Skip to main content

Flux-Lang workbench

Every Flux code block can expand into the same workbench used by /console/: a Monaco editor with Flux syntax, live diagnostics, completion, hover, formatting, structural graph projection, input editing, run output, cancellation, and effect-bound approval prompts.

The complete site and runtime ship inside the CLI:

flux docs
flux docs --model openai/gpt-5.6   # model-backed examples resolve this lazily
# docs:      http://127.0.0.1:8788/flux/
# workbench: http://127.0.0.1:8788/console/

Open the exact URL printed by the command. Its fragment contains a one-use launch secret; the page exchanges that locally for an HttpOnly, SameSite cookie and clears the fragment. Runtime and LSP routes reject other origins and sessions.

What can run​

Run appears only on examples with a server-known fixture. The examples page currently enables summarize-readme, latest-release, cached-page, wait-for-artifact, and rust-files. The first-app tutorial also exposes Part A and Part B as persistent app sessions with the handbook files from the earlier lessons. App messages reuse the same session; editing the program or switching variants starts a fresh one.

Other Flux blocks still get the editor and checker. They remain non-runnable when they rely on an undefined operation, a real Git checkout, an intentionally abridged production flow, or an external integration the tutorial cannot safely synthesize.

Safety boundary​

Each runnable block gets a private scratch project retained for its browser session. Its manifest limits both fixture files and reachable operations; the runtime performs a graph/risk preflight and then dispatches every operation through ordinary authorization, approval, redaction, and guarded IO. Approval decisions include the pending effect fingerprint, so an edit or different call cannot reuse one. Shell, plugins, the directory where you launched Flux, sub-agents, host secrets, and private network access are absent.

flux docs --bind 0.0.0.0:8788 deliberately changes posture. A non-loopback listener does not construct or mount the executor, scratch, approval, app, or LSP services; it serves static docs and structural projection only. The hosted public website similarly provides syntax-aware editing but has no backend and cannot execute.