Forge Getting Started Beta

Run your first AI Agent workflow in 5 minutes. This guide pairs with the Beta invite — registering grants you a sample dataset / rubric / workflow, so following these 5 steps gives you a feel for what Forge looks like.

Beta Scope

This is currently an invite-only private beta with 10-15 early users. For trial feedback, see §5 What to Do When You Hit a Problem at the end of this guide.

§1 Meet Forge in 30 Seconds

The AI Agent workbench: draw / run / evaluate Agent workflows in your browser, with automatic crash resume that never re-spends LLM tokens.

WAL-first durable execution

Powered by Kova (the Rust durable execution engine, crash recovery — not checkpointing, every LLM Directive is persisted to disk).

Zero external dependencies

A single runtime binary + a single WAL file, no Kafka / Redis / Cassandra required.

OpenAI-compatible gateway

LLMs go through the newapi gateway, switchable across OpenAI / Anthropic / DeepSeek / Tongyi / GLM.

Every step auditable

Every step and human approval signature is persisted to disk, meeting EU AI Act + GB/T compliance requirements.

§2 Run Your First Workflow

Prerequisites

You have received a Beta invite and completed registration and login at forge.lurus.cn.

Open /workflows/runs and click "Start New Run".
Choose the seed classify_then_route_v1, type Chinese into the input box (e.g. 今天上海天气怎么样), and click Start.
The page redirects to /workflows/runs/[id], and the timeline cards refresh in real time (passthrough → llm_call → branch → leaf, four steps). Expect < 30 seconds (when newapi.lurus.cn is online).

LLM slow / failed

When the LLM times out / fails, the run status changes to failed and shows the error — this is Kova WAL crash recovery at work, and you can later resume without re-running the earlier steps.

§3 Mid-Workflow Approval Node (HITL)

When a workflow contains an await_input step (such as the "request approval before high-risk operations" template):

It pauses when it reaches that step, and the status changes to AwaitingInput.
/approvals shows a pending row (the title is that step's prompt).
Click "Review", choose Approve / Reject / Edit and submit, and the workflow resumes automatically.

The approval decision is written into the WAL and is permanently traceable; refreshing / closing the tab does not lose state.

mermaid

sequenceDiagram
    participant U as 你 (操作员)
    participant F as Forge UI
    participant K as Kova 引擎
    participant W as WAL 文件

    U->>F: 启动 workflow
    F->>K: POST /workflows/start
    K->>W: 写 WorkflowStart (event=37)
    loop 每步
        K->>W: 写 StepCheckpoint
    end
    K->>W: 写 WorkflowAwaitInput (event=38)
    K-->>F: 状态 = AwaitingInput
    F-->>U: /approvals 出现一行待审
    U->>F: 点 Approve
    F->>K: POST /workflows/:id/resume
    K->>W: 写 WorkflowInputReceived (event=39)
    K->>W: 继续 StepCheckpoint
    K-->>F: 状态 = Completed

§4 Score (Eval) a Run

Open /eval → the Rubrics tab, and choose the seed Sample rubric (PII) or build your own.
Switch to Runs and associate the workflow_id you just finished running.
Click Score; the scorer runs in the background, and you see each criterion's score + explanation.

Available scorer types

Type	Purpose	Configuration
`pii_regex`	Detect whether the LLM output leaks ID numbers / phone numbers / emails	Write a regex pattern
`json_schema`	Check the output conforms to a JSON schema (structured generation scenarios)	Paste a JSON schema
`llm_as_judge`	Have another LLM score the main LLM's output	Write a judge prompt + pick a model + temperature
`semantic_similarity`	(WIP, not yet available — the embedding service is still being built)	—

§5 What to Do When You Hit a Problem

The workflow stays Running and never moves?

It's most likely an LLM gateway timeout (30 s). Check the last step on the /workflows/runs/[id] timeline cards; if it's llm_call, wait or cancel the run and retry.

403 You do not have permission?

You are trying to act on someone else's approval. Only the originator themselves or someone in the same tenant_id can decide — find the originator.

404 Approval not found?

The approval has been canceled or is in a terminal state. Confirm with the originator; terminal states cannot be changed.

/workflows/runs keeps loading?

kova_proxy can't reach kova-rest. Check /api/health and see whether the second segment, kova_rest, is ok.

Chinese shows as garbled text / not translated?

An i18n key is missing. Send feedback + a screenshot (see §6 Feedback).

§6 Feedback

Found a bug, want a new feature, or want a 30-minute chat with us about your use case:

Typeform form

Embedded at the bottom of the /settings page — fills out in 30 seconds, fastest for external users.

Discord

Invite link is in the footer, the first choice for developer-facing users.

forge-beta@lurus.cn, with a reply SLA within 24h.

During the Beta, all feedback goes straight into the roadmap. We look forward to your use.

Next Steps

Forge intro page — its place within the Lurus platform Kova engine docs — details of the underlying durable execution engine Forge Roadmap — what we are building next

Last updated: 2026-05-12 | Companion Beta invite playbook: 2b-bs-forge/docs/beta-invite-playbook.md

Forge Getting Started Beta ​

§1 Meet Forge in 30 Seconds ​

§2 Run Your First Workflow ​

§3 Mid-Workflow Approval Node (HITL) ​

§4 Score (Eval) a Run ​

§5 What to Do When You Hit a Problem ​

§6 Feedback ​