LexDev

Automium

Agent-native browser QA and automation platform — Playwright-backed runtime with semantic page snapshots, an MCP server for agent control, and a provider-backed test planner; the browser-runtime refactor is the latest workstream

Competes with QA Wolf

BuildingLexDev

Key metrics

Metrics being defined

Competitive Intel

0 entries

No competitive intel entries for this product yet.

Research Hub

0 types

No research documents for this product yet.

Roadmap

123 items

Planned88

`automium run --tag smoke` submits journeys, polls, and exits with correct exit code

`automium validate` reports config and connectivity issues with actionable messages

A journey authored via the API executes end-to-end and produces a pass/fail verdict

Actionability scoring correctly identifies interactive vs non-interactive elements

Actionability scoring correctly identifies interactive vs non-interactive elements *(uses engine-domain scoreActionability)*

Adapter registry resolves Claude and fixture adapters by vendor string

All adapter implementations pass contract tests matching `packages/adapters/` interfaces

All phase tests pass

API requests are traced end-to-end via OpenTelemetry

Artifact upload job processes correctly after a journey run completes

Artifacts are organized in the correct bucket structure with org/workspace scoping

Artifacts are retrievable via the API and respect tenant authorization

At least one journey completes with Claude planner producing real tool-call intents

Audit events flow through the audit-sink queue to Postgres asynchronously

AuditSinkAdapter can emit and query audit events through Postgres

Automium replay event stream is serialized as JSONL and retrievable per run

Benchmark comparison report shows meaningful metrics across fixture and Claude planners

Benchmark runner works with v2 adapters for planner comparison

BrowserRuntime interface is defined and Playwright adapter passes all interface methods

BrowserRuntime interface is defined and Playwright adapter passes all interface methods *(stub shape validated; real Playwright wiring deferred)*

Can navigate an owned benchmark product URL and produce a complete enriched snapshot

Can navigate an owned benchmark product URL and produce a complete enriched snapshot *(contract-compliant snapshot shape validated; real navigation deferred)*

CDP pipeline captures network requests, console messages, and DOM mutations in real-time

CDP pipeline captures network requests, console messages, and DOM mutations in real-time *(event types defined; real CDP subscription deferred)*

Claude adapter parses tool-call responses into PlannerIntent objects

Claude adapter produces valid tool-call prompts from enriched semantic snapshots

CLI can submit, monitor, and report on the end-to-end run with correct exit codes

CLI has zero Playwright/engine/provider SDK dependencies

Cold start meets <5s target, full 10-step journey meets <30s target

Control plane API serves artifact downloads for authenticated, authorized users

Credential CRUD endpoint encrypts values and never returns plaintext

Credential vault can store, retrieve, and rotate encrypted secrets

Drizzle schema compiles and generates migrations for all core tables

Executor can compile click, type, navigate, and assert intents into Playwright actions

Executor can compile click, type, navigate, and assert intents into Playwright actions *(action bridge validated; real Playwright execution deferred)*

Exit code 1 on test failure, exit code 2 on auth/infra error

FileStorageAdapter stores and retrieves files from R2 via S3-compatible API

Firecracker VM image boots and runs a Playwright script successfully

Firecracker VM image boots and runs a Playwright script successfully *(deferred — requires bare-metal KVM server)*

Fixture adapter returns deterministic intent sequences matching v2 contract

GitHub Actions example workflow runs successfully in a test repository

Hono server starts and serves all v1 routes

Iframe elements appear in the flattened semantic snapshot with frame metadata

JobQueueAdapter can enqueue, dequeue, and acknowledge jobs through BullMQ

Journey CRUD operations persist to and read from Postgres

JSON output contains all journey results with run IDs and replay URLs

JUnit XML output is parseable by GitHub Actions test-reporter and GitLab CI

Migrations run successfully against a Neon Postgres instance

Migrations run successfully against a Neon Postgres instance *(deferred to integration)*

No regressions in previous phase tests

OpenAPI spec is generated and accessible at /api/v1/openapi.json

Owned benchmark corpus (Altitude, Switchboard, Foundry) runs with fixture planner

Parallel submission respects concurrency limit and tenant quota

Per-journey timeout surfaces as an explicit failure in all reporters

Per-run budget enforcement aborts a journey when cumulative tokens exceed the limit

Per-step telemetry accurately reports tokens, latency, and model version from Claude API

PlannerAdapterV2 contract is defined and exported from packages/contracts/

Playwright trace.zip is captured and uploaded alongside Automium event stream

Priority ordering works correctly (high > normal > low)

Provider retry handles 429 rate limits with exponential backoff

Queue and worker tests pass with a local Redis instance

RBAC middleware enforces role-based access per the permission matrix

RealtimeTransportAdapter can publish and subscribe to Redis channels

Replay artifacts are complete: event stream, targeted crops, network log, Playwright trace

Retention tags are set on upload and R2 lifecycle rules are configured

RLS policies prevent cross-tenant data access when session variables are set

Run status endpoint returns real-time status from Postgres

Run submission enqueues a job in BullMQ and returns a run ID

SearchBackendAdapter can index and query using Postgres tsvector/tsquery

Semantic enrichment produces stable element IDs that persist across page rerenders

Semantic enrichment produces stable element IDs that persist across page rerenders *(uses engine-domain stableHash)*

Targeted vision capture produces annotated element screenshots under budget

Targeted vision capture produces annotated element screenshots under budget *(budget enforcement validated; real screenshot capture deferred)*

