.",
      "The run ends with checks performed, artifacts produced, and follow-up ticket ids or none."
    ],
    "requiredArtifacts": [],
    "deterministicChecks": [],
    "judge": {
      "enabled": false,
      "mode": "deterministic",
      "rubric": [
        "Output matches the agent purpose.",
        "No unrelated changes were made."
      ]
    }
  }
}

Set opensPr: true only if the agent is expected to commit and open a PR/MR. engine may be "claude", "codex", or "cursor"; use the user's selected engine unless the agent has a clear reason to prefer one. For recurring report/precheck agents, prefer cheap model defaults (haiku, composer-2.5-fast, or a small Codex model such as gpt-5-mini). Leave model empty only for implementation-heavy agents where the engine's coding default is intentional. Set inPlace: true only for agents that intentionally mutate the current repo without a worktree; most agents should leave it false. Set quality.deterministicChecks for commands that can verify output mechanically. Enable quality.judge only when deterministic verification is not enough; include a concrete rubric and judge model in modelPolicy.judge.

Confirmation

End with the created/updated script path, sidecar path, and how to run it from TerMinal.