agentify: install the autonomous-repo feedback loop

[jiashuoz]

Installs the autonomous-repo feedback loop into e2a via /agentify. e2a is the framework's first adopter. This PR is the human gate on the install itself — review the scaffolded files, then merge.

What each file does

autonomous-repo.config.yml — the single product-owned config. Everything else reads from it (repo, marker, labels, comms mailbox, models, fix gate). Key choices:

• fix_gate.mode: auto — triage auto-applies agent-fix for confident clean fixes → the fix lane auto-opens a PR. Merge is still the only ship gate (auto-open, never auto-merge).
• always_hitl tuned for e2a's crown jewels — email-auth (SPF/DKIM/DMARC), HMAC header/webhook signing, SMTP relay/outbound, the OpenAPI/generated-SDK contract, and domain identity. Items touching these always take the email-approval path even in auto mode.
• comms via e2a: support@e2a.dev (intake + acks + approvals); approver jszjosh@gmail.com.

.claude/skills/autonomous-repo/ — the runtime skill the lanes execute (triage / comms / fix / state-machine + email templates). .gitignore was updated to track only this subtree (the lanes read it in CI); other local .claude/ files stay ignored.

.github/workflows/feedback-{triage,comms,fix,released}.yml — the four lanes. Each no-ops loudly until its secrets exist, so merging this PR does not start anything.

scripts/

• ticket_card.sh, comms_send.sh, released_markers.sh — lane helpers (all pass _selftest).
• agentify-verify-setup.sh — fix-lane bootstrap: boots Postgres (:5433) + the migrated e2a_test DB + Mailpit so the fix agent verifies against a real running stack, not just a compile.

tools/submit-feedback-mcp/ — addon: a submit_feedback MCP tool that email-bridges into the support mailbox. Additive; the loop runs without it. Setup in AGENTIFY-ADDON-SETUP.md.

Not done here (one-time human setup — required before lanes run)

Labels are already created. Everything below needs your hands (auth is never run by the skill) — see references/setup-checklist.md:

1. GitHub App (bot identity): Issues/PRs/Contents read-write, no Workflows:write; install on repo; set github_app_login (currently e2a-support-bot[bot] — confirm/replace); add secrets AUTOREPO_APP_ID, AUTOREPO_APP_PRIVATE_KEY.
2. e2a support agent: create the support@e2a.dev agent (verified domain) + an agent-scoped API key → secret E2A_API_KEY; run with screening/HITL off.
3. Model token: claude setup-token → secret CLAUDE_CODE_OAUTH_TOKEN (or ANTHROPIC_API_KEY).
4. Repo settings: enable Actions; branch protection on main requiring the reviewer's review — load-bearing, the bot must not bypass it; (recommended) CODEOWNERS on autonomous-repo.config.yml + .github/.

Pause switch: set repo variable AUTOREPO_LANES_PAUSED=true to halt all lanes.

🤖 Generated with Claude Code

[jiashuoz]

Reverting the agentify install — closing without merge and removing the branch.