Roadmap

The remaining v1 milestones after the M0–M12 arc shipped. Granular per-task tracking lives in workplan.md; this page is the user-facing summary.

M13 — CI + verification pipeline

Get the green-checkmark story honest. The work that has to land before we can publish a release with confidence:

  • GitHub Actions workflows for the Rust workspace (cargo build, cargo test, cargo clippy --all-targets -- -D warnings), pinned to a specific stable Rust toolchain.
  • Workflow for the SwiftUI shell (swift build, run the UnibenchFFITest headless verifier).
  • Vercel project pointed at landing/ for the marketing-site deploy off main. No GitHub Actions workflow needed — Vercel watches the repo directly.
  • Workflow that regenerates docs/reference/manifest.md and asserts no diff — same content as the existing reference_doc_matches_generator test, surfaced at the CI level.
  • Integration smoke that starts the orchestrator, opens python-rag-baseline, drives one IPC call, and shuts down.
  • Live-API verification (gated behind a secret + manually dispatched) for Anthropic / OpenAI when a real DSN exists.

M14 — App Store submission

The work to ship a signed, sandboxed, App-Store-distributable build. This is non-trivial because sandbox restrictions interact with several v1 design choices.

  • shell/Unibench.xcodeproj (migrating off SwiftPM-only for signing + entitlements).
  • App-sandbox entitlements: app-sandbox, network.client, files.user-selected.read-write, files.bookmarks.app-scope.
  • Verify every v1 service type starts under sandbox restrictions (no privileged ports, no system state mutation, no launchd injection).
  • Decide the daemon-process pattern under sandbox (XPC service vs helper bundle) — separate from the current "spawn the unibench-daemon binary" model.
  • Hardened-runtime + notarization + DMG packaging for direct download distribution alongside the App Store path.
  • Privacy declarations + screenshots + review-materials for App Store submission.

M15 — v1 launch

The post-shipping work that closes the v1 box:

  • Public release announcement.
  • "Known issues at launch" page seeded from any escapes during M13's verification arc.
  • Telemetry sampling rate calibration once the first real Sentry data is flowing.
  • Roadmap planning for v1.x patch releases (audit-log-entry surface for the M9 deferred items, per-service version changelog drill-down, agent write-side tools — decision 2.39 phase 2 conversation).

Beyond v1

Items deliberately deferred during v1 that probably ship in v1.x or v2:

  • Snapshot sandbox mode (decision 2.21 v2). Today there are three sandbox modes (Resume / Fresh / Seed); snapshot was intentionally deferred.
  • Per-service "Revert to this version" action (decision 2.45). Needs the audit log's binary-version event stream to be meaningful.
  • Agent write-side tools (decision 2.39 phase 2). Today the agent is strictly read-only. The write-side path needs an audit-log surface for every action + a per-action confirmation pattern.
  • Aggregate agent feedback telemetry (decision 2.40). The thumbs feedback is captured locally; piggy-backing the provider+confidence+reaction signal on Sentry waits for v1 data to calibrate against.
  • Audit-log-entry IPC surface for secrets, recipes, samples. The orchestrator records project + service lifecycle events today; recipe-apply and secret-set events are documented but not yet recorded (deferred from M11 phase 1).

Want to nudge priority?

Open an issue at https://github.com/0xhaz/Unibench/issues. The roadmap above is the current best guess; user feedback shifts priorities in practice.