Skip to content

Latest commit

 

History

History
111 lines (88 loc) · 5.23 KB

RELEASE_NOTES.md

File metadata and controls

111 lines (88 loc) · 5.23 KB

Release Notes

v0.1.0-rc.1 — first release candidate

The first candidate cut of BackAI. It makes the hero flow real end to end and lands the P0 (guardrails) and P1 (security) work. It is a candidate, not a final release: the "known deferred" items below are tracked and gate the final v0.1.0.

The hero flow — now real end to end

af-stack init --template coding-agent scaffolds a multi-tenant coding-agent app; sign-up mints a tenant + API key; a submitted task drives the shipwright agent to clone the repo, run a coding harness (or an honest file-edit fallback when no harness binary is present), push a branch, and open a genuine pull request via the GitHub REST API; the run is metered as a per-tenant cost row. No simulated steps, no hardcoded diff. See examples/02-shipwright/.

Auth to GitHub is a tenant-supplied GH_TOKEN secret — the v1 path. The GitHub-OAuth "connect a repo" UX is explicitly backlog.

Security (P1)

  • S1 / S1b — operator-gated the DB-studio and adapter-registry surfaces, and closed 5 unauthenticated cross-tenant read leaks (logs, secrets, runs, crons, skills).
  • S2 — runtime refuses to serve as an RLS-bypassing DB role; compose provisions a restricted serving role.
  • S3 — hardened Docker sandbox defaults (cap-drop, no-new-privileges, pid limits).
  • S4 — sandbox network=restricted is rejected rather than silently allowed.
  • S5 — a default per-tenant monthly spend ceiling is wired at boot.
  • S6 — KMS preflight: the runtime fails loudly when a configured KMS can't load instead of degrading silently.
  • S8 — object storage is tenant-scoped, closing a cross-tenant leak.

Guardrails & CI (P0)

  • CI runs on push/PR; golangci-lint v2 restored as a blocking gate.
  • G2 — scripts/acceptance.sh, the hero-flow release tripwire.
  • G3 — truth-gate scaffolding in the skill verifier.

DX & deploy

  • A1 — canonical error envelope + JSON 404 for agent errors.
  • A3 — adapter boot self-check (fails fast on an unknown adapter) + a wired remote sandbox adapter.
  • X1 / X2 — SKILL.md reframed CLI/API-first; the scaffold emits a root CLAUDE.md + .mcp.json for agent discovery.
  • D2 — deploy hardening: dropped the broken Fly release_command, pinned prod-compose images off :latest, aligned the dashboard build to Node 20.
  • R1 — docs truth pass: real adapter sets, the GH_TOKEN/connect-backlog story.

Release-blockers caught by the R2 acceptance run

Running scripts/acceptance.sh against a fresh stack surfaced two first-boot defects that CI structurally cannot catch (CI's Go tests run with no Postgres, and the scaffold is only exercised via unit stubs):

  • First-boot migration crash. 00027_block1_admin_endpoints.sql had a do $$ … end $$; block without goose statement annotations, so the runner split it on an inner ; → unterminated dollar-quoted string → the runtime crash-looped and never reached /ready. Fixed by wrapping the block in -- +goose StatementBegin/StatementEnd.
  • af-stack init aborted on a fresh clone. init ran pnpm run generate:brand as a fatal step before scaffolding, so a clone without node_modules (the documented DX) aborted before the coding agent was written. Brand regeneration is now a warning — brand.yaml is the source of truth and CSS regenerates at build time — so the hero scaffold always lands.

Truth gates — enforced vs. stubbed

The skill verifier (skills/af-stack/meta/verify-skill.mjs) carries three truth gates. With ENFORCE_TRUTH_GATES=1:

  • no-stub-agents — green / enforce-ready. The hero agent is real.
  • openapi-parity — stubbed. 88 served-but-undocumented routes; enforcing is blocked until the OpenAPI spec catches up to the served surface.
  • sdk-parity — stubbed. 6 namespace diffs between the TS and Python SDKs (tracked as A2); enforcing is blocked until parity lands.

Gates flip to enforcing per-gate as the underlying reality becomes true, so the default is opt-in until then.

Known deferred (not in rc.1 — gate the final v0.1.0)

  • A2 — SDK parity across the 6 diverging namespaces (sdk-parity gate stays stubbed until closed).
  • openapi-parity — 88 served routes to document (openapi-parity gate stays stubbed until the spec catches up).
  • D1 — the customer app on all deploy targets (Helm / Fly / Render / prod-compose / Caddy); needs a domain-topology decision.
  • S1b (cost + billing) — 2 remaining principal-scoped read surfaces; needs an auth-model decision (do operators carry a tenant membership?).
  • H4 — publish @af-stack/sdk to npm; needs npm credentials.
  • GitHub connect OAuth UX — backlog; GH_TOKEN secret is the v1 path.

Running the acceptance tripwire

./scripts/acceptance.sh

Phases 1–5 (boot → runtime ready → multi-tenancy on → scaffold → agent registered) enforce today. Phases 6–8 (tenant + API key → real PR → metered cost row) require OPERATOR_TOKEN, GH_TOKEN, and ACCEPT_REPO (a throwaway repo); they SKIP until those are supplied. The tripwire is fully green only when the entire hero flow is real with credentials present.