Skip to content
Agent Dev Lab Agent Dev Lab

Inspection UI

The inspection UI (@agent-dev-lab/web) is how you start, watch, and replay runs for an ADL project. It does not execute workflows itself: it calls loadAdlProject(), then workflow.run / agent.run, and tails persisted RunEvents.

A workflow run's waterfall of steps, with a step's output selected in the inspector panel

Open the inspector

Section titled “Open the inspector”

From a project with adl.config.*:

Terminal window
bunx adl dashboard
bunx adl dashboard --serve
bunx adl dashboard --project ../other-research

The header shows the project name. Plain adl dashboard watches the project and hot-reloads agent/workflow/template edits; --serve runs the Nitro build shipped in @agent-dev-lab/web and does not watch — restart it after any change. Either way, .env* edits need a restart. --project points at another directory that contains adl.config.*.

Standalone CLI commands (adl workflow run, adl agent run, adl workflow list, etc.) are separate processes: they load the project once and exit.

Workflows

Section titled “Workflows”

Registered ids come from adl.config workflows. The sidebar lists startable workflows and past runs (title from ctx.setTitle when set).

  1. Open a workflow and start a run. If the workflow has a Zod input schema, the start dialog builds a form from it (defaults apply).
  2. The run page is a waterfall: steps, nested steps, parallel keyed steps, and agent episodes.
  3. Select a step or agent call for output, errors, and the conversation transcript for that memoryScope.
  4. Cancel calls handle.cancel(), which aborts ctx.signal, in-flight ctx.step bodies, and child agent.run / streamText calls on that run. Isolated helper runs (for example conversation titleWorkflow) are not cancelled with the parent.

Live updates use SSE (GET /api/runs/:runId/events?afterSeq=). Reconnects replay from the last applied runSeq. History is always the SQLite (or in-memory) WorkflowStore, so you can reopen a finished run later.

Helpers that you do not put in workflows: [] still persist if they run(), but they do not appear as startable targets or in the default run list. Conversation titleWorkflow uses { isolated: true } so naming a chat does not inject steps into another workflow’s tree.

Agents

Section titled “Agents”

Registered agents can be opened as conversations (standalone memoryScopes), not only as nodes inside a workflow.

  • New chat → agent.run() (AI SDK stopWhen, default stepCountIs(20)); first successful turn may set a title via titleWorkflow. Tool call/result events still fire while the model works.
  • Fork from a workflow agent episode copies that transcript into a new conversation you can continue.
  • Shared scopes show history up to the selected episode; later turns are muted so you can see what the model had at that call.
  • The agent settings panel reports effective model (id + provider when the LanguageModel exposes them), memory backend kind (sqlite / in-memory / custom), tools, stopWhen (default / custom), and title workflow id.

Event log

Section titled “Event log”

The Event log page (/events) is a process-wide tail of every RunEvent the inspector has seen — workflow runs and standalone agent conversations — not one run at a time.

  • Open it from the home sidebar or the rail. The context sidebar is hidden so the table can use the full width.
  • Live updates use SSE (GET /api/events?afterSeq=). The stream id is the process logSeq, not per-run runSeq. Reconnects replay from the last applied logSeq.
  • On inspector start, an empty in-memory buffer is hydrated from the last 100 persisted workflow runs and standalone agent episodes in WorkflowStore. Clear empties the in-memory view only; persisted runs stay. Restart hydrates again.
  • Default filters hide agent_text_delta. Add field clauses (equals / not-equals / contains on strings / exists / empty). Click a name to open that run, conversation, call, or step — conversations highlight the matching transcript slice (?call=). Right-click a value to filter; ⋯ opens the full JSON payload.
  • The footer paginates the filtered list (oldest page first; page 1 is the live tail when you stay on it).

The log is a ring buffer (default 10_000 events). It is not a durable store of its own — durability is still WorkflowStore.

What is not in the inspector yet

Section titled “What is not in the inspector yet”
  • A template playground (edit/render createTemplate markdown in the UI)
  • A dedicated raw token-debug pane (assistant text already streams via agent_text_delta in chat/run views)
  • Manual runs of adl.config.tools (registry-only today; runtime merge is createAdlRuntime({ tools }))