install-agent

name	install-agent
description	Install the Sepo agent infrastructure into another GitHub repository through a focused pull request. Use when asked to add, bootstrap, or onboard this agent backend while auditing conflicts, preserving repo-owned files, validating the diff, and guiding post-merge memory/rubrics setup.

Install Agent

Use this skill to add the Sepo agent backend to an existing repository by opening a normal PR. Keep the target repository's application code and unrelated GitHub assets intact.

For productized public cross-repository install PRs from self-evolving/repo, prefer @sepo-agent /install .... This skill remains available for manual or custom installation flows where the requester wants skill-guided adaptation.

Inputs

Confirm these before editing:

• target repository slug and default branch
• source agent repo/ref, defaulting to the current checkout
• install branch name, defaulting to agent/install-agent-infra
• preferred GitHub auth path: hosted app/OIDC, bring-your-own GitHub App, AGENT_PAT, or fallback workflow token
• model provider secret plan: OPENAI_API_KEY, CLAUDE_CODE_OAUTH_TOKEN, or both
• whether to copy any .skills/ directories; default is no
• whether to copy root AGENTS.md; default is no

Stop if the target repo, branch, or source revision is ambiguous.

Agent-Owned Scope

Install only this agent infrastructure:

• .agent/, excluding generated/dependency directories
• the source repo's .github/ agent assets as a set, merged into the target .github/ without deleting target-only content:
  • workflows, actions, prompts, and scripts/functions under source .github/
  • source prompt .md files copied by file name while preserving target-only prompts
  • existing target workflows, actions, prompts, scripts/functions, or other .github files preserved unless explicitly approved for replacement
• optional .skills/<requested-skill>/SKILL.md, only with explicit approval
• optional AGENTS.md, only with explicit approval

Do not install or overwrite target application source, repository secrets, branch protection, target-owned .github functionality, or the target root README.md unless explicitly requested.

Guardrails

• Never push directly to the target default branch or merge the PR unless asked.
• Never overwrite divergent files or remove target-only files without showing an audit and getting confirmation.
• Treat target .github/ content and .skills/ as shared/user-owned namespaces.
• Do not use rsync --delete on .github/, .github/prompts/, or .skills/. Delete target .github or skill files only after explicit file-by-file approval.
• Do not copy generated or local-only files: node_modules/, dist/, .agent/node_modules/, .agent/dist/, .git/, _worktrees/, or .claude/worktrees/.
• Merge agent-generated-output ignore rules into the target's existing .gitignore instead of replacing it: at least .agent/dist/ and .agent/node_modules/.
• Do not ask users to paste secrets into issues, PRs, comments, or committed files. Direct them to GitHub settings or gh secret set in a trusted shell.

Workflow

1. Read source setup docs first:

   • .agent/docs/setup/install-existing-repository.md
   • .agent/docs/setup/setup-guide.md

2. Prepare the target checkout.

   • Clone/open the target repo.
   • Create the install branch from the target default branch.
   • Check git status --short; stop if unrelated local changes would make the install ambiguous.

3. Audit before copying.

   • Inspect every path in Agent-Owned Scope.
   • Check related branches/refs: git ls-remote --heads origin agent/memory agent/rubrics 'agent/*'.
   • Check existing secrets/variables/workflows where permissions allow: gh secret list, gh variable list, and gh workflow list.
   • Look for legacy or overlapping scaffolds such as .flows/, .claude/, old run-claude-task / run-codex-task actions, or stale agent-* workflows.
   • Classify overlaps as absent, identical, agent-owned divergent, repo-owned/custom, or legacy removal candidate.
   • Present the audit in a small table and ask before replacing divergent files or removing anything.

4. Copy the install files conservatively.

   • Sync .agent/ with generated/dependency exclusions.
   • Preserve target-owned root files, but add the agent ignore entries .agent/dist/ and .agent/node_modules/ to the target .gitignore if missing so GitHub-hosted runner builds do not leak generated files into PRs.
   • Copy all source .github/ files/directories into the target .github/ by default, but merge rather than wholesale replace.
   • Preserve target-only .github files and replace existing .github files only when they are identical, clearly agent-owned, or explicitly approved.
   • Copy source prompt .md files without deleting target-only prompts.
   • Copy requested .skills/ directories and AGENTS.md only when approved.

5. Configure target repository guidance.

   • Do not commit secret values.
   • In the PR body, tell the owner to configure at least one provider secret: OPENAI_API_KEY and/or CLAUDE_CODE_OAUTH_TOKEN.
   • Summarize auth options from the setup guide: hosted app/OIDC, BYO GitHub App via AGENT_APP_ID + AGENT_APP_PRIVATE_KEY, AGENT_PAT, or fallback workflow token.
   • Mention useful optional variables such as AGENT_HANDLE, AGENT_RUNS_ON, AGENT_ACCESS_POLICY, AGENT_MEMORY_POLICY, AGENT_RUBRICS_POLICY, AGENT_SESSION_BUNDLE_MODE, and AGENT_STATUS_LABEL_ENABLED.

6. Review and validate before commit.

   • Confirm the diff is limited to approved agent infrastructure: git status --short, git diff --stat, and git diff -- .agent .github .skills AGENTS.md.
   • Run whitespace/staged checks: git diff --check, then stage intended files, then git diff --cached --check.
   • Because this install copies runtime code and workflow YAML, run when the target environment permits: npm --prefix .agent ci, npm --prefix .agent run build, npm --prefix .agent test, and YAML parsing for .github/workflows/*.yml.
   • If a check cannot run, report exactly why it was skipped.

7. Commit, push, and open the PR.

   • Commit message: chore: install Sepo agent infrastructure.
   • Stage only intended files, typically .agent .github, plus approved .skills/<requested-skill> and/or AGENTS.md.
   • PR body should start with summary, then Required setup after merge, then source revision, installed files, conflict audit, preserved/skipped files, validation results, and any source request link.
   • In Required setup after merge, use target-specific links for the Sepo GitHub App install page, target Actions secrets, onboarding check workflow, memory initialization workflow, rubrics initialization workflow, and setup guide where possible.
   • Do not merge the PR unless explicitly asked.

Post-Merge Guidance

After the install PR merges:

• Verify GitHub Actions are enabled and required secrets/auth settings are set.
• Check for existing branches before initializing: git ls-remote --heads https://github.com/<owner>/<repo>.git agent/memory agent/rubrics.
• If agent/memory is missing, run or ask the owner to run Agent / Memory / Initialization; do not reinitialize an existing branch.
• If the repo wants rubric steering and agent/rubrics is missing, run or ask the owner to run Agent / Rubrics / Initialization; do not reinitialize an existing branch.
• Open a smoke-test issue mentioning @sepo-agent /answer or the configured AGENT_HANDLE.
• After branches exist, use ongoing memory workflows and Agent / Rubrics / Update for future learning.

Final Response

Report the target repo/branch, source revision, PR URL or reason none was opened, audit outcome, files installed/skipped/preserved/removed, validation results including skipped checks, required setup still pending, and post-merge memory/rubrics next steps.