Menu
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.
Installer avec Codex ou Claude Copiez ce prompt, collez-le dans Codex, Claude ou un autre assistant, puis laissez-le vérifier la page du skill et l'installer pour vous.
Basé sur la classification professionnelle SOC
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'.
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.
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'.
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.
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'.
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.
| 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. |
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.
Default target is the current repo:
.agents/<id>.sh — executable script body (chmod 755).agents/<id>.json — metadata sidecarIf the user explicitly asks for a global agent, write to
~/.config/TerMinal/scripts/<id>.sh and <id>.json instead.
CLAUDE.md / AGENTS.md.agents/scripts.md if present.agents/forge.md if the agent may open a PR/MR.agents/*.sh and .agents/*.json to avoid duplicate idsid, a clear title, a one-line description, and
a lucide icon name.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:
tail -200, head -100,
rg -n -C2 ... | head -160). Prefer letting the engine open a specific path
over pasting whole files into the prompt.main/master.terminal-cli ticket for findings the agent should not fix in the same run.terminal-cli mcp list_agents or .claude/bin/list-agents before filing
tickets that need an owner agent.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.agent_id,
agent_scope, and agent_kind; multi-phase work becomes linked tickets.terminal-cli hitl only for true human blockers.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 0 for no-op success.{
"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.
End with the created/updated script path, sidecar path, and how to run it from TerMinal.