Symptom

In the mm init default interactive path (and --preset X), typing b at the MCP step (_step_mcp) should conceptually take the user back to the memory-directory step. Instead it lands on _step_provider_dirs_auto, which is a silent step (prints a "No provider memory folders detected." banner, or the "Auto-added N folder(s)" banner, but does not prompt), so it runs its click.echo calls and falls straight through to the next iteration — dropping the user back at _step_mcp with just a re-printed banner in between.

Net effect: "b" at _step_mcp appears to do almost nothing. The user has to hit "b" twice in a row to reach a prompt that can actually accept input, which no wizard documents.

Same shape as #371 but one step further along the flow.

Repro (assuming #418 lands)

$ mm init
  Choose setup style: [2]
  Directory [~/memories]: /tmp/m
  'tmp/m' doesn't exist. Create it? [Y/n] y
  No provider memory folders detected (checked: ...)
  Run 'mm init --advanced' later to add custom paths.

1. MCP Configuration     # (or whatever step_header shows for mcp)
  (b: back, q: quit)
  Select [2]: b          # user hits b expecting to go back to memory-dir

  No provider memory folders detected (checked: ...)   # banner re-prints
  Run 'mm init --advanced' later to add custom paths.

1. MCP Configuration    # same prompt reappears, no actual back-nav
  (b: back, q: quit)
  Select [2]:

Root cause

packages/memtomem/src/memtomem/cli/wizard.py:run_steps:

while i < len(steps):
    try:
        steps[i](state)
        i += 1
    except StepBack:
        if i > 0:
            i -= 1
            ...

run_steps decrements i and re-runs steps[i], but it doesn't distinguish "user hit b, went back one" from "normal forward step". If steps[i] is silent (no prompt, just echo), it runs its side effect and i += 1 advances right back to where we came from — user never sees a prompt they can act on.

Suggested directions

1. Mark silent steps explicitly. Give a step function a marker (attribute, decorator, or separate list) saying "on back-nav pass-through, not present". run_steps on StepBack keeps decrementing until it hits a non-silent step.
2. Teach silent steps to skip their side effects when visited during back-nav. Pass a direction hint from run_steps (_from_back=True) and steps can short-circuit. More plumbing.
3. Fold silent steps into adjacent interactive ones. _step_provider_dirs_auto's side effect could run inside _step_memory_dir's tail, or inside a wrapper around _step_mcp's head. Removes the silent step entirely.

Direction 1 is the narrowest fix and keeps step composability. Direction 3 would be cleaner but may complicate the advanced-flow case where provider_dirs is also a step.

Acceptance criteria

• b at _step_mcp lands the user back at the previous prompt (memory-dir), not on a silent step.
• Silent step side effects (banners) don't re-fire on a back-pass-through.
• New test that sequences CliRunner().invoke(init, input="<preset>\n<dir>\ny\nb\n...") and asserts the memory-dir prompt re-appears in the output.

Related

• mm init: "b" does not work at step 3 (Memory Directory) #371 / fix(init): "b" at Memory Directory navigates back to preset picker (#371) #418 — same UX family, one step earlier.