SuperAPI AI Directory

verasic-init

Deployment & operations

View repository

Updated 10-16 01:16

About this skill

Security: see references/scanner-notes.md and upstream SECURITY.md for expected scanner signals and trust model.

Verasic Init — Repo Setup Orchestrator

Source of truth: references/init-protocol.md (wire contract, statuses, extension guide). Profiles: references/install-profiles.md.

Orchestration (Cursor)

Plan and apply Verasic setup for this repository (confirm-first).

Step 1 — plan (no changes): run from the repository root:

bash .cursor/skills/verasic-init/scripts/init.sh   # append flags the user gave, e.g. --profile agent or --list

If the user already said --yes or --confirm, skip to step 2 with the same flags.

Step 2 — after confirmation: run again with --yes and the agreed profile:

bash .cursor/skills/verasic-init/scripts/init.sh --yes --profile cursor   # or agent / cursor-hybrid / auto

Then:

  1. Relay each report verbatim in a code block — full stdout from the first ──── rule to the last. Do not summarize, soften, or reformat. If init fails before printing a report, relay stderr instead.
  2. On the plan pass, explain the detected profile, checklist gaps, and usage section; ask which profile to apply before running --yes.
  3. If any row says action needed, walk the user through the manual steps from details, wait for them to finish, then re-run --yes to confirm.
  4. When governance posture shows hard-eligible, recommend hard apply via OpenTofu — never auto-apply; wait for explicit user approval.
  5. If verasic-github-cli-init was wired, remind them: create a fine-grained PAT, put it in .github-agent.local (chmod 600), verify with check-gh.sh.

Profiles: cursor (full Cursor), agent (skills.sh / Claude Code / Codex / Kiro / …), cursor-hybrid (skills in .agents/skills/ + Cursor slash UX). Spec: .cursor/skills/verasic-init/references/install-profiles.md

Never run gh auth login; never edit lefthook/husky configs yourself — the report prints snippets for the user instead.

Workflow

  1. Plan first (default) — from the repo root, run without --yes:
bash .cursor/skills/verasic-init/scripts/init.sh
bash .cursor/skills/verasic-init/scripts/init.sh --profile agent   # optional profile hint

Installed under a different root (e.g. .agents/skills/)? Adjust the path prefix. Append flags the user gave (--list, --check-updates, …).

  1. Relay the plan verbatim — print init's full stdout in a code block, unmodified. Explain the detected profile, scope, checklist gaps, and usage section; ask the user which profile to apply (cursor, agent, cursor-hybrid) before mutating anything.

  2. Apply after confirmation — only when the user agrees:

bash .cursor/skills/verasic-init/scripts/init.sh --yes --profile cursor

Relay that report verbatim too.

  1. If any row says action needed, walk the user through manual steps in details, wait, then re-run with --yes.
  2. If verasic-github-cli-init was wired, remind: create PAT, set GH_TOKEN in .github-agent.local, verify with check-gh.sh.

Profiles

Profile Who
cursor Full Cursor (setup.sh or skills under .cursor/skills/)
agent skills.sh, Claude Code, Codex, Kiro, Windsurf, …
cursor-hybrid npx skills add (.agents/skills/) + Cursor slash UX

Aliases: --cursor, --agent, --cursor-hybrid. Default auto-detects from repo layout.

Upstream Cursor UX (v0.2.4+): on --yes, cursor / cursor-hybrid fetch from upstream main/cursor/ unless the user set VERASIC_INIT_BUNDLE_TAG or VERASIC_INIT_REMOTE_REPO_BASE. Plan report ux upstream is the fetch base; latest bundle is informational only (not the fetch default). Do not treat a missing or old bundle tag as a fetch failure by itself.

Inspect and cherry-pick

bash .cursor/skills/verasic-init/scripts/init.sh --list
bash .cursor/skills/verasic-init/scripts/init.sh --yes --skills verasic-github-cli-init,verasic-bugbot
bash .cursor/skills/verasic-init/scripts/init.sh --yes --verify --profile cursor
bash .cursor/skills/verasic-init/scripts/init.sh --check-updates

Hard rules

  • Never run --yes without user confirmation — default plan is the safe path.
  • Relay reports verbatim; they are the user-facing deliverable.
  • Never run gh auth login or edit lefthook/husky config — wire scripts print snippets instead.
  • Init must run inside a git repository; --list and plan mode change nothing.
  • Repo-local skills only — external invoker paths get a warning, not silent cross-repo wiring.