Roadmap & testing

What shipped in the v0.3 foundation cut, what is intentionally next, and how the claude-session flows are tested.

v0.3 — foundation, done & in prod#

The runtime trio (Go API · Go daemon · Python sidecar) is correctness-closed and proven under the deck's own 8×24 load in CI:

R1 — the Python sidecar actually runs in real installs (was silently n/a).
R2.1–2.3 — lossless events: idempotent ingest → durable spool → API ack (true at-least-once). The "frozen dashboard" class is structurally closed.
R3.1/3.2 — crash-safe daemon state (fsync) + orphaned cu/SDK sessions reconciled on restart.
R4.1/4.2 — written perf budgets + a CI load harness that hard-gates zero event loss at 8 machines × 24 sessions (192 sessions).
CDN no-store on /healthz + /dl/version so deploy-monitoring and cu self-update --check can't read a stale version.

What's next#

Founder-owned, no engineering blockers: the public OSS push (tooling + a scan-clean export tree are ready) and recording the real-product killer demo.

Honestly-NEXT engineering, scoped not rushed:

Full multi-process WS fan-out harness + daemon idle CPU/MEM budget (R4.2 covered the ingest+read leg).
Transactional cu self-update with rollback.
Prod-verify the policy engine + per-session sandbox, then add them to the demo.

The full plan with deck-phase mapping lives in the repo at docs/v0.3/whats-next.md — the canonical source of truth (per AGENTS.md). This page is its reader-facing summary.

Testing the claude-session flows#

Every v0.3 dashboard flow has a formal browser QA scenario (stable id, exact steps, pass criteria, artefact) in docs/v0.3/qa-scenarios.md: discovery, live-stream-no-freeze, prompt-reaches-PTY (R2.3), kill-API-no-gap (R2), restart-reaps-orphans (R3), and never-stale version.

Automated end-to-end coverage is Playwright through the browser path (tests/e2e/qa-*.spec.ts). New for v0.3: qa-r23-prompt-delivery-001 drives the prompt textbox in a real session and asserts the prompt round-trips through the PTY. Scenarios stay 🟡 MANUAL until their spec is green against pnpm dev — the ≥98% coverage number is computed only from verified-AUTOMATED specs, so it stays truthful.