Mikie / Field notes

Shadow mode: see what Mikie would save before you change anything

Shadow mode runs Mikie alongside Claude Code or Codex without changing what the agent sees. You measure Mikie on your own work first: a labelled projection in Codex, an unchanged session to compare against in Claude Code.

How-to6 min read
Select relevant code from a repository, then use it for a focused coding task.REPOSITORYFOCUSED TASK
Less to read. More to work with.

What shadow mode does

Shadow mode is available for both Claude Code and Codex. In either one your agent works exactly as it does today.

In Codex, Mikie plans the context packet it would have built for each request, but does not hand it to your agent. Mikie records what the agent actually used next to what the packet would have needed, so you can compare the two on real tasks in your real repository.

Turn it on for Codex

With Mikie installed and set up for Codex, switch modes from your terminal:

bash
mikie mode codex shadow   # turn shadow mode on
mikie mode codex          # check which mode you are in
mikie mode codex active   # switch back when you are ready

Turn it on for Claude Code

Claude Code has shadow mode too. In it Mikie's hooks change nothing, and the session is recorded as one Mikie did not change, so you can compare it with the sessions Mikie worked in.

bash
mikie mode claude shadow   # turn shadow mode on
mikie mode claude          # check which mode you are in
mikie mode claude active   # switch back when you are ready

In either agent you can also say "turn on Mikie shadow mode" or "turn off Mikie shadow mode".

Reading the results

Shadow results appear in your Meter dashboard as what Mikie would have saved if it had been running live, marked as projected. You can also see your history from the terminal:

bash
mikie savings history

A projection is not a saving. Mikie keeps the two apart on purpose: planning-only shadow output never counts as realized savings, and nothing in your dashboard turns an estimate into a measured number.

When to switch to active

Try a small, representative task before changing your normal workflow. A useful projection is a reason to test active mode; accepted work and complete usage evidence are how you judge the result.

Choose a task you can judge twice

Use a contained bug fix with a reproducible failure, not a vague request to improve the whole application. Record the expected behavior before the session so a promising projection cannot change your definition of success.

Task checklist
Task: one reproducible bug
Acceptance: the failing case passes; existing behavior still passes
Capture: mode, model, task outcome, usage label, failures
Decision: inspect the projection; confirm active results separately

If the projection is missing or surprising

Check the selected mode and run mikie doctor before changing the task. If a usage value is unavailable, keep it unavailable. A missing projection is not a zero-saving result, and a large projection does not establish that the proposed context contains everything the fix needs.

  • Check the device and project connection before repeating a paid task.
  • Separate shadow projections from active usage in your notes.
  • If quality drops in active mode, inspect missing context before trying a more aggressive reduction.

Sources: Commands are from the installed Mikie CLI's public help. Shadow projections are labelled as projected in Meter and are never reported as realized savings.

Try it on your own repository.

Give your agent the context it needs, then see how your next session compares.