Skip to content

USAGE.md

Latest commit

 

History

History
117 lines (79 loc) · 2.42 KB

USAGE.md

File metadata and controls

117 lines (79 loc) · 2.42 KB
Raw

Usage

Launch Fakeble from the project you want reviewed:

cd /path/to/project
fakeble .

Or pass the project path:

fakeble /path/to/project

Check setup at any time:

fakeble doctor

Repair setup:

fakeble install --yes

Prompt Template

Inside the Pi host pane:

/fakeble Compare the implementation options for this task. Claude should focus on ambiguity, architecture, and failure modes. Codex should focus on repo structure, tests, edge cases, and implementation realism. Cross-review before synthesis.

You can also use plain language:

Run Fakeble on this migration plan. I want independent Claude and Codex reads, a disagreement phase, and a final recommendation with risks and verification steps.

Good Uses

  • architecture reviews
  • migration plans
  • incident diagnosis
  • high-risk implementation plans
  • design or product decisions where two independent reads are useful
  • code review before an expensive change

Bad Uses

  • one-line shell help
  • tiny factual questions
  • mechanical edits
  • anything where a single direct agent turn is obviously enough

Tmux Basics

Detach without stopping:

Ctrl-b d

List sessions:

tmux list-sessions | rg '^salon-'

Reattach:

tmux attach -t salon-SESSION

Kill a stuck session:

tmux kill-session -t salon-SESSION

Model And Effort Overrides

SALON_HOST_MODEL=openai-codex/gpt-5.4-mini \
SALON_GUEST_EFFORT=low \
SALON_CODEX_EFFORT=low \
fakeble .

For bigger reviews, leave defaults alone or set higher effort for guests.

Editing Advice

For risky repos, start with review-only prompts:

Do not edit files. Inspect independently, cross-review, then return findings and a recommended implementation plan.

If you want agents to edit, use a branch or worktree and keep the scope narrow.

When An Assistant Runs It For You

If you ask Codex or Claude to launch Fakeble, it should stay with the panel until it reaches final synthesis, final review, or a real auth/trust/MFA blocker.

A good assistant should:

  • watch the host pane every 30-60 seconds
  • confirm each guest received the task and answered
  • nudge a Codex pane stuck on [Pasted Content ...] by sending Enter
  • move the host from independent reads to cross-review to synthesis when guests are done
  • keep you updated during long runs

That is part of the Fakeble workflow, not optional ceremony.