migrate-agents

name	migrate-agents
description	Convert .agents/agents.json entries into the script-first shape — one <id>.sh + <id>.json per agent. Non-destructive (agents.json untouched). Use on /migrate-agents or to convert prompt-style agents into bash scripts.

/migrate-agents — bulk-convert agents.json → scripts

Companion to the in-app "Convert to script" button, but at directory level. For each agent entry in .agents/agents.json that does not already have a matching .agents/<id>.sh, materialize:

• .agents/<id>.sh — executable bash body that execs the prior claude -p $prompt / codex exec $prompt command. Default body is the single-LLM-call template — replace it with prechecks + escalation when you want to save tokens (see .agents/scripts.md and .agents/health.sh).
• .agents/<id>.json — sidecar metadata (title, description, icon, opensPr, engine, model, inPlace) — exactly what was previously the agents.json entry, minus prompt.

Hard rules

1. Non-destructive. Do NOT delete entries from .agents/agents.json. The runner branches to .sh when present and falls back to the prompt entry otherwise — keeping both makes the migration safe to revert.
2. Skip on conflict. If .agents/<id>.sh already exists, leave it alone and skip that agent (do not overwrite hand-edited scripts).
3. chmod 755 every script written.
4. Sidecar JSON, 2-space indent, trailing newline.

Process

set -euo pipefail

# Find the agents.json (no-op if missing)
json=".agents/agents.json"
[ -f "$json" ] || { echo "no .agents/agents.json — nothing to migrate"; exit 0; }

# For each agent entry, generate the .sh and sidecar .json.
# Iterate via jq so the prompt is properly extracted (handles newlines, quotes).
ids=$(jq -r '.[] | .id' "$json")
for id in $ids; do
  script=".agents/$id.sh"
  if [ -f "$script" ]; then
    echo "skip $id — script already exists"
    continue
  fi

  prompt=$(jq -r --arg id "$id" '.[] | select(.id == $id) | .prompt' "$json")
  engine=$(jq -r --arg id "$id" '.[] | select(.id == $id) | (.engine // "claude")' "$json")
  model=$(jq -r --arg id "$id" '.[] | select(.id == $id) | (.model // "")' "$json")

  # Build the exec line with TERMINAL_MODEL override
  if [ "$engine" = "claude" ]; then
    cmd="claude -p $(printf %q "$prompt") --permission-mode auto \${TERMINAL_MODEL:+--model \"\$TERMINAL_MODEL\"}"
  else
    cmd="codex exec -s danger-full-access -C \"\$TERMINAL_WORKTREE\" \${TERMINAL_MODEL:+--model \"\$TERMINAL_MODEL\"} $(printf %q "$prompt")"
  fi

  cat > "$script" <<EOF
#!/usr/bin/env bash
# Auto-generated by /migrate-agents from .agents/agents.json.
# Edit freely. Runtime env: TERMINAL_REPO, TERMINAL_RUN_ID, TERMINAL_BRANCH,
# TERMINAL_WORKTREE, TERMINAL_ENGINE, TERMINAL_MODEL. Helpers on PATH:
# terminal-cli ticket / hitl / activity / notify.

set -uo pipefail

exec $cmd
EOF
  chmod 755 "$script"

  # Sidecar metadata (everything from agents.json minus the prompt).
  jq --arg id "$id" '.[] | select(.id == $id) | del(.prompt)' "$json" > ".agents/$id.json"
  echo "migrated $id"
done

After running, restart TerMinal (or just open the Agents tab) and each row will render the script body in its expander with an "edit" button — replace the auto-generated single-LLM-call with a precheck-then-escalate pattern when it'd save tokens.

What this is NOT

• Not the agent designer. The designer creates new agents from a natural-language description; this skill converts existing ones in place.
• Not a sidecar editor. Once the .sh + .json exist, edit them with your configured editor (the Agents tab's "edit" button opens the .sh in Cursor/VS Code; the sidecar is plain JSON).