Terminal output streams journey completion with pass/fail indicators

Two tenants can submit runs simultaneously with no cross-tenant data access

Vision crops are included as image content blocks when flagged by the runtime

WebSocket gateway streams run progress events from Redis pub/sub

Worker lease quota enforcement prevents over-allocation

Worker process dequeues a job and reports heartbeats via HTTP

WorkOS adapter authenticates via magic-link and validates session tokens

WorkOS authentication rejects unauthenticated requests with 401

Done35

All adapter implementations pass contract tests matching `packages/adapters/` interfaces

All phase tests pass

All phase tests pass (274/274)

All phase tests pass (309/309)

Audit events flow through the audit-sink queue to Postgres asynchronously *(queue definition exists; async flow deferred)*

AuditSinkAdapter can emit and query audit events through Postgres *(stub validated)*

BrowserRuntime interface is defined and Playwright adapter passes all interface methods *(stub shape validated; real Playwright wiring deferred)*

Credential vault can store, retrieve, and rotate encrypted secrets *(encrypt/decrypt validated)*

Drizzle schema compiles and generates migrations for all core tables

Executor can compile click, type, navigate, and assert intents into Playwright actions *(action bridge validated; real Playwright execution deferred)*

Iframe elements appear in the flattened semantic snapshot with frame metadata

JobQueueAdapter can enqueue, dequeue, and acknowledge jobs through BullMQ *(stub shape validated by contract tests; BullMQ wiring deferred)*

No regressions in previous phase tests

Phase 1: Persistence Foundation (13 steps) — archived in `tasks/phases/production-phase-1.md`

Phase 2: Queue + Worker Infrastructure (8 steps) — archived in `tasks/phases/production-phase-2.md`

Priority ordering works correctly (high > normal > low)

Queue and worker tests pass with a local Redis instance *(contract tests pass without Redis; live Redis deferred to integration)*

RealtimeTransportAdapter can publish and subscribe to Redis channels *(stub shape validated; Redis wiring deferred)*

RLS policies prevent cross-tenant data access when session variables are set *(policies defined)*

SearchBackendAdapter can index and query using Postgres tsvector/tsquery *(stub validated)*

Step 3.1: **Automated** Write failing contract tests for all Phase 3 components.

Step 3.10: **Automated** Refactor pass — verify barrel exports, adapter alignment, no dead code.

Step 3.2: **Automated** Scaffold `packages/browser-runtime/` and define BrowserRuntime interface types.

Step 3.3: **Automated** Implement semantic enrichment pipeline — raw accessibility tree to contract-compliant interactive elements.

Step 3.4: **Automated** Implement frame hierarchy flattening into unified element list with frame metadata.

Step 3.5: **Automated** Implement targeted vision capture contracts and budget enforcement.

Step 3.6: **Automated** Implement executor-to-BrowserRuntime action bridge.

Step 3.7: **Automated** Implement full BrowserRuntime adapter stub with enrichment pipeline wiring.

Step 3.8: **Automated** Build contract-compliant SemanticSnapshot from enriched browser state.

Step 3.9: **Automated** Run all tests and verify they pass (green).

Worker lease quota enforcement prevents over-allocation *(existing orchestrator logic, now wired to dispatch)*

Worker process dequeues a job and reports heartbeats via HTTP *(stub lifecycle validated; real dequeue loop deferred)*

WorkOS adapter authenticates via magic-link and validates session tokens *(stub validated)*

Timeline

20 events

April 2026

chore

chore: complete browser runtime refactor pass

Apr 27static-

chore

chore: verify browser runtime green phase

Apr 27static-

feature

feat(browser-runtime): build contract semantic snapshots

Apr 26static-

feature

feat: wire browser runtime adapter stub

Apr 26static-

feature

feat(browser-runtime): add executor action bridge

Apr 26static-

feature

feat: add browser vision capture budgeting

Apr 26static-

feature

feat(browser-runtime): flatten frame hierarchy

Apr 26static-

feature

feat: add browser runtime semantic enrichment

Apr 26static-

docs

docs(tasks): mark browser runtime scaffold complete

Apr 26static-

feature

feat(browser-runtime): scaffold BrowserRuntime interface types and adapter stub (Step 3.2)

Apr 26static-

test

test(browser-runtime): add 32 red-phase contract tests for Phase 3 (Step 3.1)

Apr 26static-

docs

docs(tasks): plan Phase 3 Browser Runtime with 10 TDD steps

Apr 26static-

refactor

refactor(adapters-redis): add explicit return type, complete Phase 2 (Step 2.8)

Apr 26static-

docs

docs(tasks): mark Step 2.7 green verification complete, plan Step 2.8 (refactor pass)

Apr 26static-

feature

feat(orchestrator): add dispatchRun wiring lease-to-queue (Step 2.6)

Apr 26static-

feature

feat(worker): add WorkerProcess skeleton + HeartbeatReporter stubs (Step 2.5)

Apr 26static-

feature

feat(adapters-redis): add RealtimeTransportAdapter stub + Redis connection factory (Step 2.4)

Apr 26static-

feature

feat(adapters-bullmq): add queue definitions with priority ordering (Step 2.3)

Apr 26static-

docs

docs(tasks): mark Step 2.2 complete, plan Step 2.3 (queue definitions)

Apr 26static-

feature

feat(adapters-bullmq): scaffold JobQueueAdapter stub (Step 2.2)

Apr 26static-