LEX AI Docs Hub: Mode Reference

The ten canonical LEX AI mode keys, intent-to-mode selection rules, the kickstart and finalize tool for each mode, and how the non-step-loop shapes differ.

A
Written By Atakan KayaLast updated 3 months ago

LEX AI Docs Hub: Mode Reference

Canonical mode keys, as accepted by lex-mcp --mode and by switch_to_mode: brief, forward, backward, edit, review, test, input, deploy, mvp_generator, mvp_completion.

Mode selection rules

  • Requirements not yet written down -> brief
  • New app from scratch -> forward
  • Initial MVP only -> mvp_generator
  • Finish or expand an MVP -> mvp_completion
  • Existing app feature change -> edit
  • Audit or review request -> review
  • Reverse documentation request -> backward
  • Test suite missing or untrusted -> test
  • Input-data format changed -> input
  • Get the app running in an environment -> deploy

Session tools per mode

Mode

Kickstart

Finalize

brief

kickstart_brief

finalize_brief

forward

kickstart_workflow (new) or kickstart_run (existing)

finalize_workflow

backward

reverse_kickstart

finalize_reverse

edit

kickstart_edit

finalize_edit

review

kickstart_review

finalize_review

test

kickstart_test_run

finalize_test_run

input

kickstart_input_change

finalize_input_change

deploy

kickstart_deployment

finalize_deployment

mvp_generator

kickstart_mvp

finalize_mvp

mvp_completion

kickstart_run

finalize_workflow

Forward lifecycle

  • Kickstart with kickstart_workflow for a new repo or kickstart_run for an existing one. Both read .lex/contract.md when Brief mode has produced one.
  • Execute the step loop with get_plan_step -> step agent -> notify_step_complete.
  • Finalize in two stages: audit instructions first, then the completion call once plans/technical_docs/audit-report.md exists on disk.

Shapes other than the step loop

  • Edit, review, test, and input are flat catalogues rather than 20-step coordinator loops. They finalize into a consolidated report.
  • Backward is documentation-focused and delivers no code changes.
  • MVP generator is checkpoint-gated with explicit scope caps and local git only.
  • Deploy is sequential but has no fixed local step count: the step bodies are served by the remote deployment handbook, and synthesising one locally is not permitted.
  • MVP completion reuses the forward machinery but covers steps 9 to 19; get_plan_step refuses any step below 9.

See Mode chooser and Mode switching.

Was this helpful?

Your feedback shapes what we write next.