Skip to content

ROADMAP.md

Latest commit

 

History

History
106 lines (87 loc) · 5.54 KB

ROADMAP.md

File metadata and controls

106 lines (87 loc) · 5.54 KB
Raw

Vejas — Roadmap

Phases are increments, each validated live and end-to-end before the next. The governing discipline (VISION.md): distribution and a real, deployable demo come before the next feature — the failure mode to avoid is building the whole platform before a single user touches it.

Phase 0 — Language & runtime · DONE

  • VejasScript: lexer, parser, in-process interpreter (ADR-0001).
  • NATS JetStream, at-least-once, hot-reload (ADR-0002).
  • Business surface: literals extracted by AST, corrected in place, sample runs (ADR-0005).
  • Panel: pipeline graph, prompt→flow, script & input editing.
  • All-Rust, Python removed; one binary (ADR-0009).

Phase 1 — Composition, packages, MCP · DONE

  • Service composition, pipeline merge; pkg:service + EXPORTS private-by-default (ADR-0003, ADR-0004).
  • The runtime is its own MCP server; flow-as-tool (ADR-0006).
  • Golden-test runner + the language unit-test suite (ADR-0010); both have kept growing since.

Phase 2 — Connectors & secrets · DONE

  1. Connector SDK (ADR-0007): the Driver trait, Source/Sink families, declarative .vjs instance manifests, supervision like flows. Bundled drivers from http-in to the generic oauth-poll, plus the exec bridges (any-language over stdio, ADR-0011) — including exec-stream-source and exec-rpc, which carry the SAP connector (native Rust over libsapnwrfc, ADR-0014) and the Salesforce connector (Bulk API 2.0).
  2. Secrets/Vault (ADR-0008): SecretStore trait (Vault / file / env), secret("path") builtin, fail-closed, write-only from the panel; the panel shows references and resolve status, never values.
  3. Connector-by-prompt: vejas_new_connector + POST /connectors/new.

Exit criterion met: a real webhook → flow → sink with the credential in a vault, from docker compose.

Phase 3 — Surfaces & administration · DONE (first pass)

  • Flow-as-API — built: api "VERB /path" + respond, served under /api, OpenAPI generated at /api/openapi.json (ADR-0006).
  • Panel administration — built: connector cards (status, last error, test probe), secrets card (references + resolve status, write-only set), trace feed with sink responses.
  • Shadow-replay & approval (ADR-0005) — built (lite): propose a correction → replay the last real events from the trace ring → before/after diff → promote or discard (panel + MCP). Later: JetStream-hydrated history, audit trail, one-click rollback.

Phase 4 — Distribution & durability · IN PROGRESS

  • Remote control plane for outbound-only collectors (ADR-0013, CONTROL.md) — v1 built: NATS leaf-node uplink, closed command allowlist, status push, dual audit. v2 (content changes as locally-approved proposals — the ADR-0005 loop applied to fleet management) is specified, not built.
  • Package/connector distribution as git repos; the seam toward a marketplace (ADR-0003) — a candidate monetization surface (the paid panel: collaboration, approvals, audit, SSO).
  • vjs-test as the CI gate — built (.github/workflows/ci.yml: unit + golden + parse-check of every script and example + image build). Transport-level tests (ordering, redelivery + poison→DLQ cap, no-loss under kill -9, reconnection, anti-zombie shutdown) — built (e2e/transport/run.sh, in CI), asserting the at-least-once / publish-before- ack / anti-zombie invariants (ADR-0002) against a live nats+runtime.

Phase 5 — Operator credibility · IN PROGRESS

What turns "nice project" into "I can run this in production":

  1. Persistent dead letters · built (ADR-0015): a poisoned message stops being dropped — it lands on a dedicated DLQ stream with a death envelope (subject, unit, attempts, last error, payload), visible in the panel, re-injected explicitly after the fix. The sister loop of ADR-0005.
  2. Real observability · built (ADR-0016): OTLP/HTTP-JSON traces + a Prometheus /metrics endpoint, hand-rolled to keep the footprint. The manifesto claims it again now that it ships.
  3. Full shadow-replay · built (ADR-0018): JetStream-hydrated history, promote audit trail (git + VEJAS_AUDIT), forward-only rollback.
  4. Published, reproducible benchmarks: throughput, memory, image size, cold start — against the incumbents, same scenario. (In progress.)
  5. Rules-view (N2) · built (ADR-0019): a read-only projection of a flow's if/elif/else into readable sentences with the surface literals editable inline; a composed guard is shown verbatim and routed to the agent (N3). The rule-editing doctrine (N1 parameters / N2 reading / N3 structure, no graphical rule editor) is settled there.

Cross-cutting, ongoing

  • Keep README/MANIFESTO aligned with what is actually built — no promise the code cannot keep.
  • Purge Python-era JetStream durable consumers.
  • Publish: vejas.dev (canonical) / vejas.io are registered and the repo exists; the public flip ships with the recorded demo, not before.

How to pick up this project (for another agent)

Read VISION.md, then ARCHITECTURE.md, then the ADRs in order. The code is core/src/{vjs,main,connectors,secrets,control}.rs (~7k lines, plus the SAP and Salesforce connector crates under connectors/). Run cargo test --manifest-path core/Cargo.toml and vejas-runtime vjs-test tests/vjs to see the contract. Everything the platform can do is reachable over POST /mcp (docs/MCP.md). Do not add a transform without the ADR-0010 admission test; do not put a secret in a literal (ADR-0008); do not reintroduce Python (ADR-0009).