Roadmap
Automaintainer is the autonomous improvement platform for code repositories: a fleet of AI agents that runs a closed, verified loop — build → review → fix → re-verify — on your repos and opens merge-ready PRs you can trust. All on your own provider keys (BYOK), with no human in the per-step loop.
The wedge is maintenance, not feature autopilot. We win on trust and cadence — bounded, verifiable work that gets better every run.
Where we are today
The core loop is live in production on real repositories. Every run is:
- Correct — the Verification Gate (build → test → lint → fix → re-verify) prevents unverified work from shipping.
- On-objective — the Intent Gate reviews each PR against your stated goal before it opens.
- Cumulative — the Objective Ledger decomposes broad goals into bounded sub-tasks, tracks progress across runs, and stops when acceptance passes.
Beyond the core loop, AM now reasons about your architecture direction — not just whether CI passes. The architecture-aware review milestone checks every diff against your ADRs and style guides, flags drift, and respects no-touch zones.
What's shipped
| Capability | Description |
|---|---|
| Verification Gate | Build → verify → fix → re-verify loop. Draft PR on budget cap. |
| Playbooks | Curated, pre-gated maintenance loops (coverage, deps, TODO burn-down, lint). |
| Explore mode | Agent scouts your repo, proposes improvements, and runs the best ones — token-capped, gate-guarded. |
| Issue solving | Pick a GitHub issue, get a fix PR — no human in the loop. |
| PR watcher | Verify external PRs, comment results, auto-merge on Max tier. |
| Issue scout | Audit your repo and file GitHub issues for what AM finds. |
| Objective ledger | Decompose, track, and stop on cumulative progress toward a goal. |
| Run scorecards | Verify iterations, first-try pass rate — per run and over time. |
| Per-repo memory | INVARIANTs, LESSONs, PATTERNs that survive worker restarts. |
| Architecture-aware review | Conformance critic, no-touch zones, ADR auto-ingestion. |
| Cost per run | Token and cost tracked per run, visible in the run drawer. |
| Enterprise SSO + SCIM | OIDC, SAML 2.0, user provisioning — deployed. |
| At-rest encryption | AES-256-GCM for secrets, admin backfill tooling. |
| Worker self-update | Git-tag versioning, panel-triggered rollout. |
What's next
Phase 1 — Observability & cost (in progress)
Operators can answer "what happened, what did it cost, is the fleet healthy?" without SSH-ing into workers.
- Cost aggregation & export — filter by date range and repo, aggregate across runs, export CSV.
- Live run log streaming — WebSocket-powered, no manual refresh.
- Slack / Discord notifications — failures, budget alerts, verify failures.
- Prometheus
/metricsendpoint — fleet observability. - Budget enforcement — set spend caps, get alerted before they hit.
- Pagination — runs and repos lists at scale.
Phase 2 — UX & operator ergonomics (upcoming)
Day-to-day use of the panel feels complete — no "power user via API/SSH" gaps.
- Explore proposals UI — accept, reject, or taste proposed improvements visually.
- Worker management tab — disconnect, probe history, version drift at a glance.
- Resume interrupted runs — re-spawn a run that was cut short.
- Dashboard polish — smooth rough edges in the main dashboard.
- Per-repo model config — pick models per repository from the panel UI.
Phase 3 — Enterprise & platform API (scoping)
Teams and integrators adopt AM without bespoke ops.
- Webhook-triggered runs — event-driven maintenance (push, schedule, CI callback).
- Audit logs — all actions logged, exportable for compliance.
- Platform API expansion — richer
am-cloud platformcommands for external orchestrators. - Multi-tenant isolation hardening — as customer count grows.
This page reflects the product direction at a high level — we don't publish detailed internal plans. Track ship announcements and changelogs on the GitHub releases page.